linux/net/core/dev.c

6684 lines
162 KiB
C
Raw Normal View History

/*
* NET3 Protocol independent device support routines.
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*
* Derived from the non IP parts of dev.c 1.0.19
* Authors: Ross Biro
* Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
* Mark Evans, <evansmp@uhura.aston.ac.uk>
*
* Additional Authors:
* Florian la Roche <rzsfl@rz.uni-sb.de>
* Alan Cox <gw4pts@gw4pts.ampr.org>
* David Hinds <dahinds@users.sourceforge.net>
* Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
* Adam Sulmicki <adam@cfar.umd.edu>
* Pekka Riikonen <priikone@poesidon.pspt.fi>
*
* Changes:
* D.J. Barrow : Fixed bug where dev->refcnt gets set
* to 2 if register_netdev gets called
* before net_dev_init & also removed a
* few lines of code in the process.
* Alan Cox : device private ioctl copies fields back.
* Alan Cox : Transmit queue code does relevant
* stunts to keep the queue safe.
* Alan Cox : Fixed double lock.
* Alan Cox : Fixed promisc NULL pointer trap
* ???????? : Support the full private ioctl range
* Alan Cox : Moved ioctl permission check into
* drivers
* Tim Kordas : SIOCADDMULTI/SIOCDELMULTI
* Alan Cox : 100 backlog just doesn't cut it when
* you start doing multicast video 8)
* Alan Cox : Rewrote net_bh and list manager.
* Alan Cox : Fix ETH_P_ALL echoback lengths.
* Alan Cox : Took out transmit every packet pass
* Saved a few bytes in the ioctl handler
* Alan Cox : Network driver sets packet type before
* calling netif_rx. Saves a function
* call a packet.
* Alan Cox : Hashed net_bh()
* Richard Kooijman: Timestamp fixes.
* Alan Cox : Wrong field in SIOCGIFDSTADDR
* Alan Cox : Device lock protection.
* Alan Cox : Fixed nasty side effect of device close
* changes.
* Rudi Cilibrasi : Pass the right thing to
* set_mac_address()
* Dave Miller : 32bit quantity for the device lock to
* make it work out on a Sparc.
* Bjorn Ekwall : Added KERNELD hack.
* Alan Cox : Cleaned up the backlog initialise.
* Craig Metz : SIOCGIFCONF fix if space for under
* 1 device.
* Thomas Bogendoerfer : Return ENODEV for dev_open, if there
* is no device open function.
* Andi Kleen : Fix error reporting for SIOCGIFCONF
* Michael Chastain : Fix signed/unsigned for SIOCGIFCONF
* Cyrus Durgin : Cleaned for KMOD
* Adam Sulmicki : Bug Fix : Network Device Unload
* A network device unload needs to purge
* the backlog queue.
* Paul Rusty Russell : SIOCSIFNAME
* Pekka Riikonen : Netdev boot-time settings code
* Andrew Morton : Make unregister_netdevice wait
* indefinitely on dev->refcnt
* J Hadi Salim : - Backlog queue sampling
* - netif_rx() feedback
*/
#include <asm/uaccess.h>
#include <asm/system.h>
#include <linux/bitops.h>
#include <linux/capability.h>
#include <linux/cpu.h>
#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/hash.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 08:04:11 +00:00
#include <linux/slab.h>
#include <linux/sched.h>
#include <linux/mutex.h>
#include <linux/string.h>
#include <linux/mm.h>
#include <linux/socket.h>
#include <linux/sockios.h>
#include <linux/errno.h>
#include <linux/interrupt.h>
#include <linux/if_ether.h>
#include <linux/netdevice.h>
#include <linux/etherdevice.h>
#include <linux/ethtool.h>
#include <linux/notifier.h>
#include <linux/skbuff.h>
#include <net/net_namespace.h>
#include <net/sock.h>
#include <linux/rtnetlink.h>
#include <linux/proc_fs.h>
#include <linux/seq_file.h>
#include <linux/stat.h>
#include <net/dst.h>
#include <net/pkt_sched.h>
#include <net/checksum.h>
#include <net/xfrm.h>
#include <linux/highmem.h>
#include <linux/init.h>
#include <linux/kmod.h>
#include <linux/module.h>
#include <linux/netpoll.h>
#include <linux/rcupdate.h>
#include <linux/delay.h>
#include <net/wext.h>
#include <net/iw_handler.h>
#include <asm/current.h>
#include <linux/audit.h>
#include <linux/dmaengine.h>
#include <linux/err.h>
#include <linux/ctype.h>
#include <linux/if_arp.h>
#include <linux/if_vlan.h>
#include <linux/ip.h>
#include <net/ip.h>
#include <linux/ipv6.h>
#include <linux/in.h>
#include <linux/jhash.h>
#include <linux/random.h>
#include <trace/events/napi.h>
#include <trace/events/net.h>
#include <trace/events/skb.h>
#include <linux/pci.h>
#include <linux/inetdevice.h>
#include <linux/cpu_rmap.h>
#include <linux/if_tunnel.h>
#include <linux/if_pppox.h>
#include <linux/ppp_defs.h>
#include <linux/net_tstamp.h>
#include "net-sysfs.h"
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
/* Instead of increasing this, you should create a hash table. */
#define MAX_GRO_SKBS 8
/* This should be increased if a protocol with a bigger head is added. */
#define GRO_MAX_HEAD (MAX_HEADER + 128)
/*
* The list of packet types we will receive (as opposed to discard)
* and the routines to invoke.
*
* Why 16. Because with 16 the only overlap we get on a hash of the
* low nibble of the protocol value is RARP/SNAP/X.25.
*
* NOTE: That is no longer true with the addition of VLAN tags. Not
* sure which should go first, but I bet it won't make much
* difference if we are running VLANs. The good news is that
* this protocol won't be in the list unless compiled in, so
* the average user (w/out VLANs) will not be adversely affected.
* --BLG
*
* 0800 IP
* 8100 802.1Q VLAN
* 0001 802.3
* 0002 AX.25
* 0004 802.2
* 8035 RARP
* 0005 SNAP
* 0805 X.25
* 0806 ARP
* 8137 IPX
* 0009 Localtalk
* 86DD IPv6
*/
#define PTYPE_HASH_SIZE (16)
#define PTYPE_HASH_MASK (PTYPE_HASH_SIZE - 1)
static DEFINE_SPINLOCK(ptype_lock);
static struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
static struct list_head ptype_all __read_mostly; /* Taps */
/*
* The @dev_base_head list is protected by @dev_base_lock and the rtnl
* semaphore.
*
* Pure readers hold dev_base_lock for reading, or rcu_read_lock()
*
* Writers must hold the rtnl semaphore while they loop through the
* dev_base_head list, and hold dev_base_lock for writing when they do the
* actual updates. This allows pure readers to access the list even
* while a writer is preparing to update it.
*
* To put it another way, dev_base_lock is held for writing only to
* protect against pure readers; the rtnl semaphore provides the
* protection against other writers.
*
* See, for example usages, register_netdevice() and
* unregister_netdevice(), which must be called with the rtnl
* semaphore held.
*/
DEFINE_RWLOCK(dev_base_lock);
EXPORT_SYMBOL(dev_base_lock);
static inline void dev_base_seq_inc(struct net *net)
{
while (++net->dev_base_seq == 0);
}
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
{
unsigned hash = full_name_hash(name, strnlen(name, IFNAMSIZ));
return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
}
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
{
return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
}
static inline void rps_lock(struct softnet_data *sd)
{
#ifdef CONFIG_RPS
spin_lock(&sd->input_pkt_queue.lock);
#endif
}
static inline void rps_unlock(struct softnet_data *sd)
{
#ifdef CONFIG_RPS
spin_unlock(&sd->input_pkt_queue.lock);
#endif
}
/* Device list insertion */
static int list_netdevice(struct net_device *dev)
{
struct net *net = dev_net(dev);
ASSERT_RTNL();
write_lock_bh(&dev_base_lock);
list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
hlist_add_head_rcu(&dev->index_hlist,
dev_index_hash(net, dev->ifindex));
write_unlock_bh(&dev_base_lock);
dev_base_seq_inc(net);
return 0;
}
/* Device list removal
* caller must respect a RCU grace period before freeing/reusing dev
*/
static void unlist_netdevice(struct net_device *dev)
{
ASSERT_RTNL();
/* Unlink dev from the device chain */
write_lock_bh(&dev_base_lock);
list_del_rcu(&dev->dev_list);
hlist_del_rcu(&dev->name_hlist);
hlist_del_rcu(&dev->index_hlist);
write_unlock_bh(&dev_base_lock);
dev_base_seq_inc(dev_net(dev));
}
/*
* Our notifier list
*/
static RAW_NOTIFIER_HEAD(netdev_chain);
/*
* Device drivers call our routines to queue packets here. We empty the
* queue in the local softnet handler.
*/
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
EXPORT_PER_CPU_SYMBOL(softnet_data);
#ifdef CONFIG_LOCKDEP
/*
* register_netdevice() inits txq->_xmit_lock and sets lockdep class
* according to dev->type
*/
static const unsigned short netdev_lock_type[] =
{ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
ARPHRD_FCFABRIC, ARPHRD_IEEE802_TR, ARPHRD_IEEE80211,
ARPHRD_IEEE80211_PRISM, ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET,
ARPHRD_PHONET_PIPE, ARPHRD_IEEE802154,
ARPHRD_VOID, ARPHRD_NONE};
static const char *const netdev_lock_name[] =
{"_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
"_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
"_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
"_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
"_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
"_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
"_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
"_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
"_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
"_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
"_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
"_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
"_xmit_FCFABRIC", "_xmit_IEEE802_TR", "_xmit_IEEE80211",
"_xmit_IEEE80211_PRISM", "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET",
"_xmit_PHONET_PIPE", "_xmit_IEEE802154",
"_xmit_VOID", "_xmit_NONE"};
static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
static inline unsigned short netdev_lock_pos(unsigned short dev_type)
{
int i;
for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
if (netdev_lock_type[i] == dev_type)
return i;
/* the last key is used by default */
return ARRAY_SIZE(netdev_lock_type) - 1;
}
static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
unsigned short dev_type)
{
int i;
i = netdev_lock_pos(dev_type);
lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
netdev_lock_name[i]);
}
static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
{
int i;
i = netdev_lock_pos(dev->type);
lockdep_set_class_and_name(&dev->addr_list_lock,
&netdev_addr_lock_key[i],
netdev_lock_name[i]);
}
#else
static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
unsigned short dev_type)
{
}
static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
{
}
#endif
/*******************************************************************************
Protocol management and registration routines
*******************************************************************************/
/*
* Add a protocol ID to the list. Now that the input handler is
* smarter we can dispense with all the messy stuff that used to be
* here.
*
* BEWARE!!! Protocol handlers, mangling input packets,
* MUST BE last in hash buckets and checking protocol handlers
* MUST start from promiscuous ptype_all chain in net_bh.
* It is true now, do not change it.
* Explanation follows: if protocol handler, mangling packet, will
* be the first on list, it is not able to sense, that packet
* is cloned and should be copied-on-write, so that it will
* change it and subsequent readers will get broken packet.
* --ANK (980803)
*/
static inline struct list_head *ptype_head(const struct packet_type *pt)
{
if (pt->type == htons(ETH_P_ALL))
return &ptype_all;
else
return &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
}
/**
* dev_add_pack - add packet handler
* @pt: packet type declaration
*
* Add a protocol handler to the networking stack. The passed &packet_type
* is linked into kernel lists and may not be freed until it has been
* removed from the kernel lists.
*
* This call does not sleep therefore it can not
* guarantee all CPU's that are in middle of receiving packets
* will see the new packet type (until the next received packet).
*/
void dev_add_pack(struct packet_type *pt)
{
struct list_head *head = ptype_head(pt);
spin_lock(&ptype_lock);
list_add_rcu(&pt->list, head);
spin_unlock(&ptype_lock);
}
EXPORT_SYMBOL(dev_add_pack);
/**
* __dev_remove_pack - remove packet handler
* @pt: packet type declaration
*
* Remove a protocol handler that was previously added to the kernel
* protocol handlers by dev_add_pack(). The passed &packet_type is removed
* from the kernel lists and can be freed or reused once this function
* returns.
*
* The packet type might still be in use by receivers
* and must not be freed until after all the CPU's have gone
* through a quiescent state.
*/
void __dev_remove_pack(struct packet_type *pt)
{
struct list_head *head = ptype_head(pt);
struct packet_type *pt1;
spin_lock(&ptype_lock);
list_for_each_entry(pt1, head, list) {
if (pt == pt1) {
list_del_rcu(&pt->list);
goto out;
}
}
printk(KERN_WARNING "dev_remove_pack: %p not found.\n", pt);
out:
spin_unlock(&ptype_lock);
}
EXPORT_SYMBOL(__dev_remove_pack);
/**
* dev_remove_pack - remove packet handler
* @pt: packet type declaration
*
* Remove a protocol handler that was previously added to the kernel
* protocol handlers by dev_add_pack(). The passed &packet_type is removed
* from the kernel lists and can be freed or reused once this function
* returns.
*
* This call sleeps to guarantee that no CPU is looking at the packet
* type after return.
*/
void dev_remove_pack(struct packet_type *pt)
{
__dev_remove_pack(pt);
synchronize_net();
}
EXPORT_SYMBOL(dev_remove_pack);
/******************************************************************************
Device Boot-time Settings Routines
*******************************************************************************/
/* Boot time configuration table */
static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
/**
* netdev_boot_setup_add - add new setup entry
* @name: name of the device
* @map: configured settings for the device
*
* Adds new setup entry to the dev_boot_setup list. The function
* returns 0 on error and 1 on success. This is a generic routine to
* all netdevices.
*/
static int netdev_boot_setup_add(char *name, struct ifmap *map)
{
struct netdev_boot_setup *s;
int i;
s = dev_boot_setup;
for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
memset(s[i].name, 0, sizeof(s[i].name));
strlcpy(s[i].name, name, IFNAMSIZ);
memcpy(&s[i].map, map, sizeof(s[i].map));
break;
}
}
return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
}
/**
* netdev_boot_setup_check - check boot time settings
* @dev: the netdevice
*
* Check boot time settings for the device.
* The found settings are set for the device to be used
* later in the device probing.
* Returns 0 if no settings found, 1 if they are.
*/
int netdev_boot_setup_check(struct net_device *dev)
{
struct netdev_boot_setup *s = dev_boot_setup;
int i;
for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
!strcmp(dev->name, s[i].name)) {
dev->irq = s[i].map.irq;
dev->base_addr = s[i].map.base_addr;
dev->mem_start = s[i].map.mem_start;
dev->mem_end = s[i].map.mem_end;
return 1;
}
}
return 0;
}
EXPORT_SYMBOL(netdev_boot_setup_check);
/**
* netdev_boot_base - get address from boot time settings
* @prefix: prefix for network device
* @unit: id for network device
*
* Check boot time settings for the base address of device.
* The found settings are set for the device to be used
* later in the device probing.
* Returns 0 if no settings found.
*/
unsigned long netdev_boot_base(const char *prefix, int unit)
{
const struct netdev_boot_setup *s = dev_boot_setup;
char name[IFNAMSIZ];
int i;
sprintf(name, "%s%d", prefix, unit);
/*
* If device already registered then return base of 1
* to indicate not to probe for this interface
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (__dev_get_by_name(&init_net, name))
return 1;
for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
if (!strcmp(name, s[i].name))
return s[i].map.base_addr;
return 0;
}
/*
* Saves at boot time configured settings for any netdevice.
*/
int __init netdev_boot_setup(char *str)
{
int ints[5];
struct ifmap map;
str = get_options(str, ARRAY_SIZE(ints), ints);
if (!str || !*str)
return 0;
/* Save settings */
memset(&map, 0, sizeof(map));
if (ints[0] > 0)
map.irq = ints[1];
if (ints[0] > 1)
map.base_addr = ints[2];
if (ints[0] > 2)
map.mem_start = ints[3];
if (ints[0] > 3)
map.mem_end = ints[4];
/* Add new entry to the list */
return netdev_boot_setup_add(str, &map);
}
__setup("netdev=", netdev_boot_setup);
/*******************************************************************************
Device Interface Subroutines
*******************************************************************************/
/**
* __dev_get_by_name - find a device by its name
* @net: the applicable net namespace
* @name: name to find
*
* Find an interface by name. Must be called under RTNL semaphore
* or @dev_base_lock. If the name is found a pointer to the device
* is returned. If the name is not found then %NULL is returned. The
* reference counters are not incremented so the caller must be
* careful with locks.
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
struct net_device *__dev_get_by_name(struct net *net, const char *name)
{
struct hlist_node *p;
struct net_device *dev;
struct hlist_head *head = dev_name_hash(net, name);
hlist_for_each_entry(dev, p, head, name_hlist)
if (!strncmp(dev->name, name, IFNAMSIZ))
return dev;
return NULL;
}
EXPORT_SYMBOL(__dev_get_by_name);
/**
* dev_get_by_name_rcu - find a device by its name
* @net: the applicable net namespace
* @name: name to find
*
* Find an interface by name.
* If the name is found a pointer to the device is returned.
* If the name is not found then %NULL is returned.
* The reference counters are not incremented so the caller must be
* careful with locks. The caller must hold RCU lock.
*/
struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
{
struct hlist_node *p;
struct net_device *dev;
struct hlist_head *head = dev_name_hash(net, name);
hlist_for_each_entry_rcu(dev, p, head, name_hlist)
if (!strncmp(dev->name, name, IFNAMSIZ))
return dev;
return NULL;
}
EXPORT_SYMBOL(dev_get_by_name_rcu);
/**
* dev_get_by_name - find a device by its name
* @net: the applicable net namespace
* @name: name to find
*
* Find an interface by name. This can be called from any
* context and does its own locking. The returned handle has
* the usage count incremented and the caller must use dev_put() to
* release it when it is no longer needed. %NULL is returned if no
* matching device is found.
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
struct net_device *dev_get_by_name(struct net *net, const char *name)
{
struct net_device *dev;
rcu_read_lock();
dev = dev_get_by_name_rcu(net, name);
if (dev)
dev_hold(dev);
rcu_read_unlock();
return dev;
}
EXPORT_SYMBOL(dev_get_by_name);
/**
* __dev_get_by_index - find a device by its ifindex
* @net: the applicable net namespace
* @ifindex: index of device
*
* Search for an interface by index. Returns %NULL if the device
* is not found or a pointer to the device. The device has not
* had its reference counter increased so the caller must be careful
* about locking. The caller must hold either the RTNL semaphore
* or @dev_base_lock.
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
struct net_device *__dev_get_by_index(struct net *net, int ifindex)
{
struct hlist_node *p;
struct net_device *dev;
struct hlist_head *head = dev_index_hash(net, ifindex);
hlist_for_each_entry(dev, p, head, index_hlist)
if (dev->ifindex == ifindex)
return dev;
return NULL;
}
EXPORT_SYMBOL(__dev_get_by_index);
/**
* dev_get_by_index_rcu - find a device by its ifindex
* @net: the applicable net namespace
* @ifindex: index of device
*
* Search for an interface by index. Returns %NULL if the device
* is not found or a pointer to the device. The device has not
* had its reference counter increased so the caller must be careful
* about locking. The caller must hold RCU lock.
*/
struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
{
struct hlist_node *p;
struct net_device *dev;
struct hlist_head *head = dev_index_hash(net, ifindex);
hlist_for_each_entry_rcu(dev, p, head, index_hlist)
if (dev->ifindex == ifindex)
return dev;
return NULL;
}
EXPORT_SYMBOL(dev_get_by_index_rcu);
/**
* dev_get_by_index - find a device by its ifindex
* @net: the applicable net namespace
* @ifindex: index of device
*
* Search for an interface by index. Returns NULL if the device
* is not found or a pointer to the device. The device returned has
* had a reference added and the pointer is safe until the user calls
* dev_put to indicate they have finished with it.
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
struct net_device *dev_get_by_index(struct net *net, int ifindex)
{
struct net_device *dev;
rcu_read_lock();
dev = dev_get_by_index_rcu(net, ifindex);
if (dev)
dev_hold(dev);
rcu_read_unlock();
return dev;
}
EXPORT_SYMBOL(dev_get_by_index);
/**
* dev_getbyhwaddr_rcu - find a device by its hardware address
* @net: the applicable net namespace
* @type: media type of device
* @ha: hardware address
*
* Search for an interface by MAC address. Returns NULL if the device
* is not found or a pointer to the device.
* The caller must hold RCU or RTNL.
* The returned device has not had its ref count increased
* and the caller must therefore be careful about locking
*
*/
struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
const char *ha)
{
struct net_device *dev;
for_each_netdev_rcu(net, dev)
if (dev->type == type &&
!memcmp(dev->dev_addr, ha, dev->addr_len))
return dev;
return NULL;
}
EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
{
struct net_device *dev;
ASSERT_RTNL();
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
for_each_netdev(net, dev)
if (dev->type == type)
return dev;
return NULL;
}
EXPORT_SYMBOL(__dev_getfirstbyhwtype);
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
{
struct net_device *dev, *ret = NULL;
rcu_read_lock();
for_each_netdev_rcu(net, dev)
if (dev->type == type) {
dev_hold(dev);
ret = dev;
break;
}
rcu_read_unlock();
return ret;
}
EXPORT_SYMBOL(dev_getfirstbyhwtype);
/**
* dev_get_by_flags_rcu - find any device with given flags
* @net: the applicable net namespace
* @if_flags: IFF_* values
* @mask: bitmask of bits in if_flags to check
*
* Search for any interface with the given flags. Returns NULL if a device
* is not found or a pointer to the device. Must be called inside
* rcu_read_lock(), and result refcount is unchanged.
*/
struct net_device *dev_get_by_flags_rcu(struct net *net, unsigned short if_flags,
unsigned short mask)
{
struct net_device *dev, *ret;
ret = NULL;
for_each_netdev_rcu(net, dev) {
if (((dev->flags ^ if_flags) & mask) == 0) {
ret = dev;
break;
}
}
return ret;
}
EXPORT_SYMBOL(dev_get_by_flags_rcu);
/**
* dev_valid_name - check if name is okay for network device
* @name: name string
*
* Network device names need to be valid file names to
* to allow sysfs to work. We also disallow any kind of
* whitespace.
*/
int dev_valid_name(const char *name)
{
if (*name == '\0')
return 0;
if (strlen(name) >= IFNAMSIZ)
return 0;
if (!strcmp(name, ".") || !strcmp(name, ".."))
return 0;
while (*name) {
if (*name == '/' || isspace(*name))
return 0;
name++;
}
return 1;
}
EXPORT_SYMBOL(dev_valid_name);
/**
* __dev_alloc_name - allocate a name for a device
* @net: network namespace to allocate the device name in
* @name: name format string
* @buf: scratch buffer and result name string
*
* Passed a format string - eg "lt%d" it will try and find a suitable
* id. It scans list of devices to build up a free map, then chooses
* the first empty slot. The caller must hold the dev_base or rtnl lock
* while allocating the name and adding the device in order to avoid
* duplicates.
* Limited to bits_per_byte * page size devices (ie 32K on most platforms).
* Returns the number of the unit assigned or a negative errno code.
*/
static int __dev_alloc_name(struct net *net, const char *name, char *buf)
{
int i = 0;
const char *p;
const int max_netdevices = 8*PAGE_SIZE;
unsigned long *inuse;
struct net_device *d;
p = strnchr(name, IFNAMSIZ-1, '%');
if (p) {
/*
* Verify the string as this thing may have come from
* the user. There must be either one "%d" and no other "%"
* characters.
*/
if (p[1] != 'd' || strchr(p + 2, '%'))
return -EINVAL;
/* Use one page as a bit array of possible slots */
inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
if (!inuse)
return -ENOMEM;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
for_each_netdev(net, d) {
if (!sscanf(d->name, name, &i))
continue;
if (i < 0 || i >= max_netdevices)
continue;
/* avoid cases where sscanf is not exact inverse of printf */
snprintf(buf, IFNAMSIZ, name, i);
if (!strncmp(buf, d->name, IFNAMSIZ))
set_bit(i, inuse);
}
i = find_first_zero_bit(inuse, max_netdevices);
free_page((unsigned long) inuse);
}
if (buf != name)
snprintf(buf, IFNAMSIZ, name, i);
if (!__dev_get_by_name(net, buf))
return i;
/* It is possible to run out of possible slots
* when the name is long and there isn't enough space left
* for the digits, or if all bits are used.
*/
return -ENFILE;
}
/**
* dev_alloc_name - allocate a name for a device
* @dev: device
* @name: name format string
*
* Passed a format string - eg "lt%d" it will try and find a suitable
* id. It scans list of devices to build up a free map, then chooses
* the first empty slot. The caller must hold the dev_base or rtnl lock
* while allocating the name and adding the device in order to avoid
* duplicates.
* Limited to bits_per_byte * page size devices (ie 32K on most platforms).
* Returns the number of the unit assigned or a negative errno code.
*/
int dev_alloc_name(struct net_device *dev, const char *name)
{
char buf[IFNAMSIZ];
struct net *net;
int ret;
BUG_ON(!dev_net(dev));
net = dev_net(dev);
ret = __dev_alloc_name(net, name, buf);
if (ret >= 0)
strlcpy(dev->name, buf, IFNAMSIZ);
return ret;
}
EXPORT_SYMBOL(dev_alloc_name);
static int dev_get_valid_name(struct net_device *dev, const char *name)
{
struct net *net;
BUG_ON(!dev_net(dev));
net = dev_net(dev);
if (!dev_valid_name(name))
return -EINVAL;
if (strchr(name, '%'))
return dev_alloc_name(dev, name);
else if (__dev_get_by_name(net, name))
return -EEXIST;
else if (dev->name != name)
strlcpy(dev->name, name, IFNAMSIZ);
return 0;
}
/**
* dev_change_name - change name of a device
* @dev: device
* @newname: name (or format string) must be at least IFNAMSIZ
*
* Change name of a device, can pass format strings "eth%d".
* for wildcarding.
*/
int dev_change_name(struct net_device *dev, const char *newname)
{
char oldname[IFNAMSIZ];
int err = 0;
int ret;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
struct net *net;
ASSERT_RTNL();
BUG_ON(!dev_net(dev));
net = dev_net(dev);
if (dev->flags & IFF_UP)
return -EBUSY;
if (strncmp(newname, dev->name, IFNAMSIZ) == 0)
return 0;
memcpy(oldname, dev->name, IFNAMSIZ);
err = dev_get_valid_name(dev, newname);
if (err < 0)
return err;
rollback:
ret = device_rename(&dev->dev, dev->name);
if (ret) {
memcpy(dev->name, oldname, IFNAMSIZ);
return ret;
}
write_lock_bh(&dev_base_lock);
hlist_del_rcu(&dev->name_hlist);
write_unlock_bh(&dev_base_lock);
synchronize_rcu();
write_lock_bh(&dev_base_lock);
hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
write_unlock_bh(&dev_base_lock);
ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
ret = notifier_to_errno(ret);
if (ret) {
/* err >= 0 after dev_alloc_name() or stores the first errno */
if (err >= 0) {
err = ret;
memcpy(dev->name, oldname, IFNAMSIZ);
goto rollback;
} else {
printk(KERN_ERR
"%s: name change rollback failed: %d.\n",
dev->name, ret);
}
}
return err;
}
/**
* dev_set_alias - change ifalias of a device
* @dev: device
* @alias: name up to IFALIASZ
* @len: limit of bytes to copy from info
*
* Set ifalias for a device,
*/
int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
{
ASSERT_RTNL();
if (len >= IFALIASZ)
return -EINVAL;
if (!len) {
if (dev->ifalias) {
kfree(dev->ifalias);
dev->ifalias = NULL;
}
return 0;
}
dev->ifalias = krealloc(dev->ifalias, len + 1, GFP_KERNEL);
if (!dev->ifalias)
return -ENOMEM;
strlcpy(dev->ifalias, alias, len+1);
return len;
}
/**
* netdev_features_change - device changes features
* @dev: device to cause notification
*
* Called to indicate a device has changed features.
