net: remove release/lock_sock in tcp_splice_read
Subsystems (3)
PATCH 1/1 net: remove release/lock_sock in tcp_splice_read
When enters tcp_splice_read, tcp_splice_read will call lock_sock for sk in order to prevent other threads acquiring sk and release it before return.
But in while(tss.len) loop, it releases and re-locks sk, give the other thread a small window to lock the sk.
As a result, release/lock_sock in the while loop in tcp_splice_read may cause race condition.
--- net/ipv4/tcp.c | 2 -- 1 file changed, 2 deletions(-)
net/ipv4/tcp.c +0 -2
@@ -856,8 +856,6 @@ ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,if (!tss.len || !timeo)break ; -release_sock (sk); -lock_sock (sk); if (sk->sk_err || sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN) ||
Signature · 1 line
2.34.1
Paolo Abeni
maintainerQuoted 9 lines · sunyiqi
When enters
tcp_splice_read,tcp_splice_readwill calllock_sockfor sk in order to prevent other threads acquiring sk and release it before return.But in while(tss.len) loop, it releases and re-locks sk, give the other thread a small window to lock the sk.
As a result, release/lock_sock in the while loop in
tcp_splice_readmay cause race condition.
Which race condition exactly? do you have a backtrace?
Quoted 12 lines
Fixes 9c55e01c0cc8 ("[TCP]: Splice receive support.")Signed-off-by sunyiqi <[email protected]>--- net/ipv4/tcp.c | 2 -- 1 file changed, 2 deletions(-)net/ipv4/tcp.c +0
-0
net/ipv4/tcp.c ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, · line 860
Quoted 5 lines
net/ipv4/tcp.c +0
-2@@ -856,8 +856,6 @@ ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,if (!tss.len || !timeo)break ; -release_sock (sk); -lock_sock (sk);
This is needed to flush the sk backlog.
Somewhat related, I think we could replace the pair with sk_flush_backlog().
Thanks,
Paolo
Hello Paolo,
On Thu, Aug 15, 2024 at 6:40 PM Paolo Abeni <[email protected]> wrote:
[...]
Quoted 7 lines
net/ipv4/tcp.c ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, · line 860
Quoted 2 lines
net/ipv4/tcp.c +0
-2-release_sock (sk); -lock_sock (sk);This is needed to flush the sk backlog.
Somewhat related, I think we could replace the pair with
sk_flush_backlog().
Do you think we could do this like the following commit:
commit d41a69f1d390fa3f2546498103cdcd78b30676ff
Author: Eric Dumazet <[email protected]> Date: Fri Apr 29 14:16:53 2016 -0700
tcp: make tcp_sendmsg() aware of socket backlog
Large sendmsg()/write() hold socket lock for the duration of the call, unless sk->sk_sndbuf limit is hit. This is bad because incoming packets are parked into socket backlog for a long time.
?
Then we can avoid taking the lock too long which results in too many packets in the backlog.
Thanks,
Jason
Paolo Abeni
maintainerQuoted 16 lines · Jason Xing
On Thu, Aug 15, 2024 at 6:40 PM Paolo Abeni <[email protected]> wrote:
[...]Quoted 7 lines
net/ipv4/tcp.c ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, · line 860
Quoted 2 lines
net/ipv4/tcp.c +0
-2-release_sock (sk); -lock_sock (sk);This is needed to flush the sk backlog.
Somewhat related, I think we could replace the pair with
sk_flush_backlog().Do you think we could do this like the following commit:
commit
d41a69f1d390fa3f2546498103cdcd78b30676ffAuthor: Eric Dumazet <[email protected]> Date: Fri Apr 29 14:16:53 2016 -0700
Quoted 6 lines
tcp: make
tcp_sendmsg()aware of socket backlogLarge sendmsg()/write() hold socket lock for the duration of the call, unless
sk->sk_sndbuflimit is hit. This is bad because incoming packets are parked into socket backlog for a long time. >
?
Yep. To be more accurate I was looking at commit 93afcfd1db35882921b2521a637c78755c27b02c
In any case this should be unrelated from the supposed issue.
Cheers,
Paolo
Quoted 22 lines · Paolo Abeni
Quoted 18 lines · Jason Xing
On Thu, Aug 15, 2024 at 6:40 PM Paolo Abeni <[email protected]> wrote:
[...]Quoted 7 lines
net/ipv4/tcp.c ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos, · line 860
Quoted 2 lines
net/ipv4/tcp.c +0
-2-release_sock (sk); -lock_sock (sk);This is needed to flush the sk backlog.
Somewhat related, I think we could replace the pair with
sk_flush_backlog().Do you think we could do this like the following commit:
commit
d41a69f1d390fa3f2546498103cdcd78b30676ffAuthor: Eric Dumazet <[email protected]> Date: Fri Apr 29 14:16:53 2016 -0700tcp: make
tcp_sendmsg()aware of socket backlog
Quoted 7 lines
Quoted 4 lines
Large sendmsg()/write() hold socket lock for the duration of the call, unless
sk->sk_sndbuflimit is hit. This is bad because incoming packets are parked into socket backlog for a long time. >
?Yep. To be more accurate I was looking at commit
93afcfd1db35882921b2521a637c78755c27b02c
Thanks. It arouses my interest. Now I do believe we can do such optimization in this function.
Quoted 2 lines
In any case this should be unrelated from the supposed issue.
Sure.
Thanks,
Jason
Quoted 5 lines · sunyiqi
When enters
tcp_splice_read,tcp_splice_readwill calllock_sockfor sk in order to prevent other threads acquiring sk and release it before return.
Quoted 8 lines
But in while(tss.len) loop, it releases and re-locks sk, give the other thread a small window to lock the sk.
As a result, release/lock_sock in the while loop in
tcp_splice_readmay cause race condition.Fixes 9c55e01c0cc8 ("[TCP]: Splice receive support.")Signed-off-by sunyiqi <[email protected]>
It's more of an optimization instead of a BUG, no?
I don't consider it as a bug, unless you can prove it... Let me ask what kind of race issues could re-lock cause?
I think holding the socket lock too long is not a good idea because releasing the lock can give others breathing (having the chance to finish their own stuff). Please see release_sock().
Thanks,
Jason
Quoted 21 lines
--- net/ipv4/tcp.c | 2 -- 1 file changed, 2 deletions(-)net/ipv4/tcp.c +0
-3@@ -856,8 +856,6 @@ ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,if (!tss.len || !timeo)break ; -release_sock (sk); -lock_sock (sk); if (sk->sk_err || sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN) || --2.34.1