xref: /linux/tools/perf/builtin-stat.c (revision f8324e20f8289dffc646d64366332e05eaacab25)
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/debug.h"
47 #include "util/header.h"
48 #include "util/cpumap.h"
49 #include "util/thread.h"
50 
51 #include <sys/prctl.h>
52 #include <math.h>
53 #include <locale.h>
54 
55 static struct perf_event_attr default_attrs[] = {
56 
57   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
58   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
59   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
60   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
61 
62   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
63   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
64   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
65   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
66   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES	},
67   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES		},
68 
69 };
70 
71 static bool			system_wide			=  false;
72 static unsigned int		nr_cpus				=  0;
73 static int			run_idx				=  0;
74 
75 static int			run_count			=  1;
76 static bool			no_inherit			= false;
77 static bool			scale				=  true;
78 static pid_t			target_pid			= -1;
79 static pid_t			target_tid			= -1;
80 static pid_t			*all_tids			=  NULL;
81 static int			thread_num			=  0;
82 static pid_t			child_pid			= -1;
83 static bool			null_run			=  false;
84 static bool			big_num				=  false;
85 
86 
87 static int			*fd[MAX_NR_CPUS][MAX_COUNTERS];
88 
89 static int			event_scaled[MAX_COUNTERS];
90 
91 static volatile int done = 0;
92 
93 struct stats
94 {
95 	double n, mean, M2;
96 };
97 
98 static void update_stats(struct stats *stats, u64 val)
99 {
100 	double delta;
101 
102 	stats->n++;
103 	delta = val - stats->mean;
104 	stats->mean += delta / stats->n;
105 	stats->M2 += delta*(val - stats->mean);
106 }
107 
108 static double avg_stats(struct stats *stats)
109 {
110 	return stats->mean;
111 }
112 
113 /*
114  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
115  *
116  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
117  * s^2 = -------------------------------
118  *                  n - 1
119  *
120  * http://en.wikipedia.org/wiki/Stddev
121  *
122  * The std dev of the mean is related to the std dev by:
123  *
124  *             s
125  * s_mean = -------
126  *          sqrt(n)
127  *
128  */
129 static double stddev_stats(struct stats *stats)
130 {
131 	double variance = stats->M2 / (stats->n - 1);
132 	double variance_mean = variance / stats->n;
133 
134 	return sqrt(variance_mean);
135 }
136 
137 struct stats			event_res_stats[MAX_COUNTERS][3];
138 struct stats			runtime_nsecs_stats;
139 struct stats			walltime_nsecs_stats;
140 struct stats			runtime_cycles_stats;
141 struct stats			runtime_branches_stats;
142 
143 #define MATCH_EVENT(t, c, counter)			\
144 	(attrs[counter].type == PERF_TYPE_##t &&	\
145 	 attrs[counter].config == PERF_COUNT_##c)
146 
147 #define ERR_PERF_OPEN \
148 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
149 
150 static int create_perf_stat_counter(int counter)
151 {
152 	struct perf_event_attr *attr = attrs + counter;
153 	int thread;
154 	int ncreated = 0;
155 
156 	if (scale)
157 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
158 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
159 
160 	if (system_wide) {
161 		unsigned int cpu;
162 
163 		for (cpu = 0; cpu < nr_cpus; cpu++) {
164 			fd[cpu][counter][0] = sys_perf_event_open(attr,
165 					-1, cpumap[cpu], -1, 0);
166 			if (fd[cpu][counter][0] < 0)
167 				pr_debug(ERR_PERF_OPEN, counter,
168 					 fd[cpu][counter][0], strerror(errno));
169 			else
170 				++ncreated;
171 		}
172 	} else {
173 		attr->inherit = !no_inherit;
174 		if (target_pid == -1 && target_tid == -1) {
175 			attr->disabled = 1;
176 			attr->enable_on_exec = 1;
177 		}
178 		for (thread = 0; thread < thread_num; thread++) {
179 			fd[0][counter][thread] = sys_perf_event_open(attr,
180 				all_tids[thread], -1, -1, 0);
181 			if (fd[0][counter][thread] < 0)
182 				pr_debug(ERR_PERF_OPEN, counter,
183 					 fd[0][counter][thread],
184 					 strerror(errno));
185 			else
186 				++ncreated;
187 		}
188 	}
189 
190 	return ncreated;
191 }
192 
193 /*
194  * Does the counter have nsecs as a unit?
