linux/drivers/net/s2io.c

7749 lines
218 KiB
C
Raw Normal View History

/************************************************************************
* s2io.c: A Linux PCI-X Ethernet driver for Neterion 10GbE Server NIC
* Copyright(c) 2002-2007 Neterion Inc.
* This software may be used and distributed according to the terms of
* the GNU General Public License (GPL), incorporated herein by reference.
* Drivers based on or derived from this code fall under the GPL and must
* retain the authorship, copyright and license notice. This file is not
* a complete program and may only be used when the entire operating
* system is licensed under the GPL.
* See the file COPYING in this distribution for more information.
*
* Credits:
* Jeff Garzik : For pointing out the improper error condition
* check in the s2io_xmit routine and also some
* issues in the Tx watch dog function. Also for
* patiently answering all those innumerable
* questions regaring the 2.6 porting issues.
* Stephen Hemminger : Providing proper 2.6 porting mechanism for some
* macros available only in 2.6 Kernel.
* Francois Romieu : For pointing out all code part that were
* deprecated and also styling related comments.
* Grant Grundler : For helping me get rid of some Architecture
* dependent code.
* Christopher Hellwig : Some more 2.6 specific issues in the driver.
*
* The module loadable parameters that are supported by the driver and a brief
* explaination of all the variables.
*
* rx_ring_num : This can be used to program the number of receive rings used
* in the driver.
* rx_ring_sz: This defines the number of receive blocks each ring can have.
* This is also an array of size 8.
* rx_ring_mode: This defines the operation mode of all 8 rings. The valid
* values are 1, 2 and 3.
* tx_fifo_num: This defines the number of Tx FIFOs thats used int the driver.
* tx_fifo_len: This too is an array of 8. Each element defines the number of
* Tx descriptors that can be associated with each corresponding FIFO.
* intr_type: This defines the type of interrupt. The values can be 0(INTA),
* 1(MSI), 2(MSI_X). Default value is '0(INTA)'
* lro: Specifies whether to enable Large Receive Offload (LRO) or not.
* Possible values '1' for enable '0' for disable. Default is '0'
* lro_max_pkts: This parameter defines maximum number of packets can be
* aggregated as a single large packet
* napi: This parameter used to enable/disable NAPI (polling Rx)
* Possible values '1' for enable and '0' for disable. Default is '1'
* ufo: This parameter used to enable/disable UDP Fragmentation Offload(UFO)
* Possible values '1' for enable and '0' for disable. Default is '0'
* vlan_tag_strip: This can be used to enable or disable vlan stripping.
* Possible values '1' for enable , '0' for disable.
* Default is '2' - which means disable in promisc mode
* and enable in non-promiscuous mode.
************************************************************************/
#include <linux/module.h>
#include <linux/types.h>
#include <linux/errno.h>
#include <linux/ioport.h>
#include <linux/pci.h>
#include <linux/dma-mapping.h>
#include <linux/kernel.h>
#include <linux/netdevice.h>
#include <linux/etherdevice.h>
#include <linux/skbuff.h>
#include <linux/init.h>
#include <linux/delay.h>
#include <linux/stddef.h>
#include <linux/ioctl.h>
#include <linux/timex.h>
#include <linux/ethtool.h>
#include <linux/workqueue.h>
#include <linux/if_vlan.h>
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
#include <linux/ip.h>
#include <linux/tcp.h>
#include <net/tcp.h>
#include <asm/system.h>
#include <asm/uaccess.h>
#include <asm/io.h>
#include <asm/div64.h>
#include <asm/irq.h>
/* local include */
#include "s2io.h"
#include "s2io-regs.h"
#define DRV_VERSION "2.0.22.1"
/* S2io Driver name & version. */
static char s2io_driver_name[] = "Neterion";
static char s2io_driver_version[] = DRV_VERSION;
static int rxd_size[4] = {32,48,48,64};
static int rxd_count[4] = {127,85,85,63};
static inline int RXD_IS_UP2DT(struct RxD_t *rxdp)
{
int ret;
ret = ((!(rxdp->Control_1 & RXD_OWN_XENA)) &&
(GET_RXD_MARKER(rxdp->Control_2) != THE_RXD_MARK));
return ret;
}
/*
* Cards with following subsystem_id have a link state indication
* problem, 600B, 600C, 600D, 640B, 640C and 640D.
* macro below identifies these cards given the subsystem_id.
*/
#define CARDS_WITH_FAULTY_LINK_INDICATORS(dev_type, subid) \
(dev_type == XFRAME_I_DEVICE) ? \
((((subid >= 0x600B) && (subid <= 0x600D)) || \
((subid >= 0x640B) && (subid <= 0x640D))) ? 1 : 0) : 0
#define LINK_IS_UP(val64) (!(val64 & (ADAPTER_STATUS_RMAC_REMOTE_FAULT | \
ADAPTER_STATUS_RMAC_LOCAL_FAULT)))
#define TASKLET_IN_USE test_and_set_bit(0, (&sp->tasklet_status))
#define PANIC 1
#define LOW 2
static inline int rx_buffer_level(struct s2io_nic * sp, int rxb_size, int ring)
{
struct mac_info *mac_control;
mac_control = &sp->mac_control;
if (rxb_size <= rxd_count[sp->rxd_mode])
return PANIC;
else if ((mac_control->rings[ring].pkt_cnt - rxb_size) > 16)
return LOW;
return 0;
}
/* Ethtool related variables and Macros. */
static char s2io_gstrings[][ETH_GSTRING_LEN] = {
"Register test\t(offline)",
"Eeprom test\t(offline)",
"Link test\t(online)",
"RLDRAM test\t(offline)",
"BIST Test\t(offline)"
};
static char ethtool_xena_stats_keys[][ETH_GSTRING_LEN] = {
{"tmac_frms"},
{"tmac_data_octets"},
{"tmac_drop_frms"},
{"tmac_mcst_frms"},
{"tmac_bcst_frms"},
{"tmac_pause_ctrl_frms"},
{"tmac_ttl_octets"},
{"tmac_ucst_frms"},
{"tmac_nucst_frms"},
{"tmac_any_err_frms"},
{"tmac_ttl_less_fb_octets"},
{"tmac_vld_ip_octets"},
{"tmac_vld_ip"},
{"tmac_drop_ip"},
{"tmac_icmp"},
{"tmac_rst_tcp"},
{"tmac_tcp"},
{"tmac_udp"},
{"rmac_vld_frms"},
{"rmac_data_octets"},
{"rmac_fcs_err_frms"},
{"rmac_drop_frms"},
{"rmac_vld_mcst_frms"},
{"rmac_vld_bcst_frms"},
{"rmac_in_rng_len_err_frms"},
{"rmac_out_rng_len_err_frms"},
{"rmac_long_frms"},
{"rmac_pause_ctrl_frms"},
{"rmac_unsup_ctrl_frms"},
{"rmac_ttl_octets"},
{"rmac_accepted_ucst_frms"},
{"rmac_accepted_nucst_frms"},
{"rmac_discarded_frms"},
{"rmac_drop_events"},
{"rmac_ttl_less_fb_octets"},
{"rmac_ttl_frms"},
{"rmac_usized_frms"},
{"rmac_osized_frms"},
{"rmac_frag_frms"},
{"rmac_jabber_frms"},
{"rmac_ttl_64_frms"},
{"rmac_ttl_65_127_frms"},
{"rmac_ttl_128_255_frms"},
{"rmac_ttl_256_511_frms"},
{"rmac_ttl_512_1023_frms"},
{"rmac_ttl_1024_1518_frms"},
{"rmac_ip"},
{"rmac_ip_octets"},
{"rmac_hdr_err_ip"},
{"rmac_drop_ip"},
{"rmac_icmp"},
{"rmac_tcp"},
{"rmac_udp"},
{"rmac_err_drp_udp"},
{"rmac_xgmii_err_sym"},
{"rmac_frms_q0"},
{"rmac_frms_q1"},
{"rmac_frms_q2"},
{"rmac_frms_q3"},
{"rmac_frms_q4"},
{"rmac_frms_q5"},
{"rmac_frms_q6"},
{"rmac_frms_q7"},
{"rmac_full_q0"},
{"rmac_full_q1"},
{"rmac_full_q2"},
{"rmac_full_q3"},
{"rmac_full_q4"},
{"rmac_full_q5"},
{"rmac_full_q6"},
{"rmac_full_q7"},
{"rmac_pause_cnt"},
{"rmac_xgmii_data_err_cnt"},
{"rmac_xgmii_ctrl_err_cnt"},
{"rmac_accepted_ip"},
{"rmac_err_tcp"},
{"rd_req_cnt"},
{"new_rd_req_cnt"},
{"new_rd_req_rtry_cnt"},
{"rd_rtry_cnt"},
{"wr_rtry_rd_ack_cnt"},
{"wr_req_cnt"},
{"new_wr_req_cnt"},
{"new_wr_req_rtry_cnt"},
{"wr_rtry_cnt"},
{"wr_disc_cnt"},
{"rd_rtry_wr_ack_cnt"},
{"txp_wr_cnt"},
{"txd_rd_cnt"},
{"txd_wr_cnt"},
{"rxd_rd_cnt"},
{"rxd_wr_cnt"},
{"txf_rd_cnt"},
{"rxf_wr_cnt"}
};
static char ethtool_enhanced_stats_keys[][ETH_GSTRING_LEN] = {
{"rmac_ttl_1519_4095_frms"},
{"rmac_ttl_4096_8191_frms"},
{"rmac_ttl_8192_max_frms"},
{"rmac_ttl_gt_max_frms"},
{"rmac_osized_alt_frms"},
{"rmac_jabber_alt_frms"},
{"rmac_gt_max_alt_frms"},
{"rmac_vlan_frms"},
{"rmac_len_discard"},
{"rmac_fcs_discard"},
{"rmac_pf_discard"},
{"rmac_da_discard"},
{"rmac_red_discard"},
{"rmac_rts_discard"},
{"rmac_ingm_full_discard"},
{"link_fault_cnt"}
};
static char ethtool_driver_stats_keys[][ETH_GSTRING_LEN] = {
{"\n DRIVER STATISTICS"},
{"single_bit_ecc_errs"},
{"double_bit_ecc_errs"},
{"parity_err_cnt"},
{"serious_err_cnt"},
{"soft_reset_cnt"},
{"fifo_full_cnt"},
{"ring_full_cnt"},
("alarm_transceiver_temp_high"),
("alarm_transceiver_temp_low"),
("alarm_laser_bias_current_high"),
("alarm_laser_bias_current_low"),
("alarm_laser_output_power_high"),
("alarm_laser_output_power_low"),
("warn_transceiver_temp_high"),
("warn_transceiver_temp_low"),
("warn_laser_bias_current_high"),
("warn_laser_bias_current_low"),
("warn_laser_output_power_high"),
("warn_laser_output_power_low"),
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
("lro_aggregated_pkts"),
("lro_flush_both_count"),
("lro_out_of_sequence_pkts"),
("lro_flush_due_to_max_pkts"),
("lro_avg_aggr_pkts"),
};
#define S2IO_XENA_STAT_LEN sizeof(ethtool_xena_stats_keys)/ ETH_GSTRING_LEN
#define S2IO_ENHANCED_STAT_LEN sizeof(ethtool_enhanced_stats_keys)/ \
ETH_GSTRING_LEN
#define S2IO_DRIVER_STAT_LEN sizeof(ethtool_driver_stats_keys)/ ETH_GSTRING_LEN
#define XFRAME_I_STAT_LEN (S2IO_XENA_STAT_LEN + S2IO_DRIVER_STAT_LEN )
#define XFRAME_II_STAT_LEN (XFRAME_I_STAT_LEN + S2IO_ENHANCED_STAT_LEN )
#define XFRAME_I_STAT_STRINGS_LEN ( XFRAME_I_STAT_LEN * ETH_GSTRING_LEN )
#define XFRAME_II_STAT_STRINGS_LEN ( XFRAME_II_STAT_LEN * ETH_GSTRING_LEN )
#define S2IO_TEST_LEN sizeof(s2io_gstrings) / ETH_GSTRING_LEN
#define S2IO_STRINGS_LEN S2IO_TEST_LEN * ETH_GSTRING_LEN
#define S2IO_TIMER_CONF(timer, handle, arg, exp) \
init_timer(&timer); \
timer.function = handle; \
timer.data = (unsigned long) arg; \
mod_timer(&timer, (jiffies + exp)) \
/* Add the vlan */
static void s2io_vlan_rx_register(struct net_device *dev,
struct vlan_group *grp)
{
struct s2io_nic *nic = dev->priv;
unsigned long flags;
spin_lock_irqsave(&nic->tx_lock, flags);
nic->vlgrp = grp;
spin_unlock_irqrestore(&nic->tx_lock, flags);
}
/* A flag indicating whether 'RX_PA_CFG_STRIP_VLAN_TAG' bit is set or not */
static int vlan_strip_flag;
/* Unregister the vlan */
static void s2io_vlan_rx_kill_vid(struct net_device *dev, unsigned long vid)
{
struct s2io_nic *nic = dev->priv;
unsigned long flags;
spin_lock_irqsave(&nic->tx_lock, flags);
vlan_group_set_device(nic->vlgrp, vid, NULL);
spin_unlock_irqrestore(&nic->tx_lock, flags);
}
/*
* Constants to be programmed into the Xena's registers, to configure
* the XAUI.
*/
#define END_SIGN 0x0
2006-03-04 02:33:57 +00:00
static const u64 herc_act_dtx_cfg[] = {
/* Set address */
0x8000051536750000ULL, 0x80000515367500E0ULL,
/* Write data */
0x8000051536750004ULL, 0x80000515367500E4ULL,
/* Set address */
0x80010515003F0000ULL, 0x80010515003F00E0ULL,
/* Write data */
0x80010515003F0004ULL, 0x80010515003F00E4ULL,
/* Set address */
0x801205150D440000ULL, 0x801205150D4400E0ULL,
/* Write data */
0x801205150D440004ULL, 0x801205150D4400E4ULL,
/* Set address */
0x80020515F2100000ULL, 0x80020515F21000E0ULL,
/* Write data */
0x80020515F2100004ULL, 0x80020515F21000E4ULL,
/* Done */
END_SIGN
};
2006-03-04 02:33:57 +00:00
static const u64 xena_dtx_cfg[] = {
/* Set address */
0x8000051500000000ULL, 0x80000515000000E0ULL,
/* Write data */
0x80000515D9350004ULL, 0x80000515D93500E4ULL,
/* Set address */
0x8001051500000000ULL, 0x80010515000000E0ULL,
/* Write data */
0x80010515001E0004ULL, 0x80010515001E00E4ULL,
/* Set address */
0x8002051500000000ULL, 0x80020515000000E0ULL,
/* Write data */
0x80020515F2100004ULL, 0x80020515F21000E4ULL,
END_SIGN
};
/*
* Constants for Fixing the MacAddress problem seen mostly on
* Alpha machines.
*/
2006-03-04 02:33:57 +00:00
static const u64 fix_mac[] = {
0x0060000000000000ULL, 0x0060600000000000ULL,
0x0040600000000000ULL, 0x0000600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0060600000000000ULL,
0x0020600000000000ULL, 0x0000600000000000ULL,
0x0040600000000000ULL, 0x0060600000000000ULL,
END_SIGN
};
MODULE_LICENSE("GPL");
MODULE_VERSION(DRV_VERSION);
/* Module Loadable parameters. */
S2IO_PARM_INT(tx_fifo_num, 1);
S2IO_PARM_INT(rx_ring_num, 1);
S2IO_PARM_INT(rx_ring_mode, 1);
S2IO_PARM_INT(use_continuous_tx_intrs, 1);
S2IO_PARM_INT(rmac_pause_time, 0x100);
S2IO_PARM_INT(mc_pause_threshold_q0q3, 187);
S2IO_PARM_INT(mc_pause_threshold_q4q7, 187);
S2IO_PARM_INT(shared_splits, 0);
S2IO_PARM_INT(tmac_util_period, 5);
S2IO_PARM_INT(rmac_util_period, 5);
S2IO_PARM_INT(bimodal, 0);
S2IO_PARM_INT(l3l4hdr_size, 128);
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
/* Frequency of Rx desc syncs expressed as power of 2 */
S2IO_PARM_INT(rxsync_frequency, 3);
/* Interrupt type. Values can be 0(INTA), 1(MSI), 2(MSI_X) */
S2IO_PARM_INT(intr_type, 0);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
/* Large receive offload feature */
S2IO_PARM_INT(lro, 0);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
/* Max pkts to be aggregated by LRO at one time. If not specified,
* aggregation happens until we hit max IP pkt size(64K)
*/
S2IO_PARM_INT(lro_max_pkts, 0xFFFF);
S2IO_PARM_INT(indicate_max_pkts, 0);
S2IO_PARM_INT(napi, 1);
S2IO_PARM_INT(ufo, 0);
S2IO_PARM_INT(vlan_tag_strip, NO_STRIP_IN_PROMISC);
static unsigned int tx_fifo_len[MAX_TX_FIFOS] =
{DEFAULT_FIFO_0_LEN, [1 ...(MAX_TX_FIFOS - 1)] = DEFAULT_FIFO_1_7_LEN};
static unsigned int rx_ring_sz[MAX_RX_RINGS] =
{[0 ...(MAX_RX_RINGS - 1)] = SMALL_BLK_CNT};
static unsigned int rts_frm_len[MAX_RX_RINGS] =
{[0 ...(MAX_RX_RINGS - 1)] = 0 };
module_param_array(tx_fifo_len, uint, NULL, 0);
module_param_array(rx_ring_sz, uint, NULL, 0);
module_param_array(rts_frm_len, uint, NULL, 0);
/*
* S2IO device table.
* This table lists all the devices that this driver supports.
*/
static struct pci_device_id s2io_tbl[] __devinitdata = {
{PCI_VENDOR_ID_S2IO, PCI_DEVICE_ID_S2IO_WIN,
PCI_ANY_ID, PCI_ANY_ID},
{PCI_VENDOR_ID_S2IO, PCI_DEVICE_ID_S2IO_UNI,
PCI_ANY_ID, PCI_ANY_ID},
{PCI_VENDOR_ID_S2IO, PCI_DEVICE_ID_HERC_WIN,
PCI_ANY_ID, PCI_ANY_ID},
{PCI_VENDOR_ID_S2IO, PCI_DEVICE_ID_HERC_UNI,
PCI_ANY_ID, PCI_ANY_ID},
{0,}
};
MODULE_DEVICE_TABLE(pci, s2io_tbl);
static struct pci_driver s2io_driver = {
.name = "S2IO",
.id_table = s2io_tbl,
.probe = s2io_init_nic,
.remove = __devexit_p(s2io_rem_nic),
};
/* A simplifier macro used both by init and free shared_mem Fns(). */
#define TXD_MEM_PAGE_CNT(len, per_each) ((len+per_each - 1) / per_each)
/**
* init_shared_mem - Allocation and Initialization of Memory
* @nic: Device private variable.
* Description: The function allocates all the memory areas shared
* between the NIC and the driver. This includes Tx descriptors,
* Rx descriptors and the statistics block.
*/
static int init_shared_mem(struct s2io_nic *nic)
{
u32 size;
void *tmp_v_addr, *tmp_v_addr_next;
dma_addr_t tmp_p_addr, tmp_p_addr_next;
struct RxD_block *pre_rxd_blk = NULL;
int i, j, blk_cnt;
int lst_size, lst_per_page;
struct net_device *dev = nic->dev;
unsigned long tmp;
struct buffAdd *ba;
struct mac_info *mac_control;
struct config_param *config;
mac_control = &nic->mac_control;
config = &nic->config;
/* Allocation and initialization of TXDLs in FIOFs */
size = 0;
for (i = 0; i < config->tx_fifo_num; i++) {
size += config->tx_cfg[i].fifo_len;
}
if (size > MAX_AVAILABLE_TXDS) {
DBG_PRINT(ERR_DBG, "s2io: Requested TxDs too high, ");
DBG_PRINT(ERR_DBG, "Requested: %d, max supported: 8192\n", size);
return -EINVAL;
}
lst_size = (sizeof(struct TxD) * config->max_txds);
lst_per_page = PAGE_SIZE / lst_size;
for (i = 0; i < config->tx_fifo_num; i++) {
int fifo_len = config->tx_cfg[i].fifo_len;
int list_holder_size = fifo_len * sizeof(struct list_info_hold);
mac_control->fifos[i].list_info = kmalloc(list_holder_size,
GFP_KERNEL);
if (!mac_control->fifos[i].list_info) {
DBG_PRINT(INFO_DBG,
"Malloc failed for list_info\n");
return -ENOMEM;
}
memset(mac_control->fifos[i].list_info, 0, list_holder_size);
}
for (i = 0; i < config->tx_fifo_num; i++) {
int page_num = TXD_MEM_PAGE_CNT(config->tx_cfg[i].fifo_len,
lst_per_page);
mac_control->fifos[i].tx_curr_put_info.offset = 0;
mac_control->fifos[i].tx_curr_put_info.fifo_len =
config->tx_cfg[i].fifo_len - 1;
mac_control->fifos[i].tx_curr_get_info.offset = 0;
mac_control->fifos[i].tx_curr_get_info.fifo_len =
config->tx_cfg[i].fifo_len - 1;
mac_control->fifos[i].fifo_no = i;
mac_control->fifos[i].nic = nic;
mac_control->fifos[i].max_txds = MAX_SKB_FRAGS + 2;
for (j = 0; j < page_num; j++) {
int k = 0;
dma_addr_t tmp_p;
void *tmp_v;
tmp_v = pci_alloc_consistent(nic->pdev,
PAGE_SIZE, &tmp_p);
if (!tmp_v) {
DBG_PRINT(INFO_DBG,
"pci_alloc_consistent ");
DBG_PRINT(INFO_DBG, "failed for TxDL\n");
return -ENOMEM;
}
/* If we got a zero DMA address(can happen on
* certain platforms like PPC), reallocate.
* Store virtual address of page we don't want,
* to be freed later.
*/
if (!tmp_p) {
mac_control->zerodma_virt_addr = tmp_v;
DBG_PRINT(INIT_DBG,
"%s: Zero DMA address for TxDL. ", dev->name);
DBG_PRINT(INIT_DBG,
"Virtual address %p\n", tmp_v);
tmp_v = pci_alloc_consistent(nic->pdev,
PAGE_SIZE, &tmp_p);
if (!tmp_v) {
DBG_PRINT(INFO_DBG,
"pci_alloc_consistent ");
DBG_PRINT(INFO_DBG, "failed for TxDL\n");
return -ENOMEM;
}
}
while (k < lst_per_page) {
int l = (j * lst_per_page) + k;
if (l == config->tx_cfg[i].fifo_len)
break;
mac_control->fifos[i].list_info[l].list_virt_addr =
tmp_v + (k * lst_size);
mac_control->fifos[i].list_info[l].list_phy_addr =
tmp_p + (k * lst_size);
k++;
}
}
}
nic->ufo_in_band_v = kcalloc(size, sizeof(u64), GFP_KERNEL);
if (!nic->ufo_in_band_v)
return -ENOMEM;
/* Allocation and initialization of RXDs in Rings */
size = 0;
for (i = 0; i < config->rx_ring_num; i++) {
if (config->rx_cfg[i].num_rxd %
(rxd_count[nic->rxd_mode] + 1)) {
DBG_PRINT(ERR_DBG, "%s: RxD count of ", dev->name);
DBG_PRINT(ERR_DBG, "Ring%d is not a multiple of ",
i);
DBG_PRINT(ERR_DBG, "RxDs per Block");
return FAILURE;
}
size += config->rx_cfg[i].num_rxd;
mac_control->rings[i].block_count =
config->rx_cfg[i].num_rxd /
(rxd_count[nic->rxd_mode] + 1 );
mac_control->rings[i].pkt_cnt = config->rx_cfg[i].num_rxd -
mac_control->rings[i].block_count;
}
if (nic->rxd_mode == RXD_MODE_1)
size = (size * (sizeof(struct RxD1)));
else
size = (size * (sizeof(struct RxD3)));
for (i = 0; i < config->rx_ring_num; i++) {
mac_control->rings[i].rx_curr_get_info.block_index = 0;
mac_control->rings[i].rx_curr_get_info.offset = 0;
mac_control->rings[i].rx_curr_get_info.ring_len =
config->rx_cfg[i].num_rxd - 1;
mac_control->rings[i].rx_curr_put_info.block_index = 0;
mac_control->rings[i].rx_curr_put_info.offset = 0;
mac_control->rings[i].rx_curr_put_info.ring_len =
config->rx_cfg[i].num_rxd - 1;
mac_control->rings[i].nic = nic;
mac_control->rings[i].ring_no = i;
blk_cnt = config->rx_cfg[i].num_rxd /
(rxd_count[nic->rxd_mode] + 1);
/* Allocating all the Rx blocks */
for (j = 0; j < blk_cnt; j++) {
struct rx_block_info *rx_blocks;
int l;
rx_blocks = &mac_control->rings[i].rx_blocks[j];
size = SIZE_OF_BLOCK; //size is always page size
tmp_v_addr = pci_alloc_consistent(nic->pdev, size,
&tmp_p_addr);
if (tmp_v_addr == NULL) {
/*
* In case of failure, free_shared_mem()
* is called, which should free any
* memory that was alloced till the
* failure happened.
*/
rx_blocks->block_virt_addr = tmp_v_addr;
return -ENOMEM;
}
memset(tmp_v_addr, 0, size);
rx_blocks->block_virt_addr = tmp_v_addr;
rx_blocks->block_dma_addr = tmp_p_addr;
rx_blocks->rxds = kmalloc(sizeof(struct rxd_info)*
rxd_count[nic->rxd_mode],
GFP_KERNEL);
if (!rx_blocks->rxds)
return -ENOMEM;
for (l=0; l<rxd_count[nic->rxd_mode];l++) {
rx_blocks->rxds[l].virt_addr =
rx_blocks->block_virt_addr +
(rxd_size[nic->rxd_mode] * l);
rx_blocks->rxds[l].dma_addr =
rx_blocks->block_dma_addr +
(rxd_size[nic->rxd_mode] * l);
}
}
/* Interlinking all Rx Blocks */
for (j = 0; j < blk_cnt; j++) {
tmp_v_addr =
mac_control->rings[i].rx_blocks[j].block_virt_addr;
tmp_v_addr_next =
mac_control->rings[i].rx_blocks[(j + 1) %
blk_cnt].block_virt_addr;
tmp_p_addr =
mac_control->rings[i].rx_blocks[j].block_dma_addr;
tmp_p_addr_next =
mac_control->rings[i].rx_blocks[(j + 1) %
blk_cnt].block_dma_addr;
pre_rxd_blk = (struct RxD_block *) tmp_v_addr;
pre_rxd_blk->reserved_2_pNext_RxD_block =
(unsigned long) tmp_v_addr_next;
pre_rxd_blk->pNext_RxD_Blk_physical =
(u64) tmp_p_addr_next;
}
}
if (nic->rxd_mode >= RXD_MODE_3A) {
/*
* Allocation of Storages for buffer addresses in 2BUFF mode
* and the buffers as well.
*/
for (i = 0; i < config->rx_ring_num; i++) {
blk_cnt = config->rx_cfg[i].num_rxd /
(rxd_count[nic->rxd_mode]+ 1);
mac_control->rings[i].ba =
kmalloc((sizeof(struct buffAdd *) * blk_cnt),
GFP_KERNEL);
if (!mac_control->rings[i].ba)
return -ENOMEM;
for (j = 0; j < blk_cnt; j++) {
int k = 0;
mac_control->rings[i].ba[j] =
kmalloc((sizeof(struct buffAdd) *
(rxd_count[nic->rxd_mode] + 1)),
GFP_KERNEL);
if (!mac_control->rings[i].ba[j])
return -ENOMEM;
while (k != rxd_count[nic->rxd_mode]) {
ba = &mac_control->rings[i].ba[j][k];
ba->ba_0_org = (void *) kmalloc
(BUF0_LEN + ALIGN_SIZE, GFP_KERNEL);
if (!ba->ba_0_org)
return -ENOMEM;
tmp = (unsigned long)ba->ba_0_org;
tmp += ALIGN_SIZE;
tmp &= ~((unsigned long) ALIGN_SIZE);
ba->ba_0 = (void *) tmp;
ba->ba_1_org = (void *) kmalloc
(BUF1_LEN + ALIGN_SIZE, GFP_KERNEL);
if (!ba->ba_1_org)
return -ENOMEM;
tmp = (unsigned long) ba->ba_1_org;
tmp += ALIGN_SIZE;
tmp &= ~((unsigned long) ALIGN_SIZE);
ba->ba_1 = (void *) tmp;
k++;
}
}
}
}
/* Allocation and initialization of Statistics block */
size = sizeof(struct stat_block);
mac_control->stats_mem = pci_alloc_consistent
(nic->pdev, size, &mac_control->stats_mem_phy);
if (!mac_control->stats_mem) {
/*
* In case of failure, free_shared_mem() is called, which
* should free any memory that was alloced till the
* failure happened.
*/
return -ENOMEM;
}
mac_control->stats_mem_sz = size;
tmp_v_addr = mac_control->stats_mem;
mac_control->stats_info = (struct stat_block *) tmp_v_addr;
memset(tmp_v_addr, 0, size);
DBG_PRINT(INIT_DBG, "%s:Ring Mem PHY: 0x%llx\n", dev->name,
(unsigned long long) tmp_p_addr);
return SUCCESS;
}
/**
* free_shared_mem - Free the allocated Memory
* @nic: Device private variable.
* Description: This function is to free all memory locations allocated by
* the init_shared_mem() function and return it to the kernel.
*/
static void free_shared_mem(struct s2io_nic *nic)
{
int i, j, blk_cnt, size;
void *tmp_v_addr;
dma_addr_t tmp_p_addr;
struct mac_info *mac_control;
struct config_param *config;
int lst_size, lst_per_page;
struct net_device *dev = nic->dev;
if (!nic)
return;
mac_control = &nic->mac_control;
config = &nic->config;
lst_size = (sizeof(struct TxD) * config->max_txds);
lst_per_page = PAGE_SIZE / lst_size;
for (i = 0; i < config->tx_fifo_num; i++) {
int page_num = TXD_MEM_PAGE_CNT(config->tx_cfg[i].fifo_len,
lst_per_page);
for (j = 0; j < page_num; j++) {
int mem_blks = (j * lst_per_page);
if (!mac_control->fifos[i].list_info)
return;
if (!mac_control->fifos[i].list_info[mem_blks].
list_virt_addr)
break;
pci_free_consistent(nic->pdev, PAGE_SIZE,
mac_control->fifos[i].
list_info[mem_blks].
list_virt_addr,
mac_control->fifos[i].
list_info[mem_blks].
list_phy_addr);
}
/* If we got a zero DMA address during allocation,
* free the page now
*/
if (mac_control->zerodma_virt_addr) {
pci_free_consistent(nic->pdev, PAGE_SIZE,
mac_control->zerodma_virt_addr,
(dma_addr_t)0);
DBG_PRINT(INIT_DBG,
"%s: Freeing TxDL with zero DMA addr. ",
dev->name);
DBG_PRINT(INIT_DBG, "Virtual address %p\n",
mac_control->zerodma_virt_addr);
}
kfree(mac_control->fifos[i].list_info);
}
size = SIZE_OF_BLOCK;
for (i = 0; i < config->rx_ring_num; i++) {
blk_cnt = mac_control->rings[i].block_count;
for (j = 0; j < blk_cnt; j++) {
tmp_v_addr = mac_control->rings[i].rx_blocks[j].
block_virt_addr;
tmp_p_addr = mac_control->rings[i].rx_blocks[j].
block_dma_addr;
if (tmp_v_addr == NULL)
break;
pci_free_consistent(nic->pdev, size,
tmp_v_addr, tmp_p_addr);
kfree(mac_control->rings[i].rx_blocks[j].rxds);
}
}
if (nic->rxd_mode >= RXD_MODE_3A) {
/* Freeing buffer storage addresses in 2BUFF mode. */
for (i = 0; i < config->rx_ring_num; i++) {
blk_cnt = config->rx_cfg[i].num_rxd /
(rxd_count[nic->rxd_mode] + 1);
for (j = 0; j < blk_cnt; j++) {
int k = 0;
if (!mac_control->rings[i].ba[j])
continue;
while (k != rxd_count[nic->rxd_mode]) {
struct buffAdd *ba =
&mac_control->rings[i].ba[j][k];
kfree(ba->ba_0_org);
kfree(ba->ba_1_org);
k++;
}
kfree(mac_control->rings[i].ba[j]);
}
kfree(mac_control->rings[i].ba);
}
}
if (mac_control->stats_mem) {
pci_free_consistent(nic->pdev,
mac_control->stats_mem_sz,
mac_control->stats_mem,
mac_control->stats_mem_phy);
}
if (nic->ufo_in_band_v)
kfree(nic->ufo_in_band_v);
}
/**
* s2io_verify_pci_mode -
*/
static int s2io_verify_pci_mode(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
register u64 val64 = 0;
int mode;
val64 = readq(&bar0->pci_mode);
mode = (u8)GET_PCI_MODE(val64);
if ( val64 & PCI_MODE_UNKNOWN_MODE)
return -1; /* Unknown PCI mode */
return mode;
}
#define NEC_VENID 0x1033
#define NEC_DEVID 0x0125
static int s2io_on_nec_bridge(struct pci_dev *s2io_pdev)
{
struct pci_dev *tdev = NULL;
while ((tdev = pci_get_device(PCI_ANY_ID, PCI_ANY_ID, tdev)) != NULL) {
if (tdev->vendor == NEC_VENID && tdev->device == NEC_DEVID) {
if (tdev->bus == s2io_pdev->bus->parent)
pci_dev_put(tdev);
return 1;
}
}
return 0;
}
static int bus_speed[8] = {33, 133, 133, 200, 266, 133, 200, 266};
/**
* s2io_print_pci_mode -
*/
static int s2io_print_pci_mode(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
register u64 val64 = 0;
int mode;
struct config_param *config = &nic->config;
val64 = readq(&bar0->pci_mode);
mode = (u8)GET_PCI_MODE(val64);
if ( val64 & PCI_MODE_UNKNOWN_MODE)
return -1; /* Unknown PCI mode */
config->bus_speed = bus_speed[mode];
if (s2io_on_nec_bridge(nic->pdev)) {
DBG_PRINT(ERR_DBG, "%s: Device is on PCI-E bus\n",
nic->dev->name);
return mode;
}
if (val64 & PCI_MODE_32_BITS) {
DBG_PRINT(ERR_DBG, "%s: Device is on 32 bit ", nic->dev->name);
} else {
DBG_PRINT(ERR_DBG, "%s: Device is on 64 bit ", nic->dev->name);
}
switch(mode) {
case PCI_MODE_PCI_33:
DBG_PRINT(ERR_DBG, "33MHz PCI bus\n");
break;
case PCI_MODE_PCI_66:
DBG_PRINT(ERR_DBG, "66MHz PCI bus\n");
break;
case PCI_MODE_PCIX_M1_66:
DBG_PRINT(ERR_DBG, "66MHz PCIX(M1) bus\n");
break;
case PCI_MODE_PCIX_M1_100:
DBG_PRINT(ERR_DBG, "100MHz PCIX(M1) bus\n");
break;
case PCI_MODE_PCIX_M1_133:
DBG_PRINT(ERR_DBG, "133MHz PCIX(M1) bus\n");
break;
case PCI_MODE_PCIX_M2_66:
DBG_PRINT(ERR_DBG, "133MHz PCIX(M2) bus\n");
break;
case PCI_MODE_PCIX_M2_100:
DBG_PRINT(ERR_DBG, "200MHz PCIX(M2) bus\n");
break;
case PCI_MODE_PCIX_M2_133:
DBG_PRINT(ERR_DBG, "266MHz PCIX(M2) bus\n");
break;
default:
return -1; /* Unsupported bus speed */
}
return mode;
}
/**
* init_nic - Initialization of hardware
* @nic: device peivate variable
* Description: The function sequentially configures every block
* of the H/W from their reset values.
