Re: Stable regression with 'tcp: allow splice() to build full TSO packets'

[Willy Tarreau <w@1wt.eu>]

On Thu, May 17, 2012 at 11:50:10PM +0200, Eric Dumazet wrote:
> On Thu, 2012-05-17 at 23:40 +0200, Eric Dumazet wrote:
> 
> > I dont understand why we should tcp_push() if we sent 0 bytes in this
> > splice() call.
> > 
> > The push() should have be done already by prior splice() call, dont you
> > think ?
> > 
> > out:
> > 	if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
> > 		tcp_push(sk, flags, mss_now, tp->nonagle);
> > 
> 
> I think I now understand
> 
> One splice() syscall actually calls do_tcp_sendpages() several times
> (N).
> 
> Problem is N-1 calls are done with MSG_SENDPAGE_NOTLAST set
> 
> And last one with MSG_SENDPAGE_NOTLAST unset
> 
> So maybe we should replace this test by :
> 
> 	if (!(flags & MSG_SENDPAGE_NOTLAST))
> 		tcp_push(...);

Just tested, and it unfortunately does not fix the issue :-(

I'll see if I can log errors from lower layers.
Willy