Re: [PATCH net-next] inet: do not set backlog if listen fails

[Eric Dumazet]

On 10/03/2018 08:07 AM, Yafang Shao wrote:
> These two backlog members are not necessary set in inet_csk_listen_start().
> Regarding sk_max_ack_backlog, it will be set in the caller inet_listen
> and dccp_listen_start.
> Regraing sk_ack_backlog, it is better to put it together with
> sk_max_ack_backlog in the same function and only set on success.
> 
> Signed-off-by: Yafang Shao <laoar.shao@xxxxxxxxx>
> ---
>  net/dccp/proto.c                | 1 +
>  net/ipv4/af_inet.c              | 1 +
>  net/ipv4/inet_connection_sock.c | 3 ---
>  3 files changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/net/dccp/proto.c b/net/dccp/proto.c
> index 875858c..34d48cd 100644
> --- a/net/dccp/proto.c
> +++ b/net/dccp/proto.c
> @@ -959,6 +959,7 @@ int inet_dccp_listen(struct socket *sock, int backlog)
>  		err = dccp_listen_start(sk, backlog);
>  		if (err)
>  			goto out;
> +		sk->sk_ack_backlog = 0;


This is racy, remember that dccp and tcp have lockless listeners.

Do not change sk_ack_backlog after a TCP/DCCP socket is ready to accept new flows.