Messages in this thread Patch in this message | | | From | Namhyung Kim <> | Subject | [PATCH 3/3] perf report: Don't be bothered with locking when adding hist entries | Date | Mon, 13 May 2013 15:16:34 +0900 |
| |
From: Namhyung Kim <namhyung.kim@lge.com>
The perf report is single-threaded, so no need to grab a lock. Although the fast path of pthread_mutex_[un]lock() is very fast, there's ~3% gain by eliminating it when we have huge sample data.
$ perf record -a -F 100000 -o perf.data.bench -- perf bench sched all $ perf record -e cycles:upp -o perf.data.before -- \ > perf report -i perf.data.bench --stdio > /dev/null ... apply this patch ... $ perf record -e cycles:upp -o perf.data.after -- \ > perf report -i perf.data.bench --stdio > /dev/null $ perf diff perf.data.{before,after} | grep pthread +0.02% libpthread-2.15.so [.] _pthread_cleanup_push_defer +0.02% libpthread-2.15.so [.] _pthread_cleanup_pop_restore 0.05% -0.05% perf [.] pthread_mutex_unlock@plt 0.05% -0.05% perf [.] pthread_mutex_lock@plt 1.01% -1.01% libpthread-2.15.so [.] pthread_mutex_lock 1.68% -1.68% libpthread-2.15.so [.] __pthread_mutex_unlock_usercnt 0.05% -0.05% libpthread-2.15.so [.] pthread_mutex_unlock
Signed-off-by: Namhyung Kim <namhyung@kernel.org> --- tools/perf/builtin-report.c | 4 ---- 1 file changed, 4 deletions(-)
diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c index 63febd24e912..0f0cf2472d9d 100644 --- a/tools/perf/builtin-report.c +++ b/tools/perf/builtin-report.c @@ -312,8 +312,6 @@ static int process_sample_event(struct perf_tool *tool, if (rep->cpu_list && !test_bit(sample->cpu, rep->cpu_bitmap)) return 0; - pthread_mutex_lock(&evsel->hists.lock); - if (sort__mode == SORT_MODE__BRANCH) { ret = perf_report__add_branch_hist_entry(tool, &al, sample, evsel, machine); @@ -332,8 +330,6 @@ static int process_sample_event(struct perf_tool *tool, if (ret < 0) pr_debug("problem incrementing symbol period, skipping event\n"); } - pthread_mutex_unlock(&evsel->hists.lock); - return ret; } -- 1.7.11.7
| |