602c3c96fa1e132519b728e4d729b31718e4e7a6
[linux-3.10.git] / tools / perf / builtin-stat.c
1 /*
2  * builtin-stat.c
3  *
4  * Builtin stat command: Give a precise performance counters summary
5  * overview about any workload, CPU or specific PID.
6  *
7  * Sample output:
8
9    $ perf stat ~/hackbench 10
10    Time: 0.104
11
12     Performance counter stats for '/home/mingo/hackbench':
13
14        1255.538611  task clock ticks     #      10.143 CPU utilization factor
15              54011  context switches     #       0.043 M/sec
16                385  CPU migrations       #       0.000 M/sec
17              17755  pagefaults           #       0.014 M/sec
18         3808323185  CPU cycles           #    3033.219 M/sec
19         1575111190  instructions         #    1254.530 M/sec
20           17367895  cache references     #      13.833 M/sec
21            7674421  cache misses         #       6.112 M/sec
22
23     Wall-clock time elapsed:   123.786620 msecs
24
25  *
26  * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
27  *
28  * Improvements and fixes by:
29  *
30  *   Arjan van de Ven <arjan@linux.intel.com>
31  *   Yanmin Zhang <yanmin.zhang@intel.com>
32  *   Wu Fengguang <fengguang.wu@intel.com>
33  *   Mike Galbraith <efault@gmx.de>
34  *   Paul Mackerras <paulus@samba.org>
35  *   Jaswinder Singh Rajput <jaswinder@kernel.org>
36  *
37  * Released under the GPL v2. (and only v2, not any later version)
38  */
39
40 #include "perf.h"
41 #include "builtin.h"
42 #include "util/util.h"
43 #include "util/parse-options.h"
44 #include "util/parse-events.h"
45 #include "util/event.h"
46 #include "util/evlist.h"
47 #include "util/evsel.h"
48 #include "util/debug.h"
49 #include "util/color.h"
50 #include "util/header.h"
51 #include "util/cpumap.h"
52 #include "util/thread.h"
53 #include "util/thread_map.h"
54
55 #include <sys/prctl.h>
56 #include <math.h>
57 #include <locale.h>
58
59 #define DEFAULT_SEPARATOR       " "
60
61 static struct perf_event_attr default_attrs[] = {
62
63   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
64   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
65   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
66   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
67
68   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
69   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
70   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
71   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
72   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
73   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
74
75 };
76
77 /*
78  * Detailed stats:
79  */
80 static struct perf_event_attr detailed_attrs[] = {
81
82   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK              },
83   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES        },
84   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS          },
85   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS             },
86
87   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES              },
88   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
89   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND  },
90   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS            },
91   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS     },
92   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES           },
93
94   { .type = PERF_TYPE_HW_CACHE,
95     .config =
96          PERF_COUNT_HW_CACHE_L1D                <<  0  |
97         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
98         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
99
100   { .type = PERF_TYPE_HW_CACHE,
101     .config =
102          PERF_COUNT_HW_CACHE_L1D                <<  0  |
103         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
104         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
105
106   { .type = PERF_TYPE_HW_CACHE,
107     .config =
108          PERF_COUNT_HW_CACHE_LL                 <<  0  |
109         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
110         (PERF_COUNT_HW_CACHE_RESULT_ACCESS      << 16)                          },
111
112   { .type = PERF_TYPE_HW_CACHE,
113     .config =
114          PERF_COUNT_HW_CACHE_LL                 <<  0  |
115         (PERF_COUNT_HW_CACHE_OP_READ            <<  8) |
