xref: /linux/tools/perf/builtin-stat.c (revision 913df4453f85f1fe79b35ecf3c9a0c0b707d22a2)
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 
48 #include <sys/prctl.h>
49 #include <math.h>
50 
51 static struct perf_event_attr default_attrs[] = {
52 
53   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK	},
54   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
55   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS	},
56   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS	},
57 
58   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES	},
59   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS	},
60   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
61   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES	},
62 
63 };
64 
65 static int			system_wide			=  0;
66 static unsigned int		nr_cpus				=  0;
67 static int			run_idx				=  0;
68 
69 static int			run_count			=  1;
70 static int			inherit				=  1;
71 static int			scale				=  1;
72 static pid_t			target_pid			= -1;
73 static pid_t			child_pid			= -1;
74 static int			null_run			=  0;
75 
76 static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
77 
78 static int			event_scaled[MAX_COUNTERS];
79 
80 struct stats
81 {
82 	double n, mean, M2;
83 };
84 
85 static void update_stats(struct stats *stats, u64 val)
86 {
87 	double delta;
88 
89 	stats->n++;
90 	delta = val - stats->mean;
91 	stats->mean += delta / stats->n;
92 	stats->M2 += delta*(val - stats->mean);
93 }
94 
95 static double avg_stats(struct stats *stats)
96 {
97 	return stats->mean;
98 }
99 
100 /*
101  * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
102  *
103  *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
104  * s^2 = -------------------------------
105  *                  n - 1
106  *
107  * http://en.wikipedia.org/wiki/Stddev
108  *
109  * The std dev of the mean is related to the std dev by:
110  *
111  *             s
112  * s_mean = -------
113  *          sqrt(n)
114  *
115  */
116 static double stddev_stats(struct stats *stats)
117 {
118 	double variance = stats->M2 / (stats->n - 1);
119 	double variance_mean = variance / stats->n;
120 
121 	return sqrt(variance_mean);
122 }
123 
124 struct stats			event_res_stats[MAX_COUNTERS][3];
125 struct stats			runtime_nsecs_stats;
126 struct stats			walltime_nsecs_stats;
127 struct stats			runtime_cycles_stats;
128 
129 #define MATCH_EVENT(t, c, counter)			\
130 	(attrs[counter].type == PERF_TYPE_##t &&	\
131 	 attrs[counter].config == PERF_COUNT_##c)
132 
133 #define ERR_PERF_OPEN \
134 "Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
135 
136 static void create_perf_stat_counter(int counter, int pid)
137 {
138 	struct perf_event_attr *attr = attrs + counter;
139 
140 	if (scale)
141 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
142 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
143 
144 	if (system_wide) {
145 		unsigned int cpu;
146 
147 		for (cpu = 0; cpu < nr_cpus; cpu++) {
148 			fd[cpu][counter] = sys_perf_event_open(attr, -1, cpu, -1, 0);
149 			if (fd[cpu][counter] < 0 && verbose)
150 				fprintf(stderr, ERR_PERF_OPEN, counter,
151 					fd[cpu][counter], strerror(errno));
152 		}
153 	} else {
154 		attr->inherit	     = inherit;
155 		attr->disabled	     = 1;
156 		attr->enable_on_exec = 1;
157 
158 		fd[0][counter] = sys_perf_event_open(attr, pid, -1, -1, 0);
159 		if (fd[0][counter] < 0 && verbose)
160 			fprintf(stderr, ERR_PERF_OPEN, counter,
161 				fd[0][counter], strerror(errno));
162 	}
163 }
164 
165 /*
166  * Does the counter have nsecs as a unit?