* Return Value: SUCCESS on success and
* '-1' on failure (endian settings incorrect).
*/
static int init_nic(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
struct net_device *dev = nic->dev;
register u64 val64 = 0;
void __iomem *add;
u32 time;
int i, j;
struct mac_info *mac_control;
struct config_param *config;
int dtx_cnt = 0;
unsigned long long mem_share;
int mem_size;
mac_control = &nic->mac_control;
config = &nic->config;
/* to set the swapper controle on the card */
if(s2io_set_swapper(nic)) {
DBG_PRINT(ERR_DBG,"ERROR: Setting Swapper failed\n");
return -1;
}
/*
* Herc requires EOI to be removed from reset before XGXS, so..
*/
if (nic->device_type & XFRAME_II_DEVICE) {
val64 = 0xA500000000ULL;
writeq(val64, &bar0->sw_reset);
msleep(500);
val64 = readq(&bar0->sw_reset);
}
/* Remove XGXS from reset state */
val64 = 0;
writeq(val64, &bar0->sw_reset);
msleep(500);
val64 = readq(&bar0->sw_reset);
/* Enable Receiving broadcasts */
add = &bar0->mac_cfg;
val64 = readq(&bar0->mac_cfg);
val64 |= MAC_RMAC_BCAST_ENABLE;
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) val64, add);
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) (val64 >> 32), (add + 4));
/* Read registers in all blocks */
val64 = readq(&bar0->mac_int_mask);
val64 = readq(&bar0->mc_int_mask);
val64 = readq(&bar0->xgxs_int_mask);
/* Set MTU */
val64 = dev->mtu;
writeq(vBIT(val64, 2, 14), &bar0->rmac_max_pyld_len);
if (nic->device_type & XFRAME_II_DEVICE) {
while (herc_act_dtx_cfg[dtx_cnt] != END_SIGN) {
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
SPECIAL_REG_WRITE(herc_act_dtx_cfg[dtx_cnt],
&bar0->dtx_control, UF);
if (dtx_cnt & 0x1)
msleep(1); /* Necessary!! */
dtx_cnt++;
}
} else {
while (xena_dtx_cfg[dtx_cnt] != END_SIGN) {
SPECIAL_REG_WRITE(xena_dtx_cfg[dtx_cnt],
&bar0->dtx_control, UF);
val64 = readq(&bar0->dtx_control);
dtx_cnt++;
}
}
/* Tx DMA Initialization */
val64 = 0;
writeq(val64, &bar0->tx_fifo_partition_0);
writeq(val64, &bar0->tx_fifo_partition_1);
writeq(val64, &bar0->tx_fifo_partition_2);
writeq(val64, &bar0->tx_fifo_partition_3);
for (i = 0, j = 0; i < config->tx_fifo_num; i++) {
val64 |=
vBIT(config->tx_cfg[i].fifo_len - 1, ((i * 32) + 19),
13) | vBIT(config->tx_cfg[i].fifo_priority,
((i * 32) + 5), 3);
if (i == (config->tx_fifo_num - 1)) {
if (i % 2 == 0)
i++;
}
switch (i) {
case 1:
writeq(val64, &bar0->tx_fifo_partition_0);
val64 = 0;
break;
case 3:
writeq(val64, &bar0->tx_fifo_partition_1);
val64 = 0;
break;
case 5:
writeq(val64, &bar0->tx_fifo_partition_2);
val64 = 0;
break;
case 7:
writeq(val64, &bar0->tx_fifo_partition_3);
break;
}
}
/*
* Disable 4 PCCs for Xena1, 2 and 3 as per H/W bug
* SXE-008 TRANSMIT DMA ARBITRATION ISSUE.
*/
if ((nic->device_type == XFRAME_I_DEVICE) &&
(get_xena_rev_id(nic->pdev) < 4))
writeq(PCC_ENABLE_FOUR, &bar0->pcc_enable);
val64 = readq(&bar0->tx_fifo_partition_0);
DBG_PRINT(INIT_DBG, "Fifo partition at: 0x%p is: 0x%llx\n",
&bar0->tx_fifo_partition_0, (unsigned long long) val64);
/*
* Initialization of Tx_PA_CONFIG register to ignore packet
* integrity checking.
*/
val64 = readq(&bar0->tx_pa_cfg);
val64 |= TX_PA_CFG_IGNORE_FRM_ERR | TX_PA_CFG_IGNORE_SNAP_OUI |
TX_PA_CFG_IGNORE_LLC_CTRL | TX_PA_CFG_IGNORE_L2_ERR;
writeq(val64, &bar0->tx_pa_cfg);
/* Rx DMA intialization. */
val64 = 0;
for (i = 0; i < config->rx_ring_num; i++) {
val64 |=
vBIT(config->rx_cfg[i].ring_priority, (5 + (i * 8)),
3);
}
writeq(val64, &bar0->rx_queue_priority);
/*
* Allocating equal share of memory to all the
* configured Rings.
*/
val64 = 0;
if (nic->device_type & XFRAME_II_DEVICE)
mem_size = 32;
else
mem_size = 64;
for (i = 0; i < config->rx_ring_num; i++) {
switch (i) {
case 0:
mem_share = (mem_size / config->rx_ring_num +
mem_size % config->rx_ring_num);
val64 |= RX_QUEUE_CFG_Q0_SZ(mem_share);
continue;
case 1:
mem_share = (mem_size / config->rx_ring_num);
val64 |= RX_QUEUE_CFG_Q1_SZ(mem_share);
continue;
case 2:
mem_share = (mem_size / config->rx_ring_num);
val64 |= RX_QUEUE_CFG_Q2_SZ(mem_share);
continue;
case 3:
mem_share = (mem_size / config->rx_ring_num);
val64 |= RX_QUEUE_CFG_Q3_SZ(mem_share);
continue;
case 4:
mem_share = (mem_size / config->rx_ring_num);
val64 |= RX_QUEUE_CFG_Q4_SZ(mem_share);
continue;
case 5:
mem_share = (mem_size / config->rx_ring_num);
val64 |= RX_QUEUE_CFG_Q5_SZ(mem_share);
continue;
case 6:
mem_share = (mem_size / config->rx_ring_num);
val64 |= RX_QUEUE_CFG_Q6_SZ(mem_share);
continue;
case 7:
mem_share = (mem_size / config->rx_ring_num);
val64 |= RX_QUEUE_CFG_Q7_SZ(mem_share);
continue;
}
}
writeq(val64, &bar0->rx_queue_cfg);
/*
* Filling Tx round robin registers
* as per the number of FIFOs
*/
switch (config->tx_fifo_num) {
case 1:
val64 = 0x0000000000000000ULL;
writeq(val64, &bar0->tx_w_round_robin_0);
writeq(val64, &bar0->tx_w_round_robin_1);
writeq(val64, &bar0->tx_w_round_robin_2);
writeq(val64, &bar0->tx_w_round_robin_3);
writeq(val64, &bar0->tx_w_round_robin_4);
break;
case 2:
val64 = 0x0000010000010000ULL;
writeq(val64, &bar0->tx_w_round_robin_0);
val64 = 0x0100000100000100ULL;
writeq(val64, &bar0->tx_w_round_robin_1);
val64 = 0x0001000001000001ULL;
writeq(val64, &bar0->tx_w_round_robin_2);
val64 = 0x0000010000010000ULL;
writeq(val64, &bar0->tx_w_round_robin_3);
val64 = 0x0100000000000000ULL;
writeq(val64, &bar0->tx_w_round_robin_4);
break;
case 3:
val64 = 0x0001000102000001ULL;
writeq(val64, &bar0->tx_w_round_robin_0);
val64 = 0x0001020000010001ULL;
writeq(val64, &bar0->tx_w_round_robin_1);
val64 = 0x0200000100010200ULL;
writeq(val64, &bar0->tx_w_round_robin_2);
val64 = 0x0001000102000001ULL;
writeq(val64, &bar0->tx_w_round_robin_3);
val64 = 0x0001020000000000ULL;
writeq(val64, &bar0->tx_w_round_robin_4);
break;
case 4:
val64 = 0x0001020300010200ULL;
writeq(val64, &bar0->tx_w_round_robin_0);
val64 = 0x0100000102030001ULL;
writeq(val64, &bar0->tx_w_round_robin_1);
val64 = 0x0200010000010203ULL;
writeq(val64, &bar0->tx_w_round_robin_2);
val64 = 0x0001020001000001ULL;
writeq(val64, &bar0->tx_w_round_robin_3);
val64 = 0x0203000100000000ULL;
writeq(val64, &bar0->tx_w_round_robin_4);
break;
case 5:
val64 = 0x0001000203000102ULL;
writeq(val64, &bar0->tx_w_round_robin_0);
val64 = 0x0001020001030004ULL;
writeq(val64, &bar0->tx_w_round_robin_1);
val64 = 0x0001000203000102ULL;
writeq(val64, &bar0->tx_w_round_robin_2);
val64 = 0x0001020001030004ULL;
writeq(val64, &bar0->tx_w_round_robin_3);
val64 = 0x0001000000000000ULL;
writeq(val64, &bar0->tx_w_round_robin_4);
break;
case 6:
val64 = 0x0001020304000102ULL;
writeq(val64, &bar0->tx_w_round_robin_0);
val64 = 0x0304050001020001ULL;
writeq(val64, &bar0->tx_w_round_robin_1);
val64 = 0x0203000100000102ULL;
writeq(val64, &bar0->tx_w_round_robin_2);
val64 = 0x0304000102030405ULL;
writeq(val64, &bar0->tx_w_round_robin_3);
val64 = 0x0001000200000000ULL;
writeq(val64, &bar0->tx_w_round_robin_4);
break;
case 7:
val64 = 0x0001020001020300ULL;
writeq(val64, &bar0->tx_w_round_robin_0);
val64 = 0x0102030400010203ULL;
writeq(val64, &bar0->tx_w_round_robin_1);
val64 = 0x0405060001020001ULL;
writeq(val64, &bar0->tx_w_round_robin_2);
val64 = 0x0304050000010200ULL;
writeq(val64, &bar0->tx_w_round_robin_3);
val64 = 0x0102030000000000ULL;
writeq(val64, &bar0->tx_w_round_robin_4);
break;
case 8:
val64 = 0x0001020300040105ULL;
writeq(val64, &bar0->tx_w_round_robin_0);
val64 = 0x0200030106000204ULL;
writeq(val64, &bar0->tx_w_round_robin_1);
val64 = 0x0103000502010007ULL;
writeq(val64, &bar0->tx_w_round_robin_2);
val64 = 0x0304010002060500ULL;
writeq(val64, &bar0->tx_w_round_robin_3);
val64 = 0x0103020400000000ULL;
writeq(val64, &bar0->tx_w_round_robin_4);
break;
}
/* Enable all configured Tx FIFO partitions */
val64 = readq(&bar0->tx_fifo_partition_0);
val64 |= (TX_FIFO_PARTITION_EN);
writeq(val64, &bar0->tx_fifo_partition_0);
/* Filling the Rx round robin registers as per the
* number of Rings and steering based on QoS.
*/
switch (config->rx_ring_num) {
case 1:
val64 = 0x8080808080808080ULL;
writeq(val64, &bar0->rts_qos_steering);
break;
case 2:
val64 = 0x0000010000010000ULL;
writeq(val64, &bar0->rx_w_round_robin_0);
val64 = 0x0100000100000100ULL;
writeq(val64, &bar0->rx_w_round_robin_1);
val64 = 0x0001000001000001ULL;
writeq(val64, &bar0->rx_w_round_robin_2);
val64 = 0x0000010000010000ULL;
writeq(val64, &bar0->rx_w_round_robin_3);
val64 = 0x0100000000000000ULL;
writeq(val64, &bar0->rx_w_round_robin_4);
val64 = 0x8080808040404040ULL;
writeq(val64, &bar0->rts_qos_steering);
break;
case 3:
val64 = 0x0001000102000001ULL;
writeq(val64, &bar0->rx_w_round_robin_0);
val64 = 0x0001020000010001ULL;
writeq(val64, &bar0->rx_w_round_robin_1);
val64 = 0x0200000100010200ULL;
writeq(val64, &bar0->rx_w_round_robin_2);
val64 = 0x0001000102000001ULL;
writeq(val64, &bar0->rx_w_round_robin_3);
val64 = 0x0001020000000000ULL;
writeq(val64, &bar0->rx_w_round_robin_4);
val64 = 0x8080804040402020ULL;
writeq(val64, &bar0->rts_qos_steering);
break;
case 4:
val64 = 0x0001020300010200ULL;
writeq(val64, &bar0->rx_w_round_robin_0);
val64 = 0x0100000102030001ULL;
writeq(val64, &bar0->rx_w_round_robin_1);
val64 = 0x0200010000010203ULL;
writeq(val64, &bar0->rx_w_round_robin_2);
val64 = 0x0001020001000001ULL;
writeq(val64, &bar0->rx_w_round_robin_3);
val64 = 0x0203000100000000ULL;
writeq(val64, &bar0->rx_w_round_robin_4);
val64 = 0x8080404020201010ULL;
writeq(val64, &bar0->rts_qos_steering);
break;
case 5:
val64 = 0x0001000203000102ULL;
writeq(val64, &bar0->rx_w_round_robin_0);
val64 = 0x0001020001030004ULL;
writeq(val64, &bar0->rx_w_round_robin_1);
val64 = 0x0001000203000102ULL;
writeq(val64, &bar0->rx_w_round_robin_2);
val64 = 0x0001020001030004ULL;
writeq(val64, &bar0->rx_w_round_robin_3);
val64 = 0x0001000000000000ULL;
writeq(val64, &bar0->rx_w_round_robin_4);
val64 = 0x8080404020201008ULL;
writeq(val64, &bar0->rts_qos_steering);
break;
case 6:
val64 = 0x0001020304000102ULL;
writeq(val64, &bar0->rx_w_round_robin_0);
val64 = 0x0304050001020001ULL;
writeq(val64, &bar0->rx_w_round_robin_1);
val64 = 0x0203000100000102ULL;
writeq(val64, &bar0->rx_w_round_robin_2);
val64 = 0x0304000102030405ULL;
writeq(val64, &bar0->rx_w_round_robin_3);
val64 = 0x0001000200000000ULL;
writeq(val64, &bar0->rx_w_round_robin_4);
val64 = 0x8080404020100804ULL;
writeq(val64, &bar0->rts_qos_steering);
break;
case 7:
val64 = 0x0001020001020300ULL;
writeq(val64, &bar0->rx_w_round_robin_0);
val64 = 0x0102030400010203ULL;
writeq(val64, &bar0->rx_w_round_robin_1);
val64 = 0x0405060001020001ULL;
writeq(val64, &bar0->rx_w_round_robin_2);
val64 = 0x0304050000010200ULL;
writeq(val64, &bar0->rx_w_round_robin_3);
val64 = 0x0102030000000000ULL;
writeq(val64, &bar0->rx_w_round_robin_4);
val64 = 0x8080402010080402ULL;
writeq(val64, &bar0->rts_qos_steering);
break;
case 8:
val64 = 0x0001020300040105ULL;
writeq(val64, &bar0->rx_w_round_robin_0);
val64 = 0x0200030106000204ULL;
writeq(val64, &bar0->rx_w_round_robin_1);
val64 = 0x0103000502010007ULL;
writeq(val64, &bar0->rx_w_round_robin_2);
val64 = 0x0304010002060500ULL;
writeq(val64, &bar0->rx_w_round_robin_3);
val64 = 0x0103020400000000ULL;
writeq(val64, &bar0->rx_w_round_robin_4);
val64 = 0x8040201008040201ULL;
writeq(val64, &bar0->rts_qos_steering);
break;
}
/* UDP Fix */
val64 = 0;
for (i = 0; i < 8; i++)
writeq(val64, &bar0->rts_frm_len_n[i]);
/* Set the default rts frame length for the rings configured */
val64 = MAC_RTS_FRM_LEN_SET(dev->mtu+22);
for (i = 0 ; i < config->rx_ring_num ; i++)
writeq(val64, &bar0->rts_frm_len_n[i]);
/* Set the frame length for the configured rings
* desired by the user
*/
for (i = 0; i < config->rx_ring_num; i++) {
/* If rts_frm_len[i] == 0 then it is assumed that user not
* specified frame length steering.
* If the user provides the frame length then program
* the rts_frm_len register for those values or else
* leave it as it is.
*/
if (rts_frm_len[i] != 0) {
writeq(MAC_RTS_FRM_LEN_SET(rts_frm_len[i]),
&bar0->rts_frm_len_n[i]);
}
}
/* Disable differentiated services steering logic */
for (i = 0; i < 64; i++) {
if (rts_ds_steer(nic, i, 0) == FAILURE) {
DBG_PRINT(ERR_DBG, "%s: failed rts ds steering",
dev->name);
DBG_PRINT(ERR_DBG, "set on codepoint %d\n", i);
return FAILURE;
}
}
/* Program statistics memory */
writeq(mac_control->stats_mem_phy, &bar0->stat_addr);
if (nic->device_type == XFRAME_II_DEVICE) {
val64 = STAT_BC(0x320);
writeq(val64, &bar0->stat_byte_cnt);
}
/*
* Initializing the sampling rate for the device to calculate the
* bandwidth utilization.
*/
val64 = MAC_TX_LINK_UTIL_VAL(tmac_util_period) |
MAC_RX_LINK_UTIL_VAL(rmac_util_period);
writeq(val64, &bar0->mac_link_util);
/*
* Initializing the Transmit and Receive Traffic Interrupt
* Scheme.
*/
/*
* TTI Initialization. Default Tx timer gets us about
* 250 interrupts per sec. Continuous interrupts are enabled
* by default.
*/
if (nic->device_type == XFRAME_II_DEVICE) {
int count = (nic->config.bus_speed * 125)/2;
val64 = TTI_DATA1_MEM_TX_TIMER_VAL(count);
} else {
val64 = TTI_DATA1_MEM_TX_TIMER_VAL(0x2078);
}
val64 |= TTI_DATA1_MEM_TX_URNG_A(0xA) |
TTI_DATA1_MEM_TX_URNG_B(0x10) |
TTI_DATA1_MEM_TX_URNG_C(0x30) | TTI_DATA1_MEM_TX_TIMER_AC_EN;
if (use_continuous_tx_intrs)
val64 |= TTI_DATA1_MEM_TX_TIMER_CI_EN;
writeq(val64, &bar0->tti_data1_mem);
val64 = TTI_DATA2_MEM_TX_UFC_A(0x10) |
TTI_DATA2_MEM_TX_UFC_B(0x20) |
TTI_DATA2_MEM_TX_UFC_C(0x40) | TTI_DATA2_MEM_TX_UFC_D(0x80);
writeq(val64, &bar0->tti_data2_mem);
val64 = TTI_CMD_MEM_WE | TTI_CMD_MEM_STROBE_NEW_CMD;
writeq(val64, &bar0->tti_command_mem);
/*
* Once the operation completes, the Strobe bit of the command
* register will be reset. We poll for this particular condition
* We wait for a maximum of 500ms for the operation to complete,
* if it's not complete by then we return error.
*/
time = 0;
while (TRUE) {
val64 = readq(&bar0->tti_command_mem);
if (!(val64 & TTI_CMD_MEM_STROBE_NEW_CMD)) {
break;
}
if (time > 10) {
DBG_PRINT(ERR_DBG, "%s: TTI init Failed\n",
dev->name);
return -1;
}
msleep(50);
time++;
}
if (nic->config.bimodal) {
int k = 0;
for (k = 0; k < config->rx_ring_num; k++) {
val64 = TTI_CMD_MEM_WE | TTI_CMD_MEM_STROBE_NEW_CMD;
val64 |= TTI_CMD_MEM_OFFSET(0x38+k);
writeq(val64, &bar0->tti_command_mem);
/*
* Once the operation completes, the Strobe bit of the command
* register will be reset. We poll for this particular condition
* We wait for a maximum of 500ms for the operation to complete,
* if it's not complete by then we return error.
*/
time = 0;
while (TRUE) {
val64 = readq(&bar0->tti_command_mem);
if (!(val64 & TTI_CMD_MEM_STROBE_NEW_CMD)) {
break;
}
if (time > 10) {
DBG_PRINT(ERR_DBG,
"%s: TTI init Failed\n",
dev->name);
return -1;
}
time++;
msleep(50);
}
}
} else {
/* RTI Initialization */
if (nic->device_type == XFRAME_II_DEVICE) {
/*
* Programmed to generate Apprx 500 Intrs per
* second
*/
int count = (nic->config.bus_speed * 125)/4;
val64 = RTI_DATA1_MEM_RX_TIMER_VAL(count);
} else {
val64 = RTI_DATA1_MEM_RX_TIMER_VAL(0xFFF);
}
val64 |= RTI_DATA1_MEM_RX_URNG_A(0xA) |
RTI_DATA1_MEM_RX_URNG_B(0x10) |
RTI_DATA1_MEM_RX_URNG_C(0x30) | RTI_DATA1_MEM_RX_TIMER_AC_EN;
writeq(val64, &bar0->rti_data1_mem);
val64 = RTI_DATA2_MEM_RX_UFC_A(0x1) |
RTI_DATA2_MEM_RX_UFC_B(0x2) ;
if (nic->intr_type == MSI_X)
val64 |= (RTI_DATA2_MEM_RX_UFC_C(0x20) | \
RTI_DATA2_MEM_RX_UFC_D(0x40));
else
val64 |= (RTI_DATA2_MEM_RX_UFC_C(0x40) | \
RTI_DATA2_MEM_RX_UFC_D(0x80));
writeq(val64, &bar0->rti_data2_mem);
for (i = 0; i < config->rx_ring_num; i++) {
val64 = RTI_CMD_MEM_WE | RTI_CMD_MEM_STROBE_NEW_CMD
| RTI_CMD_MEM_OFFSET(i);
writeq(val64, &bar0->rti_command_mem);
/*
* Once the operation completes, the Strobe bit of the
* command register will be reset. We poll for this
* particular condition. We wait for a maximum of 500ms
* for the operation to complete, if it's not complete
* by then we return error.
*/
time = 0;
while (TRUE) {
val64 = readq(&bar0->rti_command_mem);
if (!(val64 & RTI_CMD_MEM_STROBE_NEW_CMD)) {
break;
}
if (time > 10) {
DBG_PRINT(ERR_DBG, "%s: RTI init Failed\n",
dev->name);
return -1;
}
time++;
msleep(50);
}
}
}
/*
* Initializing proper values as Pause threshold into all
* the 8 Queues on Rx side.
*/
writeq(0xffbbffbbffbbffbbULL, &bar0->mc_pause_thresh_q0q3);
writeq(0xffbbffbbffbbffbbULL, &bar0->mc_pause_thresh_q4q7);
/* Disable RMAC PAD STRIPPING */
add = &bar0->mac_cfg;
val64 = readq(&bar0->mac_cfg);
val64 &= ~(MAC_CFG_RMAC_STRIP_PAD);
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) (val64), add);
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) (val64 >> 32), (add + 4));
val64 = readq(&bar0->mac_cfg);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
/* Enable FCS stripping by adapter */
add = &bar0->mac_cfg;
val64 = readq(&bar0->mac_cfg);
val64 |= MAC_CFG_RMAC_STRIP_FCS;
if (nic->device_type == XFRAME_II_DEVICE)
writeq(val64, &bar0->mac_cfg);
else {
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) (val64), add);
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) (val64 >> 32), (add + 4));
}
/*
* Set the time value to be inserted in the pause frame
* generated by xena.
*/
val64 = readq(&bar0->rmac_pause_cfg);
val64 &= ~(RMAC_PAUSE_HG_PTIME(0xffff));
val64 |= RMAC_PAUSE_HG_PTIME(nic->mac_control.rmac_pause_time);
writeq(val64, &bar0->rmac_pause_cfg);
/*
* Set the Threshold Limit for Generating the pause frame
* If the amount of data in any Queue exceeds ratio of
* (mac_control.mc_pause_threshold_q0q3 or q4q7)/256
* pause frame is generated
*/
val64 = 0;
for (i = 0; i < 4; i++) {
val64 |=
(((u64) 0xFF00 | nic->mac_control.
mc_pause_threshold_q0q3)
<< (i * 2 * 8));
}
writeq(val64, &bar0->mc_pause_thresh_q0q3);
val64 = 0;
for (i = 0; i < 4; i++) {
val64 |=
(((u64) 0xFF00 | nic->mac_control.
mc_pause_threshold_q4q7)
<< (i * 2 * 8));
}
writeq(val64, &bar0->mc_pause_thresh_q4q7);
/*
* TxDMA will stop Read request if the number of read split has
* exceeded the limit pointed by shared_splits
*/
val64 = readq(&bar0->pic_control);
val64 |= PIC_CNTL_SHARED_SPLITS(shared_splits);
writeq(val64, &bar0->pic_control);
if (nic->config.bus_speed == 266) {
writeq(TXREQTO_VAL(0x7f) | TXREQTO_EN, &bar0->txreqtimeout);
writeq(0x0, &bar0->read_retry_delay);
writeq(0x0, &bar0->write_retry_delay);
}
/*
* Programming the Herc to split every write transaction
* that does not start on an ADB to reduce disconnects.
*/
if (nic->device_type == XFRAME_II_DEVICE) {
val64 = FAULT_BEHAVIOUR | EXT_REQ_EN |
MISC_LINK_STABILITY_PRD(3);
writeq(val64, &bar0->misc_control);
val64 = readq(&bar0->pic_control2);
val64 &= ~(BIT(13)|BIT(14)|BIT(15));
writeq(val64, &bar0->pic_control2);
}
if (strstr(nic->product_name, "CX4")) {
val64 = TMAC_AVG_IPG(0x17);
writeq(val64, &bar0->tmac_avg_ipg);
}
return SUCCESS;
}
#define LINK_UP_DOWN_INTERRUPT 1
#define MAC_RMAC_ERR_TIMER 2
static int s2io_link_fault_indication(struct s2io_nic *nic)
{
if (nic->intr_type != INTA)
return MAC_RMAC_ERR_TIMER;
if (nic->device_type == XFRAME_II_DEVICE)
return LINK_UP_DOWN_INTERRUPT;
else
return MAC_RMAC_ERR_TIMER;
}
/**
* en_dis_able_nic_intrs - Enable or Disable the interrupts
* @nic: device private variable,
* @mask: A mask indicating which Intr block must be modified and,
* @flag: A flag indicating whether to enable or disable the Intrs.
* Description: This function will either disable or enable the interrupts
* depending on the flag argument. The mask argument can be used to
* enable/disable any Intr block.
* Return Value: NONE.
*/
static void en_dis_able_nic_intrs(struct s2io_nic *nic, u16 mask, int flag)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
register u64 val64 = 0, temp64 = 0;
/* Top level interrupt classification */
/* PIC Interrupts */
if ((mask & (TX_PIC_INTR | RX_PIC_INTR))) {
/* Enable PIC Intrs in the general intr mask register */
val64 = TXPIC_INT_M;
if (flag == ENABLE_INTRS) {
temp64 = readq(&bar0->general_int_mask);
temp64 &= ~((u64) val64);
writeq(temp64, &bar0->general_int_mask);
/*
* If Hercules adapter enable GPIO otherwise
* disable all PCIX, Flash, MDIO, IIC and GPIO
* interrupts for now.
* TODO
*/
if (s2io_link_fault_indication(nic) ==
LINK_UP_DOWN_INTERRUPT ) {
temp64 = readq(&bar0->pic_int_mask);
temp64 &= ~((u64) PIC_INT_GPIO);
writeq(temp64, &bar0->pic_int_mask);
temp64 = readq(&bar0->gpio_int_mask);
temp64 &= ~((u64) GPIO_INT_MASK_LINK_UP);
writeq(temp64, &bar0->gpio_int_mask);
} else {
writeq(DISABLE_ALL_INTRS, &bar0->pic_int_mask);
}
/*
* No MSI Support is available presently, so TTI and
* RTI interrupts are also disabled.
*/
} else if (flag == DISABLE_INTRS) {
/*
* Disable PIC Intrs in the general
* intr mask register
*/
writeq(DISABLE_ALL_INTRS, &bar0->pic_int_mask);
temp64 = readq(&bar0->general_int_mask);
val64 |= temp64;
writeq(val64, &bar0->general_int_mask);
}
}
/* MAC Interrupts */
/* Enabling/Disabling MAC interrupts */
if (mask & (TX_MAC_INTR | RX_MAC_INTR)) {
val64 = TXMAC_INT_M | RXMAC_INT_M;
if (flag == ENABLE_INTRS) {
temp64 = readq(&bar0->general_int_mask);
temp64 &= ~((u64) val64);
writeq(temp64, &bar0->general_int_mask);
/*
* All MAC block error interrupts are disabled for now
* TODO
*/
} else if (flag == DISABLE_INTRS) {
/*
* Disable MAC Intrs in the general intr mask register
*/
writeq(DISABLE_ALL_INTRS, &bar0->mac_int_mask);
writeq(DISABLE_ALL_INTRS,
&bar0->mac_rmac_err_mask);
temp64 = readq(&bar0->general_int_mask);
val64 |= temp64;
writeq(val64, &bar0->general_int_mask);
}
}
/* Tx traffic interrupts */
if (mask & TX_TRAFFIC_INTR) {
val64 = TXTRAFFIC_INT_M;
if (flag == ENABLE_INTRS) {
temp64 = readq(&bar0->general_int_mask);
temp64 &= ~((u64) val64);
writeq(temp64, &bar0->general_int_mask);
/*
* Enable all the Tx side interrupts
* writing 0 Enables all 64 TX interrupt levels
*/
writeq(0x0, &bar0->tx_traffic_mask);
} else if (flag == DISABLE_INTRS) {
/*
* Disable Tx Traffic Intrs in the general intr mask
* register.
*/
writeq(DISABLE_ALL_INTRS, &bar0->tx_traffic_mask);
temp64 = readq(&bar0->general_int_mask);
val64 |= temp64;
writeq(val64, &bar0->general_int_mask);
}
}
/* Rx traffic interrupts */
if (mask & RX_TRAFFIC_INTR) {
val64 = RXTRAFFIC_INT_M;
if (flag == ENABLE_INTRS) {
temp64 = readq(&bar0->general_int_mask);
temp64 &= ~((u64) val64);
writeq(temp64, &bar0->general_int_mask);
/* writing 0 Enables all 8 RX interrupt levels */
writeq(0x0, &bar0->rx_traffic_mask);
} else if (flag == DISABLE_INTRS) {
/*
* Disable Rx Traffic Intrs in the general intr mask
* register.
*/
writeq(DISABLE_ALL_INTRS, &bar0->rx_traffic_mask);
temp64 = readq(&bar0->general_int_mask);
val64 |= temp64;
writeq(val64, &bar0->general_int_mask);
}
}
}
/**
* verify_pcc_quiescent- Checks for PCC quiescent state
* Return: 1 If PCC is quiescence
* 0 If PCC is not quiescence
*/
static int verify_pcc_quiescent(struct s2io_nic *sp, int flag)
{
int ret = 0, herc;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64 = readq(&bar0->adapter_status);
herc = (sp->device_type == XFRAME_II_DEVICE);
if (flag == FALSE) {
if ((!herc && (get_xena_rev_id(sp->pdev) >= 4)) || herc) {
if (!(val64 & ADAPTER_STATUS_RMAC_PCC_IDLE))
ret = 1;
} else {
if (!(val64 & ADAPTER_STATUS_RMAC_PCC_FOUR_IDLE))
ret = 1;
}
} else {
if ((!herc && (get_xena_rev_id(sp->pdev) >= 4)) || herc) {
if (((val64 & ADAPTER_STATUS_RMAC_PCC_IDLE) ==
ADAPTER_STATUS_RMAC_PCC_IDLE))
ret = 1;
} else {
if (((val64 & ADAPTER_STATUS_RMAC_PCC_FOUR_IDLE) ==
ADAPTER_STATUS_RMAC_PCC_FOUR_IDLE))
ret = 1;
}
}
return ret;
}
/**
* verify_xena_quiescence - Checks whether the H/W is ready
* Description: Returns whether the H/W is ready to go or not. Depending
* on whether adapter enable bit was written or not the comparison
* differs and the calling function passes the input argument flag to
* indicate this.
* Return: 1 If xena is quiescence
* 0 If Xena is not quiescence
*/
static int verify_xena_quiescence(struct s2io_nic *sp)
{
int mode;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64 = readq(&bar0->adapter_status);
mode = s2io_verify_pci_mode(sp);
if (!(val64 & ADAPTER_STATUS_TDMA_READY)) {
DBG_PRINT(ERR_DBG, "%s", "TDMA is not ready!");
return 0;
}
if (!(val64 & ADAPTER_STATUS_RDMA_READY)) {
DBG_PRINT(ERR_DBG, "%s", "RDMA is not ready!");
return 0;
}
if (!(val64 & ADAPTER_STATUS_PFC_READY)) {
DBG_PRINT(ERR_DBG, "%s", "PFC is not ready!");
return 0;
}
if (!(val64 & ADAPTER_STATUS_TMAC_BUF_EMPTY)) {
DBG_PRINT(ERR_DBG, "%s", "TMAC BUF is not empty!");
return 0;
}
if (!(val64 & ADAPTER_STATUS_PIC_QUIESCENT)) {
DBG_PRINT(ERR_DBG, "%s", "PIC is not QUIESCENT!");
return 0;
}
if (!(val64 & ADAPTER_STATUS_MC_DRAM_READY)) {
DBG_PRINT(ERR_DBG, "%s", "MC_DRAM is not ready!");
return 0;
}
if (!(val64 & ADAPTER_STATUS_MC_QUEUES_READY)) {
DBG_PRINT(ERR_DBG, "%s", "MC_QUEUES is not ready!");
return 0;
}
if (!(val64 & ADAPTER_STATUS_M_PLL_LOCK)) {
DBG_PRINT(ERR_DBG, "%s", "M_PLL is not locked!");
return 0;
}
/*
* In PCI 33 mode, the P_PLL is not used, and therefore,
* the the P_PLL_LOCK bit in the adapter_status register will
* not be asserted.
