net/smc: Prevent smc_release() from long blocking

In nginx/wrk benchmark, there's a hung problem with high probability
on case likes that: (client will last several minutes to exit)

server: smc_run nginx

client: smc_run wrk -c 10000 -t 1 http://server

Client hangs with the following backtrace:

0 [ffffa7ce8Of3bbf8] __schedule at ffffffff9f9eOd5f
1 [ffffa7ce8Of3bc88] schedule at ffffffff9f9eløe6
2 [ffffa7ce8Of3bcaO] schedule_timeout at ffffffff9f9e3f3c
3 [ffffa7ce8Of3bd2O] wait_for_common at ffffffff9f9el9de
4 [ffffa7ce8Of3bd8O] __flush_work at ffffffff9fOfeOl3
5 [ffffa7ce8øf3bdfO] smc_release at ffffffffcO697d24 [smc]
6 [ffffa7ce8Of3be2O] __sock_release at ffffffff9f8O2e2d
7 [ffffa7ce8Of3be4ø] sock_close at ffffffff9f8ø2ebl
8 [ffffa7ce8øf3be48] __fput at ffffffff9f334f93
9 [ffffa7ce8Of3be78] task_work_run at ffffffff9flOlff5
10 [ffffa7ce8Of3beaO] do_exit at ffffffff9fOe5Ol2
11 [ffffa7ce8Of3bflO] do_group_exit at ffffffff9fOe592a
12 [ffffa7ce8Of3bf38] __x64_sys_exit_group at ffffffff9fOe5994
13 [ffffa7ce8Of3bf4O] do_syscall_64 at ffffffff9f9d4373
14 [ffffa7ce8Of3bfsO] entry_SYSCALL_64_after_hwframe at ffffffff9fa0007c

This issue dues to flush_work(), which is used to wait for
smc_connect_work() to finish in smc_release(). Once lots of
smc_connect_work() was pending or all executing work dangling,
smc_release() has to block until one worker comes to free, which
is equivalent to wait another smc_connnect_work() to finish.

In order to fix this, There are two changes:

1. For those idle smc_connect_work(), cancel it from the workqueue; for
   executing smc_connect_work(), waiting for it to finish. For that
   purpose, replace flush_work() with cancel_work_sync().

2. Since smc_connect() hold a reference for passive closing, if
   smc_connect_work() has been cancelled, release the reference.

Fixes: 67f98918e75b ("net/smc: rebuild nonblocking connect")
Reported-by: Tony Lu <tonylu@linux.alibaba.com>
Tested-by: Dust Li <dust.li@linux.alibaba.com>
Reviewed-by: Dust Li <dust.li@linux.alibaba.com>
Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: D. Wythe <alibuda@linux.alibaba.com>
Acked-by: Karsten Graul <kgraul@linux.ibm.com>
Link: https://lore.kernel.org/r/1639571361-101128-1-git-send-email-alibuda@linux.alibaba.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index 230072f9ec48e3bf3cd43f9aa7fafa4a096ac402..1c9289f56dc471d33b08f6138bee3bea8aabe26f 100644 (file)
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c
@@ -194,7 +194,9 @@ static int smc_release(struct socket *sock)
        /* cleanup for a dangling non-blocking connect */
        if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
                tcp_abort(smc->clcsock->sk, ECONNABORTED);
-       flush_work(&smc->connect_work);
+
+       if (cancel_work_sync(&smc->connect_work))
+               sock_put(&smc->sk); /* sock_hold in smc_connect for passive closing */
 
        if (sk->sk_state == SMC_LISTEN)
                /* smc_close_non_accepted() is called and acquires