On Mon, Sep 21, 2015 at 03:20:14PM +0200, Emeric Brun wrote:
> On 09/21/2015 03:03 PM, Willy Tarreau wrote:
> > On Mon, Sep 21, 2015 at 06:25:39PM +0530, Pradeep Jindal wrote:
> >> I will give you more info in a while. I did some debugging,
> >> "peers->sync_task = NULL;" seems to be the culprit here, First,
> >> peers->sync_task is being set to NULL and then being deleted & freed which
> >> I think would obviously fail. I moved "peers->sync_task = NULL;" to after
> >> task_free call in the code below and it works as expected now. Let me know
> >> if this makes sense.
> >>
> >> ==== peers.c ====
> >> static struct task *process_peer_sync(struct task * task)
> >> {
> >> struct peers *peers = (struct peers *)task->context;
> >> struct peer *ps;
> >> struct shared_table *st;
> >>
> >> task->expire = TICK_ETERNITY;
> >>
> >> if (!peers->peers_fe) {
> >> /* this one was never started, kill it */
> >> signal_unregister_handler(peers->sighandler);
> >> peers->sync_task = NULL;
> >> task_delete(peers->sync_task);
> >> task_free(peers->sync_task);
> >> return NULL;
> >> }
> >> ==== peers.c ====
> >
> > Absolutely! This is definitely a bug. CCing Emeric who may have an idea
> > when this may happen.
> >
> > Thanks,
> > Willy
> >
>
> Yep, it's a bug! I think the issue should happen when a 'peers' section isn't
> used by any 'frontend/backend' bound on current process.
I confirm. 5 out of 6 processes die with Pradeep's config. Since I didn't
know what to look for, I didn't notice.
I'm applying your fix, Pradeep. Thanks!
Willy
