On Mon, 2023-03-06 at 11:36 +0800, D. Wythe wrote:
> From: "D. Wythe" <alibuda@xxxxxxxxxxxxxxxxx>
>
> When performing a stress test on SMC-R by rmmod mlx5_ib driver
> during the wrk/nginx test, we found that there is a probability
> of triggering a panic while terminating all link groups.
>
> This issue dues to the race between smc_smcr_terminate_all()
> and smc_buf_create().
>
> smc_smcr_terminate_all
>
> smc_buf_create
> /* init */
> conn->sndbuf_desc = NULL;
> ...
>
> __smc_lgr_terminate
> smc_conn_kill
> smc_close_abort
> smc_cdc_get_slot_and_msg_send
>
> __softirqentry_text_start
> smc_wr_tx_process_cqe
> smc_cdc_tx_handler
> READ(conn->sndbuf_desc->len);
> /* panic dues to NULL sndbuf_desc */
>
> conn->sndbuf_desc = xxx;
>
> This patch tries to fix the issue by always to check the sndbuf_desc
> before send any cdc msg, to make sure that no null pointer is
> seen during cqe processing.
>
> Fixes: 0b29ec643613 ("net/smc: immediate termination for SMCR link groups")
> Signed-off-by: D. Wythe <alibuda@xxxxxxxxxxxxxxxxx>
Looking at the code for __smc_buf_create it seems like you might have
more issues hiding in the code. From what I can tell smc_buf_get_slot
can only return a pointer or NULL but it is getting checked for being
being a PTR_ERR or IS_ERR in several spots that are likely all dead
code.
> ---
> net/smc/smc_cdc.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/net/smc/smc_cdc.c b/net/smc/smc_cdc.c
> index 53f63bf..2f0e2ee 100644
> --- a/net/smc/smc_cdc.c
> +++ b/net/smc/smc_cdc.c
> @@ -114,6 +114,9 @@ int smc_cdc_msg_send(struct smc_connection *conn,
> union smc_host_cursor cfed;
> int rc;
>
> + if (unlikely(!READ_ONCE(conn->sndbuf_desc)))
> + return -EINVAL;
> +
This return value doesn't seem right to me. Rather than en EINVAL
should this be something like a ENOBUFS just to make it easier to debug
when this issue is encountered?
> smc_cdc_add_pending_send(conn, pend);
>
> conn->tx_cdc_seq++;
