The RDMA CM currently infers the QP type from the port space selected
by the user. In the future (eg with RDMA_PS_IB or XRC), there may not
be a 1-1 correspondence between port space and QP type. For netlink
export of RDMA CM state, we want to export the QP type to userspace,
so it is cleaner to explicitly associate a QP type to an ID.
Modify rdma_create_id() to allow the user to specify the QP type, and
use it to make our selections of datagram versus connected mode.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
We want udev to create a device node under /dev/infiniband with
permission 0666 for rdma_cm, so add that info to our struct miscdevice.
Signed-off-by: Roland Dreier <roland@purestorage.com>
Acked-by: Sean Hefty <sean.hefty@intel.com>
Lustre requires that clients bind to a privileged port number before
connecting to a remote server. On larger clusters (typically more
than about 1000 nodes), the number of privileged ports is exhausted,
resulting in lustre being unusable.
To handle this, we add support for reusable addresses to the rdma_cm.
This mimics the behavior of the socket option SO_REUSEADDR. A user
may set an rdma_cm_id to reuse an address before calling
rdma_bind_addr() (explicitly or implicitly). If set, other
rdma_cm_id's may be bound to the same address, provided that they all
have reuse enabled, and there are no active listens.
If rdma_listen() is called on an rdma_cm_id that has reuse enabled, it
will only succeed if there are no other id's bound to that same
address. The reuse option is exported to user space. The behavior of
the kernel reuse implementation was verified against that given by
sockets.
This patch is derived from a path by Ira Weiny <weiny2@llnl.gov>
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
For iWARP rdma_cm ids, the "route" information is the L2 src and
next hop addresses.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
Add 802.1q VLAN support to IBoE. The VLAN tag is encoded within the
GID derived from a link local address in the following way:
GID[11] GID[12] contain the VLAN ID when the GID contains a VLAN.
The 3 bits user priority field of the packets are identical to the 3
bits of the SL.
In case of rdma_cm apps, the TOS field is used to generate the SL
field by doing a shift right of 5 bits effectively taking to 3 MS bits
of the TOS field.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
For iWARP connections, the connect request is carried in a TCP payload
on an already established TCP connection. So if the ucma's backlog is
full, the connection request is transmitted and acked at the TCP level
by the time the connect request gets dropped in the ucma. The end
result is the connection gets rejected by the iWARP provider.
Further, a 32 node 256NP OpenMPI job will generate > 128 connect
requests on some ranks.
This patch increases the default max backlog to 1024, and adds a
sysctl variable so the backlog can be adjusted at run time.
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add support for IBoE device binding and IP --> GID resolution. Path
resolving and multicast joining are implemented within cma.c by
filling in the responses and running callbacks in the CMA work queue.
IP --> GID resolution always yields IPv6 link local addresses; remote
GIDs are derived from the destination MAC address of the remote port.
Multicast GIDs are always mapped to multicast MACs as is done in IPv6.
(IPv4 multicast is enabled by translating IPv4 multicast addresses to
IPv6 multicast as described in
<http://www.mail-archive.com/ipng@sunroof.eng.sun.com/msg02134.html>.)
Some helper functions are added to ib_addr.h.
Signed-off-by: Eli Cohen <eli@mellanox.co.il>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Several RDMA user-access drivers have file_operations structures with
no .llseek method set. None of the drivers actually do anything with
f_pos, so this means llseek is essentially a NOP, instead of returning
an error as leaving other file_operations methods unimplemented would
do. This is mostly harmless, except that a NULL .llseek means that
default_llseek() is used, and this function grabs the BKL, which we
would like to avoid.
Since llseek does nothing useful on these files, we would like it to
return an error to userspace instead of silently grabbing the BKL and
succeeding. For nearly all of the file types, we take the
belt-and-suspenders approach of setting the .llseek method to
no_llseek and also calling nonseekable_open(); the exception is the
uverbs_event files, which are created with anon_inode_getfile(), which
already sets f_mode the same way as nonseekable_open() would.
This work is motivated by Arnd Bergmann's bkl-removal tree.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
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>
The RDMA CM is intended to support the use of a loopback address
when establishing a connection; however, the behavior of the CM
when loopback addresses are used is confusing and does not always
work, depending on whether loopback was specified by the server,
the client, or both.
The defined behavior of rdma_bind_addr is to associate an RDMA
device with an rdma_cm_id, as long as the user specified a non-
zero address. (ie they weren't just trying to reserve a port)
Currently, if the loopback address is passed to rdam_bind_addr,
no device is associated with the rdma_cm_id. Fix this.
If a loopback address is specified by the client as the destination
address for a connection, it will fail to establish a connection.
This is true even if the server is listing across all addresses or
on the loopback address itself. The issue is that the server tries
to translate the IP address carried in the REQ message to a local
net_device address, which fails. The translation is not needed in
this case, since the REQ carries the actual HW address that should
be used.