*/
void netdev_features_change(struct net_device *dev)
{
call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
}
EXPORT_SYMBOL(netdev_features_change);
/**
* netdev_state_change - device changes state
* @dev: device to cause notification
*
* Called to indicate a device has changed state. This function calls
* the notifier chains for netdev_chain and sends a NEWLINK message
* to the routing socket.
*/
void netdev_state_change(struct net_device *dev)
{
if (dev->flags & IFF_UP) {
call_netdevice_notifiers(NETDEV_CHANGE, dev);
rtmsg_ifinfo(RTM_NEWLINK, dev, 0);
}
}
EXPORT_SYMBOL(netdev_state_change);
int netdev_bonding_change(struct net_device *dev, unsigned long event)
{
return call_netdevice_notifiers(event, dev);
}
EXPORT_SYMBOL(netdev_bonding_change);
/**
* dev_load - load a network module
* @net: the applicable net namespace
* @name: name of interface
*
* If a network interface is not present and the process has suitable
* privileges this function loads the module. If module loading is not
* available in this kernel then it becomes a nop.
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
void dev_load(struct net *net, const char *name)
{
struct net_device *dev;
net: don't allow CAP_NET_ADMIN to load non-netdev kernel modules Since a8f80e8ff94ecba629542d9b4b5f5a8ee3eb565c any process with CAP_NET_ADMIN may load any module from /lib/modules/. This doesn't mean that CAP_NET_ADMIN is a superset of CAP_SYS_MODULE as modules are limited to /lib/modules/**. However, CAP_NET_ADMIN capability shouldn't allow anybody load any module not related to networking. This patch restricts an ability of autoloading modules to netdev modules with explicit aliases. This fixes CVE-2011-1019. Arnd Bergmann suggested to leave untouched the old pre-v2.6.32 behavior of loading netdev modules by name (without any prefix) for processes with CAP_SYS_MODULE to maintain the compatibility with network scripts that use autoloading netdev modules by aliases like "eth0", "wlan0". Currently there are only three users of the feature in the upstream kernel: ipip, ip_gre and sit. root@albatros:~# capsh --drop=$(seq -s, 0 11),$(seq -s, 13 34) -- root@albatros:~# grep Cap /proc/$$/status CapInh: 0000000000000000 CapPrm: fffffff800001000 CapEff: fffffff800001000 CapBnd: fffffff800001000 root@albatros:~# modprobe xfs FATAL: Error inserting xfs (/lib/modules/2.6.38-rc6-00001-g2bf4ca3/kernel/fs/xfs/xfs.ko): Operation not permitted root@albatros:~# lsmod | grep xfs root@albatros:~# ifconfig xfs xfs: error fetching interface information: Device not found root@albatros:~# lsmod | grep xfs root@albatros:~# lsmod | grep sit root@albatros:~# ifconfig sit sit: error fetching interface information: Device not found root@albatros:~# lsmod | grep sit root@albatros:~# ifconfig sit0 sit0 Link encap:IPv6-in-IPv4 NOARP MTU:1480 Metric:1 root@albatros:~# lsmod | grep sit sit 10457 0 tunnel4 2957 1 sit For CAP_SYS_MODULE module loading is still relaxed: root@albatros:~# grep Cap /proc/$$/status CapInh: 0000000000000000 CapPrm: ffffffffffffffff CapEff: ffffffffffffffff CapBnd: ffffffffffffffff root@albatros:~# ifconfig xfs xfs: error fetching interface information: Device not found root@albatros:~# lsmod | grep xfs xfs 745319 0 Reference: https://lkml.org/lkml/2011/2/24/203 Signed-off-by: Vasiliy Kulikov <segoon@openwall.com> Signed-off-by: Michael Tokarev <mjt@tls.msk.ru> Acked-by: David S. Miller <davem@davemloft.net> Acked-by: Kees Cook <kees.cook@canonical.com> Signed-off-by: James Morris <jmorris@namei.org>
2011-03-01 21:33:13 +00:00
int no_module;
rcu_read_lock();
dev = dev_get_by_name_rcu(net, name);
rcu_read_unlock();
net: don't allow CAP_NET_ADMIN to load non-netdev kernel modules Since a8f80e8ff94ecba629542d9b4b5f5a8ee3eb565c any process with CAP_NET_ADMIN may load any module from /lib/modules/. This doesn't mean that CAP_NET_ADMIN is a superset of CAP_SYS_MODULE as modules are limited to /lib/modules/**. However, CAP_NET_ADMIN capability shouldn't allow anybody load any module not related to networking. This patch restricts an ability of autoloading modules to netdev modules with explicit aliases. This fixes CVE-2011-1019. Arnd Bergmann suggested to leave untouched the old pre-v2.6.32 behavior of loading netdev modules by name (without any prefix) for processes with CAP_SYS_MODULE to maintain the compatibility with network scripts that use autoloading netdev modules by aliases like "eth0", "wlan0". Currently there are only three users of the feature in the upstream kernel: ipip, ip_gre and sit. root@albatros:~# capsh --drop=$(seq -s, 0 11),$(seq -s, 13 34) -- root@albatros:~# grep Cap /proc/$$/status CapInh: 0000000000000000 CapPrm: fffffff800001000 CapEff: fffffff800001000 CapBnd: fffffff800001000 root@albatros:~# modprobe xfs FATAL: Error inserting xfs (/lib/modules/2.6.38-rc6-00001-g2bf4ca3/kernel/fs/xfs/xfs.ko): Operation not permitted root@albatros:~# lsmod | grep xfs root@albatros:~# ifconfig xfs xfs: error fetching interface information: Device not found root@albatros:~# lsmod | grep xfs root@albatros:~# lsmod | grep sit root@albatros:~# ifconfig sit sit: error fetching interface information: Device not found root@albatros:~# lsmod | grep sit root@albatros:~# ifconfig sit0 sit0 Link encap:IPv6-in-IPv4 NOARP MTU:1480 Metric:1 root@albatros:~# lsmod | grep sit sit 10457 0 tunnel4 2957 1 sit For CAP_SYS_MODULE module loading is still relaxed: root@albatros:~# grep Cap /proc/$$/status CapInh: 0000000000000000 CapPrm: ffffffffffffffff CapEff: ffffffffffffffff CapBnd: ffffffffffffffff root@albatros:~# ifconfig xfs xfs: error fetching interface information: Device not found root@albatros:~# lsmod | grep xfs xfs 745319 0 Reference: https://lkml.org/lkml/2011/2/24/203 Signed-off-by: Vasiliy Kulikov <segoon@openwall.com> Signed-off-by: Michael Tokarev <mjt@tls.msk.ru> Acked-by: David S. Miller <davem@davemloft.net> Acked-by: Kees Cook <kees.cook@canonical.com> Signed-off-by: James Morris <jmorris@namei.org>
2011-03-01 21:33:13 +00:00
no_module = !dev;
if (no_module && capable(CAP_NET_ADMIN))
no_module = request_module("netdev-%s", name);
if (no_module && capable(CAP_SYS_MODULE)) {
if (!request_module("%s", name))
pr_err("Loading kernel module for a network device "
"with CAP_SYS_MODULE (deprecated). Use CAP_NET_ADMIN and alias netdev-%s "
"instead\n", name);
}
}
EXPORT_SYMBOL(dev_load);
static int __dev_open(struct net_device *dev)
{
const struct net_device_ops *ops = dev->netdev_ops;
int ret;
ASSERT_RTNL();
if (!netif_device_present(dev))
return -ENODEV;
ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
ret = notifier_to_errno(ret);
if (ret)
return ret;
set_bit(__LINK_STATE_START, &dev->state);
if (ops->ndo_validate_addr)
ret = ops->ndo_validate_addr(dev);
if (!ret && ops->ndo_open)
ret = ops->ndo_open(dev);
if (ret)
clear_bit(__LINK_STATE_START, &dev->state);
else {
dev->flags |= IFF_UP;
net_dmaengine_get();
dev_set_rx_mode(dev);
dev_activate(dev);
}
return ret;
}
/**
* dev_open - prepare an interface for use.
* @dev: device to open
*
* Takes a device from down to up state. The device's private open
* function is invoked and then the multicast lists are loaded. Finally
* the device is moved into the up state and a %NETDEV_UP message is
* sent to the netdev notifier chain.
*
* Calling this function on an active interface is a nop. On a failure
* a negative errno code is returned.
*/
int dev_open(struct net_device *dev)
{
int ret;
if (dev->flags & IFF_UP)
return 0;
ret = __dev_open(dev);
if (ret < 0)
return ret;
rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING);
call_netdevice_notifiers(NETDEV_UP, dev);
return ret;
}
EXPORT_SYMBOL(dev_open);
static int __dev_close_many(struct list_head *head)
{
struct net_device *dev;
ASSERT_RTNL();
might_sleep();
list_for_each_entry(dev, head, unreg_list) {
call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
clear_bit(__LINK_STATE_START, &dev->state);
/* Synchronize to scheduled poll. We cannot touch poll list, it
* can be even on different cpu. So just clear netif_running().
*
* dev->stop() will invoke napi_disable() on all of it's
* napi_struct instances on this device.
*/
smp_mb__after_clear_bit(); /* Commit netif_running(). */
}
dev_deactivate_many(head);
list_for_each_entry(dev, head, unreg_list) {
const struct net_device_ops *ops = dev->netdev_ops;
/*
* Call the device specific close. This cannot fail.
* Only if device is UP
*
* We allow it to be called even after a DETACH hot-plug
* event.
*/
if (ops->ndo_stop)
ops->ndo_stop(dev);
dev->flags &= ~IFF_UP;
net_dmaengine_put();
}
return 0;
}
static int __dev_close(struct net_device *dev)
{
int retval;
LIST_HEAD(single);
list_add(&dev->unreg_list, &single);
retval = __dev_close_many(&single);
list_del(&single);
return retval;
}
static int dev_close_many(struct list_head *head)
{
struct net_device *dev, *tmp;
LIST_HEAD(tmp_list);
list_for_each_entry_safe(dev, tmp, head, unreg_list)
if (!(dev->flags & IFF_UP))
list_move(&dev->unreg_list, &tmp_list);
__dev_close_many(head);
list_for_each_entry(dev, head, unreg_list) {
rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING);
call_netdevice_notifiers(NETDEV_DOWN, dev);
}
/* rollback_registered_many needs the complete original list */
list_splice(&tmp_list, head);
return 0;
}
/**
* dev_close - shutdown an interface.
* @dev: device to shutdown
*
* This function moves an active device into down state. A
* %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
* is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
* chain.
*/
int dev_close(struct net_device *dev)
{
if (dev->flags & IFF_UP) {
LIST_HEAD(single);
list_add(&dev->unreg_list, &single);
dev_close_many(&single);
list_del(&single);
}
return 0;
}
EXPORT_SYMBOL(dev_close);
/**
* dev_disable_lro - disable Large Receive Offload on a device
* @dev: device
*
* Disable Large Receive Offload (LRO) on a net device. Must be
* called under RTNL. This is needed if received packets may be
* forwarded to another interface.
*/
void dev_disable_lro(struct net_device *dev)
{
u32 flags;
/*
* If we're trying to disable lro on a vlan device
* use the underlying physical device instead
*/
if (is_vlan_dev(dev))
dev = vlan_dev_real_dev(dev);
if (dev->ethtool_ops && dev->ethtool_ops->get_flags)
flags = dev->ethtool_ops->get_flags(dev);
else
flags = ethtool_op_get_flags(dev);
if (!(flags & ETH_FLAG_LRO))
return;
__ethtool_set_flags(dev, flags & ~ETH_FLAG_LRO);
if (unlikely(dev->features & NETIF_F_LRO))
netdev_WARN(dev, "failed to disable LRO!\n");
}
EXPORT_SYMBOL(dev_disable_lro);
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
static int dev_boot_phase = 1;
/**
* register_netdevice_notifier - register a network notifier block
* @nb: notifier
*
* Register a notifier to be called when network device events occur.
* The notifier passed is linked into the kernel structures and must
* not be reused until it has been unregistered. A negative errno code
* is returned on a failure.
*
* When registered all registration and up events are replayed
* to the new notifier to allow device to have a race free
* view of the network device list.
*/
int register_netdevice_notifier(struct notifier_block *nb)
{
struct net_device *dev;
struct net_device *last;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
struct net *net;
int err;
rtnl_lock();
err = raw_notifier_chain_register(&netdev_chain, nb);
if (err)
goto unlock;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (dev_boot_phase)
goto unlock;
for_each_net(net) {
for_each_netdev(net, dev) {
err = nb->notifier_call(nb, NETDEV_REGISTER, dev);
err = notifier_to_errno(err);
if (err)
goto rollback;
if (!(dev->flags & IFF_UP))
continue;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
nb->notifier_call(nb, NETDEV_UP, dev);
}
}
unlock:
rtnl_unlock();
return err;
rollback:
last = dev;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
for_each_net(net) {
for_each_netdev(net, dev) {
if (dev == last)
break;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (dev->flags & IFF_UP) {
nb->notifier_call(nb, NETDEV_GOING_DOWN, dev);
nb->notifier_call(nb, NETDEV_DOWN, dev);
}
nb->notifier_call(nb, NETDEV_UNREGISTER, dev);
nb->notifier_call(nb, NETDEV_UNREGISTER_BATCH, dev);
}
}
raw_notifier_chain_unregister(&netdev_chain, nb);
goto unlock;
}
EXPORT_SYMBOL(register_netdevice_notifier);
/**
* unregister_netdevice_notifier - unregister a network notifier block
* @nb: notifier
*
* Unregister a notifier previously registered by
* register_netdevice_notifier(). The notifier is unlinked into the
* kernel structures and may then be reused. A negative errno code
* is returned on a failure.
*/
int unregister_netdevice_notifier(struct notifier_block *nb)
{
int err;
rtnl_lock();
err = raw_notifier_chain_unregister(&netdev_chain, nb);
rtnl_unlock();
return err;
}
EXPORT_SYMBOL(unregister_netdevice_notifier);
/**
* call_netdevice_notifiers - call all network notifier blocks
* @val: value passed unmodified to notifier function
* @dev: net_device pointer passed unmodified to notifier function
*
* Call all network notifier blocks. Parameters and return value
* are as for raw_notifier_call_chain().
*/
int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
{
ASSERT_RTNL();
return raw_notifier_call_chain(&netdev_chain, val, dev);
}
EXPORT_SYMBOL(call_netdevice_notifiers);
/* When > 0 there are consumers of rx skb time stamps */
static atomic_t netstamp_needed = ATOMIC_INIT(0);
void net_enable_timestamp(void)
{
atomic_inc(&netstamp_needed);
}
EXPORT_SYMBOL(net_enable_timestamp);
void net_disable_timestamp(void)
{
atomic_dec(&netstamp_needed);
}
EXPORT_SYMBOL(net_disable_timestamp);
static inline void net_timestamp_set(struct sk_buff *skb)
{
if (atomic_read(&netstamp_needed))
__net_timestamp(skb);
else
skb->tstamp.tv64 = 0;
}
static inline void net_timestamp_check(struct sk_buff *skb)
{
if (!skb->tstamp.tv64 && atomic_read(&netstamp_needed))
__net_timestamp(skb);
}
static int net_hwtstamp_validate(struct ifreq *ifr)
{
struct hwtstamp_config cfg;
enum hwtstamp_tx_types tx_type;
enum hwtstamp_rx_filters rx_filter;
int tx_type_valid = 0;
int rx_filter_valid = 0;
if (copy_from_user(&cfg, ifr->ifr_data, sizeof(cfg)))
return -EFAULT;
if (cfg.flags) /* reserved for future extensions */
return -EINVAL;
tx_type = cfg.tx_type;
rx_filter = cfg.rx_filter;
switch (tx_type) {
case HWTSTAMP_TX_OFF:
case HWTSTAMP_TX_ON:
case HWTSTAMP_TX_ONESTEP_SYNC:
tx_type_valid = 1;
break;
}
switch (rx_filter) {
case HWTSTAMP_FILTER_NONE:
case HWTSTAMP_FILTER_ALL:
case HWTSTAMP_FILTER_SOME:
case HWTSTAMP_FILTER_PTP_V1_L4_EVENT:
case HWTSTAMP_FILTER_PTP_V1_L4_SYNC:
case HWTSTAMP_FILTER_PTP_V1_L4_DELAY_REQ:
case HWTSTAMP_FILTER_PTP_V2_L4_EVENT:
case HWTSTAMP_FILTER_PTP_V2_L4_SYNC:
case HWTSTAMP_FILTER_PTP_V2_L4_DELAY_REQ:
case HWTSTAMP_FILTER_PTP_V2_L2_EVENT:
case HWTSTAMP_FILTER_PTP_V2_L2_SYNC:
case HWTSTAMP_FILTER_PTP_V2_L2_DELAY_REQ:
case HWTSTAMP_FILTER_PTP_V2_EVENT:
case HWTSTAMP_FILTER_PTP_V2_SYNC:
case HWTSTAMP_FILTER_PTP_V2_DELAY_REQ:
rx_filter_valid = 1;
break;
}
if (!tx_type_valid || !rx_filter_valid)
return -ERANGE;
return 0;
}
static inline bool is_skb_forwardable(struct net_device *dev,
struct sk_buff *skb)
{
unsigned int len;
if (!(dev->flags & IFF_UP))
return false;
len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
if (skb->len <= len)
return true;
/* if TSO is enabled, we don't care about the length as the packet
* could be forwarded without being segmented before
*/
if (skb_is_gso(skb))
return true;
return false;
}
/**
* dev_forward_skb - loopback an skb to another netif
*
* @dev: destination network device
* @skb: buffer to forward
*
* return values:
* NET_RX_SUCCESS (no congestion)
* NET_RX_DROP (packet was dropped, but freed)
*
* dev_forward_skb can be used for injecting an skb from the
* start_xmit function of one device into the receive queue
* of another device.
*
* The receiving device may be in another namespace, so
* we have to clear all information in the skb that could
* impact namespace isolation.
*/
int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
{
if (skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) {
if (skb_copy_ubufs(skb, GFP_ATOMIC)) {
atomic_long_inc(&dev->rx_dropped);
kfree_skb(skb);
return NET_RX_DROP;
}
}
skb_orphan(skb);
nf_reset(skb);
if (unlikely(!is_skb_forwardable(dev, skb))) {
atomic_long_inc(&dev->rx_dropped);
kfree_skb(skb);
return NET_RX_DROP;
}
skb_set_dev(skb, dev);
skb->tstamp.tv64 = 0;
skb->pkt_type = PACKET_HOST;
skb->protocol = eth_type_trans(skb, dev);
return netif_rx(skb);
}
EXPORT_SYMBOL_GPL(dev_forward_skb);
static inline int deliver_skb(struct sk_buff *skb,
struct packet_type *pt_prev,
struct net_device *orig_dev)
{
atomic_inc(&skb->users);
return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
}
/*
* Support routine. Sends outgoing frames to any network
* taps currently in use.
*/
static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
{
struct packet_type *ptype;
struct sk_buff *skb2 = NULL;
struct packet_type *pt_prev = NULL;
rcu_read_lock();
list_for_each_entry_rcu(ptype, &ptype_all, list) {
/* Never send packets back to the socket
* they originated from - MvS (miquels@drinkel.ow.org)
*/
if ((ptype->dev == dev || !ptype->dev) &&
(ptype->af_packet_priv == NULL ||
(struct sock *)ptype->af_packet_priv != skb->sk)) {
if (pt_prev) {
deliver_skb(skb2, pt_prev, skb->dev);
pt_prev = ptype;
continue;
}
skb2 = skb_clone(skb, GFP_ATOMIC);
if (!skb2)
break;
net_timestamp_set(skb2);
/* skb->nh should be correctly
set by sender, so that the second statement is
just protection against buggy protocols.
*/
skb_reset_mac_header(skb2);
if (skb_network_header(skb2) < skb2->data ||
skb2->network_header > skb2->tail) {
if (net_ratelimit())
printk(KERN_CRIT "protocol %04x is "
"buggy, dev %s\n",
ntohs(skb2->protocol),
dev->name);
skb_reset_network_header(skb2);
}
skb2->transport_header = skb2->network_header;
skb2->pkt_type = PACKET_OUTGOING;
pt_prev = ptype;
}
}
if (pt_prev)
pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
rcu_read_unlock();
}
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
/* netif_setup_tc - Handle tc mappings on real_num_tx_queues change
* @dev: Network device
* @txq: number of queues available
*
* If real_num_tx_queues is changed the tc mappings may no longer be
* valid. To resolve this verify the tc mapping remains valid and if
* not NULL the mapping. With no priorities mapping to this
* offset/count pair it will no longer be used. In the worst case TC0
* is invalid nothing can be done so disable priority mappings. If is
* expected that drivers will fix this mapping if they can before
* calling netif_set_real_num_tx_queues.
*/
static void netif_setup_tc(struct net_device *dev, unsigned int txq)
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
{
int i;
struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
/* If TC0 is invalidated disable TC mapping */
if (tc->offset + tc->count > txq) {
pr_warning("Number of in use tx queues changed "
"invalidating tc mappings. Priority "
"traffic classification disabled!\n");
dev->num_tc = 0;
return;
}
/* Invalidated prio to tc mappings set to TC0 */
for (i = 1; i < TC_BITMASK + 1; i++) {
int q = netdev_get_prio_tc_map(dev, i);
tc = &dev->tc_to_txq[q];
if (tc->offset + tc->count > txq) {
pr_warning("Number of in use tx queues "
"changed. Priority %i to tc "
"mapping %i is no longer valid "
"setting map to 0\n",
i, q);
netdev_set_prio_tc_map(dev, i, 0);
}
}
}
/*
* Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
* greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
*/
int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
{
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
int rc;
if (txq < 1 || txq > dev->num_tx_queues)
return -EINVAL;
if (dev->reg_state == NETREG_REGISTERED ||
dev->reg_state == NETREG_UNREGISTERING) {
ASSERT_RTNL();
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
txq);
if (rc)
return rc;
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
if (dev->num_tc)
netif_setup_tc(dev, txq);
if (txq < dev->real_num_tx_queues)
qdisc_reset_all_tx_gt(dev, txq);
}
dev->real_num_tx_queues = txq;
return 0;
}
EXPORT_SYMBOL(netif_set_real_num_tx_queues);
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
#ifdef CONFIG_RPS
/**
* netif_set_real_num_rx_queues - set actual number of RX queues used
* @dev: Network device
* @rxq: Actual number of RX queues
*
* This must be called either with the rtnl_lock held or before
* registration of the net device. Returns 0 on success, or a
* negative error code. If called before registration, it always
* succeeds.
*/
int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
{
int rc;
if (rxq < 1 || rxq > dev->num_rx_queues)
return -EINVAL;
if (dev->reg_state == NETREG_REGISTERED) {
ASSERT_RTNL();
rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
rxq);
if (rc)
return rc;
}
dev->real_num_rx_queues = rxq;
return 0;
}
EXPORT_SYMBOL(netif_set_real_num_rx_queues);
#endif
static inline void __netif_reschedule(struct Qdisc *q)
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
{
struct softnet_data *sd;
unsigned long flags;
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
local_irq_save(flags);
sd = &__get_cpu_var(softnet_data);
q->next_sched = NULL;
*sd->output_queue_tailp = q;
sd->output_queue_tailp = &q->next_sched;
raise_softirq_irqoff(NET_TX_SOFTIRQ);
local_irq_restore(flags);
}
void __netif_schedule(struct Qdisc *q)
{
if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
__netif_reschedule(q);
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
}
EXPORT_SYMBOL(__netif_schedule);
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
void dev_kfree_skb_irq(struct sk_buff *skb)
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
{
if (atomic_dec_and_test(&skb->users)) {
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
struct softnet_data *sd;
unsigned long flags;
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
local_irq_save(flags);
sd = &__get_cpu_var(softnet_data);
skb->next = sd->completion_queue;
sd->completion_queue = skb;
raise_softirq_irqoff(NET_TX_SOFTIRQ);
local_irq_restore(flags);
}
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
}
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
EXPORT_SYMBOL(dev_kfree_skb_irq);
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
void dev_kfree_skb_any(struct sk_buff *skb)
{
if (in_irq() || irqs_disabled())
dev_kfree_skb_irq(skb);
else
dev_kfree_skb(skb);
}
EXPORT_SYMBOL(dev_kfree_skb_any);
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
/**
* netif_device_detach - mark device as removed
* @dev: network device
*
* Mark device as removed from system and therefore no longer available.
*/
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
void netif_device_detach(struct net_device *dev)
{
if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
netif_running(dev)) {
netif_tx_stop_all_queues(dev);
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
}
}
EXPORT_SYMBOL(netif_device_detach);
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
/**
* netif_device_attach - mark device as attached
* @dev: network device
*
* Mark device as attached from system and restart if needed.
*/
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
void netif_device_attach(struct net_device *dev)
{
if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
netif_running(dev)) {
netif_tx_wake_all_queues(dev);
__netdev_watchdog_up(dev);
[NET]: Deinline some larger functions from netdevice.h On a allyesconfig'ured kernel: Size Uses Wasted Name and definition ===== ==== ====== ================================================ 95 162 12075 netif_wake_queue include/linux/netdevice.h 129 86 9265 dev_kfree_skb_any include/linux/netdevice.h 127 56 5885 netif_device_attach include/linux/netdevice.h 73 86 4505 dev_kfree_skb_irq include/linux/netdevice.h 46 60 1534 netif_device_detach include/linux/netdevice.h 119 16 1485 __netif_rx_schedule include/linux/netdevice.h 143 5 492 netif_rx_schedule include/linux/netdevice.h 81 7 366 netif_schedule include/linux/netdevice.h netif_wake_queue is big because __netif_schedule is a big inline: static inline void __netif_schedule(struct net_device *dev) { if (!test_and_set_bit(__LINK_STATE_SCHED, &dev->state)) { unsigned long flags; struct softnet_data *sd; local_irq_save(flags); sd = &__get_cpu_var(softnet_data); dev->next_sched = sd->output_queue; sd->output_queue = dev; raise_softirq_irqoff(NET_TX_SOFTIRQ); local_irq_restore(flags); } } static inline void netif_wake_queue(struct net_device *dev) { #ifdef CONFIG_NETPOLL_TRAP if (netpoll_trap()) return; #endif if (test_and_clear_bit(__LINK_STATE_XOFF, &dev->state)) __netif_schedule(dev); } By de-inlining __netif_schedule we are saving a lot of text at each callsite of netif_wake_queue and netif_schedule. __netif_rx_schedule is also big, and it makes more sense to keep both of them out of line. Patch also deinlines dev_kfree_skb_any. We can deinline dev_kfree_skb_irq instead... oh well. netif_device_attach/detach are not hot paths, we can deinline them too. Signed-off-by: Denis Vlasenko <vda@ilport.com.ua> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-03-29 23:57:29 +00:00
}
}
EXPORT_SYMBOL(netif_device_attach);
/**
* skb_dev_set -- assign a new device to a buffer
* @skb: buffer for the new device
* @dev: network device
*
* If an skb is owned by a device already, we have to reset
* all data private to the namespace a device belongs to
* before assigning it a new device.
*/
#ifdef CONFIG_NET_NS
void skb_set_dev(struct sk_buff *skb, struct net_device *dev)
{
skb_dst_drop(skb);
if (skb->dev && !net_eq(dev_net(skb->dev), dev_net(dev))) {
secpath_reset(skb);
nf_reset(skb);
skb_init_secmark(skb);
skb->mark = 0;
skb->priority = 0;
skb->nf_trace = 0;
skb->ipvs_property = 0;
#ifdef CONFIG_NET_SCHED
skb->tc_index = 0;
#endif
}
skb->dev = dev;
}
EXPORT_SYMBOL(skb_set_dev);
#endif /* CONFIG_NET_NS */
/*
* Invalidate hardware checksum when packet is to be mangled, and
* complete checksum manually on outgoing path.
