Re: [PATCH 1/2] irq_work: allow certain work in hard irq context

[Mike Galbraith <bitbucket@online.de>]

On Mon, 2014-02-03 at 09:31 +0100, Sebastian Andrzej Siewior wrote: 
> On 02/03/2014 05:00 AM, Mike Galbraith wrote:
> > On Sun, 2014-02-02 at 21:10 +0100, Sebastian Andrzej Siewior wrote:
> > 
> >> According to the backtrace both of them are trying to access the
> >> per-cpu hrtimer (sched_timer) in order to cancel but they seem to fail
> >> to get the timer lock here. They shouldn't spin there for minutes, I
> >> have no idea why they did so…
> > 
> > Hm. per-cpu...
> > 
> > I've been chasing an rt hotplug heisenbug that is pointing to per-cpu
> > oddness.  During sched domain re-construction while running Steven's
> > stress script on 64 core box, we hit a freshly constructed domain with
> > _no span_, build_sched_groups()->get_group() explodes when we meeting
> > it.  But if you try to watch the thing appear... it just doesn't.
> > 
> > static int build_sched_domains(const struct cpumask *cpu_map,
> >                                struct sched_domain_attr *attr)
> > {
> >         enum s_alloc alloc_state;
> >         struct sched_domain *sd;
> >         struct s_data d;
> >         int i, ret = -ENOMEM;
> > 
> >         alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
> >         if (alloc_state != sa_rootdomain)
> >                 goto error;
> > 
> >         /* Set up domains for cpus specified by the cpu_map. */
> >         for_each_cpu(i, cpu_map) {
> >                 struct sched_domain_topology_level *tl;
> > 
> >                 sd = NULL;
> >                 for_each_sd_topology(tl) {
> >                         sd = build_sched_domain(tl, cpu_map, attr, sd, i);
> > BUG_ON(sd == spanless-alien) here..
> 
> spanless-alien is?

An sd spanning zero CPUs, cpumask is empty.

> BUG_ON() is actually _very_ cheap. It shouldn't even create any kind of
> compiler barrier which would reload variables / registers. It should
> evaluate sd and "spanless-alien", do the compare and then go on.

Nonetheless, the bad thing then refuses to happen.

> >                         if (tl == sched_domain_topology)
> >                                 *per_cpu_ptr(d.sd, i) = sd;
> >                         if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
> >                                 sd->flags |= SD_OVERLAP;
> >                         if (cpumask_equal(cpu_map, sched_domain_span(sd)))
> >                                 break;
> >                 }
> >         }
> > 
> >         /* Build the groups for the domains */
> >         for_each_cpu(i, cpu_map) {
> >                 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
> >                         sd->span_weight = cpumask_weight(sched_domain_span(sd));
> >                         if (sd->flags & SD_OVERLAP) {
> >                                 if (build_overlap_sched_groups(sd, i))
> >                                         goto error;
> >                         } else {
> >                                 if (build_sched_groups(sd, i))
> > ..prevents meeting that alien here.. while hotplug locked.
> 
> my copy of build_sched_groups() always returns 0 so it never goes to
> the error marker.

It explodes before return, in the call to get_group(), trying to write
to lala land.

> Did you consider a compiler bug? I could try to
> rebuild your source + config on two different compilers just to see if
> it makes a difference.

Possible, but seems very unlikely.  I suspect it's something to do with
the rather unusual topology of this box.  8 sockets, but 1 NUMA node
with a whole 8GB ram.  Lots of CPU, no way to use it well.  I suppose I
should try beating up a less crippled big box, see if its reproducible
elsewhere.  It takes a while to explode on this box, and may not ever
explode in a box that doesn't have 64 CPUs all poking the same couple
sticks of ram.

-Mike