Winbind using 100% CPU

[Jeremy Allison]

On Tue, Apr 16, 2013 at 01:46:14PM -0400, Jim McDonough wrote:
> On Tue, Apr 16, 2013 at 10:00 AM, Jim McDonough <jmcd at samba.org> wrote:
> > On Thu, Apr 11, 2013 at 6:59 PM, Jeremy Allison <jra at samba.org> wrote:
> >> On Thu, Apr 11, 2013 at 09:58:33AM -0400, Dylan Klomparens wrote:
> >>> (Re-posting on this email list per Jeremy Allison's request.)
> >>>
> >>> I am trying to figure out why winbind is using 100% CPU on my file server.
> >>> I am using Samba version 4.0.4. Everything is fine for a few minutes when I
> >>> start winbind, however after a while it begins using 100% CPU. I haven't
> >>> been able to narrow down what triggers this CPU usage spike, but I did
> >>> attach the GNU debugger to find out what's going on in the process. The
> >>> backtrace revealed this information:
> >>>
> >>> #0  0x000000000041cf30 in _talloc_free at plt ()
> >>> #1  0x0000000000452320 in winbindd_reinit_after_fork ()
> >>> #2  0x00000000004524e6 in fork_domain_child ()
> >>> #3  0x0000000000453585 in wb_child_request_trigger ()
> >>> #4  0x000000381d2048e2 in tevent_common_loop_immediate () from
> >>> /lib64/libtevent.so.0
> >>> #5  0x00007fbed6b98e17 in run_events_poll () from /lib64/libsmbconf.so.0
> >>> #6  0x00007fbed6b9922e in s3_event_loop_once () from /lib64/libsmbconf.so.0
> >>> #7  0x000000381d204060 in _tevent_loop_once () from /lib64/libtevent.so.0
> >>> #8  0x000000000042049a in main ()
> >>>
> >>> Apparently it's stuck in the winbindd_reinit_after_fork (and more
> >>> specifically the _talloc_free function). This code resides in
> >>> $SOURCE_HOME\source3\winbindd\winbindd_dual.c.
> >>
> >> That looks like corrupted memory - probably a loop
> >> in the talloc tree.
> > I've got a user who sees this and we're adding the same dlinklist
> > element twice, creating a loop in the winbind child list.
> >
> > I've got a broken wrist so responses take a while, but that's my
> > current hint.  On 3.6.3 and 3.6.13.
> >
> 
> I see this in the parent winbind log.  The last 3 entries are changes
> I made to the dlist macros within winbind only.  you can see a 2
> second delay and then a second add of the same item to the child
> winbind list.  No entries in between (and a production system so there
> is reluctance to increase the debug level).
> 
> [2013/03/14 09:08:57.795585,  3]
> winbindd/winbindd_getpwnam.c:56(winbindd_getpwnam_send)
>   getpwnam xx+xxxxxxxxx-112$
> [2013/03/14 09:08:57.796185,  3]
> winbindd/winbindd_getpwnam.c:56(winbindd_getpwnam_send)
>   getpwnam xx+yyyyyyy
> [2013/03/14 09:09:00.077660,  0]
> winbindd/winbindd_dual.c:1399(fork_domain_child)
>   adding 0x7ff91105a000 to list at 0x7ff911053510
> [2013/03/14 09:09:02.435367,  0]
> winbindd/winbindd_dual.c:1399(fork_domain_child)
>   adding 0x7ff91105a000 to list at 0x7ff91105a000
> [2013/03/14 09:09:02.435510,  0] lib/util.c:1117(smb_panic)
>   PANIC (pid 28628): duplicate

Jim, do your changes lead to a smb_panic on an
DLIST_ADD of a duplicate ? If so, can you post
a complete backtrace ?

Jeremy.