*/
int skb_checksum_help(struct sk_buff *skb)
{
__wsum csum;
int ret = 0, offset;
if (skb->ip_summed == CHECKSUM_COMPLETE)
goto out_set_summed;
if (unlikely(skb_shinfo(skb)->gso_size)) {
/* Let GSO fix up the checksum. */
goto out_set_summed;
}
offset = skb_checksum_start_offset(skb);
BUG_ON(offset >= skb_headlen(skb));
csum = skb_checksum(skb, offset, skb->len - offset, 0);
offset += skb->csum_offset;
BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
if (skb_cloned(skb) &&
!skb_clone_writable(skb, offset + sizeof(__sum16))) {
ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
if (ret)
goto out;
}
*(__sum16 *)(skb->data + offset) = csum_fold(csum);
out_set_summed:
skb->ip_summed = CHECKSUM_NONE;
out:
return ret;
}
EXPORT_SYMBOL(skb_checksum_help);
/**
* skb_gso_segment - Perform segmentation on skb.
* @skb: buffer to segment
* @features: features for the output path (see dev->features)
*
* This function segments the given skb and returns a list of segments.
*
* It may return NULL if the skb requires no segmentation. This is
* only possible when GSO is used for verifying header integrity.
*/
struct sk_buff *skb_gso_segment(struct sk_buff *skb, u32 features)
{
struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
struct packet_type *ptype;
__be16 type = skb->protocol;
int vlan_depth = ETH_HLEN;
int err;
while (type == htons(ETH_P_8021Q)) {
struct vlan_hdr *vh;
if (unlikely(!pskb_may_pull(skb, vlan_depth + VLAN_HLEN)))
return ERR_PTR(-EINVAL);
vh = (struct vlan_hdr *)(skb->data + vlan_depth);
type = vh->h_vlan_encapsulated_proto;
vlan_depth += VLAN_HLEN;
}
skb_reset_mac_header(skb);
skb->mac_len = skb->network_header - skb->mac_header;
__skb_pull(skb, skb->mac_len);
if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
struct net_device *dev = skb->dev;
struct ethtool_drvinfo info = {};
if (dev && dev->ethtool_ops && dev->ethtool_ops->get_drvinfo)
dev->ethtool_ops->get_drvinfo(dev, &info);
WARN(1, "%s: caps=(0x%lx, 0x%lx) len=%d data_len=%d ip_summed=%d\n",
info.driver, dev ? dev->features : 0L,
skb->sk ? skb->sk->sk_route_caps : 0L,
skb->len, skb->data_len, skb->ip_summed);
if (skb_header_cloned(skb) &&
(err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC)))
return ERR_PTR(err);
}
rcu_read_lock();
list_for_each_entry_rcu(ptype,
&ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
if (ptype->type == type && !ptype->dev && ptype->gso_segment) {
if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
err = ptype->gso_send_check(skb);
segs = ERR_PTR(err);
if (err || skb_gso_ok(skb, features))
break;
__skb_push(skb, (skb->data -
skb_network_header(skb)));
}
segs = ptype->gso_segment(skb, features);
break;
}
}
rcu_read_unlock();
__skb_push(skb, skb->data - skb_mac_header(skb));
return segs;
}
EXPORT_SYMBOL(skb_gso_segment);
/* Take action when hardware reception checksum errors are detected. */
#ifdef CONFIG_BUG
void netdev_rx_csum_fault(struct net_device *dev)
{
if (net_ratelimit()) {
printk(KERN_ERR "%s: hw csum failure.\n",
dev ? dev->name : "<unknown>");
dump_stack();
}
}
EXPORT_SYMBOL(netdev_rx_csum_fault);
#endif
/* Actually, we should eliminate this check as soon as we know, that:
* 1. IOMMU is present and allows to map all the memory.
* 2. No high memory really exists on this machine.
*/
static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
{
#ifdef CONFIG_HIGHMEM
int i;
if (!(dev->features & NETIF_F_HIGHDMA)) {
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
if (PageHighMem(skb_frag_page(frag)))
return 1;
}
}
if (PCI_DMA_BUS_IS_PHYS) {
struct device *pdev = dev->dev.parent;
if (!pdev)
return 0;
for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
dma_addr_t addr = page_to_phys(skb_frag_page(frag));
if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
return 1;
}
}
#endif
return 0;
}
struct dev_gso_cb {
void (*destructor)(struct sk_buff *skb);
};
#define DEV_GSO_CB(skb) ((struct dev_gso_cb *)(skb)->cb)
static void dev_gso_skb_destructor(struct sk_buff *skb)
{
struct dev_gso_cb *cb;
do {
struct sk_buff *nskb = skb->next;
skb->next = nskb->next;
nskb->next = NULL;
kfree_skb(nskb);
} while (skb->next);
cb = DEV_GSO_CB(skb);
if (cb->destructor)
cb->destructor(skb);
}
/**
* dev_gso_segment - Perform emulated hardware segmentation on skb.
* @skb: buffer to segment
* @features: device features as applicable to this skb
*
* This function segments the given skb and stores the list of segments
* in skb->next.
*/
static int dev_gso_segment(struct sk_buff *skb, int features)
{
struct sk_buff *segs;
segs = skb_gso_segment(skb, features);
/* Verifying header integrity only. */
if (!segs)
return 0;
if (IS_ERR(segs))
return PTR_ERR(segs);
skb->next = segs;
DEV_GSO_CB(skb)->destructor = skb->destructor;
skb->destructor = dev_gso_skb_destructor;
return 0;
}
/*
* Try to orphan skb early, right before transmission by the device.
* We cannot orphan skb if tx timestamp is requested or the sk-reference
* is needed on driver level for other reasons, e.g. see net/can/raw.c
*/
static inline void skb_orphan_try(struct sk_buff *skb)
{
struct sock *sk = skb->sk;
if (sk && !skb_shinfo(skb)->tx_flags) {
/* skb_tx_hash() wont be able to get sk.
* We copy sk_hash into skb->rxhash
*/
if (!skb->rxhash)
skb->rxhash = sk->sk_hash;
skb_orphan(skb);
}
}
static bool can_checksum_protocol(unsigned long features, __be16 protocol)
{
return ((features & NETIF_F_GEN_CSUM) ||
((features & NETIF_F_V4_CSUM) &&
protocol == htons(ETH_P_IP)) ||
((features & NETIF_F_V6_CSUM) &&
protocol == htons(ETH_P_IPV6)) ||
((features & NETIF_F_FCOE_CRC) &&
protocol == htons(ETH_P_FCOE)));
}
static u32 harmonize_features(struct sk_buff *skb, __be16 protocol, u32 features)
{
if (!can_checksum_protocol(features, protocol)) {
features &= ~NETIF_F_ALL_CSUM;
features &= ~NETIF_F_SG;
} else if (illegal_highdma(skb->dev, skb)) {
features &= ~NETIF_F_SG;
}
return features;
}
u32 netif_skb_features(struct sk_buff *skb)
{
__be16 protocol = skb->protocol;
u32 features = skb->dev->features;
if (protocol == htons(ETH_P_8021Q)) {
struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
protocol = veh->h_vlan_encapsulated_proto;
} else if (!vlan_tx_tag_present(skb)) {
return harmonize_features(skb, protocol, features);
}
features &= (skb->dev->vlan_features | NETIF_F_HW_VLAN_TX);
if (protocol != htons(ETH_P_8021Q)) {
return harmonize_features(skb, protocol, features);
} else {
features &= NETIF_F_SG | NETIF_F_HIGHDMA | NETIF_F_FRAGLIST |
NETIF_F_GEN_CSUM | NETIF_F_HW_VLAN_TX;
return harmonize_features(skb, protocol, features);
}
}
EXPORT_SYMBOL(netif_skb_features);
/*
* Returns true if either:
* 1. skb has frag_list and the device doesn't support FRAGLIST, or
* 2. skb is fragmented and the device does not support SG, or if
* at least one of fragments is in highmem and device does not
* support DMA from it.
*/
static inline int skb_needs_linearize(struct sk_buff *skb,
int features)
{
return skb_is_nonlinear(skb) &&
((skb_has_frag_list(skb) &&
!(features & NETIF_F_FRAGLIST)) ||
(skb_shinfo(skb)->nr_frags &&
!(features & NETIF_F_SG)));
}
int dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev,
struct netdev_queue *txq)
{
const struct net_device_ops *ops = dev->netdev_ops;
int rc = NETDEV_TX_OK;
unsigned int skb_len;
if (likely(!skb->next)) {
u32 features;
/*
* If device doesn't need skb->dst, release it right now while
* its hot in this cpu cache
*/
if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
skb_dst_drop(skb);
if (!list_empty(&ptype_all))
dev_queue_xmit_nit(skb, dev);
skb_orphan_try(skb);
features = netif_skb_features(skb);
if (vlan_tx_tag_present(skb) &&
!(features & NETIF_F_HW_VLAN_TX)) {
skb = __vlan_put_tag(skb, vlan_tx_tag_get(skb));
if (unlikely(!skb))
goto out;
skb->vlan_tci = 0;
}
if (netif_needs_gso(skb, features)) {
if (unlikely(dev_gso_segment(skb, features)))
goto out_kfree_skb;
if (skb->next)
goto gso;
} else {
if (skb_needs_linearize(skb, features) &&
__skb_linearize(skb))
goto out_kfree_skb;
/* If packet is not checksummed and device does not
* support checksumming for this protocol, complete
* checksumming here.
*/
if (skb->ip_summed == CHECKSUM_PARTIAL) {
skb_set_transport_header(skb,
skb_checksum_start_offset(skb));
if (!(features & NETIF_F_ALL_CSUM) &&
skb_checksum_help(skb))
goto out_kfree_skb;
}
}
skb_len = skb->len;
rc = ops->ndo_start_xmit(skb, dev);
trace_net_dev_xmit(skb, rc, dev, skb_len);
if (rc == NETDEV_TX_OK)
txq_trans_update(txq);
return rc;
}
gso:
do {
struct sk_buff *nskb = skb->next;
skb->next = nskb->next;
nskb->next = NULL;
/*
* If device doesn't need nskb->dst, release it right now while
* its hot in this cpu cache
*/
if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
skb_dst_drop(nskb);
skb_len = nskb->len;
rc = ops->ndo_start_xmit(nskb, dev);
trace_net_dev_xmit(nskb, rc, dev, skb_len);
if (unlikely(rc != NETDEV_TX_OK)) {
if (rc & ~NETDEV_TX_MASK)
goto out_kfree_gso_skb;
nskb->next = skb->next;
skb->next = nskb;
return rc;
}
txq_trans_update(txq);
if (unlikely(netif_tx_queue_stopped(txq) && skb->next))
return NETDEV_TX_BUSY;
} while (skb->next);
out_kfree_gso_skb:
if (likely(skb->next == NULL))
skb->destructor = DEV_GSO_CB(skb)->destructor;
out_kfree_skb:
kfree_skb(skb);
out:
return rc;
}
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
static u32 hashrnd __read_mostly;
/*
* Returns a Tx hash based on the given packet descriptor a Tx queues' number
* to be used as a distribution range.
*/
u16 __skb_tx_hash(const struct net_device *dev, const struct sk_buff *skb,
unsigned int num_tx_queues)
{
u32 hash;
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
u16 qoffset = 0;
u16 qcount = num_tx_queues;
if (skb_rx_queue_recorded(skb)) {
hash = skb_get_rx_queue(skb);
while (unlikely(hash >= num_tx_queues))
hash -= num_tx_queues;
return hash;
}
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
if (dev->num_tc) {
u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
qoffset = dev->tc_to_txq[tc].offset;
qcount = dev->tc_to_txq[tc].count;
}
if (skb->sk && skb->sk->sk_hash)
hash = skb->sk->sk_hash;
else
hash = (__force u16) skb->protocol ^ skb->rxhash;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
hash = jhash_1word(hash, hashrnd);
net: implement mechanism for HW based QOS This patch provides a mechanism for lower layer devices to steer traffic using skb->priority to tx queues. This allows for hardware based QOS schemes to use the default qdisc without incurring the penalties related to global state and the qdisc lock. While reliably receiving skbs on the correct tx ring to avoid head of line blocking resulting from shuffling in the LLD. Finally, all the goodness from txq caching and xps/rps can still be leveraged. Many drivers and hardware exist with the ability to implement QOS schemes in the hardware but currently these drivers tend to rely on firmware to reroute specific traffic, a driver specific select_queue or the queue_mapping action in the qdisc. By using select_queue for this drivers need to be updated for each and every traffic type and we lose the goodness of much of the upstream work. Firmware solutions are inherently inflexible. And finally if admins are expected to build a qdisc and filter rules to steer traffic this requires knowledge of how the hardware is currently configured. The number of tx queues and the queue offsets may change depending on resources. Also this approach incurs all the overhead of a qdisc with filters. With the mechanism in this patch users can set skb priority using expected methods ie setsockopt() or the stack can set the priority directly. Then the skb will be steered to the correct tx queues aligned with hardware QOS traffic classes. In the normal case with single traffic class and all queues in this class everything works as is until the LLD enables multiple tcs. To steer the skb we mask out the lower 4 bits of the priority and allow the hardware to configure upto 15 distinct classes of traffic. This is expected to be sufficient for most applications at any rate it is more then the 8021Q spec designates and is equal to the number of prio bands currently implemented in the default qdisc. This in conjunction with a userspace application such as lldpad can be used to implement 8021Q transmission selection algorithms one of these algorithms being the extended transmission selection algorithm currently being used for DCB. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-01-17 08:06:04 +00:00
return (u16) (((u64) hash * qcount) >> 32) + qoffset;
}
EXPORT_SYMBOL(__skb_tx_hash);
static inline u16 dev_cap_txqueue(struct net_device *dev, u16 queue_index)
{
if (unlikely(queue_index >= dev->real_num_tx_queues)) {
if (net_ratelimit()) {
pr_warning("%s selects TX queue %d, but "
"real number of TX queues is %d\n",
dev->name, queue_index, dev->real_num_tx_queues);
}
return 0;
}
return queue_index;
}
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
{
#ifdef CONFIG_XPS
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
struct xps_dev_maps *dev_maps;
struct xps_map *map;
int queue_index = -1;
rcu_read_lock();
dev_maps = rcu_dereference(dev->xps_maps);
if (dev_maps) {
map = rcu_dereference(
dev_maps->cpu_map[raw_smp_processor_id()]);
if (map) {
if (map->len == 1)
queue_index = map->queues[0];
else {
u32 hash;
if (skb->sk && skb->sk->sk_hash)
hash = skb->sk->sk_hash;
else
hash = (__force u16) skb->protocol ^
skb->rxhash;
hash = jhash_1word(hash, hashrnd);
queue_index = map->queues[
((u64)hash * map->len) >> 32];
}
if (unlikely(queue_index >= dev->real_num_tx_queues))
queue_index = -1;
}
}
rcu_read_unlock();
return queue_index;
#else
return -1;
#endif
}
static struct netdev_queue *dev_pick_tx(struct net_device *dev,
struct sk_buff *skb)
{
int queue_index;
const struct net_device_ops *ops = dev->netdev_ops;
if (dev->real_num_tx_queues == 1)
queue_index = 0;
else if (ops->ndo_select_queue) {
queue_index = ops->ndo_select_queue(dev, skb);
queue_index = dev_cap_txqueue(dev, queue_index);
} else {
struct sock *sk = skb->sk;
queue_index = sk_tx_queue_get(sk);
if (queue_index < 0 || skb->ooo_okay ||
queue_index >= dev->real_num_tx_queues) {
int old_index = queue_index;
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
queue_index = get_xps_queue(dev, skb);
if (queue_index < 0)
queue_index = skb_tx_hash(dev, skb);
if (queue_index != old_index && sk) {
struct dst_entry *dst =
rcu_dereference_check(sk->sk_dst_cache, 1);
if (dst && skb_dst(skb) == dst)
sk_tx_queue_set(sk, queue_index);
}
}
}
skb_set_queue_mapping(skb, queue_index);
return netdev_get_tx_queue(dev, queue_index);
}
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
struct net_device *dev,
struct netdev_queue *txq)
{
spinlock_t *root_lock = qdisc_lock(q);
bool contended;
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
int rc;
qdisc_skb_cb(skb)->pkt_len = skb->len;
qdisc_calculate_pkt_len(skb, q);
/*
* Heuristic to force contended enqueues to serialize on a
* separate lock before trying to get qdisc main lock.
* This permits __QDISC_STATE_RUNNING owner to get the lock more often
* and dequeue packets faster.
*/
contended = qdisc_is_running(q);
if (unlikely(contended))
spin_lock(&q->busylock);
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
spin_lock(root_lock);
if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
kfree_skb(skb);
rc = NET_XMIT_DROP;
} else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
qdisc_run_begin(q)) {
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
/*
* This is a work-conserving queue; there are no old skbs
* waiting to be sent out; and the qdisc is not running -
* xmit the skb directly.
*/
if (!(dev->priv_flags & IFF_XMIT_DST_RELEASE))
skb_dst_force(skb);
qdisc_bstats_update(q, skb);
if (sch_direct_xmit(skb, q, dev, txq, root_lock)) {
if (unlikely(contended)) {
spin_unlock(&q->busylock);
contended = false;
}
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
__qdisc_run(q);
} else
qdisc_run_end(q);
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
rc = NET_XMIT_SUCCESS;
} else {
skb_dst_force(skb);
rc = q->enqueue(skb, q) & NET_XMIT_MASK;
if (qdisc_run_begin(q)) {
if (unlikely(contended)) {
spin_unlock(&q->busylock);
contended = false;
}
__qdisc_run(q);
}
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
}
spin_unlock(root_lock);
if (unlikely(contended))
spin_unlock(&q->busylock);
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
return rc;
}
static DEFINE_PER_CPU(int, xmit_recursion);
#define RECURSION_LIMIT 10
/**
* dev_queue_xmit - transmit a buffer
* @skb: buffer to transmit
*
* Queue a buffer for transmission to a network device. The caller must
* have set the device and priority and built the buffer before calling
* this function. The function can be called from an interrupt.
*
* A negative errno code is returned on a failure. A success does not
* guarantee the frame will be transmitted as it may be dropped due
* to congestion or traffic shaping.
*
* -----------------------------------------------------------------------------------
* I notice this method can also return errors from the queue disciplines,
* including NET_XMIT_DROP, which is a positive value. So, errors can also
* be positive.
*
* Regardless of the return value, the skb is consumed, so it is currently
* difficult to retry a send to this method. (You can bump the ref count
* before sending to hold a reference for retry if you are careful.)
*
* When calling this method, interrupts MUST be enabled. This is because
* the BH enable code must have IRQs enabled so that it will not deadlock.
* --BLG
*/
int dev_queue_xmit(struct sk_buff *skb)
{
struct net_device *dev = skb->dev;
struct netdev_queue *txq;
struct Qdisc *q;
int rc = -ENOMEM;
/* Disable soft irqs for various locks below. Also
* stops preemption for RCU.
*/
rcu_read_lock_bh();
txq = dev_pick_tx(dev, skb);
q = rcu_dereference_bh(txq->qdisc);
#ifdef CONFIG_NET_CLS_ACT
skb->tc_verd = SET_TC_AT(skb->tc_verd, AT_EGRESS);
#endif
trace_net_dev_queue(skb);
if (q->enqueue) {
net: Avoid enqueuing skb for default qdiscs dev_queue_xmit enqueue's a skb and calls qdisc_run which dequeue's the skb and xmits it. In most cases, the skb that is enqueue'd is the same one that is dequeue'd (unless the queue gets stopped or multiple cpu's write to the same queue and ends in a race with qdisc_run). For default qdiscs, we can remove the redundant enqueue/dequeue and simply xmit the skb since the default qdisc is work-conserving. The patch uses a new flag - TCQ_F_CAN_BYPASS to identify the default fast queue. The controversial part of the patch is incrementing qlen when a skb is requeued - this is to avoid checks like the second line below: + } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && >> !q->gso_skb && + !test_and_set_bit(__QDISC_STATE_RUNNING, &q->state)) { Results of a 2 hour testing for multiple netperf sessions (1, 2, 4, 8, 12 sessions on a 4 cpu system-X). The BW numbers are aggregate Mb/s across iterations tested with this version on System-X boxes with Chelsio 10gbps cards: ---------------------------------- Size | ORG BW NEW BW | ---------------------------------- 128K | 156964 159381 | 256K | 158650 162042 | ---------------------------------- Changes from ver1: 1. Move sch_direct_xmit declaration from sch_generic.h to pkt_sched.h 2. Update qdisc basic statistics for direct xmit path. 3. Set qlen to zero in qdisc_reset. 4. Changed some function names to more meaningful ones. Signed-off-by: Krishna Kumar <krkumar2@in.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-08-06 01:44:21 +00:00
rc = __dev_xmit_skb(skb, q, dev, txq);
goto out;
}
/* The device has no queue. Common case for software devices:
loopback, all the sorts of tunnels...
[NET]: Add netif_tx_lock Various drivers use xmit_lock internally to synchronise with their transmission routines. They do so without setting xmit_lock_owner. This is fine as long as netpoll is not in use. With netpoll it is possible for deadlocks to occur if xmit_lock_owner isn't set. This is because if a printk occurs while xmit_lock is held and xmit_lock_owner is not set can cause netpoll to attempt to take xmit_lock recursively. While it is possible to resolve this by getting netpoll to use trylock, it is suboptimal because netpoll's sole objective is to maximise the chance of getting the printk out on the wire. So delaying or dropping the message is to be avoided as much as possible. So the only alternative is to always set xmit_lock_owner. The following patch does this by introducing the netif_tx_lock family of functions that take care of setting/unsetting xmit_lock_owner. I renamed xmit_lock to _xmit_lock to indicate that it should not be used directly. I didn't provide irq versions of the netif_tx_lock functions since xmit_lock is meant to be a BH-disabling lock. This is pretty much a straight text substitution except for a small bug fix in winbond. It currently uses netif_stop_queue/spin_unlock_wait to stop transmission. This is unsafe as an IRQ can potentially wake up the queue. So it is safer to use netif_tx_disable. The hamradio bits used spin_lock_irq but it is unnecessary as xmit_lock must never be taken in an IRQ handler. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-06-09 19:20:56 +00:00
Really, it is unlikely that netif_tx_lock protection is necessary
here. (f.e. loopback and IP tunnels are clean ignoring statistics
counters.)
However, it is possible, that they rely on protection
made by us here.
Check this and shot the lock. It is not prone from deadlocks.
Either shot noqueue qdisc, it is even simpler 8)
*/
if (dev->flags & IFF_UP) {
int cpu = smp_processor_id(); /* ok because BHs are off */
if (txq->xmit_lock_owner != cpu) {
if (__this_cpu_read(xmit_recursion) > RECURSION_LIMIT)
goto recursion_alert;
HARD_TX_LOCK(dev, txq, cpu);
if (!netif_tx_queue_stopped(txq)) {
__this_cpu_inc(xmit_recursion);
rc = dev_hard_start_xmit(skb, dev, txq);
__this_cpu_dec(xmit_recursion);
if (dev_xmit_complete(rc)) {
HARD_TX_UNLOCK(dev, txq);
goto out;
}
}
HARD_TX_UNLOCK(dev, txq);
if (net_ratelimit())
printk(KERN_CRIT "Virtual device %s asks to "
"queue packet!\n", dev->name);
} else {
/* Recursion is detected! It is possible,
* unfortunately
*/
recursion_alert:
if (net_ratelimit())
printk(KERN_CRIT "Dead loop on virtual device "
"%s, fix it urgently!\n", dev->name);
}
}
rc = -ENETDOWN;
rcu_read_unlock_bh();
kfree_skb(skb);
return rc;
out:
rcu_read_unlock_bh();
return rc;
}
EXPORT_SYMBOL(dev_queue_xmit);
/*=======================================================================
Receiver routines
=======================================================================*/
int netdev_max_backlog __read_mostly = 1000;
int netdev_tstamp_prequeue __read_mostly = 1;
int netdev_budget __read_mostly = 300;
int weight_p __read_mostly = 64; /* old backlog weight */
/* Called with irq disabled */
static inline void ____napi_schedule(struct softnet_data *sd,
struct napi_struct *napi)
{
list_add_tail(&napi->poll_list, &sd->poll_list);
__raise_softirq_irqoff(NET_RX_SOFTIRQ);
}
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
/*
* __skb_get_rxhash: calculate a flow hash based on src/dst addresses
* and src/dst port numbers. Sets rxhash in skb to non-zero hash value
* on success, zero indicates no valid hash. Also, sets l4_rxhash in skb
* if hash is a canonical 4-tuple hash over transport ports.
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
*/
void __skb_get_rxhash(struct sk_buff *skb)
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
{
int nhoff, hash = 0, poff;
const struct ipv6hdr *ip6;
const struct iphdr *ip;
const struct vlan_hdr *vlan;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
u8 ip_proto;
u32 addr1, addr2;
u16 proto;
union {
u32 v32;
u16 v16[2];
} ports;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
nhoff = skb_network_offset(skb);
proto = skb->protocol;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
again:
switch (proto) {
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
case __constant_htons(ETH_P_IP):
ip:
if (!pskb_may_pull(skb, sizeof(*ip) + nhoff))
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
goto done;
ip = (const struct iphdr *) (skb->data + nhoff);
if (ip_is_fragment(ip))
ip_proto = 0;
else
ip_proto = ip->protocol;
addr1 = (__force u32) ip->saddr;
addr2 = (__force u32) ip->daddr;
nhoff += ip->ihl * 4;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
break;
case __constant_htons(ETH_P_IPV6):
ipv6:
if (!pskb_may_pull(skb, sizeof(*ip6) + nhoff))
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
goto done;
ip6 = (const struct ipv6hdr *) (skb->data + nhoff);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
ip_proto = ip6->nexthdr;
addr1 = (__force u32) ip6->saddr.s6_addr32[3];
addr2 = (__force u32) ip6->daddr.s6_addr32[3];
nhoff += 40;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
break;
case __constant_htons(ETH_P_8021Q):
if (!pskb_may_pull(skb, sizeof(*vlan) + nhoff))
goto done;
vlan = (const struct vlan_hdr *) (skb->data + nhoff);
proto = vlan->h_vlan_encapsulated_proto;
nhoff += sizeof(*vlan);
goto again;
case __constant_htons(ETH_P_PPP_SES):
if (!pskb_may_pull(skb, PPPOE_SES_HLEN + nhoff))
goto done;
proto = *((__be16 *) (skb->data + nhoff +
sizeof(struct pppoe_hdr)));
nhoff += PPPOE_SES_HLEN;
switch (proto) {
case __constant_htons(PPP_IP):
goto ip;
case __constant_htons(PPP_IPV6):
goto ipv6;
default:
goto done;
}
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
default:
goto done;
}
switch (ip_proto) {
case IPPROTO_GRE:
if (pskb_may_pull(skb, nhoff + 16)) {
u8 *h = skb->data + nhoff;
__be16 flags = *(__be16 *)h;
/*
* Only look inside GRE if version zero and no
* routing
*/
if (!(flags & (GRE_VERSION|GRE_ROUTING))) {
proto = *(__be16 *)(h + 2);
nhoff += 4;
if (flags & GRE_CSUM)
nhoff += 4;
if (flags & GRE_KEY)
nhoff += 4;
if (flags & GRE_SEQ)
nhoff += 4;
goto again;
}
}
break;
case IPPROTO_IPIP:
goto again;
default:
break;
}
ports.v32 = 0;
poff = proto_ports_offset(ip_proto);
if (poff >= 0) {
nhoff += poff;
if (pskb_may_pull(skb, nhoff + 4)) {
ports.v32 = * (__force u32 *) (skb->data + nhoff);
if (ports.v16[1] < ports.v16[0])
swap(ports.v16[0], ports.v16[1]);
skb->l4_rxhash = 1;
}
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
}
/* get a consistent hash (same value on both flow directions) */
if (addr2 < addr1)
swap(addr1, addr2);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
hash = jhash_3words(addr1, addr2, ports.v32, hashrnd);
if (!hash)
hash = 1;
done:
skb->rxhash = hash;
}
EXPORT_SYMBOL(__skb_get_rxhash);
#ifdef CONFIG_RPS
/* One global table that all flow-based protocols share. */
struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
EXPORT_SYMBOL(rps_sock_flow_table);
static struct rps_dev_flow *
set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
struct rps_dev_flow *rflow, u16 next_cpu)
{
if (next_cpu != RPS_NO_CPU) {
#ifdef CONFIG_RFS_ACCEL
struct netdev_rx_queue *rxqueue;
struct rps_dev_flow_table *flow_table;
struct rps_dev_flow *old_rflow;
u32 flow_id;
u16 rxq_index;
int rc;
/* Should we steer this flow to a different hardware queue? */
if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
!(dev->features & NETIF_F_NTUPLE))
goto out;
rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
if (rxq_index == skb_get_rx_queue(skb))
goto out;
rxqueue = dev->_rx + rxq_index;
flow_table = rcu_dereference(rxqueue->rps_flow_table);
if (!flow_table)
goto out;
flow_id = skb->rxhash & flow_table->mask;
rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
rxq_index, flow_id);
if (rc < 0)
goto out;
old_rflow = rflow;
rflow = &flow_table->flows[flow_id];
rflow->filter = rc;
if (old_rflow->filter == rflow->filter)
old_rflow->filter = RPS_NO_FILTER;
out:
#endif
rflow->last_qtail =
per_cpu(softnet_data, next_cpu).input_queue_head;
}
rflow->cpu = next_cpu;
return rflow;
}
/*
* get_rps_cpu is called from netif_receive_skb and returns the target
* CPU from the RPS map of the receiving queue for a given skb.
