From 5c15b3123f65f8fbb1b445d9a7e8812e0e435df2 Mon Sep 17 00:00:00 2001 From: "D. Wythe" Date: Wed, 15 Dec 2021 20:29:21 +0800 Subject: net/smc: Prevent smc_release() from long blocking MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit 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: 24ac3a08e658 ("net/smc: rebuild nonblocking connect") Reported-by: Tony Lu Tested-by: Dust Li Reviewed-by: Dust Li Reviewed-by: Tony Lu Signed-off-by: D. Wythe Acked-by: Karsten Graul Link: https://lore.kernel.org/r/1639571361-101128-1-git-send-email-alibuda@linux.alibaba.com Signed-off-by: Jakub Kicinski --- net/smc/af_smc.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) (limited to 'net') diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c index 230072f9ec48..1c9289f56dc4 100644 --- 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 -- cgit