Re: [PATCH] af_packet: fix efficiency issues in packet_read_pending

[Eric Dumazet <edumazet@google.com>]

On Wed, Apr 6, 2022 at 4:49 AM lianglixue <lixue.liang5086@gmail.com> wrote:
>
> In packet_read_pengding, even if the pending_refcnt of the first CPU
> is not 0, the pending_refcnt of all CPUs will be traversed,
> and the long delay of cross-cpu access in NUMA significantly reduces
> the performance of packet sending; especially in tpacket_destruct_skb.
>
> When pending_refcnt is not 0, it returns without counting the number of
> all pending packets, which significantly reduces the traversal time.
>

Can you describe the use case ?

You are changing the slow path.

Perhaps you do not use the interface in the most efficient way.

Your patch is wrong, pending_refcnt increments and decrements can
happen on different cpus.



> Signed-off-by: lianglixue <lianglixue@greatwall.com.cn>
> ---
>  net/packet/af_packet.c | 9 +++++----
>  1 file changed, 5 insertions(+), 4 deletions(-)
>
> diff --git a/net/packet/af_packet.c b/net/packet/af_packet.c
> index c39c09899fd0..c04f49e44a33 100644
> --- a/net/packet/af_packet.c
> +++ b/net/packet/af_packet.c
> @@ -1210,17 +1210,18 @@ static void packet_dec_pending(struct packet_ring_buffer *rb)
>
>  static unsigned int packet_read_pending(const struct packet_ring_buffer *rb)
>  {
> -       unsigned int refcnt = 0;
>         int cpu;
>
>         /* We don't use pending refcount in rx_ring. */
>         if (rb->pending_refcnt == NULL)
>                 return 0;
>
> -       for_each_possible_cpu(cpu)
> -               refcnt += *per_cpu_ptr(rb->pending_refcnt, cpu);
> +       for_each_possible_cpu(cpu) {
> +               if (*per_cpu_ptr(rb->pending_refcnt, cpu) != 0)
> +                       return 1;
> +       }
>
> -       return refcnt;
> +       return 0;
>  }
>
>  static int packet_alloc_pending(struct packet_sock *po)
> --
> 2.27.0
>