perf stat: Issue a HW watchdog disable hint
authorBorislav Petkov <bp@suse.de>
Tue, 7 Feb 2017 00:40:05 +0000 (01:40 +0100)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Fri, 3 Mar 2017 22:07:13 +0000 (19:07 -0300)
When using perf stat on an AMD F15h system with the default hw events
attributes, some of the events don't get counted:

 Performance counter stats for 'sleep 1':

          0.749208      task-clock (msec)         #    0.001 CPUs utilized
                 1      context-switches          #    0.001 M/sec
                 0      cpu-migrations            #    0.000 K/sec
                54      page-faults               #    0.072 M/sec
         1,122,815      cycles                    #    1.499 GHz
           286,740      stalled-cycles-frontend   #   25.54% frontend cycles idle
     <not counted>      stalled-cycles-backend                                        (0.00%)
     ^^^^^^^^^^^^
     <not counted>      instructions                                                  (0.00%)
     ^^^^^^^^^^^^
     <not counted>      branches                                                      (0.00%)
     <not counted>      branch-misses                                                 (0.00%)

       1.001550070 seconds time elapsed

The reason is that we have the HW watchdog consuming one PMU counter and
when perf tries to schedule 6 events on 6 counters and some of those
counters are constrained to only a specific subset of PMCs by the
hardware, the event scheduling fails.

So issue a hint to disable the HW watchdog around a perf stat session.

Committer note:

Testing it...

  # perf stat -d usleep 1

   Performance counter stats for 'usleep 1':

          1.180203      task-clock (msec)         #    0.490 CPUs utilized
                 1      context-switches          #    0.847 K/sec
                 0      cpu-migrations            #    0.000 K/sec
                54      page-faults               #    0.046 M/sec
           184,754      cycles                    #    0.157 GHz
           714,553      instructions              #    3.87  insn per cycle
           154,661      branches                  #  131.046 M/sec
             7,247      branch-misses             #    4.69% of all branches
           219,984      L1-dcache-loads           #  186.395 M/sec
            17,600      L1-dcache-load-misses     #    8.00% of all L1-dcache hits    (90.16%)
     <not counted>      LLC-loads                                                     (0.00%)
     <not counted>      LLC-load-misses                                               (0.00%)

       0.002406823 seconds time elapsed

  Some events weren't counted. Try disabling the NMI watchdog:
echo 0 > /proc/sys/kernel/nmi_watchdog
perf stat ...
echo 1 > /proc/sys/kernel/nmi_watchdog
  #

Signed-off-by: Borislav Petkov <bp@suse.de>
Acked-by: Ingo Molnar <mingo@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Robert Richter <rric@kernel.org>
Cc: Vince Weaver <vince@deater.net>
Link: http://lkml.kernel.org/r/20170211183218.ijnvb5f7ciyuunx4@pd.tnic
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/builtin-stat.c

index 13b54999ad79ecd4f765d557bf57764486f1fac8..f4f555a67e9b682f29d5fed0468079f729f7872d 100644 (file)
@@ -146,6 +146,7 @@ static aggr_get_id_t                aggr_get_id;
 static bool                    append_file;
 static const char              *output_name;
 static int                     output_fd;
+static int                     print_free_counters_hint;
 
 struct perf_stat {
        bool                     record;
@@ -1109,6 +1110,9 @@ static void printout(int id, int nr, struct perf_evsel *counter, double uval,
                        counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
                        csv_sep);
 
+               if (counter->supported)
+                       print_free_counters_hint = 1;
+
                fprintf(stat_config.output, "%-*s%s",
                        csv_output ? 0 : unit_width,
                        counter->unit, csv_sep);
@@ -1477,6 +1481,13 @@ static void print_footer(void)
                                avg_stats(&walltime_nsecs_stats));
        }
        fprintf(output, "\n\n");
+
+       if (print_free_counters_hint)
+               fprintf(output,
+"Some events weren't counted. Try disabling the NMI watchdog:\n"
+"      echo 0 > /proc/sys/kernel/nmi_watchdog\n"
+"      perf stat ...\n"
+"      echo 1 > /proc/sys/kernel/nmi_watchdog\n");
 }
 
 static void print_counters(struct timespec *ts, int argc, const char **argv)