Re: [PATCH v3] perf/core: Set event shadow time for inactive events too

From: Song Liu <songliubraving@fb.com>
To: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@kernel.org>,
	Arnaldo Carvalho de Melo <acme@kernel.org>,
	Jiri Olsa <jolsa@redhat.com>, Mark Rutland <mark.rutland@arm.com>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	LKML <linux-kernel@vger.kernel.org>,
	Stephane Eranian <eranian@google.com>,
	Andi Kleen <ak@linux.intel.com>,
	"Ian Rogers" <irogers@google.com>
Subject: Re: [PATCH v3] perf/core: Set event shadow time for inactive events too
Date: Mon, 6 Dec 2021 23:11:34 +0000	[thread overview]
Message-ID: <A629A4F4-1BA4-4DCF-B4F0-F0E80B55738A@fb.com> (raw)
In-Reply-To: <20211205224843.1503081-1-namhyung@kernel.org>


> On Dec 5, 2021, at 2:48 PM, Namhyung Kim <namhyung@kernel.org> wrote:
> 
> While commit f79256532682 ("perf/core: fix userpage->time_enabled of
> inactive events") fixed this problem for user rdpmc usage, bperf (perf
> stat with BPF) still has the same problem that accessing inactive perf
> events from BPF using bpf_perf_event_read_value().
> 
> You can reproduce this problem easily.  As this is about a small
> window with multiplexing, we need a large number of events and short
> duration like below:
> 
>  # perf stat -a -v --bpf-counters -e instructions,branches,branch-misses \
>    -e cache-references,cache-misses,bus-cycles,ref-cycles,cycles sleep 0.1
> 
>  Control descriptor is not initialized
>  instructions: 19616489 431324015 360374366
>  branches: 3685346 417640114 344175443
>  branch-misses: 75714 404089360 336145421
>  cache-references: 438667 390474289 327444074
>  cache-misses: 49279 349333164 272835067
>  bus-cycles: 631887 283423953 165164214
>  ref-cycles: 2578771111104847872 18446744069443110306 182116355
>  cycles: 1785221016051271680 18446744071682768912 115821694
> 
>   Performance counter stats for 'system wide':
> 
>          19,616,489      instructions              #    0.00  insn per cycle           ( 83.55%)
>           3,685,346      branches                                                      ( 82.41%)
>              75,714      branch-misses             #    2.05% of all branches          ( 83.19%)
>             438,667      cache-references                                              ( 83.86%)
>              49,279      cache-misses              #   11.234 % of all cache refs      ( 78.10%)
>             631,887      bus-cycles                                                    ( 58.27%)
>  2,578,771,111,104,847,872      ref-cycles                                                     (0.00%)
>  1,785,221,016,051,271,680      cycles                                                         (0.00%)
> 
>       0.010824702 seconds time elapsed
> 
> As you can see, it shows invalid values for the last two events.
> The -v option shows that the enabled time is way bigger than the
> running time.  So it scaled the counter values using the ratio
> between the two and resulted in that.  This problem can get worse
> if users want no-aggregation or cgroup aggregation with a small
> interval.
> 
> Actually 18446744069443110306 is 0xffffffff01b345a2 so it seems to
> have a negative enabled time.  In fact, bperf keeps values returned by
> bpf_perf_event_read_value() which calls perf_event_read_local(), and
> accumulates delta between two calls.  When event->shadow_ctx_time is
> not set, it'd return invalid enabled time which is bigger than normal.
> Later, the shadow time is set and the function starts to return a
> valid time.  At the moment, the recent value is smaller than before so
> the delta in the bperf can be negative.
> 
> I think we need to set the shadow time even the events are inactive so
> that BPF programs (or other potential users) can see valid time values
> anytime.
> 
> Cc: Song Liu <songliubraving@fb.com>
> Signed-off-by: Namhyung Kim <namhyung@kernel.org>

Acked-by: Song Liu <song@kernel.org>