Re: [PATCH net 4/5] net/smc: protect connection state transitions in listen work
From: Wenjia Zhang <wenjia@linux.ibm.com>
Date: 2023-10-12 17:14:48
Also in:
linux-rdma, linux-s390
On 11.10.23 09:33, D. Wythe wrote:
From: "D. Wythe" <alibuda@linux.alibaba.com> Consider the following scenario: smc_close_passive_work smc_listen_out_connected lock_sock() if (state == SMC_INIT) if (state == SMC_INIT) state = SMC_APPCLOSEWAIT1; state = SMC_ACTIVE release_sock() This would cause the state machine of the connection to be corrupted. Also, this issue can occur in smc_listen_out_err(). To solve this problem, we can protect the state transitions under the lock of sock to avoid collision.
To this fix, I have to repeat the question from Alexandra. Did the scenario occur in real life? Or is it just kind of potencial problem you found during the code review? If it is the former one, could you please show us the corresponding message, e.g. from dmesg? If it is the latter one, I'd like to deal with it more carefully. Going from this scenario, I noticed that there could also be other similar places where we need to make sure that no race happens. Thus, it would make more sense to find a systematic approach.
quoted hunk ↗ jump to hunk
Fixes: 3b2dec2603d5 ("net/smc: restructure client and server code in af_smc") Signed-off-by: D. Wythe <alibuda@linux.alibaba.com> --- net/smc/af_smc.c | 5 +++++ 1 file changed, 5 insertions(+)diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c index 5ad2a9f..3bb8265 100644 --- a/net/smc/af_smc.c +++ b/net/smc/af_smc.c@@ -1926,8 +1926,10 @@ static void smc_listen_out_connected(struct smc_sock *new_smc) { struct sock *newsmcsk = &new_smc->sk; + lock_sock(newsmcsk); if (newsmcsk->sk_state == SMC_INIT) newsmcsk->sk_state = SMC_ACTIVE; + release_sock(newsmcsk); smc_listen_out(new_smc); }@@ -1939,9 +1941,12 @@ static void smc_listen_out_err(struct smc_sock *new_smc) struct net *net = sock_net(newsmcsk); this_cpu_inc(net->smc.smc_stats->srv_hshake_err_cnt); + + lock_sock(newsmcsk); if (newsmcsk->sk_state == SMC_INIT) sock_put(&new_smc->sk); /* passive closing */ newsmcsk->sk_state = SMC_CLOSED; + release_sock(newsmcsk); smc_listen_out(new_smc); }