On Mon, 2017-01-23 at 11:56 -0500, Jason Baron wrote:
> On 01/23/2017 09:30 AM, Oleg Nesterov wrote:
> > Hello,
> >
> > smp_mb__after_atomic() looks wrong and misleading, sock_reset_flag() does
> > the
> > non-atomic __clear_bit() and thus it can not guarantee
> > test_bit(SOCK_NOSPACE)
> > (non-atomic too) won't be reordered.
> >
>
> Indeed. Here's a bit of discussion on it:
> http://marc.info/?l=linux-netdev&m=146662325920596&w=2
>
> > It was added by 3c7151275c0c9a "tcp: add memory barriers to write space
> > paths"
> > and the patch looks correct in that we need the barriers in
> > tcp_check_space()
> > and tcp_poll() in theory, so it seems tcp_check_space() needs smp_mb() ?
> >
>
> Yes, I think it should be upgraded to an smp_mb() there. If you agree
> with this analysis, I will send a patch to upgrade it. Note, I did not
> actually run into this race in practice.
SOCK_QUEUE_SHRUNK is used locally in TCP, it is not used by tcp_poll().
(Otherwise it would be using atomic set/clear operations)
I do not see obvious reason why we have this smp_mb__after_atomic() in
tcp_check_space().
But looking at this code, it seems we lack one barrier if sk_sndbuf is
ever increased. Fortunately this almost never happen during TCP session
lifetime...
diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index
bfa165cc455ad0a9aea44964aa663dbe6085aebd..3692e9f4c852cebf8c4d46c141f112e75e4ae66d
100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c
@@ -331,8 +331,13 @@ static void tcp_sndbuf_expand(struct sock *sk)
sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2;
sndmem *= nr_segs * per_mss;
- if (sk->sk_sndbuf < sndmem)
+ if (sk->sk_sndbuf < sndmem) {
sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]);
+ /* Paired with second sk_stream_is_writeable(sk)
+ * test from tcp_poll()
+ */
+ smp_wmb();
+ }
}
/* 2. Tuning advertised window (window_clamp, rcv_ssthresh)
