linux/net/smc/smc_tx.c

735 lines
21 KiB
C
Raw Normal View History

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 22:07:57 +08:00
// SPDX-License-Identifier: GPL-2.0
/*
* Shared Memory Communications over RDMA (SMC-R) and RoCE
*
* Manage send buffer.
* Producer:
* Copy user space data into send buffer, if send buffer space available.
* Consumer:
* Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available.
*
* Copyright IBM Corp. 2016
*
* Author(s): Ursula Braun <ubraun@linux.vnet.ibm.com>
*/
#include <linux/net.h>
#include <linux/rcupdate.h>
#include <linux/workqueue.h>
sched/headers: Move task_struct::signal and task_struct::sighand types and accessors into <linux/sched/signal.h> task_struct::signal and task_struct::sighand are pointers, which would normally make it straightforward to not define those types in sched.h. That is not so, because the types are accompanied by a myriad of APIs (macros and inline functions) that dereference them. Split the types and the APIs out of sched.h and move them into a new header, <linux/sched/signal.h>. With this change sched.h does not know about 'struct signal' and 'struct sighand' anymore, trying to put accessors into sched.h as a test fails the following way: ./include/linux/sched.h: In function ‘test_signal_types’: ./include/linux/sched.h:2461:18: error: dereferencing pointer to incomplete type ‘struct signal_struct’ ^ This reduces the size and complexity of sched.h significantly. Update all headers and .c code that relied on getting the signal handling functionality from <linux/sched.h> to include <linux/sched/signal.h>. The list of affected files in the preparatory patch was partly generated by grepping for the APIs, and partly by doing coverage build testing, both all[yes|mod|def|no]config builds on 64-bit and 32-bit x86, and an array of cross-architecture builds. Nevertheless some (trivial) build breakage is still expected related to rare Kconfig combinations and in-flight patches to various kernel code, but most of it should be handled by this patch. Acked-by: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-kernel@vger.kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-02-02 15:35:14 +08:00
#include <linux/sched/signal.h>
#include <net/sock.h>
#include <net/tcp.h>
#include "smc.h"
#include "smc_wr.h"
#include "smc_cdc.h"
#include "smc_close.h"
#include "smc_ism.h"
#include "smc_tx.h"
#include "smc_stats.h"
#include "smc_tracepoint.h"
#define SMC_TX_WORK_DELAY 0
/***************************** sndbuf producer *******************************/
/* callback implementation for sk.sk_write_space()
* to wakeup sndbuf producers that blocked with smc_tx_wait().
* called under sk_socket lock.
*/
static void smc_tx_write_space(struct sock *sk)
{
struct socket *sock = sk->sk_socket;
struct smc_sock *smc = smc_sk(sk);
struct socket_wq *wq;
/* similar to sk_stream_write_space */
if (atomic_read(&smc->conn.sndbuf_space) && sock) {
if (test_bit(SOCK_NOSPACE, &sock->flags))
SMC_STAT_RMB_TX_FULL(smc, !smc->conn.lnk);
clear_bit(SOCK_NOSPACE, &sock->flags);
rcu_read_lock();
wq = rcu_dereference(sk->sk_wq);
if (skwq_has_sleeper(wq))
wake_up_interruptible_poll(&wq->wait,
EPOLLOUT | EPOLLWRNORM |
EPOLLWRBAND);
if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))
sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT);
rcu_read_unlock();
}
}
/* Wakeup sndbuf producers that blocked with smc_tx_wait().
* Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
*/
void smc_tx_sndbuf_nonfull(struct smc_sock *smc)
{
if (smc->sk.sk_socket &&
test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))
smc->sk.sk_write_space(&smc->sk);
}
/* blocks sndbuf producer until at least one byte of free space available
* or urgent Byte was consumed
*/
static int smc_tx_wait(struct smc_sock *smc, int flags)
{
DEFINE_WAIT_FUNC(wait, woken_wake_function);
struct smc_connection *conn = &smc->conn;
struct sock *sk = &smc->sk;
long timeo;
int rc = 0;
/* similar to sk_stream_wait_memory */
timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
add_wait_queue(sk_sleep(sk), &wait);
while (1) {
sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
if (sk->sk_err ||
(sk->sk_shutdown & SEND_SHUTDOWN) ||
conn->killed ||
conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {
rc = -EPIPE;
break;
}
if (smc_cdc_rxed_any_close(conn)) {
rc = -ECONNRESET;
break;
}
if (!timeo) {
/* ensure EPOLLOUT is subsequently generated */
set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
rc = -EAGAIN;
break;
}
if (signal_pending(current)) {
rc = sock_intr_errno(timeo);
break;
}
sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
if (atomic_read(&conn->sndbuf_space) && !conn->urg_tx_pend)
break; /* at least 1 byte of free & no urgent data */
set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
sk_wait_event(sk, &timeo,
net: deal with most data-races in sk_wait_event() __condition is evaluated twice in sk_wait_event() macro. First invocation is lockless, and reads can race with writes, as spotted by syzbot. BUG: KCSAN: data-race in sk_stream_wait_connect / tcp_disconnect write to 0xffff88812d83d6a0 of 4 bytes by task 9065 on cpu 1: tcp_disconnect+0x2cd/0xdb0 inet_shutdown+0x19e/0x1f0 net/ipv4/af_inet.c:911 __sys_shutdown_sock net/socket.c:2343 [inline] __sys_shutdown net/socket.c:2355 [inline] __do_sys_shutdown net/socket.c:2363 [inline] __se_sys_shutdown+0xf8/0x140 net/socket.c:2361 __x64_sys_shutdown+0x31/0x40 net/socket.c:2361 do_syscall_x64 arch/x86/entry/common.c:50 [inline] do_syscall_64+0x41/0xc0 arch/x86/entry/common.c:80 entry_SYSCALL_64_after_hwframe+0x63/0xcd read to 0xffff88812d83d6a0 of 4 bytes by task 9040 on cpu 0: sk_stream_wait_connect+0x1de/0x3a0 net/core/stream.c:75 tcp_sendmsg_locked+0x2e4/0x2120 net/ipv4/tcp.c:1266 tcp_sendmsg+0x30/0x50 net/ipv4/tcp.c:1484 inet6_sendmsg+0x63/0x80 net/ipv6/af_inet6.c:651 sock_sendmsg_nosec net/socket.c:724 [inline] sock_sendmsg net/socket.c:747 [inline] __sys_sendto+0x246/0x300 net/socket.c:2142 __do_sys_sendto net/socket.c:2154 [inline] __se_sys_sendto net/socket.c:2150 [inline] __x64_sys_sendto+0x78/0x90 net/socket.c:2150 do_syscall_x64 arch/x86/entry/common.c:50 [inline] do_syscall_64+0x41/0xc0 arch/x86/entry/common.c:80 entry_SYSCALL_64_after_hwframe+0x63/0xcd value changed: 0x00000000 -> 0x00000068 Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Reported-by: syzbot <syzkaller@googlegroups.com> Signed-off-by: Eric Dumazet <edumazet@google.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-05-10 02:29:48 +08:00
READ_ONCE(sk->sk_err) ||
(READ_ONCE(sk->sk_shutdown) & SEND_SHUTDOWN) ||
smc_cdc_rxed_any_close(conn) ||
(atomic_read(&conn->sndbuf_space) &&
!conn->urg_tx_pend),
&wait);
}
remove_wait_queue(sk_sleep(sk), &wait);
return rc;
}
static bool smc_tx_is_corked(struct smc_sock *smc)
{
struct tcp_sock *tp = tcp_sk(smc->clcsock->sk);
return (tp->nonagle & TCP_NAGLE_CORK) ? true : false;
}
net/smc: add autocorking support This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-03-01 17:43:57 +08:00
/* If we have pending CDC messages, do not send:
* Because CQE of this CDC message will happen shortly, it gives
* a chance to coalesce future sendmsg() payload in to one RDMA Write,
* without need for a timer, and with no latency trade off.