167  */
168 static inline int nsec_counter(int counter)
169 {
170 	if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
171 	    MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
172 		return 1;
173 
174 	return 0;
175 }
176 
177 /*
178  * Read out the results of a single counter:
179  */
180 static void read_counter(int counter)
181 {
182 	u64 count[3], single_count[3];
183 	unsigned int cpu;
184 	size_t res, nv;
185 	int scaled;
186 	int i;
187 
188 	count[0] = count[1] = count[2] = 0;
189 
190 	nv = scale ? 3 : 1;
191 	for (cpu = 0; cpu < nr_cpus; cpu++) {
192 		if (fd[cpu][counter] < 0)
193 			continue;
194 
195 		res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
196 		assert(res == nv * sizeof(u64));
197 
198 		close(fd[cpu][counter]);
199 		fd[cpu][counter] = -1;
200 
201 		count[0] += single_count[0];
202 		if (scale) {
203 			count[1] += single_count[1];
204 			count[2] += single_count[2];
205 		}
206 	}
207 
208 	scaled = 0;
209 	if (scale) {
210 		if (count[2] == 0) {
211 			event_scaled[counter] = -1;
212 			count[0] = 0;
213 			return;
214 		}
215 
216 		if (count[2] < count[1]) {
217 			event_scaled[counter] = 1;
218 			count[0] = (unsigned long long)
219 				((double)count[0] * count[1] / count[2] + 0.5);
220 		}
221 	}
222 
223 	for (i = 0; i < 3; i++)
224 		update_stats(&event_res_stats[counter][i], count[i]);
225 
226 	if (verbose) {
227 		fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
228 				count[0], count[1], count[2]);
229 	}
230 
231 	/*
232 	 * Save the full runtime - to allow normalization during printout:
233 	 */
234 	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
235 		update_stats(&runtime_nsecs_stats, count[0]);
236 	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
237 		update_stats(&runtime_cycles_stats, count[0]);
238 }
239 
240 static int run_perf_stat(int argc __used, const char **argv)
241 {
242 	unsigned long long t0, t1;
243 	int status = 0;
244 	int counter;
245 	int pid;
246 	int child_ready_pipe[2], go_pipe[2];
247 	char buf;
248 
249 	if (!system_wide)
250 		nr_cpus = 1;
251 
252 	if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
253 		perror("failed to create pipes");
254 		exit(1);
255 	}
256 
257 	if ((pid = fork()) < 0)
258 		perror("failed to fork");
259 
260 	if (!pid) {
261 		close(child_ready_pipe[0]);
262 		close(go_pipe[1]);
263 		fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
264 
265 		/*
266 		 * Do a dummy execvp to get the PLT entry resolved,
267 		 * so we avoid the resolver overhead on the real
268 		 * execvp call.
269 		 */
270 		execvp("", (char **)argv);
271 
272 		/*
273 		 * Tell the parent we're ready to go
274 		 */
275 		close(child_ready_pipe[1]);
276 
277 		/*
278 		 * Wait until the parent tells us to go.
279 		 */
280 		if (read(go_pipe[0], &buf, 1) == -1)
281 			perror("unable to read pipe");
282 
283 		execvp(argv[0], (char **)argv);
284 
285 		perror(argv[0]);
286 		exit(-1);
287 	}
288 
289 	child_pid = pid;
290 
291 	/*
292 	 * Wait for the child to be ready to exec.
293 	 */
294 	close(child_ready_pipe[1]);
295 	close(go_pipe[0]);
296 	if (read(child_ready_pipe[0], &buf, 1) == -1)
297 		perror("unable to read pipe");
298 	close(child_ready_pipe[0]);
299 
300 	for (counter = 0; counter < nr_counters; counter++)
301 		create_perf_stat_counter(counter, pid);
302 
303 	/*
304 	 * Enable counters and exec the command:
305 	 */
306 	t0 = rdclock();
307 
308 	close(go_pipe[1]);
309 	wait(&status);
310 
311 	t1 = rdclock();
312 
313 	update_stats(&walltime_nsecs_stats, t1 - t0);
314 
315 	for (counter = 0; counter < nr_counters; counter++)
316 		read_counter(counter);
317 
318 	return WEXITSTATUS(status);
319 }
320 
321 static void print_noise(int counter, double avg)
322 {
323 	if (run_count == 1)
324 		return;
325 
326 	fprintf(stderr, "   ( +- %7.3f%% )",
327 			100 * stddev_stats(&event_res_stats[counter][0]) / avg);
328 }
329 
330 static void nsec_printout(int counter, double avg)
331 {
332 	double msecs = avg / 1e6;
333 
334 	fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
335 
336 	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
337 		fprintf(stderr, " # %10.3f CPUs ",
338 				avg / avg_stats(&walltime_nsecs_stats));
339 	}
340 }
341 
342 static void abs_printout(int counter, double avg)
343 {
344 	double total, ratio = 0.0;
345 
346 	fprintf(stderr, " %14.0f  %-24s", avg, event_name(counter));
347 
348 	if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
349 		total = avg_stats(&runtime_cycles_stats);
350 
351 		if (total)
352 			ratio = avg / total;
353 