195  */
196 static inline int nsec_counter(int counter)
197 {
198 	if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
199 	    MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
200 		return 1;
201 
202 	return 0;
203 }
204 
205 /*
206  * Read out the results of a single counter:
207  */
208 static void read_counter(int counter)
209 {
210 	u64 count[3], single_count[3];
211 	unsigned int cpu;
212 	size_t res, nv;
213 	int scaled;
214 	int i, thread;
215 
216 	count[0] = count[1] = count[2] = 0;
217 
218 	nv = scale ? 3 : 1;
219 	for (cpu = 0; cpu < nr_cpus; cpu++) {
220 		for (thread = 0; thread < thread_num; thread++) {
221 			if (fd[cpu][counter][thread] < 0)
222 				continue;
223 
224 			res = read(fd[cpu][counter][thread],
225 					single_count, nv * sizeof(u64));
226 			assert(res == nv * sizeof(u64));
227 
228 			close(fd[cpu][counter][thread]);
229 			fd[cpu][counter][thread] = -1;
230 
231 			count[0] += single_count[0];
232 			if (scale) {
233 				count[1] += single_count[1];
234 				count[2] += single_count[2];
235 			}
236 		}
237 	}
238 
239 	scaled = 0;
240 	if (scale) {
241 		if (count[2] == 0) {
242 			event_scaled[counter] = -1;
243 			count[0] = 0;
244 			return;
245 		}
246 
247 		if (count[2] < count[1]) {
248 			event_scaled[counter] = 1;
249 			count[0] = (unsigned long long)
250 				((double)count[0] * count[1] / count[2] + 0.5);
251 		}
252 	}
253 
254 	for (i = 0; i < 3; i++)
255 		update_stats(&event_res_stats[counter][i], count[i]);
256 
257 	if (verbose) {
258 		fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
259 				count[0], count[1], count[2]);
260 	}
261 
262 	/*
263 	 * Save the full runtime - to allow normalization during printout:
264 	 */
265 	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
266 		update_stats(&runtime_nsecs_stats, count[0]);
267 	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
268 		update_stats(&runtime_cycles_stats, count[0]);
269 	if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
270 		update_stats(&runtime_branches_stats, count[0]);
271 }
272 
273 static int run_perf_stat(int argc __used, const char **argv)
274 {
275 	unsigned long long t0, t1;
276 	int status = 0;
277 	int counter, ncreated = 0;
278 	int child_ready_pipe[2], go_pipe[2];
279 	const bool forks = (argc > 0);
280 	char buf;
281 
282 	if (!system_wide)
283 		nr_cpus = 1;
284 
285 	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
286 		perror("failed to create pipes");
287 		exit(1);
288 	}
289 
290 	if (forks) {
291 		if ((child_pid = fork()) < 0)
292 			perror("failed to fork");
293 
294 		if (!child_pid) {
295 			close(child_ready_pipe[0]);
296 			close(go_pipe[1]);
297 			fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
298 
299 			/*
300 			 * Do a dummy execvp to get the PLT entry resolved,
301 			 * so we avoid the resolver overhead on the real
302 			 * execvp call.
303 			 */
304 			execvp("", (char **)argv);
305 
306 			/*
307 			 * Tell the parent we're ready to go
308 			 */
309 			close(child_ready_pipe[1]);
310 
311 			/*
312 			 * Wait until the parent tells us to go.
313 			 */
314 			if (read(go_pipe[0], &buf, 1) == -1)
315 				perror("unable to read pipe");
316 
317 			execvp(argv[0], (char **)argv);
318 
319 			perror(argv[0]);
320 			exit(-1);
321 		}
322 
323 		if (target_tid == -1 && target_pid == -1 && !system_wide)
324 			all_tids[0] = child_pid;
325 
326 		/*
327 		 * Wait for the child to be ready to exec.
