perf stat: clarify unsupported events from uncounted events
David Ahern [Mon, 30 May 2011 14:55:59 +0000 (08:55 -0600)]
perf stat continues running even if the event list contains counters
that are not supported. The resulting output then contains <not counted>
for those events which gets confusing as to which events are supported,
but not counted and which are not supported.

Before:

perf stat -ddd -- sleep 1

      Performance counter stats for 'sleep 1':

          0.571283 task-clock                #    0.001 CPUs utilized
                 1 context-switches          #    0.002 M/sec
                 0 CPU-migrations            #    0.000 M/sec
               157 page-faults               #    0.275 M/sec
         1,037,707 cycles                    #    1.816 GHz
     <not counted> stalled-cycles-frontend
     <not counted> stalled-cycles-backend
           654,499 instructions              #    0.63  insns per cycle
           136,129 branches                  #  238.286 M/sec
     <not counted> branch-misses
     <not counted> L1-dcache-loads
     <not counted> L1-dcache-load-misses
     <not counted> LLC-loads
     <not counted> LLC-load-misses
     <not counted> L1-icache-loads
     <not counted> L1-icache-load-misses
     <not counted> dTLB-loads
     <not counted> dTLB-load-misses
     <not counted> iTLB-loads
     <not counted> iTLB-load-misses
     <not counted> L1-dcache-prefetches
     <not counted> L1-dcache-prefetch-misses

       1.001004836 seconds time elapsed

After:

perf stat -ddd -- sleep 1

 Performance counter stats for 'sleep 1':

          1.350326 task-clock                #    0.001 CPUs utilized
                 2 context-switches          #    0.001 M/sec
                 0 CPU-migrations            #    0.000 M/sec
               157 page-faults               #    0.116 M/sec
            11,986 cycles                    #    0.009 GHz
   <not supported> stalled-cycles-frontend
   <not supported> stalled-cycles-backend
           496,986 instructions              #   41.46  insns per cycle
           138,065 branches                  #  102.246 M/sec
             7,245 branch-misses             #    5.25% of all branches
     <not counted> L1-dcache-loads
     <not counted> L1-dcache-load-misses
     <not counted> LLC-loads
     <not counted> LLC-load-misses
     <not counted> L1-icache-loads
     <not counted> L1-icache-load-misses
     <not counted> dTLB-loads
     <not counted> dTLB-load-misses
     <not counted> iTLB-loads
     <not counted> iTLB-load-misses
     <not counted> L1-dcache-prefetches
   <not supported> L1-dcache-prefetch-misses

       1.002397333 seconds time elapsed

v1->v2:
changed supported type from int to bool

v2->v3
fixed vertical alignment of new struct element

Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/1306767359-13221-1-git-send-email-dsahern@gmail.com
Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>

tools/perf/builtin-stat.c
tools/perf/util/evsel.h

index a9f0671..784ed6d 100644 (file)
@@ -61,6 +61,8 @@
 #include <locale.h>
 
 #define DEFAULT_SEPARATOR      " "
+#define CNTR_NOT_SUPPORTED     "<not supported>"
+#define CNTR_NOT_COUNTED       "<not counted>"
 
 static struct perf_event_attr default_attrs[] = {
 
@@ -448,6 +450,7 @@ static int run_perf_stat(int argc __used, const char **argv)
                                if (verbose)
                                        ui__warning("%s event is not supported by the kernel.\n",
                                                    event_name(counter));
+                               counter->supported = false;
                                continue;
                        }
 
@@ -466,6 +469,7 @@ static int run_perf_stat(int argc __used, const char **argv)
                        die("Not all events could be opened.\n");
                        return -1;
                }
+               counter->supported = true;
        }
 
        if (perf_evlist__set_filters(evsel_list)) {
@@ -861,7 +865,7 @@ static void print_counter_aggr(struct perf_evsel *counter)
        if (scaled == -1) {
                fprintf(stderr, "%*s%s%*s",
                        csv_output ? 0 : 18,
-                       "<not counted>",
+                       counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
                        csv_sep,
                        csv_output ? 0 : -24,
                        event_name(counter));
@@ -914,7 +918,8 @@ static void print_counter(struct perf_evsel *counter)
                                csv_output ? 0 : -4,
                                evsel_list->cpus->map[cpu], csv_sep,
                                csv_output ? 0 : 18,
-                               "<not counted>", csv_sep,
+                               counter->supported ? CNTR_NOT_COUNTED : CNTR_NOT_SUPPORTED,
+                               csv_sep,
                                csv_output ? 0 : -24,
                                event_name(counter));
 
index 7e9366e..e9a3155 100644 (file)
@@ -61,6 +61,7 @@ struct perf_evsel {
                off_t           id_offset;
        };
        struct cgroup_sel       *cgrp;
+       bool                    supported;
 };
 
 struct cpu_map;