Re: possible regression in perf

[Jiri Olsa <olsajiri@gmail.com>]

On Mon, Dec 11, 2023 at 07:00:51PM +0100, trinity pointard wrote:
> Hi,
> 
> I think I found a regression in perf related to the control fd.
> At its introduction (bee328cb71eb0b38ab128d7c475209d973a13f92), it
> looks like enabling or disabling counters would print the current
> counters. However when `enable/disable <counter>` was introduced
> (991ae4eb36911fe9a99bef1c22b9578ceec3896a), it changed what
> `evlist__ctlfd_process()` would return on a successful enable/disable,
> which made the call to `process_interval()` in `process_evlist()`
> essentially dead code.
> 
> I was in the process of patching perf so that `snapshot` would print
> current counters (in which case I would send enable, sleep a bit, and
> send disable+snapshot, to do my measurements). But given what I saw, I
> think just sending enable and disable was enough at one point.
> 
> Is this a known issue? Would a patch which makes
> `evlist__ctlfd_enable()` return 1 on success be an acceptable fix?
> It's somewhat more verbose than what I wanted initially (printing both
> on enable and disable, printing on enable isn't useful to me), but
> that's reasonably easy to post-process (or enable could be made to not
> print current counters).
> 
> Regards,
> trinity-1686a/trinity Pointard
> 
> P.S.: this is my first time posting to LKML, please tell me if I did
> anything wrong :-)

hi,
thanks for the report, could you please send proposed change as patch
and steps to reproduce the issue?

jirka