Re: [PATCH] inet: race in wait for connect.

[Stephen Hemminger <shemminger@linux-foundation.org>]

On Mon, 29 Oct 2007 15:38:42 -0700 (PDT)
David Miller <davem@davemloft.net> wrote:

> From: Stephen Hemminger <shemminger@linux-foundation.org>
> Date: Mon, 29 Oct 2007 14:37:14 -0700
> 
> > On Mon, 29 Oct 2007 19:29:06 -0200
> > "Arnaldo Carvalho de Melo" <acme@redhat.com> wrote:
> > 
> > > Em Mon, Oct 29, 2007 at 01:52:22PM -0700, Stephen Hemminger escreveu:
> > > > Fix possible race while waiting for connections in accept. I don't
> > > > know of a test case that could reproduce this directly.
> > > > 
> > > > The state of the socket should be checked before checking the queue.
> > > > If the socket has left the TCP_LISTEN state, then the accept queue
> > > > is no longer valid.
> > > 
> > > Well if it left from LISTEN to CLOSED inet_csk_listen_stop must have
> > > been called, and that calls reqsk_queue_yank_acceptq, that sets it to
> > > NULL, reqsk_queue_empty(&icsk->icsk_accept_queue) returns true and we
> > > get to if (sk->sk_state != TCP_LISTEN), returning -EINVAL as with your
> > > patch.  So I can't see a race, just one branch less in one case :-)
> > > 
> > > - Arnaldo
> > 
> > Yeah, your right. The listen queue is garbage at this point but the
> > accept queue is always empty.
> 
> Is there some failing test case you are aware of which led
> you to this piece of code or did you happen to notice it
> while scanning around?

While instrumenting the snot out of accept logic for looking at alternatives
to the MT-accept vs. close issue.

-- 
Stephen Hemminger <shemminger@linux-foundation.org>