[PATCH] s3:events: Call all ready fd event handlers on each iteration of the main loop

[simo]

On Tue, 2010-07-06 at 11:05 +0200, Volker Lendecke wrote:
> On Mon, Jul 05, 2010 at 06:01:17PM -0700, Steven Danneman wrote:
> > Attached is a proposed patch to make the main message loop in S3 more
> > fair to fd based events.  Currently, network traffic can starve out all
> > other fd based events. We hit this at Isilon with our fd based oplock
> > upcalls, but this issue should also affect iNotify on Linux and oplocks
> > on Irix.
> > 
> > As this is a major change to the server I wanted to pass it by the list
> > first.
> 
> Sorry, but I think this is the wrong fix. We deliberately
> only handle one single event in run_events. The reason is
> that while handling events the state of other pending events
> can change. This has caused us big pain in the past.
> 
> Yes, this is indeed a problem that I had to badly work
> around with d5cf648, but the fix needs to look a bit
> differently I think. We need to remove a file descriptor off
> the list of interested fd's whenever it is being handled.
> When either no fd's are left or in a select call nobody is
> active, we need to push all ones on the list again. This
> should give us round-robin handling of fds.
> 
> I am not sure if only select is affected. It might be the
> case that Linux under epoll is already fair in the sense
> that it never starves sockets.
> 
> Anybody around knowing for sure?

I've experimentally seen sockets starve under linux using epoll.

Simo.

-- 
Simo Sorce
Samba Team GPL Compliance Officer <simo at samba.org>
Principal Software Engineer at Red Hat, Inc. <simo at redhat.com>