On 10/6/11 2:56 PM, Arun Sharma wrote:

> i.e. select() should be weighted by 4x vs nanosleep() as confirmed via:
>
> perf script | grep comm=foo
>
> foo 15516 [006] 2291.187831: sched_stat_sleep: comm=foo pid=15516
> delay=4054262 [ns]
> foo 15516 [006] 2291.187832: sched_stat_sleep: comm=foo pid=15516
> delay=4054262 [ns]
> foo 15516 [006] 2291.188895: sched_stat_sleep: comm=foo pid=15516
> delay=1053565 [ns]
> foo 15516 [006] 2291.188896: sched_stat_sleep: comm=foo pid=15516
> delay=1053565 [ns]
> foo 15516 [006] 2291.188897: sched_stat_sleep: comm=foo pid=15516
> delay=1053565 [ns]
>
> Andrew, are you already working on user space patches?

Attached is a quick hack (mainly meant as a RFC to get the discussion 
going).

After the patch I get the expected result:

      0.16%              foo  [unknown]          [k] 0
                        |
                        --- schedule
                           |
                           |--79.26%-- schedule_hrtimeout_range_clock
                           |          schedule_hrtimeout_range
                           |          poll_schedule_timeout
                           |          do_select
                           |          core_sys_select
                           |          sys_select
                           |          system_call_fastpath
                           |
                           |--16.87%-- do_nanosleep
                           |          hrtimer_nanosleep
                           |          sys_nanosleep
                           |          system_call_fastpath
                           |
                            --3.88%-- pipe_wait
                                      pipe_read
                                      do_sync_read
                                      vfs_read
                                      sys_read
                                      system_call_fastpath

We still have the issue that per-process profiling doesn't work.

  -Arun