Re: [PATCH bpf-next 1/5] perf/core: Add PERF_FORMAT_LOST read_format

From: Daniel Xu
Date: Fri Sep 27 2019 - 17:28:49 EST


Hi Jiri,

On Tue Sep 24, 2019 at 10:33 AM Jiri Olsa wrote:
> On Tue, Sep 17, 2019 at 06:30:52AM -0700, Daniel Xu wrote:
>
> SNIP
>
> > + PERF_FORMAT_MAX = 1U << 5, /* non-ABI */
> > };
> >
> > #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
> > diff --git a/kernel/events/core.c b/kernel/events/core.c
> > index 0463c1151bae..ee08d3ed6299 100644
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -1715,6 +1715,9 @@ static void __perf_event_read_size(struct perf_event *event, int nr_siblings)
> > if (event->attr.read_format & PERF_FORMAT_ID)
> > entry += sizeof(u64);
> >
> > + if (event->attr.read_format & PERF_FORMAT_LOST)
> > + entry += sizeof(u64);
> > +
> > if (event->attr.read_format & PERF_FORMAT_GROUP) {
> > nr += nr_siblings;
> > size += sizeof(u64);
> > @@ -4734,6 +4737,24 @@ u64 perf_event_read_value(struct perf_event *event, u64 *enabled, u64 *running)
> > }
> > EXPORT_SYMBOL_GPL(perf_event_read_value);
> >
> > +static struct pmu perf_kprobe;
> > +static u64 perf_event_lost(struct perf_event *event)
> > +{
> > + struct ring_buffer *rb;
> > + u64 lost = 0;
> > +
> > + rcu_read_lock();
> > + rb = rcu_dereference(event->rb);
> > + if (likely(!!rb))
> > + lost += local_read(&rb->lost);
> > + rcu_read_unlock();
> > +
> > + if (event->attr.type == perf_kprobe.type)
> > + lost += perf_kprobe_missed(event);
>
> not sure what was the peterz's suggestion, but here you are mixing
> ring buffer's lost count with kprobes missed count, seems wrong

To be honest, I'm not 100% sure what the correct semantics here should
be. I thought it might be less misleading if we included ring buffer
related misses as well.

Regardless, I am ok with either.

> maybe we could add PERF_FORMAT_KPROBE_MISSED

I think the feedback from the last patchset was that we want to keep
the misses unified.

Peter, do you have any thoughts?

Thanks,
Daniel