Re: [PATCH net-next v3] net/smc: avoid atomic_set and smp_wmb in the tx path when possible

[Alexandra Winter <wintera@linux.ibm.com>]

On 17.11.23 12:16, Li RongQing wrote:
> There is rare possibility that conn->tx_pushing is not 1, since
> tx_pushing is just checked with 1, so move the setting tx_pushing
> to 1 after atomic_dec_and_test() return false, to avoid atomic_set
> and smp_wmb in tx path
> 
> Reviewed-by: Dust Li <dust.li@linux.alibaba.com>
> Signed-off-by: Li RongQing <lirongqing@baidu.com>
> ---
> diff v3: improvements in the commit body and comments
> diff v2: fix a typo in commit body and add net-next subject-prefix
>  net/smc/smc_tx.c | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
> 
> diff --git a/net/smc/smc_tx.c b/net/smc/smc_tx.c
> index 3b0ff3b..2c2933f 100644
> --- a/net/smc/smc_tx.c
> +++ b/net/smc/smc_tx.c
> @@ -667,8 +667,6 @@ int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
>  		return 0;
>  
>  again:
> -	atomic_set(&conn->tx_pushing, 1);
> -	smp_wmb(); /* Make sure tx_pushing is 1 before real send */
>  	rc = __smc_tx_sndbuf_nonempty(conn);
>  
>  	/* We need to check whether someone else have added some data into
> @@ -677,8 +675,11 @@ int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
>  	 * If so, we need to push again to prevent those data hang in the send
>  	 * queue.
>  	 */
> -	if (unlikely(!atomic_dec_and_test(&conn->tx_pushing)))
> +	if (unlikely(!atomic_dec_and_test(&conn->tx_pushing))) {
> +		atomic_set(&conn->tx_pushing, 1);
> +		smp_wmb(); /* Make sure tx_pushing is 1 before send again */
>  		goto again;
> +	}
>  
>  	return rc;
>  }

It seems to me that the purpose of conn->tx_pushing is
a) Serve as a mutex, so only one thread per conn will call __smc_tx_sndbuf_nonempty().
b) Repeat, in case some other thread has added data to sndbuf concurrently.

I agree that this patch does not change the behaviour of this function and removes an
atomic_set() in the likely path.

I wonder however: All callers of smc_tx_sndbuf_nonempty() must hold the socket lock.
So how can we ever run in a concurrency situation?
Is this handling of conn->tx_pushing necessary at all?