Re: [PATCH 3/4] perf: teach perf inject to merge sched_stat_* and sched_switch events

[Arun Sharma <asharma@fb.com>]

On 11/2/11 3:59 PM, Andrew Vagin wrote:

> +		event_sw = (union perf_event *)&ent->header;

This causes problems with -Wstrict-aliasing=3.

I noticed that the suggested cmdline is:

perf script record sched-stat -- -e sched:sched_stat_sleep

I'd much prefer something along the lines of:

perf script record sched-stat -e <events> -- sleep 1

Also, the perf inject command is a bit unforgiving when the user makes 
some mistakes (eg: didn't specify -e sched:sched_stat_sleep). The 
failure mode is hard to understand (perf report -i perf.data.d fails 
with "no samples").

Doing it directly:

# perf record -ag -e sched:sched_switch --filter "prev_state == 1 || 
prev_state == 2" -e 
sched:sched_process_exit,sched:sched_stat_sleep,sched:sched_stat_iowait 
-- sleep 1
# perf inject -v -s -i perf.data -o perf.data.d

results in a whole bunch of:

Could not find sched_switch for pid 16188
Could not find sched_switch for pid 16212
Could not find sched_switch for pid 1500
Could not find sched_switch for pid 16195
Could not find sched_switch for pid 16179

I think this is the same issue you referenced in "Known issues".

Although I'm able to get some profiling info out of perf.data.d, the 
process I really wanted to trace got zero samples.

  -Arun