Re: [PATCH v3 5/9] perf build-id: Mark DSO in sample callchains

[Namhyung Kim <namhyung@kernel.org>]

On Wed, May 28, 2025 at 01:54:41PM -0700, Ian Rogers wrote:
> On Wed, May 28, 2025 at 1:41 PM Namhyung Kim <namhyung@kernel.org> wrote:
> >
> > On Mon, Apr 28, 2025 at 02:34:04PM -0700, Ian Rogers wrote:
> > > Previously only the sample IP's map DSO would be marked hit for the
> > > purposes of populating the build ID cache. Walk the call chain to mark
> > > all IPs and DSOs.
> >
> > I think this is correct, but I'm afraid it'd also increase the processing
> > time.  Do you happen to have any numbers?
> 
> It increases time spent processing the data file but to get a large
> data file I had to run for multiple seconds and I struggled to get the
> performance cost of this to be in the milliseconds (ie a tiny fraction
> of the record time). Ultimately I found the change imperceptible and
> couldn't think of a good command line to make it perceptible.

The worst case would be dwarf unwinding.  Maybe we can skip the
processing if it takes too long..

> 
> If the time is spent populating ~/.debug because more DSOs are marked
> then this is fixing a bug and isn't a problem with the patch.

Right, it's a good thing.

> 
> My personal opinion is that it is somewhat surprising `perf record` is
> post-processing the perf.data file at all, and -B and -N would be my
> expected defaults - just as --buildid-mmap implies --no-buildid (-B).

Otherwise nobody will run perf buildid-cache to add the info. :)


> I didn't want to modify existing behaviors in these changes, however,
> in this case I was just trying to make the existing behavior correct,
> similar to fixing the same bug in `perf inject`.

Thanks for your work,
Namhyung