[PATCH] perf stat: balance opening and closing of events

[Mark Rutland]

In create_perf_stat_counter(), we open events on a potentially per-cpu
or per-thread basis, depending on if we have a target CPU list. The
number of FDs we allocate depends on whether we're in per-thread or
per-cpu mode.

Subsequently, we close events using perf_evlist__close(), which assumes
that if an evsel has a cpu map, we've opened events per-cpu-per-thread
for the evsel, and each of the FDs needs to be closed. If we'd actually
opened events purely per-thread, this can result in erroneously
treating unallocated memory as FDs, which we attempt to close, then set
to -1.

This has been observed to corrupt the datastructures used by the libc
memory allocator, resulting in segfaults when closing events.

To fix this, this patch adds a close_perf_stat_counter() function that
matches the logic in create_perf_stat_counter(), ensuring that we open
and close the same number of counters. Helpers are added for closing an
evsel per-thread or per-cpu.

Signed-off-by: Mark Rutland <mark.rutland@xxxxxxx>
Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
Cc: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
Cc: Ingo Molnar <mingo@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: linux-kernel@xxxxxxxxxxxxxxx
---
 tools/perf/builtin-stat.c | 12 +++++++++++-
 tools/perf/util/evsel.c   | 12 ++++++++++++
 tools/perf/util/evsel.h   |  4 ++++
 3 files changed, 27 insertions(+), 1 deletion(-)

diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index a02f2e9..80e2cf4 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -237,6 +237,14 @@ static int create_perf_stat_counter(struct perf_evsel *evsel)
 	return perf_evsel__open_per_thread(evsel, evsel_list->threads);
 }
 
+static void close_perf_stat_counter(struct perf_evsel *evsel)
+{
+	if (target__has_cpu(&target))
+		perf_evsel__close_per_cpu(evsel, perf_evsel__cpus(evsel));
+
+	perf_evsel__close_per_thread(evsel, evsel_list->threads);
+}
+
 /*
  * Does the counter have nsecs as a unit?
  */
@@ -686,7 +694,9 @@ static int __run_perf_stat(int argc, const char **argv)
 	 * group leaders.
 	 */
 	read_counters();
-	perf_evlist__close(evsel_list);
+
+	evlist__for_each_entry(evsel_list, counter)
+		close_perf_stat_counter(counter);
 
 	return WEXITSTATUS(status);
 }
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index 04e536a..18da665 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -1655,6 +1655,18 @@ void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
 	perf_evsel__free_fd(evsel);
 }
 
+void perf_evsel__close_per_cpu(struct perf_evsel *evsel,
+			       struct cpu_map *cpus)
+{
+	perf_evsel__close(evsel, cpus->nr, 1);
+}
+
+void perf_evsel__close_per_thread(struct perf_evsel *evsel,
+			       struct thread_map *threads)
+{
+	perf_evsel__close(evsel, 1, threads->nr);
+}
+
 static struct {
 	struct cpu_map map;
 	int cpus[1];
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index 06ef6f2..02bea43 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -252,6 +252,10 @@ int perf_evsel__open_per_thread(struct perf_evsel *evsel,
 				struct thread_map *threads);
 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 		     struct thread_map *threads);
+void perf_evsel__close_per_cpu(struct perf_evsel *evsel,
+			       struct cpu_map *cpus);
+void perf_evsel__close_per_thread(struct perf_evsel *evsel,
+				  struct thread_map *threads);
 void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads);
 
 struct perf_sample;
-- 
1.9.1