perf inject: Fix file corruption due to event deletion
"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.
Removed the now unused 'evsel' variable, update the comment about the
evsel removal not being performed anymore, and apply the patch manually
as it failed with this warning:
warning: Patch sent with format=flowed; space at the end of lines might be lost.
Testing it with:
$ perf bench internals inject-build-id
# Running 'internals/inject-build-id' benchmark:
Average build-id injection took: 8.543 msec (+- 0.130 msec)
Average time per event: 0.838 usec (+- 0.013 usec)
Average memory usage: 12717 KB (+- 9 KB)
Average build-id-all injection took: 5.710 msec (+- 0.058 msec)
Average time per event: 0.560 usec (+- 0.006 usec)
Average memory usage: 12079 KB (+- 7 KB)
Signed-off-by: Al Grant <firstname.lastname@example.org>
Acked-by: Adrian Hunter <email@example.com>
Acked-by: Namhyung Kim <firstname.lastname@example.org>
Cc: Alexander Shishkin <email@example.com>
Cc: Jiri Olsa <firstname.lastname@example.org>
Cc: Mark Rutland <email@example.com>
Cc: Peter Zijlstra <firstname.lastname@example.org>
Signed-off-by: Arnaldo Carvalho de Melo <email@example.com>
1 file changed