*/
if (!(val64 & ADAPTER_STATUS_P_PLL_LOCK) &&
sp->device_type == XFRAME_II_DEVICE && mode !=
PCI_MODE_PCI_33) {
DBG_PRINT(ERR_DBG, "%s", "P_PLL is not locked!");
return 0;
}
if (!((val64 & ADAPTER_STATUS_RC_PRC_QUIESCENT) ==
ADAPTER_STATUS_RC_PRC_QUIESCENT)) {
DBG_PRINT(ERR_DBG, "%s", "RC_PRC is not QUIESCENT!");
return 0;
}
return 1;
}
/**
* fix_mac_address - Fix for Mac addr problem on Alpha platforms
* @sp: Pointer to device specifc structure
* Description :
* New procedure to clear mac address reading problems on Alpha platforms
*
*/
static void fix_mac_address(struct s2io_nic * sp)
{
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64;
int i = 0;
while (fix_mac[i] != END_SIGN) {
writeq(fix_mac[i++], &bar0->gpio_control);
udelay(10);
val64 = readq(&bar0->gpio_control);
}
}
/**
* start_nic - Turns the device on
* @nic : device private variable.
* Description:
* This function actually turns the device on. Before this function is
* called,all Registers are configured from their reset states
* and shared memory is allocated but the NIC is still quiescent. On
* calling this function, the device interrupts are cleared and the NIC is
* literally switched on by writing into the adapter control register.
* Return Value:
* SUCCESS on success and -1 on failure.
*/
static int start_nic(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
struct net_device *dev = nic->dev;
register u64 val64 = 0;
u16 subid, i;
struct mac_info *mac_control;
struct config_param *config;
mac_control = &nic->mac_control;
config = &nic->config;
/* PRC Initialization and configuration */
for (i = 0; i < config->rx_ring_num; i++) {
writeq((u64) mac_control->rings[i].rx_blocks[0].block_dma_addr,
&bar0->prc_rxd0_n[i]);
val64 = readq(&bar0->prc_ctrl_n[i]);
if (nic->config.bimodal)
val64 |= PRC_CTRL_BIMODAL_INTERRUPT;
if (nic->rxd_mode == RXD_MODE_1)
val64 |= PRC_CTRL_RC_ENABLED;
else
val64 |= PRC_CTRL_RC_ENABLED | PRC_CTRL_RING_MODE_3;
if (nic->device_type == XFRAME_II_DEVICE)
val64 |= PRC_CTRL_GROUP_READS;
val64 &= ~PRC_CTRL_RXD_BACKOFF_INTERVAL(0xFFFFFF);
val64 |= PRC_CTRL_RXD_BACKOFF_INTERVAL(0x1000);
writeq(val64, &bar0->prc_ctrl_n[i]);
}
if (nic->rxd_mode == RXD_MODE_3B) {
/* Enabling 2 buffer mode by writing into Rx_pa_cfg reg. */
val64 = readq(&bar0->rx_pa_cfg);
val64 |= RX_PA_CFG_IGNORE_L2_ERR;
writeq(val64, &bar0->rx_pa_cfg);
}
if (vlan_tag_strip == 0) {
val64 = readq(&bar0->rx_pa_cfg);
val64 &= ~RX_PA_CFG_STRIP_VLAN_TAG;
writeq(val64, &bar0->rx_pa_cfg);
vlan_strip_flag = 0;
}
/*
* Enabling MC-RLDRAM. After enabling the device, we timeout
* for around 100ms, which is approximately the time required
* for the device to be ready for operation.
*/
val64 = readq(&bar0->mc_rldram_mrs);
val64 |= MC_RLDRAM_QUEUE_SIZE_ENABLE | MC_RLDRAM_MRS_ENABLE;
SPECIAL_REG_WRITE(val64, &bar0->mc_rldram_mrs, UF);
val64 = readq(&bar0->mc_rldram_mrs);
msleep(100); /* Delay by around 100 ms. */
/* Enabling ECC Protection. */
val64 = readq(&bar0->adapter_control);
val64 &= ~ADAPTER_ECC_EN;
writeq(val64, &bar0->adapter_control);
/*
* Clearing any possible Link state change interrupts that
* could have popped up just before Enabling the card.
*/
val64 = readq(&bar0->mac_rmac_err_reg);
if (val64)
writeq(val64, &bar0->mac_rmac_err_reg);
/*
* Verify if the device is ready to be enabled, if so enable
* it.
*/
val64 = readq(&bar0->adapter_status);
if (!verify_xena_quiescence(nic)) {
DBG_PRINT(ERR_DBG, "%s: device is not ready, ", dev->name);
DBG_PRINT(ERR_DBG, "Adapter status reads: 0x%llx\n",
(unsigned long long) val64);
return FAILURE;
}
/*
* With some switches, link might be already up at this point.
* Because of this weird behavior, when we enable laser,
* we may not get link. We need to handle this. We cannot
* figure out which switch is misbehaving. So we are forced to
* make a global change.
*/
/* Enabling Laser. */
val64 = readq(&bar0->adapter_control);
val64 |= ADAPTER_EOI_TX_ON;
writeq(val64, &bar0->adapter_control);
if (s2io_link_fault_indication(nic) == MAC_RMAC_ERR_TIMER) {
/*
* Dont see link state interrupts initally on some switches,
* so directly scheduling the link state task here.
*/
schedule_work(&nic->set_link_task);
}
/* SXE-002: Initialize link and activity LED */
subid = nic->pdev->subsystem_device;
if (((subid & 0xFF) >= 0x07) &&
(nic->device_type == XFRAME_I_DEVICE)) {
val64 = readq(&bar0->gpio_control);
val64 |= 0x0000800000000000ULL;
writeq(val64, &bar0->gpio_control);
val64 = 0x0411040400000000ULL;
writeq(val64, (void __iomem *)bar0 + 0x2700);
}
return SUCCESS;
}
/**
* s2io_txdl_getskb - Get the skb from txdl, unmap and return skb
*/
static struct sk_buff *s2io_txdl_getskb(struct fifo_info *fifo_data, struct \
TxD *txdlp, int get_off)
{
struct s2io_nic *nic = fifo_data->nic;
struct sk_buff *skb;
struct TxD *txds;
u16 j, frg_cnt;
txds = txdlp;
if (txds->Host_Control == (u64)(long)nic->ufo_in_band_v) {
pci_unmap_single(nic->pdev, (dma_addr_t)
txds->Buffer_Pointer, sizeof(u64),
PCI_DMA_TODEVICE);
txds++;
}
skb = (struct sk_buff *) ((unsigned long)
txds->Host_Control);
if (!skb) {
memset(txdlp, 0, (sizeof(struct TxD) * fifo_data->max_txds));
return NULL;
}
pci_unmap_single(nic->pdev, (dma_addr_t)
txds->Buffer_Pointer,
skb->len - skb->data_len,
PCI_DMA_TODEVICE);
frg_cnt = skb_shinfo(skb)->nr_frags;
if (frg_cnt) {
txds++;
for (j = 0; j < frg_cnt; j++, txds++) {
skb_frag_t *frag = &skb_shinfo(skb)->frags[j];
if (!txds->Buffer_Pointer)
break;
pci_unmap_page(nic->pdev, (dma_addr_t)
txds->Buffer_Pointer,
frag->size, PCI_DMA_TODEVICE);
}
}
memset(txdlp,0, (sizeof(struct TxD) * fifo_data->max_txds));
return(skb);
}
/**
* free_tx_buffers - Free all queued Tx buffers
* @nic : device private variable.
* Description:
* Free all queued Tx buffers.
* Return Value: void
*/
static void free_tx_buffers(struct s2io_nic *nic)
{
struct net_device *dev = nic->dev;
struct sk_buff *skb;
struct TxD *txdp;
int i, j;
struct mac_info *mac_control;
struct config_param *config;
int cnt = 0;
mac_control = &nic->mac_control;
config = &nic->config;
for (i = 0; i < config->tx_fifo_num; i++) {
for (j = 0; j < config->tx_cfg[i].fifo_len - 1; j++) {
txdp = (struct TxD *) mac_control->fifos[i].list_info[j].
list_virt_addr;
skb = s2io_txdl_getskb(&mac_control->fifos[i], txdp, j);
if (skb) {
dev_kfree_skb(skb);
cnt++;
}
}
DBG_PRINT(INTR_DBG,
"%s:forcibly freeing %d skbs on FIFO%d\n",
dev->name, cnt, i);
mac_control->fifos[i].tx_curr_get_info.offset = 0;
mac_control->fifos[i].tx_curr_put_info.offset = 0;
}
}
/**
* stop_nic - To stop the nic
* @nic ; device private variable.
* Description:
* This function does exactly the opposite of what the start_nic()
* function does. This function is called to stop the device.
* Return Value:
* void.
*/
static void stop_nic(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
register u64 val64 = 0;
u16 interruptible;
struct mac_info *mac_control;
struct config_param *config;
mac_control = &nic->mac_control;
config = &nic->config;
/* Disable all interrupts */
interruptible = TX_TRAFFIC_INTR | RX_TRAFFIC_INTR;
interruptible |= TX_PIC_INTR | RX_PIC_INTR;
interruptible |= TX_MAC_INTR | RX_MAC_INTR;
en_dis_able_nic_intrs(nic, interruptible, DISABLE_INTRS);
/* Clearing Adapter_En bit of ADAPTER_CONTROL Register */
val64 = readq(&bar0->adapter_control);
val64 &= ~(ADAPTER_CNTL_EN);
writeq(val64, &bar0->adapter_control);
}
static int fill_rxd_3buf(struct s2io_nic *nic, struct RxD_t *rxdp, struct \
sk_buff *skb)
{
struct net_device *dev = nic->dev;
struct sk_buff *frag_list;
void *tmp;
/* Buffer-1 receives L3/L4 headers */
((struct RxD3*)rxdp)->Buffer1_ptr = pci_map_single
(nic->pdev, skb->data, l3l4hdr_size + 4,
PCI_DMA_FROMDEVICE);
/* skb_shinfo(skb)->frag_list will have L4 data payload */
skb_shinfo(skb)->frag_list = dev_alloc_skb(dev->mtu + ALIGN_SIZE);
if (skb_shinfo(skb)->frag_list == NULL) {
DBG_PRINT(INFO_DBG, "%s: dev_alloc_skb failed\n ", dev->name);
return -ENOMEM ;
}
frag_list = skb_shinfo(skb)->frag_list;
skb->truesize += frag_list->truesize;
frag_list->next = NULL;
tmp = (void *)ALIGN((long)frag_list->data, ALIGN_SIZE + 1);
frag_list->data = tmp;
skb_reset_tail_pointer(frag_list);
/* Buffer-2 receives L4 data payload */
((struct RxD3*)rxdp)->Buffer2_ptr = pci_map_single(nic->pdev,
frag_list->data, dev->mtu,
PCI_DMA_FROMDEVICE);
rxdp->Control_2 |= SET_BUFFER1_SIZE_3(l3l4hdr_size + 4);
rxdp->Control_2 |= SET_BUFFER2_SIZE_3(dev->mtu);
return SUCCESS;
}
/**
* fill_rx_buffers - Allocates the Rx side skbs
* @nic: device private variable
* @ring_no: ring number
* Description:
* The function allocates Rx side skbs and puts the physical
* address of these buffers into the RxD buffer pointers, so that the NIC
* can DMA the received frame into these locations.
* The NIC supports 3 receive modes, viz
* 1. single buffer,
* 2. three buffer and
* 3. Five buffer modes.
* Each mode defines how many fragments the received frame will be split
* up into by the NIC. The frame is split into L3 header, L4 Header,
* L4 payload in three buffer mode and in 5 buffer mode, L4 payload itself
* is split into 3 fragments. As of now only single buffer mode is
* supported.
* Return Value:
* SUCCESS on success or an appropriate -ve value on failure.
*/
static int fill_rx_buffers(struct s2io_nic *nic, int ring_no)
{
struct net_device *dev = nic->dev;
struct sk_buff *skb;
struct RxD_t *rxdp;
int off, off1, size, block_no, block_no1;
u32 alloc_tab = 0;
u32 alloc_cnt;
struct mac_info *mac_control;
struct config_param *config;
u64 tmp;
struct buffAdd *ba;
unsigned long flags;
struct RxD_t *first_rxdp = NULL;
u64 Buffer0_ptr = 0, Buffer1_ptr = 0;
mac_control = &nic->mac_control;
config = &nic->config;
alloc_cnt = mac_control->rings[ring_no].pkt_cnt -
atomic_read(&nic->rx_bufs_left[ring_no]);
block_no1 = mac_control->rings[ring_no].rx_curr_get_info.block_index;
off1 = mac_control->rings[ring_no].rx_curr_get_info.offset;
while (alloc_tab < alloc_cnt) {
block_no = mac_control->rings[ring_no].rx_curr_put_info.
block_index;
off = mac_control->rings[ring_no].rx_curr_put_info.offset;
rxdp = mac_control->rings[ring_no].
rx_blocks[block_no].rxds[off].virt_addr;
if ((block_no == block_no1) && (off == off1) &&
(rxdp->Host_Control)) {
DBG_PRINT(INTR_DBG, "%s: Get and Put",
dev->name);
DBG_PRINT(INTR_DBG, " info equated\n");
goto end;
}
if (off && (off == rxd_count[nic->rxd_mode])) {
mac_control->rings[ring_no].rx_curr_put_info.
block_index++;
if (mac_control->rings[ring_no].rx_curr_put_info.
block_index == mac_control->rings[ring_no].
block_count)
mac_control->rings[ring_no].rx_curr_put_info.
block_index = 0;
block_no = mac_control->rings[ring_no].
rx_curr_put_info.block_index;
if (off == rxd_count[nic->rxd_mode])
off = 0;
mac_control->rings[ring_no].rx_curr_put_info.
offset = off;
rxdp = mac_control->rings[ring_no].
rx_blocks[block_no].block_virt_addr;
DBG_PRINT(INTR_DBG, "%s: Next block at: %p\n",
dev->name, rxdp);
}
if(!napi) {
spin_lock_irqsave(&nic->put_lock, flags);
mac_control->rings[ring_no].put_pos =
(block_no * (rxd_count[nic->rxd_mode] + 1)) + off;
spin_unlock_irqrestore(&nic->put_lock, flags);
} else {
mac_control->rings[ring_no].put_pos =
(block_no * (rxd_count[nic->rxd_mode] + 1)) + off;
}
if ((rxdp->Control_1 & RXD_OWN_XENA) &&
((nic->rxd_mode >= RXD_MODE_3A) &&
(rxdp->Control_2 & BIT(0)))) {
mac_control->rings[ring_no].rx_curr_put_info.
offset = off;
goto end;
}
/* calculate size of skb based on ring mode */
size = dev->mtu + HEADER_ETHERNET_II_802_3_SIZE +
HEADER_802_2_SIZE + HEADER_SNAP_SIZE;
if (nic->rxd_mode == RXD_MODE_1)
size += NET_IP_ALIGN;
else if (nic->rxd_mode == RXD_MODE_3B)
size = dev->mtu + ALIGN_SIZE + BUF0_LEN + 4;
else
size = l3l4hdr_size + ALIGN_SIZE + BUF0_LEN + 4;
/* allocate skb */
skb = dev_alloc_skb(size);
if(!skb) {
DBG_PRINT(INFO_DBG, "%s: Out of ", dev->name);
DBG_PRINT(INFO_DBG, "memory to allocate SKBs\n");
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
if (first_rxdp) {
wmb();
first_rxdp->Control_1 |= RXD_OWN_XENA;
}
return -ENOMEM ;
}
if (nic->rxd_mode == RXD_MODE_1) {
/* 1 buffer mode - normal operation mode */
memset(rxdp, 0, sizeof(struct RxD1));
skb_reserve(skb, NET_IP_ALIGN);
((struct RxD1*)rxdp)->Buffer0_ptr = pci_map_single
(nic->pdev, skb->data, size - NET_IP_ALIGN,
PCI_DMA_FROMDEVICE);
rxdp->Control_2 = SET_BUFFER0_SIZE_1(size - NET_IP_ALIGN);
} else if (nic->rxd_mode >= RXD_MODE_3A) {
/*
* 2 or 3 buffer mode -
* Both 2 buffer mode and 3 buffer mode provides 128
* byte aligned receive buffers.
*
* 3 buffer mode provides header separation where in
* skb->data will have L3/L4 headers where as
* skb_shinfo(skb)->frag_list will have the L4 data
* payload
*/
/* save the buffer pointers to avoid frequent dma mapping */
Buffer0_ptr = ((struct RxD3*)rxdp)->Buffer0_ptr;
Buffer1_ptr = ((struct RxD3*)rxdp)->Buffer1_ptr;
memset(rxdp, 0, sizeof(struct RxD3));
/* restore the buffer pointers for dma sync*/
((struct RxD3*)rxdp)->Buffer0_ptr = Buffer0_ptr;
((struct RxD3*)rxdp)->Buffer1_ptr = Buffer1_ptr;
ba = &mac_control->rings[ring_no].ba[block_no][off];
skb_reserve(skb, BUF0_LEN);
tmp = (u64)(unsigned long) skb->data;
tmp += ALIGN_SIZE;
tmp &= ~ALIGN_SIZE;
skb->data = (void *) (unsigned long)tmp;
skb_reset_tail_pointer(skb);
if (!(((struct RxD3*)rxdp)->Buffer0_ptr))
((struct RxD3*)rxdp)->Buffer0_ptr =
pci_map_single(nic->pdev, ba->ba_0, BUF0_LEN,
PCI_DMA_FROMDEVICE);
else
pci_dma_sync_single_for_device(nic->pdev,
(dma_addr_t) ((struct RxD3*)rxdp)->Buffer0_ptr,
BUF0_LEN, PCI_DMA_FROMDEVICE);
rxdp->Control_2 = SET_BUFFER0_SIZE_3(BUF0_LEN);
if (nic->rxd_mode == RXD_MODE_3B) {
/* Two buffer mode */
/*
* Buffer2 will have L3/L4 header plus
* L4 payload
*/
((struct RxD3*)rxdp)->Buffer2_ptr = pci_map_single
(nic->pdev, skb->data, dev->mtu + 4,
PCI_DMA_FROMDEVICE);
/* Buffer-1 will be dummy buffer. Not used */
if (!(((struct RxD3*)rxdp)->Buffer1_ptr)) {
((struct RxD3*)rxdp)->Buffer1_ptr =
pci_map_single(nic->pdev,
ba->ba_1, BUF1_LEN,
PCI_DMA_FROMDEVICE);
}
rxdp->Control_2 |= SET_BUFFER1_SIZE_3(1);
rxdp->Control_2 |= SET_BUFFER2_SIZE_3
(dev->mtu + 4);
} else {
/* 3 buffer mode */
if (fill_rxd_3buf(nic, rxdp, skb) == -ENOMEM) {
dev_kfree_skb_irq(skb);
if (first_rxdp) {
wmb();
first_rxdp->Control_1 |=
RXD_OWN_XENA;
}
return -ENOMEM ;
}
}
rxdp->Control_2 |= BIT(0);
}
rxdp->Host_Control = (unsigned long) (skb);
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
if (alloc_tab & ((1 << rxsync_frequency) - 1))
rxdp->Control_1 |= RXD_OWN_XENA;
off++;
if (off == (rxd_count[nic->rxd_mode] + 1))
off = 0;
mac_control->rings[ring_no].rx_curr_put_info.offset = off;
rxdp->Control_2 |= SET_RXD_MARKER;
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
if (!(alloc_tab & ((1 << rxsync_frequency) - 1))) {
if (first_rxdp) {
wmb();
first_rxdp->Control_1 |= RXD_OWN_XENA;
}
first_rxdp = rxdp;
}
atomic_inc(&nic->rx_bufs_left[ring_no]);
alloc_tab++;
}
end:
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
/* Transfer ownership of first descriptor to adapter just before
* exiting. Before that, use memory barrier so that ownership
* and other fields are seen by adapter correctly.
*/
if (first_rxdp) {
wmb();
first_rxdp->Control_1 |= RXD_OWN_XENA;
}
return SUCCESS;
}
static void free_rxd_blk(struct s2io_nic *sp, int ring_no, int blk)
{
struct net_device *dev = sp->dev;
int j;
struct sk_buff *skb;
struct RxD_t *rxdp;
struct mac_info *mac_control;
struct buffAdd *ba;
mac_control = &sp->mac_control;
for (j = 0 ; j < rxd_count[sp->rxd_mode]; j++) {
rxdp = mac_control->rings[ring_no].
rx_blocks[blk].rxds[j].virt_addr;
skb = (struct sk_buff *)
((unsigned long) rxdp->Host_Control);
if (!skb) {
continue;
}
if (sp->rxd_mode == RXD_MODE_1) {
pci_unmap_single(sp->pdev, (dma_addr_t)
((struct RxD1*)rxdp)->Buffer0_ptr,
dev->mtu +
HEADER_ETHERNET_II_802_3_SIZE
+ HEADER_802_2_SIZE +
HEADER_SNAP_SIZE,
PCI_DMA_FROMDEVICE);
memset(rxdp, 0, sizeof(struct RxD1));
} else if(sp->rxd_mode == RXD_MODE_3B) {
ba = &mac_control->rings[ring_no].
ba[blk][j];
pci_unmap_single(sp->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer0_ptr,
BUF0_LEN,
PCI_DMA_FROMDEVICE);
pci_unmap_single(sp->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer1_ptr,
BUF1_LEN,
PCI_DMA_FROMDEVICE);
pci_unmap_single(sp->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer2_ptr,
dev->mtu + 4,
PCI_DMA_FROMDEVICE);
memset(rxdp, 0, sizeof(struct RxD3));
} else {
pci_unmap_single(sp->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer0_ptr, BUF0_LEN,
PCI_DMA_FROMDEVICE);
pci_unmap_single(sp->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer1_ptr,
l3l4hdr_size + 4,
PCI_DMA_FROMDEVICE);
pci_unmap_single(sp->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer2_ptr, dev->mtu,
PCI_DMA_FROMDEVICE);
memset(rxdp, 0, sizeof(struct RxD3));
}
dev_kfree_skb(skb);
atomic_dec(&sp->rx_bufs_left[ring_no]);
}
}
/**
* free_rx_buffers - Frees all Rx buffers
* @sp: device private variable.
* Description:
* This function will free all Rx buffers allocated by host.
* Return Value:
* NONE.
*/
static void free_rx_buffers(struct s2io_nic *sp)
{
struct net_device *dev = sp->dev;
int i, blk = 0, buf_cnt = 0;
struct mac_info *mac_control;
struct config_param *config;
mac_control = &sp->mac_control;
config = &sp->config;
for (i = 0; i < config->rx_ring_num; i++) {
for (blk = 0; blk < rx_ring_sz[i]; blk++)
free_rxd_blk(sp,i,blk);
mac_control->rings[i].rx_curr_put_info.block_index = 0;
mac_control->rings[i].rx_curr_get_info.block_index = 0;
mac_control->rings[i].rx_curr_put_info.offset = 0;
mac_control->rings[i].rx_curr_get_info.offset = 0;
atomic_set(&sp->rx_bufs_left[i], 0);
DBG_PRINT(INIT_DBG, "%s:Freed 0x%x Rx Buffers on ring%d\n",
dev->name, buf_cnt, i);
}
}
/**
* s2io_poll - Rx interrupt handler for NAPI support
* @dev : pointer to the device structure.
* @budget : The number of packets that were budgeted to be processed
* during one pass through the 'Poll" function.
* Description:
* Comes into picture only if NAPI support has been incorporated. It does
* the same thing that rx_intr_handler does, but not in a interrupt context
* also It will process only a given number of packets.
* Return value:
* 0 on success and 1 if there are No Rx packets to be processed.
*/
static int s2io_poll(struct net_device *dev, int *budget)
{
struct s2io_nic *nic = dev->priv;
int pkt_cnt = 0, org_pkts_to_process;
struct mac_info *mac_control;
struct config_param *config;
struct XENA_dev_config __iomem *bar0 = nic->bar0;
int i;
atomic_inc(&nic->isr_cnt);
mac_control = &nic->mac_control;
config = &nic->config;
nic->pkts_to_process = *budget;
if (nic->pkts_to_process > dev->quota)
nic->pkts_to_process = dev->quota;
org_pkts_to_process = nic->pkts_to_process;
writeq(S2IO_MINUS_ONE, &bar0->rx_traffic_int);
readl(&bar0->rx_traffic_int);
for (i = 0; i < config->rx_ring_num; i++) {
rx_intr_handler(&mac_control->rings[i]);
pkt_cnt = org_pkts_to_process - nic->pkts_to_process;
if (!nic->pkts_to_process) {
/* Quota for the current iteration has been met */
goto no_rx;
}
}
if (!pkt_cnt)
pkt_cnt = 1;
dev->quota -= pkt_cnt;
*budget -= pkt_cnt;
netif_rx_complete(dev);
for (i = 0; i < config->rx_ring_num; i++) {
if (fill_rx_buffers(nic, i) == -ENOMEM) {
DBG_PRINT(INFO_DBG, "%s:Out of memory", dev->name);
DBG_PRINT(INFO_DBG, " in Rx Poll!!\n");
break;
}
}
/* Re enable the Rx interrupts. */
writeq(0x0, &bar0->rx_traffic_mask);
readl(&bar0->rx_traffic_mask);
atomic_dec(&nic->isr_cnt);
return 0;
no_rx:
dev->quota -= pkt_cnt;
*budget -= pkt_cnt;
for (i = 0; i < config->rx_ring_num; i++) {
if (fill_rx_buffers(nic, i) == -ENOMEM) {
DBG_PRINT(INFO_DBG, "%s:Out of memory", dev->name);
DBG_PRINT(INFO_DBG, " in Rx Poll!!\n");
break;
}
}
atomic_dec(&nic->isr_cnt);
return 1;
}
#ifdef CONFIG_NET_POLL_CONTROLLER
/**
* s2io_netpoll - netpoll event handler entry point
* @dev : pointer to the device structure.
* Description:
* This function will be called by upper layer to check for events on the
* interface in situations where interrupts are disabled. It is used for
* specific in-kernel networking tasks, such as remote consoles and kernel
* debugging over the network (example netdump in RedHat).
*/
static void s2io_netpoll(struct net_device *dev)
{
struct s2io_nic *nic = dev->priv;
struct mac_info *mac_control;
struct config_param *config;
struct XENA_dev_config __iomem *bar0 = nic->bar0;
u64 val64 = 0xFFFFFFFFFFFFFFFFULL;
int i;
disable_irq(dev->irq);
atomic_inc(&nic->isr_cnt);
mac_control = &nic->mac_control;
config = &nic->config;
writeq(val64, &bar0->rx_traffic_int);
writeq(val64, &bar0->tx_traffic_int);
/* we need to free up the transmitted skbufs or else netpoll will
* run out of skbs and will fail and eventually netpoll application such
* as netdump will fail.
*/
for (i = 0; i < config->tx_fifo_num; i++)
tx_intr_handler(&mac_control->fifos[i]);
/* check for received packet and indicate up to network */
for (i = 0; i < config->rx_ring_num; i++)
rx_intr_handler(&mac_control->rings[i]);
for (i = 0; i < config->rx_ring_num; i++) {
if (fill_rx_buffers(nic, i) == -ENOMEM) {
DBG_PRINT(INFO_DBG, "%s:Out of memory", dev->name);
DBG_PRINT(INFO_DBG, " in Rx Netpoll!!\n");
break;
}
}
atomic_dec(&nic->isr_cnt);
enable_irq(dev->irq);
return;
}
#endif
/**
* rx_intr_handler - Rx interrupt handler
* @nic: device private variable.
* Description:
* If the interrupt is because of a received frame or if the
* receive ring contains fresh as yet un-processed frames,this function is
* called. It picks out the RxD at which place the last Rx processing had
* stopped and sends the skb to the OSM's Rx handler and then increments
* the offset.
* Return Value:
* NONE.
*/
static void rx_intr_handler(struct ring_info *ring_data)
{
struct s2io_nic *nic = ring_data->nic;
struct net_device *dev = (struct net_device *) nic->dev;
int get_block, put_block, put_offset;
struct rx_curr_get_info get_info, put_info;
struct RxD_t *rxdp;
struct sk_buff *skb;
int pkt_cnt = 0;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
int i;
spin_lock(&nic->rx_lock);
if (atomic_read(&nic->card_state) == CARD_DOWN) {
DBG_PRINT(INTR_DBG, "%s: %s going down for reset\n",
__FUNCTION__, dev->name);
spin_unlock(&nic->rx_lock);
return;
}
get_info = ring_data->rx_curr_get_info;
get_block = get_info.block_index;
memcpy(&put_info, &ring_data->rx_curr_put_info, sizeof(put_info));
put_block = put_info.block_index;
rxdp = ring_data->rx_blocks[get_block].rxds[get_info.offset].virt_addr;
if (!napi) {
spin_lock(&nic->put_lock);
put_offset = ring_data->put_pos;
spin_unlock(&nic->put_lock);
} else
put_offset = ring_data->put_pos;
while (RXD_IS_UP2DT(rxdp)) {
/*
* If your are next to put index then it's
* FIFO full condition
*/
if ((get_block == put_block) &&
(get_info.offset + 1) == put_info.offset) {
DBG_PRINT(INTR_DBG, "%s: Ring Full\n",dev->name);
break;
}
skb = (struct sk_buff *) ((unsigned long)rxdp->Host_Control);
if (skb == NULL) {
DBG_PRINT(ERR_DBG, "%s: The skb is ",
dev->name);
DBG_PRINT(ERR_DBG, "Null in Rx Intr\n");
spin_unlock(&nic->rx_lock);
return;
}
if (nic->rxd_mode == RXD_MODE_1) {
pci_unmap_single(nic->pdev, (dma_addr_t)
((struct RxD1*)rxdp)->Buffer0_ptr,
dev->mtu +
HEADER_ETHERNET_II_802_3_SIZE +
HEADER_802_2_SIZE +
HEADER_SNAP_SIZE,
PCI_DMA_FROMDEVICE);
} else if (nic->rxd_mode == RXD_MODE_3B) {
pci_dma_sync_single_for_cpu(nic->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer0_ptr,
BUF0_LEN, PCI_DMA_FROMDEVICE);
pci_unmap_single(nic->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer2_ptr,
dev->mtu + 4,
PCI_DMA_FROMDEVICE);
} else {
pci_dma_sync_single_for_cpu(nic->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer0_ptr, BUF0_LEN,
PCI_DMA_FROMDEVICE);
pci_unmap_single(nic->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer1_ptr,
l3l4hdr_size + 4,
PCI_DMA_FROMDEVICE);
pci_unmap_single(nic->pdev, (dma_addr_t)
((struct RxD3*)rxdp)->Buffer2_ptr,
dev->mtu, PCI_DMA_FROMDEVICE);
}
prefetch(skb->data);
rx_osm_handler(ring_data, rxdp);
get_info.offset++;
ring_data->rx_curr_get_info.offset = get_info.offset;
rxdp = ring_data->rx_blocks[get_block].
rxds[get_info.offset].virt_addr;
if (get_info.offset == rxd_count[nic->rxd_mode]) {
get_info.offset = 0;
ring_data->rx_curr_get_info.offset = get_info.offset;
get_block++;
if (get_block == ring_data->block_count)
get_block = 0;
ring_data->rx_curr_get_info.block_index = get_block;
rxdp = ring_data->rx_blocks[get_block].block_virt_addr;
}
nic->pkts_to_process -= 1;
if ((napi) && (!nic->pkts_to_process))
break;
pkt_cnt++;
if ((indicate_max_pkts) && (pkt_cnt > indicate_max_pkts))
break;
}
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (nic->lro) {
/* Clear all LRO sessions before exiting */
for (i=0; i<MAX_LRO_SESSIONS; i++) {
struct lro *lro = &nic->lro0_n[i];
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (lro->in_use) {
update_L3L4_header(nic, lro);
queue_rx_frame(lro->parent);
clear_lro_session(lro);
}
}
}
spin_unlock(&nic->rx_lock);
}
/**
* tx_intr_handler - Transmit interrupt handler
* @nic : device private variable
* Description:
* If an interrupt was raised to indicate DMA complete of the
* Tx packet, this function is called. It identifies the last TxD
* whose buffer was freed and frees all skbs whose data have already
* DMA'ed into the NICs internal memory.