354 		fprintf(stderr, " # %10.3f IPC  ", ratio);
355 	} else {
356 		total = avg_stats(&runtime_nsecs_stats);
357 
358 		if (total)
359 			ratio = 1000.0 * avg / total;
360 
361 		fprintf(stderr, " # %10.3f M/sec", ratio);
362 	}
363 }
364 
365 /*
366  * Print out the results of a single counter:
367  */
368 static void print_counter(int counter)
369 {
370 	double avg = avg_stats(&event_res_stats[counter][0]);
371 	int scaled = event_scaled[counter];
372 
373 	if (scaled == -1) {
374 		fprintf(stderr, " %14s  %-24s\n",
375 			"<not counted>", event_name(counter));
376 		return;
377 	}
378 
379 	if (nsec_counter(counter))
380 		nsec_printout(counter, avg);
381 	else
382 		abs_printout(counter, avg);
383 
384 	print_noise(counter, avg);
385 
386 	if (scaled) {
387 		double avg_enabled, avg_running;
388 
389 		avg_enabled = avg_stats(&event_res_stats[counter][1]);
390 		avg_running = avg_stats(&event_res_stats[counter][2]);
391 
392 		fprintf(stderr, "  (scaled from %.2f%%)",
393 				100 * avg_running / avg_enabled);
394 	}
395 
396 	fprintf(stderr, "\n");
397 }
398 
399 static void print_stat(int argc, const char **argv)
400 {
401 	int i, counter;
402 
403 	fflush(stdout);
404 
405 	fprintf(stderr, "\n");
406 	fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
407 
408 	for (i = 1; i < argc; i++)
409 		fprintf(stderr, " %s", argv[i]);
410 
411 	fprintf(stderr, "\'");
412 	if (run_count > 1)
413 		fprintf(stderr, " (%d runs)", run_count);
414 	fprintf(stderr, ":\n\n");
415 
416 	for (counter = 0; counter < nr_counters; counter++)
417 		print_counter(counter);
418 
419 	fprintf(stderr, "\n");
420 	fprintf(stderr, " %14.9f  seconds time elapsed",
421 			avg_stats(&walltime_nsecs_stats)/1e9);
422 	if (run_count > 1) {
423 		fprintf(stderr, "   ( +- %7.3f%% )",
424 				100*stddev_stats(&walltime_nsecs_stats) /
425 				avg_stats(&walltime_nsecs_stats));
426 	}
427 	fprintf(stderr, "\n\n");
428 }
429 
430 static volatile int signr = -1;
431 
432 static void skip_signal(int signo)
433 {
434 	signr = signo;
435 }
436 
437 static void sig_atexit(void)
438 {
439 	if (child_pid != -1)
440 		kill(child_pid, SIGTERM);
441 
442 	if (signr == -1)
443 		return;
444 
445 	signal(signr, SIG_DFL);
446 	kill(getpid(), signr);
447 }
448 
449 static const char * const stat_usage[] = {
450 	"perf stat [<options>] <command>",
451 	NULL
452 };
453 
454 static const struct option options[] = {
455 	OPT_CALLBACK('e', "event", NULL, "event",
456 		     "event selector. use 'perf list' to list available events",
457 		     parse_events),
458 	OPT_BOOLEAN('i', "inherit", &inherit,
459 		    "child tasks inherit counters"),
460 	OPT_INTEGER('p', "pid", &target_pid,
461 		    "stat events on existing pid"),
462 	OPT_BOOLEAN('a', "all-cpus", &system_wide,
463 		    "system-wide collection from all CPUs"),
464 	OPT_BOOLEAN('c', "scale", &scale,
465 		    "scale/normalize counters"),
466 	OPT_BOOLEAN('v', "verbose", &verbose,
467 		    "be more verbose (show counter open errors, etc)"),
468 	OPT_INTEGER('r', "repeat", &run_count,
469 		    "repeat command and print average + stddev (max: 100)"),
470 	OPT_BOOLEAN('n', "null", &null_run,
471 		    "null run - dont start any counters"),
472 	OPT_END()
473 };
474 
475 int cmd_stat(int argc, const char **argv, const char *prefix __used)
476 {
477 	int status;
478 
479 	argc = parse_options(argc, argv, options, stat_usage,
480 		PARSE_OPT_STOP_AT_NON_OPTION);
481 	if (!argc)
482 		usage_with_options(stat_usage, options);
483 	if (run_count <= 0)
484 		usage_with_options(stat_usage, options);
485 
486 	/* Set attrs and nr_counters if no event is selected and !null_run */
487 	if (!null_run && !nr_counters) {
488 		memcpy(attrs, default_attrs, sizeof(default_attrs));
489 		nr_counters = ARRAY_SIZE(default_attrs);
490 	}
491 
492 	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
493 	assert(nr_cpus <= MAX_NR_CPUS);
494 	assert((int)nr_cpus >= 0);
495 
496 	/*
497 	 * We dont want to block the signals - that would cause
498 	 * child tasks to inherit that and Ctrl-C would not work.
499 	 * What we want is for Ctrl-C to work in the exec()-ed
500 	 * task, but being ignored by perf stat itself:
501 	 */
502 	atexit(sig_atexit);
503 	signal(SIGINT,  skip_signal);
504 	signal(SIGALRM, skip_signal);
505 	signal(SIGABRT, skip_signal);
506 
507 	status = 0;
508 	for (run_idx = 0; run_idx < run_count; run_idx++) {
509 		if (run_count != 1 && verbose)
510 			fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
511 		status = run_perf_stat(argc, argv);
512 	}
513 
514 	print_stat(argc, argv);
515 
516 	return status;
517 }
518