perf stat: Ignore error thread when enabling system-wide --per-thread
authorJin Yao <yao.jin@linux.intel.com>
Tue, 16 Jan 2018 15:43:08 +0000 (23:43 +0800)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Tue, 27 Feb 2018 14:29:21 +0000 (11:29 -0300)
If we execute 'perf stat --per-thread' with non-root account (even set
kernel.perf_event_paranoid = -1 yet), it reports the error:

  jinyao@skl:~$ perf stat --per-thread
  Error:
  You may not have permission to collect system-wide stats.

  Consider tweaking /proc/sys/kernel/perf_event_paranoid,
  which controls use of the performance events system by
  unprivileged users (without CAP_SYS_ADMIN).

  The current value is 2:

    -1: Allow use of (almost) all events by all users
        Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK
  >= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN
        Disallow raw tracepoint access by users without CAP_SYS_ADMIN
  >= 1: Disallow CPU event access by users without CAP_SYS_ADMIN
  >= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN

  To make this setting permanent, edit /etc/sysctl.conf too, e.g.:

          kernel.perf_event_paranoid = -1

Perhaps the ptrace rule doesn't allow to trace some processes. But anyway
the global --per-thread mode had better ignore such errors and continue
working on other threads.

This patch will record the index of error thread in perf_evsel__open()
and remove this thread before retrying.

For example (run with non-root, kernel.perf_event_paranoid isn't set):

  jinyao@skl:~$ perf stat --per-thread
  ^C
   Performance counter stats for 'system wide':

         vmstat-3458    6.171984   cpu-clock:u (msec) #  0.000 CPUs utilized
           perf-3670    0.515599   cpu-clock:u (msec) #  0.000 CPUs utilized
         vmstat-3458   1,163,643   cycles:u           #  0.189 GHz
           perf-3670      40,881   cycles:u           #  0.079 GHz
         vmstat-3458   1,410,238   instructions:u     #  1.21  insn per cycle
           perf-3670       3,536   instructions:u     #  0.09  insn per cycle
         vmstat-3458     288,937   branches:u         # 46.814 M/sec
           perf-3670         936   branches:u         #  1.815 M/sec
         vmstat-3458      15,195   branch-misses:u    #  5.26% of all branches
           perf-3670          76   branch-misses:u    #  8.12% of all branches

        12.651675247 seconds time elapsed

Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/1516117388-10120-1-git-send-email-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/builtin-stat.c
tools/perf/util/evsel.c
tools/perf/util/thread_map.c
tools/perf/util/thread_map.h

index fadcff52cd0944e7501b761a855e18e8771e2965..6214d2b220b29db937e8f2faf8082d243cb1afe2 100644 (file)
@@ -637,7 +637,19 @@ try_again:
                                 if (verbose > 0)
                                         ui__warning("%s\n", msg);
                                 goto try_again;
-                        }
+                       } else if (target__has_per_thread(&target) &&
+                                  evsel_list->threads &&
+                                  evsel_list->threads->err_thread != -1) {
+                               /*
+                                * For global --per-thread case, skip current
+                                * error thread.
+                                */
+                               if (!thread_map__remove(evsel_list->threads,
+                                                       evsel_list->threads->err_thread)) {
+                                       evsel_list->threads->err_thread = -1;
+                                       goto try_again;
+                               }
+                       }
 
                        perf_evsel__open_strerror(counter, &target,
                                                  errno, msg, sizeof(msg));
index ef351688b79798a1c46b95094fb078c6daee14aa..b56e1c2ddaee2e320a7ce437434557f901a8dc47 100644 (file)
@@ -1915,6 +1915,9 @@ try_fallback:
                goto fallback_missing_features;
        }
 out_close:
+       if (err)
+               threads->err_thread = thread;
+
        do {
                while (--thread >= 0) {
                        close(FD(evsel, cpu, thread));
index 729dad8f412d62ac700a8801ffc79c67c8025bc6..5d467d8ae9abf9bee19b5cb02b01d1849af5f32d 100644 (file)
@@ -32,6 +32,7 @@ static void thread_map__reset(struct thread_map *map, int start, int nr)
        size_t size = (nr - start) * sizeof(map->map[0]);
 
        memset(&map->map[start], 0, size);
+       map->err_thread = -1;
 }
 
 static struct thread_map *thread_map__realloc(struct thread_map *map, int nr)
index 5ec91cfd18696b9f1efccdf824bfc4a9ec32b122..2f689c90a8c6d59c5e52a7d2627c9c793df55834 100644 (file)
@@ -14,6 +14,7 @@ struct thread_map_data {
 struct thread_map {
        refcount_t refcnt;
        int nr;
+       int err_thread;
        struct thread_map_data map[];
 };