* Return Value:
* NONE
*/
static void tx_intr_handler(struct fifo_info *fifo_data)
{
struct s2io_nic *nic = fifo_data->nic;
struct net_device *dev = (struct net_device *) nic->dev;
struct tx_curr_get_info get_info, put_info;
struct sk_buff *skb;
struct TxD *txdlp;
get_info = fifo_data->tx_curr_get_info;
memcpy(&put_info, &fifo_data->tx_curr_put_info, sizeof(put_info));
txdlp = (struct TxD *) fifo_data->list_info[get_info.offset].
list_virt_addr;
while ((!(txdlp->Control_1 & TXD_LIST_OWN_XENA)) &&
(get_info.offset != put_info.offset) &&
(txdlp->Host_Control)) {
/* Check for TxD errors */
if (txdlp->Control_1 & TXD_T_CODE) {
unsigned long long err;
err = txdlp->Control_1 & TXD_T_CODE;
if (err & 0x1) {
nic->mac_control.stats_info->sw_stat.
parity_err_cnt++;
}
if ((err >> 48) == 0xA) {
DBG_PRINT(TX_DBG, "TxD returned due \
to loss of link\n");
}
else {
DBG_PRINT(ERR_DBG, "***TxD error %llx\n", err);
}
}
skb = s2io_txdl_getskb(fifo_data, txdlp, get_info.offset);
if (skb == NULL) {
DBG_PRINT(ERR_DBG, "%s: Null skb ",
__FUNCTION__);
DBG_PRINT(ERR_DBG, "in Tx Free Intr\n");
return;
}
/* Updating the statistics block */
nic->stats.tx_bytes += skb->len;
dev_kfree_skb_irq(skb);
get_info.offset++;
if (get_info.offset == get_info.fifo_len + 1)
get_info.offset = 0;
txdlp = (struct TxD *) fifo_data->list_info
[get_info.offset].list_virt_addr;
fifo_data->tx_curr_get_info.offset =
get_info.offset;
}
spin_lock(&nic->tx_lock);
if (netif_queue_stopped(dev))
netif_wake_queue(dev);
spin_unlock(&nic->tx_lock);
}
/**
* s2io_mdio_write - Function to write in to MDIO registers
* @mmd_type : MMD type value (PMA/PMD/WIS/PCS/PHYXS)
* @addr : address value
* @value : data value
* @dev : pointer to net_device structure
* Description:
* This function is used to write values to the MDIO registers
* NONE
*/
static void s2io_mdio_write(u32 mmd_type, u64 addr, u16 value, struct net_device *dev)
{
u64 val64 = 0x0;
struct s2io_nic *sp = dev->priv;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
//address transaction
val64 = val64 | MDIO_MMD_INDX_ADDR(addr)
| MDIO_MMD_DEV_ADDR(mmd_type)
| MDIO_MMS_PRT_ADDR(0x0);
writeq(val64, &bar0->mdio_control);
val64 = val64 | MDIO_CTRL_START_TRANS(0xE);
writeq(val64, &bar0->mdio_control);
udelay(100);
//Data transaction
val64 = 0x0;
val64 = val64 | MDIO_MMD_INDX_ADDR(addr)
| MDIO_MMD_DEV_ADDR(mmd_type)
| MDIO_MMS_PRT_ADDR(0x0)
| MDIO_MDIO_DATA(value)
| MDIO_OP(MDIO_OP_WRITE_TRANS);
writeq(val64, &bar0->mdio_control);
val64 = val64 | MDIO_CTRL_START_TRANS(0xE);
writeq(val64, &bar0->mdio_control);
udelay(100);
val64 = 0x0;
val64 = val64 | MDIO_MMD_INDX_ADDR(addr)
| MDIO_MMD_DEV_ADDR(mmd_type)
| MDIO_MMS_PRT_ADDR(0x0)
| MDIO_OP(MDIO_OP_READ_TRANS);
writeq(val64, &bar0->mdio_control);
val64 = val64 | MDIO_CTRL_START_TRANS(0xE);
writeq(val64, &bar0->mdio_control);
udelay(100);
}
/**
* s2io_mdio_read - Function to write in to MDIO registers
* @mmd_type : MMD type value (PMA/PMD/WIS/PCS/PHYXS)
* @addr : address value
* @dev : pointer to net_device structure
* Description:
* This function is used to read values to the MDIO registers
* NONE
*/
static u64 s2io_mdio_read(u32 mmd_type, u64 addr, struct net_device *dev)
{
u64 val64 = 0x0;
u64 rval64 = 0x0;
struct s2io_nic *sp = dev->priv;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
/* address transaction */
val64 = val64 | MDIO_MMD_INDX_ADDR(addr)
| MDIO_MMD_DEV_ADDR(mmd_type)
| MDIO_MMS_PRT_ADDR(0x0);
writeq(val64, &bar0->mdio_control);
val64 = val64 | MDIO_CTRL_START_TRANS(0xE);
writeq(val64, &bar0->mdio_control);
udelay(100);
/* Data transaction */
val64 = 0x0;
val64 = val64 | MDIO_MMD_INDX_ADDR(addr)
| MDIO_MMD_DEV_ADDR(mmd_type)
| MDIO_MMS_PRT_ADDR(0x0)
| MDIO_OP(MDIO_OP_READ_TRANS);
writeq(val64, &bar0->mdio_control);
val64 = val64 | MDIO_CTRL_START_TRANS(0xE);
writeq(val64, &bar0->mdio_control);
udelay(100);
/* Read the value from regs */
rval64 = readq(&bar0->mdio_control);
rval64 = rval64 & 0xFFFF0000;
rval64 = rval64 >> 16;
return rval64;
}
/**
* s2io_chk_xpak_counter - Function to check the status of the xpak counters
* @counter : couter value to be updated
* @flag : flag to indicate the status
* @type : counter type
* Description:
* This function is to check the status of the xpak counters value
* NONE
*/
static void s2io_chk_xpak_counter(u64 *counter, u64 * regs_stat, u32 index, u16 flag, u16 type)
{
u64 mask = 0x3;
u64 val64;
int i;
for(i = 0; i <index; i++)
mask = mask << 0x2;
if(flag > 0)
{
*counter = *counter + 1;
val64 = *regs_stat & mask;
val64 = val64 >> (index * 0x2);
val64 = val64 + 1;
if(val64 == 3)
{
switch(type)
{
case 1:
DBG_PRINT(ERR_DBG, "Take Xframe NIC out of "
"service. Excessive temperatures may "
"result in premature transceiver "
"failure \n");
break;
case 2:
DBG_PRINT(ERR_DBG, "Take Xframe NIC out of "
"service Excessive bias currents may "
"indicate imminent laser diode "
"failure \n");
break;
case 3:
DBG_PRINT(ERR_DBG, "Take Xframe NIC out of "
"service Excessive laser output "
"power may saturate far-end "
"receiver\n");
break;
default:
DBG_PRINT(ERR_DBG, "Incorrect XPAK Alarm "
"type \n");
}
val64 = 0x0;
}
val64 = val64 << (index * 0x2);
*regs_stat = (*regs_stat & (~mask)) | (val64);
} else {
*regs_stat = *regs_stat & (~mask);
}
}
/**
* s2io_updt_xpak_counter - Function to update the xpak counters
* @dev : pointer to net_device struct
* Description:
* This function is to upate the status of the xpak counters value
* NONE
*/
static void s2io_updt_xpak_counter(struct net_device *dev)
{
u16 flag = 0x0;
u16 type = 0x0;
u16 val16 = 0x0;
u64 val64 = 0x0;
u64 addr = 0x0;
struct s2io_nic *sp = dev->priv;
struct stat_block *stat_info = sp->mac_control.stats_info;
/* Check the communication with the MDIO slave */
addr = 0x0000;
val64 = 0x0;
val64 = s2io_mdio_read(MDIO_MMD_PMA_DEV_ADDR, addr, dev);
if((val64 == 0xFFFF) || (val64 == 0x0000))
{
DBG_PRINT(ERR_DBG, "ERR: MDIO slave access failed - "
"Returned %llx\n", (unsigned long long)val64);
return;
}
/* Check for the expecte value of 2040 at PMA address 0x0000 */
if(val64 != 0x2040)
{
DBG_PRINT(ERR_DBG, "Incorrect value at PMA address 0x0000 - ");
DBG_PRINT(ERR_DBG, "Returned: %llx- Expected: 0x2040\n",
(unsigned long long)val64);
return;
}
/* Loading the DOM register to MDIO register */
addr = 0xA100;
s2io_mdio_write(MDIO_MMD_PMA_DEV_ADDR, addr, val16, dev);
val64 = s2io_mdio_read(MDIO_MMD_PMA_DEV_ADDR, addr, dev);
/* Reading the Alarm flags */
addr = 0xA070;
val64 = 0x0;
val64 = s2io_mdio_read(MDIO_MMD_PMA_DEV_ADDR, addr, dev);
flag = CHECKBIT(val64, 0x7);
type = 1;
s2io_chk_xpak_counter(&stat_info->xpak_stat.alarm_transceiver_temp_high,
&stat_info->xpak_stat.xpak_regs_stat,
0x0, flag, type);
if(CHECKBIT(val64, 0x6))
stat_info->xpak_stat.alarm_transceiver_temp_low++;
flag = CHECKBIT(val64, 0x3);
type = 2;
s2io_chk_xpak_counter(&stat_info->xpak_stat.alarm_laser_bias_current_high,
&stat_info->xpak_stat.xpak_regs_stat,
0x2, flag, type);
if(CHECKBIT(val64, 0x2))
stat_info->xpak_stat.alarm_laser_bias_current_low++;
flag = CHECKBIT(val64, 0x1);
type = 3;
s2io_chk_xpak_counter(&stat_info->xpak_stat.alarm_laser_output_power_high,
&stat_info->xpak_stat.xpak_regs_stat,
0x4, flag, type);
if(CHECKBIT(val64, 0x0))
stat_info->xpak_stat.alarm_laser_output_power_low++;
/* Reading the Warning flags */
addr = 0xA074;
val64 = 0x0;
val64 = s2io_mdio_read(MDIO_MMD_PMA_DEV_ADDR, addr, dev);
if(CHECKBIT(val64, 0x7))
stat_info->xpak_stat.warn_transceiver_temp_high++;
if(CHECKBIT(val64, 0x6))
stat_info->xpak_stat.warn_transceiver_temp_low++;
if(CHECKBIT(val64, 0x3))
stat_info->xpak_stat.warn_laser_bias_current_high++;
if(CHECKBIT(val64, 0x2))
stat_info->xpak_stat.warn_laser_bias_current_low++;
if(CHECKBIT(val64, 0x1))
stat_info->xpak_stat.warn_laser_output_power_high++;
if(CHECKBIT(val64, 0x0))
stat_info->xpak_stat.warn_laser_output_power_low++;
}
/**
* alarm_intr_handler - Alarm Interrrupt handler
* @nic: device private variable
* Description: If the interrupt was neither because of Rx packet or Tx
* complete, this function is called. If the interrupt was to indicate
* a loss of link, the OSM link status handler is invoked for any other
* alarm interrupt the block that raised the interrupt is displayed
* and a H/W reset is issued.
* Return Value:
* NONE
*/
static void alarm_intr_handler(struct s2io_nic *nic)
{
struct net_device *dev = (struct net_device *) nic->dev;
struct XENA_dev_config __iomem *bar0 = nic->bar0;
register u64 val64 = 0, err_reg = 0;
u64 cnt;
int i;
if (atomic_read(&nic->card_state) == CARD_DOWN)
return;
nic->mac_control.stats_info->sw_stat.ring_full_cnt = 0;
/* Handling the XPAK counters update */
if(nic->mac_control.stats_info->xpak_stat.xpak_timer_count < 72000) {
/* waiting for an hour */
nic->mac_control.stats_info->xpak_stat.xpak_timer_count++;
} else {
s2io_updt_xpak_counter(dev);
/* reset the count to zero */
nic->mac_control.stats_info->xpak_stat.xpak_timer_count = 0;
}
/* Handling link status change error Intr */
if (s2io_link_fault_indication(nic) == MAC_RMAC_ERR_TIMER) {
err_reg = readq(&bar0->mac_rmac_err_reg);
writeq(err_reg, &bar0->mac_rmac_err_reg);
if (err_reg & RMAC_LINK_STATE_CHANGE_INT) {
schedule_work(&nic->set_link_task);
}
}
/* Handling Ecc errors */
val64 = readq(&bar0->mc_err_reg);
writeq(val64, &bar0->mc_err_reg);
if (val64 & (MC_ERR_REG_ECC_ALL_SNG | MC_ERR_REG_ECC_ALL_DBL)) {
if (val64 & MC_ERR_REG_ECC_ALL_DBL) {
nic->mac_control.stats_info->sw_stat.
double_ecc_errs++;
DBG_PRINT(INIT_DBG, "%s: Device indicates ",
dev->name);
DBG_PRINT(INIT_DBG, "double ECC error!!\n");
if (nic->device_type != XFRAME_II_DEVICE) {
/* Reset XframeI only if critical error */
if (val64 & (MC_ERR_REG_MIRI_ECC_DB_ERR_0 |
MC_ERR_REG_MIRI_ECC_DB_ERR_1)) {
netif_stop_queue(dev);
schedule_work(&nic->rst_timer_task);
nic->mac_control.stats_info->sw_stat.
soft_reset_cnt++;
}
}
} else {
nic->mac_control.stats_info->sw_stat.
single_ecc_errs++;
}
}
/* In case of a serious error, the device will be Reset. */
val64 = readq(&bar0->serr_source);
if (val64 & SERR_SOURCE_ANY) {
nic->mac_control.stats_info->sw_stat.serious_err_cnt++;
DBG_PRINT(ERR_DBG, "%s: Device indicates ", dev->name);
DBG_PRINT(ERR_DBG, "serious error %llx!!\n",
(unsigned long long)val64);
netif_stop_queue(dev);
schedule_work(&nic->rst_timer_task);
nic->mac_control.stats_info->sw_stat.soft_reset_cnt++;
}
/*
* Also as mentioned in the latest Errata sheets if the PCC_FB_ECC
* Error occurs, the adapter will be recycled by disabling the
* adapter enable bit and enabling it again after the device
* becomes Quiescent.
*/
val64 = readq(&bar0->pcc_err_reg);
writeq(val64, &bar0->pcc_err_reg);
if (val64 & PCC_FB_ECC_DB_ERR) {
u64 ac = readq(&bar0->adapter_control);
ac &= ~(ADAPTER_CNTL_EN);
writeq(ac, &bar0->adapter_control);
ac = readq(&bar0->adapter_control);
schedule_work(&nic->set_link_task);
}
/* Check for data parity error */
val64 = readq(&bar0->pic_int_status);
if (val64 & PIC_INT_GPIO) {
val64 = readq(&bar0->gpio_int_reg);
if (val64 & GPIO_INT_REG_DP_ERR_INT) {
nic->mac_control.stats_info->sw_stat.parity_err_cnt++;
schedule_work(&nic->rst_timer_task);
nic->mac_control.stats_info->sw_stat.soft_reset_cnt++;
}
}
/* Check for ring full counter */
if (nic->device_type & XFRAME_II_DEVICE) {
val64 = readq(&bar0->ring_bump_counter1);
for (i=0; i<4; i++) {
cnt = ( val64 & vBIT(0xFFFF,(i*16),16));
cnt >>= 64 - ((i+1)*16);
nic->mac_control.stats_info->sw_stat.ring_full_cnt
+= cnt;
}
val64 = readq(&bar0->ring_bump_counter2);
for (i=0; i<4; i++) {
cnt = ( val64 & vBIT(0xFFFF,(i*16),16));
cnt >>= 64 - ((i+1)*16);
nic->mac_control.stats_info->sw_stat.ring_full_cnt
+= cnt;
}
}
/* Other type of interrupts are not being handled now, TODO */
}
/**
* wait_for_cmd_complete - waits for a command to complete.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* Description: Function that waits for a command to Write into RMAC
* ADDR DATA registers to be completed and returns either success or
* error depending on whether the command was complete or not.
* Return value:
* SUCCESS on success and FAILURE on failure.
*/
static int wait_for_cmd_complete(void __iomem *addr, u64 busy_bit,
int bit_state)
{
int ret = FAILURE, cnt = 0, delay = 1;
u64 val64;
if ((bit_state != S2IO_BIT_RESET) && (bit_state != S2IO_BIT_SET))
return FAILURE;
do {
val64 = readq(addr);
if (bit_state == S2IO_BIT_RESET) {
if (!(val64 & busy_bit)) {
ret = SUCCESS;
break;
}
} else {
if (!(val64 & busy_bit)) {
ret = SUCCESS;
break;
}
}
if(in_interrupt())
mdelay(delay);
else
msleep(delay);
if (++cnt >= 10)
delay = 50;
} while (cnt < 20);
return ret;
}
/*
* check_pci_device_id - Checks if the device id is supported
* @id : device id
* Description: Function to check if the pci device id is supported by driver.
* Return value: Actual device id if supported else PCI_ANY_ID
*/
static u16 check_pci_device_id(u16 id)
{
switch (id) {
case PCI_DEVICE_ID_HERC_WIN:
case PCI_DEVICE_ID_HERC_UNI:
return XFRAME_II_DEVICE;
case PCI_DEVICE_ID_S2IO_UNI:
case PCI_DEVICE_ID_S2IO_WIN:
return XFRAME_I_DEVICE;
default:
return PCI_ANY_ID;
}
}
/**
* s2io_reset - Resets the card.
* @sp : private member of the device structure.
* Description: Function to Reset the card. This function then also
* restores the previously saved PCI configuration space registers as
* the card reset also resets the configuration space.
* Return value:
* void.
*/
static void s2io_reset(struct s2io_nic * sp)
{
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64;
u16 subid, pci_cmd;
int i;
u16 val16;
unsigned long long reset_cnt = 0;
DBG_PRINT(INIT_DBG,"%s - Resetting XFrame card %s\n",
__FUNCTION__, sp->dev->name);
/* Back up the PCI-X CMD reg, dont want to lose MMRBC, OST settings */
pci_read_config_word(sp->pdev, PCIX_COMMAND_REGISTER, &(pci_cmd));
if (sp->device_type == XFRAME_II_DEVICE) {
int ret;
ret = pci_set_power_state(sp->pdev, 3);
if (!ret)
ret = pci_set_power_state(sp->pdev, 0);
else {
DBG_PRINT(ERR_DBG,"%s PME based SW_Reset failed!\n",
__FUNCTION__);
goto old_way;
}
msleep(20);
goto new_way;
}
old_way:
val64 = SW_RESET_ALL;
writeq(val64, &bar0->sw_reset);
new_way:
if (strstr(sp->product_name, "CX4")) {
msleep(750);
}
msleep(250);
for (i = 0; i < S2IO_MAX_PCI_CONFIG_SPACE_REINIT; i++) {
/* Restore the PCI state saved during initialization. */
pci_restore_state(sp->pdev);
pci_read_config_word(sp->pdev, 0x2, &val16);
if (check_pci_device_id(val16) != (u16)PCI_ANY_ID)
break;
msleep(200);
}
if (check_pci_device_id(val16) == (u16)PCI_ANY_ID) {
DBG_PRINT(ERR_DBG,"%s SW_Reset failed!\n", __FUNCTION__);
}
pci_write_config_word(sp->pdev, PCIX_COMMAND_REGISTER, pci_cmd);
s2io_init_pci(sp);
/* Set swapper to enable I/O register access */
s2io_set_swapper(sp);
/* Restore the MSIX table entries from local variables */
restore_xmsi_data(sp);
/* Clear certain PCI/PCI-X fields after reset */
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
if (sp->device_type == XFRAME_II_DEVICE) {
/* Clear "detected parity error" bit */
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
pci_write_config_word(sp->pdev, PCI_STATUS, 0x8000);
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
/* Clearing PCIX Ecc status register */
pci_write_config_dword(sp->pdev, 0x68, 0x7C);
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
/* Clearing PCI_STATUS error reflected here */
writeq(BIT(62), &bar0->txpic_int_reg);
}
/* Reset device statistics maintained by OS */
memset(&sp->stats, 0, sizeof (struct net_device_stats));
/* save reset count */
reset_cnt = sp->mac_control.stats_info->sw_stat.soft_reset_cnt;
memset(sp->mac_control.stats_info, 0, sizeof(struct stat_block));
/* restore reset count */
sp->mac_control.stats_info->sw_stat.soft_reset_cnt = reset_cnt;
/* SXE-002: Configure link and activity LED to turn it off */
subid = sp->pdev->subsystem_device;
if (((subid & 0xFF) >= 0x07) &&
(sp->device_type == XFRAME_I_DEVICE)) {
val64 = readq(&bar0->gpio_control);
val64 |= 0x0000800000000000ULL;
writeq(val64, &bar0->gpio_control);
val64 = 0x0411040400000000ULL;
writeq(val64, (void __iomem *)bar0 + 0x2700);
}
/*
* Clear spurious ECC interrupts that would have occured on
* XFRAME II cards after reset.
*/
if (sp->device_type == XFRAME_II_DEVICE) {
val64 = readq(&bar0->pcc_err_reg);
writeq(val64, &bar0->pcc_err_reg);
}
/* restore the previously assigned mac address */
s2io_set_mac_addr(sp->dev, (u8 *)&sp->def_mac_addr[0].mac_addr);
sp->device_enabled_once = FALSE;
}
/**
* s2io_set_swapper - to set the swapper controle on the card
* @sp : private member of the device structure,
* pointer to the s2io_nic structure.
* Description: Function to set the swapper control on the card
* correctly depending on the 'endianness' of the system.
* Return value:
* SUCCESS on success and FAILURE on failure.
*/
static int s2io_set_swapper(struct s2io_nic * sp)
{
struct net_device *dev = sp->dev;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64, valt, valr;
/*
* Set proper endian settings and verify the same by reading
* the PIF Feed-back register.
*/
val64 = readq(&bar0->pif_rd_swapper_fb);
if (val64 != 0x0123456789ABCDEFULL) {
int i = 0;
u64 value[] = { 0xC30000C3C30000C3ULL, /* FE=1, SE=1 */
0x8100008181000081ULL, /* FE=1, SE=0 */
0x4200004242000042ULL, /* FE=0, SE=1 */
0}; /* FE=0, SE=0 */
while(i<4) {
writeq(value[i], &bar0->swapper_ctrl);
val64 = readq(&bar0->pif_rd_swapper_fb);
if (val64 == 0x0123456789ABCDEFULL)
break;
i++;
}
if (i == 4) {
DBG_PRINT(ERR_DBG, "%s: Endian settings are wrong, ",
dev->name);
DBG_PRINT(ERR_DBG, "feedback read %llx\n",
(unsigned long long) val64);
return FAILURE;
}
valr = value[i];
} else {
valr = readq(&bar0->swapper_ctrl);
}
valt = 0x0123456789ABCDEFULL;
writeq(valt, &bar0->xmsi_address);
val64 = readq(&bar0->xmsi_address);
if(val64 != valt) {
int i = 0;
u64 value[] = { 0x00C3C30000C3C300ULL, /* FE=1, SE=1 */
0x0081810000818100ULL, /* FE=1, SE=0 */
0x0042420000424200ULL, /* FE=0, SE=1 */
0}; /* FE=0, SE=0 */
while(i<4) {
writeq((value[i] | valr), &bar0->swapper_ctrl);
writeq(valt, &bar0->xmsi_address);
val64 = readq(&bar0->xmsi_address);
if(val64 == valt)
break;
i++;
}
if(i == 4) {
unsigned long long x = val64;
DBG_PRINT(ERR_DBG, "Write failed, Xmsi_addr ");
DBG_PRINT(ERR_DBG, "reads:0x%llx\n", x);
return FAILURE;
}
}
val64 = readq(&bar0->swapper_ctrl);
val64 &= 0xFFFF000000000000ULL;
#ifdef __BIG_ENDIAN
/*
* The device by default set to a big endian format, so a
* big endian driver need not set anything.
*/
val64 |= (SWAPPER_CTRL_TXP_FE |
SWAPPER_CTRL_TXP_SE |
SWAPPER_CTRL_TXD_R_FE |
SWAPPER_CTRL_TXD_W_FE |
SWAPPER_CTRL_TXF_R_FE |
SWAPPER_CTRL_RXD_R_FE |
SWAPPER_CTRL_RXD_W_FE |
SWAPPER_CTRL_RXF_W_FE |
SWAPPER_CTRL_XMSI_FE |
SWAPPER_CTRL_STATS_FE | SWAPPER_CTRL_STATS_SE);
if (sp->intr_type == INTA)
val64 |= SWAPPER_CTRL_XMSI_SE;
writeq(val64, &bar0->swapper_ctrl);
#else
/*
* Initially we enable all bits to make it accessible by the
* driver, then we selectively enable only those bits that
* we want to set.
*/
val64 |= (SWAPPER_CTRL_TXP_FE |
SWAPPER_CTRL_TXP_SE |
SWAPPER_CTRL_TXD_R_FE |
SWAPPER_CTRL_TXD_R_SE |
SWAPPER_CTRL_TXD_W_FE |
SWAPPER_CTRL_TXD_W_SE |
SWAPPER_CTRL_TXF_R_FE |
SWAPPER_CTRL_RXD_R_FE |
SWAPPER_CTRL_RXD_R_SE |
SWAPPER_CTRL_RXD_W_FE |
SWAPPER_CTRL_RXD_W_SE |
SWAPPER_CTRL_RXF_W_FE |
SWAPPER_CTRL_XMSI_FE |
SWAPPER_CTRL_STATS_FE | SWAPPER_CTRL_STATS_SE);
if (sp->intr_type == INTA)
val64 |= SWAPPER_CTRL_XMSI_SE;
writeq(val64, &bar0->swapper_ctrl);
#endif
val64 = readq(&bar0->swapper_ctrl);
/*
* Verifying if endian settings are accurate by reading a
* feedback register.
*/
val64 = readq(&bar0->pif_rd_swapper_fb);
if (val64 != 0x0123456789ABCDEFULL) {
/* Endian settings are incorrect, calls for another dekko. */
DBG_PRINT(ERR_DBG, "%s: Endian settings are wrong, ",
dev->name);
DBG_PRINT(ERR_DBG, "feedback read %llx\n",
(unsigned long long) val64);
return FAILURE;
}
return SUCCESS;
}
static int wait_for_msix_trans(struct s2io_nic *nic, int i)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
u64 val64;
int ret = 0, cnt = 0;
do {
val64 = readq(&bar0->xmsi_access);
if (!(val64 & BIT(15)))
break;
mdelay(1);
cnt++;
} while(cnt < 5);
if (cnt == 5) {
DBG_PRINT(ERR_DBG, "XMSI # %d Access failed\n", i);
ret = 1;
}
return ret;
}
static void restore_xmsi_data(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
u64 val64;
int i;
for (i=0; i < MAX_REQUESTED_MSI_X; i++) {
writeq(nic->msix_info[i].addr, &bar0->xmsi_address);
writeq(nic->msix_info[i].data, &bar0->xmsi_data);
val64 = (BIT(7) | BIT(15) | vBIT(i, 26, 6));
writeq(val64, &bar0->xmsi_access);
if (wait_for_msix_trans(nic, i)) {
DBG_PRINT(ERR_DBG, "failed in %s\n", __FUNCTION__);
continue;
}
}
}
static void store_xmsi_data(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
u64 val64, addr, data;
int i;
/* Store and display */
for (i=0; i < MAX_REQUESTED_MSI_X; i++) {
val64 = (BIT(15) | vBIT(i, 26, 6));
writeq(val64, &bar0->xmsi_access);
if (wait_for_msix_trans(nic, i)) {
DBG_PRINT(ERR_DBG, "failed in %s\n", __FUNCTION__);
continue;
}
addr = readq(&bar0->xmsi_address);
data = readq(&bar0->xmsi_data);
if (addr && data) {
nic->msix_info[i].addr = addr;
nic->msix_info[i].data = data;
}
}
}
int s2io_enable_msi(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
u16 msi_ctrl, msg_val;
struct config_param *config = &nic->config;
struct net_device *dev = nic->dev;
u64 val64, tx_mat, rx_mat;
int i, err;
val64 = readq(&bar0->pic_control);
val64 &= ~BIT(1);
writeq(val64, &bar0->pic_control);
err = pci_enable_msi(nic->pdev);
if (err) {
DBG_PRINT(ERR_DBG, "%s: enabling MSI failed\n",
nic->dev->name);
return err;
}
/*
* Enable MSI and use MSI-1 in stead of the standard MSI-0
* for interrupt handling.
*/
pci_read_config_word(nic->pdev, 0x4c, &msg_val);
msg_val ^= 0x1;
pci_write_config_word(nic->pdev, 0x4c, msg_val);
pci_read_config_word(nic->pdev, 0x4c, &msg_val);
pci_read_config_word(nic->pdev, 0x42, &msi_ctrl);
msi_ctrl |= 0x10;
pci_write_config_word(nic->pdev, 0x42, msi_ctrl);
/* program MSI-1 into all usable Tx_Mat and Rx_Mat fields */
tx_mat = readq(&bar0->tx_mat0_n[0]);
for (i=0; i<config->tx_fifo_num; i++) {
tx_mat |= TX_MAT_SET(i, 1);
}
writeq(tx_mat, &bar0->tx_mat0_n[0]);
rx_mat = readq(&bar0->rx_mat);
for (i=0; i<config->rx_ring_num; i++) {
rx_mat |= RX_MAT_SET(i, 1);
}
writeq(rx_mat, &bar0->rx_mat);
dev->irq = nic->pdev->irq;
return 0;
}
static int s2io_enable_msi_x(struct s2io_nic *nic)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
u64 tx_mat, rx_mat;
u16 msi_control; /* Temp variable */
int ret, i, j, msix_indx = 1;
nic->entries = kmalloc(MAX_REQUESTED_MSI_X * sizeof(struct msix_entry),
GFP_KERNEL);
if (nic->entries == NULL) {
DBG_PRINT(INFO_DBG, "%s: Memory allocation failed\n", __FUNCTION__);
return -ENOMEM;
}
memset(nic->entries, 0, MAX_REQUESTED_MSI_X * sizeof(struct msix_entry));
nic->s2io_entries =
kmalloc(MAX_REQUESTED_MSI_X * sizeof(struct s2io_msix_entry),
GFP_KERNEL);
if (nic->s2io_entries == NULL) {
DBG_PRINT(INFO_DBG, "%s: Memory allocation failed\n", __FUNCTION__);
kfree(nic->entries);
return -ENOMEM;
}
memset(nic->s2io_entries, 0,
MAX_REQUESTED_MSI_X * sizeof(struct s2io_msix_entry));
for (i=0; i< MAX_REQUESTED_MSI_X; i++) {
nic->entries[i].entry = i;
nic->s2io_entries[i].entry = i;
nic->s2io_entries[i].arg = NULL;
nic->s2io_entries[i].in_use = 0;
}
tx_mat = readq(&bar0->tx_mat0_n[0]);
for (i=0; i<nic->config.tx_fifo_num; i++, msix_indx++) {
tx_mat |= TX_MAT_SET(i, msix_indx);
nic->s2io_entries[msix_indx].arg = &nic->mac_control.fifos[i];
nic->s2io_entries[msix_indx].type = MSIX_FIFO_TYPE;
nic->s2io_entries[msix_indx].in_use = MSIX_FLG;
}
writeq(tx_mat, &bar0->tx_mat0_n[0]);
if (!nic->config.bimodal) {
rx_mat = readq(&bar0->rx_mat);
for (j=0; j<nic->config.rx_ring_num; j++, msix_indx++) {
rx_mat |= RX_MAT_SET(j, msix_indx);
nic->s2io_entries[msix_indx].arg = &nic->mac_control.rings[j];
nic->s2io_entries[msix_indx].type = MSIX_RING_TYPE;
nic->s2io_entries[msix_indx].in_use = MSIX_FLG;
}
writeq(rx_mat, &bar0->rx_mat);
} else {
tx_mat = readq(&bar0->tx_mat0_n[7]);
for (j=0; j<nic->config.rx_ring_num; j++, msix_indx++) {
tx_mat |= TX_MAT_SET(i, msix_indx);
nic->s2io_entries[msix_indx].arg = &nic->mac_control.rings[j];
nic->s2io_entries[msix_indx].type = MSIX_RING_TYPE;
nic->s2io_entries[msix_indx].in_use = MSIX_FLG;
}
writeq(tx_mat, &bar0->tx_mat0_n[7]);
}
nic->avail_msix_vectors = 0;
ret = pci_enable_msix(nic->pdev, nic->entries, MAX_REQUESTED_MSI_X);
/* We fail init if error or we get less vectors than min required */
if (ret >= (nic->config.tx_fifo_num + nic->config.rx_ring_num + 1)) {
nic->avail_msix_vectors = ret;
ret = pci_enable_msix(nic->pdev, nic->entries, ret);
}
if (ret) {
DBG_PRINT(ERR_DBG, "%s: Enabling MSIX failed\n", nic->dev->name);
kfree(nic->entries);
kfree(nic->s2io_entries);
nic->entries = NULL;
nic->s2io_entries = NULL;
nic->avail_msix_vectors = 0;
return -ENOMEM;
}
if (!nic->avail_msix_vectors)
nic->avail_msix_vectors = MAX_REQUESTED_MSI_X;
/*
* To enable MSI-X, MSI also needs to be enabled, due to a bug
* in the herc NIC. (Temp change, needs to be removed later)
*/
pci_read_config_word(nic->pdev, 0x42, &msi_control);
msi_control |= 0x1; /* Enable MSI */
pci_write_config_word(nic->pdev, 0x42, msi_control);
return 0;
}
/* ********************************************************* *
* Functions defined below concern the OS part of the driver *
* ********************************************************* */
/**
* s2io_open - open entry point of the driver
* @dev : pointer to the device structure.
* Description:
* This function is the open entry point of the driver. It mainly calls a
* function to allocate Rx buffers and inserts them into the buffer
* descriptors and then enables the Rx part of the NIC.
* Return value:
* 0 on success and an appropriate (-)ve integer as defined in errno.h
* file on failure.
*/
static int s2io_open(struct net_device *dev)
{
struct s2io_nic *sp = dev->priv;
int err = 0;
/*
* Make sure you have link off by default every time
* Nic is initialized
*/
netif_carrier_off(dev);
sp->last_link_state = 0;
/* Initialize H/W and enable interrupts */
err = s2io_card_up(sp);
if (err) {
DBG_PRINT(ERR_DBG, "%s: H/W initialization failed\n",
dev->name);
goto hw_init_failed;
}
if (s2io_set_mac_addr(dev, dev->dev_addr) == FAILURE) {
DBG_PRINT(ERR_DBG, "Set Mac Address Failed\n");
s2io_card_down(sp);
err = -ENODEV;
goto hw_init_failed;
}
netif_start_queue(dev);
return 0;
hw_init_failed:
if (sp->intr_type == MSI_X) {
if (sp->entries)
kfree(sp->entries);
if (sp->s2io_entries)
kfree(sp->s2io_entries);
}
return err;
}
/**
* s2io_close -close entry point of the driver
* @dev : device pointer.
* Description:
* This is the stop entry point of the driver. It needs to undo exactly
* whatever was done by the open entry point,thus it's usually referred to
* as the close function.Among other things this function mainly stops the
* Rx side of the NIC and frees all the Rx buffers in the Rx rings.
* Return value:
* 0 on success and an appropriate (-)ve integer as defined in errno.h
* file on failure.
*/
static int s2io_close(struct net_device *dev)
{
struct s2io_nic *sp = dev->priv;
netif_stop_queue(dev);
/* Reset card, kill tasklet and free Tx and Rx buffers. */
s2io_card_down(sp);
sp->device_close_flag = TRUE; /* Device is shut down. */
return 0;
}
/**
* s2io_xmit - Tx entry point of te driver
* @skb : the socket buffer containing the Tx data.
* @dev : device pointer.
* Description :
* This function is the Tx entry point of the driver. S2IO NIC supports
* certain protocol assist features on Tx side, namely CSO, S/G, LSO.
* NOTE: when device cant queue the pkt,just the trans_start variable will
* not be upadted.
* Return value:
* 0 on success & 1 on failure.