328 		 */
329 		close(child_ready_pipe[1]);
330 		close(go_pipe[0]);
331 		if (read(child_ready_pipe[0], &buf, 1) == -1)
332 			perror("unable to read pipe");
333 		close(child_ready_pipe[0]);
334 	}
335 
336 	for (counter = 0; counter < nr_counters; counter++)
337 		ncreated += create_perf_stat_counter(counter);
338 
339 	if (ncreated == 0) {
340 		pr_err("No permission to collect %sstats.\n"
341 		       "Consider tweaking /proc/sys/kernel/perf_event_paranoid.\n",
342 		       system_wide ? "system-wide " : "");
343 		if (child_pid != -1)
344 			kill(child_pid, SIGTERM);
345 		return -1;
346 	}
347 
348 	/*
349 	 * Enable counters and exec the command:
350 	 */
351 	t0 = rdclock();
352 
353 	if (forks) {
354 		close(go_pipe[1]);
355 		wait(&status);
356 	} else {
357 		while(!done) sleep(1);
358 	}
359 
360 	t1 = rdclock();
361 
362 	update_stats(&walltime_nsecs_stats, t1 - t0);
363 
364 	for (counter = 0; counter < nr_counters; counter++)
365 		read_counter(counter);
366 
367 	return WEXITSTATUS(status);
368 }
369 
370 static void print_noise(int counter, double avg)
371 {
372 	if (run_count == 1)
373 		return;
374 
375 	fprintf(stderr, "   ( +- %7.3f%% )",
376 			100 * stddev_stats(&event_res_stats[counter][0]) / avg);
377 }
378 
379 static void nsec_printout(int counter, double avg)
380 {
381 	double msecs = avg / 1e6;
382 
383 	fprintf(stderr, " %18.6f  %-24s", msecs, event_name(counter));
384 
385 	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
386 		fprintf(stderr, " # %10.3f CPUs ",
387 				avg / avg_stats(&walltime_nsecs_stats));
388 	}
389 }
390 
391 static void abs_printout(int counter, double avg)
392 {
393 	double total, ratio = 0.0;
394 
395 	if (big_num)
396 		fprintf(stderr, " %'18.0f  %-24s", avg, event_name(counter));
397 	else
398 		fprintf(stderr, " %18.0f  %-24s", avg, event_name(counter));
399 
400 	if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
401 		total = avg_stats(&runtime_cycles_stats);
402 
403 		if (total)
404 			ratio = avg / total;
405 
406 		fprintf(stderr, " # %10.3f IPC  ", ratio);
407 	} else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) &&
408 			runtime_branches_stats.n != 0) {
409 		total = avg_stats(&runtime_branches_stats);
410 
411 		if (total)
412 			ratio = avg * 100 / total;
413 
414 		fprintf(stderr, " # %10.3f %%    ", ratio);
415 
416 	} else if (runtime_nsecs_stats.n != 0) {
417 		total = avg_stats(&runtime_nsecs_stats);
418 
419 		if (total)
420 			ratio = 1000.0 * avg / total;
421 
422 		fprintf(stderr, " # %10.3f M/sec", ratio);
423 	}
424 }
425 
426 /*
427  * Print out the results of a single counter:
428  */
429 static void print_counter(int counter)
430 {
431 	double avg = avg_stats(&event_res_stats[counter][0]);
432 	int scaled = event_scaled[counter];
433 
434 	if (scaled == -1) {
435 		fprintf(stderr, " %18s  %-24s\n",
436 			"<not counted>", event_name(counter));
437 		return;
438 	}
439 
440 	if (nsec_counter(counter))
441 		nsec_printout(counter, avg);
442 	else
443 		abs_printout(counter, avg);
444 
445 	print_noise(counter, avg);
446 
447 	if (scaled) {
448 		double avg_enabled, avg_running;
449 
450 		avg_enabled = avg_stats(&event_res_stats[counter][1]);
451 		avg_running = avg_stats(&event_res_stats[counter][2]);
452 
453 		fprintf(stderr, "  (scaled from %.2f%%)",
454 				100 * avg_running / avg_enabled);
455 	}
456 
457 	fprintf(stderr, "\n");
458 }
459 
460 static void print_stat(int argc, const char **argv)
461 {
462 	int i, counter;
463 
464 	fflush(stdout);
465 
466 	fprintf(stderr, "\n");
467 	fprintf(stderr, " Performance counter stats for ");
468 	if(target_pid == -1 && target_tid == -1) {
469 		fprintf(stderr, "\'%s", argv[0]);
470 		for (i = 1; i < argc; i++)
471 			fprintf(stderr, " %s", argv[i]);
472 	} else if (target_pid != -1)
473 		fprintf(stderr, "process id \'%d", target_pid);
474 	else
475 		fprintf(stderr, "thread id \'%d", target_tid);
476 
477 	fprintf(stderr, "\'");
478 	if (run_count > 1)
479 		fprintf(stderr, " (%d runs)", run_count);
480 	fprintf(stderr, ":\n\n");
481 
