[PATCH v2 net-next] inet: do not set backlog if listen fails

We don't need to set the backlog if listen fails.
The sk_max_ack_backlog will be set in the caller inet_listen() and
dccp_listen_start() if inet_csk_listen_start() return without error.
So just remove this line to avoid this unnecessary operation.

Regarding sk_ack_backlog, we have to set it before a TCP/DCCP socket is
ready to accept new flows to avoid race, because dccp and tcp have lockless
listeners

Signed-off-by: Yafang Shao <redacted>
---
 net/ipv4/inet_connection_sock.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index dfd5009..cdd5c95 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c

@@ -871,7 +871,6 @@ int inet_csk_listen_start(struct sock *sk, int backlog)
 
 	reqsk_queue_alloc(&icsk->icsk_accept_queue);
 
-	sk->sk_max_ack_backlog = backlog;
 	sk->sk_ack_backlog = 0;
 	inet_csk_delack_init(sk);
 

-- 
1.8.3.1

On 10/07/2018 06:36 AM, Yafang Shao wrote:

We don't need to set the backlog if listen fails.
The sk_max_ack_backlog will be set in the caller inet_listen() and
dccp_listen_start() if inet_csk_listen_start() return without error.
So just remove this line to avoid this unnecessary operation.

Regarding sk_ack_backlog, we have to set it before a TCP/DCCP socket is
ready to accept new flows to avoid race, because dccp and tcp have lockless
listeners

Really could you not add irrelevant stuff in the changelog ?

This patch simply removes one redundant setting, you do not have to explain
about dccp/tcp being lockless and that sk_ack_backlog is not touched by this patch.

Also the title is misleading, since we will still set the backlog even if the listen
fails.

If you really want sk_max_ack_backlog being not changed if listen() fails,
then I am afraid you will need to submit a different patch.

On Mon, Oct 8, 2018 at 1:39 AM Eric Dumazet [off-list ref] wrote:

On 10/07/2018 06:36 AM, Yafang Shao wrote:

quoted

We don't need to set the backlog if listen fails.
The sk_max_ack_backlog will be set in the caller inet_listen() and
dccp_listen_start() if inet_csk_listen_start() return without error.
So just remove this line to avoid this unnecessary operation.

Regarding sk_ack_backlog, we have to set it before a TCP/DCCP socket is
ready to accept new flows to avoid race, because dccp and tcp have lockless
listeners

Really could you not add irrelevant stuff in the changelog ?

Sure, I will remove it.

This patch simply removes one redundant setting, you do not have to explain
about dccp/tcp being lockless and that sk_ack_backlog is not touched by this patch.

Also the title is misleading, since we will still set the backlog even if the listen
fails.

If you really want sk_max_ack_backlog being not changed if listen() fails,
then I am afraid you will need to submit a different patch.

After this patch, sk_max_ack_backlog will only be set after the sk is
successfully added to listening hash table,
and then 0 is returned, that means listen()  is successful.

Take inet_listen() for example,

if (old_state != TCP_LISTEN) {
    err = inet_csk_listen_start(sk, backlog);
    if (err)
        goto out;
}

// it can only be set if err is 0.
sk->sk_max_ack_backlog = backlog;
err = 0;

out:
    return err;


Pls. correct me if I missed something.

Thanks
Yafang

On Mon, Oct 8, 2018 at 3:47 AM Yafang Shao [off-list ref] wrote:

Pls. correct me if I missed something.

You missed that this variable must be set before listener is setup.
Same feedback than prior version really (when you trived to move
around sk_ack_backlog clearing)

You are changing inet_csk_listen_start() but this is wrong...

On Mon, Oct 8, 2018 at 10:48 PM Eric Dumazet [off-list ref] wrote:

On Mon, Oct 8, 2018 at 3:47 AM Yafang Shao [off-list ref] wrote:

quoted

Pls. correct me if I missed something.

You missed that this variable must be set before listener is setup.
Same feedback than prior version really (when you trived to move
around sk_ack_backlog clearing)

You are changing inet_csk_listen_start() but this is wrong...

Got it.
Thank you.