*/
static int s2io_xmit(struct sk_buff *skb, struct net_device *dev)
{
struct s2io_nic *sp = dev->priv;
u16 frg_cnt, frg_len, i, queue, queue_len, put_off, get_off;
register u64 val64;
struct TxD *txdp;
struct TxFIFO_element __iomem *tx_fifo;
unsigned long flags;
u16 vlan_tag = 0;
int vlan_priority = 0;
struct mac_info *mac_control;
struct config_param *config;
int offload_type;
mac_control = &sp->mac_control;
config = &sp->config;
DBG_PRINT(TX_DBG, "%s: In Neterion Tx routine\n", dev->name);
spin_lock_irqsave(&sp->tx_lock, flags);
if (atomic_read(&sp->card_state) == CARD_DOWN) {
DBG_PRINT(TX_DBG, "%s: Card going down for reset\n",
dev->name);
spin_unlock_irqrestore(&sp->tx_lock, flags);
dev_kfree_skb(skb);
return 0;
}
queue = 0;
/* Get Fifo number to Transmit based on vlan priority */
if (sp->vlgrp && vlan_tx_tag_present(skb)) {
vlan_tag = vlan_tx_tag_get(skb);
vlan_priority = vlan_tag >> 13;
queue = config->fifo_mapping[vlan_priority];
}
put_off = (u16) mac_control->fifos[queue].tx_curr_put_info.offset;
get_off = (u16) mac_control->fifos[queue].tx_curr_get_info.offset;
txdp = (struct TxD *) mac_control->fifos[queue].list_info[put_off].
list_virt_addr;
queue_len = mac_control->fifos[queue].tx_curr_put_info.fifo_len + 1;
/* Avoid "put" pointer going beyond "get" pointer */
if (txdp->Host_Control ||
((put_off+1) == queue_len ? 0 : (put_off+1)) == get_off) {
DBG_PRINT(TX_DBG, "Error in xmit, No free TXDs.\n");
netif_stop_queue(dev);
dev_kfree_skb(skb);
spin_unlock_irqrestore(&sp->tx_lock, flags);
return 0;
}
/* A buffer with no data will be dropped */
if (!skb->len) {
DBG_PRINT(TX_DBG, "%s:Buffer has no data..\n", dev->name);
dev_kfree_skb(skb);
spin_unlock_irqrestore(&sp->tx_lock, flags);
return 0;
}
offload_type = s2io_offload_type(skb);
if (offload_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)) {
txdp->Control_1 |= TXD_TCP_LSO_EN;
txdp->Control_1 |= TXD_TCP_LSO_MSS(s2io_tcp_mss(skb));
}
if (skb->ip_summed == CHECKSUM_PARTIAL) {
txdp->Control_2 |=
(TXD_TX_CKO_IPV4_EN | TXD_TX_CKO_TCP_EN |
TXD_TX_CKO_UDP_EN);
}
txdp->Control_1 |= TXD_GATHER_CODE_FIRST;
txdp->Control_1 |= TXD_LIST_OWN_XENA;
txdp->Control_2 |= config->tx_intr_type;
if (sp->vlgrp && vlan_tx_tag_present(skb)) {
txdp->Control_2 |= TXD_VLAN_ENABLE;
txdp->Control_2 |= TXD_VLAN_TAG(vlan_tag);
}
frg_len = skb->len - skb->data_len;
if (offload_type == SKB_GSO_UDP) {
int ufo_size;
ufo_size = s2io_udp_mss(skb);
ufo_size &= ~7;
txdp->Control_1 |= TXD_UFO_EN;
txdp->Control_1 |= TXD_UFO_MSS(ufo_size);
txdp->Control_1 |= TXD_BUFFER0_SIZE(8);
#ifdef __BIG_ENDIAN
sp->ufo_in_band_v[put_off] =
(u64)skb_shinfo(skb)->ip6_frag_id;
#else
sp->ufo_in_band_v[put_off] =
(u64)skb_shinfo(skb)->ip6_frag_id << 32;
#endif
txdp->Host_Control = (unsigned long)sp->ufo_in_band_v;
txdp->Buffer_Pointer = pci_map_single(sp->pdev,
sp->ufo_in_band_v,
sizeof(u64), PCI_DMA_TODEVICE);
txdp++;
}
txdp->Buffer_Pointer = pci_map_single
(sp->pdev, skb->data, frg_len, PCI_DMA_TODEVICE);
txdp->Host_Control = (unsigned long) skb;
txdp->Control_1 |= TXD_BUFFER0_SIZE(frg_len);
if (offload_type == SKB_GSO_UDP)
txdp->Control_1 |= TXD_UFO_EN;
frg_cnt = skb_shinfo(skb)->nr_frags;
/* For fragmented SKB. */
for (i = 0; i < frg_cnt; i++) {
skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
/* A '0' length fragment will be ignored */
if (!frag->size)
continue;
txdp++;
txdp->Buffer_Pointer = (u64) pci_map_page
(sp->pdev, frag->page, frag->page_offset,
frag->size, PCI_DMA_TODEVICE);
txdp->Control_1 = TXD_BUFFER0_SIZE(frag->size);
if (offload_type == SKB_GSO_UDP)
txdp->Control_1 |= TXD_UFO_EN;
}
txdp->Control_1 |= TXD_GATHER_CODE_LAST;
if (offload_type == SKB_GSO_UDP)
frg_cnt++; /* as Txd0 was used for inband header */
tx_fifo = mac_control->tx_FIFO_start[queue];
val64 = mac_control->fifos[queue].list_info[put_off].list_phy_addr;
writeq(val64, &tx_fifo->TxDL_Pointer);
val64 = (TX_FIFO_LAST_TXD_NUM(frg_cnt) | TX_FIFO_FIRST_LIST |
TX_FIFO_LAST_LIST);
if (offload_type)
val64 |= TX_FIFO_SPECIAL_FUNC;
writeq(val64, &tx_fifo->List_Control);
[PATCH] S2io: Errors found during review Hi, This is a patch to incorporate comments from earlier 12 patches. It also fixes a few issues we found during this time. Following is a list of changes in this patch. Item 1 incorporates earlier comments. Issues addressed in items 2 to 4 were discovered recently. 1. wmb() call in s2io_xmit() replaced with mmiowb(). 2. The dtx_control register was earlier programmed incorrectly for Xframe II adapter. 3. As suggested by hardware team, after a reset, in case of Xframe II adapter, we clear certain spurious errors by clearing PCI-X ECC status register, "detected parity error" bit in PCI_STATUS register and PCI_STATUS bit in txpic_int register. 4. On IBM PPC platforms, we found that in the Rx buffer replenish function, two memory writes(one to the the descriptor length and another to the ownership) were getting reordered. This was causing the adapter to see the ownership transfered to it before the length was updated. One solution was to add a wmb() but since this would turnout expensive on some platforms if called for every descriptor, we set the ownership bit and other fields of '2' to 'N' Rx descriptors followed by a wmb() and then set the ownership of first descriptor ('1'). Here the value 'N' is configurable by making it a module loadable parameter (rxsync_frequency). (NOTE: This parameter is a power of 2). 5. Bumped up the driver version no. to 2.0.2.1 Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Raghavendra Koushik <raghavendra.koushik@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2005-08-03 19:41:38 +00:00
mmiowb();
put_off++;
if (put_off == mac_control->fifos[queue].tx_curr_put_info.fifo_len + 1)
put_off = 0;
mac_control->fifos[queue].tx_curr_put_info.offset = put_off;
/* Avoid "put" pointer going beyond "get" pointer */
if (((put_off+1) == queue_len ? 0 : (put_off+1)) == get_off) {
sp->mac_control.stats_info->sw_stat.fifo_full_cnt++;
DBG_PRINT(TX_DBG,
"No free TxDs for xmit, Put: 0x%x Get:0x%x\n",
put_off, get_off);
netif_stop_queue(dev);
}
dev->trans_start = jiffies;
spin_unlock_irqrestore(&sp->tx_lock, flags);
return 0;
}
static void
s2io_alarm_handle(unsigned long data)
{
struct s2io_nic *sp = (struct s2io_nic *)data;
alarm_intr_handler(sp);
mod_timer(&sp->alarm_timer, jiffies + HZ / 2);
}
static int s2io_chk_rx_buffers(struct s2io_nic *sp, int rng_n)
{
int rxb_size, level;
if (!sp->lro) {
rxb_size = atomic_read(&sp->rx_bufs_left[rng_n]);
level = rx_buffer_level(sp, rxb_size, rng_n);
if ((level == PANIC) && (!TASKLET_IN_USE)) {
int ret;
DBG_PRINT(INTR_DBG, "%s: Rx BD hit ", __FUNCTION__);
DBG_PRINT(INTR_DBG, "PANIC levels\n");
if ((ret = fill_rx_buffers(sp, rng_n)) == -ENOMEM) {
DBG_PRINT(INFO_DBG, "Out of memory in %s",
__FUNCTION__);
clear_bit(0, (&sp->tasklet_status));
return -1;
}
clear_bit(0, (&sp->tasklet_status));
} else if (level == LOW)
tasklet_schedule(&sp->task);
} else if (fill_rx_buffers(sp, rng_n) == -ENOMEM) {
DBG_PRINT(INFO_DBG, "%s:Out of memory", sp->dev->name);
DBG_PRINT(INFO_DBG, " in Rx Intr!!\n");
}
return 0;
}
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
static irqreturn_t s2io_msi_handle(int irq, void *dev_id)
{
struct net_device *dev = (struct net_device *) dev_id;
struct s2io_nic *sp = dev->priv;
int i;
struct mac_info *mac_control;
struct config_param *config;
atomic_inc(&sp->isr_cnt);
mac_control = &sp->mac_control;
config = &sp->config;
DBG_PRINT(INTR_DBG, "%s: MSI handler\n", __FUNCTION__);
/* If Intr is because of Rx Traffic */
for (i = 0; i < config->rx_ring_num; i++)
rx_intr_handler(&mac_control->rings[i]);
/* If Intr is because of Tx Traffic */
for (i = 0; i < config->tx_fifo_num; i++)
tx_intr_handler(&mac_control->fifos[i]);
/*
* If the Rx buffer count is below the panic threshold then
* reallocate the buffers from the interrupt handler itself,
* else schedule a tasklet to reallocate the buffers.
*/
for (i = 0; i < config->rx_ring_num; i++)
s2io_chk_rx_buffers(sp, i);
atomic_dec(&sp->isr_cnt);
return IRQ_HANDLED;
}
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
static irqreturn_t s2io_msix_ring_handle(int irq, void *dev_id)
{
struct ring_info *ring = (struct ring_info *)dev_id;
struct s2io_nic *sp = ring->nic;
atomic_inc(&sp->isr_cnt);
rx_intr_handler(ring);
s2io_chk_rx_buffers(sp, ring->ring_no);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
atomic_dec(&sp->isr_cnt);
return IRQ_HANDLED;
}
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
static irqreturn_t s2io_msix_fifo_handle(int irq, void *dev_id)
{
struct fifo_info *fifo = (struct fifo_info *)dev_id;
struct s2io_nic *sp = fifo->nic;
atomic_inc(&sp->isr_cnt);
tx_intr_handler(fifo);
atomic_dec(&sp->isr_cnt);
return IRQ_HANDLED;
}
static void s2io_txpic_intr_handle(struct s2io_nic *sp)
{
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64;
val64 = readq(&bar0->pic_int_status);
if (val64 & PIC_INT_GPIO) {
val64 = readq(&bar0->gpio_int_reg);
if ((val64 & GPIO_INT_REG_LINK_DOWN) &&
(val64 & GPIO_INT_REG_LINK_UP)) {
/*
* This is unstable state so clear both up/down
* interrupt and adapter to re-evaluate the link state.
*/
val64 |= GPIO_INT_REG_LINK_DOWN;
val64 |= GPIO_INT_REG_LINK_UP;
writeq(val64, &bar0->gpio_int_reg);
val64 = readq(&bar0->gpio_int_mask);
val64 &= ~(GPIO_INT_MASK_LINK_UP |
GPIO_INT_MASK_LINK_DOWN);
writeq(val64, &bar0->gpio_int_mask);
}
else if (val64 & GPIO_INT_REG_LINK_UP) {
val64 = readq(&bar0->adapter_status);
/* Enable Adapter */
val64 = readq(&bar0->adapter_control);
val64 |= ADAPTER_CNTL_EN;
writeq(val64, &bar0->adapter_control);
val64 |= ADAPTER_LED_ON;
writeq(val64, &bar0->adapter_control);
if (!sp->device_enabled_once)
sp->device_enabled_once = 1;
s2io_link(sp, LINK_UP);
/*
* unmask link down interrupt and mask link-up
* intr
*/
val64 = readq(&bar0->gpio_int_mask);
val64 &= ~GPIO_INT_MASK_LINK_DOWN;
val64 |= GPIO_INT_MASK_LINK_UP;
writeq(val64, &bar0->gpio_int_mask);
}else if (val64 & GPIO_INT_REG_LINK_DOWN) {
val64 = readq(&bar0->adapter_status);
s2io_link(sp, LINK_DOWN);
/* Link is down so unmaks link up interrupt */
val64 = readq(&bar0->gpio_int_mask);
val64 &= ~GPIO_INT_MASK_LINK_UP;
val64 |= GPIO_INT_MASK_LINK_DOWN;
writeq(val64, &bar0->gpio_int_mask);
/* turn off LED */
val64 = readq(&bar0->adapter_control);
val64 = val64 &(~ADAPTER_LED_ON);
writeq(val64, &bar0->adapter_control);
}
}
val64 = readq(&bar0->gpio_int_mask);
}
/**
* s2io_isr - ISR handler of the device .
* @irq: the irq of the device.
* @dev_id: a void pointer to the dev structure of the NIC.
* Description: This function is the ISR handler of the device. It
* identifies the reason for the interrupt and calls the relevant
* service routines. As a contongency measure, this ISR allocates the
* recv buffers, if their numbers are below the panic value which is
* presently set to 25% of the original number of rcv buffers allocated.
* Return value:
* IRQ_HANDLED: will be returned if IRQ was handled by this routine
* IRQ_NONE: will be returned if interrupt is not from our device
*/
IRQ: Maintain regs pointer globally rather than passing to IRQ handlers Maintain a per-CPU global "struct pt_regs *" variable which can be used instead of passing regs around manually through all ~1800 interrupt handlers in the Linux kernel. The regs pointer is used in few places, but it potentially costs both stack space and code to pass it around. On the FRV arch, removing the regs parameter from all the genirq function results in a 20% speed up of the IRQ exit path (ie: from leaving timer_interrupt() to leaving do_IRQ()). Where appropriate, an arch may override the generic storage facility and do something different with the variable. On FRV, for instance, the address is maintained in GR28 at all times inside the kernel as part of general exception handling. Having looked over the code, it appears that the parameter may be handed down through up to twenty or so layers of functions. Consider a USB character device attached to a USB hub, attached to a USB controller that posts its interrupts through a cascaded auxiliary interrupt controller. A character device driver may want to pass regs to the sysrq handler through the input layer which adds another few layers of parameter passing. I've build this code with allyesconfig for x86_64 and i386. I've runtested the main part of the code on FRV and i386, though I can't test most of the drivers. I've also done partial conversion for powerpc and MIPS - these at least compile with minimal configurations. This will affect all archs. Mostly the changes should be relatively easy. Take do_IRQ(), store the regs pointer at the beginning, saving the old one: struct pt_regs *old_regs = set_irq_regs(regs); And put the old one back at the end: set_irq_regs(old_regs); Don't pass regs through to generic_handle_irq() or __do_IRQ(). In timer_interrupt(), this sort of change will be necessary: - update_process_times(user_mode(regs)); - profile_tick(CPU_PROFILING, regs); + update_process_times(user_mode(get_irq_regs())); + profile_tick(CPU_PROFILING); I'd like to move update_process_times()'s use of get_irq_regs() into itself, except that i386, alone of the archs, uses something other than user_mode(). Some notes on the interrupt handling in the drivers: (*) input_dev() is now gone entirely. The regs pointer is no longer stored in the input_dev struct. (*) finish_unlinks() in drivers/usb/host/ohci-q.c needs checking. It does something different depending on whether it's been supplied with a regs pointer or not. (*) Various IRQ handler function pointers have been moved to type irq_handler_t. Signed-Off-By: David Howells <dhowells@redhat.com> (cherry picked from 1b16e7ac850969f38b375e511e3fa2f474a33867 commit)
2006-10-05 13:55:46 +00:00
static irqreturn_t s2io_isr(int irq, void *dev_id)
{
struct net_device *dev = (struct net_device *) dev_id;
struct s2io_nic *sp = dev->priv;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
int i;
u64 reason = 0;
struct mac_info *mac_control;
struct config_param *config;
atomic_inc(&sp->isr_cnt);
mac_control = &sp->mac_control;
config = &sp->config;
/*
* Identify the cause for interrupt and call the appropriate
* interrupt handler. Causes for the interrupt could be;
* 1. Rx of packet.
* 2. Tx complete.
* 3. Link down.
* 4. Error in any functional blocks of the NIC.
*/
reason = readq(&bar0->general_int_status);
if (!reason) {
/* The interrupt was not raised by us. */
atomic_dec(&sp->isr_cnt);
return IRQ_NONE;
}
else if (unlikely(reason == S2IO_MINUS_ONE) ) {
/* Disable device and get out */
atomic_dec(&sp->isr_cnt);
return IRQ_NONE;
}
if (napi) {
if (reason & GEN_INTR_RXTRAFFIC) {
if ( likely ( netif_rx_schedule_prep(dev)) ) {
__netif_rx_schedule(dev);
writeq(S2IO_MINUS_ONE, &bar0->rx_traffic_mask);
}
else
writeq(S2IO_MINUS_ONE, &bar0->rx_traffic_int);
}
} else {
/*
* Rx handler is called by default, without checking for the
* cause of interrupt.
* rx_traffic_int reg is an R1 register, writing all 1's
* will ensure that the actual interrupt causing bit get's
* cleared and hence a read can be avoided.
*/
if (reason & GEN_INTR_RXTRAFFIC)
writeq(S2IO_MINUS_ONE, &bar0->rx_traffic_int);
for (i = 0; i < config->rx_ring_num; i++) {
rx_intr_handler(&mac_control->rings[i]);
}
}
/*
* tx_traffic_int reg is an R1 register, writing all 1's
* will ensure that the actual interrupt causing bit get's
* cleared and hence a read can be avoided.
*/
if (reason & GEN_INTR_TXTRAFFIC)
writeq(S2IO_MINUS_ONE, &bar0->tx_traffic_int);
for (i = 0; i < config->tx_fifo_num; i++)
tx_intr_handler(&mac_control->fifos[i]);
if (reason & GEN_INTR_TXPIC)
s2io_txpic_intr_handle(sp);
/*
* If the Rx buffer count is below the panic threshold then
* reallocate the buffers from the interrupt handler itself,
* else schedule a tasklet to reallocate the buffers.
*/
if (!napi) {
for (i = 0; i < config->rx_ring_num; i++)
s2io_chk_rx_buffers(sp, i);
}
writeq(0, &bar0->general_int_mask);
readl(&bar0->general_int_status);
atomic_dec(&sp->isr_cnt);
return IRQ_HANDLED;
}
/**
* s2io_updt_stats -
*/
static void s2io_updt_stats(struct s2io_nic *sp)
{
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64;
int cnt = 0;
if (atomic_read(&sp->card_state) == CARD_UP) {
/* Apprx 30us on a 133 MHz bus */
val64 = SET_UPDT_CLICKS(10) |
STAT_CFG_ONE_SHOT_EN | STAT_CFG_STAT_EN;
writeq(val64, &bar0->stat_cfg);
do {
udelay(100);
val64 = readq(&bar0->stat_cfg);
if (!(val64 & BIT(0)))
break;
cnt++;
if (cnt == 5)
break; /* Updt failed */
} while(1);
}
}
/**
* s2io_get_stats - Updates the device statistics structure.
* @dev : pointer to the device structure.
* Description:
* This function updates the device statistics structure in the s2io_nic
* structure and returns a pointer to the same.
* Return value:
* pointer to the updated net_device_stats structure.
*/
static struct net_device_stats *s2io_get_stats(struct net_device *dev)
{
struct s2io_nic *sp = dev->priv;
struct mac_info *mac_control;
struct config_param *config;
mac_control = &sp->mac_control;
config = &sp->config;
/* Configure Stats for immediate updt */
s2io_updt_stats(sp);
sp->stats.tx_packets =
le32_to_cpu(mac_control->stats_info->tmac_frms);
sp->stats.tx_errors =
le32_to_cpu(mac_control->stats_info->tmac_any_err_frms);
sp->stats.rx_errors =
le64_to_cpu(mac_control->stats_info->rmac_drop_frms);
sp->stats.multicast =
le32_to_cpu(mac_control->stats_info->rmac_vld_mcst_frms);
sp->stats.rx_length_errors =
le64_to_cpu(mac_control->stats_info->rmac_long_frms);
return (&sp->stats);
}
/**
* s2io_set_multicast - entry point for multicast address enable/disable.
* @dev : pointer to the device structure
* Description:
* This function is a driver entry point which gets called by the kernel
* whenever multicast addresses must be enabled/disabled. This also gets
* called to set/reset promiscuous mode. Depending on the deivce flag, we
* determine, if multicast address must be enabled or if promiscuous mode
* is to be disabled etc.
* Return value:
* void.
*/
static void s2io_set_multicast(struct net_device *dev)
{
int i, j, prev_cnt;
struct dev_mc_list *mclist;
struct s2io_nic *sp = dev->priv;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64 = 0, multi_mac = 0x010203040506ULL, mask =
0xfeffffffffffULL;
u64 dis_addr = 0xffffffffffffULL, mac_addr = 0;
void __iomem *add;
if ((dev->flags & IFF_ALLMULTI) && (!sp->m_cast_flg)) {
/* Enable all Multicast addresses */
writeq(RMAC_ADDR_DATA0_MEM_ADDR(multi_mac),
&bar0->rmac_addr_data0_mem);
writeq(RMAC_ADDR_DATA1_MEM_MASK(mask),
&bar0->rmac_addr_data1_mem);
val64 = RMAC_ADDR_CMD_MEM_WE |
RMAC_ADDR_CMD_MEM_STROBE_NEW_CMD |
RMAC_ADDR_CMD_MEM_OFFSET(MAC_MC_ALL_MC_ADDR_OFFSET);
writeq(val64, &bar0->rmac_addr_cmd_mem);
/* Wait till command completes */
wait_for_cmd_complete(&bar0->rmac_addr_cmd_mem,
RMAC_ADDR_CMD_MEM_STROBE_CMD_EXECUTING,
S2IO_BIT_RESET);
sp->m_cast_flg = 1;
sp->all_multi_pos = MAC_MC_ALL_MC_ADDR_OFFSET;
} else if ((dev->flags & IFF_ALLMULTI) && (sp->m_cast_flg)) {
/* Disable all Multicast addresses */
writeq(RMAC_ADDR_DATA0_MEM_ADDR(dis_addr),
&bar0->rmac_addr_data0_mem);
writeq(RMAC_ADDR_DATA1_MEM_MASK(0x0),
&bar0->rmac_addr_data1_mem);
val64 = RMAC_ADDR_CMD_MEM_WE |
RMAC_ADDR_CMD_MEM_STROBE_NEW_CMD |
RMAC_ADDR_CMD_MEM_OFFSET(sp->all_multi_pos);
writeq(val64, &bar0->rmac_addr_cmd_mem);
/* Wait till command completes */
wait_for_cmd_complete(&bar0->rmac_addr_cmd_mem,
RMAC_ADDR_CMD_MEM_STROBE_CMD_EXECUTING,
S2IO_BIT_RESET);
sp->m_cast_flg = 0;
sp->all_multi_pos = 0;
}
if ((dev->flags & IFF_PROMISC) && (!sp->promisc_flg)) {
/* Put the NIC into promiscuous mode */
add = &bar0->mac_cfg;
val64 = readq(&bar0->mac_cfg);
val64 |= MAC_CFG_RMAC_PROM_ENABLE;
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) val64, add);
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) (val64 >> 32), (add + 4));
if (vlan_tag_strip != 1) {
val64 = readq(&bar0->rx_pa_cfg);
val64 &= ~RX_PA_CFG_STRIP_VLAN_TAG;
writeq(val64, &bar0->rx_pa_cfg);
vlan_strip_flag = 0;
}
val64 = readq(&bar0->mac_cfg);
sp->promisc_flg = 1;
DBG_PRINT(INFO_DBG, "%s: entered promiscuous mode\n",
dev->name);
} else if (!(dev->flags & IFF_PROMISC) && (sp->promisc_flg)) {
/* Remove the NIC from promiscuous mode */
add = &bar0->mac_cfg;
val64 = readq(&bar0->mac_cfg);
val64 &= ~MAC_CFG_RMAC_PROM_ENABLE;
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) val64, add);
writeq(RMAC_CFG_KEY(0x4C0D), &bar0->rmac_cfg_key);
writel((u32) (val64 >> 32), (add + 4));
if (vlan_tag_strip != 0) {
val64 = readq(&bar0->rx_pa_cfg);
val64 |= RX_PA_CFG_STRIP_VLAN_TAG;
writeq(val64, &bar0->rx_pa_cfg);
vlan_strip_flag = 1;
}
val64 = readq(&bar0->mac_cfg);
sp->promisc_flg = 0;
DBG_PRINT(INFO_DBG, "%s: left promiscuous mode\n",
dev->name);
}
/* Update individual M_CAST address list */
if ((!sp->m_cast_flg) && dev->mc_count) {
if (dev->mc_count >
(MAX_ADDRS_SUPPORTED - MAC_MC_ADDR_START_OFFSET - 1)) {
DBG_PRINT(ERR_DBG, "%s: No more Rx filters ",
dev->name);
DBG_PRINT(ERR_DBG, "can be added, please enable ");
DBG_PRINT(ERR_DBG, "ALL_MULTI instead\n");
return;
}
prev_cnt = sp->mc_addr_count;
sp->mc_addr_count = dev->mc_count;
/* Clear out the previous list of Mc in the H/W. */
for (i = 0; i < prev_cnt; i++) {
writeq(RMAC_ADDR_DATA0_MEM_ADDR(dis_addr),
&bar0->rmac_addr_data0_mem);
writeq(RMAC_ADDR_DATA1_MEM_MASK(0ULL),
&bar0->rmac_addr_data1_mem);
val64 = RMAC_ADDR_CMD_MEM_WE |
RMAC_ADDR_CMD_MEM_STROBE_NEW_CMD |
RMAC_ADDR_CMD_MEM_OFFSET
(MAC_MC_ADDR_START_OFFSET + i);
writeq(val64, &bar0->rmac_addr_cmd_mem);
/* Wait for command completes */
if (wait_for_cmd_complete(&bar0->rmac_addr_cmd_mem,
RMAC_ADDR_CMD_MEM_STROBE_CMD_EXECUTING,
S2IO_BIT_RESET)) {
DBG_PRINT(ERR_DBG, "%s: Adding ",
dev->name);
DBG_PRINT(ERR_DBG, "Multicasts failed\n");
return;
}
}
/* Create the new Rx filter list and update the same in H/W. */
for (i = 0, mclist = dev->mc_list; i < dev->mc_count;
i++, mclist = mclist->next) {
memcpy(sp->usr_addrs[i].addr, mclist->dmi_addr,
ETH_ALEN);
mac_addr = 0;
for (j = 0; j < ETH_ALEN; j++) {
mac_addr |= mclist->dmi_addr[j];
mac_addr <<= 8;
}
mac_addr >>= 8;
writeq(RMAC_ADDR_DATA0_MEM_ADDR(mac_addr),
&bar0->rmac_addr_data0_mem);
writeq(RMAC_ADDR_DATA1_MEM_MASK(0ULL),
&bar0->rmac_addr_data1_mem);
val64 = RMAC_ADDR_CMD_MEM_WE |
RMAC_ADDR_CMD_MEM_STROBE_NEW_CMD |
RMAC_ADDR_CMD_MEM_OFFSET
(i + MAC_MC_ADDR_START_OFFSET);
writeq(val64, &bar0->rmac_addr_cmd_mem);
/* Wait for command completes */
if (wait_for_cmd_complete(&bar0->rmac_addr_cmd_mem,
RMAC_ADDR_CMD_MEM_STROBE_CMD_EXECUTING,
S2IO_BIT_RESET)) {
DBG_PRINT(ERR_DBG, "%s: Adding ",
dev->name);
DBG_PRINT(ERR_DBG, "Multicasts failed\n");
return;
}
}
}
}
/**
* s2io_set_mac_addr - Programs the Xframe mac address
* @dev : pointer to the device structure.
* @addr: a uchar pointer to the new mac address which is to be set.
* Description : This procedure will program the Xframe to receive
* frames with new Mac Address
* Return value: SUCCESS on success and an appropriate (-)ve integer
* as defined in errno.h file on failure.
*/
static int s2io_set_mac_addr(struct net_device *dev, u8 * addr)
{
struct s2io_nic *sp = dev->priv;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
register u64 val64, mac_addr = 0;
int i;
u64 old_mac_addr = 0;
/*
* Set the new MAC address as the new unicast filter and reflect this
* change on the device address registered with the OS. It will be
* at offset 0.
*/
for (i = 0; i < ETH_ALEN; i++) {
mac_addr <<= 8;
mac_addr |= addr[i];
old_mac_addr <<= 8;
old_mac_addr |= sp->def_mac_addr[0].mac_addr[i];
}
if(0 == mac_addr)
return SUCCESS;
/* Update the internal structure with this new mac address */
if(mac_addr != old_mac_addr) {
memset(sp->def_mac_addr[0].mac_addr, 0, sizeof(ETH_ALEN));
sp->def_mac_addr[0].mac_addr[5] = (u8) (mac_addr);
sp->def_mac_addr[0].mac_addr[4] = (u8) (mac_addr >> 8);
sp->def_mac_addr[0].mac_addr[3] = (u8) (mac_addr >> 16);
sp->def_mac_addr[0].mac_addr[2] = (u8) (mac_addr >> 24);
sp->def_mac_addr[0].mac_addr[1] = (u8) (mac_addr >> 32);
sp->def_mac_addr[0].mac_addr[0] = (u8) (mac_addr >> 40);
}
writeq(RMAC_ADDR_DATA0_MEM_ADDR(mac_addr),
&bar0->rmac_addr_data0_mem);
val64 =
RMAC_ADDR_CMD_MEM_WE | RMAC_ADDR_CMD_MEM_STROBE_NEW_CMD |
RMAC_ADDR_CMD_MEM_OFFSET(0);
writeq(val64, &bar0->rmac_addr_cmd_mem);
/* Wait till command completes */
if (wait_for_cmd_complete(&bar0->rmac_addr_cmd_mem,
RMAC_ADDR_CMD_MEM_STROBE_CMD_EXECUTING, S2IO_BIT_RESET)) {
DBG_PRINT(ERR_DBG, "%s: set_mac_addr failed\n", dev->name);
return FAILURE;
}
return SUCCESS;
}
/**
* s2io_ethtool_sset - Sets different link parameters.
* @sp : private member of the device structure, which is a pointer to the * s2io_nic structure.
* @info: pointer to the structure with parameters given by ethtool to set
* link information.
* Description:
* The function sets different link parameters provided by the user onto
* the NIC.
* Return value:
* 0 on success.
*/
static int s2io_ethtool_sset(struct net_device *dev,
struct ethtool_cmd *info)
{
struct s2io_nic *sp = dev->priv;
if ((info->autoneg == AUTONEG_ENABLE) ||
(info->speed != SPEED_10000) || (info->duplex != DUPLEX_FULL))
return -EINVAL;
else {
s2io_close(sp->dev);
s2io_open(sp->dev);
}
return 0;
}
/**
* s2io_ethtol_gset - Return link specific information.
* @sp : private member of the device structure, pointer to the
* s2io_nic structure.
* @info : pointer to the structure with parameters given by ethtool
* to return link information.
* Description:
* Returns link specific information like speed, duplex etc.. to ethtool.
* Return value :
* return 0 on success.
*/
static int s2io_ethtool_gset(struct net_device *dev, struct ethtool_cmd *info)
{
struct s2io_nic *sp = dev->priv;
info->supported = (SUPPORTED_10000baseT_Full | SUPPORTED_FIBRE);
info->advertising = (SUPPORTED_10000baseT_Full | SUPPORTED_FIBRE);
info->port = PORT_FIBRE;
/* info->transceiver?? TODO */
if (netif_carrier_ok(sp->dev)) {
info->speed = 10000;
info->duplex = DUPLEX_FULL;
} else {
info->speed = -1;
info->duplex = -1;
}
info->autoneg = AUTONEG_DISABLE;
return 0;
}
/**
* s2io_ethtool_gdrvinfo - Returns driver specific information.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @info : pointer to the structure with parameters given by ethtool to
* return driver information.
* Description:
* Returns driver specefic information like name, version etc.. to ethtool.
* Return value:
* void
*/
static void s2io_ethtool_gdrvinfo(struct net_device *dev,
struct ethtool_drvinfo *info)
{
struct s2io_nic *sp = dev->priv;
strncpy(info->driver, s2io_driver_name, sizeof(info->driver));
strncpy(info->version, s2io_driver_version, sizeof(info->version));
strncpy(info->fw_version, "", sizeof(info->fw_version));
strncpy(info->bus_info, pci_name(sp->pdev), sizeof(info->bus_info));
info->regdump_len = XENA_REG_SPACE;
info->eedump_len = XENA_EEPROM_SPACE;
info->testinfo_len = S2IO_TEST_LEN;
if (sp->device_type == XFRAME_I_DEVICE)
info->n_stats = XFRAME_I_STAT_LEN;
else
info->n_stats = XFRAME_II_STAT_LEN;
}
/**
* s2io_ethtool_gregs - dumps the entire space of Xfame into the buffer.
* @sp: private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @regs : pointer to the structure with parameters given by ethtool for
* dumping the registers.
* @reg_space: The input argumnet into which all the registers are dumped.
* Description:
* Dumps the entire register space of xFrame NIC into the user given
* buffer area.
* Return value :
* void .
*/
static void s2io_ethtool_gregs(struct net_device *dev,
struct ethtool_regs *regs, void *space)
{
int i;
u64 reg;
u8 *reg_space = (u8 *) space;
struct s2io_nic *sp = dev->priv;
regs->len = XENA_REG_SPACE;
regs->version = sp->pdev->subsystem_device;
for (i = 0; i < regs->len; i += 8) {
reg = readq(sp->bar0 + i);
memcpy((reg_space + i), &reg, 8);
}
}
/**
* s2io_phy_id - timer function that alternates adapter LED.
* @data : address of the private member of the device structure, which
* is a pointer to the s2io_nic structure, provided as an u32.
* Description: This is actually the timer function that alternates the
* adapter LED bit of the adapter control bit to set/reset every time on
* invocation. The timer is set for 1/2 a second, hence tha NIC blinks
* once every second.
*/
static void s2io_phy_id(unsigned long data)
{
struct s2io_nic *sp = (struct s2io_nic *) data;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64 = 0;
u16 subid;
subid = sp->pdev->subsystem_device;
if ((sp->device_type == XFRAME_II_DEVICE) ||
((subid & 0xFF) >= 0x07)) {
val64 = readq(&bar0->gpio_control);
val64 ^= GPIO_CTRL_GPIO_0;
writeq(val64, &bar0->gpio_control);
} else {
val64 = readq(&bar0->adapter_control);
val64 ^= ADAPTER_LED_ON;
writeq(val64, &bar0->adapter_control);
}
mod_timer(&sp->id_timer, jiffies + HZ / 2);
}
/**
* s2io_ethtool_idnic - To physically identify the nic on the system.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @id : pointer to the structure with identification parameters given by
* ethtool.
* Description: Used to physically identify the NIC on the system.
* The Link LED will blink for a time specified by the user for
* identification.
* NOTE: The Link has to be Up to be able to blink the LED. Hence
* identification is possible only if it's link is up.
* Return value:
* int , returns 0 on success
*/
static int s2io_ethtool_idnic(struct net_device *dev, u32 data)
{
u64 val64 = 0, last_gpio_ctrl_val;
struct s2io_nic *sp = dev->priv;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u16 subid;
subid = sp->pdev->subsystem_device;
last_gpio_ctrl_val = readq(&bar0->gpio_control);
if ((sp->device_type == XFRAME_I_DEVICE) &&
((subid & 0xFF) < 0x07)) {
val64 = readq(&bar0->adapter_control);
if (!(val64 & ADAPTER_CNTL_EN)) {
printk(KERN_ERR
"Adapter Link down, cannot blink LED\n");
return -EFAULT;
}
}
if (sp->id_timer.function == NULL) {
init_timer(&sp->id_timer);
sp->id_timer.function = s2io_phy_id;
sp->id_timer.data = (unsigned long) sp;
}
mod_timer(&sp->id_timer, jiffies);
if (data)
msleep_interruptible(data * HZ);
else
msleep_interruptible(MAX_FLICKER_TIME);
del_timer_sync(&sp->id_timer);
if (CARDS_WITH_FAULTY_LINK_INDICATORS(sp->device_type, subid)) {
writeq(last_gpio_ctrl_val, &bar0->gpio_control);
last_gpio_ctrl_val = readq(&bar0->gpio_control);
}
return 0;
}
/**
* s2io_ethtool_getpause_data -Pause frame frame generation and reception.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @ep : pointer to the structure with pause parameters given by ethtool.
