Re: [PATCH 2/2] perf report: report module before querying isactivation in dwfl unwind

From	Jan Kratochvil <jan.kratochvil@redhat.com>
Newsgroups	linux.kernel
Subject	Re: [PATCH 2/2] perf report: report module before querying isactivation in dwfl unwind
Date	2017-06-15 21:30 +0200
Message-ID	<tSICS-6z6-17@gated-at.bofh.it> (permalink)
References	<tNVU5-2oY-3@gated-at.bofh.it> <tNVU5-2oY-11@gated-at.bofh.it> <tNWGu-2XT-15@gated-at.bofh.it> <tOXK9-1L4-9@gated-at.bofh.it>
Organization	linux.* mail to news gateway

Show all headers | View raw

On Mon, 05 Jun 2017 12:47:20 +0200, Jiri Olsa wrote:
> On Fri, Jun 02, 2017 at 12:25:08PM -0300, Arnaldo Carvalho de Melo wrote:
> > Em Fri, Jun 02, 2017 at 04:37:53PM +0200, Milian Wolff escreveu:
> > > The PC returned by dwfl_frame_pc may map into a not-yet-reported
> > > module. We have to report it before we continue unwinding. But when
> > > we query for the isactivation flag in dwfl_frame_pc, libdw will
> > > actually do one more unwinding step internally which can then break
> > > and lead to missed frames or broken stacks.
> > > 
> > > With libunwind we get e.g.:
> 
> sounds like a libdw design issue.. Jan is there a specific way
> to handle the case that Milian described?

I agree with the patch.

Just be aware for performance then the __report_module() call in entry() is no
longer useful as the '--pc' adjustment should never cross a module boundary.

Although for perf unwinding performance there is still a wide gap there
(such as caching the loaded modules for multiple backtraces.)


Thanks,
Jan

Back to linux.kernel | Previous | Next | Find similar | Unroll thread

Thread

Re: [PATCH 2/2] perf report: report module before querying  isactivation in dwfl unwind Jan Kratochvil <jan.kratochvil@redhat.com> - 2017-06-15 21:30 +0200

csiph-web