Re: [PATCH 4/4] [x86] perf: fix accidentally ack'ing a second eventon intel perf counter

[Stephane Eranian]

On Thu, Sep 2, 2010 at 5:47 PM, Don Zickus <dzickus@xxxxxxxxxx> wrote:
> On Thu, Sep 02, 2010 at 04:39:02PM +0200, Stephane Eranian wrote:
>> Don,
>>
>> On Thu, Sep 2, 2010 at 4:19 PM, Don Zickus <dzickus@xxxxxxxxxx> wrote:
>> > On Thu, Sep 02, 2010 at 10:13:19AM +0200, Stephane Eranian wrote:
>> >> Robert,
>> >>
>> >> Do you have the test program you used to test this?
>> >> I believe the NHM hack does not solve the problem, it
>> >> just makes it harder to appear.
>> >
>> > Could be.
>> >
>> >>
>> >> I suspect the real issue is that the GLOBAL_STATUS
>> >> bitmask cannot be trusted. I'd like to verify this.
>> >>
>> >> Has the problem appear only on Nehalem or also on
>> >> Westmere?
>> >
>> > I was able to duplicate on
>> >
>> > Intel(R) Core(TM) i5 CPU Â Â Â Â 650 Â@ 3.20GHz
>> > Intel(R) Xeon(R) CPU Â Â Â Â Â X5560 Â@ 2.80GHz
>> >
>> I managed to reproduce on core i7 860 (without patch4).
>> Looking at the code again, I am dubious you ever execute
>> the retry goto. If the PMU is disabled and you've just
>> cleared the OVF_STAT, then I don't see where the new
>> overflows would come from. But that's a separate problem.
>
> I agree with you, but a printk before the goto proved otherwise! :-)
>

Well, I was wrong about my status validation. I had my test wrong.
I am seeing discrepancies between OVF_STATUS (status)
and manual inspection of each counter (status2) compared to
active_mask (act):

[  489.338564] CPU2 status=0x6 status2=0x4 act=0x7
[  501.484235] CPU2 status=0x6 status2=0x4 act=0x7
[  541.686414] CPU2 status=0x6 status2=0x4 act=0x7
[  561.819575] CPU2 status=0x6 status2=0x4 act=0x7

Note that I have the NMI watchdog on, and running perf record
with 2 events only.

OVF_STATUS reports more than what has actually occurred.
I think that is fine because in the loop there is a check for against
active_mask. Nevertheless, I think there is a problem right there.

I am trying to narrow down when it happens. It seems to occur
only when I start a new perf record session. And possibly when
perf runs on a CPU that is not monitored:

taskset -c 4   perf record -e cycles -e instructions -F 2000 -a  -C 2
-o /dev/null -- sleep 600

> And a printk of the status bit that triggered the goto happened to be the
> same one that we initially cleared. ÂLike I said when I initially posted
> the patch, I am not sure why it works but it does do something to stem the
> NMI.

Yes, I suspect the clearing does not work well. Let's try to do it
multiple times
in a loop.

>
> There is probably a deeper problem here, I was just trying to get the
> external/unknown nmis working again.
>
I suspect that if we were not to use OVF_STATUS, then we would not need
the back-to-back nmi logic. NMi cannot be nested. So just looking at
the counters
should be reliable (excl. PEBS for now).


> Cheers,
> Don
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/