[PATCH] Cache messaging dgm connections

[Ralph Böhme]

On Fri, Sep 16, 2016 at 10:53:29PM -0700, Jeremy Allison wrote:
> On Fri, Sep 16, 2016 at 10:06:56PM -0700, Jeremy Allison wrote:
> > On Fri, Sep 16, 2016 at 02:03:42PM -0700, Jeremy Allison wrote:
> > > On Fri, Sep 16, 2016 at 09:05:23AM -0700, Jeremy Allison wrote:
> > > > On Fri, Sep 16, 2016 at 06:00:44PM +0200, Ralph Böhme wrote:
> > > > > On Fri, Sep 16, 2016 at 08:36:33AM -0700, Jeremy Allison wrote:
> > > > > > On Fri, Sep 16, 2016 at 11:29:06AM +0200, Ralph Böhme wrote:
> > > > > > > On Thu, Sep 15, 2016 at 10:09:18AM -0700, Jeremy Allison wrote:
> > > > > > > > On Thu, Sep 15, 2016 at 06:38:25PM +0200, Ralph Böhme wrote:
> > > > > > > > > 
> > > > > > > > > I got one through with these two patches. Running more...
> > > > > > > > > 
> > > > > > > > > Cheerio!
> > > > > > > > 
> > > > > > > > Oh, this looks more elegant than the reparent struct poll_funcs_tevent_context
> > > > > > > > to ev patch I just posted...
> > > > > > > > 
> > > > > > > > The:
> > > > > > > > 
> > > > > > > >  +     TALLOC_FREE(msg->msg_dgm_ref);
> > > > > > > > 
> > > > > > > > looks very clever :-).
> > > > > > > 
> > > > > > > it passed three more autobuilds. Shall we push it?
> > > > > > 
> > > > > > Yes please - the version we discussed on the phone
> > > > > > with the free of the timer events followed by context,
> > > > > > plus the TALLOC_FREE(msg->msg_dgm_ref); change !
> > > > > 
> > > > > Then the attached is the one. Please push if ok.
> > > > 
> > > > 
> > > > It's missing the:
> > > > 
> > > > +     size_t num_contexts = talloc_array_length(state->contexts);
> > > > 
> > > > in the second patch. With that added, pushed !
> > > 
> > > Still got the crash with this in autobuild, now with the backtrace
> > > below. So we're still not catching everything (or something else
> > > is freeing ev behind our backs :-).

is it still the same test that triggers it? From the SBT it looks like.

> > > As I know we're going to be co-located on Monday, let's
> > > schedule a time to look at this directly at the SNIA
> > > conf.
> > 
> > In the meantime, here's what I'm running a private autobuild
> > with in the hope of catching the place where ev gets deleted.
> 
> Slightly better canary code that catches ev addition and
> destruction.

I'm now running autobuilds with it on sn-devel.

Cheerio!
-slow