116         (PERF_COUNT_HW_CACHE_RESULT_MISS        << 16)                          },
117 };
118
119 struct perf_evlist              *evsel_list;
120
121 static bool                     system_wide                     =  false;
122 static int                      run_idx                         =  0;
123
124 static int                      run_count                       =  1;
125 static bool                     no_inherit                      = false;
126 static bool                     scale                           =  true;
127 static bool                     no_aggr                         = false;
128 static pid_t                    target_pid                      = -1;
129 static pid_t                    target_tid                      = -1;
130 static pid_t                    child_pid                       = -1;
131 static bool                     null_run                        =  false;
132 static bool                     detailed_run                    =  false;
133 static bool                     sync_run                        =  false;
134 static bool                     big_num                         =  true;
135 static int                      big_num_opt                     =  -1;
136 static const char               *cpu_list;
137 static const char               *csv_sep                        = NULL;
138 static bool                     csv_output                      = false;
139
140 static volatile int done = 0;
141
142 struct stats
143 {
144         double n, mean, M2;
145 };
146
147 struct perf_stat {
148         struct stats      res_stats[3];
149 };
150
151 static int perf_evsel__alloc_stat_priv(struct perf_evsel *evsel)
152 {
153         evsel->priv = zalloc(sizeof(struct perf_stat));
154         return evsel->priv == NULL ? -ENOMEM : 0;
155 }
156
157 static void perf_evsel__free_stat_priv(struct perf_evsel *evsel)
158 {
159         free(evsel->priv);
160         evsel->priv = NULL;
161 }
162
163 static void update_stats(struct stats *stats, u64 val)
164 {
165         double delta;
166
167         stats->n++;
168         delta = val - stats->mean;
169         stats->mean += delta / stats->n;
170         stats->M2 += delta*(val - stats->mean);
171 }
172
173 static double avg_stats(struct stats *stats)
174 {
175         return stats->mean;
176 }
177
178 /*
179  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
180  *
181  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
182  * s^2 = -------------------------------
183  *                  n - 1
184  *
185  * http://en.wikipedia.org/wiki/Stddev
186  *
187  * The std dev of the mean is related to the std dev by:
188  *
189  *             s
190  * s_mean = -------
191  *          sqrt(n)
192  *
193  */
194 static double stddev_stats(struct stats *stats)
195 {
196         double variance = stats->M2 / (stats->n - 1);
197         double variance_mean = variance / stats->n;
198
199         return sqrt(variance_mean);
200 }
201
202 struct stats                    runtime_nsecs_stats[MAX_NR_CPUS];
203 struct stats                    runtime_cycles_stats[MAX_NR_CPUS];
204 struct stats                    runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
205 struct stats                    runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
206 struct stats                    runtime_branches_stats[MAX_NR_CPUS];
207 struct stats                    runtime_cacherefs_stats[MAX_NR_CPUS];
208 struct stats                    runtime_l1_dcache_stats[MAX_NR_CPUS];
209 struct stats                    walltime_nsecs_stats;
210
211 static int create_perf_stat_counter(struct perf_evsel *evsel)
212 {
213         struct perf_event_attr *attr = &evsel->attr;
214
215         if (scale)
216                 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
217                                     PERF_FORMAT_TOTAL_TIME_RUNNING;
218
219         attr->inherit = !no_inherit;
220
221         if (system_wide)
222                 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false);
223
224         if (target_pid == -1 && target_tid == -1) {
225                 attr->disabled = 1;
226                 attr->enable_on_exec = 1;
227         }
228
229         return perf_evsel__open_per_thread(evsel, evsel_list->threads, false);
230 }
231
232 /*
233  * Does the counter have nsecs as a unit?