482 	for (counter = 0; counter < nr_counters; counter++)
483 		print_counter(counter);
484 
485 	fprintf(stderr, "\n");
486 	fprintf(stderr, " %18.9f  seconds time elapsed",
487 			avg_stats(&walltime_nsecs_stats)/1e9);
488 	if (run_count > 1) {
489 		fprintf(stderr, "   ( +- %7.3f%% )",
490 				100*stddev_stats(&walltime_nsecs_stats) /
491 				avg_stats(&walltime_nsecs_stats));
492 	}
493 	fprintf(stderr, "\n\n");
494 }
495 
496 static volatile int signr = -1;
497 
498 static void skip_signal(int signo)
499 {
500 	if(child_pid == -1)
501 		done = 1;
502 
503 	signr = signo;
504 }
505 
506 static void sig_atexit(void)
507 {
508 	if (child_pid != -1)
509 		kill(child_pid, SIGTERM);
510 
511 	if (signr == -1)
512 		return;
513 
514 	signal(signr, SIG_DFL);
515 	kill(getpid(), signr);
516 }
517 
518 static const char * const stat_usage[] = {
519 	"perf stat [<options>] [<command>]",
520 	NULL
521 };
522 
523 static const struct option options[] = {
524 	OPT_CALLBACK('e', "event", NULL, "event",
525 		     "event selector. use 'perf list' to list available events",
526 		     parse_events),
527 	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
528 		    "child tasks do not inherit counters"),
529 	OPT_INTEGER('p', "pid", &target_pid,
530 		    "stat events on existing process id"),
531 	OPT_INTEGER('t', "tid", &target_tid,
532 		    "stat events on existing thread id"),
533 	OPT_BOOLEAN('a', "all-cpus", &system_wide,
534 		    "system-wide collection from all CPUs"),
535 	OPT_BOOLEAN('c', "scale", &scale,
536 		    "scale/normalize counters"),
537 	OPT_INCR('v', "verbose", &verbose,
538 		    "be more verbose (show counter open errors, etc)"),
539 	OPT_INTEGER('r', "repeat", &run_count,
540 		    "repeat command and print average + stddev (max: 100)"),
541 	OPT_BOOLEAN('n', "null", &null_run,
542 		    "null run - dont start any counters"),
543 	OPT_BOOLEAN('B', "big-num", &big_num,
544 		    "print large numbers with thousands\' separators"),
545 	OPT_END()
546 };
547 
548 int cmd_stat(int argc, const char **argv, const char *prefix __used)
549 {
550 	int status;
551 	int i,j;
552 
553 	setlocale(LC_ALL, "");
554 
555 	argc = parse_options(argc, argv, options, stat_usage,
556 		PARSE_OPT_STOP_AT_NON_OPTION);
557 	if (!argc && target_pid == -1 && target_tid == -1)
558 		usage_with_options(stat_usage, options);
559 	if (run_count <= 0)
560 		usage_with_options(stat_usage, options);
561 
562 	/* Set attrs and nr_counters if no event is selected and !null_run */
563 	if (!null_run && !nr_counters) {
564 		memcpy(attrs, default_attrs, sizeof(default_attrs));
565 		nr_counters = ARRAY_SIZE(default_attrs);
566 	}
567 
568 	if (system_wide)
569 		nr_cpus = read_cpu_map();
570 	else
571 		nr_cpus = 1;
572 
573 	if (target_pid != -1) {
574 		target_tid = target_pid;
575 		thread_num = find_all_tid(target_pid, &all_tids);
576 		if (thread_num <= 0) {
577 			fprintf(stderr, "Can't find all threads of pid %d\n",
578 					target_pid);
579 			usage_with_options(stat_usage, options);
580 		}
581 	} else {
582 		all_tids=malloc(sizeof(pid_t));
583 		if (!all_tids)
584 			return -ENOMEM;
585 
586 		all_tids[0] = target_tid;
587 		thread_num = 1;
588 	}
589 
590 	for (i = 0; i < MAX_NR_CPUS; i++) {
591 		for (j = 0; j < MAX_COUNTERS; j++) {
592 			fd[i][j] = malloc(sizeof(int)*thread_num);
593 			if (!fd[i][j])
594 				return -ENOMEM;
595 		}
596 	}
597 
598 	/*
599 	 * We dont want to block the signals - that would cause
600 	 * child tasks to inherit that and Ctrl-C would not work.
601 	 * What we want is for Ctrl-C to work in the exec()-ed
602 	 * task, but being ignored by perf stat itself:
603 	 */
604 	atexit(sig_atexit);
605 	signal(SIGINT,  skip_signal);
606 	signal(SIGALRM, skip_signal);
607 	signal(SIGABRT, skip_signal);
608 
609 	status = 0;
610 	for (run_idx = 0; run_idx < run_count; run_idx++) {
611 		if (run_count != 1 && verbose)
612 			fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
613 		status = run_perf_stat(argc, argv);
614 	}
615 
616 	if (status != -1)
617 		print_stat(argc, argv);
618 
619 	return status;
620 }
621