Re: [RFC] perf/core: Add an ioctl to get a number of lost samples

[Namhyung Kim <namhyung@kernel.org>]

On Wed, Aug 11, 2021 at 12:57 PM Jiri Olsa <jolsa@redhat.com> wrote:
>
> On Wed, Aug 11, 2021 at 12:33:38PM -0700, Stephane Eranian wrote:
> > On Wed, Aug 11, 2021 at 8:04 AM Jiri Olsa <jolsa@redhat.com> wrote:
> > >
> > > On Tue, Aug 10, 2021 at 11:21:35PM -0700, Namhyung Kim wrote:
> > > > Sometimes we want to know an accurate number of samples even if it's
> > > > lost.  Currenlty PERF_RECORD_LOST is generated for a ring-buffer which
> > > > might be shared with other events.  So it's hard to know per-event
> > > > lost count.
> > > >
> > > > Add event->lost_samples field and PERF_EVENT_IOC_LOST_SAMPLES to
> > > > retrieve it from userspace.
> > > >
> > > > Signed-off-by: Namhyung Kim <namhyung@kernel.org>
> > > > ---
> > > >  include/linux/perf_event.h      | 2 ++
> > > >  include/uapi/linux/perf_event.h | 1 +
> > > >  kernel/events/core.c            | 9 +++++++++
> > > >  kernel/events/ring_buffer.c     | 5 ++++-
> > > >  4 files changed, 16 insertions(+), 1 deletion(-)
> > > >
> > > > diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
> > > > index f5a6a2f069ed..44d72079c77a 100644
> > > > --- a/include/linux/perf_event.h
> > > > +++ b/include/linux/perf_event.h
> > > > @@ -756,6 +756,8 @@ struct perf_event {
> > > >       struct pid_namespace            *ns;
> > > >       u64                             id;
> > > >
> > > > +     atomic_t                        lost_samples;
> > > > +
> > > >       u64                             (*clock)(void);
> > > >       perf_overflow_handler_t         overflow_handler;
> > > >       void                            *overflow_handler_context;
> > > > diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
> > > > index bf8143505c49..24397799127d 100644
> > > > --- a/include/uapi/linux/perf_event.h
> > > > +++ b/include/uapi/linux/perf_event.h
> > > > @@ -505,6 +505,7 @@ struct perf_event_query_bpf {
> > > >  #define PERF_EVENT_IOC_PAUSE_OUTPUT          _IOW('$', 9, __u32)
> > > >  #define PERF_EVENT_IOC_QUERY_BPF             _IOWR('$', 10, struct perf_event_query_bpf *)
> > > >  #define PERF_EVENT_IOC_MODIFY_ATTRIBUTES     _IOW('$', 11, struct perf_event_attr *)
> > > > +#define PERF_EVENT_IOC_LOST_SAMPLES          _IOR('$', 12, __u64 *)
> > >
> > > would it be better to use the read syscall for that?
> > >   https://lore.kernel.org/lkml/20210622153918.688500-5-jolsa@kernel.org/
> > >
> > > that patchset ended up on me not having a way to reproduce the
> > > issue you guys wanted the fix for ;-) the lost count is there
> > > as well
> > >
> > Does the read format approach succeed even when the event is in error state?
>
> nope..
>
>         /*
>          * Return end-of-file for a read on an event that is in
>          * error state (i.e. because it was pinned but it couldn't be
>          * scheduled on to the CPU at some point).
>          */
>         if (event->state == PERF_EVENT_STATE_ERROR)
>                 return 0;
>

By the way, it'd be nice if the kernel would provide a way for
better error reporting.  There are many cases return -EINVAL
and it's hard to know what's the problem exactly.

Thanks,
Namhyung