perf/core: Fix scheduling regression of pinned groups
authorAlexander Shishkin <alexander.shishkin@linux.intel.com>
Tue, 18 Jul 2017 11:08:34 +0000 (14:08 +0300)
committerIngo Molnar <mingo@kernel.org>
Thu, 20 Jul 2017 07:43:02 +0000 (09:43 +0200)
commit3bda69c1c3993a2bddbae01397d12bfef6054011
treef68537445d68484381849c1f105e4bb69a51f9ea
parentdc853e26f73e903e0c87e24f2695b5dcf33b3bc1
perf/core: Fix scheduling regression of pinned groups

Vince Weaver reported:

> I was tracking down some regressions in my perf_event_test testsuite.
> Some of the tests broke in the 4.11-rc1 timeframe.
>
> I've bisected one of them, this report is about
> tests/overflow/simul_oneshot_group_overflow
> This test creates an event group containing two sampling events, set
> to overflow to a signal handler (which disables and then refreshes the
> event).
>
> On a good kernel you get the following:
>  Event perf::instructions with period 1000000
>  Event perf::instructions with period 2000000
>  fd 3 overflows: 946 (perf::instructions/1000000)
>  fd 4 overflows: 473 (perf::instructions/2000000)
>  Ending counts:
>  Count 0: 946379875
>  Count 1: 946365218
>
> With the broken kernels you get:
>  Event perf::instructions with period 1000000
>  Event perf::instructions with period 2000000
>  fd 3 overflows: 938 (perf::instructions/1000000)
>  fd 4 overflows: 318 (perf::instructions/2000000)
>  Ending counts:
>  Count 0: 946373080
>  Count 1: 653373058

The root cause of the bug is that the following commit:

  487f05e18a ("perf/core: Optimize event rescheduling on active contexts")

erronously assumed that event's 'pinned' setting determines whether the
event belongs to a pinned group or not, but in fact, it's the group
leader's pinned state that matters.

This was discovered by Vince in the test case described above, where two instruction
counters are grouped, the group leader is pinned, but the other event is not;
in the regressed case the counters were off by 33% (the difference between events'
periods), but should be the same within the error margin.

Fix the problem by looking at the group leader's pinning.

Reported-by: Vince Weaver <vincent.weaver@maine.edu>
Tested-by: Vince Weaver <vincent.weaver@maine.edu>
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Fixes: 487f05e18a ("perf/core: Optimize event rescheduling on active contexts")
Link: http://lkml.kernel.org/r/87lgnmvw7h.fsf@ashishki-desk.ger.corp.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/events/core.c