Re: perf trace + BPF

From: Arnaldo Carvalho de Melo
Date: Fri Apr 01 2016 - 09:50:48 EST


Em Fri, Apr 01, 2016 at 09:29:52PM +0800, Wangnan (F) escreveu:
>
>
> On 2016/4/1 2:08, Arnaldo Carvalho de Melo wrote:
> >Hi Wang,
> >
> > Trying to get back at working with 'perf trace' + BPF and I'm
> >noticing that the perf_event_attr->sample_type for the BPF events are
> >different than the ones for the raw_syscalls:sys_{enter,exit} or to
> >other tracepoint events we may ask 'perf trace' to set up, for instance,
> >for:
> >
> >perf trace -vv -T --ev sched:sched_switch --ev bpf-output/no-inherit,name=evt/ --ev /home/acme/bpf/test_bpf_trace.c/map:channel.event=evt/ usleep 10
> >
> >We get these sample types (full perf_event_attr dump at the end of this
> >message):
> >
> >sched:sched_switch:
> >sample_type IP|TID|TIME|CPU|PERIOD|RAW|IDENTIFIER
> >
> >[root@jouet ~]# perf probe -l
> > perf_bpf_probe:func_begin (on SyS_nanosleep@linux/kernel/time/hrtimer.c)
> > perf_bpf_probe:func_end (on SyS_nanosleep%return@linux/kernel/time/hrtimer.c)
> >
> > sample_type IP|TID|RAW|IDENTIFIER
> >
> >raw_syscalls:sys_{enter,exit}:
> >
> > sample_type IP|TID|TIME|CPU|PERIOD|RAW|IDENTIFIER
> >
> >I guess this is an interaction with that "no-inherit" part, probably we need to
> >set PERF_SAMPLE_TIME by some other means...
>
> Please see [1]. Not related to no-inherit, just because the bpf-output event
> is not a tracepoint.

Right, that was a dead end of mine when investigating this, thanks for
fixing this up!

> And glad to see you restart working on my patches again!

Yeah, hopefully this time we'll get most of what is outstanding merged.

I want to at some time experiment with doing things 'perf trace' needs
with BPF, behind the scenes :-)

> Thank you.
>
> [1] http://lkml.kernel.org/g/1459517202-42320-1-git-send-email-wangnan0@xxxxxxxxxx
>
>
>
> >- Arnaldo
> >
>