* rcu_read_lock must be held on entry.
*/
static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
struct rps_dev_flow **rflowp)
{
struct netdev_rx_queue *rxqueue;
struct rps_map *map;
struct rps_dev_flow_table *flow_table;
struct rps_sock_flow_table *sock_flow_table;
int cpu = -1;
u16 tcpu;
if (skb_rx_queue_recorded(skb)) {
u16 index = skb_get_rx_queue(skb);
if (unlikely(index >= dev->real_num_rx_queues)) {
WARN_ONCE(dev->real_num_rx_queues > 1,
"%s received packet on queue %u, but number "
"of RX queues is %u\n",
dev->name, index, dev->real_num_rx_queues);
goto done;
}
rxqueue = dev->_rx + index;
} else
rxqueue = dev->_rx;
map = rcu_dereference(rxqueue->rps_map);
if (map) {
if (map->len == 1 &&
!rcu_access_pointer(rxqueue->rps_flow_table)) {
tcpu = map->cpus[0];
if (cpu_online(tcpu))
cpu = tcpu;
goto done;
}
} else if (!rcu_access_pointer(rxqueue->rps_flow_table)) {
goto done;
}
skb_reset_network_header(skb);
if (!skb_get_rxhash(skb))
goto done;
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
flow_table = rcu_dereference(rxqueue->rps_flow_table);
sock_flow_table = rcu_dereference(rps_sock_flow_table);
if (flow_table && sock_flow_table) {
u16 next_cpu;
struct rps_dev_flow *rflow;
rflow = &flow_table->flows[skb->rxhash & flow_table->mask];
tcpu = rflow->cpu;
next_cpu = sock_flow_table->ents[skb->rxhash &
sock_flow_table->mask];
/*
* If the desired CPU (where last recvmsg was done) is
* different from current CPU (one in the rx-queue flow
* table entry), switch if one of the following holds:
* - Current CPU is unset (equal to RPS_NO_CPU).
* - Current CPU is offline.
* - The current CPU's queue tail has advanced beyond the
* last packet that was enqueued using this table entry.
* This guarantees that all previous packets for the flow
* have been dequeued, thus preserving in order delivery.
*/
if (unlikely(tcpu != next_cpu) &&
(tcpu == RPS_NO_CPU || !cpu_online(tcpu) ||
((int)(per_cpu(softnet_data, tcpu).input_queue_head -
rflow->last_qtail)) >= 0))
rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
if (tcpu != RPS_NO_CPU && cpu_online(tcpu)) {
*rflowp = rflow;
cpu = tcpu;
goto done;
}
}
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
if (map) {
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
tcpu = map->cpus[((u64) skb->rxhash * map->len) >> 32];
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
if (cpu_online(tcpu)) {
cpu = tcpu;
goto done;
}
}
done:
return cpu;
}
#ifdef CONFIG_RFS_ACCEL
/**
* rps_may_expire_flow - check whether an RFS hardware filter may be removed
* @dev: Device on which the filter was set
* @rxq_index: RX queue index
* @flow_id: Flow ID passed to ndo_rx_flow_steer()
* @filter_id: Filter ID returned by ndo_rx_flow_steer()
*
* Drivers that implement ndo_rx_flow_steer() should periodically call
* this function for each installed filter and remove the filters for
* which it returns %true.
*/
bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
u32 flow_id, u16 filter_id)
{
struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
struct rps_dev_flow_table *flow_table;
struct rps_dev_flow *rflow;
bool expire = true;
int cpu;
rcu_read_lock();
flow_table = rcu_dereference(rxqueue->rps_flow_table);
if (flow_table && flow_id <= flow_table->mask) {
rflow = &flow_table->flows[flow_id];
cpu = ACCESS_ONCE(rflow->cpu);
if (rflow->filter == filter_id && cpu != RPS_NO_CPU &&
((int)(per_cpu(softnet_data, cpu).input_queue_head -
rflow->last_qtail) <
(int)(10 * flow_table->mask)))
expire = false;
}
rcu_read_unlock();
return expire;
}
EXPORT_SYMBOL(rps_may_expire_flow);
#endif /* CONFIG_RFS_ACCEL */
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
/* Called from hardirq (IPI) context */
static void rps_trigger_softirq(void *data)
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
{
struct softnet_data *sd = data;
____napi_schedule(sd, &sd->backlog);
sd->received_rps++;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
}
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
#endif /* CONFIG_RPS */
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
/*
* Check if this softnet_data structure is another cpu one
* If yes, queue it to our IPI list and return 1
* If no, return 0
*/
static int rps_ipi_queued(struct softnet_data *sd)
{
#ifdef CONFIG_RPS
struct softnet_data *mysd = &__get_cpu_var(softnet_data);
if (sd != mysd) {
sd->rps_ipi_next = mysd->rps_ipi_list;
mysd->rps_ipi_list = sd;
__raise_softirq_irqoff(NET_RX_SOFTIRQ);
return 1;
}
#endif /* CONFIG_RPS */
return 0;
}
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
/*
* enqueue_to_backlog is called to queue an skb to a per CPU backlog
* queue (may be a remote CPU queue).
*/
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
unsigned int *qtail)
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
{
struct softnet_data *sd;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
unsigned long flags;
sd = &per_cpu(softnet_data, cpu);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
local_irq_save(flags);
rps_lock(sd);
if (skb_queue_len(&sd->input_pkt_queue) <= netdev_max_backlog) {
if (skb_queue_len(&sd->input_pkt_queue)) {
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
enqueue:
__skb_queue_tail(&sd->input_pkt_queue, skb);
input_queue_tail_incr_save(sd, qtail);
rps_unlock(sd);
local_irq_restore(flags);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
return NET_RX_SUCCESS;
}
/* Schedule NAPI for backlog device
* We can use non atomic operation since we own the queue lock
*/
if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
if (!rps_ipi_queued(sd))
____napi_schedule(sd, &sd->backlog);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
}
goto enqueue;
}
sd->dropped++;
rps_unlock(sd);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
local_irq_restore(flags);
atomic_long_inc(&skb->dev->rx_dropped);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
kfree_skb(skb);
return NET_RX_DROP;
}
/**
* netif_rx - post buffer to the network code
* @skb: buffer to post
*
* This function receives a packet from a device driver and queues it for
* the upper (protocol) levels to process. It always succeeds. The buffer
* may be dropped during processing for congestion control or by the
* protocol layers.
*
* return values:
* NET_RX_SUCCESS (no congestion)
* NET_RX_DROP (packet was dropped)
*
*/
int netif_rx(struct sk_buff *skb)
{
int ret;
/* if netpoll wants it, pretend we never saw it */
if (netpoll_rx(skb))
return NET_RX_DROP;
if (netdev_tstamp_prequeue)
net_timestamp_check(skb);
trace_netif_rx(skb);
#ifdef CONFIG_RPS
{
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
struct rps_dev_flow voidflow, *rflow = &voidflow;
int cpu;
preempt_disable();
rcu_read_lock();
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
cpu = get_rps_cpu(skb->dev, skb, &rflow);
if (cpu < 0)
cpu = smp_processor_id();
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
rcu_read_unlock();
preempt_enable();
}
#else
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
{
unsigned int qtail;
ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
put_cpu();
}
#endif
return ret;
}
EXPORT_SYMBOL(netif_rx);
int netif_rx_ni(struct sk_buff *skb)
{
int err;
preempt_disable();
err = netif_rx(skb);
if (local_softirq_pending())
do_softirq();
preempt_enable();
return err;
}
EXPORT_SYMBOL(netif_rx_ni);
static void net_tx_action(struct softirq_action *h)
{
struct softnet_data *sd = &__get_cpu_var(softnet_data);
if (sd->completion_queue) {
struct sk_buff *clist;
local_irq_disable();
clist = sd->completion_queue;
sd->completion_queue = NULL;
local_irq_enable();
while (clist) {
struct sk_buff *skb = clist;
clist = clist->next;
WARN_ON(atomic_read(&skb->users));
trace_kfree_skb(skb, net_tx_action);
__kfree_skb(skb);
}
}
if (sd->output_queue) {
struct Qdisc *head;
local_irq_disable();
head = sd->output_queue;
sd->output_queue = NULL;
sd->output_queue_tailp = &sd->output_queue;
local_irq_enable();
while (head) {
struct Qdisc *q = head;
spinlock_t *root_lock;
head = head->next_sched;
root_lock = qdisc_lock(q);
if (spin_trylock(root_lock)) {
smp_mb__before_clear_bit();
clear_bit(__QDISC_STATE_SCHED,
&q->state);
qdisc_run(q);
spin_unlock(root_lock);
} else {
if (!test_bit(__QDISC_STATE_DEACTIVATED,
&q->state)) {
__netif_reschedule(q);
} else {
smp_mb__before_clear_bit();
clear_bit(__QDISC_STATE_SCHED,
&q->state);
}
}
}
}
}
#if (defined(CONFIG_BRIDGE) || defined(CONFIG_BRIDGE_MODULE)) && \
(defined(CONFIG_ATM_LANE) || defined(CONFIG_ATM_LANE_MODULE))
/* This hook is defined here for ATM LANE */
int (*br_fdb_test_addr_hook)(struct net_device *dev,
unsigned char *addr) __read_mostly;
EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
#endif
#ifdef CONFIG_NET_CLS_ACT
/* TODO: Maybe we should just force sch_ingress to be compiled in
* when CONFIG_NET_CLS_ACT is? otherwise some useless instructions
* a compare and 2 stores extra right now if we dont have it on
* but have CONFIG_NET_CLS_ACT
* NOTE: This doesn't stop any functionality; if you dont have
* the ingress scheduler, you just can't add policies on ingress.
*
*/
static int ing_filter(struct sk_buff *skb, struct netdev_queue *rxq)
{
struct net_device *dev = skb->dev;
u32 ttl = G_TC_RTTL(skb->tc_verd);
int result = TC_ACT_OK;
struct Qdisc *q;
if (unlikely(MAX_RED_LOOP < ttl++)) {
if (net_ratelimit())
pr_warning( "Redir loop detected Dropping packet (%d->%d)\n",
skb->skb_iif, dev->ifindex);
return TC_ACT_SHOT;
}
skb->tc_verd = SET_TC_RTTL(skb->tc_verd, ttl);
skb->tc_verd = SET_TC_AT(skb->tc_verd, AT_INGRESS);
q = rxq->qdisc;
if (q != &noop_qdisc) {
spin_lock(qdisc_lock(q));
if (likely(!test_bit(__QDISC_STATE_DEACTIVATED, &q->state)))
result = qdisc_enqueue_root(skb, q);
spin_unlock(qdisc_lock(q));
}
return result;
}
static inline struct sk_buff *handle_ing(struct sk_buff *skb,
struct packet_type **pt_prev,
int *ret, struct net_device *orig_dev)
{
struct netdev_queue *rxq = rcu_dereference(skb->dev->ingress_queue);
if (!rxq || rxq->qdisc == &noop_qdisc)
goto out;
if (*pt_prev) {
*ret = deliver_skb(skb, *pt_prev, orig_dev);
*pt_prev = NULL;
}
switch (ing_filter(skb, rxq)) {
case TC_ACT_SHOT:
case TC_ACT_STOLEN:
kfree_skb(skb);
return NULL;
}
out:
skb->tc_verd = 0;
return skb;
}
#endif
/**
* netdev_rx_handler_register - register receive handler
* @dev: device to register a handler for
* @rx_handler: receive handler to register
* @rx_handler_data: data pointer that is used by rx handler
*
* Register a receive hander for a device. This handler will then be
* called from __netif_receive_skb. A negative errno code is returned
* on a failure.
*
* The caller must hold the rtnl_mutex.
*
* For a general description of rx_handler, see enum rx_handler_result.
*/
int netdev_rx_handler_register(struct net_device *dev,
rx_handler_func_t *rx_handler,
void *rx_handler_data)
{
ASSERT_RTNL();
if (dev->rx_handler)
return -EBUSY;
rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
rcu_assign_pointer(dev->rx_handler, rx_handler);
return 0;
}
EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
/**
* netdev_rx_handler_unregister - unregister receive handler
* @dev: device to unregister a handler from
*
* Unregister a receive hander from a device.
*
* The caller must hold the rtnl_mutex.
*/
void netdev_rx_handler_unregister(struct net_device *dev)
{
ASSERT_RTNL();
RCU_INIT_POINTER(dev->rx_handler, NULL);
RCU_INIT_POINTER(dev->rx_handler_data, NULL);
}
EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
static int __netif_receive_skb(struct sk_buff *skb)
{
struct packet_type *ptype, *pt_prev;
rx_handler_func_t *rx_handler;
struct net_device *orig_dev;
struct net_device *null_or_dev;
bool deliver_exact = false;
int ret = NET_RX_DROP;
__be16 type;
if (!netdev_tstamp_prequeue)
net_timestamp_check(skb);
trace_netif_receive_skb(skb);
/* if we've gotten here through NAPI, check netpoll */
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
if (netpoll_receive_skb(skb))
return NET_RX_DROP;
if (!skb->skb_iif)
skb->skb_iif = skb->dev->ifindex;
orig_dev = skb->dev;
skb_reset_network_header(skb);
skb_reset_transport_header(skb);
skb_reset_mac_len(skb);
pt_prev = NULL;
rcu_read_lock();
another_round:
__this_cpu_inc(softnet_data.processed);
if (skb->protocol == cpu_to_be16(ETH_P_8021Q)) {
skb = vlan_untag(skb);
if (unlikely(!skb))
goto out;
}
#ifdef CONFIG_NET_CLS_ACT
if (skb->tc_verd & TC_NCLS) {
skb->tc_verd = CLR_TC_NCLS(skb->tc_verd);
goto ncls;
}
#endif
list_for_each_entry_rcu(ptype, &ptype_all, list) {
if (!ptype->dev || ptype->dev == skb->dev) {
if (pt_prev)
ret = deliver_skb(skb, pt_prev, orig_dev);
pt_prev = ptype;
}
}
#ifdef CONFIG_NET_CLS_ACT
skb = handle_ing(skb, &pt_prev, &ret, orig_dev);
if (!skb)
goto out;
ncls:
#endif
rx_handler = rcu_dereference(skb->dev->rx_handler);
net: allow vlan traffic to be received under bond The following configuration used to work as I expected. At least we could use the fcoe interfaces to do MPIO and the bond0 iface to do load balancing or failover. ---eth2.228-fcoe | eth2 -----| | |---- bond0 | eth3 -----| | ---eth3.228-fcoe This worked because of a change we added to allow inactive slaves to rx 'exact' matches. This functionality was kept intact with the rx_handler mechanism. However now the vlan interface attached to the active slave never receives traffic because the bonding rx_handler updates the skb->dev and goto's another_round. Previously, the vlan_do_receive() logic was called before the bonding rx_handler. Now by the time vlan_do_receive calls vlan_find_dev() the skb->dev is set to bond0 and it is clear no vlan is attached to this iface. The vlan lookup fails. This patch moves the VLAN check above the rx_handler. A VLAN tagged frame is now routed to the eth2.228-fcoe iface in the above schematic. Untagged frames continue to the bond0 as normal. This case also remains intact, eth2 --> bond0 --> vlan.228 Here the skb is VLAN tagged but the vlan lookup fails on eth2 causing the bonding rx_handler to be called. On the second pass the vlan lookup is on the bond0 iface and completes as expected. Putting a VLAN.228 on both the bond0 and eth2 device will result in eth2.228 receiving the skb. I don't think this is completely unexpected and was the result prior to the rx_handler result. Note, the same setup is also used for other storage traffic that MPIO is used with eg. iSCSI and similar setups can be contrived without storage protocols. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Acked-by: Jesse Gross <jesse@nicira.com> Reviewed-by: Jiri Pirko <jpirko@redhat.com> Tested-by: Hans Schillstrom <hams.schillstrom@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-10-10 09:16:41 +00:00
if (vlan_tx_tag_present(skb)) {
if (pt_prev) {
ret = deliver_skb(skb, pt_prev, orig_dev);
pt_prev = NULL;
}
if (vlan_do_receive(&skb, !rx_handler))
net: allow vlan traffic to be received under bond The following configuration used to work as I expected. At least we could use the fcoe interfaces to do MPIO and the bond0 iface to do load balancing or failover. ---eth2.228-fcoe | eth2 -----| | |---- bond0 | eth3 -----| | ---eth3.228-fcoe This worked because of a change we added to allow inactive slaves to rx 'exact' matches. This functionality was kept intact with the rx_handler mechanism. However now the vlan interface attached to the active slave never receives traffic because the bonding rx_handler updates the skb->dev and goto's another_round. Previously, the vlan_do_receive() logic was called before the bonding rx_handler. Now by the time vlan_do_receive calls vlan_find_dev() the skb->dev is set to bond0 and it is clear no vlan is attached to this iface. The vlan lookup fails. This patch moves the VLAN check above the rx_handler. A VLAN tagged frame is now routed to the eth2.228-fcoe iface in the above schematic. Untagged frames continue to the bond0 as normal. This case also remains intact, eth2 --> bond0 --> vlan.228 Here the skb is VLAN tagged but the vlan lookup fails on eth2 causing the bonding rx_handler to be called. On the second pass the vlan lookup is on the bond0 iface and completes as expected. Putting a VLAN.228 on both the bond0 and eth2 device will result in eth2.228 receiving the skb. I don't think this is completely unexpected and was the result prior to the rx_handler result. Note, the same setup is also used for other storage traffic that MPIO is used with eg. iSCSI and similar setups can be contrived without storage protocols. Signed-off-by: John Fastabend <john.r.fastabend@intel.com> Acked-by: Jesse Gross <jesse@nicira.com> Reviewed-by: Jiri Pirko <jpirko@redhat.com> Tested-by: Hans Schillstrom <hams.schillstrom@ericsson.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-10-10 09:16:41 +00:00
goto another_round;
else if (unlikely(!skb))
goto out;
}
if (rx_handler) {
if (pt_prev) {
ret = deliver_skb(skb, pt_prev, orig_dev);
pt_prev = NULL;
}
switch (rx_handler(&skb)) {
case RX_HANDLER_CONSUMED:
goto out;
case RX_HANDLER_ANOTHER:
goto another_round;
case RX_HANDLER_EXACT:
deliver_exact = true;
case RX_HANDLER_PASS:
break;
default:
BUG();
}
}
/* deliver only exact match when indicated */
null_or_dev = deliver_exact ? skb->dev : NULL;
bonding: allow arp_ip_targets on separate vlans to use arp validation This allows a bond device to specify an arp_ip_target as a host that is not on the same vlan as the base bond device and still use arp validation. A configuration like this, now works: BONDING_OPTS="mode=active-backup arp_interval=1000 arp_ip_target=10.0.100.1 arp_validate=3" 1: lo: <LOOPBACK,UP,LOWER_UP> mtu 16436 qdisc noqueue link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00 inet 127.0.0.1/8 scope host lo inet6 ::1/128 scope host valid_lft forever preferred_lft forever 2: eth1: <BROADCAST,MULTICAST,SLAVE,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast master bond0 qlen 1000 link/ether 00:13:21:be:33:e9 brd ff:ff:ff:ff:ff:ff 3: eth0: <BROADCAST,MULTICAST,SLAVE,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast master bond0 qlen 1000 link/ether 00:13:21:be:33:e9 brd ff:ff:ff:ff:ff:ff 8: bond0: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc noqueue link/ether 00:13:21:be:33:e9 brd ff:ff:ff:ff:ff:ff inet6 fe80::213:21ff:febe:33e9/64 scope link valid_lft forever preferred_lft forever 9: bond0.100@bond0: <BROADCAST,MULTICAST,MASTER,UP,LOWER_UP> mtu 1500 qdisc noqueue link/ether 00:13:21:be:33:e9 brd ff:ff:ff:ff:ff:ff inet 10.0.100.2/24 brd 10.0.100.255 scope global bond0.100 inet6 fe80::213:21ff:febe:33e9/64 scope link valid_lft forever preferred_lft forever Ethernet Channel Bonding Driver: v3.6.0 (September 26, 2009) Bonding Mode: fault-tolerance (active-backup) Primary Slave: None Currently Active Slave: eth1 MII Status: up MII Polling Interval (ms): 0 Up Delay (ms): 0 Down Delay (ms): 0 ARP Polling Interval (ms): 1000 ARP IP target/s (n.n.n.n form): 10.0.100.1 Slave Interface: eth1 MII Status: up Link Failure Count: 1 Permanent HW addr: 00:40:05:30:ff:30 Slave Interface: eth0 MII Status: up Link Failure Count: 0 Permanent HW addr: 00:13:21:be:33:e9 Signed-off-by: Andy Gospodarek <andy@greyhouse.net> Signed-off-by: Jay Vosburgh <fubar@us.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-12-14 10:48:58 +00:00
type = skb->protocol;
list_for_each_entry_rcu(ptype,
&ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
if (ptype->type == type &&
(ptype->dev == null_or_dev || ptype->dev == skb->dev ||
ptype->dev == orig_dev)) {
if (pt_prev)
ret = deliver_skb(skb, pt_prev, orig_dev);
pt_prev = ptype;
}
}
if (pt_prev) {
ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
} else {
atomic_long_inc(&skb->dev->rx_dropped);
kfree_skb(skb);
/* Jamal, now you will not able to escape explaining
* me how you were going to use this. :-)
*/
ret = NET_RX_DROP;
}
out:
rcu_read_unlock();
return ret;
}
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
/**
* netif_receive_skb - process receive buffer from network
* @skb: buffer to process
*
* netif_receive_skb() is the main receive data processing function.
* It always succeeds. The buffer may be dropped during processing
* for congestion control or by the protocol layers.
*
* This function may only be called from softirq context and interrupts
* should be enabled.
*
* Return values (usually ignored):
* NET_RX_SUCCESS: no congestion
* NET_RX_DROP: packet was dropped
*/
int netif_receive_skb(struct sk_buff *skb)
{
if (netdev_tstamp_prequeue)
net_timestamp_check(skb);
if (skb_defer_rx_timestamp(skb))
return NET_RX_SUCCESS;
#ifdef CONFIG_RPS
{
struct rps_dev_flow voidflow, *rflow = &voidflow;
int cpu, ret;
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
rcu_read_lock();
cpu = get_rps_cpu(skb->dev, skb, &rflow);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
if (cpu >= 0) {
ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
rcu_read_unlock();
} else {
rcu_read_unlock();
ret = __netif_receive_skb(skb);
}
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
return ret;
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
}
#else
return __netif_receive_skb(skb);
#endif
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
}
EXPORT_SYMBOL(netif_receive_skb);
/* Network device is going away, flush any packets still pending
* Called with irqs disabled.
