Re: [PATCH net] net/smc: fix potential sleeping issue in smc_switch_conns

[Guangguan Wang <guangguan.wang@linux.alibaba.com>]

On 2024/4/13 11:51, Zhengchao Shao wrote:
> Potential sleeping issue exists in the following processes:
> smc_switch_conns
>   spin_lock_bh(&conn->send_lock)
>   smc_switch_link_and_count
>     smcr_link_put
>       __smcr_link_clear
>         smc_lgr_put
>           __smc_lgr_free
>             smc_lgr_free_bufs
>               __smc_lgr_free_bufs
>                 smc_buf_free
>                   smcr_buf_free
>                     smcr_buf_unmap_link
>                       smc_ib_put_memory_region
>                         ib_dereg_mr
>                           ib_dereg_mr_user
>                             mr->device->ops.dereg_mr
> If scheduling exists when the IB driver implements .dereg_mr hook
> function, the bug "scheduling while atomic" will occur. For example,
> cxgb4 and efa driver. Use mutex lock instead of spin lock to fix it.
> 
> Fixes: 20c9398d3309 ("net/smc: Resolve the race between SMC-R link access and clear")
> Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
> ---
>  net/smc/af_smc.c   |  2 +-
>  net/smc/smc.h      |  2 +-
>  net/smc/smc_cdc.c  | 14 +++++++-------
>  net/smc/smc_core.c |  8 ++++----
>  net/smc/smc_tx.c   |  8 ++++----
>  5 files changed, 17 insertions(+), 17 deletions(-)
> 

Hi Zhengchao,

I doubt whether this bug really exists, as efa supports SRD QP while SMC-R relies on RC QP,
cxgb4 is a IWARP adaptor while SMC-R relies on ROCE adaptor.

Thanks,
Guangguan Wang