xref: /linux/tools/perf/builtin-stat.c (revision 5bdef865eb358b6f3760e25e591ae115e9eeddef)
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 
46 #include <sys/prctl.h>
47 #include <math.h>
48 
49 static struct perf_counter_attr default_attrs[] = {
50 
51   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK	},
52   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
53   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS	},
54   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS	},
55 
56   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES	},
57   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS	},
58   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
59   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES	},
60 
61 };
62 
63 #define MAX_RUN			100
64 
65 static int			system_wide			=  0;
66 static int			verbose				=  0;
67 static unsigned int		nr_cpus				=  0;
68 static int			run_idx				=  0;
69 
70 static int			run_count			=  1;
71 static int			inherit				=  1;
72 static int			scale				=  1;
73 static int			target_pid			= -1;
74 static int			null_run			=  0;
75 
76 static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
77 
78 static u64			runtime_nsecs[MAX_RUN];
79 static u64			walltime_nsecs[MAX_RUN];
80 static u64			runtime_cycles[MAX_RUN];
81 
82 static u64			event_res[MAX_RUN][MAX_COUNTERS][3];
83 static u64			event_scaled[MAX_RUN][MAX_COUNTERS];
84 
85 static u64			event_res_avg[MAX_COUNTERS][3];
86 static u64			event_res_noise[MAX_COUNTERS][3];
87 
88 static u64			event_scaled_avg[MAX_COUNTERS];
89 
90 static u64			runtime_nsecs_avg;
91 static u64			runtime_nsecs_noise;
92 
93 static u64			walltime_nsecs_avg;
94 static u64			walltime_nsecs_noise;
95 
96 static u64			runtime_cycles_avg;
97 static u64			runtime_cycles_noise;
98 
99 #define MATCH_EVENT(t, c, counter)			\
100 	(attrs[counter].type == PERF_TYPE_##t &&	\
101 	 attrs[counter].config == PERF_COUNT_##c)
102 
103 #define ERR_PERF_OPEN \
104 "Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n"
105 
106 static void create_perf_stat_counter(int counter, int pid)
107 {
108 	struct perf_counter_attr *attr = attrs + counter;
109 
110 	if (scale)
111 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
112 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
113 
114 	if (system_wide) {
115 		unsigned int cpu;
116 
117 		for (cpu = 0; cpu < nr_cpus; cpu++) {
118 			fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
119 			if (fd[cpu][counter] < 0 && verbose)
120 				fprintf(stderr, ERR_PERF_OPEN, counter,
121 					fd[cpu][counter], strerror(errno));
122 		}
123 	} else {
124 		attr->inherit	     = inherit;
125 		attr->disabled	     = 1;
126 		attr->enable_on_exec = 1;
127 
128 		fd[0][counter] = sys_perf_counter_open(attr, pid, -1, -1, 0);
129 		if (fd[0][counter] < 0 && verbose)
130 			fprintf(stderr, ERR_PERF_OPEN, counter,
131 				fd[0][counter], strerror(errno));
132 	}
133 }
134 
135 /*
136  * Does the counter have nsecs as a unit?
