Re: [PATCH v2] perf/core: fix a possible deadlock scenario

[Peter Zijlstra]

On Mon, Jul 23, 2018 at 06:44:43PM -0700, Cong Wang wrote:
> On Mon, Jul 23, 2018 at 6:35 PM Cong Wang <xiyou.wangcong@xxxxxxxxx> wrote:
> >
> > Hi, Peter, Andi
> >
> > While reviewing the deadlock, I find out it looks like we could have the
> > following infinite recursion too:
> >
> > perf_event_account_interrupt()
> > __perf_event_account_interrupt()
> > perf_adjust_period()
> > event->pmu->stop
> > x86_pmu_stop()
> > x86_pmu.disable()
> 
> Hmm, x86_pmu_stop() calls __test_and_clear_bit(), so
> we should not call x86_pmu.disable() twice here.

Right, but since we set HES_UPTODATE after calling
x86_perf_event_update() that can in fact recurse.

Now, I don't think that'll be fatal, but it might be good to test that.

If you pick these patches:

  https://lkml.kernel.org/r/20170928121823.430053219@xxxxxxxxxxxxx

use force_early_printk (and actually configure a serial early_printk)
and put a printk() in x86_pmu_stop() and then run the perf_fuzzer or
something to try and reproduce.

But paranoia suggets moving that HES_UPTODATE thing one line up.

> > intel_pmu_disable_event()
> > intel_pmu_pebs_disable()
> > intel_pmu_drain_pebs_buffer()
> > intel_pmu_drain_pebs_nhm()
> > <repeat....>
> >
> > This time is pure hardware events, attr.freq must be non-zero.
> >
> > And, we could enter this infinite recursion in NMI handler too:
> >
> > intel_pmu_handle_irq()
> > perf_event_overflow()
> > __perf_event_overflow()
> > __perf_event_account_interrupt()
> > ....
> >
> > Or this is impossible too?

I'm not sure I see this second one.. can you be a little more specific?