* Algorithm here:
* 1. First message should never cork
* 2. If we have pending Tx CDC messages, wait for the first CDC
* message's completion
* 3. Don't cork to much data in a single RDMA Write to prevent burst
* traffic, total corked message should not exceed sendbuf/2
*/
static bool smc_should_autocork(struct smc_sock *smc)
{
struct smc_connection *conn = &smc->conn;
int corking_size;
corking_size = min_t(unsigned int, conn->sndbuf_desc->len >> 1,
sock_net(&smc->sk)->smc.sysctl_autocorking_size);
net/smc: add autocorking support This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-03-01 17:43:57 +08:00
if (atomic_read(&conn->cdc_pend_tx_wr) == 0 ||
smc_tx_prepared_sends(conn) > corking_size)
return false;
return true;
}
static bool smc_tx_should_cork(struct smc_sock *smc, struct msghdr *msg)
{
struct smc_connection *conn = &smc->conn;
if (smc_should_autocork(smc))
return true;
/* for a corked socket defer the RDMA writes if
* sndbuf_space is still available. The applications
* should known how/when to uncork it.
*/
if ((msg->msg_flags & MSG_MORE ||
smc_tx_is_corked(smc)) &&
net/smc: add autocorking support This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-03-01 17:43:57 +08:00
atomic_read(&conn->sndbuf_space))
return true;
return false;
}
/* sndbuf producer: main API called by socket layer.
* called under sock lock.
*/
int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
{
size_t copylen, send_done = 0, send_remaining = len;
size_t chunk_len, chunk_off, chunk_len_sum;
struct smc_connection *conn = &smc->conn;
union smc_host_cursor prep;
struct sock *sk = &smc->sk;
char *sndbuf_base;
int tx_cnt_prep;
int writespace;
int rc, chunk;
/* This should be in poll */
sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
rc = -EPIPE;
goto out_err;
}
if (sk->sk_state == SMC_INIT)
return -ENOTCONN;
if (len > conn->sndbuf_desc->len)
SMC_STAT_RMB_TX_SIZE_SMALL(smc, !conn->lnk);
if (len > conn->peer_rmbe_size)
SMC_STAT_RMB_TX_PEER_SIZE_SMALL(smc, !conn->lnk);
if (msg->msg_flags & MSG_OOB)
SMC_STAT_INC(smc, urg_data_cnt);
while (msg_data_left(msg)) {
if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
(smc->sk.sk_err == ECONNABORTED) ||
conn->killed)
return -EPIPE;
if (smc_cdc_rxed_any_close(conn))
return send_done ?: -ECONNRESET;
if (msg->msg_flags & MSG_OOB)
conn->local_tx_ctrl.prod_flags.urg_data_pending = 1;
if (!atomic_read(&conn->sndbuf_space) || conn->urg_tx_pend) {
if (send_done)
return send_done;
rc = smc_tx_wait(smc, msg->msg_flags);
if (rc)
goto out_err;
continue;
}
/* initialize variables for 1st iteration of subsequent loop */
/* could be just 1 byte, even after smc_tx_wait above */
writespace = atomic_read(&conn->sndbuf_space);
/* not more than what user space asked for */
copylen = min_t(size_t, send_remaining, writespace);
/* determine start of sndbuf */
sndbuf_base = conn->sndbuf_desc->cpu_addr;
smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
tx_cnt_prep = prep.count;
/* determine chunks where to write into sndbuf */
/* either unwrapped case, or 1st chunk of wrapped case */
chunk_len = min_t(size_t, copylen, conn->sndbuf_desc->len -
tx_cnt_prep);
chunk_len_sum = chunk_len;
chunk_off = tx_cnt_prep;
for (chunk = 0; chunk < 2; chunk++) {
rc = memcpy_from_msg(sndbuf_base + chunk_off,
msg, chunk_len);
if (rc) {
smc_sndbuf_sync_sg_for_device(conn);
if (send_done)
return send_done;
goto out_err;
}
send_done += chunk_len;
send_remaining -= chunk_len;
if (chunk_len_sum == copylen)
break; /* either on 1st or 2nd iteration */
/* prepare next (== 2nd) iteration */
chunk_len = copylen - chunk_len; /* remainder */
chunk_len_sum += chunk_len;
chunk_off = 0; /* modulo offset in send ring buffer */
}
smc_sndbuf_sync_sg_for_device(conn);
/* update cursors */
smc_curs_add(conn->sndbuf_desc->len, &prep, copylen);