234  */
235 static inline int nsec_counter(struct perf_evsel *evsel)
236 {
237         if (perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK) ||
238             perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
239                 return 1;
240
241         return 0;
242 }
243
244 /*
245  * Update various tracking values we maintain to print
246  * more semantic information such as miss/hit ratios,
247  * instruction rates, etc:
248  */
249 static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
250 {
251         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
252                 update_stats(&runtime_nsecs_stats[0], count[0]);
253         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
254                 update_stats(&runtime_cycles_stats[0], count[0]);
255         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
256                 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
257         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
258                 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
259         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
260                 update_stats(&runtime_branches_stats[0], count[0]);
261         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
262                 update_stats(&runtime_cacherefs_stats[0], count[0]);
263         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
264                 update_stats(&runtime_l1_dcache_stats[0], count[0]);
265 }
266
267 /*
268  * Read out the results of a single counter:
269  * aggregate counts across CPUs in system-wide mode
270  */
271 static int read_counter_aggr(struct perf_evsel *counter)
272 {
273         struct perf_stat *ps = counter->priv;
274         u64 *count = counter->counts->aggr.values;
275         int i;
276
277         if (__perf_evsel__read(counter, evsel_list->cpus->nr,
278                                evsel_list->threads->nr, scale) < 0)
279                 return -1;
280
281         for (i = 0; i < 3; i++)
282                 update_stats(&ps->res_stats[i], count[i]);
283
284         if (verbose) {
285                 fprintf(stderr, "%s: %" PRIu64 " %" PRIu64 " %" PRIu64 "\n",
286                         event_name(counter), count[0], count[1], count[2]);
287         }
288
289         /*
290          * Save the full runtime - to allow normalization during printout:
291          */
292         update_shadow_stats(counter, count);
293
294         return 0;
295 }
296
297 /*
298  * Read out the results of a single counter:
299  * do not aggregate counts across CPUs in system-wide mode
300  */
301 static int read_counter(struct perf_evsel *counter)
302 {
303         u64 *count;
304         int cpu;
305
306         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
307                 if (__perf_evsel__read_on_cpu(counter, cpu, 0, scale) < 0)
308                         return -1;
309
310                 count = counter->counts->cpu[cpu].values;
311
312                 update_shadow_stats(counter, count);
313         }
314
315         return 0;
316 }
317
318 static int run_perf_stat(int argc __used, const char **argv)
319 {
320         unsigned long long t0, t1;
321         struct perf_evsel *counter;
322         int status = 0;
323         int child_ready_pipe[2], go_pipe[2];
324         const bool forks = (argc > 0);
325         char buf;
326
327         if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
328                 perror("failed to create pipes");
329                 exit(1);
330         }
331
332         if (forks) {
333                 if ((child_pid = fork()) < 0)
334                         perror("failed to fork");
335
336                 if (!child_pid) {
337                         close(child_ready_pipe[0]);
338                         close(go_pipe[1]);
339                         fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
340
341                         /*
342                          * Do a dummy execvp to get the PLT entry resolved,
343                          * so we avoid the resolver overhead on the real
344                          * execvp call.
345                          */
346                         execvp("", (char **)argv);
347
348                         /*
349                          * Tell the parent we're ready to go
350                          */
351                         close(child_ready_pipe[1]);
352
353                         /*
354                          * Wait until the parent tells us to go.
355                          */
356                         if (read(go_pipe[0], &buf, 1) == -1)
357                                 perror("unable to read pipe");
358
359                         execvp(argv[0], (char **)argv);
360
361                         perror(argv[0]);
362                         exit(-1);
363                 }
364
365                 if (target_tid == -1 && target_pid == -1 && !system_wide)
366                         evsel_list->threads->map[0] = child_pid;
367
368                 /*
369                  * Wait for the child to be ready to exec.