137  */
138 static inline int nsec_counter(int counter)
139 {
140 	if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
141 	    MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
142 		return 1;
143 
144 	return 0;
145 }
146 
147 /*
148  * Read out the results of a single counter:
149  */
150 static void read_counter(int counter)
151 {
152 	u64 *count, single_count[3];
153 	unsigned int cpu;
154 	size_t res, nv;
155 	int scaled;
156 
157 	count = event_res[run_idx][counter];
158 
159 	count[0] = count[1] = count[2] = 0;
160 
161 	nv = scale ? 3 : 1;
162 	for (cpu = 0; cpu < nr_cpus; cpu++) {
163 		if (fd[cpu][counter] < 0)
164 			continue;
165 
166 		res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
167 		assert(res == nv * sizeof(u64));
168 
169 		close(fd[cpu][counter]);
170 		fd[cpu][counter] = -1;
171 
172 		count[0] += single_count[0];
173 		if (scale) {
174 			count[1] += single_count[1];
175 			count[2] += single_count[2];
176 		}
177 	}
178 
179 	scaled = 0;
180 	if (scale) {
181 		if (count[2] == 0) {
182 			event_scaled[run_idx][counter] = -1;
183 			count[0] = 0;
184 			return;
185 		}
186 
187 		if (count[2] < count[1]) {
188 			event_scaled[run_idx][counter] = 1;
189 			count[0] = (unsigned long long)
190 				((double)count[0] * count[1] / count[2] + 0.5);
191 		}
192 	}
193 	/*
194 	 * Save the full runtime - to allow normalization during printout:
195 	 */
196 	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
197 		runtime_nsecs[run_idx] = count[0];
198 	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
199 		runtime_cycles[run_idx] = count[0];
200 }
201 
202 static int run_perf_stat(int argc __used, const char **argv)
203 {
204 	unsigned long long t0, t1;
205 	int status = 0;
206 	int counter;
207 	int pid;
208 	int child_ready_pipe[2], go_pipe[2];
209 	char buf;
210 
211 	if (!system_wide)
212 		nr_cpus = 1;
213 
214 	if (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0) {
215 		perror("failed to create pipes");
216 		exit(1);
217 	}
218 
219 	if ((pid = fork()) < 0)
220 		perror("failed to fork");
221 
222 	if (!pid) {
223 		close(child_ready_pipe[0]);
224 		close(go_pipe[1]);
225 		fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
226 
227 		/*
228 		 * Do a dummy execvp to get the PLT entry resolved,
229 		 * so we avoid the resolver overhead on the real
230 		 * execvp call.
231 		 */
232 		execvp("", (char **)argv);
233 
234 		/*
235 		 * Tell the parent we're ready to go
236 		 */
237 		close(child_ready_pipe[1]);
238 
239 		/*
240 		 * Wait until the parent tells us to go.
241 		 */
242 		if (read(go_pipe[0], &buf, 1) == -1)
243 			perror("unable to read pipe");
244 
245 		execvp(argv[0], (char **)argv);
246 
247 		perror(argv[0]);
248 		exit(-1);
249 	}
250 
251 	/*
252 	 * Wait for the child to be ready to exec.
253 	 */
254 	close(child_ready_pipe[1]);
255 	close(go_pipe[0]);
256 	if (read(child_ready_pipe[0], &buf, 1) == -1)
257 		perror("unable to read pipe");
258 	close(child_ready_pipe[0]);
259 
260 	for (counter = 0; counter < nr_counters; counter++)
261 		create_perf_stat_counter(counter, pid);
262 
263 	/*
264 	 * Enable counters and exec the command:
265 	 */
266 	t0 = rdclock();
267 
268 	close(go_pipe[1]);
269 	wait(&status);
270 
271 	t1 = rdclock();
272 
273 	walltime_nsecs[run_idx] = t1 - t0;
274 
275 	for (counter = 0; counter < nr_counters; counter++)
276 		read_counter(counter);
277 
278 	return WEXITSTATUS(status);
279 }
280 
281 static void print_noise(u64 *count, u64 *noise)
282 {
283 	if (run_count > 1)
284 		fprintf(stderr, "   ( +- %7.3f%% )",
285 			(double)noise[0]/(count[0]+1)*100.0);
286 }
287 
288 static void nsec_printout(int counter, u64 *count, u64 *noise)
289 {
290 	double msecs = (double)count[0] / 1000000;
291 
292 	fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
