lkml.org 
[lkml]   [2016]   [Dec]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH 2/3] perf/x86/pebs: add workaround for broken OVFL status on HSW
On Thu, Dec 15, 2016 at 9:10 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> On Thu, Dec 15, 2016 at 08:59:56AM -0800, Stephane Eranian wrote:
>> On Thu, Dec 15, 2016 at 12:42 AM, Peter Zijlstra <peterz@infradead.org> wrote:
>> > On Wed, Dec 14, 2016 at 11:26:49PM -0800, Stephane Eranian wrote:
>> >> On Wed, Dec 14, 2016 at 9:55 AM, Peter Zijlstra <peterz@infradead.org> wrote:
>> >> >
>> >> > Just spotted this again, ping?
>> >> >
>> >> Ok, on what processor running what command, so I can try and reproduce?
>> >
>> > For me its more of a correctness issue, i've not actually spotted a
>> > problem as such.
>> >
>> > But every time I read this code it makes me wonder.
>> >
>> > Supposing that the hardware sets the CTRL overflow flags but hasn't
>> > generated the PEBS record yet (or not enough records to reach the PEBS
>> > buffer threshold) we still don't want to process these events as if they
>> > were !PEBS.
>> >
>> I am suspicious about the case where you have multiple PEBS events and
>> they do not quite fire at the same time but close enough that you may have
>> PEBS in-flight by the time you enter handle_irq.
>>
>> Last night I ran a simple test on SKL using tip.git:
>>
>> $ perf record --e
>> cpu/event=0xd0,umask=0x81/upp,cpu/event=0xc0,umask=1/upp,cpu/event=0xd0,umask=0x81/upp
>> multichase; perf report -D | fgrep SAMPLE | grep -v 'IP, 0x4' | grep
>> -v events
>>
>> Basically, looking for samples missing the EXACT tag, i.e., samples
>> processed a regular event when I only have PEBS events. Over 8h, I got
>> about 3 or 4 such samples. So there is still a condition where we see
>> the overflow as regular and not PEBS. So we need to examine that code
>> again looking for possible race with PEBS in flight and not having the
>> PEBS overflow bits yet.
>
> Isn't that exactly the case I was talking about? and would be avoided by
> the proposed patch?
>
>
> So semantically the counter overflows and then arms PEBS to record a
> record on the next event once its armed (and this can be multiple events
> after the overflow, since arming takes a while too).
>
>
> Now, if the chip manages to raise the regular overflow bit during that
> time, you get exactly what is described.
>
> meaning we should unconditionally clear the pebs_enabled.

If we unconditionally clear the pebs_enabled counter from status, then
we guarantee
these counters will never be processed as regular overflowed counters.
I am okay with this.
I am testing it right now.

But the part I still don't get is why is the status bitmask showing
some pebs counters set when
the counter are explicitly program with their PMI bit cleared. I need
to check whether somehow
the PMI bit gets set.

\
 
 \ /
  Last update: 2016-12-16 09:39    [W:0.095 / U:0.404 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site