Finally, cleanup loopback support to be more transport neutral.
Replace separate calls to get/set the sgid and dgid from the
device address to a single call that behaves correctly depending
on the format of the device address. And support both IPv4 and
IPv6 address formats.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
[ Fixed RDS build by s/ib_addr_get/rdma_addr_get/ - Roland ]
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Export rdma_set_ib_paths to user space to allow applications to
manually set the IB path used for connections. This allows
alternative ways for a user space application or library to obtain
path record information, including retrieving path information
from cached data, avoiding direct interaction with the IB SA.
The IB SA is a single, centralized entity that can limit scaling
on large clusters running MPI applications.
Future changes to the rdma cm can expand on this framework to
support the full range of features allowed by the IB CM, such as
separate forward and reverse paths and APM.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Reviewed-By: Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
After m68k's task_thread_info() doesn't refer to current,
it's possible to remove sched.h from interrupt.h and not break m68k!
Many thanks to Heiko Carstens for allowing this.
Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com>
In case of error, the function ucma_alloc_multicast() returns a NULL
pointer, but never returns an ERR pointer. So after a call to this
function, an IS_ERR test should be replaced by a NULL test.
The semantic match that finds this problem is as follows:
(http://www.emn.fr/x-info/coccinelle/)
// <smpl>
@match bad_is_err_test@
expression x, E;
@@
x = ucma_alloc_multicast(...)
... when != x = E
IS_ERR(x)
// </smpl>
Signed-off-by: Julien Brunel <brunel@diku.dk>
Signed-off-by: Julia Lawall <julia@diku.dk>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
There are a few places where the RDMA CM code handles IPv6 by doing
struct sockaddr addr;
u8 pad[sizeof(struct sockaddr_in6) -
sizeof(struct sockaddr)];
This is fragile and ugly; handle this in a better way with just
struct sockaddr_storage addr;
[ Also roll in patch from Aleksey Senin <alekseys@voltaire.com> to
switch to struct sockaddr_storage and get rid of padding arrays in
struct rdma_addr. ]
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Add __force cast of node_guid to __u64, since we are sticking it into a
structure whose definition is shared with userspace.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
This is based on user feedback from Doug Ledford at RedHat:
Events that occur on an rdma_cm_id are reported to userspace through an
event channel. Connection request events are reported on the event
channel associated with the listen. When the connection is accepted, a
new rdma_cm_id is created and automatically uses the listen event
channel. This is suboptimal where the user only wants listen events on
that channel.
Additionally, it may be desirable to have events related to connection
establishment use a different event channel than those related to
already established connections.
Allow the user to migrate an rdma_cm_id between event channels. All
pending events associated with the rdma_cm_id are moved to the new event
channel.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Export the ability to set the type of service to user space. Model
the interface after setsockopt.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Change the returned error code to ENOMEM if the connection event
backlog is full. This prevents the ib_cm from issuing a reject
on the connection, which can allow retries to succeed.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Extend rdma_cm to support multicast communication. Multicast support
is added to the existing RDMA_PS_UDP port space, as well as a new
RDMA_PS_IPOIB port space. The latter port space allows joining the
multicast groups used by IPoIB, which enables offloading IPoIB traffic
to a separate QP. The port space determines the signature used in the
MGID when joining the group. The newly added RDMA_PS_IPOIB also
allows for unicast operations, similar to RDMA_PS_UDP.
Supporting the RDMA_PS_IPOIB requires changing how UD QPs are initialized,
since we can no longer assume that the qkey is constant. This requires
saving the Q_Key to use when attaching to a device, so that it is
available when creating the QP. The Q_Key information is exported to
the user through the existing rdma_init_qp_attr() interface.
Multicast support is also exported to userspace through the rdma_ucm.
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Many struct file_operations in the kernel can be "const". Marking them const
moves these to the .rodata section, which avoids false sharing with potential
dirty data. In addition it'll catch accidental writes at compile time to
these shared resources.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There's a problem with how rdma cm events are reported to userspace
that can lead to application crashes.
When a new connection request arrives, a context for the connection is
allocated in the kernel. The connection event is then reported to
userspace. The userspace library retrieves the event and allocates
its own context for the connection. The userspace context is
associated with the kernel's context when accepting. This allows the
kernel to give userspace context with other events.
A problem occurs if a second event for the same connection occurs
before the user has had a chance to call accept. The userspace
context has not yet been set, which causes the librdmacm to crash.
(This has been seen when the app takes too long to call accept,
resulting in the remote side timing out and rejecting the connection)
Fix this by ignoring events for new connections until userspace has
set their context. This can only happen if an error occurs on a new
connection before the user accepts it. This is okay, since the accept
will just fail later.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
We discard new connection requests while the listen backlog is full,
but leak a struct ucma_event in the process. Free the structure in
this case.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Export the rdma cm interfaces to userspace via a misc device.
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Roland Dreier