* Description:
* Returns the Pause frame generation and reception capability of the NIC.
* Return value:
* void
*/
static void s2io_ethtool_getpause_data(struct net_device *dev,
struct ethtool_pauseparam *ep)
{
u64 val64;
struct s2io_nic *sp = dev->priv;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
val64 = readq(&bar0->rmac_pause_cfg);
if (val64 & RMAC_PAUSE_GEN_ENABLE)
ep->tx_pause = TRUE;
if (val64 & RMAC_PAUSE_RX_ENABLE)
ep->rx_pause = TRUE;
ep->autoneg = FALSE;
}
/**
* s2io_ethtool_setpause_data - set/reset pause frame generation.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @ep : pointer to the structure with pause parameters given by ethtool.
* Description:
* It can be used to set or reset Pause frame generation or reception
* support of the NIC.
* Return value:
* int, returns 0 on Success
*/
static int s2io_ethtool_setpause_data(struct net_device *dev,
struct ethtool_pauseparam *ep)
{
u64 val64;
struct s2io_nic *sp = dev->priv;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
val64 = readq(&bar0->rmac_pause_cfg);
if (ep->tx_pause)
val64 |= RMAC_PAUSE_GEN_ENABLE;
else
val64 &= ~RMAC_PAUSE_GEN_ENABLE;
if (ep->rx_pause)
val64 |= RMAC_PAUSE_RX_ENABLE;
else
val64 &= ~RMAC_PAUSE_RX_ENABLE;
writeq(val64, &bar0->rmac_pause_cfg);
return 0;
}
/**
* read_eeprom - reads 4 bytes of data from user given offset.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @off : offset at which the data must be written
* @data : Its an output parameter where the data read at the given
* offset is stored.
* Description:
* Will read 4 bytes of data from the user given offset and return the
* read data.
* NOTE: Will allow to read only part of the EEPROM visible through the
* I2C bus.
* Return value:
* -1 on failure and 0 on success.
*/
#define S2IO_DEV_ID 5
static int read_eeprom(struct s2io_nic * sp, int off, u64 * data)
{
int ret = -1;
u32 exit_cnt = 0;
u64 val64;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
if (sp->device_type == XFRAME_I_DEVICE) {
val64 = I2C_CONTROL_DEV_ID(S2IO_DEV_ID) | I2C_CONTROL_ADDR(off) |
I2C_CONTROL_BYTE_CNT(0x3) | I2C_CONTROL_READ |
I2C_CONTROL_CNTL_START;
SPECIAL_REG_WRITE(val64, &bar0->i2c_control, LF);
while (exit_cnt < 5) {
val64 = readq(&bar0->i2c_control);
if (I2C_CONTROL_CNTL_END(val64)) {
*data = I2C_CONTROL_GET_DATA(val64);
ret = 0;
break;
}
msleep(50);
exit_cnt++;
}
}
if (sp->device_type == XFRAME_II_DEVICE) {
val64 = SPI_CONTROL_KEY(0x9) | SPI_CONTROL_SEL1 |
SPI_CONTROL_BYTECNT(0x3) |
SPI_CONTROL_CMD(0x3) | SPI_CONTROL_ADDR(off);
SPECIAL_REG_WRITE(val64, &bar0->spi_control, LF);
val64 |= SPI_CONTROL_REQ;
SPECIAL_REG_WRITE(val64, &bar0->spi_control, LF);
while (exit_cnt < 5) {
val64 = readq(&bar0->spi_control);
if (val64 & SPI_CONTROL_NACK) {
ret = 1;
break;
} else if (val64 & SPI_CONTROL_DONE) {
*data = readq(&bar0->spi_data);
*data &= 0xffffff;
ret = 0;
break;
}
msleep(50);
exit_cnt++;
}
}
return ret;
}
/**
* write_eeprom - actually writes the relevant part of the data value.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @off : offset at which the data must be written
* @data : The data that is to be written
* @cnt : Number of bytes of the data that are actually to be written into
* the Eeprom. (max of 3)
* Description:
* Actually writes the relevant part of the data value into the Eeprom
* through the I2C bus.
* Return value:
* 0 on success, -1 on failure.
*/
static int write_eeprom(struct s2io_nic * sp, int off, u64 data, int cnt)
{
int exit_cnt = 0, ret = -1;
u64 val64;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
if (sp->device_type == XFRAME_I_DEVICE) {
val64 = I2C_CONTROL_DEV_ID(S2IO_DEV_ID) | I2C_CONTROL_ADDR(off) |
I2C_CONTROL_BYTE_CNT(cnt) | I2C_CONTROL_SET_DATA((u32)data) |
I2C_CONTROL_CNTL_START;
SPECIAL_REG_WRITE(val64, &bar0->i2c_control, LF);
while (exit_cnt < 5) {
val64 = readq(&bar0->i2c_control);
if (I2C_CONTROL_CNTL_END(val64)) {
if (!(val64 & I2C_CONTROL_NACK))
ret = 0;
break;
}
msleep(50);
exit_cnt++;
}
}
if (sp->device_type == XFRAME_II_DEVICE) {
int write_cnt = (cnt == 8) ? 0 : cnt;
writeq(SPI_DATA_WRITE(data,(cnt<<3)), &bar0->spi_data);
val64 = SPI_CONTROL_KEY(0x9) | SPI_CONTROL_SEL1 |
SPI_CONTROL_BYTECNT(write_cnt) |
SPI_CONTROL_CMD(0x2) | SPI_CONTROL_ADDR(off);
SPECIAL_REG_WRITE(val64, &bar0->spi_control, LF);
val64 |= SPI_CONTROL_REQ;
SPECIAL_REG_WRITE(val64, &bar0->spi_control, LF);
while (exit_cnt < 5) {
val64 = readq(&bar0->spi_control);
if (val64 & SPI_CONTROL_NACK) {
ret = 1;
break;
} else if (val64 & SPI_CONTROL_DONE) {
ret = 0;
break;
}
msleep(50);
exit_cnt++;
}
}
return ret;
}
static void s2io_vpd_read(struct s2io_nic *nic)
{
u8 *vpd_data;
u8 data;
int i=0, cnt, fail = 0;
int vpd_addr = 0x80;
if (nic->device_type == XFRAME_II_DEVICE) {
strcpy(nic->product_name, "Xframe II 10GbE network adapter");
vpd_addr = 0x80;
}
else {
strcpy(nic->product_name, "Xframe I 10GbE network adapter");
vpd_addr = 0x50;
}
strcpy(nic->serial_num, "NOT AVAILABLE");
vpd_data = kmalloc(256, GFP_KERNEL);
if (!vpd_data)
return;
for (i = 0; i < 256; i +=4 ) {
pci_write_config_byte(nic->pdev, (vpd_addr + 2), i);
pci_read_config_byte(nic->pdev, (vpd_addr + 2), &data);
pci_write_config_byte(nic->pdev, (vpd_addr + 3), 0);
for (cnt = 0; cnt <5; cnt++) {
msleep(2);
pci_read_config_byte(nic->pdev, (vpd_addr + 3), &data);
if (data == 0x80)
break;
}
if (cnt >= 5) {
DBG_PRINT(ERR_DBG, "Read of VPD data failed\n");
fail = 1;
break;
}
pci_read_config_dword(nic->pdev, (vpd_addr + 4),
(u32 *)&vpd_data[i]);
}
if(!fail) {
/* read serial number of adapter */
for (cnt = 0; cnt < 256; cnt++) {
if ((vpd_data[cnt] == 'S') &&
(vpd_data[cnt+1] == 'N') &&
(vpd_data[cnt+2] < VPD_STRING_LEN)) {
memset(nic->serial_num, 0, VPD_STRING_LEN);
memcpy(nic->serial_num, &vpd_data[cnt + 3],
vpd_data[cnt+2]);
break;
}
}
}
if ((!fail) && (vpd_data[1] < VPD_STRING_LEN)) {
memset(nic->product_name, 0, vpd_data[1]);
memcpy(nic->product_name, &vpd_data[3], vpd_data[1]);
}
kfree(vpd_data);
}
/**
* s2io_ethtool_geeprom - reads the value stored in the Eeprom.
* @sp : private member of the device structure, which is a pointer to the * s2io_nic structure.
* @eeprom : pointer to the user level structure provided by ethtool,
* containing all relevant information.
* @data_buf : user defined value to be written into Eeprom.
* Description: Reads the values stored in the Eeprom at given offset
* for a given length. Stores these values int the input argument data
* buffer 'data_buf' and returns these to the caller (ethtool.)
* Return value:
* int 0 on success
*/
static int s2io_ethtool_geeprom(struct net_device *dev,
struct ethtool_eeprom *eeprom, u8 * data_buf)
{
u32 i, valid;
u64 data;
struct s2io_nic *sp = dev->priv;
eeprom->magic = sp->pdev->vendor | (sp->pdev->device << 16);
if ((eeprom->offset + eeprom->len) > (XENA_EEPROM_SPACE))
eeprom->len = XENA_EEPROM_SPACE - eeprom->offset;
for (i = 0; i < eeprom->len; i += 4) {
if (read_eeprom(sp, (eeprom->offset + i), &data)) {
DBG_PRINT(ERR_DBG, "Read of EEPROM failed\n");
return -EFAULT;
}
valid = INV(data);
memcpy((data_buf + i), &valid, 4);
}
return 0;
}
/**
* s2io_ethtool_seeprom - tries to write the user provided value in Eeprom
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @eeprom : pointer to the user level structure provided by ethtool,
* containing all relevant information.
* @data_buf ; user defined value to be written into Eeprom.
* Description:
* Tries to write the user provided value in the Eeprom, at the offset
* given by the user.
* Return value:
* 0 on success, -EFAULT on failure.
*/
static int s2io_ethtool_seeprom(struct net_device *dev,
struct ethtool_eeprom *eeprom,
u8 * data_buf)
{
int len = eeprom->len, cnt = 0;
u64 valid = 0, data;
struct s2io_nic *sp = dev->priv;
if (eeprom->magic != (sp->pdev->vendor | (sp->pdev->device << 16))) {
DBG_PRINT(ERR_DBG,
"ETHTOOL_WRITE_EEPROM Err: Magic value ");
DBG_PRINT(ERR_DBG, "is wrong, Its not 0x%x\n",
eeprom->magic);
return -EFAULT;
}
while (len) {
data = (u32) data_buf[cnt] & 0x000000FF;
if (data) {
valid = (u32) (data << 24);
} else
valid = data;
if (write_eeprom(sp, (eeprom->offset + cnt), valid, 0)) {
DBG_PRINT(ERR_DBG,
"ETHTOOL_WRITE_EEPROM Err: Cannot ");
DBG_PRINT(ERR_DBG,
"write into the specified offset\n");
return -EFAULT;
}
cnt++;
len--;
}
return 0;
}
/**
* s2io_register_test - reads and writes into all clock domains.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @data : variable that returns the result of each of the test conducted b
* by the driver.
* Description:
* Read and write into all clock domains. The NIC has 3 clock domains,
* see that registers in all the three regions are accessible.
* Return value:
* 0 on success.
*/
static int s2io_register_test(struct s2io_nic * sp, uint64_t * data)
{
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64 = 0, exp_val;
int fail = 0;
val64 = readq(&bar0->pif_rd_swapper_fb);
if (val64 != 0x123456789abcdefULL) {
fail = 1;
DBG_PRINT(INFO_DBG, "Read Test level 1 fails\n");
}
val64 = readq(&bar0->rmac_pause_cfg);
if (val64 != 0xc000ffff00000000ULL) {
fail = 1;
DBG_PRINT(INFO_DBG, "Read Test level 2 fails\n");
}
val64 = readq(&bar0->rx_queue_cfg);
if (sp->device_type == XFRAME_II_DEVICE)
exp_val = 0x0404040404040404ULL;
else
exp_val = 0x0808080808080808ULL;
if (val64 != exp_val) {
fail = 1;
DBG_PRINT(INFO_DBG, "Read Test level 3 fails\n");
}
val64 = readq(&bar0->xgxs_efifo_cfg);
if (val64 != 0x000000001923141EULL) {
fail = 1;
DBG_PRINT(INFO_DBG, "Read Test level 4 fails\n");
}
val64 = 0x5A5A5A5A5A5A5A5AULL;
writeq(val64, &bar0->xmsi_data);
val64 = readq(&bar0->xmsi_data);
if (val64 != 0x5A5A5A5A5A5A5A5AULL) {
fail = 1;
DBG_PRINT(ERR_DBG, "Write Test level 1 fails\n");
}
val64 = 0xA5A5A5A5A5A5A5A5ULL;
writeq(val64, &bar0->xmsi_data);
val64 = readq(&bar0->xmsi_data);
if (val64 != 0xA5A5A5A5A5A5A5A5ULL) {
fail = 1;
DBG_PRINT(ERR_DBG, "Write Test level 2 fails\n");
}
*data = fail;
return fail;
}
/**
* s2io_eeprom_test - to verify that EEprom in the xena can be programmed.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @data:variable that returns the result of each of the test conducted by
* the driver.
* Description:
* Verify that EEPROM in the xena can be programmed using I2C_CONTROL
* register.
* Return value:
* 0 on success.
*/
static int s2io_eeprom_test(struct s2io_nic * sp, uint64_t * data)
{
int fail = 0;
u64 ret_data, org_4F0, org_7F0;
u8 saved_4F0 = 0, saved_7F0 = 0;
struct net_device *dev = sp->dev;
/* Test Write Error at offset 0 */
/* Note that SPI interface allows write access to all areas
* of EEPROM. Hence doing all negative testing only for Xframe I.
*/
if (sp->device_type == XFRAME_I_DEVICE)
if (!write_eeprom(sp, 0, 0, 3))
fail = 1;
/* Save current values at offsets 0x4F0 and 0x7F0 */
if (!read_eeprom(sp, 0x4F0, &org_4F0))
saved_4F0 = 1;
if (!read_eeprom(sp, 0x7F0, &org_7F0))
saved_7F0 = 1;
/* Test Write at offset 4f0 */
if (write_eeprom(sp, 0x4F0, 0x012345, 3))
fail = 1;
if (read_eeprom(sp, 0x4F0, &ret_data))
fail = 1;
if (ret_data != 0x012345) {
DBG_PRINT(ERR_DBG, "%s: eeprom test error at offset 0x4F0. "
"Data written %llx Data read %llx\n",
dev->name, (unsigned long long)0x12345,
(unsigned long long)ret_data);
fail = 1;
}
/* Reset the EEPROM data go FFFF */
write_eeprom(sp, 0x4F0, 0xFFFFFF, 3);
/* Test Write Request Error at offset 0x7c */
if (sp->device_type == XFRAME_I_DEVICE)
if (!write_eeprom(sp, 0x07C, 0, 3))
fail = 1;
/* Test Write Request at offset 0x7f0 */
if (write_eeprom(sp, 0x7F0, 0x012345, 3))
fail = 1;
if (read_eeprom(sp, 0x7F0, &ret_data))
fail = 1;
if (ret_data != 0x012345) {
DBG_PRINT(ERR_DBG, "%s: eeprom test error at offset 0x7F0. "
"Data written %llx Data read %llx\n",
dev->name, (unsigned long long)0x12345,
(unsigned long long)ret_data);
fail = 1;
}
/* Reset the EEPROM data go FFFF */
write_eeprom(sp, 0x7F0, 0xFFFFFF, 3);
if (sp->device_type == XFRAME_I_DEVICE) {
/* Test Write Error at offset 0x80 */
if (!write_eeprom(sp, 0x080, 0, 3))
fail = 1;
/* Test Write Error at offset 0xfc */
if (!write_eeprom(sp, 0x0FC, 0, 3))
fail = 1;
/* Test Write Error at offset 0x100 */
if (!write_eeprom(sp, 0x100, 0, 3))
fail = 1;
/* Test Write Error at offset 4ec */
if (!write_eeprom(sp, 0x4EC, 0, 3))
fail = 1;
}
/* Restore values at offsets 0x4F0 and 0x7F0 */
if (saved_4F0)
write_eeprom(sp, 0x4F0, org_4F0, 3);
if (saved_7F0)
write_eeprom(sp, 0x7F0, org_7F0, 3);
*data = fail;
return fail;
}
/**
* s2io_bist_test - invokes the MemBist test of the card .
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @data:variable that returns the result of each of the test conducted by
* the driver.
* Description:
* This invokes the MemBist test of the card. We give around
* 2 secs time for the Test to complete. If it's still not complete
* within this peiod, we consider that the test failed.
* Return value:
* 0 on success and -1 on failure.
*/
static int s2io_bist_test(struct s2io_nic * sp, uint64_t * data)
{
u8 bist = 0;
int cnt = 0, ret = -1;
pci_read_config_byte(sp->pdev, PCI_BIST, &bist);
bist |= PCI_BIST_START;
pci_write_config_word(sp->pdev, PCI_BIST, bist);
while (cnt < 20) {
pci_read_config_byte(sp->pdev, PCI_BIST, &bist);
if (!(bist & PCI_BIST_START)) {
*data = (bist & PCI_BIST_CODE_MASK);
ret = 0;
break;
}
msleep(100);
cnt++;
}
return ret;
}
/**
* s2io-link_test - verifies the link state of the nic
* @sp ; private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @data: variable that returns the result of each of the test conducted by
* the driver.
* Description:
* The function verifies the link state of the NIC and updates the input
* argument 'data' appropriately.
* Return value:
* 0 on success.
*/
static int s2io_link_test(struct s2io_nic * sp, uint64_t * data)
{
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64;
val64 = readq(&bar0->adapter_status);
if(!(LINK_IS_UP(val64)))
*data = 1;
else
*data = 0;
return *data;
}
/**
* s2io_rldram_test - offline test for access to the RldRam chip on the NIC
* @sp - private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @data - variable that returns the result of each of the test
* conducted by the driver.
* Description:
* This is one of the offline test that tests the read and write
* access to the RldRam chip on the NIC.
* Return value:
* 0 on success.
*/
static int s2io_rldram_test(struct s2io_nic * sp, uint64_t * data)
{
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64;
int cnt, iteration = 0, test_fail = 0;
val64 = readq(&bar0->adapter_control);
val64 &= ~ADAPTER_ECC_EN;
writeq(val64, &bar0->adapter_control);
val64 = readq(&bar0->mc_rldram_test_ctrl);
val64 |= MC_RLDRAM_TEST_MODE;
SPECIAL_REG_WRITE(val64, &bar0->mc_rldram_test_ctrl, LF);
val64 = readq(&bar0->mc_rldram_mrs);
val64 |= MC_RLDRAM_QUEUE_SIZE_ENABLE;
SPECIAL_REG_WRITE(val64, &bar0->mc_rldram_mrs, UF);
val64 |= MC_RLDRAM_MRS_ENABLE;
SPECIAL_REG_WRITE(val64, &bar0->mc_rldram_mrs, UF);
while (iteration < 2) {
val64 = 0x55555555aaaa0000ULL;
if (iteration == 1) {
val64 ^= 0xFFFFFFFFFFFF0000ULL;
}
writeq(val64, &bar0->mc_rldram_test_d0);
val64 = 0xaaaa5a5555550000ULL;
if (iteration == 1) {
val64 ^= 0xFFFFFFFFFFFF0000ULL;
}
writeq(val64, &bar0->mc_rldram_test_d1);
val64 = 0x55aaaaaaaa5a0000ULL;
if (iteration == 1) {
val64 ^= 0xFFFFFFFFFFFF0000ULL;
}
writeq(val64, &bar0->mc_rldram_test_d2);
val64 = (u64) (0x0000003ffffe0100ULL);
writeq(val64, &bar0->mc_rldram_test_add);
val64 = MC_RLDRAM_TEST_MODE | MC_RLDRAM_TEST_WRITE |
MC_RLDRAM_TEST_GO;
SPECIAL_REG_WRITE(val64, &bar0->mc_rldram_test_ctrl, LF);
for (cnt = 0; cnt < 5; cnt++) {
val64 = readq(&bar0->mc_rldram_test_ctrl);
if (val64 & MC_RLDRAM_TEST_DONE)
break;
msleep(200);
}
if (cnt == 5)
break;
val64 = MC_RLDRAM_TEST_MODE | MC_RLDRAM_TEST_GO;
SPECIAL_REG_WRITE(val64, &bar0->mc_rldram_test_ctrl, LF);
for (cnt = 0; cnt < 5; cnt++) {
val64 = readq(&bar0->mc_rldram_test_ctrl);
if (val64 & MC_RLDRAM_TEST_DONE)
break;
msleep(500);
}
if (cnt == 5)
break;
val64 = readq(&bar0->mc_rldram_test_ctrl);
if (!(val64 & MC_RLDRAM_TEST_PASS))
test_fail = 1;
iteration++;
}
*data = test_fail;
/* Bring the adapter out of test mode */
SPECIAL_REG_WRITE(0, &bar0->mc_rldram_test_ctrl, LF);
return test_fail;
}
/**
* s2io_ethtool_test - conducts 6 tsets to determine the health of card.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @ethtest : pointer to a ethtool command specific structure that will be
* returned to the user.
* @data : variable that returns the result of each of the test
* conducted by the driver.
* Description:
* This function conducts 6 tests ( 4 offline and 2 online) to determine
* the health of the card.
* Return value:
* void
*/
static void s2io_ethtool_test(struct net_device *dev,
struct ethtool_test *ethtest,
uint64_t * data)
{
struct s2io_nic *sp = dev->priv;
int orig_state = netif_running(sp->dev);
if (ethtest->flags == ETH_TEST_FL_OFFLINE) {
/* Offline Tests. */
if (orig_state)
s2io_close(sp->dev);
if (s2io_register_test(sp, &data[0]))
ethtest->flags |= ETH_TEST_FL_FAILED;
s2io_reset(sp);
if (s2io_rldram_test(sp, &data[3]))
ethtest->flags |= ETH_TEST_FL_FAILED;
s2io_reset(sp);
if (s2io_eeprom_test(sp, &data[1]))
ethtest->flags |= ETH_TEST_FL_FAILED;
if (s2io_bist_test(sp, &data[4]))
ethtest->flags |= ETH_TEST_FL_FAILED;
if (orig_state)
s2io_open(sp->dev);
data[2] = 0;
} else {
/* Online Tests. */
if (!orig_state) {
DBG_PRINT(ERR_DBG,
"%s: is not up, cannot run test\n",
dev->name);
data[0] = -1;
data[1] = -1;
data[2] = -1;
data[3] = -1;
data[4] = -1;
}
if (s2io_link_test(sp, &data[2]))
ethtest->flags |= ETH_TEST_FL_FAILED;
data[0] = 0;
data[1] = 0;
data[3] = 0;
data[4] = 0;
}
}
static void s2io_get_ethtool_stats(struct net_device *dev,
struct ethtool_stats *estats,
u64 * tmp_stats)
{
int i = 0;
struct s2io_nic *sp = dev->priv;
struct stat_block *stat_info = sp->mac_control.stats_info;
s2io_updt_stats(sp);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_frms_oflow) << 32 |
le32_to_cpu(stat_info->tmac_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_data_octets_oflow) << 32 |
le32_to_cpu(stat_info->tmac_data_octets);
tmp_stats[i++] = le64_to_cpu(stat_info->tmac_drop_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_mcst_frms_oflow) << 32 |
le32_to_cpu(stat_info->tmac_mcst_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_bcst_frms_oflow) << 32 |
le32_to_cpu(stat_info->tmac_bcst_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->tmac_pause_ctrl_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_ttl_octets_oflow) << 32 |
le32_to_cpu(stat_info->tmac_ttl_octets);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_ucst_frms_oflow) << 32 |
le32_to_cpu(stat_info->tmac_ucst_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_nucst_frms_oflow) << 32 |
le32_to_cpu(stat_info->tmac_nucst_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_any_err_frms_oflow) << 32 |
le32_to_cpu(stat_info->tmac_any_err_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->tmac_ttl_less_fb_octets);
tmp_stats[i++] = le64_to_cpu(stat_info->tmac_vld_ip_octets);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_vld_ip_oflow) << 32 |
le32_to_cpu(stat_info->tmac_vld_ip);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_drop_ip_oflow) << 32 |
le32_to_cpu(stat_info->tmac_drop_ip);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_icmp_oflow) << 32 |
le32_to_cpu(stat_info->tmac_icmp);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->tmac_rst_tcp_oflow) << 32 |
le32_to_cpu(stat_info->tmac_rst_tcp);
tmp_stats[i++] = le64_to_cpu(stat_info->tmac_tcp);
tmp_stats[i++] = (u64)le32_to_cpu(stat_info->tmac_udp_oflow) << 32 |
le32_to_cpu(stat_info->tmac_udp);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_vld_frms_oflow) << 32 |
le32_to_cpu(stat_info->rmac_vld_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_data_octets_oflow) << 32 |
le32_to_cpu(stat_info->rmac_data_octets);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_fcs_err_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_drop_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_vld_mcst_frms_oflow) << 32 |
le32_to_cpu(stat_info->rmac_vld_mcst_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_vld_bcst_frms_oflow) << 32 |
le32_to_cpu(stat_info->rmac_vld_bcst_frms);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_in_rng_len_err_frms);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_out_rng_len_err_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_long_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_pause_ctrl_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_unsup_ctrl_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_ttl_octets_oflow) << 32 |
le32_to_cpu(stat_info->rmac_ttl_octets);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_accepted_ucst_frms_oflow)
<< 32 | le32_to_cpu(stat_info->rmac_accepted_ucst_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_accepted_nucst_frms_oflow)
<< 32 | le32_to_cpu(stat_info->rmac_accepted_nucst_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_discarded_frms_oflow) << 32 |
le32_to_cpu(stat_info->rmac_discarded_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_drop_events_oflow)
<< 32 | le32_to_cpu(stat_info->rmac_drop_events);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_less_fb_octets);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_usized_frms_oflow) << 32 |
le32_to_cpu(stat_info->rmac_usized_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_osized_frms_oflow) << 32 |
le32_to_cpu(stat_info->rmac_osized_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_frag_frms_oflow) << 32 |
le32_to_cpu(stat_info->rmac_frag_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_jabber_frms_oflow) << 32 |
le32_to_cpu(stat_info->rmac_jabber_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_64_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_65_127_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_128_255_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_256_511_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_512_1023_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_1024_1518_frms);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_ip_oflow) << 32 |
le32_to_cpu(stat_info->rmac_ip);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ip_octets);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_hdr_err_ip);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_drop_ip_oflow) << 32 |
le32_to_cpu(stat_info->rmac_drop_ip);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_icmp_oflow) << 32 |
le32_to_cpu(stat_info->rmac_icmp);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_tcp);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_udp_oflow) << 32 |
le32_to_cpu(stat_info->rmac_udp);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_err_drp_udp_oflow) << 32 |
le32_to_cpu(stat_info->rmac_err_drp_udp);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_xgmii_err_sym);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_frms_q0);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_frms_q1);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_frms_q2);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_frms_q3);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_frms_q4);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_frms_q5);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_frms_q6);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_frms_q7);
tmp_stats[i++] = le16_to_cpu(stat_info->rmac_full_q0);
tmp_stats[i++] = le16_to_cpu(stat_info->rmac_full_q1);
tmp_stats[i++] = le16_to_cpu(stat_info->rmac_full_q2);
tmp_stats[i++] = le16_to_cpu(stat_info->rmac_full_q3);
tmp_stats[i++] = le16_to_cpu(stat_info->rmac_full_q4);
tmp_stats[i++] = le16_to_cpu(stat_info->rmac_full_q5);
tmp_stats[i++] = le16_to_cpu(stat_info->rmac_full_q6);
tmp_stats[i++] = le16_to_cpu(stat_info->rmac_full_q7);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_pause_cnt_oflow) << 32 |
le32_to_cpu(stat_info->rmac_pause_cnt);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_xgmii_data_err_cnt);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_xgmii_ctrl_err_cnt);
tmp_stats[i++] =
(u64)le32_to_cpu(stat_info->rmac_accepted_ip_oflow) << 32 |
le32_to_cpu(stat_info->rmac_accepted_ip);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_err_tcp);
tmp_stats[i++] = le32_to_cpu(stat_info->rd_req_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->new_rd_req_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->new_rd_req_rtry_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->rd_rtry_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->wr_rtry_rd_ack_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->wr_req_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->new_wr_req_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->new_wr_req_rtry_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->wr_rtry_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->wr_disc_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->rd_rtry_wr_ack_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->txp_wr_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->txd_rd_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->txd_wr_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->rxd_rd_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->rxd_wr_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->txf_rd_cnt);
tmp_stats[i++] = le32_to_cpu(stat_info->rxf_wr_cnt);
/* Enhanced statistics exist only for Hercules */
if(sp->device_type == XFRAME_II_DEVICE) {
tmp_stats[i++] =
le64_to_cpu(stat_info->rmac_ttl_1519_4095_frms);
tmp_stats[i++] =
le64_to_cpu(stat_info->rmac_ttl_4096_8191_frms);
tmp_stats[i++] =
le64_to_cpu(stat_info->rmac_ttl_8192_max_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_ttl_gt_max_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_osized_alt_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_jabber_alt_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_gt_max_alt_frms);
tmp_stats[i++] = le64_to_cpu(stat_info->rmac_vlan_frms);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_len_discard);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_fcs_discard);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_pf_discard);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_da_discard);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_red_discard);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_rts_discard);
tmp_stats[i++] = le32_to_cpu(stat_info->rmac_ingm_full_discard);
tmp_stats[i++] = le32_to_cpu(stat_info->link_fault_cnt);
}
tmp_stats[i++] = 0;
tmp_stats[i++] = stat_info->sw_stat.single_ecc_errs;
tmp_stats[i++] = stat_info->sw_stat.double_ecc_errs;
tmp_stats[i++] = stat_info->sw_stat.parity_err_cnt;
tmp_stats[i++] = stat_info->sw_stat.serious_err_cnt;
tmp_stats[i++] = stat_info->sw_stat.soft_reset_cnt;
tmp_stats[i++] = stat_info->sw_stat.fifo_full_cnt;
tmp_stats[i++] = stat_info->sw_stat.ring_full_cnt;
tmp_stats[i++] = stat_info->xpak_stat.alarm_transceiver_temp_high;
tmp_stats[i++] = stat_info->xpak_stat.alarm_transceiver_temp_low;
tmp_stats[i++] = stat_info->xpak_stat.alarm_laser_bias_current_high;
tmp_stats[i++] = stat_info->xpak_stat.alarm_laser_bias_current_low;
tmp_stats[i++] = stat_info->xpak_stat.alarm_laser_output_power_high;
tmp_stats[i++] = stat_info->xpak_stat.alarm_laser_output_power_low;
tmp_stats[i++] = stat_info->xpak_stat.warn_transceiver_temp_high;
tmp_stats[i++] = stat_info->xpak_stat.warn_transceiver_temp_low;
tmp_stats[i++] = stat_info->xpak_stat.warn_laser_bias_current_high;
tmp_stats[i++] = stat_info->xpak_stat.warn_laser_bias_current_low;
tmp_stats[i++] = stat_info->xpak_stat.warn_laser_output_power_high;
tmp_stats[i++] = stat_info->xpak_stat.warn_laser_output_power_low;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
tmp_stats[i++] = stat_info->sw_stat.clubbed_frms_cnt;
tmp_stats[i++] = stat_info->sw_stat.sending_both;
tmp_stats[i++] = stat_info->sw_stat.outof_sequence_pkts;
tmp_stats[i++] = stat_info->sw_stat.flush_max_pkts;
if (stat_info->sw_stat.num_aggregations) {
u64 tmp = stat_info->sw_stat.sum_avg_pkts_aggregated;
int count = 0;
/*
* Since 64-bit divide does not work on all platforms,
* do repeated subtraction.
*/
while (tmp >= stat_info->sw_stat.num_aggregations) {
tmp -= stat_info->sw_stat.num_aggregations;
count++;
}
tmp_stats[i++] = count;
}
else
tmp_stats[i++] = 0;
}
static int s2io_ethtool_get_regs_len(struct net_device *dev)
{
return (XENA_REG_SPACE);
}
static u32 s2io_ethtool_get_rx_csum(struct net_device * dev)
{
struct s2io_nic *sp = dev->priv;
return (sp->rx_csum);
}
static int s2io_ethtool_set_rx_csum(struct net_device *dev, u32 data)
{
struct s2io_nic *sp = dev->priv;
if (data)
sp->rx_csum = 1;
else
sp->rx_csum = 0;
return 0;
}
static int s2io_get_eeprom_len(struct net_device *dev)
{
return (XENA_EEPROM_SPACE);
}
static int s2io_ethtool_self_test_count(struct net_device *dev)
{
return (S2IO_TEST_LEN);
}
static void s2io_ethtool_get_strings(struct net_device *dev,
u32 stringset, u8 * data)
{
int stat_size = 0;
struct s2io_nic *sp = dev->priv;
switch (stringset) {
case ETH_SS_TEST:
memcpy(data, s2io_gstrings, S2IO_STRINGS_LEN);
break;
case ETH_SS_STATS:
stat_size = sizeof(ethtool_xena_stats_keys);
memcpy(data, &ethtool_xena_stats_keys,stat_size);
if(sp->device_type == XFRAME_II_DEVICE) {
memcpy(data + stat_size,
&ethtool_enhanced_stats_keys,
sizeof(ethtool_enhanced_stats_keys));
stat_size += sizeof(ethtool_enhanced_stats_keys);
}
memcpy(data + stat_size, &ethtool_driver_stats_keys,
sizeof(ethtool_driver_stats_keys));
}
}
static int s2io_ethtool_get_stats_count(struct net_device *dev)
{
struct s2io_nic *sp = dev->priv;
int stat_count = 0;
switch(sp->device_type) {
case XFRAME_I_DEVICE:
stat_count = XFRAME_I_STAT_LEN;
break;
case XFRAME_II_DEVICE:
stat_count = XFRAME_II_STAT_LEN;
break;
}
return stat_count;
}
static int s2io_ethtool_op_set_tx_csum(struct net_device *dev, u32 data)
{
if (data)
dev->features |= NETIF_F_IP_CSUM;
else
dev->features &= ~NETIF_F_IP_CSUM;
return 0;
}
static u32 s2io_ethtool_op_get_tso(struct net_device *dev)
{
return (dev->features & NETIF_F_TSO) != 0;
}
static int s2io_ethtool_op_set_tso(struct net_device *dev, u32 data)
{
if (data)
dev->features |= (NETIF_F_TSO | NETIF_F_TSO6);
else
dev->features &= ~(NETIF_F_TSO | NETIF_F_TSO6);
return 0;
}
static const struct ethtool_ops netdev_ethtool_ops = {
.get_settings = s2io_ethtool_gset,
.set_settings = s2io_ethtool_sset,
.get_drvinfo = s2io_ethtool_gdrvinfo,
.get_regs_len = s2io_ethtool_get_regs_len,
.get_regs = s2io_ethtool_gregs,
.get_link = ethtool_op_get_link,
.get_eeprom_len = s2io_get_eeprom_len,
.get_eeprom = s2io_ethtool_geeprom,
.set_eeprom = s2io_ethtool_seeprom,
.get_pauseparam = s2io_ethtool_getpause_data,
.set_pauseparam = s2io_ethtool_setpause_data,
.get_rx_csum = s2io_ethtool_get_rx_csum,
.set_rx_csum = s2io_ethtool_set_rx_csum,
.get_tx_csum = ethtool_op_get_tx_csum,
.set_tx_csum = s2io_ethtool_op_set_tx_csum,
.get_sg = ethtool_op_get_sg,
.set_sg = ethtool_op_set_sg,
.get_tso = s2io_ethtool_op_get_tso,
.set_tso = s2io_ethtool_op_set_tso,
.get_ufo = ethtool_op_get_ufo,
.set_ufo = ethtool_op_set_ufo,
.self_test_count = s2io_ethtool_self_test_count,
.self_test = s2io_ethtool_test,
.get_strings = s2io_ethtool_get_strings,
.phys_id = s2io_ethtool_idnic,
.get_stats_count = s2io_ethtool_get_stats_count,
.get_ethtool_stats = s2io_get_ethtool_stats
};
/**
* s2io_ioctl - Entry point for the Ioctl
* @dev : Device pointer.
