Re: [PATCH 02/10] perf/x86: Improve HT workaround GP counter constraint

[Stephane Eranian]

On Fri, May 22, 2015 at 6:05 AM, Stephane Eranian <eranian@xxxxxxxxxx> wrote:
> On Fri, May 22, 2015 at 5:59 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> On Fri, May 22, 2015 at 05:55:32AM -0700, Stephane Eranian wrote:
>>> On Fri, May 22, 2015 at 5:53 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>> > On Fri, May 22, 2015 at 05:35:14AM -0700, Stephane Eranian wrote:
>>> >> On Fri, May 22, 2015 at 4:28 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>>> >> > On Fri, May 22, 2015 at 01:21:46PM +0200, Peter Zijlstra wrote:
>>> >> >> @@ -821,8 +828,24 @@ int x86_schedule_events(struct cpu_hw_ev
>>> >> >>
>>> >> >>       /* slow path */
>>> >> >>       if (i != n) {
>>> >> >> +             int gpmax = x86_pmu.num_counters / 2;
>>> >> >> +
>>> >> >> +             /*
>>> >> >> +              * Do not allow scheduling of more than half the available
>>> >> >> +              * generic counters.
>>> >> >> +              *
>>> >> >> +              * This helps avoid counter starvation of sibling thread by
>>> >> >> +              * ensuring at most half the counters cannot be in exclusive
>>> >> >> +              * mode. There is no designated counters for the limits. Any
>>> >> >> +              * N/2 counters can be used. This helps with events with
>>> >> >> +              * specific counter constraints.
>>> >> >> +              */
>>> >> >> +             if (is_ht_workaround_enabled() && !cpuc->is_fake &&
>>> >> >> +                 READ_ONCE(cpuc->excl_cntrs->exclusive_present))
>>> >> >> +                     gpmax /= 2;
>>> >> >> +
>>> >> >>               unsched = perf_assign_events(cpuc->event_constraint, n, wmin,
>>> >> >> -                                          wmax, assign);
>>> >> >> +                                          wmax, gpmax, assign);
>>> >> >>       }
>>> >> >>
>>> >> >
>>> >> > Hmm, I divide by 2 twice.. no wonder it doesn't quite work as expected.
>>> >>
>>> >> Yes, that's what I said. Other problem is, with no watchdog, measuring
>>> >> a non-corrupting event is still multiplexing when more than 2 instances
>>> >> are passed:
>>> >>  $ perf stat -a -C 0 -e r20cc,r20cc,r20cc,r20cc -I 1000 sleep 100
>>> >>
>>> >> I get 50% scheduling, only 2 out of 4 events scheduled at any time.
>>> >>
>>> >> There is nothing running on the sibling thread, so it should let me run with 4
>>> >> instances as per your patch.
>>> >
>>> > Ah, I limited it to n/2 if either of the siblings has an exclusive event
>>> > on.
>>> >
>>> But in my test case above, there was no exclusive event at all on either
>>> sibling and yet it limited the non-excl to 2.
>>
>> I bet you tested the exclusive events earlier :-) Its one of the bugs,
>> the n_excl accounting is leaking up. Once !0 it stays !0.
>
> So you're saying intel_put_excl_constraint() does not do the --n_excl?
> Could it be that the flags is not showing PERF_X86_EVENT_EXCL?
> Cannot be related to cpuc_fake because you have it in both the ++
> and -- functions.

One other thing I noticed is that the --n_excl needs to be protected by the
excl_cntrs->lock in put_excl_constraints().
--
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/