370                  */
371                 close(child_ready_pipe[1]);
372                 close(go_pipe[0]);
373                 if (read(child_ready_pipe[0], &buf, 1) == -1)
374                         perror("unable to read pipe");
375                 close(child_ready_pipe[0]);
376         }
377
378         list_for_each_entry(counter, &evsel_list->entries, node) {
379                 if (create_perf_stat_counter(counter) < 0) {
380                         if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
381                                 if (verbose)
382                                         ui__warning("%s event is not supported by the kernel.\n",
383                                                     event_name(counter));
384                                 continue;
385                         }
386
387                         if (errno == EPERM || errno == EACCES) {
388                                 error("You may not have permission to collect %sstats.\n"
389                                       "\t Consider tweaking"
390                                       " /proc/sys/kernel/perf_event_paranoid or running as root.",
391                                       system_wide ? "system-wide " : "");
392                         } else {
393                                 error("open_counter returned with %d (%s). "
394                                       "/bin/dmesg may provide additional information.\n",
395                                        errno, strerror(errno));
396                         }
397                         if (child_pid != -1)
398                                 kill(child_pid, SIGTERM);
399                         die("Not all events could be opened.\n");
400                         return -1;
401                 }
402         }
403
404         if (perf_evlist__set_filters(evsel_list)) {
405                 error("failed to set filter with %d (%s)\n", errno,
406                         strerror(errno));
407                 return -1;
408         }
409
410         /*
411          * Enable counters and exec the command:
412          */
413         t0 = rdclock();
414
415         if (forks) {
416                 close(go_pipe[1]);
417                 wait(&status);
418         } else {
419                 while(!done) sleep(1);
420         }
421
422         t1 = rdclock();
423
424         update_stats(&walltime_nsecs_stats, t1 - t0);
425
426         if (no_aggr) {
427                 list_for_each_entry(counter, &evsel_list->entries, node) {
428                         read_counter(counter);
429                         perf_evsel__close_fd(counter, evsel_list->cpus->nr, 1);
430                 }
431         } else {
432                 list_for_each_entry(counter, &evsel_list->entries, node) {
433                         read_counter_aggr(counter);
434                         perf_evsel__close_fd(counter, evsel_list->cpus->nr,
435                                              evsel_list->threads->nr);
436                 }
437         }
438
439         return WEXITSTATUS(status);
440 }
441
442 static void print_noise_pct(double total, double avg)
443 {
444         double pct = 0.0;
445
446         if (avg)
447                 pct = 100.0*total/avg;
448
449         fprintf(stderr, "  ( +-%6.2f%% )", pct);
450 }
451
452 static void print_noise(struct perf_evsel *evsel, double avg)
453 {
454         struct perf_stat *ps;
455
456         if (run_count == 1)
457                 return;
458
459         ps = evsel->priv;
460         print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
461 }
462
463 static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
464 {
465         double msecs = avg / 1e6;
466         char cpustr[16] = { '\0', };
467         const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s";
468
469         if (no_aggr)
470                 sprintf(cpustr, "CPU%*d%s",
471                         csv_output ? 0 : -4,
472                         evsel_list->cpus->map[cpu], csv_sep);
473
474         fprintf(stderr, fmt, cpustr, msecs, csv_sep, event_name(evsel));
475
476         if (evsel->cgrp)
477                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
478
479         if (csv_output)
480                 return;
481
482         if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
483                 fprintf(stderr, " # %8.3f CPUs utilized          ", avg / avg_stats(&walltime_nsecs_stats));
484 }
485
486 static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
487 {
488         double total, ratio = 0.0;
489         const char *color;
490
491         total = avg_stats(&runtime_cycles_stats[cpu]);
492
493         if (total)
494                 ratio = avg / total * 100.0;
495
496         color = PERF_COLOR_NORMAL;
497         if (ratio > 50.0)
498                 color = PERF_COLOR_RED;