* @ifr : An IOCTL specefic structure, that can contain a pointer to
* a proprietary structure used to pass information to the driver.
* @cmd : This is used to distinguish between the different commands that
* can be passed to the IOCTL functions.
* Description:
* Currently there are no special functionality supported in IOCTL, hence
* function always return EOPNOTSUPPORTED
*/
static int s2io_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
{
return -EOPNOTSUPP;
}
/**
* s2io_change_mtu - entry point to change MTU size for the device.
* @dev : device pointer.
* @new_mtu : the new MTU size for the device.
* Description: A driver entry point to change MTU size for the device.
* Before changing the MTU the device must be stopped.
* Return value:
* 0 on success and an appropriate (-)ve integer as defined in errno.h
* file on failure.
*/
static int s2io_change_mtu(struct net_device *dev, int new_mtu)
{
struct s2io_nic *sp = dev->priv;
if ((new_mtu < MIN_MTU) || (new_mtu > S2IO_JUMBO_SIZE)) {
DBG_PRINT(ERR_DBG, "%s: MTU size is invalid.\n",
dev->name);
return -EPERM;
}
dev->mtu = new_mtu;
if (netif_running(dev)) {
s2io_card_down(sp);
netif_stop_queue(dev);
if (s2io_card_up(sp)) {
DBG_PRINT(ERR_DBG, "%s: Device bring up failed\n",
__FUNCTION__);
}
if (netif_queue_stopped(dev))
netif_wake_queue(dev);
} else { /* Device is down */
struct XENA_dev_config __iomem *bar0 = sp->bar0;
u64 val64 = new_mtu;
writeq(vBIT(val64, 2, 14), &bar0->rmac_max_pyld_len);
}
return 0;
}
/**
* s2io_tasklet - Bottom half of the ISR.
* @dev_adr : address of the device structure in dma_addr_t format.
* Description:
* This is the tasklet or the bottom half of the ISR. This is
* an extension of the ISR which is scheduled by the scheduler to be run
* when the load on the CPU is low. All low priority tasks of the ISR can
* be pushed into the tasklet. For now the tasklet is used only to
* replenish the Rx buffers in the Rx buffer descriptors.
* Return value:
* void.
*/
static void s2io_tasklet(unsigned long dev_addr)
{
struct net_device *dev = (struct net_device *) dev_addr;
struct s2io_nic *sp = dev->priv;
int i, ret;
struct mac_info *mac_control;
struct config_param *config;
mac_control = &sp->mac_control;
config = &sp->config;
if (!TASKLET_IN_USE) {
for (i = 0; i < config->rx_ring_num; i++) {
ret = fill_rx_buffers(sp, i);
if (ret == -ENOMEM) {
DBG_PRINT(INFO_DBG, "%s: Out of ",
dev->name);
DBG_PRINT(ERR_DBG, "memory in tasklet\n");
break;
} else if (ret == -EFILL) {
DBG_PRINT(INFO_DBG,
"%s: Rx Ring %d is full\n",
dev->name, i);
break;
}
}
clear_bit(0, (&sp->tasklet_status));
}
}
/**
* s2io_set_link - Set the LInk status
* @data: long pointer to device private structue
* Description: Sets the link status for the adapter
*/
static void s2io_set_link(struct work_struct *work)
{
struct s2io_nic *nic = container_of(work, struct s2io_nic, set_link_task);
struct net_device *dev = nic->dev;
struct XENA_dev_config __iomem *bar0 = nic->bar0;
register u64 val64;
u16 subid;
rtnl_lock();
if (!netif_running(dev))
goto out_unlock;
if (test_and_set_bit(0, &(nic->link_state))) {
/* The card is being reset, no point doing anything */
goto out_unlock;
}
subid = nic->pdev->subsystem_device;
if (s2io_link_fault_indication(nic) == MAC_RMAC_ERR_TIMER) {
/*
* Allow a small delay for the NICs self initiated
* cleanup to complete.
*/
msleep(100);
}
val64 = readq(&bar0->adapter_status);
if (LINK_IS_UP(val64)) {
if (!(readq(&bar0->adapter_control) & ADAPTER_CNTL_EN)) {
if (verify_xena_quiescence(nic)) {
val64 = readq(&bar0->adapter_control);
val64 |= ADAPTER_CNTL_EN;
writeq(val64, &bar0->adapter_control);
if (CARDS_WITH_FAULTY_LINK_INDICATORS(
nic->device_type, subid)) {
val64 = readq(&bar0->gpio_control);
val64 |= GPIO_CTRL_GPIO_0;
writeq(val64, &bar0->gpio_control);
val64 = readq(&bar0->gpio_control);
} else {
val64 |= ADAPTER_LED_ON;
writeq(val64, &bar0->adapter_control);
}
nic->device_enabled_once = TRUE;
} else {
DBG_PRINT(ERR_DBG, "%s: Error: ", dev->name);
DBG_PRINT(ERR_DBG, "device is not Quiescent\n");
netif_stop_queue(dev);
}
}
val64 = readq(&bar0->adapter_status);
if (!LINK_IS_UP(val64)) {
DBG_PRINT(ERR_DBG, "%s:", dev->name);
DBG_PRINT(ERR_DBG, " Link down after enabling ");
DBG_PRINT(ERR_DBG, "device \n");
} else
s2io_link(nic, LINK_UP);
} else {
if (CARDS_WITH_FAULTY_LINK_INDICATORS(nic->device_type,
subid)) {
val64 = readq(&bar0->gpio_control);
val64 &= ~GPIO_CTRL_GPIO_0;
writeq(val64, &bar0->gpio_control);
val64 = readq(&bar0->gpio_control);
}
s2io_link(nic, LINK_DOWN);
}
clear_bit(0, &(nic->link_state));
out_unlock:
rtnl_unlock();
}
static int set_rxd_buffer_pointer(struct s2io_nic *sp, struct RxD_t *rxdp,
struct buffAdd *ba,
struct sk_buff **skb, u64 *temp0, u64 *temp1,
u64 *temp2, int size)
{
struct net_device *dev = sp->dev;
struct sk_buff *frag_list;
if ((sp->rxd_mode == RXD_MODE_1) && (rxdp->Host_Control == 0)) {
/* allocate skb */
if (*skb) {
DBG_PRINT(INFO_DBG, "SKB is not NULL\n");
/*
* As Rx frame are not going to be processed,
* using same mapped address for the Rxd
* buffer pointer
*/
((struct RxD1*)rxdp)->Buffer0_ptr = *temp0;
} else {
*skb = dev_alloc_skb(size);
if (!(*skb)) {
DBG_PRINT(INFO_DBG, "%s: Out of ", dev->name);
DBG_PRINT(INFO_DBG, "memory to allocate SKBs\n");
return -ENOMEM ;
}
/* storing the mapped addr in a temp variable
* such it will be used for next rxd whose
* Host Control is NULL
*/
((struct RxD1*)rxdp)->Buffer0_ptr = *temp0 =
pci_map_single( sp->pdev, (*skb)->data,
size - NET_IP_ALIGN,
PCI_DMA_FROMDEVICE);
rxdp->Host_Control = (unsigned long) (*skb);
}
} else if ((sp->rxd_mode == RXD_MODE_3B) && (rxdp->Host_Control == 0)) {
/* Two buffer Mode */
if (*skb) {
((struct RxD3*)rxdp)->Buffer2_ptr = *temp2;
((struct RxD3*)rxdp)->Buffer0_ptr = *temp0;
((struct RxD3*)rxdp)->Buffer1_ptr = *temp1;
} else {
*skb = dev_alloc_skb(size);
if (!(*skb)) {
DBG_PRINT(INFO_DBG, "%s: dev_alloc_skb failed\n",
dev->name);
return -ENOMEM;
}
((struct RxD3*)rxdp)->Buffer2_ptr = *temp2 =
pci_map_single(sp->pdev, (*skb)->data,
dev->mtu + 4,
PCI_DMA_FROMDEVICE);
((struct RxD3*)rxdp)->Buffer0_ptr = *temp0 =
pci_map_single( sp->pdev, ba->ba_0, BUF0_LEN,
PCI_DMA_FROMDEVICE);
rxdp->Host_Control = (unsigned long) (*skb);
/* Buffer-1 will be dummy buffer not used */
((struct RxD3*)rxdp)->Buffer1_ptr = *temp1 =
pci_map_single(sp->pdev, ba->ba_1, BUF1_LEN,
PCI_DMA_FROMDEVICE);
}
} else if ((rxdp->Host_Control == 0)) {
/* Three buffer mode */
if (*skb) {
((struct RxD3*)rxdp)->Buffer0_ptr = *temp0;
((struct RxD3*)rxdp)->Buffer1_ptr = *temp1;
((struct RxD3*)rxdp)->Buffer2_ptr = *temp2;
} else {
*skb = dev_alloc_skb(size);
if (!(*skb)) {
DBG_PRINT(INFO_DBG, "%s: dev_alloc_skb failed\n",
dev->name);
return -ENOMEM;
}
((struct RxD3*)rxdp)->Buffer0_ptr = *temp0 =
pci_map_single(sp->pdev, ba->ba_0, BUF0_LEN,
PCI_DMA_FROMDEVICE);
/* Buffer-1 receives L3/L4 headers */
((struct RxD3*)rxdp)->Buffer1_ptr = *temp1 =
pci_map_single( sp->pdev, (*skb)->data,
l3l4hdr_size + 4,
PCI_DMA_FROMDEVICE);
/*
* skb_shinfo(skb)->frag_list will have L4
* data payload
*/
skb_shinfo(*skb)->frag_list = dev_alloc_skb(dev->mtu +
ALIGN_SIZE);
if (skb_shinfo(*skb)->frag_list == NULL) {
DBG_PRINT(ERR_DBG, "%s: dev_alloc_skb \
failed\n ", dev->name);
return -ENOMEM ;
}
frag_list = skb_shinfo(*skb)->frag_list;
frag_list->next = NULL;
/*
* Buffer-2 receives L4 data payload
*/
((struct RxD3*)rxdp)->Buffer2_ptr = *temp2 =
pci_map_single( sp->pdev, frag_list->data,
dev->mtu, PCI_DMA_FROMDEVICE);
}
}
return 0;
}
static void set_rxd_buffer_size(struct s2io_nic *sp, struct RxD_t *rxdp,
int size)
{
struct net_device *dev = sp->dev;
if (sp->rxd_mode == RXD_MODE_1) {
rxdp->Control_2 = SET_BUFFER0_SIZE_1( size - NET_IP_ALIGN);
} else if (sp->rxd_mode == RXD_MODE_3B) {
rxdp->Control_2 = SET_BUFFER0_SIZE_3(BUF0_LEN);
rxdp->Control_2 |= SET_BUFFER1_SIZE_3(1);
rxdp->Control_2 |= SET_BUFFER2_SIZE_3( dev->mtu + 4);
} else {
rxdp->Control_2 = SET_BUFFER0_SIZE_3(BUF0_LEN);
rxdp->Control_2 |= SET_BUFFER1_SIZE_3(l3l4hdr_size + 4);
rxdp->Control_2 |= SET_BUFFER2_SIZE_3(dev->mtu);
}
}
static int rxd_owner_bit_reset(struct s2io_nic *sp)
{
int i, j, k, blk_cnt = 0, size;
struct mac_info * mac_control = &sp->mac_control;
struct config_param *config = &sp->config;
struct net_device *dev = sp->dev;
struct RxD_t *rxdp = NULL;
struct sk_buff *skb = NULL;
struct buffAdd *ba = NULL;
u64 temp0_64 = 0, temp1_64 = 0, temp2_64 = 0;
/* Calculate the size based on ring mode */
size = dev->mtu + HEADER_ETHERNET_II_802_3_SIZE +
HEADER_802_2_SIZE + HEADER_SNAP_SIZE;
if (sp->rxd_mode == RXD_MODE_1)
size += NET_IP_ALIGN;
else if (sp->rxd_mode == RXD_MODE_3B)
size = dev->mtu + ALIGN_SIZE + BUF0_LEN + 4;
else
size = l3l4hdr_size + ALIGN_SIZE + BUF0_LEN + 4;
for (i = 0; i < config->rx_ring_num; i++) {
blk_cnt = config->rx_cfg[i].num_rxd /
(rxd_count[sp->rxd_mode] +1);
for (j = 0; j < blk_cnt; j++) {
for (k = 0; k < rxd_count[sp->rxd_mode]; k++) {
rxdp = mac_control->rings[i].
rx_blocks[j].rxds[k].virt_addr;
if(sp->rxd_mode >= RXD_MODE_3A)
ba = &mac_control->rings[i].ba[j][k];
if (set_rxd_buffer_pointer(sp, rxdp, ba,
&skb,(u64 *)&temp0_64,
(u64 *)&temp1_64,
(u64 *)&temp2_64,
size) == ENOMEM) {
return 0;
}
set_rxd_buffer_size(sp, rxdp, size);
wmb();
/* flip the Ownership bit to Hardware */
rxdp->Control_1 |= RXD_OWN_XENA;
}
}
}
return 0;
}
static int s2io_add_isr(struct s2io_nic * sp)
{
int ret = 0;
struct net_device *dev = sp->dev;
int err = 0;
if (sp->intr_type == MSI)
ret = s2io_enable_msi(sp);
else if (sp->intr_type == MSI_X)
ret = s2io_enable_msi_x(sp);
if (ret) {
DBG_PRINT(ERR_DBG, "%s: Defaulting to INTA\n", dev->name);
sp->intr_type = INTA;
}
/* Store the values of the MSIX table in the struct s2io_nic structure */
store_xmsi_data(sp);
/* After proper initialization of H/W, register ISR */
if (sp->intr_type == MSI) {
err = request_irq((int) sp->pdev->irq, s2io_msi_handle,
IRQF_SHARED, sp->name, dev);
if (err) {
pci_disable_msi(sp->pdev);
DBG_PRINT(ERR_DBG, "%s: MSI registration failed\n",
dev->name);
return -1;
}
}
if (sp->intr_type == MSI_X) {
int i, msix_tx_cnt=0,msix_rx_cnt=0;
for (i=1; (sp->s2io_entries[i].in_use == MSIX_FLG); i++) {
if (sp->s2io_entries[i].type == MSIX_FIFO_TYPE) {
sprintf(sp->desc[i], "%s:MSI-X-%d-TX",
dev->name, i);
err = request_irq(sp->entries[i].vector,
s2io_msix_fifo_handle, 0, sp->desc[i],
sp->s2io_entries[i].arg);
/* If either data or addr is zero print it */
if(!(sp->msix_info[i].addr &&
sp->msix_info[i].data)) {
DBG_PRINT(ERR_DBG, "%s @ Addr:0x%llx"
"Data:0x%lx\n",sp->desc[i],
(unsigned long long)
sp->msix_info[i].addr,
(unsigned long)
ntohl(sp->msix_info[i].data));
} else {
msix_tx_cnt++;
}
} else {
sprintf(sp->desc[i], "%s:MSI-X-%d-RX",
dev->name, i);
err = request_irq(sp->entries[i].vector,
s2io_msix_ring_handle, 0, sp->desc[i],
sp->s2io_entries[i].arg);
/* If either data or addr is zero print it */
if(!(sp->msix_info[i].addr &&
sp->msix_info[i].data)) {
DBG_PRINT(ERR_DBG, "%s @ Addr:0x%llx"
"Data:0x%lx\n",sp->desc[i],
(unsigned long long)
sp->msix_info[i].addr,
(unsigned long)
ntohl(sp->msix_info[i].data));
} else {
msix_rx_cnt++;
}
}
if (err) {
DBG_PRINT(ERR_DBG,"%s:MSI-X-%d registration "
"failed\n", dev->name, i);
DBG_PRINT(ERR_DBG, "Returned: %d\n", err);
return -1;
}
sp->s2io_entries[i].in_use = MSIX_REGISTERED_SUCCESS;
}
printk("MSI-X-TX %d entries enabled\n",msix_tx_cnt);
printk("MSI-X-RX %d entries enabled\n",msix_rx_cnt);
}
if (sp->intr_type == INTA) {
err = request_irq((int) sp->pdev->irq, s2io_isr, IRQF_SHARED,
sp->name, dev);
if (err) {
DBG_PRINT(ERR_DBG, "%s: ISR registration failed\n",
dev->name);
return -1;
}
}
return 0;
}
static void s2io_rem_isr(struct s2io_nic * sp)
{
int cnt = 0;
struct net_device *dev = sp->dev;
if (sp->intr_type == MSI_X) {
int i;
u16 msi_control;
for (i=1; (sp->s2io_entries[i].in_use ==
MSIX_REGISTERED_SUCCESS); i++) {
int vector = sp->entries[i].vector;
void *arg = sp->s2io_entries[i].arg;
free_irq(vector, arg);
}
pci_read_config_word(sp->pdev, 0x42, &msi_control);
msi_control &= 0xFFFE; /* Disable MSI */
pci_write_config_word(sp->pdev, 0x42, msi_control);
pci_disable_msix(sp->pdev);
} else {
free_irq(sp->pdev->irq, dev);
if (sp->intr_type == MSI) {
u16 val;
pci_disable_msi(sp->pdev);
pci_read_config_word(sp->pdev, 0x4c, &val);
val ^= 0x1;
pci_write_config_word(sp->pdev, 0x4c, val);
}
}
/* Waiting till all Interrupt handlers are complete */
cnt = 0;
do {
msleep(10);
if (!atomic_read(&sp->isr_cnt))
break;
cnt++;
} while(cnt < 5);
}
static void s2io_card_down(struct s2io_nic * sp)
{
int cnt = 0;
struct XENA_dev_config __iomem *bar0 = sp->bar0;
unsigned long flags;
register u64 val64 = 0;
del_timer_sync(&sp->alarm_timer);
/* If s2io_set_link task is executing, wait till it completes. */
while (test_and_set_bit(0, &(sp->link_state))) {
msleep(50);
}
atomic_set(&sp->card_state, CARD_DOWN);
/* disable Tx and Rx traffic on the NIC */
stop_nic(sp);
s2io_rem_isr(sp);
/* Kill tasklet. */
tasklet_kill(&sp->task);
/* Check if the device is Quiescent and then Reset the NIC */
do {
/* As per the HW requirement we need to replenish the
* receive buffer to avoid the ring bump. Since there is
* no intention of processing the Rx frame at this pointwe are
* just settting the ownership bit of rxd in Each Rx
* ring to HW and set the appropriate buffer size
* based on the ring mode
*/
rxd_owner_bit_reset(sp);
val64 = readq(&bar0->adapter_status);
if (verify_xena_quiescence(sp)) {
if(verify_pcc_quiescent(sp, sp->device_enabled_once))
break;
}
msleep(50);
cnt++;
if (cnt == 10) {
DBG_PRINT(ERR_DBG,
"s2io_close:Device not Quiescent ");
DBG_PRINT(ERR_DBG, "adaper status reads 0x%llx\n",
(unsigned long long) val64);
break;
}
} while (1);
s2io_reset(sp);
spin_lock_irqsave(&sp->tx_lock, flags);
/* Free all Tx buffers */
free_tx_buffers(sp);
spin_unlock_irqrestore(&sp->tx_lock, flags);
/* Free all Rx buffers */
spin_lock_irqsave(&sp->rx_lock, flags);
free_rx_buffers(sp);
spin_unlock_irqrestore(&sp->rx_lock, flags);
clear_bit(0, &(sp->link_state));
}
static int s2io_card_up(struct s2io_nic * sp)
{
int i, ret = 0;
struct mac_info *mac_control;
struct config_param *config;
struct net_device *dev = (struct net_device *) sp->dev;
u16 interruptible;
/* Initialize the H/W I/O registers */
if (init_nic(sp) != 0) {
DBG_PRINT(ERR_DBG, "%s: H/W initialization failed\n",
dev->name);
s2io_reset(sp);
return -ENODEV;
}
/*
* Initializing the Rx buffers. For now we are considering only 1
* Rx ring and initializing buffers into 30 Rx blocks
*/
mac_control = &sp->mac_control;
config = &sp->config;
for (i = 0; i < config->rx_ring_num; i++) {
if ((ret = fill_rx_buffers(sp, i))) {
DBG_PRINT(ERR_DBG, "%s: Out of memory in Open\n",
dev->name);
s2io_reset(sp);
free_rx_buffers(sp);
return -ENOMEM;
}
DBG_PRINT(INFO_DBG, "Buf in ring:%d is %d:\n", i,
atomic_read(&sp->rx_bufs_left[i]));
}
/* Maintain the state prior to the open */
if (sp->promisc_flg)
sp->promisc_flg = 0;
if (sp->m_cast_flg) {
sp->m_cast_flg = 0;
sp->all_multi_pos= 0;
}
/* Setting its receive mode */
s2io_set_multicast(dev);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (sp->lro) {
/* Initialize max aggregatable pkts per session based on MTU */
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
sp->lro_max_aggr_per_sess = ((1<<16) - 1) / dev->mtu;
/* Check if we can use(if specified) user provided value */
if (lro_max_pkts < sp->lro_max_aggr_per_sess)
sp->lro_max_aggr_per_sess = lro_max_pkts;
}
/* Enable Rx Traffic and interrupts on the NIC */
if (start_nic(sp)) {
DBG_PRINT(ERR_DBG, "%s: Starting NIC failed\n", dev->name);
s2io_reset(sp);
free_rx_buffers(sp);
return -ENODEV;
}
/* Add interrupt service routine */
if (s2io_add_isr(sp) != 0) {
if (sp->intr_type == MSI_X)
s2io_rem_isr(sp);
s2io_reset(sp);
free_rx_buffers(sp);
return -ENODEV;
}
S2IO_TIMER_CONF(sp->alarm_timer, s2io_alarm_handle, sp, (HZ/2));
/* Enable tasklet for the device */
tasklet_init(&sp->task, s2io_tasklet, (unsigned long) dev);
/* Enable select interrupts */
if (sp->intr_type != INTA)
en_dis_able_nic_intrs(sp, ENA_ALL_INTRS, DISABLE_INTRS);
else {
interruptible = TX_TRAFFIC_INTR | RX_TRAFFIC_INTR;
interruptible |= TX_PIC_INTR | RX_PIC_INTR;
interruptible |= TX_MAC_INTR | RX_MAC_INTR;
en_dis_able_nic_intrs(sp, interruptible, ENABLE_INTRS);
}
atomic_set(&sp->card_state, CARD_UP);
return 0;
}
/**
* s2io_restart_nic - Resets the NIC.
* @data : long pointer to the device private structure
* Description:
* This function is scheduled to be run by the s2io_tx_watchdog
* function after 0.5 secs to reset the NIC. The idea is to reduce
* the run time of the watch dog routine which is run holding a
* spin lock.
*/
static void s2io_restart_nic(struct work_struct *work)
{
struct s2io_nic *sp = container_of(work, struct s2io_nic, rst_timer_task);
struct net_device *dev = sp->dev;
rtnl_lock();
if (!netif_running(dev))
goto out_unlock;
s2io_card_down(sp);
if (s2io_card_up(sp)) {
DBG_PRINT(ERR_DBG, "%s: Device bring up failed\n",
dev->name);
}
netif_wake_queue(dev);
DBG_PRINT(ERR_DBG, "%s: was reset by Tx watchdog timer\n",
dev->name);
out_unlock:
rtnl_unlock();
}
/**
* s2io_tx_watchdog - Watchdog for transmit side.
* @dev : Pointer to net device structure
* Description:
* This function is triggered if the Tx Queue is stopped
* for a pre-defined amount of time when the Interface is still up.
* If the Interface is jammed in such a situation, the hardware is
* reset (by s2io_close) and restarted again (by s2io_open) to
* overcome any problem that might have been caused in the hardware.
* Return value:
* void
*/
static void s2io_tx_watchdog(struct net_device *dev)
{
struct s2io_nic *sp = dev->priv;
if (netif_carrier_ok(dev)) {
schedule_work(&sp->rst_timer_task);
sp->mac_control.stats_info->sw_stat.soft_reset_cnt++;
}
}
/**
* rx_osm_handler - To perform some OS related operations on SKB.
* @sp: private member of the device structure,pointer to s2io_nic structure.
* @skb : the socket buffer pointer.
* @len : length of the packet
* @cksum : FCS checksum of the frame.
* @ring_no : the ring from which this RxD was extracted.
* Description:
* This function is called by the Rx interrupt serivce routine to perform
* some OS related operations on the SKB before passing it to the upper
* layers. It mainly checks if the checksum is OK, if so adds it to the
* SKBs cksum variable, increments the Rx packet count and passes the SKB
* to the upper layer. If the checksum is wrong, it increments the Rx
* packet error count, frees the SKB and returns error.
* Return value:
* SUCCESS on success and -1 on failure.
*/
static int rx_osm_handler(struct ring_info *ring_data, struct RxD_t * rxdp)
{
struct s2io_nic *sp = ring_data->nic;
struct net_device *dev = (struct net_device *) sp->dev;
struct sk_buff *skb = (struct sk_buff *)
((unsigned long) rxdp->Host_Control);
int ring_no = ring_data->ring_no;
u16 l3_csum, l4_csum;
unsigned long long err = rxdp->Control_1 & RXD_T_CODE;
struct lro *lro;
skb->dev = dev;
if (err) {
/* Check for parity error */
if (err & 0x1) {
sp->mac_control.stats_info->sw_stat.parity_err_cnt++;
}
/*
* Drop the packet if bad transfer code. Exception being
* 0x5, which could be due to unsupported IPv6 extension header.
* In this case, we let stack handle the packet.
* Note that in this case, since checksum will be incorrect,
* stack will validate the same.
*/
if (err && ((err >> 48) != 0x5)) {
DBG_PRINT(ERR_DBG, "%s: Rx error Value: 0x%llx\n",
dev->name, err);
sp->stats.rx_crc_errors++;
dev_kfree_skb(skb);
atomic_dec(&sp->rx_bufs_left[ring_no]);
rxdp->Host_Control = 0;
return 0;
}
}
/* Updating statistics */
rxdp->Host_Control = 0;
sp->stats.rx_packets++;
if (sp->rxd_mode == RXD_MODE_1) {
int len = RXD_GET_BUFFER0_SIZE_1(rxdp->Control_2);
sp->stats.rx_bytes += len;
skb_put(skb, len);
} else if (sp->rxd_mode >= RXD_MODE_3A) {
int get_block = ring_data->rx_curr_get_info.block_index;
int get_off = ring_data->rx_curr_get_info.offset;
int buf0_len = RXD_GET_BUFFER0_SIZE_3(rxdp->Control_2);
int buf2_len = RXD_GET_BUFFER2_SIZE_3(rxdp->Control_2);
unsigned char *buff = skb_push(skb, buf0_len);
struct buffAdd *ba = &ring_data->ba[get_block][get_off];
sp->stats.rx_bytes += buf0_len + buf2_len;
memcpy(buff, ba->ba_0, buf0_len);
if (sp->rxd_mode == RXD_MODE_3A) {
int buf1_len = RXD_GET_BUFFER1_SIZE_3(rxdp->Control_2);
skb_put(skb, buf1_len);
skb->len += buf2_len;
skb->data_len += buf2_len;
skb_put(skb_shinfo(skb)->frag_list, buf2_len);
sp->stats.rx_bytes += buf1_len;
} else
skb_put(skb, buf2_len);
}
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if ((rxdp->Control_1 & TCP_OR_UDP_FRAME) && ((!sp->lro) ||
(sp->lro && (!(rxdp->Control_1 & RXD_FRAME_IP_FRAG)))) &&
(sp->rx_csum)) {
l3_csum = RXD_GET_L3_CKSUM(rxdp->Control_1);
l4_csum = RXD_GET_L4_CKSUM(rxdp->Control_1);
if ((l3_csum == L3_CKSUM_OK) && (l4_csum == L4_CKSUM_OK)) {
/*
* NIC verifies if the Checksum of the received
* frame is Ok or not and accordingly returns
* a flag in the RxD.
*/
skb->ip_summed = CHECKSUM_UNNECESSARY;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (sp->lro) {
u32 tcp_len;
u8 *tcp;
int ret = 0;
ret = s2io_club_tcp_session(skb->data, &tcp,
&tcp_len, &lro, rxdp, sp);
switch (ret) {
case 3: /* Begin anew */
lro->parent = skb;
goto aggregate;
case 1: /* Aggregate */
{
lro_append_pkt(sp, lro,
skb, tcp_len);
goto aggregate;
}
case 4: /* Flush session */
{
lro_append_pkt(sp, lro,
skb, tcp_len);
queue_rx_frame(lro->parent);
clear_lro_session(lro);
sp->mac_control.stats_info->
sw_stat.flush_max_pkts++;
goto aggregate;
}
case 2: /* Flush both */
lro->parent->data_len =
lro->frags_len;
sp->mac_control.stats_info->
sw_stat.sending_both++;
queue_rx_frame(lro->parent);
clear_lro_session(lro);
goto send_up;
case 0: /* sessions exceeded */
case -1: /* non-TCP or not
* L2 aggregatable
*/
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
case 5: /*
* First pkt in session not
* L3/L4 aggregatable
*/
break;
default:
DBG_PRINT(ERR_DBG,
"%s: Samadhana!!\n",
__FUNCTION__);
BUG();
}
}
} else {
/*
* Packet with erroneous checksum, let the
* upper layers deal with it.
*/
skb->ip_summed = CHECKSUM_NONE;
}
} else {
skb->ip_summed = CHECKSUM_NONE;
}
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (!sp->lro) {
skb->protocol = eth_type_trans(skb, dev);
if ((sp->vlgrp && RXD_GET_VLAN_TAG(rxdp->Control_2) &&
vlan_strip_flag)) {
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
/* Queueing the vlan frame to the upper layer */
if (napi)
vlan_hwaccel_receive_skb(skb, sp->vlgrp,
RXD_GET_VLAN_TAG(rxdp->Control_2));
else
vlan_hwaccel_rx(skb, sp->vlgrp,
RXD_GET_VLAN_TAG(rxdp->Control_2));
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
} else {
if (napi)
netif_receive_skb(skb);
else
netif_rx(skb);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
}
} else {
send_up:
queue_rx_frame(skb);
}
dev->last_rx = jiffies;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
aggregate:
atomic_dec(&sp->rx_bufs_left[ring_no]);
return SUCCESS;
}
/**
* s2io_link - stops/starts the Tx queue.
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* @link : inidicates whether link is UP/DOWN.
* Description:
* This function stops/starts the Tx queue depending on whether the link
* status of the NIC is is down or up. This is called by the Alarm
* interrupt handler whenever a link change interrupt comes up.
* Return value:
* void.
*/
static void s2io_link(struct s2io_nic * sp, int link)
{
struct net_device *dev = (struct net_device *) sp->dev;
if (link != sp->last_link_state) {
if (link == LINK_DOWN) {
DBG_PRINT(ERR_DBG, "%s: Link down\n", dev->name);
netif_carrier_off(dev);
} else {
DBG_PRINT(ERR_DBG, "%s: Link Up\n", dev->name);
netif_carrier_on(dev);
}
}
sp->last_link_state = link;
}
/**
* get_xena_rev_id - to identify revision ID of xena.
* @pdev : PCI Dev structure
* Description:
* Function to identify the Revision ID of xena.
* Return value:
* returns the revision ID of the device.
*/
static int get_xena_rev_id(struct pci_dev *pdev)
{
u8 id = 0;
int ret;
ret = pci_read_config_byte(pdev, PCI_REVISION_ID, (u8 *) & id);
return id;
}
/**
* s2io_init_pci -Initialization of PCI and PCI-X configuration registers .
* @sp : private member of the device structure, which is a pointer to the
* s2io_nic structure.
* Description:
* This function initializes a few of the PCI and PCI-X configuration registers
* with recommended values.
* Return value:
* void
*/
static void s2io_init_pci(struct s2io_nic * sp)
{
u16 pci_cmd = 0, pcix_cmd = 0;
/* Enable Data Parity Error Recovery in PCI-X command register. */
pci_read_config_word(sp->pdev, PCIX_COMMAND_REGISTER,
&(pcix_cmd));
pci_write_config_word(sp->pdev, PCIX_COMMAND_REGISTER,
(pcix_cmd | 1));
pci_read_config_word(sp->pdev, PCIX_COMMAND_REGISTER,
&(pcix_cmd));
/* Set the PErr Response bit in PCI command register. */
pci_read_config_word(sp->pdev, PCI_COMMAND, &pci_cmd);
pci_write_config_word(sp->pdev, PCI_COMMAND,
(pci_cmd | PCI_COMMAND_PARITY));
pci_read_config_word(sp->pdev, PCI_COMMAND, &pci_cmd);
}
static int s2io_verify_parm(struct pci_dev *pdev, u8 *dev_intr_type)
{
if ( tx_fifo_num > 8) {
DBG_PRINT(ERR_DBG, "s2io: Requested number of Tx fifos not "
"supported\n");
DBG_PRINT(ERR_DBG, "s2io: Default to 8 Tx fifos\n");
tx_fifo_num = 8;
}
if ( rx_ring_num > 8) {
DBG_PRINT(ERR_DBG, "s2io: Requested number of Rx rings not "
"supported\n");
DBG_PRINT(ERR_DBG, "s2io: Default to 8 Rx rings\n");
rx_ring_num = 8;
}
if (*dev_intr_type != INTA)
napi = 0;
#ifndef CONFIG_PCI_MSI
if (*dev_intr_type != INTA) {
DBG_PRINT(ERR_DBG, "s2io: This kernel does not support"
"MSI/MSI-X. Defaulting to INTA\n");
*dev_intr_type = INTA;
}
#else
if (*dev_intr_type > MSI_X) {
DBG_PRINT(ERR_DBG, "s2io: Wrong intr_type requested. "
"Defaulting to INTA\n");
*dev_intr_type = INTA;
}
#endif
if ((*dev_intr_type == MSI_X) &&
((pdev->device != PCI_DEVICE_ID_HERC_WIN) &&
(pdev->device != PCI_DEVICE_ID_HERC_UNI))) {
DBG_PRINT(ERR_DBG, "s2io: Xframe I does not support MSI_X. "
"Defaulting to INTA\n");
*dev_intr_type = INTA;
}
if (rx_ring_mode > 3) {
DBG_PRINT(ERR_DBG, "s2io: Requested ring mode not supported\n");
DBG_PRINT(ERR_DBG, "s2io: Defaulting to 3-buffer mode\n");
rx_ring_mode = 3;
}
return SUCCESS;
}
/**
* rts_ds_steer - Receive traffic steering based on IPv4 or IPv6 TOS
* or Traffic class respectively.