*/
static void flush_backlog(void *arg)
{
struct net_device *dev = arg;
struct softnet_data *sd = &__get_cpu_var(softnet_data);
struct sk_buff *skb, *tmp;
rps_lock(sd);
skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
if (skb->dev == dev) {
__skb_unlink(skb, &sd->input_pkt_queue);
kfree_skb(skb);
input_queue_head_incr(sd);
}
}
rps_unlock(sd);
skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
if (skb->dev == dev) {
__skb_unlink(skb, &sd->process_queue);
kfree_skb(skb);
input_queue_head_incr(sd);
}
}
}
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
static int napi_gro_complete(struct sk_buff *skb)
{
struct packet_type *ptype;
__be16 type = skb->protocol;
struct list_head *head = &ptype_base[ntohs(type) & PTYPE_HASH_MASK];
int err = -ENOENT;
if (NAPI_GRO_CB(skb)->count == 1) {
skb_shinfo(skb)->gso_size = 0;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
goto out;
}
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
rcu_read_lock();
list_for_each_entry_rcu(ptype, head, list) {
if (ptype->type != type || ptype->dev || !ptype->gro_complete)
continue;
err = ptype->gro_complete(skb);
break;
}
rcu_read_unlock();
if (err) {
WARN_ON(&ptype->list == head);
kfree_skb(skb);
return NET_RX_SUCCESS;
}
out:
return netif_receive_skb(skb);
}
inline void napi_gro_flush(struct napi_struct *napi)
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
{
struct sk_buff *skb, *next;
for (skb = napi->gro_list; skb; skb = next) {
next = skb->next;
skb->next = NULL;
napi_gro_complete(skb);
}
napi->gro_count = 0;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
napi->gro_list = NULL;
}
EXPORT_SYMBOL(napi_gro_flush);
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
{
struct sk_buff **pp = NULL;
struct packet_type *ptype;
__be16 type = skb->protocol;
struct list_head *head = &ptype_base[ntohs(type) & PTYPE_HASH_MASK];
int same_flow;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
int mac_len;
enum gro_result ret;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
if (!(skb->dev->features & NETIF_F_GRO) || netpoll_rx_on(skb))
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
goto normal;
if (skb_is_gso(skb) || skb_has_frag_list(skb))
goto normal;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
rcu_read_lock();
list_for_each_entry_rcu(ptype, head, list) {
if (ptype->type != type || ptype->dev || !ptype->gro_receive)
continue;
skb_set_network_header(skb, skb_gro_offset(skb));
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
mac_len = skb->network_header - skb->mac_header;
skb->mac_len = mac_len;
NAPI_GRO_CB(skb)->same_flow = 0;
NAPI_GRO_CB(skb)->flush = 0;
NAPI_GRO_CB(skb)->free = 0;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
pp = ptype->gro_receive(&napi->gro_list, skb);
break;
}
rcu_read_unlock();
if (&ptype->list == head)
goto normal;
same_flow = NAPI_GRO_CB(skb)->same_flow;
ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
if (pp) {
struct sk_buff *nskb = *pp;
*pp = nskb->next;
nskb->next = NULL;
napi_gro_complete(nskb);
napi->gro_count--;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
}
if (same_flow)
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
goto ok;
if (NAPI_GRO_CB(skb)->flush || napi->gro_count >= MAX_GRO_SKBS)
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
goto normal;
napi->gro_count++;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
NAPI_GRO_CB(skb)->count = 1;
skb_shinfo(skb)->gso_size = skb_gro_len(skb);
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
skb->next = napi->gro_list;
napi->gro_list = skb;
ret = GRO_HELD;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
pull:
if (skb_headlen(skb) < skb_gro_offset(skb)) {
int grow = skb_gro_offset(skb) - skb_headlen(skb);
BUG_ON(skb->end - skb->tail < grow);
memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
skb->tail += grow;
skb->data_len -= grow;
skb_shinfo(skb)->frags[0].page_offset += grow;
skb_frag_size_sub(&skb_shinfo(skb)->frags[0], grow);
if (unlikely(!skb_frag_size(&skb_shinfo(skb)->frags[0]))) {
skb_frag_unref(skb, 0);
memmove(skb_shinfo(skb)->frags,
skb_shinfo(skb)->frags + 1,
--skb_shinfo(skb)->nr_frags * sizeof(skb_frag_t));
}
}
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
ok:
return ret;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
normal:
ret = GRO_NORMAL;
goto pull;
}
EXPORT_SYMBOL(dev_gro_receive);
static inline gro_result_t
__napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
{
struct sk_buff *p;
for (p = napi->gro_list; p; p = p->next) {
unsigned long diffs;
diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
vlan: Centralize handling of hardware acceleration. Currently each driver that is capable of vlan hardware acceleration must be aware of the vlan groups that are configured and then pass the stripped tag to a specialized receive function. This is different from other types of hardware offload in that it places a significant amount of knowledge in the driver itself rather keeping it in the networking core. This makes vlan offloading function more similarly to other forms of offloading (such as checksum offloading or TSO) by doing the following: * On receive, stripped vlans are passed directly to the network core, without attempting to check for vlan groups or reconstructing the header if no group * vlans are made less special by folding the logic into the main receive routines * On transmit, the device layer will add the vlan header in software if the hardware doesn't support it, instead of spreading that logic out in upper layers, such as bonding. There are a number of advantages to this: * Fixes all bugs with drivers incorrectly dropping vlan headers at once. * Avoids having to disable VLAN acceleration when in promiscuous mode (good for bridging since it always puts devices in promiscuous mode). * Keeps VLAN tag separate until given to ultimate consumer, which avoids needing to do header reconstruction as in tg3 unless absolutely necessary. * Consolidates common code in core networking. Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-10-20 13:56:06 +00:00
diffs |= p->vlan_tci ^ skb->vlan_tci;
diffs |= compare_ether_header(skb_mac_header(p),
skb_gro_mac_header(skb));
NAPI_GRO_CB(p)->same_flow = !diffs;
NAPI_GRO_CB(p)->flush = 0;
}
return dev_gro_receive(napi, skb);
}
gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
{
switch (ret) {
case GRO_NORMAL:
if (netif_receive_skb(skb))
ret = GRO_DROP;
break;
case GRO_DROP:
case GRO_MERGED_FREE:
kfree_skb(skb);
break;
case GRO_HELD:
case GRO_MERGED:
break;
}
return ret;
}
EXPORT_SYMBOL(napi_skb_finish);
void skb_gro_reset_offset(struct sk_buff *skb)
{
NAPI_GRO_CB(skb)->data_offset = 0;
NAPI_GRO_CB(skb)->frag0 = NULL;
NAPI_GRO_CB(skb)->frag0_len = 0;
if (skb->mac_header == skb->tail &&
!PageHighMem(skb_frag_page(&skb_shinfo(skb)->frags[0]))) {
NAPI_GRO_CB(skb)->frag0 =
skb_frag_address(&skb_shinfo(skb)->frags[0]);
NAPI_GRO_CB(skb)->frag0_len = skb_frag_size(&skb_shinfo(skb)->frags[0]);
}
}
EXPORT_SYMBOL(skb_gro_reset_offset);
gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
{
skb_gro_reset_offset(skb);
return napi_skb_finish(__napi_gro_receive(napi, skb), skb);
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
}
EXPORT_SYMBOL(napi_gro_receive);
static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
{
__skb_pull(skb, skb_headlen(skb));
skb_reserve(skb, NET_IP_ALIGN - skb_headroom(skb));
vlan: Centralize handling of hardware acceleration. Currently each driver that is capable of vlan hardware acceleration must be aware of the vlan groups that are configured and then pass the stripped tag to a specialized receive function. This is different from other types of hardware offload in that it places a significant amount of knowledge in the driver itself rather keeping it in the networking core. This makes vlan offloading function more similarly to other forms of offloading (such as checksum offloading or TSO) by doing the following: * On receive, stripped vlans are passed directly to the network core, without attempting to check for vlan groups or reconstructing the header if no group * vlans are made less special by folding the logic into the main receive routines * On transmit, the device layer will add the vlan header in software if the hardware doesn't support it, instead of spreading that logic out in upper layers, such as bonding. There are a number of advantages to this: * Fixes all bugs with drivers incorrectly dropping vlan headers at once. * Avoids having to disable VLAN acceleration when in promiscuous mode (good for bridging since it always puts devices in promiscuous mode). * Keeps VLAN tag separate until given to ultimate consumer, which avoids needing to do header reconstruction as in tg3 unless absolutely necessary. * Consolidates common code in core networking. Signed-off-by: Jesse Gross <jesse@nicira.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-10-20 13:56:06 +00:00
skb->vlan_tci = 0;
skb->dev = napi->dev;
skb->skb_iif = 0;
napi->skb = skb;
}
struct sk_buff *napi_get_frags(struct napi_struct *napi)
{
struct sk_buff *skb = napi->skb;
if (!skb) {
skb = netdev_alloc_skb_ip_align(napi->dev, GRO_MAX_HEAD);
if (skb)
napi->skb = skb;
}
return skb;
}
EXPORT_SYMBOL(napi_get_frags);
gro_result_t napi_frags_finish(struct napi_struct *napi, struct sk_buff *skb,
gro_result_t ret)
{
switch (ret) {
case GRO_NORMAL:
case GRO_HELD:
skb->protocol = eth_type_trans(skb, skb->dev);
if (ret == GRO_HELD)
skb_gro_pull(skb, -ETH_HLEN);
else if (netif_receive_skb(skb))
ret = GRO_DROP;
break;
case GRO_DROP:
case GRO_MERGED_FREE:
napi_reuse_skb(napi, skb);
break;
case GRO_MERGED:
break;
}
return ret;
}
EXPORT_SYMBOL(napi_frags_finish);
struct sk_buff *napi_frags_skb(struct napi_struct *napi)
{
struct sk_buff *skb = napi->skb;
struct ethhdr *eth;
unsigned int hlen;
unsigned int off;
napi->skb = NULL;
skb_reset_mac_header(skb);
skb_gro_reset_offset(skb);
off = skb_gro_offset(skb);
hlen = off + sizeof(*eth);
eth = skb_gro_header_fast(skb, off);
if (skb_gro_header_hard(skb, hlen)) {
eth = skb_gro_header_slow(skb, hlen, off);
if (unlikely(!eth)) {
napi_reuse_skb(napi, skb);
skb = NULL;
goto out;
}
}
skb_gro_pull(skb, sizeof(*eth));
/*
* This works because the only protocols we care about don't require
* special handling. We'll fix it up properly at the end.
*/
skb->protocol = eth->h_proto;
out:
return skb;
}
EXPORT_SYMBOL(napi_frags_skb);
gro_result_t napi_gro_frags(struct napi_struct *napi)
{
struct sk_buff *skb = napi_frags_skb(napi);
if (!skb)
return GRO_DROP;
return napi_frags_finish(napi, skb, __napi_gro_receive(napi, skb));
}
EXPORT_SYMBOL(napi_gro_frags);
/*
* net_rps_action sends any pending IPI's for rps.
* Note: called with local irq disabled, but exits with local irq enabled.
*/
static void net_rps_action_and_irq_enable(struct softnet_data *sd)
{
#ifdef CONFIG_RPS
struct softnet_data *remsd = sd->rps_ipi_list;
if (remsd) {
sd->rps_ipi_list = NULL;
local_irq_enable();
/* Send pending IPI's to kick RPS processing on remote cpus. */
while (remsd) {
struct softnet_data *next = remsd->rps_ipi_next;
if (cpu_online(remsd->cpu))
__smp_call_function_single(remsd->cpu,
&remsd->csd, 0);
remsd = next;
}
} else
#endif
local_irq_enable();
}
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
static int process_backlog(struct napi_struct *napi, int quota)
{
int work = 0;
struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
#ifdef CONFIG_RPS
/* Check if we have pending ipi, its better to send them now,
* not waiting net_rx_action() end.
*/
if (sd->rps_ipi_list) {
local_irq_disable();
net_rps_action_and_irq_enable(sd);
}
#endif
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
napi->weight = weight_p;
local_irq_disable();
while (work < quota) {
struct sk_buff *skb;
unsigned int qlen;
while ((skb = __skb_dequeue(&sd->process_queue))) {
local_irq_enable();
__netif_receive_skb(skb);
local_irq_disable();
input_queue_head_incr(sd);
if (++work >= quota) {
local_irq_enable();
return work;
}
}
rps_lock(sd);
qlen = skb_queue_len(&sd->input_pkt_queue);
if (qlen)
skb_queue_splice_tail_init(&sd->input_pkt_queue,
&sd->process_queue);
if (qlen < quota - work) {
/*
* Inline a custom version of __napi_complete().
* only current cpu owns and manipulates this napi,
* and NAPI_STATE_SCHED is the only possible flag set on backlog.
* we can use a plain write instead of clear_bit(),
* and we dont need an smp_mb() memory barrier.
*/
list_del(&napi->poll_list);
napi->state = 0;
quota = work + qlen;
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
}
rps_unlock(sd);
}
local_irq_enable();
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
return work;
}
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
/**
* __napi_schedule - schedule for receive
* @n: entry to schedule
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
*
* The entry's receive function will be scheduled to run
*/
void __napi_schedule(struct napi_struct *n)
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
{
unsigned long flags;
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
local_irq_save(flags);
____napi_schedule(&__get_cpu_var(softnet_data), n);
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
local_irq_restore(flags);
}
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
EXPORT_SYMBOL(__napi_schedule);
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
void __napi_complete(struct napi_struct *n)
{
BUG_ON(!test_bit(NAPI_STATE_SCHED, &n->state));
BUG_ON(n->gro_list);
list_del(&n->poll_list);
smp_mb__before_clear_bit();
clear_bit(NAPI_STATE_SCHED, &n->state);
}
EXPORT_SYMBOL(__napi_complete);
void napi_complete(struct napi_struct *n)
{
unsigned long flags;
/*
* don't let napi dequeue from the cpu poll list
* just in case its running on a different cpu
*/
if (unlikely(test_bit(NAPI_STATE_NPSVC, &n->state)))
return;
napi_gro_flush(n);
local_irq_save(flags);
__napi_complete(n);
local_irq_restore(flags);
}
EXPORT_SYMBOL(napi_complete);
void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
int (*poll)(struct napi_struct *, int), int weight)
{
INIT_LIST_HEAD(&napi->poll_list);
napi->gro_count = 0;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
napi->gro_list = NULL;
napi->skb = NULL;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
napi->poll = poll;
napi->weight = weight;
list_add(&napi->dev_list, &dev->napi_list);
napi->dev = dev;
#ifdef CONFIG_NETPOLL
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
spin_lock_init(&napi->poll_lock);
napi->poll_owner = -1;
#endif
set_bit(NAPI_STATE_SCHED, &napi->state);
}
EXPORT_SYMBOL(netif_napi_add);
void netif_napi_del(struct napi_struct *napi)
{
struct sk_buff *skb, *next;
list_del_init(&napi->dev_list);
napi_free_frags(napi);
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
for (skb = napi->gro_list; skb; skb = next) {
next = skb->next;
skb->next = NULL;
kfree_skb(skb);
}
napi->gro_list = NULL;
napi->gro_count = 0;
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
}
EXPORT_SYMBOL(netif_napi_del);
static void net_rx_action(struct softirq_action *h)
{
struct softnet_data *sd = &__get_cpu_var(softnet_data);
unsigned long time_limit = jiffies + 2;
int budget = netdev_budget;
void *have;
local_irq_disable();
while (!list_empty(&sd->poll_list)) {
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
struct napi_struct *n;
int work, weight;
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
/* If softirq window is exhuasted then punt.
* Allow this to run for 2 jiffies since which will allow
* an average latency of 1.5/HZ.
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
*/
if (unlikely(budget <= 0 || time_after(jiffies, time_limit)))
goto softnet_break;
local_irq_enable();
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
/* Even though interrupts have been re-enabled, this
* access is safe because interrupts can only add new
* entries to the tail of this list, and only ->poll()
* calls can remove this head entry from the list.
*/
n = list_first_entry(&sd->poll_list, struct napi_struct, poll_list);
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
have = netpoll_poll_lock(n);
weight = n->weight;
/* This NAPI_STATE_SCHED test is for avoiding a race
* with netpoll's poll_napi(). Only the entity which
* obtains the lock and sees NAPI_STATE_SCHED set will
* actually make the ->poll() call. Therefore we avoid
* accidentally calling ->poll() when NAPI is not scheduled.
*/
work = 0;
if (test_bit(NAPI_STATE_SCHED, &n->state)) {
work = n->poll(n, weight);
trace_napi_poll(n);
}
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
WARN_ON_ONCE(work > weight);
budget -= work;
local_irq_disable();
/* Drivers must not modify the NAPI state if they
* consume the entire weight. In such cases this code
* still "owns" the NAPI instance and therefore can
* move the instance around on the list at-will.
*/
if (unlikely(work == weight)) {
if (unlikely(napi_disable_pending(n))) {
local_irq_enable();
napi_complete(n);
local_irq_disable();
} else
list_move_tail(&n->poll_list, &sd->poll_list);
}
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
netpoll_poll_unlock(have);
}
out:
net_rps_action_and_irq_enable(sd);
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
#ifdef CONFIG_NET_DMA
/*
* There may not be any more sk_buffs coming right now, so push
* any pending DMA copies to hardware
*/
dma_issue_pending_all();
#endif
[NET]: Make NAPI polling independent of struct net_device objects. Several devices have multiple independant RX queues per net device, and some have a single interrupt doorbell for several queues. In either case, it's easier to support layouts like that if the structure representing the poll is independant from the net device itself. The signature of the ->poll() call back goes from: int foo_poll(struct net_device *dev, int *budget) to int foo_poll(struct napi_struct *napi, int budget) The caller is returned the number of RX packets processed (or the number of "NAPI credits" consumed if you want to get abstract). The callee no longer messes around bumping dev->quota, *budget, etc. because that is all handled in the caller upon return. The napi_struct is to be embedded in the device driver private data structures. Furthermore, it is the driver's responsibility to disable all NAPI instances in it's ->stop() device close handler. Since the napi_struct is privatized into the driver's private data structures, only the driver knows how to get at all of the napi_struct instances it may have per-device. With lots of help and suggestions from Rusty Russell, Roland Dreier, Michael Chan, Jeff Garzik, and Jamal Hadi Salim. Bug fixes from Thomas Graf, Roland Dreier, Peter Zijlstra, Joseph Fannin, Scott Wood, Hans J. Koch, and Michael Chan. [ Ported to current tree and all drivers converted. Integrated Stephen's follow-on kerneldoc additions, and restored poll_list handling to the old style to fix mutual exclusion issues. -DaveM ] Signed-off-by: Stephen Hemminger <shemminger@linux-foundation.org> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-03 23:41:36 +00:00
return;
softnet_break:
sd->time_squeeze++;
__raise_softirq_irqoff(NET_RX_SOFTIRQ);
goto out;
}
static gifconf_func_t *gifconf_list[NPROTO];
/**
* register_gifconf - register a SIOCGIF handler
* @family: Address family
* @gifconf: Function handler
*
* Register protocol dependent address dumping routines. The handler
* that is passed must not be freed or reused until it has been replaced
* by another handler.
*/
int register_gifconf(unsigned int family, gifconf_func_t *gifconf)
{
if (family >= NPROTO)
return -EINVAL;
gifconf_list[family] = gifconf;
return 0;
}
EXPORT_SYMBOL(register_gifconf);
/*
* Map an interface index to its name (SIOCGIFNAME)
*/
/*
* We need this ioctl for efficient implementation of the
* if_indextoname() function required by the IPv6 API. Without
* it, we would have to search all the interfaces to find a
* match. --pb
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
static int dev_ifname(struct net *net, struct ifreq __user *arg)
{
struct net_device *dev;
struct ifreq ifr;
/*
* Fetch the caller's info block.
*/
if (copy_from_user(&ifr, arg, sizeof(struct ifreq)))
return -EFAULT;
rcu_read_lock();
dev = dev_get_by_index_rcu(net, ifr.ifr_ifindex);
if (!dev) {
rcu_read_unlock();
return -ENODEV;
}
strcpy(ifr.ifr_name, dev->name);
rcu_read_unlock();
if (copy_to_user(arg, &ifr, sizeof(struct ifreq)))
return -EFAULT;
return 0;
}
/*
* Perform a SIOCGIFCONF call. This structure will change
* size eventually, and there is nothing I can do about it.
* Thus we will need a 'compatibility mode'.
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
static int dev_ifconf(struct net *net, char __user *arg)
{
struct ifconf ifc;
struct net_device *dev;
char __user *pos;
int len;
int total;
int i;
/*
* Fetch the caller's info block.
*/
if (copy_from_user(&ifc, arg, sizeof(struct ifconf)))
return -EFAULT;
pos = ifc.ifc_buf;
len = ifc.ifc_len;
/*
* Loop over the interfaces, and write an info block for each.
*/
total = 0;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
for_each_netdev(net, dev) {
for (i = 0; i < NPROTO; i++) {
if (gifconf_list[i]) {
int done;
if (!pos)
done = gifconf_list[i](dev, NULL, 0);
else
done = gifconf_list[i](dev, pos + total,
len - total);
if (done < 0)
return -EFAULT;
total += done;
}
}
}
/*
* All done. Write the updated control block back to the caller.
*/
ifc.ifc_len = total;
/*
* Both BSD and Solaris return 0 here, so we do too.
*/
return copy_to_user(arg, &ifc, sizeof(struct ifconf)) ? -EFAULT : 0;
}
#ifdef CONFIG_PROC_FS
#define BUCKET_SPACE (32 - NETDEV_HASHBITS)
struct dev_iter_state {
struct seq_net_private p;
unsigned int pos; /* bucket << BUCKET_SPACE + offset */
};
#define get_bucket(x) ((x) >> BUCKET_SPACE)
#define get_offset(x) ((x) & ((1 << BUCKET_SPACE) - 1))
#define set_bucket_offset(b, o) ((b) << BUCKET_SPACE | (o))
static inline struct net_device *dev_from_same_bucket(struct seq_file *seq)
{
struct dev_iter_state *state = seq->private;
struct net *net = seq_file_net(seq);
struct net_device *dev;
struct hlist_node *p;
struct hlist_head *h;
unsigned int count, bucket, offset;
bucket = get_bucket(state->pos);
offset = get_offset(state->pos);
h = &net->dev_name_head[bucket];
count = 0;
hlist_for_each_entry_rcu(dev, p, h, name_hlist) {
if (count++ == offset) {
state->pos = set_bucket_offset(bucket, count);
return dev;
}
}
return NULL;
}
static inline struct net_device *dev_from_new_bucket(struct seq_file *seq)
{
struct dev_iter_state *state = seq->private;
struct net_device *dev;
unsigned int bucket;
bucket = get_bucket(state->pos);
do {
dev = dev_from_same_bucket(seq);
if (dev)
return dev;
bucket++;
state->pos = set_bucket_offset(bucket, 0);
} while (bucket < NETDEV_HASHENTRIES);
return NULL;
}
/*
* This is invoked by the /proc filesystem handler to display a device
* in detail.
*/
void *dev_seq_start(struct seq_file *seq, loff_t *pos)
__acquires(RCU)
{
struct dev_iter_state *state = seq->private;
rcu_read_lock();
if (!*pos)
return SEQ_START_TOKEN;
/* check for end of the hash */
if (state->pos == 0 && *pos > 1)
return NULL;
return dev_from_new_bucket(seq);
}
void *dev_seq_next(struct seq_file *seq, void *v, loff_t *pos)
{
struct net_device *dev;
++*pos;
if (v == SEQ_START_TOKEN)
return dev_from_new_bucket(seq);
dev = dev_from_same_bucket(seq);
if (dev)
return dev;
return dev_from_new_bucket(seq);
}
void dev_seq_stop(struct seq_file *seq, void *v)
__releases(RCU)
{
rcu_read_unlock();
}
static void dev_seq_printf_stats(struct seq_file *seq, struct net_device *dev)
{
struct rtnl_link_stats64 temp;
const struct rtnl_link_stats64 *stats = dev_get_stats(dev, &temp);
seq_printf(seq, "%6s: %7llu %7llu %4llu %4llu %4llu %5llu %10llu %9llu "
"%8llu %7llu %4llu %4llu %4llu %5llu %7llu %10llu\n",
dev->name, stats->rx_bytes, stats->rx_packets,
stats->rx_errors,
stats->rx_dropped + stats->rx_missed_errors,
stats->rx_fifo_errors,
stats->rx_length_errors + stats->rx_over_errors +
stats->rx_crc_errors + stats->rx_frame_errors,
stats->rx_compressed, stats->multicast,
stats->tx_bytes, stats->tx_packets,
stats->tx_errors, stats->tx_dropped,
stats->tx_fifo_errors, stats->collisions,
stats->tx_carrier_errors +
stats->tx_aborted_errors +
stats->tx_window_errors +
stats->tx_heartbeat_errors,
stats->tx_compressed);
}
/*
* Called from the PROCfs module. This now uses the new arbitrary sized
* /proc/net interface to create /proc/net/dev
*/
static int dev_seq_show(struct seq_file *seq, void *v)
{
if (v == SEQ_START_TOKEN)
seq_puts(seq, "Inter-| Receive "
" | Transmit\n"
" face |bytes packets errs drop fifo frame "
"compressed multicast|bytes packets errs "
"drop fifo colls carrier compressed\n");
else
dev_seq_printf_stats(seq, v);
return 0;
}
static struct softnet_data *softnet_get_online(loff_t *pos)
{
struct softnet_data *sd = NULL;
while (*pos < nr_cpu_ids)
if (cpu_online(*pos)) {
sd = &per_cpu(softnet_data, *pos);
break;
} else
++*pos;
return sd;
}
static void *softnet_seq_start(struct seq_file *seq, loff_t *pos)
{
return softnet_get_online(pos);
}
static void *softnet_seq_next(struct seq_file *seq, void *v, loff_t *pos)
{
++*pos;
return softnet_get_online(pos);
}
static void softnet_seq_stop(struct seq_file *seq, void *v)
{
}
static int softnet_seq_show(struct seq_file *seq, void *v)
{
struct softnet_data *sd = v;
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
seq_printf(seq, "%08x %08x %08x %08x %08x %08x %08x %08x %08x %08x\n",
sd->processed, sd->dropped, sd->time_squeeze, 0,
0, 0, 0, 0, /* was fastroute */
sd->cpu_collision, sd->received_rps);
return 0;
}
static const struct seq_operations dev_seq_ops = {
.start = dev_seq_start,
.next = dev_seq_next,
.stop = dev_seq_stop,
.show = dev_seq_show,
};
static int dev_seq_open(struct inode *inode, struct file *file)
{
return seq_open_net(inode, file, &dev_seq_ops,
sizeof(struct dev_iter_state));
}
static const struct file_operations dev_seq_fops = {
.owner = THIS_MODULE,
.open = dev_seq_open,
.read = seq_read,
.llseek = seq_lseek,
.release = seq_release_net,
};
static const struct seq_operations softnet_seq_ops = {
.start = softnet_seq_start,
.next = softnet_seq_next,
.stop = softnet_seq_stop,
.show = softnet_seq_show,
};
static int softnet_seq_open(struct inode *inode, struct file *file)
{
return seq_open(file, &softnet_seq_ops);
}
static const struct file_operations softnet_seq_fops = {
.owner = THIS_MODULE,
.open = softnet_seq_open,
.read = seq_read,
.llseek = seq_lseek,
.release = seq_release,
};
static void *ptype_get_idx(loff_t pos)
{
struct packet_type *pt = NULL;
loff_t i = 0;
int t;
list_for_each_entry_rcu(pt, &ptype_all, list) {
if (i == pos)
return pt;
++i;
}
for (t = 0; t < PTYPE_HASH_SIZE; t++) {
list_for_each_entry_rcu(pt, &ptype_base[t], list) {
if (i == pos)
return pt;
++i;
}
}
return NULL;
}
static void *ptype_seq_start(struct seq_file *seq, loff_t *pos)
__acquires(RCU)
{
rcu_read_lock();
return *pos ? ptype_get_idx(*pos - 1) : SEQ_START_TOKEN;
}
static void *ptype_seq_next(struct seq_file *seq, void *v, loff_t *pos)
{
struct packet_type *pt;
struct list_head *nxt;
int hash;
++*pos;
if (v == SEQ_START_TOKEN)
return ptype_get_idx(0);
pt = v;
nxt = pt->list.next;
if (pt->type == htons(ETH_P_ALL)) {
if (nxt != &ptype_all)
goto found;
hash = 0;
nxt = ptype_base[0].next;
} else
hash = ntohs(pt->type) & PTYPE_HASH_MASK;
while (nxt == &ptype_base[hash]) {
if (++hash >= PTYPE_HASH_SIZE)
return NULL;
nxt = ptype_base[hash].next;
}
found:
return list_entry(nxt, struct packet_type, list);
}
static void ptype_seq_stop(struct seq_file *seq, void *v)
__releases(RCU)
{
rcu_read_unlock();
}
static int ptype_seq_show(struct seq_file *seq, void *v)
{
struct packet_type *pt = v;
if (v == SEQ_START_TOKEN)
seq_puts(seq, "Type Device Function\n");
else if (pt->dev == NULL || dev_net(pt->dev) == seq_file_net(seq)) {
if (pt->type == htons(ETH_P_ALL))
seq_puts(seq, "ALL ");
else
seq_printf(seq, "%04x", ntohs(pt->type));
seq_printf(seq, " %-8s %pF\n",
pt->dev ? pt->dev->name : "", pt->func);
}
return 0;
}
static const struct seq_operations ptype_seq_ops = {
.start = ptype_seq_start,
.next = ptype_seq_next,
.stop = ptype_seq_stop,
.show = ptype_seq_show,
};
static int ptype_seq_open(struct inode *inode, struct file *file)
{
return seq_open_net(inode, file, &ptype_seq_ops,
sizeof(struct seq_net_private));
}
static const struct file_operations ptype_seq_fops = {
.owner = THIS_MODULE,
.open = ptype_seq_open,
.read = seq_read,
.llseek = seq_lseek,
.release = seq_release_net,
};
static int __net_init dev_proc_net_init(struct net *net)
{
int rc = -ENOMEM;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (!proc_net_fops_create(net, "dev", S_IRUGO, &dev_seq_fops))
goto out;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (!proc_net_fops_create(net, "softnet_stat", S_IRUGO, &softnet_seq_fops))
goto out_dev;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (!proc_net_fops_create(net, "ptype", S_IRUGO, &ptype_seq_fops))
goto out_softnet;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (wext_proc_init(net))
goto out_ptype;
rc = 0;
out:
return rc;
out_ptype:
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
proc_net_remove(net, "ptype");
out_softnet:
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
proc_net_remove(net, "softnet_stat");
out_dev:
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
proc_net_remove(net, "dev");
goto out;
}
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
static void __net_exit dev_proc_net_exit(struct net *net)
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
{
wext_proc_exit(net);
proc_net_remove(net, "ptype");
proc_net_remove(net, "softnet_stat");
proc_net_remove(net, "dev");
}
static struct pernet_operations __net_initdata dev_proc_ops = {
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
.init = dev_proc_net_init,
.exit = dev_proc_net_exit,
};
static int __init dev_proc_init(void)
{
return register_pernet_subsys(&dev_proc_ops);
}
#else
#define dev_proc_init() 0
#endif /* CONFIG_PROC_FS */
/**
* netdev_set_master - set up master pointer
* @slave: slave device
* @master: new master device
*
* Changes the master device of the slave. Pass %NULL to break the
* bonding. The caller must hold the RTNL semaphore. On a failure
* a negative errno code is returned. On success the reference counts
* are adjusted and the function returns zero.
