Re: [PATCH 2/4] tools lib perf: Handle read format in perf_evsel__read()
From: Arnaldo Carvalho de Melo
Date: Tue Aug 16 2022 - 16:03:27 EST
Em Tue, Aug 16, 2022 at 03:19:20PM +0200, Jiri Olsa escreveu:
> On Mon, Aug 15, 2022 at 12:01:04PM -0700, Namhyung Kim wrote:
> > The perf_counts_values should be increased to read the new lost data.
> > Also adjust values after read according the read format.
> >
> > This supports PERF_FORMAT_GROUP which has a different data format but
> > it's only available for leader events. Currently it doesn't have an API
> > to read sibling (member) events in the group. But users may read the
> > sibling event directly.
> >
> > Also reading from mmap would be disabled when the read format has ID or
> > LOST bit as it's not exposed via mmap.
> >
> > Signed-off-by: Namhyung Kim <namhyung@xxxxxxxxxx>
> > ---
> > tools/lib/perf/evsel.c | 72 +++++++++++++++++++++++++++++
> > tools/lib/perf/include/perf/event.h | 3 +-
> > tools/lib/perf/include/perf/evsel.h | 4 +-
> > 3 files changed, 77 insertions(+), 2 deletions(-)
> >
> > diff --git a/tools/lib/perf/evsel.c b/tools/lib/perf/evsel.c
> > index 952f3520d5c2..fc23670231cb 100644
> > --- a/tools/lib/perf/evsel.c
> > +++ b/tools/lib/perf/evsel.c
> > @@ -305,6 +305,9 @@ int perf_evsel__read_size(struct perf_evsel *evsel)
> > if (read_format & PERF_FORMAT_ID)
> > entry += sizeof(u64);
> >
> > + if (read_format & PERF_FORMAT_LOST)
> > + entry += sizeof(u64);
> > +
> > if (read_format & PERF_FORMAT_GROUP) {
> > nr = evsel->nr_members;
> > size += sizeof(u64);
> > @@ -314,24 +317,93 @@ int perf_evsel__read_size(struct perf_evsel *evsel)
> > return size;
> > }
> >
> > +/* This only reads values for the leader */
> > +static int perf_evsel__read_group(struct perf_evsel *evsel, int cpu_map_idx,
> > + int thread, struct perf_counts_values *count)
> > +{
> > + size_t size = perf_evsel__read_size(evsel);
> > + int *fd = FD(evsel, cpu_map_idx, thread);
> > + u64 read_format = evsel->attr.read_format;
> > + u64 *data;
> > + int idx = 1;
> > +
> > + if (fd == NULL || *fd < 0)
> > + return -EINVAL;
> > +
> > + data = calloc(1, size);
> > + if (data == NULL)
> > + return -ENOMEM;
> > +
> > + if (readn(*fd, data, size) <= 0) {
> > + free(data);
> > + return -errno;
> > + }
>
> could you please put in here some comment that this is intentionaly
> reading only the leader or better yet rename the function? I was lost
> before I got to read the changelog ;-)
>
> > +
> > + if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
> > + count->ena = data[idx++];
> > + if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
> > + count->run = data[idx++];
> > +
> > + /* value is always available */
> > + count->val = data[idx++];
> > + if (read_format & PERF_FORMAT_ID)
> > + count->id = data[idx++];
> > + if (read_format & PERF_FORMAT_LOST)
> > + count->lost = data[idx++];
> > +
> > + free(data);
> > + return 0;
> > +}
> > +
> > +/*
> > + * The perf read format is very flexible. It needs to set the proper
> > + * values according to the read format.
> > + */
> > +static void perf_evsel__adjust_values(struct perf_evsel *evsel,
> > + struct perf_counts_values *count)
> > +{
> > + u64 read_format = evsel->attr.read_format;
> > +
> > + if (!(read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)) {
> > + memmove(&count->values[2], &count->values[1], 24);
> > + count->ena = 0;
> > + }
> > +
> > + if (!(read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)) {
> > + memmove(&count->values[3], &count->values[2], 16);
> > + count->run = 0;
> > + }
> > +
> > + if (!(read_format & PERF_FORMAT_ID)) {
> > + memmove(&count->values[4], &count->values[3], 8);
> > + count->id = 0;
> > + }
> > +}
>
>
> could we do this the same way we read group counters.. like make read
> into local buffer and initialize perf_counts_values values based on
> format, something like:
>
> readn(fd, data ...
>
> if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
> count->ena = data[idx++];
> if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
> count->run = data[idx++];
>
> /* value is always available */
> count->val = data[idx++];
> if (read_format & PERF_FORMAT_ID)
> count->id = data[idx++];
> if (read_format & PERF_FORMAT_LOST)
> count->lost = data[idx++];
I see this now, so I'll wait for Namhyung to react to your comments,
while keeping it in my local tree just to have it build tested.
- Arnaldo
>
> and perhaps we should cancel that perf_counts_values's union and keep
> only 'val/ena/run...' fields?
>
> jirka
--
- Arnaldo