* @nic: device peivate variable
* Description: The function configures the receive steering to
* desired receive ring.
* Return Value: SUCCESS on success and
* '-1' on failure (endian settings incorrect).
*/
static int rts_ds_steer(struct s2io_nic *nic, u8 ds_codepoint, u8 ring)
{
struct XENA_dev_config __iomem *bar0 = nic->bar0;
register u64 val64 = 0;
if (ds_codepoint > 63)
return FAILURE;
val64 = RTS_DS_MEM_DATA(ring);
writeq(val64, &bar0->rts_ds_mem_data);
val64 = RTS_DS_MEM_CTRL_WE |
RTS_DS_MEM_CTRL_STROBE_NEW_CMD |
RTS_DS_MEM_CTRL_OFFSET(ds_codepoint);
writeq(val64, &bar0->rts_ds_mem_ctrl);
return wait_for_cmd_complete(&bar0->rts_ds_mem_ctrl,
RTS_DS_MEM_CTRL_STROBE_CMD_BEING_EXECUTED,
S2IO_BIT_RESET);
}
/**
* s2io_init_nic - Initialization of the adapter .
* @pdev : structure containing the PCI related information of the device.
* @pre: List of PCI devices supported by the driver listed in s2io_tbl.
* Description:
* The function initializes an adapter identified by the pci_dec structure.
* All OS related initialization including memory and device structure and
* initlaization of the device private variable is done. Also the swapper
* control register is initialized to enable read and write into the I/O
* registers of the device.
* Return value:
* returns 0 on success and negative on failure.
*/
static int __devinit
s2io_init_nic(struct pci_dev *pdev, const struct pci_device_id *pre)
{
struct s2io_nic *sp;
struct net_device *dev;
int i, j, ret;
int dma_flag = FALSE;
u32 mac_up, mac_down;
u64 val64 = 0, tmp64 = 0;
struct XENA_dev_config __iomem *bar0 = NULL;
u16 subid;
struct mac_info *mac_control;
struct config_param *config;
int mode;
u8 dev_intr_type = intr_type;
if ((ret = s2io_verify_parm(pdev, &dev_intr_type)))
return ret;
if ((ret = pci_enable_device(pdev))) {
DBG_PRINT(ERR_DBG,
"s2io_init_nic: pci_enable_device failed\n");
return ret;
}
if (!pci_set_dma_mask(pdev, DMA_64BIT_MASK)) {
DBG_PRINT(INIT_DBG, "s2io_init_nic: Using 64bit DMA\n");
dma_flag = TRUE;
if (pci_set_consistent_dma_mask
(pdev, DMA_64BIT_MASK)) {
DBG_PRINT(ERR_DBG,
"Unable to obtain 64bit DMA for \
consistent allocations\n");
pci_disable_device(pdev);
return -ENOMEM;
}
} else if (!pci_set_dma_mask(pdev, DMA_32BIT_MASK)) {
DBG_PRINT(INIT_DBG, "s2io_init_nic: Using 32bit DMA\n");
} else {
pci_disable_device(pdev);
return -ENOMEM;
}
if (dev_intr_type != MSI_X) {
if (pci_request_regions(pdev, s2io_driver_name)) {
DBG_PRINT(ERR_DBG, "Request Regions failed\n");
pci_disable_device(pdev);
return -ENODEV;
}
}
else {
if (!(request_mem_region(pci_resource_start(pdev, 0),
pci_resource_len(pdev, 0), s2io_driver_name))) {
DBG_PRINT(ERR_DBG, "bar0 Request Regions failed\n");
pci_disable_device(pdev);
return -ENODEV;
}
if (!(request_mem_region(pci_resource_start(pdev, 2),
pci_resource_len(pdev, 2), s2io_driver_name))) {
DBG_PRINT(ERR_DBG, "bar1 Request Regions failed\n");
release_mem_region(pci_resource_start(pdev, 0),
pci_resource_len(pdev, 0));
pci_disable_device(pdev);
return -ENODEV;
}
}
dev = alloc_etherdev(sizeof(struct s2io_nic));
if (dev == NULL) {
DBG_PRINT(ERR_DBG, "Device allocation failed\n");
pci_disable_device(pdev);
pci_release_regions(pdev);
return -ENODEV;
}
pci_set_master(pdev);
pci_set_drvdata(pdev, dev);
SET_MODULE_OWNER(dev);
SET_NETDEV_DEV(dev, &pdev->dev);
/* Private member variable initialized to s2io NIC structure */
sp = dev->priv;
memset(sp, 0, sizeof(struct s2io_nic));
sp->dev = dev;
sp->pdev = pdev;
sp->high_dma_flag = dma_flag;
sp->device_enabled_once = FALSE;
if (rx_ring_mode == 1)
sp->rxd_mode = RXD_MODE_1;
if (rx_ring_mode == 2)
sp->rxd_mode = RXD_MODE_3B;
if (rx_ring_mode == 3)
sp->rxd_mode = RXD_MODE_3A;
sp->intr_type = dev_intr_type;
if ((pdev->device == PCI_DEVICE_ID_HERC_WIN) ||
(pdev->device == PCI_DEVICE_ID_HERC_UNI))
sp->device_type = XFRAME_II_DEVICE;
else
sp->device_type = XFRAME_I_DEVICE;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
sp->lro = lro;
/* Initialize some PCI/PCI-X fields of the NIC. */
s2io_init_pci(sp);
/*
* Setting the device configuration parameters.
* Most of these parameters can be specified by the user during
* module insertion as they are module loadable parameters. If
* these parameters are not not specified during load time, they
* are initialized with default values.
*/
mac_control = &sp->mac_control;
config = &sp->config;
/* Tx side parameters. */
config->tx_fifo_num = tx_fifo_num;
for (i = 0; i < MAX_TX_FIFOS; i++) {
config->tx_cfg[i].fifo_len = tx_fifo_len[i];
config->tx_cfg[i].fifo_priority = i;
}
/* mapping the QoS priority to the configured fifos */
for (i = 0; i < MAX_TX_FIFOS; i++)
config->fifo_mapping[i] = fifo_map[config->tx_fifo_num][i];
config->tx_intr_type = TXD_INT_TYPE_UTILZ;
for (i = 0; i < config->tx_fifo_num; i++) {
config->tx_cfg[i].f_no_snoop =
(NO_SNOOP_TXD | NO_SNOOP_TXD_BUFFER);
if (config->tx_cfg[i].fifo_len < 65) {
config->tx_intr_type = TXD_INT_TYPE_PER_LIST;
break;
}
}
/* + 2 because one Txd for skb->data and one Txd for UFO */
config->max_txds = MAX_SKB_FRAGS + 2;
/* Rx side parameters. */
config->rx_ring_num = rx_ring_num;
for (i = 0; i < MAX_RX_RINGS; i++) {
config->rx_cfg[i].num_rxd = rx_ring_sz[i] *
(rxd_count[sp->rxd_mode] + 1);
config->rx_cfg[i].ring_priority = i;
}
for (i = 0; i < rx_ring_num; i++) {
config->rx_cfg[i].ring_org = RING_ORG_BUFF1;
config->rx_cfg[i].f_no_snoop =
(NO_SNOOP_RXD | NO_SNOOP_RXD_BUFFER);
}
/* Setting Mac Control parameters */
mac_control->rmac_pause_time = rmac_pause_time;
mac_control->mc_pause_threshold_q0q3 = mc_pause_threshold_q0q3;
mac_control->mc_pause_threshold_q4q7 = mc_pause_threshold_q4q7;
/* Initialize Ring buffer parameters. */
for (i = 0; i < config->rx_ring_num; i++)
atomic_set(&sp->rx_bufs_left[i], 0);
/* Initialize the number of ISRs currently running */
atomic_set(&sp->isr_cnt, 0);
/* initialize the shared memory used by the NIC and the host */
if (init_shared_mem(sp)) {
DBG_PRINT(ERR_DBG, "%s: Memory allocation failed\n",
dev->name);
ret = -ENOMEM;
goto mem_alloc_failed;
}
sp->bar0 = ioremap(pci_resource_start(pdev, 0),
pci_resource_len(pdev, 0));
if (!sp->bar0) {
DBG_PRINT(ERR_DBG, "%s: Neterion: cannot remap io mem1\n",
dev->name);
ret = -ENOMEM;
goto bar0_remap_failed;
}
sp->bar1 = ioremap(pci_resource_start(pdev, 2),
pci_resource_len(pdev, 2));
if (!sp->bar1) {
DBG_PRINT(ERR_DBG, "%s: Neterion: cannot remap io mem2\n",
dev->name);
ret = -ENOMEM;
goto bar1_remap_failed;
}
dev->irq = pdev->irq;
dev->base_addr = (unsigned long) sp->bar0;
/* Initializing the BAR1 address as the start of the FIFO pointer. */
for (j = 0; j < MAX_TX_FIFOS; j++) {
mac_control->tx_FIFO_start[j] = (struct TxFIFO_element __iomem *)
(sp->bar1 + (j * 0x00020000));
}
/* Driver entry points */
dev->open = &s2io_open;
dev->stop = &s2io_close;
dev->hard_start_xmit = &s2io_xmit;
dev->get_stats = &s2io_get_stats;
dev->set_multicast_list = &s2io_set_multicast;
dev->do_ioctl = &s2io_ioctl;
dev->change_mtu = &s2io_change_mtu;
SET_ETHTOOL_OPS(dev, &netdev_ethtool_ops);
dev->features |= NETIF_F_HW_VLAN_TX | NETIF_F_HW_VLAN_RX;
dev->vlan_rx_register = s2io_vlan_rx_register;
dev->vlan_rx_kill_vid = (void *)s2io_vlan_rx_kill_vid;
/*
* will use eth_mac_addr() for dev->set_mac_address
* mac address will be set every time dev->open() is called
*/
dev->poll = s2io_poll;
dev->weight = 32;
#ifdef CONFIG_NET_POLL_CONTROLLER
dev->poll_controller = s2io_netpoll;
#endif
dev->features |= NETIF_F_SG | NETIF_F_IP_CSUM;
if (sp->high_dma_flag == TRUE)
dev->features |= NETIF_F_HIGHDMA;
dev->features |= NETIF_F_TSO;
[IPV6]: Added GSO support for TCPv6 This patch adds GSO support for IPv6 and TCPv6. This is based on a patch by Ananda Raju <Ananda.Raju@neterion.com>. His original description is: This patch enables TSO over IPv6. Currently Linux network stacks restricts TSO over IPv6 by clearing of the NETIF_F_TSO bit from "dev->features". This patch will remove this restriction. This patch will introduce a new flag NETIF_F_TSO6 which will be used to check whether device supports TSO over IPv6. If device support TSO over IPv6 then we don't clear of NETIF_F_TSO and which will make the TCP layer to create TSO packets. Any device supporting TSO over IPv6 will set NETIF_F_TSO6 flag in "dev->features" along with NETIF_F_TSO. In case when user disables TSO using ethtool, NETIF_F_TSO will get cleared from "dev->features". So even if we have NETIF_F_TSO6 we don't get TSO packets created by TCP layer. SKB_GSO_TCPV4 renamed to SKB_GSO_TCP to make it generic GSO packet. SKB_GSO_UDPV4 renamed to SKB_GSO_UDP as UFO is not a IPv4 feature. UFO is supported over IPv6 also The following table shows there is significant improvement in throughput with normal frames and CPU usage for both normal and jumbo. -------------------------------------------------- | | 1500 | 9600 | | ------------------|-------------------| | | thru CPU | thru CPU | -------------------------------------------------- | TSO OFF | 2.00 5.5% id | 5.66 20.0% id | -------------------------------------------------- | TSO ON | 2.63 78.0 id | 5.67 39.0% id | -------------------------------------------------- Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: David S. Miller <davem@davemloft.net>
2006-06-30 20:37:03 +00:00
dev->features |= NETIF_F_TSO6;
if ((sp->device_type & XFRAME_II_DEVICE) && (ufo)) {
dev->features |= NETIF_F_UFO;
dev->features |= NETIF_F_HW_CSUM;
}
dev->tx_timeout = &s2io_tx_watchdog;
dev->watchdog_timeo = WATCH_DOG_TIMEOUT;
INIT_WORK(&sp->rst_timer_task, s2io_restart_nic);
INIT_WORK(&sp->set_link_task, s2io_set_link);
pci_save_state(sp->pdev);
/* Setting swapper control on the NIC, for proper reset operation */
if (s2io_set_swapper(sp)) {
DBG_PRINT(ERR_DBG, "%s:swapper settings are wrong\n",
dev->name);
ret = -EAGAIN;
goto set_swap_failed;
}
/* Verify if the Herc works on the slot its placed into */
if (sp->device_type & XFRAME_II_DEVICE) {
mode = s2io_verify_pci_mode(sp);
if (mode < 0) {
DBG_PRINT(ERR_DBG, "%s: ", __FUNCTION__);
DBG_PRINT(ERR_DBG, " Unsupported PCI bus mode\n");
ret = -EBADSLT;
goto set_swap_failed;
}
}
/* Not needed for Herc */
if (sp->device_type & XFRAME_I_DEVICE) {
/*
* Fix for all "FFs" MAC address problems observed on
* Alpha platforms
*/
fix_mac_address(sp);
s2io_reset(sp);
}
/*
* MAC address initialization.
* For now only one mac address will be read and used.
*/
bar0 = sp->bar0;
val64 = RMAC_ADDR_CMD_MEM_RD | RMAC_ADDR_CMD_MEM_STROBE_NEW_CMD |
RMAC_ADDR_CMD_MEM_OFFSET(0 + MAC_MAC_ADDR_START_OFFSET);
writeq(val64, &bar0->rmac_addr_cmd_mem);
wait_for_cmd_complete(&bar0->rmac_addr_cmd_mem,
RMAC_ADDR_CMD_MEM_STROBE_CMD_EXECUTING, S2IO_BIT_RESET);
tmp64 = readq(&bar0->rmac_addr_data0_mem);
mac_down = (u32) tmp64;
mac_up = (u32) (tmp64 >> 32);
sp->def_mac_addr[0].mac_addr[3] = (u8) (mac_up);
sp->def_mac_addr[0].mac_addr[2] = (u8) (mac_up >> 8);
sp->def_mac_addr[0].mac_addr[1] = (u8) (mac_up >> 16);
sp->def_mac_addr[0].mac_addr[0] = (u8) (mac_up >> 24);
sp->def_mac_addr[0].mac_addr[5] = (u8) (mac_down >> 16);
sp->def_mac_addr[0].mac_addr[4] = (u8) (mac_down >> 24);
/* Set the factory defined MAC address initially */
dev->addr_len = ETH_ALEN;
memcpy(dev->dev_addr, sp->def_mac_addr, ETH_ALEN);
/* reset Nic and bring it to known state */
s2io_reset(sp);
/*
* Initialize the tasklet status and link state flags
* and the card state parameter
*/
atomic_set(&(sp->card_state), 0);
sp->tasklet_status = 0;
sp->link_state = 0;
/* Initialize spinlocks */
spin_lock_init(&sp->tx_lock);
if (!napi)
spin_lock_init(&sp->put_lock);
spin_lock_init(&sp->rx_lock);
/*
* SXE-002: Configure link and activity LED to init state
* on driver load.
*/
subid = sp->pdev->subsystem_device;
if ((subid & 0xFF) >= 0x07) {
val64 = readq(&bar0->gpio_control);
val64 |= 0x0000800000000000ULL;
writeq(val64, &bar0->gpio_control);
val64 = 0x0411040400000000ULL;
writeq(val64, (void __iomem *) bar0 + 0x2700);
val64 = readq(&bar0->gpio_control);
}
sp->rx_csum = 1; /* Rx chksum verify enabled by default */
if (register_netdev(dev)) {
DBG_PRINT(ERR_DBG, "Device registration failed\n");
ret = -ENODEV;
goto register_failed;
}
s2io_vpd_read(sp);
DBG_PRINT(ERR_DBG, "Copyright(c) 2002-2007 Neterion Inc.\n");
DBG_PRINT(ERR_DBG, "%s: Neterion %s (rev %d)\n",dev->name,
sp->product_name, get_xena_rev_id(sp->pdev));
DBG_PRINT(ERR_DBG, "%s: Driver version %s\n", dev->name,
s2io_driver_version);
DBG_PRINT(ERR_DBG, "%s: MAC ADDR: "
"%02x:%02x:%02x:%02x:%02x:%02x", dev->name,
sp->def_mac_addr[0].mac_addr[0],
sp->def_mac_addr[0].mac_addr[1],
sp->def_mac_addr[0].mac_addr[2],
sp->def_mac_addr[0].mac_addr[3],
sp->def_mac_addr[0].mac_addr[4],
sp->def_mac_addr[0].mac_addr[5]);
DBG_PRINT(ERR_DBG, "SERIAL NUMBER: %s\n", sp->serial_num);
if (sp->device_type & XFRAME_II_DEVICE) {
mode = s2io_print_pci_mode(sp);
if (mode < 0) {
DBG_PRINT(ERR_DBG, " Unsupported PCI bus mode\n");
ret = -EBADSLT;
unregister_netdev(dev);
goto set_swap_failed;
}
}
switch(sp->rxd_mode) {
case RXD_MODE_1:
DBG_PRINT(ERR_DBG, "%s: 1-Buffer receive mode enabled\n",
dev->name);
break;
case RXD_MODE_3B:
DBG_PRINT(ERR_DBG, "%s: 2-Buffer receive mode enabled\n",
dev->name);
break;
case RXD_MODE_3A:
DBG_PRINT(ERR_DBG, "%s: 3-Buffer receive mode enabled\n",
dev->name);
break;
}
if (napi)
DBG_PRINT(ERR_DBG, "%s: NAPI enabled\n", dev->name);
switch(sp->intr_type) {
case INTA:
DBG_PRINT(ERR_DBG, "%s: Interrupt type INTA\n", dev->name);
break;
case MSI:
DBG_PRINT(ERR_DBG, "%s: Interrupt type MSI\n", dev->name);
break;
case MSI_X:
DBG_PRINT(ERR_DBG, "%s: Interrupt type MSI-X\n", dev->name);
break;
}
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (sp->lro)
DBG_PRINT(ERR_DBG, "%s: Large receive offload enabled\n",
dev->name);
if (ufo)
DBG_PRINT(ERR_DBG, "%s: UDP Fragmentation Offload(UFO)"
" enabled\n", dev->name);
/* Initialize device name */
sprintf(sp->name, "%s Neterion %s", dev->name, sp->product_name);
/* Initialize bimodal Interrupts */
sp->config.bimodal = bimodal;
if (!(sp->device_type & XFRAME_II_DEVICE) && bimodal) {
sp->config.bimodal = 0;
DBG_PRINT(ERR_DBG,"%s:Bimodal intr not supported by Xframe I\n",
dev->name);
}
/*
* Make Link state as off at this point, when the Link change
* interrupt comes the state will be automatically changed to
* the right state.
*/
netif_carrier_off(dev);
return 0;
register_failed:
set_swap_failed:
iounmap(sp->bar1);
bar1_remap_failed:
iounmap(sp->bar0);
bar0_remap_failed:
mem_alloc_failed:
free_shared_mem(sp);
pci_disable_device(pdev);
if (dev_intr_type != MSI_X)
pci_release_regions(pdev);
else {
release_mem_region(pci_resource_start(pdev, 0),
pci_resource_len(pdev, 0));
release_mem_region(pci_resource_start(pdev, 2),
pci_resource_len(pdev, 2));
}
pci_set_drvdata(pdev, NULL);
free_netdev(dev);
return ret;
}
/**
* s2io_rem_nic - Free the PCI device
* @pdev: structure containing the PCI related information of the device.
* Description: This function is called by the Pci subsystem to release a
* PCI device and free up all resource held up by the device. This could
* be in response to a Hot plug event or when the driver is to be removed
* from memory.
*/
static void __devexit s2io_rem_nic(struct pci_dev *pdev)
{
struct net_device *dev =
(struct net_device *) pci_get_drvdata(pdev);
struct s2io_nic *sp;
if (dev == NULL) {
DBG_PRINT(ERR_DBG, "Driver Data is NULL!!\n");
return;
}
flush_scheduled_work();
sp = dev->priv;
unregister_netdev(dev);
free_shared_mem(sp);
iounmap(sp->bar0);
iounmap(sp->bar1);
if (sp->intr_type != MSI_X)
pci_release_regions(pdev);
else {
release_mem_region(pci_resource_start(pdev, 0),
pci_resource_len(pdev, 0));
release_mem_region(pci_resource_start(pdev, 2),
pci_resource_len(pdev, 2));
}
pci_set_drvdata(pdev, NULL);
free_netdev(dev);
pci_disable_device(pdev);
}
/**
* s2io_starter - Entry point for the driver
* Description: This function is the entry point for the driver. It verifies
* the module loadable parameters and initializes PCI configuration space.
*/
int __init s2io_starter(void)
{
return pci_register_driver(&s2io_driver);
}
/**
* s2io_closer - Cleanup routine for the driver
* Description: This function is the cleanup routine for the driver. It unregist * ers the driver.
*/
static __exit void s2io_closer(void)
{
pci_unregister_driver(&s2io_driver);
DBG_PRINT(INIT_DBG, "cleanup done\n");
}
module_init(s2io_starter);
module_exit(s2io_closer);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
static int check_L2_lro_capable(u8 *buffer, struct iphdr **ip,
struct tcphdr **tcp, struct RxD_t *rxdp)
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
{
int ip_off;
u8 l2_type = (u8)((rxdp->Control_1 >> 37) & 0x7), ip_len;
if (!(rxdp->Control_1 & RXD_FRAME_PROTO_TCP)) {
DBG_PRINT(INIT_DBG,"%s: Non-TCP frames not supported for LRO\n",
__FUNCTION__);
return -1;
}
/* TODO:
* By default the VLAN field in the MAC is stripped by the card, if this
* feature is turned off in rx_pa_cfg register, then the ip_off field
* has to be shifted by a further 2 bytes
*/
switch (l2_type) {
case 0: /* DIX type */
case 4: /* DIX type with VLAN */
ip_off = HEADER_ETHERNET_II_802_3_SIZE;
break;
/* LLC, SNAP etc are considered non-mergeable */
default:
return -1;
}
*ip = (struct iphdr *)((u8 *)buffer + ip_off);
ip_len = (u8)((*ip)->ihl);
ip_len <<= 2;
*tcp = (struct tcphdr *)((unsigned long)*ip + ip_len);
return 0;
}
static int check_for_socket_match(struct lro *lro, struct iphdr *ip,
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
struct tcphdr *tcp)
{
DBG_PRINT(INFO_DBG,"%s: Been here...\n", __FUNCTION__);
if ((lro->iph->saddr != ip->saddr) || (lro->iph->daddr != ip->daddr) ||
(lro->tcph->source != tcp->source) || (lro->tcph->dest != tcp->dest))
return -1;
return 0;
}
static inline int get_l4_pyld_length(struct iphdr *ip, struct tcphdr *tcp)
{
return(ntohs(ip->tot_len) - (ip->ihl << 2) - (tcp->doff << 2));
}
static void initiate_new_session(struct lro *lro, u8 *l2h,
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
struct iphdr *ip, struct tcphdr *tcp, u32 tcp_pyld_len)
{
DBG_PRINT(INFO_DBG,"%s: Been here...\n", __FUNCTION__);
lro->l2h = l2h;
lro->iph = ip;
lro->tcph = tcp;
lro->tcp_next_seq = tcp_pyld_len + ntohl(tcp->seq);
lro->tcp_ack = ntohl(tcp->ack_seq);
lro->sg_num = 1;
lro->total_len = ntohs(ip->tot_len);
lro->frags_len = 0;
/*
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
* check if we saw TCP timestamp. Other consistency checks have
* already been done.
*/
if (tcp->doff == 8) {
u32 *ptr;
ptr = (u32 *)(tcp+1);
lro->saw_ts = 1;
lro->cur_tsval = *(ptr+1);
lro->cur_tsecr = *(ptr+2);
}
lro->in_use = 1;
}
static void update_L3L4_header(struct s2io_nic *sp, struct lro *lro)
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
{
struct iphdr *ip = lro->iph;
struct tcphdr *tcp = lro->tcph;
__sum16 nchk;
struct stat_block *statinfo = sp->mac_control.stats_info;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
DBG_PRINT(INFO_DBG,"%s: Been here...\n", __FUNCTION__);
/* Update L3 header */
ip->tot_len = htons(lro->total_len);
ip->check = 0;
nchk = ip_fast_csum((u8 *)lro->iph, ip->ihl);
ip->check = nchk;
/* Update L4 header */
tcp->ack_seq = lro->tcp_ack;
tcp->window = lro->window;
/* Update tsecr field if this session has timestamps enabled */
if (lro->saw_ts) {
u32 *ptr = (u32 *)(tcp + 1);
*(ptr+2) = lro->cur_tsecr;
}
/* Update counters required for calculation of
* average no. of packets aggregated.
*/
statinfo->sw_stat.sum_avg_pkts_aggregated += lro->sg_num;
statinfo->sw_stat.num_aggregations++;
}
static void aggregate_new_rx(struct lro *lro, struct iphdr *ip,
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
struct tcphdr *tcp, u32 l4_pyld)
{
DBG_PRINT(INFO_DBG,"%s: Been here...\n", __FUNCTION__);
lro->total_len += l4_pyld;
lro->frags_len += l4_pyld;
lro->tcp_next_seq += l4_pyld;
lro->sg_num++;
/* Update ack seq no. and window ad(from this pkt) in LRO object */
lro->tcp_ack = tcp->ack_seq;
lro->window = tcp->window;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (lro->saw_ts) {
u32 *ptr;
/* Update tsecr and tsval from this packet */
ptr = (u32 *) (tcp + 1);
lro->cur_tsval = *(ptr + 1);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
lro->cur_tsecr = *(ptr + 2);
}
}
static int verify_l3_l4_lro_capable(struct lro *l_lro, struct iphdr *ip,
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
struct tcphdr *tcp, u32 tcp_pyld_len)
{
u8 *ptr;
DBG_PRINT(INFO_DBG,"%s: Been here...\n", __FUNCTION__);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (!tcp_pyld_len) {
/* Runt frame or a pure ack */
return -1;
}
if (ip->ihl != 5) /* IP has options */
return -1;
/* If we see CE codepoint in IP header, packet is not mergeable */
if (INET_ECN_is_ce(ipv4_get_dsfield(ip)))
return -1;
/* If we see ECE or CWR flags in TCP header, packet is not mergeable */
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (tcp->urg || tcp->psh || tcp->rst || tcp->syn || tcp->fin ||
tcp->ece || tcp->cwr || !tcp->ack) {
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
/*
* Currently recognize only the ack control word and
* any other control field being set would result in
* flushing the LRO session
*/
return -1;
}
/*
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
* Allow only one TCP timestamp option. Don't aggregate if
* any other options are detected.
*/
if (tcp->doff != 5 && tcp->doff != 8)
return -1;
if (tcp->doff == 8) {
ptr = (u8 *)(tcp + 1);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
while (*ptr == TCPOPT_NOP)
ptr++;
if (*ptr != TCPOPT_TIMESTAMP || *(ptr+1) != TCPOLEN_TIMESTAMP)
return -1;
/* Ensure timestamp value increases monotonically */
if (l_lro)
if (l_lro->cur_tsval > *((u32 *)(ptr+2)))
return -1;
/* timestamp echo reply should be non-zero */
if (*((u32 *)(ptr+6)) == 0)
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
return -1;
}
return 0;
}
static int
s2io_club_tcp_session(u8 *buffer, u8 **tcp, u32 *tcp_len, struct lro **lro,
struct RxD_t *rxdp, struct s2io_nic *sp)
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
{
struct iphdr *ip;
struct tcphdr *tcph;
int ret = 0, i;
if (!(ret = check_L2_lro_capable(buffer, &ip, (struct tcphdr **)tcp,
rxdp))) {
DBG_PRINT(INFO_DBG,"IP Saddr: %x Daddr: %x\n",
ip->saddr, ip->daddr);
} else {
return ret;
}
tcph = (struct tcphdr *)*tcp;
*tcp_len = get_l4_pyld_length(ip, tcph);
for (i=0; i<MAX_LRO_SESSIONS; i++) {
struct lro *l_lro = &sp->lro0_n[i];
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (l_lro->in_use) {
if (check_for_socket_match(l_lro, ip, tcph))
continue;
/* Sock pair matched */
*lro = l_lro;
if ((*lro)->tcp_next_seq != ntohl(tcph->seq)) {
DBG_PRINT(INFO_DBG, "%s:Out of order. expected "
"0x%x, actual 0x%x\n", __FUNCTION__,
(*lro)->tcp_next_seq,
ntohl(tcph->seq));
sp->mac_control.stats_info->
sw_stat.outof_sequence_pkts++;
ret = 2;
break;
}
if (!verify_l3_l4_lro_capable(l_lro, ip, tcph,*tcp_len))
ret = 1; /* Aggregate */
else
ret = 2; /* Flush both */
break;
}
}
if (ret == 0) {
/* Before searching for available LRO objects,
* check if the pkt is L3/L4 aggregatable. If not
* don't create new LRO session. Just send this
* packet up.
*/
if (verify_l3_l4_lro_capable(NULL, ip, tcph, *tcp_len)) {
return 5;
}
for (i=0; i<MAX_LRO_SESSIONS; i++) {
struct lro *l_lro = &sp->lro0_n[i];
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
if (!(l_lro->in_use)) {
*lro = l_lro;
ret = 3; /* Begin anew */
break;
}
}
}
if (ret == 0) { /* sessions exceeded */
DBG_PRINT(INFO_DBG,"%s:All LRO sessions already in use\n",
__FUNCTION__);
*lro = NULL;
return ret;
}
switch (ret) {
case 3:
initiate_new_session(*lro, buffer, ip, tcph, *tcp_len);
break;
case 2:
update_L3L4_header(sp, *lro);
break;
case 1:
aggregate_new_rx(*lro, ip, tcph, *tcp_len);
if ((*lro)->sg_num == sp->lro_max_aggr_per_sess) {
update_L3L4_header(sp, *lro);
ret = 4; /* Flush the LRO */
}
break;
default:
DBG_PRINT(ERR_DBG,"%s:Dont know, can't say!!\n",
__FUNCTION__);
break;
}
return ret;
}
static void clear_lro_session(struct lro *lro)
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
{
static u16 lro_struct_size = sizeof(struct lro);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
memset(lro, 0, lro_struct_size);
}
static void queue_rx_frame(struct sk_buff *skb)
{
struct net_device *dev = skb->dev;
skb->protocol = eth_type_trans(skb, dev);
if (napi)
netif_receive_skb(skb);
else
netif_rx(skb);
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
}
static void lro_append_pkt(struct s2io_nic *sp, struct lro *lro,
struct sk_buff *skb,
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
u32 tcp_len)
{
struct sk_buff *first = lro->parent;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
first->len += tcp_len;
first->data_len = lro->frags_len;
skb_pull(skb, (skb->len - tcp_len));
if (skb_shinfo(first)->frag_list)
lro->last_frag->next = skb;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
else
skb_shinfo(first)->frag_list = skb;
first->truesize += skb->truesize;
lro->last_frag = skb;
[PATCH] S2io: Large Receive Offload (LRO) feature(v2) for Neterion (s2io) 10GbE Xframe PCI-X and PCI-E NICs Hi, Below is a patch for the Large Receive Offload feature. Please review and let us know your comments. LRO algorithm was described in an OLS 2005 presentation, located at ftp.s2io.com user: linuxdocs password: HALdocs The same ftp site has Programming Manual for Xframe-I ASIC. LRO feature is supported on Neterion Xframe-I, Xframe-II and Xframe-Express 10GbE NICs. Brief description: The Large Receive Offload(LRO) feature is a stateless offload that is complementary to TSO feature but on the receive path. The idea is to combine and collapse(upto 64K maximum) in the driver, in-sequence TCP packets belonging to the same session. It is mainly designed to improve 1500 mtu receive performance, since Jumbo frame performance is already close to 10GbE line rate. Some performance numbers are attached below. Implementation details: 1. Handle packet chains from multiple sessions(current default MAX_LRO_SESSSIONS=32). 2. Examine each packet for eligiblity to aggregate. A packet is considered eligible if it meets all the below criteria. a. It is a TCP/IP packet and L2 type is not LLC or SNAP. b. The packet has no checksum errors(L3 and L4). c. There are no IP options. The only TCP option supported is timestamps. d. Search and locate the LRO object corresponding to this socket and ensure packet is in TCP sequence. e. It's not a special packet(SYN, FIN, RST, URG, PSH etc. flags are not set). f. TCP payload is non-zero(It's not a pure ACK). g. It's not an IP-fragmented packet. 3. If a packet is found eligible, the LRO object is updated with information such as next sequence number expected, current length of aggregated packet and so on. If not eligible or max packets reached, update IP and TCP headers of first packet in the chain and pass it up to stack. 4. The frag_list in skb structure is used to chain packets into one large packet. Kernel changes required: None Performance results: Main focus of the initial testing was on 1500 mtu receiver, since this is a bottleneck not covered by the existing stateless offloads. There are couple disclaimers about the performance results below: 1. Your mileage will vary!!!! We initially concentrated on couple pci-x 2.0 platforms that are powerful enough to push 10 GbE NIC and do not have bottlenecks other than cpu%; testing on other platforms is still in progress. On some lower end systems we are seeing lower gains. 2. Current LRO implementation is still (for the most part) software based, and therefore performance potential of the feature is far from being realized. Full hw implementation of LRO is expected in the next version of Xframe ASIC. Performance delta(with MTU=1500) going from LRO disabled to enabled: IBM 2-way Xeon (x366) : 3.5 to 7.1 Gbps 2-way Opteron : 4.5 to 6.1 Gbps Signed-off-by: Ravinandan Arakali <ravinandan.arakali@neterion.com> Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
2006-01-25 19:53:07 +00:00
sp->mac_control.stats_info->sw_stat.clubbed_frms_cnt++;
return;
}