499         else if (ratio > 30.0)
500                 color = PERF_COLOR_MAGENTA;
501         else if (ratio > 10.0)
502                 color = PERF_COLOR_YELLOW;
503
504         fprintf(stderr, " #  ");
505         color_fprintf(stderr, color, "%6.2f%%", ratio);
506         fprintf(stderr, " frontend cycles idle   ");
507 }
508
509 static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
510 {
511         double total, ratio = 0.0;
512         const char *color;
513
514         total = avg_stats(&runtime_cycles_stats[cpu]);
515
516         if (total)
517                 ratio = avg / total * 100.0;
518
519         color = PERF_COLOR_NORMAL;
520         if (ratio > 75.0)
521                 color = PERF_COLOR_RED;
522         else if (ratio > 50.0)
523                 color = PERF_COLOR_MAGENTA;
524         else if (ratio > 20.0)
525                 color = PERF_COLOR_YELLOW;
526
527         fprintf(stderr, " #  ");
528         color_fprintf(stderr, color, "%6.2f%%", ratio);
529         fprintf(stderr, " backend  cycles idle   ");
530 }
531
532 static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
533 {
534         double total, ratio = 0.0;
535         const char *color;
536
537         total = avg_stats(&runtime_branches_stats[cpu]);
538
539         if (total)
540                 ratio = avg / total * 100.0;
541
542         color = PERF_COLOR_NORMAL;
543         if (ratio > 20.0)
544                 color = PERF_COLOR_RED;
545         else if (ratio > 10.0)
546                 color = PERF_COLOR_MAGENTA;
547         else if (ratio > 5.0)
548                 color = PERF_COLOR_YELLOW;
549
550         fprintf(stderr, " #  ");
551         color_fprintf(stderr, color, "%6.2f%%", ratio);
552         fprintf(stderr, " of all branches        ");
553 }
554
555 static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
556 {
557         double total, ratio = 0.0;
558         const char *color;
559
560         total = avg_stats(&runtime_l1_dcache_stats[cpu]);
561
562         if (total)
563                 ratio = avg / total * 100.0;
564
565         color = PERF_COLOR_NORMAL;
566         if (ratio > 20.0)
567                 color = PERF_COLOR_RED;
568         else if (ratio > 10.0)
569                 color = PERF_COLOR_MAGENTA;
570         else if (ratio > 5.0)
571                 color = PERF_COLOR_YELLOW;
572
573         fprintf(stderr, " #  ");
574         color_fprintf(stderr, color, "%6.2f%%", ratio);
575         fprintf(stderr, " of all L1-dcache hits  ");
576 }
577
578 static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
579 {
580         double total, ratio = 0.0;
581         char cpustr[16] = { '\0', };
582         const char *fmt;
583
584         if (csv_output)
585                 fmt = "%s%.0f%s%s";
586         else if (big_num)
587                 fmt = "%s%'18.0f%s%-24s";
588         else
589                 fmt = "%s%18.0f%s%-24s";
590
591         if (no_aggr)
592                 sprintf(cpustr, "CPU%*d%s",
593                         csv_output ? 0 : -4,
594                         evsel_list->cpus->map[cpu], csv_sep);
595         else
596                 cpu = 0;
597
598         fprintf(stderr, fmt, cpustr, avg, csv_sep, event_name(evsel));
599
600         if (evsel->cgrp)
601                 fprintf(stderr, "%s%s", csv_sep, evsel->cgrp->name);
602
603         if (csv_output)
604                 return;
605
606         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
607                 total = avg_stats(&runtime_cycles_stats[cpu]);
608
609                 if (total)
610                         ratio = avg / total;
611
612                 fprintf(stderr, " #   %5.2f  insns per cycle        ", ratio);
613
614                 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
615                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
616
617                 if (total && avg) {
618                         ratio = total / avg;
619                         fprintf(stderr, "\n                                            #   %5.2f  stalled cycles per insn", ratio);
620                 }
621
622         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
623                         runtime_branches_stats[cpu].n != 0) {
624                 print_branch_misses(cpu, evsel, avg);
625         } else if (
626                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
627                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
628                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
629                                         ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
630                         runtime_l1_dcache_stats[cpu].n != 0) {
631                 print_l1_dcache_misses(cpu, evsel, avg);
632         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
