[PATCH] Do not close winbind socket during use

[Andrew Bartlett]

On Tue, 2013-07-09 at 22:36 +0200, Stefan (metze) Metzmacher wrote:
> Am 09.07.2013 18:03, schrieb Stefan (metze) Metzmacher:
> > Am 09.07.2013 17:33, schrieb Stefan (metze) Metzmacher:
> >> Hi Andrew,
> >>
> >> Am 03.07.2013 09:44, schrieb Andrew Bartlett:
> >>> On Thu, 2013-06-27 at 11:42 +1000, Andrew Bartlett wrote:
> >>>> On Wed, 2013-06-26 at 20:39 +1000, Andrew Bartlett wrote:
> >>>>> On Mon, 2013-06-24 at 15:26 +0000, Philippe.Simonet at swisscom.com wrote:
> >>>>>> Hi Andrew, and by putting more num-callers : 
> >>>>>>
> >>>>>> valgrind --num-callers=50 samba -i -M single
> >>>>>
> >>>>> Thanks for getting me that.  I've managed to reproduce it here, but not
> >>>>> under valgrind, and only when I hack the code to force a timeout.  At
> >>>>> least this should help me figure out why we process the winbind socket
> >>>>> close, which is the crux of this issue.
> >>>>
> >>>> I think I've found the cause of the issue you are hitting.  There is
> >>>> still another issue with the nested event loop in the krb5 libs, but
> >>>> these two patches should help significantly.
> >>>>
> >>>> As you have had more luck than I in reproducing this in a unaltered
> >>>> setting, please let me know if this helps.
> >>>>
> >>>> Patches are for git master, but may apply to 4.0 as well.
> >>>
> >>> G'Day,
> >>>
> >>> The original reporter has confirmed to me that this removes the segfault
> >>> for him.  It changes it to a 105 sec hang, (due to the winbind client
> >>> trying for 5 second at at a time many times). 
> >>>
> >>> Can I get a review on it so we can rid master and eventually 4.0 of this
> >>> nasty crash?
> >>
> >> I've looked through this patches and have some improvements.
> >> The main problem is that we're not sure wbsrv_call_loop() is called again
> >> on the terminated connection, when the last pending request is finished.
> >> That's why I remember all broken connections and try to clean them up
> >> before accepting a new connection or processing any new request on any
> >> connection.
> >> This way we're sure the connection gets removed eventually.
> >>
> >> I'm currently running some autobuild with the attached patches,
> >> they might also fix the current flakey crashes, e.g.
> >> https://git.samba.org/autobuild.flakey/2013-07-08-0055/samba.stderr
> > 
> > Here's the next try, which hopefully don't crash in make test :-)
> 
> Ok, it passed 4 times on master and 4 times on v4-0-test,
> if you're ok with it I'll squash my changes and the missing
> Pair-programmed-with:, Signed-off-by:, Reviewed-by: tags and push it...
> 
> Are you fine with that?

Thanks, please do that.

Signed-off-by: Andrew Bartlett <abartlet at samba.org>
Reviewed-by: Andrew Bartlett <abartlet at samba.org>

Andrew Bartlett

-- 
Andrew Bartlett                                http://samba.org/~abartlet/
Authentication Developer, Samba Team           http://samba.org