*/
int netdev_set_master(struct net_device *slave, struct net_device *master)
{
struct net_device *old = slave->master;
ASSERT_RTNL();
if (master) {
if (old)
return -EBUSY;
dev_hold(master);
}
slave->master = master;
if (old)
dev_put(old);
return 0;
}
EXPORT_SYMBOL(netdev_set_master);
/**
* netdev_set_bond_master - set up bonding master/slave pair
* @slave: slave device
* @master: new master device
*
* Changes the master device of the slave. Pass %NULL to break the
* bonding. The caller must hold the RTNL semaphore. On a failure
* a negative errno code is returned. On success %RTM_NEWLINK is sent
* to the routing socket and the function returns zero.
*/
int netdev_set_bond_master(struct net_device *slave, struct net_device *master)
{
int err;
ASSERT_RTNL();
err = netdev_set_master(slave, master);
if (err)
return err;
if (master)
slave->flags |= IFF_SLAVE;
else
slave->flags &= ~IFF_SLAVE;
rtmsg_ifinfo(RTM_NEWLINK, slave, IFF_SLAVE);
return 0;
}
EXPORT_SYMBOL(netdev_set_bond_master);
static void dev_change_rx_flags(struct net_device *dev, int flags)
{
const struct net_device_ops *ops = dev->netdev_ops;
if ((dev->flags & IFF_UP) && ops->ndo_change_rx_flags)
ops->ndo_change_rx_flags(dev, flags);
}
static int __dev_set_promiscuity(struct net_device *dev, int inc)
{
unsigned short old_flags = dev->flags;
uid_t uid;
gid_t gid;
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-15 01:51:31 +00:00
ASSERT_RTNL();
dev->flags |= IFF_PROMISC;
dev->promiscuity += inc;
if (dev->promiscuity == 0) {
/*
* Avoid overflow.
* If inc causes overflow, untouch promisc and return error.
*/
if (inc < 0)
dev->flags &= ~IFF_PROMISC;
else {
dev->promiscuity -= inc;
printk(KERN_WARNING "%s: promiscuity touches roof, "
"set promiscuity failed, promiscuity feature "
"of device might be broken.\n", dev->name);
return -EOVERFLOW;
}
}
if (dev->flags != old_flags) {
printk(KERN_INFO "device %s %s promiscuous mode\n",
dev->name, (dev->flags & IFF_PROMISC) ? "entered" :
"left");
if (audit_enabled) {
current_uid_gid(&uid, &gid);
audit_log(current->audit_context, GFP_ATOMIC,
AUDIT_ANOM_PROMISCUOUS,
"dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
dev->name, (dev->flags & IFF_PROMISC),
(old_flags & IFF_PROMISC),
audit_get_loginuid(current),
uid, gid,
audit_get_sessionid(current));
}
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-15 01:51:31 +00:00
dev_change_rx_flags(dev, IFF_PROMISC);
}
return 0;
}
/**
* dev_set_promiscuity - update promiscuity count on a device
* @dev: device
* @inc: modifier
*
* Add or remove promiscuity from a device. While the count in the device
* remains above zero the interface remains promiscuous. Once it hits zero
* the device reverts back to normal filtering operation. A negative inc
* value is used to drop promiscuity on the device.
* Return 0 if successful or a negative errno code on error.
*/
int dev_set_promiscuity(struct net_device *dev, int inc)
{
unsigned short old_flags = dev->flags;
int err;
err = __dev_set_promiscuity(dev, inc);
if (err < 0)
return err;
if (dev->flags != old_flags)
dev_set_rx_mode(dev);
return err;
}
EXPORT_SYMBOL(dev_set_promiscuity);
/**
* dev_set_allmulti - update allmulti count on a device
* @dev: device
* @inc: modifier
*
* Add or remove reception of all multicast frames to a device. While the
* count in the device remains above zero the interface remains listening
* to all interfaces. Once it hits zero the device reverts back to normal
* filtering operation. A negative @inc value is used to drop the counter
* when releasing a resource needing all multicasts.
* Return 0 if successful or a negative errno code on error.
*/
int dev_set_allmulti(struct net_device *dev, int inc)
{
unsigned short old_flags = dev->flags;
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-15 01:51:31 +00:00
ASSERT_RTNL();
dev->flags |= IFF_ALLMULTI;
dev->allmulti += inc;
if (dev->allmulti == 0) {
/*
* Avoid overflow.
* If inc causes overflow, untouch allmulti and return error.
*/
if (inc < 0)
dev->flags &= ~IFF_ALLMULTI;
else {
dev->allmulti -= inc;
printk(KERN_WARNING "%s: allmulti touches roof, "
"set allmulti failed, allmulti feature of "
"device might be broken.\n", dev->name);
return -EOVERFLOW;
}
}
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-15 01:51:31 +00:00
if (dev->flags ^ old_flags) {
dev_change_rx_flags(dev, IFF_ALLMULTI);
dev_set_rx_mode(dev);
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-15 01:51:31 +00:00
}
return 0;
}
EXPORT_SYMBOL(dev_set_allmulti);
/*
* Upload unicast and multicast address lists to device and
* configure RX filtering. When the device doesn't support unicast
* filtering it is put in promiscuous mode while unicast addresses
* are present.
*/
void __dev_set_rx_mode(struct net_device *dev)
{
const struct net_device_ops *ops = dev->netdev_ops;
/* dev_open will call this function so the list will stay sane. */
if (!(dev->flags&IFF_UP))
return;
if (!netif_device_present(dev))
return;
if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
/* Unicast addresses changes may only happen under the rtnl,
* therefore calling __dev_set_promiscuity here is safe.
*/
if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
__dev_set_promiscuity(dev, 1);
dev->uc_promisc = true;
} else if (netdev_uc_empty(dev) && dev->uc_promisc) {
__dev_set_promiscuity(dev, -1);
dev->uc_promisc = false;
}
}
if (ops->ndo_set_rx_mode)
ops->ndo_set_rx_mode(dev);
}
void dev_set_rx_mode(struct net_device *dev)
{
netif_addr_lock_bh(dev);
__dev_set_rx_mode(dev);
netif_addr_unlock_bh(dev);
}
/**
* dev_get_flags - get flags reported to userspace
* @dev: device
*
* Get the combination of flag bits exported through APIs to userspace.
*/
unsigned dev_get_flags(const struct net_device *dev)
{
unsigned flags;
flags = (dev->flags & ~(IFF_PROMISC |
IFF_ALLMULTI |
IFF_RUNNING |
IFF_LOWER_UP |
IFF_DORMANT)) |
(dev->gflags & (IFF_PROMISC |
IFF_ALLMULTI));
if (netif_running(dev)) {
if (netif_oper_up(dev))
flags |= IFF_RUNNING;
if (netif_carrier_ok(dev))
flags |= IFF_LOWER_UP;
if (netif_dormant(dev))
flags |= IFF_DORMANT;
}
return flags;
}
EXPORT_SYMBOL(dev_get_flags);
int __dev_change_flags(struct net_device *dev, unsigned int flags)
{
int old_flags = dev->flags;
int ret;
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-15 01:51:31 +00:00
ASSERT_RTNL();
/*
* Set the flags on our device.
*/
dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
IFF_AUTOMEDIA)) |
(dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
IFF_ALLMULTI));
/*
* Load in the correct multicast list now the flags have changed.
*/
if ((old_flags ^ flags) & IFF_MULTICAST)
dev_change_rx_flags(dev, IFF_MULTICAST);
[NET]: Add net_device change_rx_mode callback Currently the set_multicast_list (and set_rx_mode) callbacks are responsible for configuring the device according to the IFF_PROMISC, IFF_MULTICAST and IFF_ALLMULTI flags and the mc_list (and uc_list in case of set_rx_mode). These callbacks can be invoked from BH context without the rtnl_mutex by dev_mc_add/dev_mc_delete, which makes reading the device flags and promiscous/allmulti count racy. For real hardware drivers that just commit all changes to the hardware this is not a real problem since the stack guarantees to call them for every change, so at least the final call will not race and commit the correct configuration to the hardware. For software devices that want to synchronize promiscous and multicast state to an underlying device however this can cause corruption of the underlying device's flags or promisc/allmulti counts. When the software device is concurrently put in promiscous or allmulti mode while set_multicast_list is invoked from bottem half context, the device might synchronize the change to the underlying device without holding the rtnl_mutex, which races with concurrent changes to the underlying device. Add a dev->change_rx_flags hook that is invoked when any of the flags that affect rx filtering change (under the rtnl_mutex), which allows drivers to perform synchronization immediately and only synchronize the address lists in set_multicast_list/set_rx_mode. Signed-off-by: Patrick McHardy <kaber@trash.net> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-07-15 01:51:31 +00:00
dev_set_rx_mode(dev);
/*
* Have we downed the interface. We handle IFF_UP ourselves
* according to user attempts to set it, rather than blindly
* setting it.
*/
ret = 0;
if ((old_flags ^ flags) & IFF_UP) { /* Bit is different ? */
ret = ((old_flags & IFF_UP) ? __dev_close : __dev_open)(dev);
if (!ret)
dev_set_rx_mode(dev);
}
if ((flags ^ dev->gflags) & IFF_PROMISC) {
int inc = (flags & IFF_PROMISC) ? 1 : -1;
dev->gflags ^= IFF_PROMISC;
dev_set_promiscuity(dev, inc);
}
/* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
is important. Some (broken) drivers set IFF_PROMISC, when
IFF_ALLMULTI is requested not asking us and not reporting.
*/
if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
dev->gflags ^= IFF_ALLMULTI;
dev_set_allmulti(dev, inc);
}
return ret;
}
void __dev_notify_flags(struct net_device *dev, unsigned int old_flags)
{
unsigned int changes = dev->flags ^ old_flags;
if (changes & IFF_UP) {
if (dev->flags & IFF_UP)
call_netdevice_notifiers(NETDEV_UP, dev);
else
call_netdevice_notifiers(NETDEV_DOWN, dev);
}
if (dev->flags & IFF_UP &&
(changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE)))
call_netdevice_notifiers(NETDEV_CHANGE, dev);
}
/**
* dev_change_flags - change device settings
* @dev: device
* @flags: device state flags
*
* Change settings on device based state flags. The flags are
* in the userspace exported format.
*/
int dev_change_flags(struct net_device *dev, unsigned flags)
{
int ret, changes;
int old_flags = dev->flags;
ret = __dev_change_flags(dev, flags);
if (ret < 0)
return ret;
changes = old_flags ^ dev->flags;
if (changes)
rtmsg_ifinfo(RTM_NEWLINK, dev, changes);
__dev_notify_flags(dev, old_flags);
return ret;
}
EXPORT_SYMBOL(dev_change_flags);
/**
* dev_set_mtu - Change maximum transfer unit
* @dev: device
* @new_mtu: new transfer unit
*
* Change the maximum transfer size of the network device.
*/
int dev_set_mtu(struct net_device *dev, int new_mtu)
{
const struct net_device_ops *ops = dev->netdev_ops;
int err;
if (new_mtu == dev->mtu)
return 0;
/* MTU must be positive. */
if (new_mtu < 0)
return -EINVAL;
if (!netif_device_present(dev))
return -ENODEV;
err = 0;
if (ops->ndo_change_mtu)
err = ops->ndo_change_mtu(dev, new_mtu);
else
dev->mtu = new_mtu;
if (!err && dev->flags & IFF_UP)
call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
return err;
}
EXPORT_SYMBOL(dev_set_mtu);
/**
* dev_set_group - Change group this device belongs to
* @dev: device
* @new_group: group this device should belong to
*/
void dev_set_group(struct net_device *dev, int new_group)
{
dev->group = new_group;
}
EXPORT_SYMBOL(dev_set_group);
/**
* dev_set_mac_address - Change Media Access Control Address
* @dev: device
* @sa: new address
*
* Change the hardware (MAC) address of the device
*/
int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
{
const struct net_device_ops *ops = dev->netdev_ops;
int err;
if (!ops->ndo_set_mac_address)
return -EOPNOTSUPP;
if (sa->sa_family != dev->type)
return -EINVAL;
if (!netif_device_present(dev))
return -ENODEV;
err = ops->ndo_set_mac_address(dev, sa);
if (!err)
call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
return err;
}
EXPORT_SYMBOL(dev_set_mac_address);
/*
* Perform the SIOCxIFxxx calls, inside rcu_read_lock()
*/
static int dev_ifsioc_locked(struct net *net, struct ifreq *ifr, unsigned int cmd)
{
int err;
struct net_device *dev = dev_get_by_name_rcu(net, ifr->ifr_name);
if (!dev)
return -ENODEV;
switch (cmd) {
case SIOCGIFFLAGS: /* Get interface flags */
ifr->ifr_flags = (short) dev_get_flags(dev);
return 0;
case SIOCGIFMETRIC: /* Get the metric on the interface
(currently unused) */
ifr->ifr_metric = 0;
return 0;
case SIOCGIFMTU: /* Get the MTU of a device */
ifr->ifr_mtu = dev->mtu;
return 0;
case SIOCGIFHWADDR:
if (!dev->addr_len)
memset(ifr->ifr_hwaddr.sa_data, 0, sizeof ifr->ifr_hwaddr.sa_data);
else
memcpy(ifr->ifr_hwaddr.sa_data, dev->dev_addr,
min(sizeof ifr->ifr_hwaddr.sa_data, (size_t) dev->addr_len));
ifr->ifr_hwaddr.sa_family = dev->type;
return 0;
case SIOCGIFSLAVE:
err = -EINVAL;
break;
case SIOCGIFMAP:
ifr->ifr_map.mem_start = dev->mem_start;
ifr->ifr_map.mem_end = dev->mem_end;
ifr->ifr_map.base_addr = dev->base_addr;
ifr->ifr_map.irq = dev->irq;
ifr->ifr_map.dma = dev->dma;
ifr->ifr_map.port = dev->if_port;
return 0;
case SIOCGIFINDEX:
ifr->ifr_ifindex = dev->ifindex;
return 0;
case SIOCGIFTXQLEN:
ifr->ifr_qlen = dev->tx_queue_len;
return 0;
default:
/* dev_ioctl() should ensure this case
* is never reached
*/
WARN_ON(1);
err = -ENOTTY;
break;
}
return err;
}
/*
* Perform the SIOCxIFxxx calls, inside rtnl_lock()
*/
static int dev_ifsioc(struct net *net, struct ifreq *ifr, unsigned int cmd)
{
int err;
struct net_device *dev = __dev_get_by_name(net, ifr->ifr_name);
const struct net_device_ops *ops;
if (!dev)
return -ENODEV;
ops = dev->netdev_ops;
switch (cmd) {
case SIOCSIFFLAGS: /* Set interface flags */
return dev_change_flags(dev, ifr->ifr_flags);
case SIOCSIFMETRIC: /* Set the metric on the interface
(currently unused) */
return -EOPNOTSUPP;
case SIOCSIFMTU: /* Set the MTU of a device */
return dev_set_mtu(dev, ifr->ifr_mtu);
case SIOCSIFHWADDR:
return dev_set_mac_address(dev, &ifr->ifr_hwaddr);
case SIOCSIFHWBROADCAST:
if (ifr->ifr_hwaddr.sa_family != dev->type)
return -EINVAL;
memcpy(dev->broadcast, ifr->ifr_hwaddr.sa_data,
min(sizeof ifr->ifr_hwaddr.sa_data, (size_t) dev->addr_len));
call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
return 0;
case SIOCSIFMAP:
if (ops->ndo_set_config) {
if (!netif_device_present(dev))
return -ENODEV;
return ops->ndo_set_config(dev, &ifr->ifr_map);
}
return -EOPNOTSUPP;
case SIOCADDMULTI:
if (!ops->ndo_set_rx_mode ||
ifr->ifr_hwaddr.sa_family != AF_UNSPEC)
return -EINVAL;
if (!netif_device_present(dev))
return -ENODEV;
return dev_mc_add_global(dev, ifr->ifr_hwaddr.sa_data);
case SIOCDELMULTI:
if (!ops->ndo_set_rx_mode ||
ifr->ifr_hwaddr.sa_family != AF_UNSPEC)
return -EINVAL;
if (!netif_device_present(dev))
return -ENODEV;
return dev_mc_del_global(dev, ifr->ifr_hwaddr.sa_data);
case SIOCSIFTXQLEN:
if (ifr->ifr_qlen < 0)
return -EINVAL;
dev->tx_queue_len = ifr->ifr_qlen;
return 0;
case SIOCSIFNAME:
ifr->ifr_newname[IFNAMSIZ-1] = '\0';
return dev_change_name(dev, ifr->ifr_newname);
case SIOCSHWTSTAMP:
err = net_hwtstamp_validate(ifr);
if (err)
return err;
/* fall through */
/*
* Unknown or private ioctl
*/
default:
if ((cmd >= SIOCDEVPRIVATE &&
cmd <= SIOCDEVPRIVATE + 15) ||
cmd == SIOCBONDENSLAVE ||
cmd == SIOCBONDRELEASE ||
cmd == SIOCBONDSETHWADDR ||
cmd == SIOCBONDSLAVEINFOQUERY ||
cmd == SIOCBONDINFOQUERY ||
cmd == SIOCBONDCHANGEACTIVE ||
cmd == SIOCGMIIPHY ||
cmd == SIOCGMIIREG ||
cmd == SIOCSMIIREG ||
cmd == SIOCBRADDIF ||
cmd == SIOCBRDELIF ||
cmd == SIOCSHWTSTAMP ||
cmd == SIOCWANDEV) {
err = -EOPNOTSUPP;
if (ops->ndo_do_ioctl) {
if (netif_device_present(dev))
err = ops->ndo_do_ioctl(dev, ifr, cmd);
else
err = -ENODEV;
}
} else
err = -EINVAL;
}
return err;
}
/*
* This function handles all "interface"-type I/O control requests. The actual
* 'doing' part of this is dev_ifsioc above.
*/
/**
* dev_ioctl - network device ioctl
* @net: the applicable net namespace
* @cmd: command to issue
* @arg: pointer to a struct ifreq in user space
*
* Issue ioctl functions to devices. This is normally called by the
* user space syscall interfaces but can sometimes be useful for
* other purposes. The return value is the return from the syscall if
* positive or a negative errno code on error.
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
int dev_ioctl(struct net *net, unsigned int cmd, void __user *arg)
{
struct ifreq ifr;
int ret;
char *colon;
/* One special case: SIOCGIFCONF takes ifconf argument
and requires shared lock, because it sleeps writing
to user space.
*/
if (cmd == SIOCGIFCONF) {
rtnl_lock();
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
ret = dev_ifconf(net, (char __user *) arg);
rtnl_unlock();
return ret;
}
if (cmd == SIOCGIFNAME)
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
return dev_ifname(net, (struct ifreq __user *)arg);
if (copy_from_user(&ifr, arg, sizeof(struct ifreq)))
return -EFAULT;
ifr.ifr_name[IFNAMSIZ-1] = 0;
colon = strchr(ifr.ifr_name, ':');
if (colon)
*colon = 0;
/*
* See which interface the caller is talking about.
*/
switch (cmd) {
/*
* These ioctl calls:
* - can be done by all.
* - atomic and do not require locking.
* - return a value
*/
case SIOCGIFFLAGS:
case SIOCGIFMETRIC:
case SIOCGIFMTU:
case SIOCGIFHWADDR:
case SIOCGIFSLAVE:
case SIOCGIFMAP:
case SIOCGIFINDEX:
case SIOCGIFTXQLEN:
dev_load(net, ifr.ifr_name);
rcu_read_lock();
ret = dev_ifsioc_locked(net, &ifr, cmd);
rcu_read_unlock();
if (!ret) {
if (colon)
*colon = ':';
if (copy_to_user(arg, &ifr,
sizeof(struct ifreq)))
ret = -EFAULT;
}
return ret;
case SIOCETHTOOL:
dev_load(net, ifr.ifr_name);
rtnl_lock();
ret = dev_ethtool(net, &ifr);
rtnl_unlock();
if (!ret) {
if (colon)
*colon = ':';
if (copy_to_user(arg, &ifr,
sizeof(struct ifreq)))
ret = -EFAULT;
}
return ret;
/*
* These ioctl calls:
* - require superuser power.
* - require strict serialization.
* - return a value
*/
case SIOCGMIIPHY:
case SIOCGMIIREG:
case SIOCSIFNAME:
if (!capable(CAP_NET_ADMIN))
return -EPERM;
dev_load(net, ifr.ifr_name);
rtnl_lock();
ret = dev_ifsioc(net, &ifr, cmd);
rtnl_unlock();
if (!ret) {
if (colon)
*colon = ':';
if (copy_to_user(arg, &ifr,
sizeof(struct ifreq)))
ret = -EFAULT;
}
return ret;
/*
* These ioctl calls:
* - require superuser power.
* - require strict serialization.
* - do not return a value
*/
case SIOCSIFFLAGS:
case SIOCSIFMETRIC:
case SIOCSIFMTU:
case SIOCSIFMAP:
case SIOCSIFHWADDR:
case SIOCSIFSLAVE:
case SIOCADDMULTI:
case SIOCDELMULTI:
case SIOCSIFHWBROADCAST:
case SIOCSIFTXQLEN:
case SIOCSMIIREG:
case SIOCBONDENSLAVE:
case SIOCBONDRELEASE:
case SIOCBONDSETHWADDR:
case SIOCBONDCHANGEACTIVE:
case SIOCBRADDIF:
case SIOCBRDELIF:
case SIOCSHWTSTAMP:
if (!capable(CAP_NET_ADMIN))
return -EPERM;
/* fall through */
case SIOCBONDSLAVEINFOQUERY:
case SIOCBONDINFOQUERY:
dev_load(net, ifr.ifr_name);
rtnl_lock();
ret = dev_ifsioc(net, &ifr, cmd);
rtnl_unlock();
return ret;
case SIOCGIFMEM:
/* Get the per device memory space. We can add this but
* currently do not support it */
case SIOCSIFMEM:
/* Set the per device memory buffer space.
* Not applicable in our case */
case SIOCSIFLINK:
return -ENOTTY;
/*
* Unknown or private ioctl.
*/
default:
if (cmd == SIOCWANDEV ||
(cmd >= SIOCDEVPRIVATE &&
cmd <= SIOCDEVPRIVATE + 15)) {
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
dev_load(net, ifr.ifr_name);
rtnl_lock();
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
ret = dev_ifsioc(net, &ifr, cmd);
rtnl_unlock();
if (!ret && copy_to_user(arg, &ifr,
sizeof(struct ifreq)))
ret = -EFAULT;
return ret;
}
/* Take care of Wireless Extensions */
if (cmd >= SIOCIWFIRST && cmd <= SIOCIWLAST)
return wext_handle_ioctl(net, &ifr, cmd, arg);
return -ENOTTY;
}
}
/**
* dev_new_index - allocate an ifindex
* @net: the applicable net namespace
*
* Returns a suitable unique value for a new device interface
* number. The caller must hold the rtnl semaphore or the
* dev_base_lock to be sure it remains unique.
*/
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
static int dev_new_index(struct net *net)
{
static int ifindex;
for (;;) {
if (++ifindex <= 0)
ifindex = 1;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (!__dev_get_by_index(net, ifindex))
return ifindex;
}
}
/* Delayed registration/unregisteration */
static LIST_HEAD(net_todo_list);
static void net_set_todo(struct net_device *dev)
{
list_add_tail(&dev->todo_list, &net_todo_list);
}
static void rollback_registered_many(struct list_head *head)
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
{
struct net_device *dev, *tmp;
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
BUG_ON(dev_boot_phase);
ASSERT_RTNL();
list_for_each_entry_safe(dev, tmp, head, unreg_list) {
/* Some devices call without registering
* for initialization unwind. Remove those
* devices and proceed with the remaining.
*/
if (dev->reg_state == NETREG_UNINITIALIZED) {
pr_debug("unregister_netdevice: device %s/%p never "
"was registered\n", dev->name, dev);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
WARN_ON(1);
list_del(&dev->unreg_list);
continue;
}
dev->dismantle = true;
BUG_ON(dev->reg_state != NETREG_REGISTERED);
}
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
/* If device is running, close it first. */
dev_close_many(head);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
list_for_each_entry(dev, head, unreg_list) {
/* And unlink it from device chain. */
unlist_netdevice(dev);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
dev->reg_state = NETREG_UNREGISTERING;
}
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
synchronize_net();
list_for_each_entry(dev, head, unreg_list) {
/* Shutdown queueing discipline. */
dev_shutdown(dev);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
/* Notify protocols, that we are about to destroy
this device. They should clean all the things.
*/
call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
if (!dev->rtnl_link_ops ||
dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
rtmsg_ifinfo(RTM_DELLINK, dev, ~0U);
/*
* Flush the unicast and multicast chains
*/
dev_uc_flush(dev);
dev_mc_flush(dev);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
if (dev->netdev_ops->ndo_uninit)
dev->netdev_ops->ndo_uninit(dev);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
/* Notifier chain MUST detach us from master device. */
WARN_ON(dev->master);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
/* Remove entries from kobject tree */
netdev_unregister_kobject(dev);
}
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
/* Process any work delayed until the end of the batch */
dev = list_first_entry(head, struct net_device, unreg_list);
call_netdevice_notifiers(NETDEV_UNREGISTER_BATCH, dev);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
synchronize_net();
list_for_each_entry(dev, head, unreg_list)
dev_put(dev);
}
static void rollback_registered(struct net_device *dev)
{
LIST_HEAD(single);
list_add(&dev->unreg_list, &single);
rollback_registered_many(&single);
list_del(&single);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
}
static u32 netdev_fix_features(struct net_device *dev, u32 features)
{
/* Fix illegal checksum combinations */
if ((features & NETIF_F_HW_CSUM) &&
(features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
netdev_warn(dev, "mixed HW and IP checksum settings.\n");
features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
}
if ((features & NETIF_F_NO_CSUM) &&
(features & (NETIF_F_HW_CSUM|NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
netdev_warn(dev, "mixed no checksumming and other settings.\n");
features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM|NETIF_F_HW_CSUM);
}
/* Fix illegal SG+CSUM combinations. */
if ((features & NETIF_F_SG) &&
!(features & NETIF_F_ALL_CSUM)) {
netdev_dbg(dev,
"Dropping NETIF_F_SG since no checksum feature.\n");
features &= ~NETIF_F_SG;
}
/* TSO requires that SG is present as well. */
if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
features &= ~NETIF_F_ALL_TSO;
}
/* TSO ECN requires that TSO is present as well. */
if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
features &= ~NETIF_F_TSO_ECN;
/* Software GSO depends on SG. */
if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
features &= ~NETIF_F_GSO;
}
/* UFO needs SG and checksumming */
if (features & NETIF_F_UFO) {
/* maybe split UFO into V4 and V6? */
if (!((features & NETIF_F_GEN_CSUM) ||
(features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))
== (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
netdev_dbg(dev,
"Dropping NETIF_F_UFO since no checksum offload features.\n");
features &= ~NETIF_F_UFO;
}
if (!(features & NETIF_F_SG)) {
netdev_dbg(dev,
"Dropping NETIF_F_UFO since no NETIF_F_SG feature.\n");
features &= ~NETIF_F_UFO;
}
}
return features;
}
int __netdev_update_features(struct net_device *dev)
{
u32 features;
int err = 0;
ASSERT_RTNL();
features = netdev_get_wanted_features(dev);
if (dev->netdev_ops->ndo_fix_features)
features = dev->netdev_ops->ndo_fix_features(dev, features);
/* driver might be less strict about feature dependencies */
features = netdev_fix_features(dev, features);
if (dev->features == features)
return 0;
netdev_dbg(dev, "Features changed: 0x%08x -> 0x%08x\n",
dev->features, features);
if (dev->netdev_ops->ndo_set_features)
err = dev->netdev_ops->ndo_set_features(dev, features);
if (unlikely(err < 0)) {
netdev_err(dev,
"set_features() failed (%d); wanted 0x%08x, left 0x%08x\n",
err, features, dev->features);
return -1;
}
if (!err)
dev->features = features;
return 1;
}
/**
* netdev_update_features - recalculate device features
* @dev: the device to check
*
* Recalculate dev->features set and send notifications if it
* has changed. Should be called after driver or hardware dependent
* conditions might have changed that influence the features.
