Re: perf 6.9-1 (archlinux) crashes during recording of cycles + raw_syscalls

[Namhyung Kim <namhyung@kernel.org>]

On Tue, Jun 04, 2024 at 04:02:08PM -0300, Arnaldo Carvalho de Melo wrote:
> On Tue, Jun 04, 2024 at 11:48:09AM -0700, Ian Rogers wrote:
> > On Tue, Jun 4, 2024 at 7:12 AM Arnaldo Carvalho de Melo <acme@kernel.org> wrote:
> > > Can you please try with the attached and perhaps provide your Tested-by?
> 
> > > From ab355e2c6b4cf641a9fff7af38059cf69ac712d5 Mon Sep 17 00:00:00 2001
> > > From: Arnaldo Carvalho de Melo <acme@redhat.com>
> > > Date: Tue, 4 Jun 2024 11:00:22 -0300
> > > Subject: [PATCH 1/1] Revert "perf record: Reduce memory for recording
> > >  PERF_RECORD_LOST_SAMPLES event"
> 
> > > This reverts commit 7d1405c71df21f6c394b8a885aa8a133f749fa22.
>  
> > I think we should try to fight back reverts when possible. Reverts are
> > removing something somebody poured time and attention into. When a
> 
> While in the development phase, yeah, but when we find a regression and
> the revert makes it go away, that is the way to go.
> 
> The person who poured time on the development gets notified and can
> decide if/when to try again.
> 
> Millian had to pour time to figure out why something stopped working,
> was kind enough to provide the output from multiple tools to help in
> fixing the problem and I had to do the bisect to figure out when the
> problem happened and to check if reverting it we would have the tool
> working again.
> 
> If we try to fix this for v6.10 we may end up adding yet another bug, so
> the safe thing to do at this point is to do the revert.
> 
> We can try improving this once again for v6.11.

I think I found a couple of problems with this issue. :(

 1. perf_session__set_id_hdr_size() uses the first evsel in the session
    But I think it should pick the tracking event.  I guess we assume
    all events have the same set of sample_type wrt the sample_id_all
    but I'm not sure if it's correct.

 2. With --call-graph dwarf, it seems to set unrelated sample type bits
    in the attr like ADDR and DATA_SRC.

 3. For tracepoint events, evsel__newtp_idx() sets a couple of sample
    type regardless of the configuration.  This includes RAW, TIME and
    CPU.  This one changes the format of the id headers.

 4. PERF_RECORD_LOST_SAMPLES is for the sampling event, so it should
    use the event's sample_type.  But the event parsing looks up the
    event using evlist->is_pos which is set for the first event.

 5. I think we can remove some sample type (i.e. TID and CPU) from the
    tracking event in most cases.  ID(ENTIFIER) will be used for LOST_
    SAMPLES and TIME is needed anyway.  TID is might be used for SWITCH
    but others already contain necessary information in the type.  I
    wish we could add id field to PERF_RECORD_LOST_SAMPLES and tid/pid
    to PERF_RECORD_SWITCH.

Thanks,
Namhyung

> 
> > regression has occurred then I think we should add the regression case
> > as a test.
> 
> Sure, I thought about that as well, will try and have one shell test
> with that, referring to this case, for v6.11.
> 
> - Arnaldo