[PATCH net-next v2] unix: Improve locking scheme in unix_show_fdinfo()

[PATCH net-next v2] unix: Improve locking scheme in unix_show_fdinfo()

[Kirill Tkhai]

After switching to TCP_ESTABLISHED or TCP_LISTEN sk_state, alive SOCK_STREAM
and SOCK_SEQPACKET sockets can't change it anymore (since commit 3ff8bff704f4
"unix: Fix race in SOCK_SEQPACKET's unix_dgram_sendmsg()").

Thus, we do not need to take lock here.

Signed-off-by: Kirill Tkhai <tkhai@ya.ru>
---
v2: Initialize "nr_fds = 0".

 net/unix/af_unix.c |   20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)

diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
index f0c2293f1d3b..009616fa0256 100644
--- a/net/unix/af_unix.c
+++ b/net/unix/af_unix.c
@@ -807,23 +807,23 @@ static int unix_count_nr_fds(struct sock *sk)
 static void unix_show_fdinfo(struct seq_file *m, struct socket *sock)
 {
 	struct sock *sk = sock->sk;
+	unsigned char s_state;
 	struct unix_sock *u;
-	int nr_fds;
+	int nr_fds = 0;
 
 	if (sk) {
+		s_state = READ_ONCE(sk->sk_state);
 		u = unix_sk(sk);
-		if (sock->type == SOCK_DGRAM) {
-			nr_fds = atomic_read(&u->scm_stat.nr_fds);
-			goto out_print;
-		}
 
-		unix_state_lock(sk);
-		if (sk->sk_state != TCP_LISTEN)
+		/* SOCK_STREAM and SOCK_SEQPACKET sockets never change their
+		 * sk_state after switching to TCP_ESTABLISHED or TCP_LISTEN.
+		 * SOCK_DGRAM is ordinary. So, no lock is needed.
+		 */
+		if (sock->type == SOCK_DGRAM || s_state == TCP_ESTABLISHED)
 			nr_fds = atomic_read(&u->scm_stat.nr_fds);
-		else
+		else if (s_state == TCP_LISTEN)
 			nr_fds = unix_count_nr_fds(sk);
-		unix_state_unlock(sk);
-out_print:
+
 		seq_printf(m, "scm_fds: %u\n", nr_fds);
 	}
 }

Re: [PATCH net-next v2] unix: Improve locking scheme in unix_show_fdinfo()

[Kuniyuki Iwashima]

From:   Kirill Tkhai <tkhai@ya.ru>
Date:   Sat, 14 Jan 2023 12:35:02 +0300
> After switching to TCP_ESTABLISHED or TCP_LISTEN sk_state, alive SOCK_STREAM
> and SOCK_SEQPACKET sockets can't change it anymore (since commit 3ff8bff704f4
> "unix: Fix race in SOCK_SEQPACKET's unix_dgram_sendmsg()").
> 
> Thus, we do not need to take lock here.
> 
> Signed-off-by: Kirill Tkhai <tkhai@ya.ru>
> ---
> v2: Initialize "nr_fds = 0".

Yes, this is necessary because the new if-else does not cover
(SOCK_STREAM, TCP_CLOSE), and in such a case, nr_fds is
uninitialised val with the v1 patch.

This version looks good.

Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>


> 
>  net/unix/af_unix.c |   20 ++++++++++----------
>  1 file changed, 10 insertions(+), 10 deletions(-)
> 
> diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
> index f0c2293f1d3b..009616fa0256 100644
> --- a/net/unix/af_unix.c
> +++ b/net/unix/af_unix.c
> @@ -807,23 +807,23 @@ static int unix_count_nr_fds(struct sock *sk)
>  static void unix_show_fdinfo(struct seq_file *m, struct socket *sock)
>  {
>  	struct sock *sk = sock->sk;
> +	unsigned char s_state;
>  	struct unix_sock *u;
> -	int nr_fds;
> +	int nr_fds = 0;
>  
>  	if (sk) {
> +		s_state = READ_ONCE(sk->sk_state);
>  		u = unix_sk(sk);
> -		if (sock->type == SOCK_DGRAM) {
> -			nr_fds = atomic_read(&u->scm_stat.nr_fds);
> -			goto out_print;
> -		}
>  
> -		unix_state_lock(sk);
> -		if (sk->sk_state != TCP_LISTEN)
> +		/* SOCK_STREAM and SOCK_SEQPACKET sockets never change their
> +		 * sk_state after switching to TCP_ESTABLISHED or TCP_LISTEN.
> +		 * SOCK_DGRAM is ordinary. So, no lock is needed.
> +		 */
> +		if (sock->type == SOCK_DGRAM || s_state == TCP_ESTABLISHED)
>  			nr_fds = atomic_read(&u->scm_stat.nr_fds);
> -		else
> +		else if (s_state == TCP_LISTEN)
>  			nr_fds = unix_count_nr_fds(sk);
> -		unix_state_unlock(sk);
> -out_print:
> +
>  		seq_printf(m, "scm_fds: %u\n", nr_fds);
>  	}
>  }

Re: [PATCH net-next v2] unix: Improve locking scheme in unix_show_fdinfo()

[patchwork-bot+netdevbpf]

Hello:

This patch was applied to netdev/net-next.git (master)
by David S. Miller <davem@davemloft.net>:

On Sat, 14 Jan 2023 12:35:02 +0300 you wrote:
> After switching to TCP_ESTABLISHED or TCP_LISTEN sk_state, alive SOCK_STREAM
> and SOCK_SEQPACKET sockets can't change it anymore (since commit 3ff8bff704f4
> "unix: Fix race in SOCK_SEQPACKET's unix_dgram_sendmsg()").
> 
> Thus, we do not need to take lock here.
> 
> Signed-off-by: Kirill Tkhai <tkhai@ya.ru>
> 
> [...]

Here is the summary with links:
  - [net-next,v2] unix: Improve locking scheme in unix_show_fdinfo()
    https://git.kernel.org/netdev/net-next/c/b27401a30ee4

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html