*/
void netdev_update_features(struct net_device *dev)
{
if (__netdev_update_features(dev))
netdev_features_change(dev);
}
EXPORT_SYMBOL(netdev_update_features);
/**
* netdev_change_features - recalculate device features
* @dev: the device to check
*
* Recalculate dev->features set and send notifications even
* if they have not changed. Should be called instead of
* netdev_update_features() if also dev->vlan_features might
* have changed to allow the changes to be propagated to stacked
* VLAN devices.
*/
void netdev_change_features(struct net_device *dev)
{
__netdev_update_features(dev);
netdev_features_change(dev);
}
EXPORT_SYMBOL(netdev_change_features);
/**
* netif_stacked_transfer_operstate - transfer operstate
* @rootdev: the root or lower level device to transfer state from
* @dev: the device to transfer operstate to
*
* Transfer operational state from root to device. This is normally
* called when a stacking relationship exists between the root
* device and the device(a leaf device).
*/
void netif_stacked_transfer_operstate(const struct net_device *rootdev,
struct net_device *dev)
{
if (rootdev->operstate == IF_OPER_DORMANT)
netif_dormant_on(dev);
else
netif_dormant_off(dev);
if (netif_carrier_ok(rootdev)) {
if (!netif_carrier_ok(dev))
netif_carrier_on(dev);
} else {
if (netif_carrier_ok(dev))
netif_carrier_off(dev);
}
}
EXPORT_SYMBOL(netif_stacked_transfer_operstate);
#ifdef CONFIG_RPS
static int netif_alloc_rx_queues(struct net_device *dev)
{
unsigned int i, count = dev->num_rx_queues;
struct netdev_rx_queue *rx;
BUG_ON(count < 1);
rx = kcalloc(count, sizeof(struct netdev_rx_queue), GFP_KERNEL);
if (!rx) {
pr_err("netdev: Unable to allocate %u rx queues.\n", count);
return -ENOMEM;
}
dev->_rx = rx;
for (i = 0; i < count; i++)
rx[i].dev = dev;
return 0;
}
#endif
static void netdev_init_one_queue(struct net_device *dev,
struct netdev_queue *queue, void *_unused)
{
/* Initialize queue lock */
spin_lock_init(&queue->_xmit_lock);
netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
queue->xmit_lock_owner = -1;
netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
queue->dev = dev;
}
static int netif_alloc_netdev_queues(struct net_device *dev)
{
unsigned int count = dev->num_tx_queues;
struct netdev_queue *tx;
BUG_ON(count < 1);
tx = kcalloc(count, sizeof(struct netdev_queue), GFP_KERNEL);
if (!tx) {
pr_err("netdev: Unable to allocate %u tx queues.\n",
count);
return -ENOMEM;
}
dev->_tx = tx;
xps: Transmit Packet Steering This patch implements transmit packet steering (XPS) for multiqueue devices. XPS selects a transmit queue during packet transmission based on configuration. This is done by mapping the CPU transmitting the packet to a queue. This is the transmit side analogue to RPS-- where RPS is selecting a CPU based on receive queue, XPS selects a queue based on the CPU (previously there was an XPS patch from Eric Dumazet, but that might more appropriately be called transmit completion steering). Each transmit queue can be associated with a number of CPUs which will use the queue to send packets. This is configured as a CPU mask on a per queue basis in: /sys/class/net/eth<n>/queues/tx-<n>/xps_cpus The mappings are stored per device in an inverted data structure that maps CPUs to queues. In the netdevice structure this is an array of num_possible_cpu structures where each structure holds and array of queue_indexes for queues which that CPU can use. The benefits of XPS are improved locality in the per queue data structures. Also, transmit completions are more likely to be done nearer to the sending thread, so this should promote locality back to the socket on free (e.g. UDP). The benefits of XPS are dependent on cache hierarchy, application load, and other factors. XPS would nominally be configured so that a queue would only be shared by CPUs which are sharing a cache, the degenerative configuration woud be that each CPU has it's own queue. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. bnx2x on 16 core AMD XPS (16 queues, 1 TX queue per CPU) 1234K at 100% CPU No XPS (16 queues) 996K at 100% CPU Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-11-21 13:17:27 +00:00
netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
spin_lock_init(&dev->tx_global_lock);
return 0;
}
/**
* register_netdevice - register a network device
* @dev: device to register
*
* Take a completed network device structure and add it to the kernel
* interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
* chain. 0 is returned on success. A negative errno code is returned
* on a failure to set up the device, or if the name is a duplicate.
*
* Callers must hold the rtnl semaphore. You may want
* register_netdev() instead of this.
*
* BUGS:
* The locking appears insufficient to guarantee two parallel registers
* will not get the same name.
*/
int register_netdevice(struct net_device *dev)
{
int ret;
struct net *net = dev_net(dev);
BUG_ON(dev_boot_phase);
ASSERT_RTNL();
might_sleep();
/* When net_device's are persistent, this will be fatal. */
BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
BUG_ON(!net);
spin_lock_init(&dev->addr_list_lock);
netdev_set_addr_lockdep_class(dev);
dev->iflink = -1;
net:set valid name before calling ndo_init() In commit 1c5cae815d19 (net: call dev_alloc_name from register_netdevice), a bug of bonding was involved, see example 1 and 2. In register_netdevice(), the name of net_device is not valid until dev_get_valid_name() is called. But dev->netdev_ops->ndo_init(that is bond_init) is called before dev_get_valid_name(), and it uses the invalid name of net_device. I think register_netdevice() should make sure that the name of net_device is valid before calling ndo_init(). example 1: modprobe bonding ls /proc/net/bonding/bond%d ps -eLf root 3398 2 3398 0 1 21:34 ? 00:00:00 [bond%d] example 2: modprobe bonding max_bonds=3 [ 170.100292] bonding: Ethernet Channel Bonding Driver: v3.7.1 (April 27, 2011) [ 170.101090] bonding: Warning: either miimon or arp_interval and arp_ip_target module parameters must be specified, otherwise bonding will not detect link failures! see bonding.txt for details. [ 170.102469] ------------[ cut here ]------------ [ 170.103150] WARNING: at /home/pwp/net-next-2.6/fs/proc/generic.c:586 proc_register+0x126/0x157() [ 170.104075] Hardware name: VirtualBox [ 170.105065] proc_dir_entry 'bonding/bond%d' already registered [ 170.105613] Modules linked in: bonding(+) sunrpc ipv6 uinput microcode ppdev parport_pc parport joydev e1000 pcspkr i2c_piix4 i2c_core [last unloaded: bonding] [ 170.108397] Pid: 3457, comm: modprobe Not tainted 2.6.39-rc2+ #14 [ 170.108935] Call Trace: [ 170.109382] [<c0438f3b>] warn_slowpath_common+0x6a/0x7f [ 170.109911] [<c051a42a>] ? proc_register+0x126/0x157 [ 170.110329] [<c0438fc3>] warn_slowpath_fmt+0x2b/0x2f [ 170.110846] [<c051a42a>] proc_register+0x126/0x157 [ 170.111870] [<c051a4dd>] proc_create_data+0x82/0x98 [ 170.112335] [<f94e6af6>] bond_create_proc_entry+0x3f/0x73 [bonding] [ 170.112905] [<f94dd806>] bond_init+0x77/0xa5 [bonding] [ 170.113319] [<c0721ac6>] register_netdevice+0x8c/0x1d3 [ 170.113848] [<f94e0e30>] bond_create+0x6c/0x90 [bonding] [ 170.114322] [<f94f4763>] bonding_init+0x763/0x7b1 [bonding] [ 170.114879] [<c0401240>] do_one_initcall+0x76/0x122 [ 170.115317] [<f94f4000>] ? 0xf94f3fff [ 170.115799] [<c0463f1e>] sys_init_module+0x1286/0x140d [ 170.116879] [<c07c6d9f>] sysenter_do_call+0x12/0x28 [ 170.117404] ---[ end trace 64e4fac3ae5fff1a ]--- [ 170.117924] bond%d: Warning: failed to register to debugfs [ 170.128728] ------------[ cut here ]------------ [ 170.129360] WARNING: at /home/pwp/net-next-2.6/fs/proc/generic.c:586 proc_register+0x126/0x157() [ 170.130323] Hardware name: VirtualBox [ 170.130797] proc_dir_entry 'bonding/bond%d' already registered [ 170.131315] Modules linked in: bonding(+) sunrpc ipv6 uinput microcode ppdev parport_pc parport joydev e1000 pcspkr i2c_piix4 i2c_core [last unloaded: bonding] [ 170.133731] Pid: 3457, comm: modprobe Tainted: G W 2.6.39-rc2+ #14 [ 170.134308] Call Trace: [ 170.134743] [<c0438f3b>] warn_slowpath_common+0x6a/0x7f [ 170.135305] [<c051a42a>] ? proc_register+0x126/0x157 [ 170.135820] [<c0438fc3>] warn_slowpath_fmt+0x2b/0x2f [ 170.137168] [<c051a42a>] proc_register+0x126/0x157 [ 170.137700] [<c051a4dd>] proc_create_data+0x82/0x98 [ 170.138174] [<f94e6af6>] bond_create_proc_entry+0x3f/0x73 [bonding] [ 170.138745] [<f94dd806>] bond_init+0x77/0xa5 [bonding] [ 170.139278] [<c0721ac6>] register_netdevice+0x8c/0x1d3 [ 170.139828] [<f94e0e30>] bond_create+0x6c/0x90 [bonding] [ 170.140361] [<f94f4763>] bonding_init+0x763/0x7b1 [bonding] [ 170.140927] [<c0401240>] do_one_initcall+0x76/0x122 [ 170.141494] [<f94f4000>] ? 0xf94f3fff [ 170.141975] [<c0463f1e>] sys_init_module+0x1286/0x140d [ 170.142463] [<c07c6d9f>] sysenter_do_call+0x12/0x28 [ 170.142974] ---[ end trace 64e4fac3ae5fff1b ]--- [ 170.144949] bond%d: Warning: failed to register to debugfs Signed-off-by: Weiping Pan <panweiping3@gmail.com> Reviewed-by: Jiri Pirko <jpirko@redhat.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-05-12 15:46:56 +00:00
ret = dev_get_valid_name(dev, dev->name);
if (ret < 0)
goto out;
/* Init, if this function is available */
if (dev->netdev_ops->ndo_init) {
ret = dev->netdev_ops->ndo_init(dev);
if (ret) {
if (ret > 0)
ret = -EIO;
goto out;
}
}
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
dev->ifindex = dev_new_index(net);
if (dev->iflink == -1)
dev->iflink = dev->ifindex;
/* Transfer changeable features to wanted_features and enable
* software offloads (GSO and GRO).
*/
dev->hw_features |= NETIF_F_SOFT_FEATURES;
dev->features |= NETIF_F_SOFT_FEATURES;
dev->wanted_features = dev->features & dev->hw_features;
net: Allow no-cache copy from user on transmit This patch uses __copy_from_user_nocache on transmit to bypass data cache for a performance improvement. skb_add_data_nocache and skb_copy_to_page_nocache can be called by sendmsg functions to use this feature, initial support is in tcp_sendmsg. This functionality is configurable per device using ethtool. Presumably, this feature would only be useful when the driver does not touch the data. The feature is turned on by default if a device indicates that it does some form of checksum offload; it is off by default for devices that do no checksum offload or indicate no checksum is necessary. For the former case copy-checksum is probably done anyway, in the latter case the device is likely loopback in which case the no cache copy is probably not beneficial. This patch was tested using 200 instances of netperf TCP_RR with 1400 byte request and one byte reply. Platform is 16 core AMD x86. No-cache copy disabled: 672703 tps, 97.13% utilization 50/90/99% latency:244.31 484.205 1028.41 No-cache copy enabled: 702113 tps, 96.16% utilization, 50/90/99% latency 238.56 467.56 956.955 Using 14000 byte request and response sizes demonstrate the effects more dramatically: No-cache copy disabled: 79571 tps, 34.34 %utlization 50/90/95% latency 1584.46 2319.59 5001.76 No-cache copy enabled: 83856 tps, 34.81% utilization 50/90/95% latency 2508.42 2622.62 2735.88 Note especially the effect on latency tail (95th percentile). This seems to provide a nice performance improvement and is consistent in the tests I ran. Presumably, this would provide the greatest benfits in the presence of an application workload stressing the cache and a lot of transmit data happening. Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-04-05 05:30:30 +00:00
/* Turn on no cache copy if HW is doing checksum */
dev->hw_features |= NETIF_F_NOCACHE_COPY;
if ((dev->features & NETIF_F_ALL_CSUM) &&
!(dev->features & NETIF_F_NO_CSUM)) {
dev->wanted_features |= NETIF_F_NOCACHE_COPY;
dev->features |= NETIF_F_NOCACHE_COPY;
}
/* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
*/
dev->vlan_features |= NETIF_F_HIGHDMA;
ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
ret = notifier_to_errno(ret);
if (ret)
goto err_uninit;
ret = netdev_register_kobject(dev);
if (ret)
goto err_uninit;
dev->reg_state = NETREG_REGISTERED;
__netdev_update_features(dev);
/*
* Default initial state at registry is that the
* device is present.
*/
set_bit(__LINK_STATE_PRESENT, &dev->state);
dev_init_scheduler(dev);
dev_hold(dev);
list_netdevice(dev);
/* Notify protocols, that a new device appeared. */
ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
ret = notifier_to_errno(ret);
[NET]: Fix free_netdev on register_netdev failure. Point 1: The unregistering of a network device schedule a netdev_run_todo. This function calls dev->destructor when it is set and the destructor calls free_netdev. Point 2: In the case of an initialization of a network device the usual code is: * alloc_netdev * register_netdev -> if this one fails, call free_netdev and exit with error. Point 3: In the register_netdevice function at the later state, when the device is at the registered state, a call to the netdevice_notifiers is made. If one of the notification falls into an error, a rollback to the registered state is done using unregister_netdevice. Conclusion: When a network device fails to register during initialization because one network subsystem returned an error during a notification call chain, the network device is freed twice because of fact 1 and fact 2. The second free_netdev will be done with an invalid pointer. Proposed solution: The following patch move all the code of unregister_netdevice *except* the call to net_set_todo, to a new function "rollback_registered". The following functions are changed in this way: * register_netdevice: calls rollback_registered when a notification fails * unregister_netdevice: calls rollback_register + net_set_todo, the call order to net_set_todo is changed because it is the latest now. Since it justs add an element to a list that should not break anything. Signed-off-by: Daniel Lezcano <dlezcano@fr.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-10-30 22:38:18 +00:00
if (ret) {
rollback_registered(dev);
dev->reg_state = NETREG_UNREGISTERED;
}
/*
* Prevent userspace races by waiting until the network
* device is fully setup before sending notifications.
*/
if (!dev->rtnl_link_ops ||
dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U);
out:
return ret;
err_uninit:
if (dev->netdev_ops->ndo_uninit)
dev->netdev_ops->ndo_uninit(dev);
goto out;
}
EXPORT_SYMBOL(register_netdevice);
/**
* init_dummy_netdev - init a dummy network device for NAPI
* @dev: device to init
*
* This takes a network device structure and initialize the minimum
* amount of fields so it can be used to schedule NAPI polls without
* registering a full blown interface. This is to be used by drivers
* that need to tie several hardware interfaces to a single NAPI
* poll scheduler due to HW limitations.
*/
int init_dummy_netdev(struct net_device *dev)
{
/* Clear everything. Note we don't initialize spinlocks
* are they aren't supposed to be taken by any of the
* NAPI code and this dummy netdev is supposed to be
* only ever used for NAPI polls
*/
memset(dev, 0, sizeof(struct net_device));
/* make sure we BUG if trying to hit standard
* register/unregister code path
*/
dev->reg_state = NETREG_DUMMY;
/* NAPI wants this */
INIT_LIST_HEAD(&dev->napi_list);
/* a dummy interface is started by default */
set_bit(__LINK_STATE_PRESENT, &dev->state);
set_bit(__LINK_STATE_START, &dev->state);
/* Note : We dont allocate pcpu_refcnt for dummy devices,
* because users of this 'device' dont need to change
* its refcount.
*/
return 0;
}
EXPORT_SYMBOL_GPL(init_dummy_netdev);
/**
* register_netdev - register a network device
* @dev: device to register
*
* Take a completed network device structure and add it to the kernel
* interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
* chain. 0 is returned on success. A negative errno code is returned
* on a failure to set up the device, or if the name is a duplicate.
*
* This is a wrapper around register_netdevice that takes the rtnl semaphore
* and expands the device name if you passed a format string to
* alloc_netdev.
*/
int register_netdev(struct net_device *dev)
{
int err;
rtnl_lock();
err = register_netdevice(dev);
rtnl_unlock();
return err;
}
EXPORT_SYMBOL(register_netdev);
int netdev_refcnt_read(const struct net_device *dev)
{
int i, refcnt = 0;
for_each_possible_cpu(i)
refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
return refcnt;
}
EXPORT_SYMBOL(netdev_refcnt_read);
/*
* netdev_wait_allrefs - wait until all references are gone.
*
* This is called when unregistering network devices.
*
* Any protocol or device that holds a reference should register
* for netdevice notification, and cleanup and put back the
* reference if they receive an UNREGISTER event.
* We can get stuck here if buggy protocols don't correctly
* call dev_put.
*/
static void netdev_wait_allrefs(struct net_device *dev)
{
unsigned long rebroadcast_time, warning_time;
int refcnt;
linkwatch: linkwatch_forget_dev() to speedup device dismantle Herbert Xu a écrit : > On Tue, Nov 17, 2009 at 04:26:04AM -0800, David Miller wrote: >> Really, the link watch stuff is just due for a redesign. I don't >> think a simple hack is going to cut it this time, sorry Eric :-) > > I have no objections against any redesigns, but since the only > caller of linkwatch_forget_dev runs in process context with the > RTNL, it could also legally emit those events. Thanks guys, here an updated version then, before linkwatch surgery ? In this version, I force the event to be sent synchronously. [PATCH net-next-2.6] linkwatch: linkwatch_forget_dev() to speedup device dismantle time ip link del eth3.103 ; time ip link del eth3.104 ; time ip link del eth3.105 real 0m0.266s user 0m0.000s sys 0m0.001s real 0m0.770s user 0m0.000s sys 0m0.000s real 0m1.022s user 0m0.000s sys 0m0.000s One problem of current schem in vlan dismantle phase is the holding of device done by following chain : vlan_dev_stop() -> netif_carrier_off(dev) -> linkwatch_fire_event(dev) -> dev_hold() ... And __linkwatch_run_queue() runs up to one second later... A generic fix to this problem is to add a linkwatch_forget_dev() method to unlink the device from the list of watched devices. dev->link_watch_next becomes dev->link_watch_list (and use a bit more memory), to be able to unlink device in O(1). After patch : time ip link del eth3.103 ; time ip link del eth3.104 ; time ip link del eth3.105 real 0m0.024s user 0m0.000s sys 0m0.000s real 0m0.032s user 0m0.000s sys 0m0.001s real 0m0.033s user 0m0.000s sys 0m0.000s Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2009-11-17 05:59:21 +00:00
linkwatch_forget_dev(dev);
rebroadcast_time = warning_time = jiffies;
refcnt = netdev_refcnt_read(dev);
while (refcnt != 0) {
if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
rtnl_lock();
/* Rebroadcast unregister notification */
call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
/* don't resend NETDEV_UNREGISTER_BATCH, _BATCH users
* should have already handle it the first time */
if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
&dev->state)) {
/* We must not have linkwatch events
* pending on unregister. If this
* happens, we simply run the queue
* unscheduled, resulting in a noop
* for this device.
*/
linkwatch_run_queue();
}
__rtnl_unlock();
rebroadcast_time = jiffies;
}
msleep(250);
refcnt = netdev_refcnt_read(dev);
if (time_after(jiffies, warning_time + 10 * HZ)) {
printk(KERN_EMERG "unregister_netdevice: "
"waiting for %s to become free. Usage "
"count = %d\n",
dev->name, refcnt);
warning_time = jiffies;
}
}
}
/* The sequence is:
*
* rtnl_lock();
* ...
* register_netdevice(x1);
* register_netdevice(x2);
* ...
* unregister_netdevice(y1);
* unregister_netdevice(y2);
* ...
* rtnl_unlock();
* free_netdev(y1);
* free_netdev(y2);
*
* We are invoked by rtnl_unlock().
* This allows us to deal with problems:
* 1) We can delete sysfs objects which invoke hotplug
* without deadlocking with linkwatch via keventd.
* 2) Since we run with the RTNL semaphore not held, we can sleep
* safely in order to wait for the netdev refcnt to drop to zero.
*
* We must not return until all unregister events added during
* the interval the lock was held have been completed.
*/
void netdev_run_todo(void)
{
struct list_head list;
/* Snapshot list, allow later requests */
list_replace_init(&net_todo_list, &list);
__rtnl_unlock();
/* Wait for rcu callbacks to finish before attempting to drain
* the device list. This usually avoids a 250ms wait.
*/
if (!list_empty(&list))
rcu_barrier();
while (!list_empty(&list)) {
struct net_device *dev
= list_first_entry(&list, struct net_device, todo_list);
list_del(&dev->todo_list);
if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
printk(KERN_ERR "network todo '%s' but state %d\n",
dev->name, dev->reg_state);
dump_stack();
continue;
}
dev->reg_state = NETREG_UNREGISTERED;
on_each_cpu(flush_backlog, dev, 1);
netdev_wait_allrefs(dev);
/* paranoia */
BUG_ON(netdev_refcnt_read(dev));
WARN_ON(rcu_access_pointer(dev->ip_ptr));
WARN_ON(rcu_access_pointer(dev->ip6_ptr));
WARN_ON(dev->dn_ptr);
if (dev->destructor)
dev->destructor(dev);
/* Free network device */
kobject_put(&dev->dev.kobj);
}
}
/* Convert net_device_stats to rtnl_link_stats64. They have the same
* fields in the same order, with only the type differing.
*/
static void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
const struct net_device_stats *netdev_stats)
{
#if BITS_PER_LONG == 64
BUILD_BUG_ON(sizeof(*stats64) != sizeof(*netdev_stats));
memcpy(stats64, netdev_stats, sizeof(*stats64));
#else
size_t i, n = sizeof(*stats64) / sizeof(u64);
const unsigned long *src = (const unsigned long *)netdev_stats;
u64 *dst = (u64 *)stats64;
BUILD_BUG_ON(sizeof(*netdev_stats) / sizeof(unsigned long) !=
sizeof(*stats64) / sizeof(u64));
for (i = 0; i < n; i++)
dst[i] = src[i];
#endif
}
/**
* dev_get_stats - get network device statistics
* @dev: device to get statistics from
* @storage: place to store stats
*
* Get network statistics from device. Return @storage.
* The device driver may provide its own method by setting
* dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
* otherwise the internal statistics structure is used.
*/
struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
struct rtnl_link_stats64 *storage)
{
const struct net_device_ops *ops = dev->netdev_ops;
if (ops->ndo_get_stats64) {
memset(storage, 0, sizeof(*storage));
ops->ndo_get_stats64(dev, storage);
} else if (ops->ndo_get_stats) {
netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
} else {
netdev_stats_to_stats64(storage, &dev->stats);
}
storage->rx_dropped += atomic_long_read(&dev->rx_dropped);
return storage;
}
EXPORT_SYMBOL(dev_get_stats);
struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
{
struct netdev_queue *queue = dev_ingress_queue(dev);
#ifdef CONFIG_NET_CLS_ACT
if (queue)
return queue;
queue = kzalloc(sizeof(*queue), GFP_KERNEL);
if (!queue)
return NULL;
netdev_init_one_queue(dev, queue, NULL);
queue->qdisc = &noop_qdisc;
queue->qdisc_sleeping = &noop_qdisc;
rcu_assign_pointer(dev->ingress_queue, queue);
#endif
return queue;
}
/**
* alloc_netdev_mqs - allocate network device
* @sizeof_priv: size of private data to allocate space for
* @name: device name format string
* @setup: callback to initialize device
* @txqs: the number of TX subqueues to allocate
* @rxqs: the number of RX subqueues to allocate
*
* Allocates a struct net_device with private data area for driver use
* and performs basic initialization. Also allocates subquue structs
* for each queue on the device.
*/
struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
void (*setup)(struct net_device *),
unsigned int txqs, unsigned int rxqs)
{
struct net_device *dev;
size_t alloc_size;
struct net_device *p;
BUG_ON(strlen(name) >= sizeof(dev->name));
if (txqs < 1) {
pr_err("alloc_netdev: Unable to allocate device "
"with zero queues.\n");
return NULL;
}
#ifdef CONFIG_RPS
if (rxqs < 1) {
pr_err("alloc_netdev: Unable to allocate device "
"with zero RX queues.\n");
return NULL;
}
#endif
alloc_size = sizeof(struct net_device);
if (sizeof_priv) {
/* ensure 32-byte alignment of private area */
alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
alloc_size += sizeof_priv;
}
/* ensure 32-byte alignment of whole construct */
alloc_size += NETDEV_ALIGN - 1;
p = kzalloc(alloc_size, GFP_KERNEL);
if (!p) {
printk(KERN_ERR "alloc_netdev: Unable to allocate device.\n");
return NULL;
}
dev = PTR_ALIGN(p, NETDEV_ALIGN);
dev->padded = (char *)dev - (char *)p;
dev->pcpu_refcnt = alloc_percpu(int);
if (!dev->pcpu_refcnt)
goto free_p;
if (dev_addr_init(dev))
goto free_pcpu;
dev_mc_init(dev);
dev_uc_init(dev);
dev_net_set(dev, &init_net);
dev->gso_max_size = GSO_MAX_SIZE;
INIT_LIST_HEAD(&dev->napi_list);
INIT_LIST_HEAD(&dev->unreg_list);
INIT_LIST_HEAD(&dev->link_watch_list);
dev->priv_flags = IFF_XMIT_DST_RELEASE;
setup(dev);
dev->num_tx_queues = txqs;
dev->real_num_tx_queues = txqs;
if (netif_alloc_netdev_queues(dev))
goto free_all;
#ifdef CONFIG_RPS
dev->num_rx_queues = rxqs;
dev->real_num_rx_queues = rxqs;
if (netif_alloc_rx_queues(dev))
goto free_all;
#endif
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
strcpy(dev->name, name);
dev->group = INIT_NETDEV_GROUP;
return dev;
free_all:
free_netdev(dev);
return NULL;
free_pcpu:
free_percpu(dev->pcpu_refcnt);
kfree(dev->_tx);
#ifdef CONFIG_RPS
kfree(dev->_rx);
#endif
free_p:
kfree(p);
return NULL;
}
EXPORT_SYMBOL(alloc_netdev_mqs);
/**
* free_netdev - free network device
* @dev: device
*
* This function does the last stage of destroying an allocated device
* interface. The reference to the device object is released.
* If this is the last reference then it will be freed.