smc_curs_copy(&conn->tx_curs_prep, &prep, conn);
/* increased in send tasklet smc_cdc_tx_handler() */
smp_mb__before_atomic();
atomic_sub(copylen, &conn->sndbuf_space);
/* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
smp_mb__after_atomic();
/* since we just produced more new data into sndbuf,
* trigger sndbuf consumer: RDMA write into peer RMBE and CDC
*/
if ((msg->msg_flags & MSG_OOB) && !send_remaining)
conn->urg_tx_pend = true;
net/smc: add autocorking support This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-03-01 17:43:57 +08:00
/* If we need to cork, do nothing and wait for the next
* sendmsg() call or push on tx completion
*/
net/smc: add autocorking support This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-03-01 17:43:57 +08:00
if (!smc_tx_should_cork(smc, msg))
smc_tx_sndbuf_nonempty(conn);
trace_smc_tx_sendmsg(smc, copylen);
} /* while (msg_data_left(msg)) */
return send_done;
out_err:
rc = sk_stream_error(sk, msg->msg_flags, rc);
/* make sure we wake any epoll edge trigger waiter */
if (unlikely(rc == -EAGAIN))
sk->sk_write_space(sk);
return rc;
}
/***************************** sndbuf consumer *******************************/
/* sndbuf consumer: actual data transfer of one target chunk with ISM write */
int smcd_tx_ism_write(struct smc_connection *conn, void *data, size_t len,
u32 offset, int signal)
{
int rc;
rc = smc_ism_write(conn->lgr->smcd, conn->peer_token,
conn->peer_rmbe_idx, signal, conn->tx_off + offset,
data, len);
if (rc)
conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
return rc;
}
/* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
int num_sges, struct ib_rdma_wr *rdma_wr)
{
struct smc_link_group *lgr = conn->lgr;
struct smc_link *link = conn->lnk;
int rc;
rdma_wr->wr.wr_id = smc_wr_tx_get_next_wr_id(link);
rdma_wr->wr.num_sge = num_sges;
rdma_wr->remote_addr =
lgr->rtokens[conn->rtoken_idx][link->link_idx].dma_addr +
/* RMBE within RMB */
conn->tx_off +
/* offset within RMBE */
peer_rmbe_offset;
rdma_wr->rkey = lgr->rtokens[conn->rtoken_idx][link->link_idx].rkey;
rc = ib_post_send(link->roce_qp, &rdma_wr->wr, NULL);
if (rc)
smcr_link_down_cond_sched(link);
return rc;
}
/* sndbuf consumer */
static inline void smc_tx_advance_cursors(struct smc_connection *conn,
union smc_host_cursor *prod,
union smc_host_cursor *sent,
size_t len)
{
smc_curs_add(conn->peer_rmbe_size, prod, len);
/* increased in recv tasklet smc_cdc_msg_rcv() */
smp_mb__before_atomic();
/* data in flight reduces usable snd_wnd */
atomic_sub(len, &conn->peer_rmbe_space);
/* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
smp_mb__after_atomic();
smc_curs_add(conn->sndbuf_desc->len, sent, len);
}
/* SMC-R helper for smc_tx_rdma_writes() */
static int smcr_tx_rdma_writes(struct smc_connection *conn, size_t len,
size_t src_off, size_t src_len,
size_t dst_off, size_t dst_len,
struct smc_rdma_wr *wr_rdma_buf)
{
struct smc_link *link = conn->lnk;
dma_addr_t dma_addr =
sg_dma_address(conn->sndbuf_desc->sgt[link->link_idx].sgl);
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 17:44:04 +08:00
u64 virt_addr = (uintptr_t)conn->sndbuf_desc->cpu_addr;
int src_len_sum = src_len, dst_len_sum = dst_len;
int sent_count = src_off;
int srcchunk, dstchunk;
int num_sges;
int rc;
for (dstchunk = 0; dstchunk < 2; dstchunk++) {
struct ib_rdma_wr *wr = &wr_rdma_buf->wr_tx_rdma[dstchunk];
struct ib_sge *sge = wr->wr.sg_list;
u64 base_addr = dma_addr;
if (dst_len < link->qp_attr.cap.max_inline_data) {
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 17:44:04 +08:00
base_addr = virt_addr;
wr->wr.send_flags |= IB_SEND_INLINE;
} else {
wr->wr.send_flags &= ~IB_SEND_INLINE;
}
num_sges = 0;
for (srcchunk = 0; srcchunk < 2; srcchunk++) {
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 17:44:04 +08:00
sge[srcchunk].addr = conn->sndbuf_desc->is_vm ?