633                         runtime_cacherefs_stats[cpu].n != 0) {
634                 total = avg_stats(&runtime_cacherefs_stats[cpu]);
635
636                 if (total)
637                         ratio = avg * 100 / total;
638
639                 fprintf(stderr, " # %8.3f %% of all cache refs    ", ratio);
640
641         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
642                 print_stalled_cycles_frontend(cpu, evsel, avg);
643         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
644                 print_stalled_cycles_backend(cpu, evsel, avg);
645         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
646                 total = avg_stats(&runtime_nsecs_stats[cpu]);
647
648                 if (total)
649                         ratio = 1.0 * avg / total;
650
651                 fprintf(stderr, " # %8.3f GHz                    ", ratio);
652         } else if (runtime_nsecs_stats[cpu].n != 0) {
653                 total = avg_stats(&runtime_nsecs_stats[cpu]);
654
655                 if (total)
656                         ratio = 1000.0 * avg / total;
657
658                 fprintf(stderr, " # %8.3f M/sec                  ", ratio);
659         } else {
660                 fprintf(stderr, "                                   ");
661         }
662 }
663
664 /*
665  * Print out the results of a single counter:
666  * aggregated counts in system-wide mode
667  */
668 static void print_counter_aggr(struct perf_evsel *counter)
669 {
670         struct perf_stat *ps = counter->priv;
671         double avg = avg_stats(&ps->res_stats[0]);
672         int scaled = counter->counts->scaled;
673
674         if (scaled == -1) {
675                 fprintf(stderr, "%*s%s%*s",
676                         csv_output ? 0 : 18,
677                         "<not counted>",
678                         csv_sep,
679                         csv_output ? 0 : -24,
680                         event_name(counter));
681
682                 if (counter->cgrp)
683                         fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
684
685                 fputc('\n', stderr);
686                 return;
687         }
688
689         if (nsec_counter(counter))
690                 nsec_printout(-1, counter, avg);
691         else
692                 abs_printout(-1, counter, avg);
693
694         if (csv_output) {
695                 fputc('\n', stderr);
696                 return;
697         }
698
699         print_noise(counter, avg);
700
701         if (scaled) {
702                 double avg_enabled, avg_running;
703
704                 avg_enabled = avg_stats(&ps->res_stats[1]);
705                 avg_running = avg_stats(&ps->res_stats[2]);
706
707                 fprintf(stderr, "  (%.2f%%)", 100 * avg_running / avg_enabled);
708         }
709         fprintf(stderr, "\n");
710 }
711
712 /*
713  * Print out the results of a single counter:
714  * does not use aggregated count in system-wide
715  */
716 static void print_counter(struct perf_evsel *counter)
717 {
718         u64 ena, run, val;
719         int cpu;
720
721         for (cpu = 0; cpu < evsel_list->cpus->nr; cpu++) {
722                 val = counter->counts->cpu[cpu].val;
723                 ena = counter->counts->cpu[cpu].ena;
724                 run = counter->counts->cpu[cpu].run;
725                 if (run == 0 || ena == 0) {
726                         fprintf(stderr, "CPU%*d%s%*s%s%*s",
727                                 csv_output ? 0 : -4,
728                                 evsel_list->cpus->map[cpu], csv_sep,
729                                 csv_output ? 0 : 18,
730                                 "<not counted>", csv_sep,
731                                 csv_output ? 0 : -24,
732                                 event_name(counter));
733
734                         if (counter->cgrp)
735                                 fprintf(stderr, "%s%s", csv_sep, counter->cgrp->name);
736
737                         fputc('\n', stderr);
738                         continue;
739                 }
740
741                 if (nsec_counter(counter))
742                         nsec_printout(cpu, counter, val);
743                 else
744                         abs_printout(cpu, counter, val);
745
746                 if (!csv_output) {
747                         print_noise(counter, 1.0);
748
749                         if (run != ena)
750                                 fprintf(stderr, "  (%.2f%%)", 100.0 * run / ena);
751                 }
752                 fputc('\n', stderr);
753         }
754 }
755
756 static void print_stat(int argc, const char **argv)
757 {
758         struct perf_evsel *counter;
759         int i;
760
761         fflush(stdout);
762
763         if (!csv_output) {
764                 fprintf(stderr, "\n");
