Hi Leo,

On Fri, Jun 11, 2021 at 5:27 PM Leo Yan [off-list ref] wrote:

On Fri, Jun 11, 2021 at 01:31:41AM -0700, Denis Nikitin wrote:

quoted

Hi Suzuki,

On Thu, Jun 10, 2021 at 9:04 AM Suzuki K Poulose [off-list ref] wrote:

quoted

[...]

quoted

Are you able to confirm if the patch 3 alone fixes the above issue ?
I am not convinced that Patch 1 is necessary.

Yes. "perf cs-etm: Remove callback cs_etm_find_snapshot()" alone fixes
the issue.

Based on current testing result, we should give high priority for
patches 2 and 3.

The patch 1 is controversial for how to handle the trace data kept
in multiple AUX buffers; essentially it's up to how we understand the
snapshot definition.  I confirmed Intel-PT and CoreSight have the same
behaviour for capturing trace data from multiple AUX ring buffers when
snapshot occurs.

I'd like to leave patch 1/3 out, and resend it if we get conclusion.
At the meantime, @Denis, if you have observed any profiling result
(or profiling quality) difference caused by patch 1, the feedback would
be very valuable.

I evaluated AutoFDO profiles with benchmarks but I was only focused
on the system-wide mode. And as I understood patch 1 fixes the issue
in non system-wide mode.
Currently I'm OoO so I won't be able to do further evaluation.

- Denis

Thanks a lot for Denis' testing and insight review from Suzuki!

Leo