(virt_addr + src_off) : (base_addr + src_off);
sge[srcchunk].length = src_len;
net/smc: Allow virtually contiguous sndbufs or RMBs for SMC-R On long-running enterprise production servers, high-order contiguous memory pages are usually very rare and in most cases we can only get fragmented pages. When replacing TCP with SMC-R in such production scenarios, attempting to allocate high-order physically contiguous sndbufs and RMBs may result in frequent memory compaction, which will cause unexpected hung issue and further stability risks. So this patch is aimed to allow SMC-R link group to use virtually contiguous sndbufs and RMBs to avoid potential issues mentioned above. Whether to use physically or virtually contiguous buffers can be set by sysctl smcr_buf_type. Note that using virtually contiguous buffers will bring an acceptable performance regression, which can be mainly divided into two parts: 1) regression in data path, which is brought by additional address translation of sndbuf by RNIC in Tx. But in general, translating address through MTT is fast. Taking 256KB sndbuf and RMB as an example, the comparisons in qperf latency and bandwidth test with physically and virtually contiguous buffers are as follows: - client: smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\ -t 5 -vu tcp_{bw|lat} - server: smc_run taskset -c <cpu> qperf [latency] msgsize tcp smcr smcr-use-virt-buf 1 11.17 us 7.56 us 7.51 us (-0.67%) 2 10.65 us 7.74 us 7.56 us (-2.31%) 4 11.11 us 7.52 us 7.59 us ( 0.84%) 8 10.83 us 7.55 us 7.51 us (-0.48%) 16 11.21 us 7.46 us 7.51 us ( 0.71%) 32 10.65 us 7.53 us 7.58 us ( 0.61%) 64 10.95 us 7.74 us 7.80 us ( 0.76%) 128 11.14 us 7.83 us 7.87 us ( 0.47%) 256 10.97 us 7.94 us 7.92 us (-0.28%) 512 11.23 us 7.94 us 8.20 us ( 3.25%) 1024 11.60 us 8.12 us 8.20 us ( 0.96%) 2048 14.04 us 8.30 us 8.51 us ( 2.49%) 4096 16.88 us 9.13 us 9.07 us (-0.64%) 8192 22.50 us 10.56 us 11.22 us ( 6.26%) 16384 28.99 us 12.88 us 13.83 us ( 7.37%) 32768 40.13 us 16.76 us 16.95 us ( 1.16%) 65536 68.70 us 24.68 us 24.85 us ( 0.68%) [bandwidth] msgsize tcp smcr smcr-use-virt-buf 1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%) 2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%) 4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%) 8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%) 16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%) 32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%) 64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%) 128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%) 256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%) 512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%) 1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%) 2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%) 4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%) 8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%) 16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%) 32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%) 65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%) 2) regression in buffer initialization and destruction path, which is brought by additional MR operations of sndbufs. But thanks to link group buffer reuse mechanism, the impact of this kind of regression decreases as times of buffer reuse increases. Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R buffer-related function obtained by bpftrace are as follows: Function Phys-bufs Virt-bufs smcr_new_buf_create() 67154 ns 79164 ns smc_ib_buf_map_sg() 525 ns 928 ns smc_ib_get_memory_region() 162294 ns 161191 ns smc_wr_reg_send() 9957 ns 9635 ns smc_ib_put_memory_region() 203548 ns 198374 ns smc_ib_buf_unmap_sg() 508 ns 1158 ns ------------ Test environment notes: 1. Above tests run on 2 VMs within the same Host. 2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to the each VM respectively. 3. VMs' vCPUs are binded to different physical CPUs, and the binded physical CPUs are isolated by `isolcpus=xxx` cmdline. 4. NICs' queue number are set to 1. Signed-off-by: Wen Gu <guwen@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-07-14 17:44:04 +08:00
if (conn->sndbuf_desc->is_vm)
sge[srcchunk].lkey =
conn->sndbuf_desc->mr[link->link_idx]->lkey;
num_sges++;
src_off += src_len;
if (src_off >= conn->sndbuf_desc->len)
src_off -= conn->sndbuf_desc->len;
/* modulo in send ring */
if (src_len_sum == dst_len)
break; /* either on 1st or 2nd iteration */
/* prepare next (== 2nd) iteration */
src_len = dst_len - src_len; /* remainder */
src_len_sum += src_len;
}
rc = smc_tx_rdma_write(conn, dst_off, num_sges, wr);
if (rc)
return rc;
if (dst_len_sum == len)
break; /* either on 1st or 2nd iteration */
/* prepare next (== 2nd) iteration */
dst_off = 0; /* modulo offset in RMBE ring buffer */
dst_len = len - dst_len; /* remainder */
dst_len_sum += dst_len;
src_len = min_t(int, dst_len, conn->sndbuf_desc->len -
sent_count);
src_len_sum = src_len;
}
return 0;
}
/* SMC-D helper for smc_tx_rdma_writes() */
static int smcd_tx_rdma_writes(struct smc_connection *conn, size_t len,
size_t src_off, size_t src_len,
size_t dst_off, size_t dst_len)
{
int src_len_sum = src_len, dst_len_sum = dst_len;
int srcchunk, dstchunk;
int rc;
for (dstchunk = 0; dstchunk < 2; dstchunk++) {
for (srcchunk = 0; srcchunk < 2; srcchunk++) {
void *data = conn->sndbuf_desc->cpu_addr + src_off;
rc = smcd_tx_ism_write(conn, data, src_len, dst_off +
sizeof(struct smcd_cdc_msg), 0);
if (rc)
return rc;
dst_off += src_len;
src_off += src_len;
if (src_off >= conn->sndbuf_desc->len)
src_off -= conn->sndbuf_desc->len;
/* modulo in send ring */
if (src_len_sum == dst_len)
break; /* either on 1st or 2nd iteration */
/* prepare next (== 2nd) iteration */
src_len = dst_len - src_len; /* remainder */
src_len_sum += src_len;
}
if (dst_len_sum == len)
break; /* either on 1st or 2nd iteration */
/* prepare next (== 2nd) iteration */
dst_off = 0; /* modulo offset in RMBE ring buffer */
dst_len = len - dst_len; /* remainder */
dst_len_sum += dst_len;
src_len = min_t(int, dst_len, conn->sndbuf_desc->len - src_off);
src_len_sum = src_len;
}
return 0;
}
/* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
* usable snd_wnd as max transmit
*/
static int smc_tx_rdma_writes(struct smc_connection *conn,
struct smc_rdma_wr *wr_rdma_buf)
{
size_t len, src_len, dst_off, dst_len; /* current chunk values */
union smc_host_cursor sent, prep, prod, cons;
struct smc_cdc_producer_flags *pflags;
int to_send, rmbespace;
int rc;
/* source: sndbuf */
smc_curs_copy(&sent, &conn->tx_curs_sent, conn);
smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
/* cf. wmem_alloc - (snd_max - snd_una) */
to_send = smc_curs_diff(conn->sndbuf_desc->len, &sent, &prep);
if (to_send <= 0)
return 0;
/* destination: RMBE */
/* cf. snd_wnd */
rmbespace = atomic_read(&conn->peer_rmbe_space);
if (rmbespace <= 0) {
struct smc_sock *smc = container_of(conn, struct smc_sock,
conn);
SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
