Re: [PATCH] perf, x86: catch spurious interrupts after disabling counters

[Cyrill Gorcunov <gorcunov@gmail.com>]

On Wed, Sep 15, 2010 at 06:47:51PM +0200, Stephane Eranian wrote:
> On Wed, Sep 15, 2010 at 6:46 PM, Cyrill Gorcunov <gorcunov@gmail.com> wrote:
> > On Wed, Sep 15, 2010 at 06:20:34PM +0200, Robert Richter wrote:
> >> On 14.09.10 19:41:32, Robert Richter wrote:
> >> > I found the reason why we get the unknown nmi. For some reason
> >> > cpuc->active_mask in x86_pmu_handle_irq() is zero. Thus, no counters
> >> > are handled when we get an nmi. It seems there is somewhere a race
> >> > accessing the active_mask. So far I don't have a fix available.
> >> > Changing x86_pmu_stop() did not help:
> >>
> >> The patch below for tip/perf/urgent fixes this.
> >>
> >> -Robert
> >>
> >> From 4206a086f5b37efc1b4d94f1d90b55802b299ca0 Mon Sep 17 00:00:00 2001
> >> From: Robert Richter <robert.richter@amd.com>
> >> Date: Wed, 15 Sep 2010 16:12:59 +0200
> >> Subject: [PATCH] perf, x86: catch spurious interrupts after disabling counters
> >>
> >> Some cpus still deliver spurious interrupts after disabling a counter.
> >> This caused 'undelivered NMI' messages. This patch fixes this.
> >>
> >> Signed-off-by: Robert Richter <robert.richter@amd.com>
> >> ---
> > ...
> >
> > Hi Robert, thanks a lot for tracking this issue! I might be missing
> > something but why don't you clean this ->running mask bits on pmu-stop?
> > What if counter gets disabled/freed or whatever before issue any nmis?
> > Another question I have still -- is this an hardware issue in general?
> >
> what's for sure, is that you can have an interrupt in flight by the time
> you disable.
> 

I fear you can x86_pmu_stop()

	if (__test_and_clear_bit(hwc->idx, cpuc->active_mask)) {

--->		active_mask will be cleared here for sure
--->		but counter still ticks, say nmi happens active_mask
--->		is cleared, but NMI can still happen and gets buffered
--->		before you disable counter in real

		x86_pmu.disable(event);
		cpuc->events[hwc->idx] = NULL;
		WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
		hwc->state |= PERF_HES_STOPPED;
	}

No?

	-- Cyrill