On Tue, 2016-08-16 at 13:19 -0700, Cong Wang wrote:
> On Tue, Aug 16, 2016 at 11:39 AM, Eric Dumazet <eric.duma...@gmail.com> wrote:
> > On Tue, 2016-08-16 at 08:39 -0700, Eric Dumazet wrote:
> >> On Tue, 2016-08-16 at 12:45 +0200, Greg KH wrote:
> >> > For some reason Marco's emails can't make it to netdev, so I'm
> >> > forwarding this on. Please cc: him on responses.
> >>
> >> Thanks for the report Greg and Marco.
> >>
> >> My first guess is this is caused by
> >>
> >> d41a69f1d390 tcp: make tcp_sendmsg() aware of socket backlog
> >>
> >> And a combination of funky sendmsg() flags (like FastOpen)
> >>
> >> I will look at this problem today.
> >>
> >
> > No, above commit was innocent ;)
> >
> > It looks like the bug is very old, and following patch would fix it.
> > I will submit it formally after few tests.
> >
> >
> > diff --git a/include/net/tcp.h b/include/net/tcp.h
> > index c00e7d51bb18..7717302cab91 100644
> > --- a/include/net/tcp.h
> > +++ b/include/net/tcp.h
> > @@ -1523,6 +1523,8 @@ static inline void tcp_check_send_head(struct sock
> > *sk, struct sk_buff *skb_unli
> > {
> > if (sk->sk_send_head == skb_unlinked)
> > sk->sk_send_head = NULL;
> > + if (tcp_sk(sk)->highest_sack == skb_unlinked)
> > + tcp_sk(sk)->highest_sack = NULL;
> > }
>
> Hmm, but from the stack traces it indicates the skb is freed
> inside tcp_sendmsg(), which must be:
>
>
> do_fault:
> if (!skb->len) {
> tcp_unlink_write_queue(skb, sk);
> /* It is the one place in all of TCP, except connection
> * reset, where we can be unlinking the send_head.
> */
> tcp_check_send_head(sk, skb);
> sk_wmem_free_skb(sk, skb);
> }
>
> In this case, skb->len == 0 means it is newly allocated skb by
> sk_stream_alloc_skb(), so it should not have a chance to be
> picked by tp->highest_sack yet b/c the whole function locks
> the sock?
>
> I must miss something here.
Look at skb_entail() : It calls tcp_add_write_queue_tail()
And tcp_add_write_queue_tail() looks like :
static inline void tcp_add_write_queue_tail(struct sock *sk, struct sk_buff
*skb)
{
__tcp_add_write_queue_tail(sk, skb);
/* Queue it, remembering where we must start sending. */
if (sk->sk_send_head == NULL) {
sk->sk_send_head = skb;
if (tcp_sk(sk)->highest_sack == NULL)
tcp_sk(sk)->highest_sack = skb;
}
}
So we definitely need to undo what tcp_add_write_queue_tail() did.
