Re: [RFC 00/10] perf: Add build id parsing fault detection/fix

From: Jiri Olsa
Date: Sun Jun 27 2021 - 13:25:45 EST


On Wed, Jun 23, 2021 at 12:48:30PM -0700, Namhyung Kim wrote:
> Hi Jiri,
>
> Thanks for your work!
>
> On Tue, Jun 22, 2021 at 2:19 PM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
> >
> > On Tue, Jun 22, 2021 at 03:14:04PM -0300, Arnaldo Carvalho de Melo wrote:
> > > Em Tue, Jun 22, 2021 at 10:47:54AM -0700, Ian Rogers escreveu:
> > > > On Tue, Jun 22, 2021 at 10:39 AM Arnaldo Carvalho de Melo
> > > > <acme@xxxxxxxxxx> wrote:
> > > > >
> > > > > Em Tue, Jun 22, 2021 at 05:39:08PM +0200, Jiri Olsa escreveu:
> > > > > > hi,
> > > > > > this *RFC* patchset adds support to detect faults during
> > > > > > mmap2's build id parsing and a way to fix such maps in
> > > > > > generated perf.data.
> > > > > >
> > > > > > It adds support to record build id faults count for session
> > > > > > and store it in perf.data and perf inject support to find
> > > > > > these maps and reads build ids for them in user space.
> > > > >
> > > > > > It's probably best explained by the workflow:
> > > > > >
> > > > > > Record data with --buildid-mmap option:
> > > > > >
> > > > > > # perf record --buildid-mmap ...
> > > > > > ...
> > > > > > [ perf record: Woken up 1 times to write data ]
> > > > > > [ perf record: Failed to parse 4 build ids]
> > > > > > [ perf record: Captured and wrote 0.008 MB perf.data ]
> > > > > >
> > > > > > Check if there's any build id fault reported:
> > > > > >
> > > > > > # perf report --header-only
> > > > > > ...
> > > > > > # build id mmap stats: FAULTS 4, LOST 0, NOT FIXED
> > > > > >
> > > > > > There is, check the stats:
> > > > > >
> > > > > > # perf report --stat
> > > > > >
> > > > > > Aggregated stats:
> > > > > > TOTAL events: 104
> > > > > > ....
> > > > > > BUILD_ID fails: 4 (14.3%)
> > > > > >
> > > > > > Yep, let's fix it:
> > > > > >
> > > > > > # perf inject --buildid-mmap2 -i perf.data -o perf-fixed.data
> > > > >
> > > > > Can we make it possible to automate this with --fixup-buildids or a
> > > > > perfconfig 'record' knob?
> > > > >
> > > > > This would entail requesting that build-ids that _fail_ be sent to the
> > > > > side-band thread we have in 'perf record', this way we wouldn't have to
> > > > > traverse the whole perf.data file, be it with 'perf-record' at the end
> > > > > of a session with faulty build ids, or in a similar fashion using 'perf
> > > > > inject' as you suggest.
> > > > >
> > > > > I even think that we can have all these modes and let the user to decide
> > > > > how important is this for them and how convenient they want the whole
> > > > > process to be.
> >
> > right, that might be good to decide first.. because as I said,
> > I never hit faulted build id, so it probably needs the special
> > setup you guys are using.. could you try on your setup and check
> > how many faulted build ids you see?
>
> Did you check data mmaps? It might be easy to get faults
> from data files and we don't know if it's an ELF or not
> before reading the ELF header in the first page.

well, AFAICS the mmap event is sent right after the elf file
is loaded, so it does not have a chance to be swapped out

>
> I'm not sure if we can limit it to exec mappings, there might
> be data-only DSOs and we may want to symbolize them too.

hum, I haven't checked the data-only DSO, which we'd load
manually, not the loader.. will check

thanks,
jirka