Re: [RFC PATCH] perf/stat: Add --disable-hwdt

[Borislav Petkov]

On Sat, Feb 11, 2017 at 06:59:10PM +0100, Ingo Molnar wrote:
> So I checked the perf-list manpage and it didn't tell me much about how to disable 
> the NMI watchdog.

Oh, it is buried there:

"
EVENT GROUPS
------------

...


Globally pinned events can limit the number of counters available for
other groups. On x86 systems, the NMI watchdog pins a counter by default.
The nmi watchdog can be disabled as root with

	echo 0 > /proc/sys/kernel/nmi_watchdog"

> How about a more proactive hint, something like:
> 
> 	To disable the NMI watchdog permanently, do:

Why permanently? We want it to run and be disabled around the
measurement only. Anyway, here's something more to the point:

---
From: Borislav Petkov <bp@xxxxxxx>
Date: Tue, 7 Feb 2017 01:40:05 +0100
Subject: [PATCH -v2] perf stat: Issue a HW watchdog disable hint

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.

Signed-off-by: Borislav Petkov <bp@xxxxxxx>
---
 tools/perf/builtin-stat.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index a02f2e965628..a2763243a03d 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -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)
-- 
2.11.0

-- 
Regards/Gruss,
    Boris.

Good mailing practices for 400: avoid top-posting and trim the reply.