293 
294 	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
295 		if (walltime_nsecs_avg)
296 			fprintf(stderr, " # %10.3f CPUs ",
297 				(double)count[0] / (double)walltime_nsecs_avg);
298 	}
299 	print_noise(count, noise);
300 }
301 
302 static void abs_printout(int counter, u64 *count, u64 *noise)
303 {
304 	fprintf(stderr, " %14Ld  %-24s", count[0], event_name(counter));
305 
306 	if (runtime_cycles_avg &&
307 	    MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
308 		fprintf(stderr, " # %10.3f IPC  ",
309 			(double)count[0] / (double)runtime_cycles_avg);
310 	} else {
311 		if (runtime_nsecs_avg) {
312 			fprintf(stderr, " # %10.3f M/sec",
313 				(double)count[0]/runtime_nsecs_avg*1000.0);
314 		}
315 	}
316 	print_noise(count, noise);
317 }
318 
319 /*
320  * Print out the results of a single counter:
321  */
322 static void print_counter(int counter)
323 {
324 	u64 *count, *noise;
325 	int scaled;
326 
327 	count = event_res_avg[counter];
328 	noise = event_res_noise[counter];
329 	scaled = event_scaled_avg[counter];
330 
331 	if (scaled == -1) {
332 		fprintf(stderr, " %14s  %-24s\n",
333 			"<not counted>", event_name(counter));
334 		return;
335 	}
336 
337 	if (nsec_counter(counter))
338 		nsec_printout(counter, count, noise);
339 	else
340 		abs_printout(counter, count, noise);
341 
342 	if (scaled)
343 		fprintf(stderr, "  (scaled from %.2f%%)",
344 			(double) count[2] / count[1] * 100);
345 
346 	fprintf(stderr, "\n");
347 }
348 
349 /*
350  * normalize_noise noise values down to stddev:
351  */
352 static void normalize_noise(u64 *val)
353 {
354 	double res;
355 
356 	res = (double)*val / (run_count * sqrt((double)run_count));
357 
358 	*val = (u64)res;
359 }
360 
361 static void update_avg(const char *name, int idx, u64 *avg, u64 *val)
362 {
363 	*avg += *val;
364 
365 	if (verbose > 1)
366 		fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
367 }
368 /*
369  * Calculate the averages and noises:
370  */
371 static void calc_avg(void)
372 {
373 	int i, j;
374 
375 	if (verbose > 1)
376 		fprintf(stderr, "\n");
377 
378 	for (i = 0; i < run_count; i++) {
379 		update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
380 		update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
381 		update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
382 
383 		for (j = 0; j < nr_counters; j++) {
384 			update_avg("counter/0", j,
385 				event_res_avg[j]+0, event_res[i][j]+0);
386 			update_avg("counter/1", j,
387 				event_res_avg[j]+1, event_res[i][j]+1);
388 			update_avg("counter/2", j,
389 				event_res_avg[j]+2, event_res[i][j]+2);
390 			if (event_scaled[i][j] != (u64)-1)
391 				update_avg("scaled", j,
392 					event_scaled_avg + j, event_scaled[i]+j);
393 			else
394 				event_scaled_avg[j] = -1;
395 		}
396 	}
397 	runtime_nsecs_avg /= run_count;
398 	walltime_nsecs_avg /= run_count;
399 	runtime_cycles_avg /= run_count;
400 
401 	for (j = 0; j < nr_counters; j++) {
402 		event_res_avg[j][0] /= run_count;
403 		event_res_avg[j][1] /= run_count;
404 		event_res_avg[j][2] /= run_count;
405 	}
406 
407 	for (i = 0; i < run_count; i++) {
408 		runtime_nsecs_noise +=
409 			abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
410 		walltime_nsecs_noise +=
411 			abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
412 		runtime_cycles_noise +=
413 			abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
414 
415 		for (j = 0; j < nr_counters; j++) {
416 			event_res_noise[j][0] +=
417 				abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
418 			event_res_noise[j][1] +=
419 				abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
420 			event_res_noise[j][2] +=
421 				abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
422 		}
423 	}
424 