return 0;
}
smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
/* if usable snd_wnd closes ask peer to advertise once it opens again */
pflags = &conn->local_tx_ctrl.prod_flags;
pflags->write_blocked = (to_send >= rmbespace);
/* cf. usable snd_wnd */
len = min(to_send, rmbespace);
/* initialize variables for first iteration of subsequent nested loop */
dst_off = prod.count;
if (prod.wrap == cons.wrap) {
/* the filled destination area is unwrapped,
* hence the available free destination space is wrapped
* and we need 2 destination chunks of sum len; start with 1st
* which is limited by what's available in sndbuf
*/
dst_len = min_t(size_t,
conn->peer_rmbe_size - prod.count, len);
} else {
/* the filled destination area is wrapped,
* hence the available free destination space is unwrapped
* and we need a single destination chunk of entire len
*/
dst_len = len;
}
/* dst_len determines the maximum src_len */
if (sent.count + dst_len <= conn->sndbuf_desc->len) {
/* unwrapped src case: single chunk of entire dst_len */
src_len = dst_len;
} else {
/* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
src_len = conn->sndbuf_desc->len - sent.count;
}
if (conn->lgr->is_smcd)
rc = smcd_tx_rdma_writes(conn, len, sent.count, src_len,
dst_off, dst_len);
else
rc = smcr_tx_rdma_writes(conn, len, sent.count, src_len,
dst_off, dst_len, wr_rdma_buf);
if (rc)
return rc;
if (conn->urg_tx_pend && len == to_send)
pflags->urg_data_present = 1;
smc_tx_advance_cursors(conn, &prod, &sent, len);
/* update connection's cursors with advanced local cursors */
smc_curs_copy(&conn->local_tx_ctrl.prod, &prod, conn);
/* dst: peer RMBE */
smc_curs_copy(&conn->tx_curs_sent, &sent, conn);/* src: local sndbuf */
return 0;
}
/* Wakeup sndbuf consumers from any context (IRQ or process)
* since there is more data to transmit; usable snd_wnd as max transmit
*/
static int smcr_tx_sndbuf_nonempty(struct smc_connection *conn)
{
struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
struct smc_link *link = conn->lnk;
struct smc_rdma_wr *wr_rdma_buf;
struct smc_cdc_tx_pend *pend;
struct smc_wr_buf *wr_buf;
int rc;
if (!link || !smc_wr_tx_link_hold(link))
return -ENOLINK;
rc = smc_cdc_get_free_slot(conn, link, &wr_buf, &wr_rdma_buf, &pend);
if (rc < 0) {
smc_wr_tx_link_put(link);
if (rc == -EBUSY) {
struct smc_sock *smc =
container_of(conn, struct smc_sock, conn);
if (smc->sk.sk_err == ECONNABORTED)
return sock_error(&smc->sk);
if (conn->killed)
return -EPIPE;
rc = 0;
mod_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
SMC_TX_WORK_DELAY);
}
return rc;
}
spin_lock_bh(&conn->send_lock);
if (link != conn->lnk) {
/* link of connection changed, tx_work will restart */
smc_wr_tx_put_slot(link,
(struct smc_wr_tx_pend_priv *)pend);
rc = -ENOLINK;
goto out_unlock;
}
if (!pflags->urg_data_present) {
rc = smc_tx_rdma_writes(conn, wr_rdma_buf);
if (rc) {
smc_wr_tx_put_slot(link,
(struct smc_wr_tx_pend_priv *)pend);
goto out_unlock;
}
}
rc = smc_cdc_msg_send(conn, wr_buf, pend);
if (!rc && pflags->urg_data_present) {
pflags->urg_data_pending = 0;
pflags->urg_data_present = 0;
}
out_unlock:
spin_unlock_bh(&conn->send_lock);
smc_wr_tx_link_put(link);
return rc;
}
static int smcd_tx_sndbuf_nonempty(struct smc_connection *conn)
{
struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
int rc = 0;
spin_lock_bh(&conn->send_lock);
if (!pflags->urg_data_present)
rc = smc_tx_rdma_writes(conn, NULL);
if (!rc)
rc = smcd_cdc_msg_send(conn);