*/
void free_netdev(struct net_device *dev)
{
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
struct napi_struct *p, *n;
release_net(dev_net(dev));
kfree(dev->_tx);
#ifdef CONFIG_RPS
kfree(dev->_rx);
#endif
kfree(rcu_dereference_protected(dev->ingress_queue, 1));
/* Flush device addresses */
dev_addr_flush(dev);
net: Add Generic Receive Offload infrastructure This patch adds the top-level GRO (Generic Receive Offload) infrastructure. This is pretty similar to LRO except that this is protocol-independent. Instead of holding packets in an lro_mgr structure, they're now held in napi_struct. For drivers that intend to use this, they can set the NETIF_F_GRO bit and call napi_gro_receive instead of netif_receive_skb or just call netif_rx. The latter will call napi_receive_skb automatically. When napi_gro_receive is used, the driver must either call napi_complete/napi_rx_complete, or call napi_gro_flush in softirq context if the driver uses the primitives __napi_complete/__napi_rx_complete. Protocols will set the gro_receive and gro_complete function pointers in order to participate in this scheme. In addition to the packet, gro_receive will get a list of currently held packets. Each packet in the list has a same_flow field which is non-zero if it is a potential match for the new packet. For each packet that may match, they also have a flush field which is non-zero if the held packet must not be merged with the new packet. Once gro_receive has determined that the new skb matches a held packet, the held packet may be processed immediately if the new skb cannot be merged with it. In this case gro_receive should return the pointer to the existing skb in gro_list. Otherwise the new skb should be merged into the existing packet and NULL should be returned, unless the new skb makes it impossible for any further merges to be made (e.g., FIN packet) where the merged skb should be returned. Whenever the skb is merged into an existing entry, the gro_receive function should set NAPI_GRO_CB(skb)->same_flow. Note that if an skb merely matches an existing entry but can't be merged with it, then this shouldn't be set. If gro_receive finds it pointless to hold the new skb for future merging, it should set NAPI_GRO_CB(skb)->flush. Held packets will be flushed by napi_gro_flush which is called by napi_complete and napi_rx_complete. Currently held packets are stored in a singly liked list just like LRO. The list is limited to a maximum of 8 entries. In future, this may be expanded to use a hash table to allow more flows to be held for merging. Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2008-12-16 07:38:52 +00:00
list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
netif_napi_del(p);
free_percpu(dev->pcpu_refcnt);
dev->pcpu_refcnt = NULL;
/* Compatibility with error handling in drivers */
if (dev->reg_state == NETREG_UNINITIALIZED) {
kfree((char *)dev - dev->padded);
return;
}
BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
dev->reg_state = NETREG_RELEASED;
/* will free via device release */
put_device(&dev->dev);
}
EXPORT_SYMBOL(free_netdev);
/**
* synchronize_net - Synchronize with packet receive processing
*
* Wait for packets currently being received to be done.
* Does not block later packets from starting.
*/
void synchronize_net(void)
{
might_sleep();
net: use synchronize_rcu_expedited() synchronize_rcu() is very slow in various situations (HZ=100, CONFIG_NO_HZ=y, CONFIG_PREEMPT=n) Extract from my (mostly idle) 8 core machine : synchronize_rcu() in 99985 us synchronize_rcu() in 79982 us synchronize_rcu() in 87612 us synchronize_rcu() in 79827 us synchronize_rcu() in 109860 us synchronize_rcu() in 98039 us synchronize_rcu() in 89841 us synchronize_rcu() in 79842 us synchronize_rcu() in 80151 us synchronize_rcu() in 119833 us synchronize_rcu() in 99858 us synchronize_rcu() in 73999 us synchronize_rcu() in 79855 us synchronize_rcu() in 79853 us When we hold RTNL mutex, we would like to spend some cpu cycles but not block too long other processes waiting for this mutex. We also want to setup/dismantle network features as fast as possible at boot/shutdown time. This patch makes synchronize_net() call the expedited version if RTNL is locked. synchronize_rcu_expedited() typical delay is about 20 us on my machine. synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 20 us synchronize_rcu_expedited() in 16 us synchronize_rcu_expedited() in 20 us synchronize_rcu_expedited() in 18 us synchronize_rcu_expedited() in 18 us Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> CC: Paul E. McKenney <paulmck@linux.vnet.ibm.com> CC: Ben Greear <greearb@candelatech.com> Reviewed-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-05-23 23:07:32 +00:00
if (rtnl_is_locked())
synchronize_rcu_expedited();
else
synchronize_rcu();
}
EXPORT_SYMBOL(synchronize_net);
/**
* unregister_netdevice_queue - remove device from the kernel
* @dev: device
* @head: list
*
* This function shuts down a device interface and removes it
* from the kernel tables.
* If head not NULL, device is queued to be unregistered later.
*
* Callers must hold the rtnl semaphore. You may want
* unregister_netdev() instead of this.
*/
void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
{
ASSERT_RTNL();
if (head) {
list_move_tail(&dev->unreg_list, head);
} else {
rollback_registered(dev);
/* Finish processing unregister after unlock */
net_set_todo(dev);
}
}
EXPORT_SYMBOL(unregister_netdevice_queue);
/**
* unregister_netdevice_many - unregister many devices
* @head: list of devices
*/
void unregister_netdevice_many(struct list_head *head)
{
struct net_device *dev;
if (!list_empty(head)) {
rollback_registered_many(head);
list_for_each_entry(dev, head, unreg_list)
net_set_todo(dev);
}
}
EXPORT_SYMBOL(unregister_netdevice_many);
/**
* unregister_netdev - remove device from the kernel
* @dev: device
*
* This function shuts down a device interface and removes it
* from the kernel tables.
*
* This is just a wrapper for unregister_netdevice that takes
* the rtnl semaphore. In general you want to use this and not
* unregister_netdevice.
*/
void unregister_netdev(struct net_device *dev)
{
rtnl_lock();
unregister_netdevice(dev);
rtnl_unlock();
}
EXPORT_SYMBOL(unregister_netdev);
/**
* dev_change_net_namespace - move device to different nethost namespace
* @dev: device
* @net: network namespace
* @pat: If not NULL name pattern to try if the current device name
* is already taken in the destination network namespace.
*
* This function shuts down a device interface and moves it
* to a new network namespace. On success 0 is returned, on
* a failure a netagive errno code is returned.
*
* Callers must hold the rtnl semaphore.
*/
int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
{
int err;
ASSERT_RTNL();
/* Don't allow namespace local devices to be moved. */
err = -EINVAL;
if (dev->features & NETIF_F_NETNS_LOCAL)
goto out;
/* Ensure the device has been registrered */
err = -EINVAL;
if (dev->reg_state != NETREG_REGISTERED)
goto out;
/* Get out if there is nothing todo */
err = 0;
if (net_eq(dev_net(dev), net))
goto out;
/* Pick the destination device name, and ensure
* we can use it in the destination network namespace.
*/
err = -EEXIST;
if (__dev_get_by_name(net, dev->name)) {
/* We get here if we can't use the current device name */
if (!pat)
goto out;
if (dev_get_valid_name(dev, pat) < 0)
goto out;
}
/*
* And now a mini version of register_netdevice unregister_netdevice.
*/
/* If device is running close it first. */
dev_close(dev);
/* And unlink it from device chain */
err = -ENODEV;
unlist_netdevice(dev);
synchronize_net();
/* Shutdown queueing discipline. */
dev_shutdown(dev);
/* Notify protocols, that we are about to destroy
this device. They should clean all the things.
Note that dev->reg_state stays at NETREG_REGISTERED.
This is wanted because this way 8021q and macvlan know
the device is just moving and can keep their slaves up.
*/
call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
call_netdevice_notifiers(NETDEV_UNREGISTER_BATCH, dev);
rtmsg_ifinfo(RTM_DELLINK, dev, ~0U);
/*
* Flush the unicast and multicast chains
*/
dev_uc_flush(dev);
dev_mc_flush(dev);
/* Actually switch the network namespace */
dev_net_set(dev, net);
/* If there is an ifindex conflict assign a new one */
if (__dev_get_by_index(net, dev->ifindex)) {
int iflink = (dev->iflink == dev->ifindex);
dev->ifindex = dev_new_index(net);
if (iflink)
dev->iflink = dev->ifindex;
}
/* Fixup kobjects */
err = device_rename(&dev->dev, dev->name);
WARN_ON(err);
/* Add the device back in the hashes */
list_netdevice(dev);
/* Notify protocols, that a new device appeared. */
call_netdevice_notifiers(NETDEV_REGISTER, dev);
/*
* Prevent userspace races by waiting until the network
* device is fully setup before sending notifications.
*/
rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U);
synchronize_net();
err = 0;
out:
return err;
}
EXPORT_SYMBOL_GPL(dev_change_net_namespace);
static int dev_cpu_callback(struct notifier_block *nfb,
unsigned long action,
void *ocpu)
{
struct sk_buff **list_skb;
struct sk_buff *skb;
unsigned int cpu, oldcpu = (unsigned long)ocpu;
struct softnet_data *sd, *oldsd;
if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
return NOTIFY_OK;
local_irq_disable();
cpu = smp_processor_id();
sd = &per_cpu(softnet_data, cpu);
oldsd = &per_cpu(softnet_data, oldcpu);
/* Find end of our completion_queue. */
list_skb = &sd->completion_queue;
while (*list_skb)
list_skb = &(*list_skb)->next;
/* Append completion queue from offline CPU. */
*list_skb = oldsd->completion_queue;
oldsd->completion_queue = NULL;
/* Append output queue from offline CPU. */
if (oldsd->output_queue) {
*sd->output_queue_tailp = oldsd->output_queue;
sd->output_queue_tailp = oldsd->output_queue_tailp;
oldsd->output_queue = NULL;
oldsd->output_queue_tailp = &oldsd->output_queue;
}
/* Append NAPI poll list from offline CPU. */
if (!list_empty(&oldsd->poll_list)) {
list_splice_init(&oldsd->poll_list, &sd->poll_list);
raise_softirq_irqoff(NET_RX_SOFTIRQ);
}
raise_softirq_irqoff(NET_TX_SOFTIRQ);
local_irq_enable();
/* Process offline CPU's input_pkt_queue */
while ((skb = __skb_dequeue(&oldsd->process_queue))) {
netif_rx(skb);
input_queue_head_incr(oldsd);
rfs: Receive Flow Steering This patch implements receive flow steering (RFS). RFS steers received packets for layer 3 and 4 processing to the CPU where the application for the corresponding flow is running. RFS is an extension of Receive Packet Steering (RPS). The basic idea of RFS is that when an application calls recvmsg (or sendmsg) the application's running CPU is stored in a hash table that is indexed by the connection's rxhash which is stored in the socket structure. The rxhash is passed in skb's received on the connection from netif_receive_skb. For each received packet, the associated rxhash is used to look up the CPU in the hash table, if a valid CPU is set then the packet is steered to that CPU using the RPS mechanisms. The convolution of the simple approach is that it would potentially allow OOO packets. If threads are thrashing around CPUs or multiple threads are trying to read from the same sockets, a quickly changing CPU value in the hash table could cause rampant OOO packets-- we consider this a non-starter. To avoid OOO packets, this solution implements two types of hash tables: rps_sock_flow_table and rps_dev_flow_table. rps_sock_table is a global hash table. Each entry is just a CPU number and it is populated in recvmsg and sendmsg as described above. This table contains the "desired" CPUs for flows. rps_dev_flow_table is specific to each device queue. Each entry contains a CPU and a tail queue counter. The CPU is the "current" CPU for a matching flow. The tail queue counter holds the value of a tail queue counter for the associated CPU's backlog queue at the time of last enqueue for a flow matching the entry. Each backlog queue has a queue head counter which is incremented on dequeue, and so a queue tail counter is computed as queue head count + queue length. When a packet is enqueued on a backlog queue, the current value of the queue tail counter is saved in the hash entry of the rps_dev_flow_table. And now the trick: when selecting the CPU for RPS (get_rps_cpu) the rps_sock_flow table and the rps_dev_flow table for the RX queue are consulted. When the desired CPU for the flow (found in the rps_sock_flow table) does not match the current CPU (found in the rps_dev_flow table), the current CPU is changed to the desired CPU if one of the following is true: - The current CPU is unset (equal to RPS_NO_CPU) - Current CPU is offline - The current CPU's queue head counter >= queue tail counter in the rps_dev_flow table. This checks if the queue tail has advanced beyond the last packet that was enqueued using this table entry. This guarantees that all packets queued using this entry have been dequeued, thus preserving in order delivery. Making each queue have its own rps_dev_flow table has two advantages: 1) the tail queue counters will be written on each receive, so keeping the table local to interrupting CPU s good for locality. 2) this allows lockless access to the table-- the CPU number and queue tail counter need to be accessed together under mutual exclusion from netif_receive_skb, we assume that this is only called from device napi_poll which is non-reentrant. This patch implements RFS for TCP and connected UDP sockets. It should be usable for other flow oriented protocols. There are two configuration parameters for RFS. The "rps_flow_entries" kernel init parameter sets the number of entries in the rps_sock_flow_table, the per rxqueue sysfs entry "rps_flow_cnt" contains the number of entries in the rps_dev_flow table for the rxqueue. Both are rounded to power of two. The obvious benefit of RFS (over just RPS) is that it achieves CPU locality between the receive processing for a flow and the applications processing; this can result in increased performance (higher pps, lower latency). The benefits of RFS are dependent on cache hierarchy, application load, and other factors. On simple benchmarks, we don't necessarily see improvement and sometimes see degradation. However, for more complex benchmarks and for applications where cache pressure is much higher this technique seems to perform very well. Below are some benchmark results which show the potential benfit of this patch. The netperf test has 500 instances of netperf TCP_RR test with 1 byte req. and resp. The RPC test is an request/response test similar in structure to netperf RR test ith 100 threads on each host, but does more work in userspace that netperf. e1000e on 8 core Intel No RFS or RPS 104K tps at 30% CPU No RFS (best RPS config): 290K tps at 63% CPU RFS 303K tps at 61% CPU RPC test tps CPU% 50/90/99% usec latency Latency StdDev No RFS/RPS 103K 48% 757/900/3185 4472.35 RPS only: 174K 73% 415/993/2468 491.66 RFS 223K 73% 379/651/1382 315.61 Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-04-16 23:01:27 +00:00
}
while ((skb = __skb_dequeue(&oldsd->input_pkt_queue))) {
netif_rx(skb);
input_queue_head_incr(oldsd);
}
return NOTIFY_OK;
}
/**
* netdev_increment_features - increment feature set by one
* @all: current feature set
* @one: new feature set
* @mask: mask feature set
*
* Computes a new feature set after adding a device with feature set
* @one to the master device with current feature set @all. Will not
* enable anything that is off in @mask. Returns the new feature set.
*/
u32 netdev_increment_features(u32 all, u32 one, u32 mask)
{
if (mask & NETIF_F_GEN_CSUM)
mask |= NETIF_F_ALL_CSUM;
mask |= NETIF_F_VLAN_CHALLENGED;
all |= one & (NETIF_F_ONE_FOR_ALL|NETIF_F_ALL_CSUM) & mask;
all &= one | ~NETIF_F_ALL_FOR_ALL;
net: Allow no-cache copy from user on transmit This patch uses __copy_from_user_nocache on transmit to bypass data cache for a performance improvement. skb_add_data_nocache and skb_copy_to_page_nocache can be called by sendmsg functions to use this feature, initial support is in tcp_sendmsg. This functionality is configurable per device using ethtool. Presumably, this feature would only be useful when the driver does not touch the data. The feature is turned on by default if a device indicates that it does some form of checksum offload; it is off by default for devices that do no checksum offload or indicate no checksum is necessary. For the former case copy-checksum is probably done anyway, in the latter case the device is likely loopback in which case the no cache copy is probably not beneficial. This patch was tested using 200 instances of netperf TCP_RR with 1400 byte request and one byte reply. Platform is 16 core AMD x86. No-cache copy disabled: 672703 tps, 97.13% utilization 50/90/99% latency:244.31 484.205 1028.41 No-cache copy enabled: 702113 tps, 96.16% utilization, 50/90/99% latency 238.56 467.56 956.955 Using 14000 byte request and response sizes demonstrate the effects more dramatically: No-cache copy disabled: 79571 tps, 34.34 %utlization 50/90/95% latency 1584.46 2319.59 5001.76 No-cache copy enabled: 83856 tps, 34.81% utilization 50/90/95% latency 2508.42 2622.62 2735.88 Note especially the effect on latency tail (95th percentile). This seems to provide a nice performance improvement and is consistent in the tests I ran. Presumably, this would provide the greatest benfits in the presence of an application workload stressing the cache and a lot of transmit data happening. Signed-off-by: Tom Herbert <therbert@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2011-04-05 05:30:30 +00:00
/* If device needs checksumming, downgrade to it. */
if (all & (NETIF_F_ALL_CSUM & ~NETIF_F_NO_CSUM))
all &= ~NETIF_F_NO_CSUM;
/* If one device supports hw checksumming, set for all. */
if (all & NETIF_F_GEN_CSUM)
all &= ~(NETIF_F_ALL_CSUM & ~NETIF_F_GEN_CSUM);
return all;
}
EXPORT_SYMBOL(netdev_increment_features);
static struct hlist_head *netdev_create_hash(void)
{
int i;
struct hlist_head *hash;
hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
if (hash != NULL)
for (i = 0; i < NETDEV_HASHENTRIES; i++)
INIT_HLIST_HEAD(&hash[i]);
return hash;
}
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
/* Initialize per network namespace state */
static int __net_init netdev_init(struct net *net)
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
{
INIT_LIST_HEAD(&net->dev_base_head);
net->dev_name_head = netdev_create_hash();
if (net->dev_name_head == NULL)
goto err_name;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
net->dev_index_head = netdev_create_hash();
if (net->dev_index_head == NULL)
goto err_idx;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
return 0;
err_idx:
kfree(net->dev_name_head);
err_name:
return -ENOMEM;
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
}
/**
* netdev_drivername - network driver for the device
* @dev: network device
*
* Determine network driver for device.
*/
const char *netdev_drivername(const struct net_device *dev)
{
const struct device_driver *driver;
const struct device *parent;
const char *empty = "";
parent = dev->dev.parent;
if (!parent)
return empty;
driver = parent->driver;
if (driver && driver->name)
return driver->name;
return empty;
}
int __netdev_printk(const char *level, const struct net_device *dev,
struct va_format *vaf)
{
int r;
if (dev && dev->dev.parent)
r = dev_printk(level, dev->dev.parent, "%s: %pV",
netdev_name(dev), vaf);
else if (dev)
r = printk("%s%s: %pV", level, netdev_name(dev), vaf);
else
r = printk("%s(NULL net_device): %pV", level, vaf);
return r;
}
EXPORT_SYMBOL(__netdev_printk);
int netdev_printk(const char *level, const struct net_device *dev,
const char *format, ...)
{
struct va_format vaf;
va_list args;
int r;
va_start(args, format);
vaf.fmt = format;
vaf.va = &args;
r = __netdev_printk(level, dev, &vaf);
va_end(args);
return r;
}
EXPORT_SYMBOL(netdev_printk);
#define define_netdev_printk_level(func, level) \
int func(const struct net_device *dev, const char *fmt, ...) \
{ \
int r; \
struct va_format vaf; \
va_list args; \
\
va_start(args, fmt); \
\
vaf.fmt = fmt; \
vaf.va = &args; \
\
r = __netdev_printk(level, dev, &vaf); \
va_end(args); \
\
return r; \
} \
EXPORT_SYMBOL(func);
define_netdev_printk_level(netdev_emerg, KERN_EMERG);
define_netdev_printk_level(netdev_alert, KERN_ALERT);
define_netdev_printk_level(netdev_crit, KERN_CRIT);
define_netdev_printk_level(netdev_err, KERN_ERR);
define_netdev_printk_level(netdev_warn, KERN_WARNING);
define_netdev_printk_level(netdev_notice, KERN_NOTICE);
define_netdev_printk_level(netdev_info, KERN_INFO);
static void __net_exit netdev_exit(struct net *net)
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
{
kfree(net->dev_name_head);
kfree(net->dev_index_head);
}
static struct pernet_operations __net_initdata netdev_net_ops = {
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
.init = netdev_init,
.exit = netdev_exit,
};
static void __net_exit default_device_exit(struct net *net)
{
struct net_device *dev, *aux;
/*
* Push all migratable network devices back to the
* initial network namespace
*/
rtnl_lock();
for_each_netdev_safe(net, dev, aux) {
int err;
char fb_name[IFNAMSIZ];
/* Ignore unmoveable devices (i.e. loopback) */
if (dev->features & NETIF_F_NETNS_LOCAL)
continue;
/* Leave virtual devices for the generic cleanup */
if (dev->rtnl_link_ops)
continue;
/* Push remaining network devices to init_net */
snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
err = dev_change_net_namespace(dev, &init_net, fb_name);
if (err) {
printk(KERN_EMERG "%s: failed to move %s to init_net: %d\n",
__func__, dev->name, err);
BUG();
}
}
rtnl_unlock();
}
static void __net_exit default_device_exit_batch(struct list_head *net_list)
{
/* At exit all network devices most be removed from a network
* namespace. Do this in the reverse order of registration.
* Do this across as many network namespaces as possible to
* improve batching efficiency.
*/
struct net_device *dev;
struct net *net;
LIST_HEAD(dev_kill_list);
rtnl_lock();
list_for_each_entry(net, net_list, exit_list) {
for_each_netdev_reverse(net, dev) {
if (dev->rtnl_link_ops)
dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
else
unregister_netdevice_queue(dev, &dev_kill_list);
}
}
unregister_netdevice_many(&dev_kill_list);
list_del(&dev_kill_list);
rtnl_unlock();
}
static struct pernet_operations __net_initdata default_device_ops = {
.exit = default_device_exit,
.exit_batch = default_device_exit_batch,
};
/*
* Initialize the DEV module. At boot time this walks the device list and
* unhooks any devices that fail to initialise (normally hardware not
* present) and leaves us with a valid list of present and active devices.
*
*/
/*
* This is called single threaded during boot, so no need
* to take the rtnl semaphore.
*/
static int __init net_dev_init(void)
{
int i, rc = -ENOMEM;
BUG_ON(!dev_boot_phase);
if (dev_proc_init())
goto out;
if (netdev_kobject_init())
goto out;
INIT_LIST_HEAD(&ptype_all);
for (i = 0; i < PTYPE_HASH_SIZE; i++)
INIT_LIST_HEAD(&ptype_base[i]);
[NET]: Make the device list and device lookups per namespace. This patch makes most of the generic device layer network namespace safe. This patch makes dev_base_head a network namespace variable, and then it picks up a few associated variables. The functions: dev_getbyhwaddr dev_getfirsthwbytype dev_get_by_flags dev_get_by_name __dev_get_by_name dev_get_by_index __dev_get_by_index dev_ioctl dev_ethtool dev_load wireless_process_ioctl were modified to take a network namespace argument, and deal with it. vlan_ioctl_set and brioctl_set were modified so their hooks will receive a network namespace argument. So basically anthing in the core of the network stack that was affected to by the change of dev_base was modified to handle multiple network namespaces. The rest of the network stack was simply modified to explicitly use &init_net the initial network namespace. This can be fixed when those components of the network stack are modified to handle multiple network namespaces. For now the ifindex generator is left global. Fundametally ifindex numbers are per namespace, or else we will have corner case problems with migration when we get that far. At the same time there are assumptions in the network stack that the ifindex of a network device won't change. Making the ifindex number global seems a good compromise until the network stack can cope with ifindex changes when you change namespaces, and the like. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2007-09-17 18:56:21 +00:00
if (register_pernet_subsys(&netdev_net_ops))
goto out;
/*
* Initialise the packet receive queues.
*/
for_each_possible_cpu(i) {
struct softnet_data *sd = &per_cpu(softnet_data, i);
memset(sd, 0, sizeof(*sd));
skb_queue_head_init(&sd->input_pkt_queue);
skb_queue_head_init(&sd->process_queue);
sd->completion_queue = NULL;
INIT_LIST_HEAD(&sd->poll_list);
sd->output_queue = NULL;
sd->output_queue_tailp = &sd->output_queue;
#ifdef CONFIG_RPS
sd->csd.func = rps_trigger_softirq;
sd->csd.info = sd;
sd->csd.flags = 0;
sd->cpu = i;
#endif
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
sd->backlog.poll = process_backlog;
sd->backlog.weight = weight_p;
sd->backlog.gro_list = NULL;
sd->backlog.gro_count = 0;
}
dev_boot_phase = 0;
/* The loopback device is special if any other network devices
* is present in a network namespace the loopback device must
* be present. Since we now dynamically allocate and free the
* loopback device ensure this invariant is maintained by
* keeping the loopback device as the first device on the
* list of network devices. Ensuring the loopback devices
* is the first device that appears and the last network device
* that disappears.
*/
if (register_pernet_device(&loopback_net_ops))
goto out;
if (register_pernet_device(&default_device_ops))
goto out;
open_softirq(NET_TX_SOFTIRQ, net_tx_action);
open_softirq(NET_RX_SOFTIRQ, net_rx_action);
hotcpu_notifier(dev_cpu_callback, 0);
dst_init();
dev_mcast_init();
rc = 0;
out:
return rc;
}
subsys_initcall(net_dev_init);
static int __init initialize_hashrnd(void)
{
rps: Receive Packet Steering This patch implements software receive side packet steering (RPS). RPS distributes the load of received packet processing across multiple CPUs. Problem statement: Protocol processing done in the NAPI context for received packets is serialized per device queue and becomes a bottleneck under high packet load. This substantially limits pps that can be achieved on a single queue NIC and provides no scaling with multiple cores. This solution queues packets early on in the receive path on the backlog queues of other CPUs. This allows protocol processing (e.g. IP and TCP) to be performed on packets in parallel. For each device (or each receive queue in a multi-queue device) a mask of CPUs is set to indicate the CPUs that can process packets. A CPU is selected on a per packet basis by hashing contents of the packet header (e.g. the TCP or UDP 4-tuple) and using the result to index into the CPU mask. The IPI mechanism is used to raise networking receive softirqs between CPUs. This effectively emulates in software what a multi-queue NIC can provide, but is generic requiring no device support. Many devices now provide a hash over the 4-tuple on a per packet basis (e.g. the Toeplitz hash). This patch allow drivers to set the HW reported hash in an skb field, and that value in turn is used to index into the RPS maps. Using the HW generated hash can avoid cache misses on the packet when steering it to a remote CPU. The CPU mask is set on a per device and per queue basis in the sysfs variable /sys/class/net/<device>/queues/rx-<n>/rps_cpus. This is a set of canonical bit maps for receive queues in the device (numbered by <n>). If a device does not support multi-queue, a single variable is used for the device (rx-0). Generally, we have found this technique increases pps capabilities of a single queue device with good CPU utilization. Optimal settings for the CPU mask seem to depend on architectures and cache hierarcy. Below are some results running 500 instances of netperf TCP_RR test with 1 byte req. and resp. Results show cumulative transaction rate and system CPU utilization. e1000e on 8 core Intel Without RPS: 108K tps at 33% CPU With RPS: 311K tps at 64% CPU forcedeth on 16 core AMD Without RPS: 156K tps at 15% CPU With RPS: 404K tps at 49% CPU bnx2x on 16 core AMD Without RPS 567K tps at 61% CPU (4 HW RX queues) Without RPS 738K tps at 96% CPU (8 HW RX queues) With RPS: 854K tps at 76% CPU (4 HW RX queues) Caveats: - The benefits of this patch are dependent on architecture and cache hierarchy. Tuning the masks to get best performance is probably necessary. - This patch adds overhead in the path for processing a single packet. In a lightly loaded server this overhead may eliminate the advantages of increased parallelism, and possibly cause some relative performance degradation. We have found that masks that are cache aware (share same caches with the interrupting CPU) mitigate much of this. - The RPS masks can be changed dynamically, however whenever the mask is changed this introduces the possibility of generating out of order packets. It's probably best not change the masks too frequently. Signed-off-by: Tom Herbert <therbert@google.com> include/linux/netdevice.h | 32 ++++- include/linux/skbuff.h | 3 + net/core/dev.c | 335 +++++++++++++++++++++++++++++++++++++-------- net/core/net-sysfs.c | 225 ++++++++++++++++++++++++++++++- net/core/skbuff.c | 2 + 5 files changed, 538 insertions(+), 59 deletions(-) Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2010-03-16 08:03:29 +00:00
get_random_bytes(&hashrnd, sizeof(hashrnd));
return 0;
}
late_initcall_sync(initialize_hashrnd);