Re: [PATCH] perf inject corrupts file by deleting event

[Namhyung Kim <namhyung@kernel.org>]

Hello,

On Sat, Nov 14, 2020 at 5:38 AM Al Grant <al.grant@foss.arm.com> wrote:
>
> "perf inject" can create corrupt files when synthesizing sample
> events from AUX data. This happens when in the input file, the
> first event (for the AUX data) has a different sample_type from
> the second event (generally dummy). Specifically, they differ in
> the bits that indicate the standard fields appended to perf
> records in the mmap buffer. "perf inject" deletes the first event
> and moves up the second event to first position. The problem is
> with the synthetic PERF_RECORD_MMAP (etc.) events created by
> "perf record". Since these are synthetic versions of events which
> are normally produced by the kernel, they have to have the
> standard fields appended as described by sample_type. "perf record"
> fills these in with zeroes, including the IDENTIFIER field;
> perf readers interpret records with zero IDENTIFIER using the
> descriptor for the first event in the file. Since "perf inject"
> changes the first event, these synthetic records are then
> processed with the wrong value of sample_type, and the perf
> reader reads bad data, reports on incorrect length records etc.
>
> Mismatching sample_types are seen with "perf record -e cs_etm//",
> where the AUX event has TID|TIME|CPU|IDENTIFIER and the dummy
> event has TID|TIME|IDENTIFIER. Perhaps they could be the same,
> but it isn't normally a problem if they aren't - perf has
> no problems reading the file. The sample_types have to agree on
> the position of IDENTIFIER, because that's how perf finds the
> right event descriptor in the first place, but they don't normally
> have to agree on other fields, and perf doesn't check that they do.
> The problem is specific to the way "perf inject" reorganizes the
> events and the way synthetic MMAP events are recorded with a zero
> identifier. A simple solution is to stop "perf inject" deleting
> the tracing event.
>
> Signed-off-by: Al Grant <al.grant@arm.com>
> Acked-by: Adrian Hunter <adrian.hunter@intel.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> Cc: Mark Rutland <mark.rutland@arm.com>
> Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
> Cc: Jiri Olsa <jolsa@redhat.com>
> Cc: Namhyung Kim <namhyung@kernel.org>

While I'm ok with this change, I think we can put the dummy events
to the front of the evlist (during record) so that we can make sure that
tracking records would refer to them in order to parse the data.

And I also think that we should omit the dummy events from the
perf report output.

Thanks,
Namhyung


> ---
>   tools/perf/builtin-inject.c | 7 -------
>   1 file changed, 7 deletions(-)
>
> diff --git a/tools/perf/builtin-inject.c b/tools/perf/builtin-inject.c
> index 452a75fe68e5..f4968ebf5f3a 100644
> --- a/tools/perf/builtin-inject.c
> +++ b/tools/perf/builtin-inject.c
> @@ -791,13 +791,6 @@ static int __cmd_inject(struct perf_inject *inject)
>                              inject->itrace_synth_opts.add_last_branch)
>                                  perf_header__set_feat(&session->header,
>                                                        HEADER_BRANCH_STACK);
> -                       evsel = perf_evlist__id2evsel_strict(session->evlist,
> -                                                            inject->aux_id);
> -                       if (evsel) {
> -                               pr_debug("Deleting %s\n", evsel__name(evsel));
> -                               evlist__remove(session->evlist, evsel);
> -                               evsel__delete(evsel);
> -                       }
>                  }
>                  session->header.data_offset = output_data_offset;
>                  session->header.data_size = inject->bytes_written;