[-next,v2,0/2] perf stat: Support inherit events for bperf

Message

Tengda Wu Sept. 5, 2024, 11:59 a.m. UTC

Hi,

bperf (perf-stat --bpf-counter) has not supported inherit events
during fork() since it was first introduced.

This patch series tries to add this support by:
 1) adding two new bpf programs to monitor task lifecycle;
 2) recording new tasks in the filter map dynamically;
 3) reusing `accum_key` of parent task for new tasks.

Thanks,
Tengda


Changelog:
---------
v2: (Address comments from Namhyung)
 * Remove the unused init_filter_entries in follower bpf, declare
   a global filter_entry_count in bpf_counter instead
 * Attach on_newtask and on_exittask progs only if the filter type
   is either PID or TGID

v1: https://lore.kernel.org/all/[e-mail]/


Tengda Wu (2):
  perf stat: Support inherit events during fork() for bperf
  perf test: Use sqrtloop workload to test bperf event

 tools/perf/tests/shell/stat_bpf_counters.sh   |  2 +-
 tools/perf/util/bpf_counter.c                 | 32 +++++++--
 tools/perf/util/bpf_skel/bperf_follower.bpf.c | 70 +++++++++++++++++--
 tools/perf/util/bpf_skel/bperf_u.h            |  5 ++
 4 files changed, 95 insertions(+), 14 deletions(-)

Comments

Tengda Wu Sept. 14, 2024, 6:58 a.m. UTC | #1

Hi Song and Numhyung,

Gentle ping.

This patch set provides event inheritance support for bperf during fork(). Currently, based on numhyung's feedback, the v2 version has been modified. Could you please take a look?

Any comment or suggestion is appreciated.

Thanks!
Tengda

On 2024/9/5 19:59, Tengda Wu wrote:
> Hi,
> 
> bperf (perf-stat --bpf-counter) has not supported inherit events
> during fork() since it was first introduced.
> 
> This patch series tries to add this support by:
>  1) adding two new bpf programs to monitor task lifecycle;
>  2) recording new tasks in the filter map dynamically;
>  3) reusing `accum_key` of parent task for new tasks.
> 
> Thanks,
> Tengda
> 
> 
> Changelog:
> ---------
> v2: (Address comments from Namhyung)
>  * Remove the unused init_filter_entries in follower bpf, declare
>    a global filter_entry_count in bpf_counter instead
>  * Attach on_newtask and on_exittask progs only if the filter type
>    is either PID or TGID
> 
> v1: https://lore.kernel.org/all/[e-mail]/
> 
> 
> Tengda Wu (2):
>   perf stat: Support inherit events during fork() for bperf
>   perf test: Use sqrtloop workload to test bperf event
> 
>  tools/perf/tests/shell/stat_bpf_counters.sh   |  2 +-
>  tools/perf/util/bpf_counter.c                 | 32 +++++++--
>  tools/perf/util/bpf_skel/bperf_follower.bpf.c | 70 +++++++++++++++++--
>  tools/perf/util/bpf_skel/bperf_u.h            |  5 ++
>  4 files changed, 95 insertions(+), 14 deletions(-)
>