Re: x86/mce: machine check warning during poweroff

[Suresh Siddha <suresh.b.siddha@intel.com>]

On Tue, 2012-01-17 at 15:22 +0530, Srivatsa S. Bhat wrote:
> Thanks for the patch, but unfortunately it doesn't fix the problem!
> Exactly the same stack traces are seen during a CPU Hotplug stress test.
> (I didn't even have to stress it - it is so fragile that just a script
> to offline all cpus except the boot cpu was good enough to reproduce the
> problem easily.)

hmm, that's weird. with the patch, sched_ilb_notifier() should have
cleared the cpu going offline from the nohz.idle_cpus_mask. And this
should have happened after that cpu is removed from active mask. So
no-one else should add that cpu back to the nohz.idle_cpus_mask and this
should prevent the issue from happening.

I could reproduce the problem easily with out the patch but when I
applied the patch I couldn't recreate the issue. Srivatsa, can you
please re-check the kernel you tested indeed has the fix?

re-Reviewing the code/patch also doesn't give me a hint.

> I have a few questions regarding the synchronization with CPU Hotplug.
> What guarantees that the code which selects and IPIs the new ilb is totally
> race-free with respect to CPU hotplug and we will never IPI an offline CPU?

So, nohz_balancer_kick() gets called only from interrupts disabled.
During that time (from selecting the ilb_cpu to sending the IPI), no cpu
can go offline. As the offline happens from the stop-machine process
context with interrupts disabled.

Only thing we need to make sure is the offlined cpu shouldn't be part of
the nohz.idle_cpus_mask and for post 3.2 code, posted patch ensures
that.

For 3.2 and before, when a cpu exits tickless idle, it gets removed from
the nohz.idle_cpus_mask (and also from the nohz.load_balancer). And if
the cpu is not in the active mask (while going offline), subsequent
calls to select_nohz_load_balancer() ensures that the cpu going down
doesn't update the nohz structures. So I thought 3.2 shouldn't exhibit
this problem.


> (As demonstrated above, this issue is in 3.2-rc7
> as well.)

hmm, don't think we ran into this before 3.2. So, what am I missing from
the above? I will try to reproduce it on 3.2 too.

thanks,
suresh