425 	normalize_noise(&runtime_nsecs_noise);
426 	normalize_noise(&walltime_nsecs_noise);
427 	normalize_noise(&runtime_cycles_noise);
428 
429 	for (j = 0; j < nr_counters; j++) {
430 		normalize_noise(&event_res_noise[j][0]);
431 		normalize_noise(&event_res_noise[j][1]);
432 		normalize_noise(&event_res_noise[j][2]);
433 	}
434 }
435 
436 static void print_stat(int argc, const char **argv)
437 {
438 	int i, counter;
439 
440 	calc_avg();
441 
442 	fflush(stdout);
443 
444 	fprintf(stderr, "\n");
445 	fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
446 
447 	for (i = 1; i < argc; i++)
448 		fprintf(stderr, " %s", argv[i]);
449 
450 	fprintf(stderr, "\'");
451 	if (run_count > 1)
452 		fprintf(stderr, " (%d runs)", run_count);
453 	fprintf(stderr, ":\n\n");
454 
455 	for (counter = 0; counter < nr_counters; counter++)
456 		print_counter(counter);
457 
458 	fprintf(stderr, "\n");
459 	fprintf(stderr, " %14.9f  seconds time elapsed",
460 			(double)walltime_nsecs_avg/1e9);
461 	if (run_count > 1) {
462 		fprintf(stderr, "   ( +- %7.3f%% )",
463 			100.0*(double)walltime_nsecs_noise/(double)walltime_nsecs_avg);
464 	}
465 	fprintf(stderr, "\n\n");
466 }
467 
468 static volatile int signr = -1;
469 
470 static void skip_signal(int signo)
471 {
472 	signr = signo;
473 }
474 
475 static void sig_atexit(void)
476 {
477 	if (signr == -1)
478 		return;
479 
480 	signal(signr, SIG_DFL);
481 	kill(getpid(), signr);
482 }
483 
484 static const char * const stat_usage[] = {
485 	"perf stat [<options>] <command>",
486 	NULL
487 };
488 
489 static const struct option options[] = {
490 	OPT_CALLBACK('e', "event", NULL, "event",
491 		     "event selector. use 'perf list' to list available events",
492 		     parse_events),
493 	OPT_BOOLEAN('i', "inherit", &inherit,
494 		    "child tasks inherit counters"),
495 	OPT_INTEGER('p', "pid", &target_pid,
496 		    "stat events on existing pid"),
497 	OPT_BOOLEAN('a', "all-cpus", &system_wide,
498 		    "system-wide collection from all CPUs"),
499 	OPT_BOOLEAN('S', "scale", &scale,
500 		    "scale/normalize counters"),
501 	OPT_BOOLEAN('v', "verbose", &verbose,
502 		    "be more verbose (show counter open errors, etc)"),
503 	OPT_INTEGER('r', "repeat", &run_count,
504 		    "repeat command and print average + stddev (max: 100)"),
505 	OPT_BOOLEAN('n', "null", &null_run,
506 		    "null run - dont start any counters"),
507 	OPT_END()
508 };
509 
510 int cmd_stat(int argc, const char **argv, const char *prefix __used)
511 {
512 	int status;
513 
514 	argc = parse_options(argc, argv, options, stat_usage,
515 		PARSE_OPT_STOP_AT_NON_OPTION);
516 	if (!argc)
517 		usage_with_options(stat_usage, options);
518 	if (run_count <= 0 || run_count > MAX_RUN)
519 		usage_with_options(stat_usage, options);
520 
521 	/* Set attrs and nr_counters if no event is selected and !null_run */
522 	if (!null_run && !nr_counters) {
523 		memcpy(attrs, default_attrs, sizeof(default_attrs));
524 		nr_counters = ARRAY_SIZE(default_attrs);
525 	}
526 
527 	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
528 	assert(nr_cpus <= MAX_NR_CPUS);
529 	assert((int)nr_cpus >= 0);
530 
531 	/*
532 	 * We dont want to block the signals - that would cause
533 	 * child tasks to inherit that and Ctrl-C would not work.
534 	 * What we want is for Ctrl-C to work in the exec()-ed
535 	 * task, but being ignored by perf stat itself:
536 	 */
537 	atexit(sig_atexit);
538 	signal(SIGINT,  skip_signal);
539 	signal(SIGALRM, skip_signal);
540 	signal(SIGABRT, skip_signal);
541 
542 	status = 0;
543 	for (run_idx = 0; run_idx < run_count; run_idx++) {
544 		if (run_count != 1 && verbose)
545 			fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
546 		status = run_perf_stat(argc, argv);
547 	}
548 
549 	print_stat(argc, argv);
550 
551 	return status;
552 }
553