Re: [PATCH v2 0/6] perf tools: Use SIGUSR2 control data dumpping

From: Wangnan (F)
Date: Fri Apr 15 2016 - 09:41:41 EST




On 2016/4/15 21:35, Arnaldo Carvalho de Melo wrote:
Em Fri, Apr 15, 2016 at 04:14:26PM +0300, Adrian Hunter escreveu:
On 15/04/16 03:19, Arnaldo Carvalho de Melo wrote:
Em Thu, Apr 14, 2016 at 02:21:59PM +0000, Wang Nan escreveu:
This patch set is based on perf/core.

v1 -> v2: Fix a bug which triggers output switching without
'--switch-output' provided.

Patch 5/10 in v1 has a bug: with following cmdline:
# perf record -e intel_bts// --per-thread --snapshot -p 13588
Ok, I moved the trigger stuff to a new file, util/trigger.h and added
missing man page entries, pushed to acme/perf/core, please take a look
if all is well before I push this to Ingo, tomorrow.

Adrian/others, if you can, please take a look.
Looks OK. I gave intel_pt a quick test with snapshot mode which didn't seem
to be broken.

Using --switch-output with intel_pt gives data that does not decode in the
2nd and subsequent files because there is no continuity of switch events.
Also decoding the last file segfaults because the auxtrace index has not
been split between the files.
Probably need to prevent people using --switch-output with intel_pt, but
I'll look at it more next week.
But the workflow of continuous tracing with regular analysis to go on
dumping uninteresting sections of the trace seems interesting? With the
massive volumes of data with Intel PT continuous tracing is deemded
impossible?

Another question: is it possible to make auxtrace (intel_bts, intel_bt
and coresight) writing to overwritable ring buffer, so we can continuous
trace them and silently drop uninteresting data?

Thank you.

- Arnaldo
- Arnaldo
Sending 'SIGUSR2' to perf triggers output switching, which us
undesirable. The reason is in signal handler perf doesn't consider if
user provided --switch-output explicitly.

With this patchset applied:

# /perf record --switch-output -e intel_bts// --per-thread --snapshot -p 13588 &
[1] 15116
# kill -s SIGUSR2 15116
# [ perf record: dump data: Woken up 5 times ]
[ perf record: Dump perf.data.2016041506170000 ]
# kill -s SIGUSR2 15116
# [ perf record: dump data: Woken up 3 times ]
[ perf record: Dump perf.data.2016041506170447 ]
# fg
/perf record --switch-output -e intel_bts// --per-thread --snapshot -p 13588
^C[ perf record: Woken up 3 times to write data ]
[ perf record: Dump perf.data.2016041506170785 ]
[ perf record: Captured and wrote 0.012 MB perf.data.<timestamp> ]
# ls -l ./perf.data*
-rw------- 1 root root 4212888 Apr 15 06:17 ./perf.data.2016041506170000
-rw------- 1 root root 4212704 Apr 15 06:17 ./perf.data.2016041506170447
-rw------- 1 root root 18368 Apr 15 06:17 ./perf.data.2016041506170785

The first 2 perf.data works fine. The final one has no samples.

And:
# rm ./perf.data*
# /perf record -e intel_bts// --per-thread --snapshot -p 13588 &
[1] 15396
# kill -s SIGUSR2 15396
# kill -s SIGUSR2 15396
# fg
/perf record -e intel_bts// --per-thread --snapshot -p 13588
^C[ perf record: Woken up 9 times to write data ]
[ perf record: Captured and wrote 8.012 MB perf.data ]
# ls -l ./perf.data*
-rw------- 1 root root 8417552 Apr 15 06:20 ./perf.data

(In v1, the second test trigger output switching incorrectly)

Patch 1-2/6 in this patchset makes switch output trigger similar to
auxtrace snapshot trigger. Patch 1/6 introduces a 'trigger' class for
them. It would be better to merge patch 1/6 with commit c0bdc1c ("perf
record: Turns auxtrace_snapshot_enable into 3 states"). However, Ingo
has collected that commit, so make it a separated patch.

Signed-off-by: Wang Nan <wangnan0@xxxxxxxxxx>
Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Cc: Adrian Hunter <adrian.hunter@xxxxxxxxx>
Cc: Jiri Olsa <jolsa@xxxxxxxxxx>
Cc: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
Cc: Zefan Li <lizefan@xxxxxxxxxx>
Cc: pi3orama@xxxxxxx
Cc: He Kuang <hekuang@xxxxxxxxxx>

Wang Nan (6):
perf tools: Extrace trigger class from auxtrace_snapshot
perf record: Split output into multiple files via '--switch-output'
perf record: Force enable --timestamp-filename when --switch-output is
provided
perf record: Disable buildid cache options by default in switch output
mode
perf record: Re-synthesize tracking events after output switching
perf record: Generate tracking events for process forked by perf

tools/perf/builtin-record.c | 147 +++++++++++++++++++++++++++++++-------------
tools/perf/util/util.h | 60 ++++++++++++++++++
2 files changed, 163 insertions(+), 44 deletions(-)

--
1.8.3.4