Re: [PATCH] perf report: Show zero counters as well in 'perf report --stat'

From: Ingo Molnar
Date: Fri Mar 09 2018 - 03:38:27 EST



* Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:

> > STAT events: 0
> > STAT_ROUND events: 0
> > EVENT_UPDATE events: 0
> > TIME_CONV events: 1
> > FEATURE events: 0
> >
> > It's pretty clear at a glance that LOST_SAMPLES is present but zero.
>
> Your wording confused me a bit, that "is present" part, because there
> are some PERF_RECORD_ events that will only be "present" if we actually
> explicitely ask them to be, by setting flags in perf_event_attr, like:
>
> mmap : 1, /* include mmap data */
> comm : 1, /* include comm data */
> task : 1, /* trace fork/exit */
> mmap_data : 1, /* non-exec mmap data */
> sample_id_all : 1, /* sample_type all events */
> mmap2 : 1, /* include mmap with inode data */
> comm_exec : 1, /* flag comm events that are due to an exec */
> context_switch : 1, /* context switch data */
> namespaces : 1, /* include namespaces data */
>
> So, for PERF_RECORD_MMAP2 events, one can say that it is "present but
> zero" if we have an event with perf_event_attr.mmap2 = 1 and no
> PERF_RECORD_MMAP2 events recorded.
>
> But for PERF_RECORD_LOST, that will always be present, if we lose
> records.

Yeah, the triple ambiguity between:

- did the event disappear due to 'we did not lose any records',
- or did it disappear because it's somehow a conditional stat field,
- or did it disappear because I'm blind and/or mis-remembering the field name.

... is what was causing trouble to me personally, to the level that I had to go
and look into the tooling code to make sure I'm interpreting it correctly.

> So perhaps to make all this clean we can add your patch, and on top of
> it another that shows the perf_record_attr flag for the ones that are
> not on all the time, something like:
>
> fomalhaut:~> perf report --stat
>
> Aggregated stats:
> TOTAL events: 495984
> MMAP events: 85 attr.mmap: 1
> LOST events: 0
> COMM events: 3389 attr.comm: 1, attr.comm_exec: 0
> EXIT events: 1605 attr.task: 1
> THROTTLE events: 2
> UNTHROTTLE events: 2
> FORK events: 3377 attr.task: 1
> READ events: 0
> SAMPLE events: 472629
> MMAP2 events: 14753 attr.mmap2: 1
> AUX events: 0 have to look at what enables this, etc
> ITRACE_START events: 0 ditto
> LOST_SAMPLES events: 0
> SWITCH events: 0 attr.context_switch: 0
> SWITCH_CPU_WIDE events: 0 attr.context_switch: 0
> NAMESPACES events: 0 attr.namespaces: 0
> ATTR events: 0
> EVENT_TYPE events: 0
> TRACING_DATA events: 0
> BUILD_ID events: 0 this implies disabling in 'perf record' command line
> FINISHED_ROUND events: 139
> ID_INDEX events: 0
> AUXTRACE_INFO events: 0
> AUXTRACE events: 0
> AUXTRACE_ERROR events: 0
> THREAD_MAP events: 1
> CPU_MAP events: 1
> STAT_CONFIG events: 0
> STAT events: 0
> STAT_ROUND events: 0
> EVENT_UPDATE events: 0
> TIME_CONV events: 1 have to check
> FEATURE events: 0

Yeah, I'm fine with this too if it's easy enough to implement - although I suspect
in most cases the knowledge that a stat field not present must be due to an
environment/setup dependency (and not a runtime/workload dependency) is enough.

It's the uncertainty of 'stat line can disappear because it was zero' that was my
main beef :-)

Thanks,

Ingo