765                 fprintf(stderr, " Performance counter stats for ");
766                 if(target_pid == -1 && target_tid == -1) {
767                         fprintf(stderr, "\'%s", argv[0]);
768                         for (i = 1; i < argc; i++)
769                                 fprintf(stderr, " %s", argv[i]);
770                 } else if (target_pid != -1)
771                         fprintf(stderr, "process id \'%d", target_pid);
772                 else
773                         fprintf(stderr, "thread id \'%d", target_tid);
774
775                 fprintf(stderr, "\'");
776                 if (run_count > 1)
777                         fprintf(stderr, " (%d runs)", run_count);
778                 fprintf(stderr, ":\n\n");
779         }
780
781         if (no_aggr) {
782                 list_for_each_entry(counter, &evsel_list->entries, node)
783                         print_counter(counter);
784         } else {
785                 list_for_each_entry(counter, &evsel_list->entries, node)
786                         print_counter_aggr(counter);
787         }
788
789         if (!csv_output) {
790                 fprintf(stderr, "\n");
791                 fprintf(stderr, " %18.9f  seconds time elapsed",
792                                 avg_stats(&walltime_nsecs_stats)/1e9);
793                 if (run_count > 1) {
794                         print_noise_pct(stddev_stats(&walltime_nsecs_stats),
795                                         avg_stats(&walltime_nsecs_stats));
796                 }
797                 fprintf(stderr, "\n\n");
798         }
799 }
800
801 static volatile int signr = -1;
802
803 static void skip_signal(int signo)
804 {
805         if(child_pid == -1)
806                 done = 1;
807
808         signr = signo;
809 }
810
811 static void sig_atexit(void)
812 {
813         if (child_pid != -1)
814                 kill(child_pid, SIGTERM);
815
816         if (signr == -1)
817                 return;
818
819         signal(signr, SIG_DFL);
820         kill(getpid(), signr);
821 }
822
823 static const char * const stat_usage[] = {
824         "perf stat [<options>] [<command>]",
825         NULL
826 };
827
828 static int stat__set_big_num(const struct option *opt __used,
829                              const char *s __used, int unset)
830 {
831         big_num_opt = unset ? 0 : 1;
832         return 0;
833 }
834
835 static const struct option options[] = {
836         OPT_CALLBACK('e', "event", &evsel_list, "event",
837                      "event selector. use 'perf list' to list available events",
838                      parse_events),
839         OPT_CALLBACK(0, "filter", &evsel_list, "filter",
840                      "event filter", parse_filter),
841         OPT_BOOLEAN('i', "no-inherit", &no_inherit,
842                     "child tasks do not inherit counters"),
843         OPT_INTEGER('p', "pid", &target_pid,
844                     "stat events on existing process id"),
845         OPT_INTEGER('t', "tid", &target_tid,
846                     "stat events on existing thread id"),
847         OPT_BOOLEAN('a', "all-cpus", &system_wide,
848                     "system-wide collection from all CPUs"),
849         OPT_BOOLEAN('c', "scale", &scale,
850                     "scale/normalize counters"),
851         OPT_INCR('v', "verbose", &verbose,
852                     "be more verbose (show counter open errors, etc)"),
853         OPT_INTEGER('r', "repeat", &run_count,
854                     "repeat command and print average + stddev (max: 100)"),
855         OPT_BOOLEAN('n', "null", &null_run,
856                     "null run - dont start any counters"),
857         OPT_BOOLEAN('d', "detailed", &detailed_run,
858                     "detailed run - start a lot of events"),
859         OPT_BOOLEAN('S', "sync", &sync_run,
860                     "call sync() before starting a run"),
861         OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 
862                            "print large numbers with thousands\' separators",
863                            stat__set_big_num),
864         OPT_STRING('C', "cpu", &cpu_list, "cpu",
865                     "list of cpus to monitor in system-wide"),
866         OPT_BOOLEAN('A', "no-aggr", &no_aggr,
867                     "disable CPU count aggregation"),
868         OPT_STRING('x', "field-separator", &csv_sep, "separator",
869                    "print counts with custom separator"),
870         OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
871                      "monitor event in cgroup name only",
872                      parse_cgroups),
873         OPT_END()
874 };
875
876 int cmd_stat(int argc, const char **argv, const char *prefix __used)
877 {
878         struct perf_evsel *pos;
879         int status = -ENOMEM;
880
881         setlocale(LC_ALL, "");
882
