Re: [-next regression] TCP window full with EPOLLWAKEUP

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Monday, May 21, 2012, Jiri Slaby wrote:
> On 05/20/2012 08:34 PM, Rafael J. Wysocki wrote:
> > On Sunday, May 20, 2012, Rafael J. Wysocki wrote:
> >> On Sunday, May 20, 2012, Rafael J. Wysocki wrote:
> >>> On Sunday, May 20, 2012, Jiri Slaby wrote:
> >>>> Hi,
> >>>>
> >>>> a bisection shows that with the following commit from -next:
> >>>> commit 4d7e30d98939a0340022ccd49325a3d70f7e0238
> >>>> Author: Arve Hjønnevåg <arve@android.com>
> >>>> Date:   Tue May 1 21:33:34 2012 +0200
> >>>>
> >>>>     epoll: Add a flag, EPOLLWAKEUP, to prevent suspend while epoll
> >>>> events are ready
> >>>>
> >>>> ====
> >>>>
> >>>> one of mono programs I use stops receiving data from the network.
> >>>> Wireshark shows that the TCP window of a connection is filled. This
> >>>> means the program does not read the data fast enough after requesting
> >>>> the data.
> >>>>
> >>>> If I revert that commit on the top of -next (20120518), everything works
> >>>> as expected.
> >>>
> >>> Hmm.  I suppose that the failing program doesn't set EPOLLWAKEUP by mistake,
> >>> does it?
> >>
> >> If it doesn't, we can assume that epi-ws is always NULL and all of the added
> >> overhead comes from the function calls.  So, I wonder if the appended patch
> >> makes any difference?
> > 
> > Having thought more about this I have to say this doesn't seem to make much
> > sense, because in that case you'd see some progress, although probably a bit
> > slower than before.
> > 
> > So, I think what happens is that the application tries to set EPOLLWAKEUP,
> > but doesn't have the capability, so the entire operation fails for it, but
> > it doesn't check the return value.
> > 
> > I wonder if the following helps, then.
> > 
> > Thanks,
> > Rafael
> > 
> > 
> > ---
> >  fs/eventpoll.c |    2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > Index: linux/fs/eventpoll.c
> > ===================================================================
> > --- linux.orig/fs/eventpoll.c
> > +++ linux/fs/eventpoll.c
> > @@ -1711,7 +1711,7 @@ SYSCALL_DEFINE4(epoll_ctl, int, epfd, in
> >  
> >  	/* Check if EPOLLWAKEUP is allowed */
> >  	if ((epds.events & EPOLLWAKEUP) && !capable(CAP_EPOLLWAKEUP))
> > -		goto error_tgt_fput;
> > +		epds.events &= ~EPOLLWAKEUP;
> 
> Yes, this fixed the issue.

Cool, thanks for testing!

I'll resend it in a while with a changelog and tags.

Thanks,
Rafael