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 <[email protected]> Cc: Alexander Shishkin <[email protected]> Cc: Arnaldo Carvalho de Melo <[email protected]> Cc: Ingo Molnar <[email protected]> Cc: Peter Zijlstra <[email protected]> Cc: [email protected] --- 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