if (!rc && pflags->urg_data_present) {
pflags->urg_data_pending = 0;
pflags->urg_data_present = 0;
}
spin_unlock_bh(&conn->send_lock);
return rc;
}
int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
{
net/smc: add autocorking support This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-03-01 17:43:57 +08:00
struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
int rc = 0;
/* No data in the send queue */
if (unlikely(smc_tx_prepared_sends(conn) <= 0))
goto out;
/* Peer don't have RMBE space */
if (unlikely(atomic_read(&conn->peer_rmbe_space) <= 0)) {
SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
goto out;
}
if (conn->killed ||
net/smc: add autocorking support This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-03-01 17:43:57 +08:00
conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
rc = -EPIPE; /* connection being aborted */
goto out;
}
if (conn->lgr->is_smcd)
rc = smcd_tx_sndbuf_nonempty(conn);
else
rc = smcr_tx_sndbuf_nonempty(conn);
if (!rc) {
/* trigger socket release if connection is closing */
smc_close_wake_tx_prepared(smc);
}
net/smc: add autocorking support This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: Dust Li <dust.li@linux.alibaba.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2022-03-01 17:43:57 +08:00
out:
return rc;
}
/* Wakeup sndbuf consumers from process context
* since there is more data to transmit. The caller
* must hold sock lock.
*/
void smc_tx_pending(struct smc_connection *conn)
{
struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
int rc;
if (smc->sk.sk_err)
return;
rc = smc_tx_sndbuf_nonempty(conn);
if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
!atomic_read(&conn->bytes_to_rcv))
conn->local_rx_ctrl.prod_flags.write_blocked = 0;
}
/* Wakeup sndbuf consumers from process context
* since there is more data to transmit in locked
* sock.
*/
void smc_tx_work(struct work_struct *work)
{
struct smc_connection *conn = container_of(to_delayed_work(work),
struct smc_connection,
tx_work);
struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
lock_sock(&smc->sk);
smc_tx_pending(conn);
release_sock(&smc->sk);
}
void smc_tx_consumer_update(struct smc_connection *conn, bool force)
{
union smc_host_cursor cfed, cons, prod;
int sender_free = conn->rmb_desc->len;
int to_confirm;
smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
smc_curs_copy(&cfed, &conn->rx_curs_confirmed, conn);
to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);
if (to_confirm > conn->rmbe_update_limit) {
smc_curs_copy(&prod, &conn->local_rx_ctrl.prod, conn);
sender_free = conn->rmb_desc->len -
smc_curs_diff_large(conn->rmb_desc->len,
&cfed, &prod);
}
if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
force ||
((to_confirm > conn->rmbe_update_limit) &&
((sender_free <= (conn->rmb_desc->len / 2)) ||
conn->local_rx_ctrl.prod_flags.write_blocked))) {
if (conn->killed ||
conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
return;
if ((smc_cdc_get_slot_and_msg_send(conn) < 0) &&
!conn->killed) {
queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
SMC_TX_WORK_DELAY);
return;
}
}
if (conn->local_rx_ctrl.prod_flags.write_blocked &&
!atomic_read(&conn->bytes_to_rcv))
conn->local_rx_ctrl.prod_flags.write_blocked = 0;
}
/***************************** send initialize *******************************/
/* Initialize send properties on connection establishment. NB: not __init! */
void smc_tx_init(struct smc_sock *smc)
{
smc->sk.sk_write_space = smc_tx_write_space;
}