883         evsel_list = perf_evlist__new(NULL, NULL);
884         if (evsel_list == NULL)
885                 return -ENOMEM;
886
887         argc = parse_options(argc, argv, options, stat_usage,
888                 PARSE_OPT_STOP_AT_NON_OPTION);
889
890         if (csv_sep)
891                 csv_output = true;
892         else
893                 csv_sep = DEFAULT_SEPARATOR;
894
895         /*
896          * let the spreadsheet do the pretty-printing
897          */
898         if (csv_output) {
899                 /* User explicitely passed -B? */
900                 if (big_num_opt == 1) {
901                         fprintf(stderr, "-B option not supported with -x\n");
902                         usage_with_options(stat_usage, options);
903                 } else /* Nope, so disable big number formatting */
904                         big_num = false;
905         } else if (big_num_opt == 0) /* User passed --no-big-num */
906                 big_num = false;
907
908         if (!argc && target_pid == -1 && target_tid == -1)
909                 usage_with_options(stat_usage, options);
910         if (run_count <= 0)
911                 usage_with_options(stat_usage, options);
912
913         /* no_aggr, cgroup are for system-wide only */
914         if ((no_aggr || nr_cgroups) && !system_wide) {
915                 fprintf(stderr, "both cgroup and no-aggregation "
916                         "modes only available in system-wide mode\n");
917
918                 usage_with_options(stat_usage, options);
919         }
920
921         /* Set attrs and nr_counters if no event is selected and !null_run */
922         if (detailed_run) {
923                 size_t c;
924
925                 for (c = 0; c < ARRAY_SIZE(detailed_attrs); ++c) {
926                         pos = perf_evsel__new(&detailed_attrs[c], c);
927                         if (pos == NULL)
928                                 goto out;
929                         perf_evlist__add(evsel_list, pos);
930                 }
931         }
932         /* Set attrs and nr_counters if no event is selected and !null_run */
933         if (!detailed_run && !null_run && !evsel_list->nr_entries) {
934                 size_t c;
935
936                 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
937                         pos = perf_evsel__new(&default_attrs[c], c);
938                         if (pos == NULL)
939                                 goto out;
940                         perf_evlist__add(evsel_list, pos);
941                 }
942         }
943
944         if (target_pid != -1)
945                 target_tid = target_pid;
946
947         evsel_list->threads = thread_map__new(target_pid, target_tid);
948         if (evsel_list->threads == NULL) {
949                 pr_err("Problems finding threads of monitor\n");
950                 usage_with_options(stat_usage, options);
951         }
952
953         if (system_wide)
954                 evsel_list->cpus = cpu_map__new(cpu_list);
955         else
956                 evsel_list->cpus = cpu_map__dummy_new();
957
958         if (evsel_list->cpus == NULL) {
959                 perror("failed to parse CPUs map");
960                 usage_with_options(stat_usage, options);
961                 return -1;
962         }
963
964         list_for_each_entry(pos, &evsel_list->entries, node) {
965                 if (perf_evsel__alloc_stat_priv(pos) < 0 ||
966                     perf_evsel__alloc_counts(pos, evsel_list->cpus->nr) < 0 ||
967                     perf_evsel__alloc_fd(pos, evsel_list->cpus->nr, evsel_list->threads->nr) < 0)
968                         goto out_free_fd;
969         }
970
971         /*
972          * We dont want to block the signals - that would cause
973          * child tasks to inherit that and Ctrl-C would not work.
974          * What we want is for Ctrl-C to work in the exec()-ed
975          * task, but being ignored by perf stat itself:
976          */
977         atexit(sig_atexit);
978         signal(SIGINT,  skip_signal);
979         signal(SIGALRM, skip_signal);
980         signal(SIGABRT, skip_signal);
981
982         status = 0;
983         for (run_idx = 0; run_idx < run_count; run_idx++) {
984                 if (run_count != 1 && verbose)
985                         fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
986
987                 if (sync_run)
988                         sync();
989
990                 status = run_perf_stat(argc, argv);
991         }
992
993         if (status != -1)
994                 print_stat(argc, argv);
995 out_free_fd:
996         list_for_each_entry(pos, &evsel_list->entries, node)
997                 perf_evsel__free_stat_priv(pos);
998         perf_evlist__delete_maps(evsel_list);
999 out:
1000         perf_evlist__delete(evsel_list);
1001         return status;
1002 }