xref: /linux/tools/perf/builtin-stat.c (revision 44eeab67416711db9b84610ef18c99a60415dff8)
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  *
36  * Released under the GPL v2. (and only v2, not any later version)
37  */
38 
39 #include "perf.h"
40 #include "builtin.h"
41 #include "util/util.h"
42 #include "util/parse-options.h"
43 #include "util/parse-events.h"
44 
45 #include <sys/prctl.h>
46 #include <math.h>
47 
48 static struct perf_counter_attr default_attrs[MAX_COUNTERS] = {
49 
50   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK	},
51   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES},
52   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS	},
53   { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS	},
54 
55   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES	},
56   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS	},
57   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES},
58   { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES	},
59 
60 };
61 
62 static int			system_wide			=  0;
63 static int			inherit				=  1;
64 static int			verbose				=  0;
65 
66 static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
67 
68 static int			target_pid			= -1;
69 static int			nr_cpus				=  0;
70 static unsigned int		page_size;
71 
72 static int			scale				=  1;
73 
74 static const unsigned int default_count[] = {
75 	1000000,
76 	1000000,
77 	  10000,
78 	  10000,
79 	1000000,
80 	  10000,
81 };
82 
83 #define MAX_RUN 100
84 
85 static int			run_count		=  1;
86 static int			run_idx			=  0;
87 
88 static u64			event_res[MAX_RUN][MAX_COUNTERS][3];
89 static u64			event_scaled[MAX_RUN][MAX_COUNTERS];
90 
91 //static u64			event_hist[MAX_RUN][MAX_COUNTERS][3];
92 
93 
94 static u64			runtime_nsecs[MAX_RUN];
95 static u64			walltime_nsecs[MAX_RUN];
96 static u64			runtime_cycles[MAX_RUN];
97 
98 static u64			event_res_avg[MAX_COUNTERS][3];
99 static u64			event_res_noise[MAX_COUNTERS][3];
100 
101 static u64			event_scaled_avg[MAX_COUNTERS];
102 
103 static u64			runtime_nsecs_avg;
104 static u64			runtime_nsecs_noise;
105 
106 static u64			walltime_nsecs_avg;
107 static u64			walltime_nsecs_noise;
108 
109 static u64			runtime_cycles_avg;
110 static u64			runtime_cycles_noise;
111 
112 static void create_perf_stat_counter(int counter)
113 {
114 	struct perf_counter_attr *attr = attrs + counter;
115 
116 	if (scale)
117 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
118 				    PERF_FORMAT_TOTAL_TIME_RUNNING;
119 
120 	if (system_wide) {
121 		int cpu;
122 		for (cpu = 0; cpu < nr_cpus; cpu ++) {
123 			fd[cpu][counter] = sys_perf_counter_open(attr, -1, cpu, -1, 0);
124 			if (fd[cpu][counter] < 0 && verbose) {
125 				printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[cpu][counter], strerror(errno));
126 			}
127 		}
128 	} else {
129 		attr->inherit	= inherit;
130 		attr->disabled	= 1;
131 
132 		fd[0][counter] = sys_perf_counter_open(attr, 0, -1, -1, 0);
133 		if (fd[0][counter] < 0 && verbose) {
134 			printf("Error: counter %d, sys_perf_counter_open() syscall returned with %d (%s)\n", counter, fd[0][counter], strerror(errno));
135 		}
136 	}
137 }
138 
139 /*
140  * Does the counter have nsecs as a unit?
141  */
142 static inline int nsec_counter(int counter)
143 {
144 	if (attrs[counter].type != PERF_TYPE_SOFTWARE)
145 		return 0;
146 
147 	if (attrs[counter].config == PERF_COUNT_SW_CPU_CLOCK)
148 		return 1;
149 
150 	if (attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
151 		return 1;
152 
153 	return 0;
154 }
155 
156 /*
157  * Read out the results of a single counter:
158  */
159 static void read_counter(int counter)
160 {
161 	u64 *count, single_count[3];
162 	ssize_t res;
163 	int cpu, nv;
164 	int scaled;
165 
166 	count = event_res[run_idx][counter];
167 
168 	count[0] = count[1] = count[2] = 0;
169 
170 	nv = scale ? 3 : 1;
171 	for (cpu = 0; cpu < nr_cpus; cpu ++) {
172 		if (fd[cpu][counter] < 0)
173 			continue;
174 
175 		res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
176 		assert(res == nv * sizeof(u64));
177 		close(fd[cpu][counter]);
178 		fd[cpu][counter] = -1;
179 
180 		count[0] += single_count[0];
181 		if (scale) {
182 			count[1] += single_count[1];
183 			count[2] += single_count[2];
184 		}
185 	}
186 
187 	scaled = 0;
188 	if (scale) {
189 		if (count[2] == 0) {
190 			event_scaled[run_idx][counter] = -1;
191 			count[0] = 0;
192 			return;
193 		}
194 
195 		if (count[2] < count[1]) {
196 			event_scaled[run_idx][counter] = 1;
197 			count[0] = (unsigned long long)
198 				((double)count[0] * count[1] / count[2] + 0.5);
199 		}
200 	}
201 	/*
202 	 * Save the full runtime - to allow normalization during printout:
203 	 */
204 	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
205 		attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK)
206 		runtime_nsecs[run_idx] = count[0];
207 	if (attrs[counter].type == PERF_TYPE_HARDWARE &&
208 		attrs[counter].config == PERF_COUNT_HW_CPU_CYCLES)
209 		runtime_cycles[run_idx] = count[0];
210 }
211 
212 static int run_perf_stat(int argc, const char **argv)
213 {
214 	unsigned long long t0, t1;
215 	int status = 0;
216 	int counter;
217 	int pid;
218 
219 	if (!system_wide)
220 		nr_cpus = 1;
221 
222 	for (counter = 0; counter < nr_counters; counter++)
223 		create_perf_stat_counter(counter);
224 
225 	/*
226 	 * Enable counters and exec the command:
227 	 */
228 	t0 = rdclock();
229 	prctl(PR_TASK_PERF_COUNTERS_ENABLE);
230 
231 	if ((pid = fork()) < 0)
232 		perror("failed to fork");
233 
234 	if (!pid) {
235 		if (execvp(argv[0], (char **)argv)) {
236 			perror(argv[0]);
237 			exit(-1);
238 		}
239 	}
240 
241 	wait(&status);
242 
243 	prctl(PR_TASK_PERF_COUNTERS_DISABLE);
244 	t1 = rdclock();
245 
246 	walltime_nsecs[run_idx] = t1 - t0;
247 
248 	for (counter = 0; counter < nr_counters; counter++)
249 		read_counter(counter);
250 
251 	return WEXITSTATUS(status);
252 }
253 
254 static void print_noise(u64 *count, u64 *noise)
255 {
256 	if (run_count > 1)
257 		fprintf(stderr, "   ( +- %7.3f%% )",
258 			(double)noise[0]/(count[0]+1)*100.0);
259 }
260 
261 static void nsec_printout(int counter, u64 *count, u64 *noise)
262 {
263 	double msecs = (double)count[0] / 1000000;
264 
265 	fprintf(stderr, " %14.6f  %-20s", msecs, event_name(counter));
266 
267 	if (attrs[counter].type == PERF_TYPE_SOFTWARE &&
268 		attrs[counter].config == PERF_COUNT_SW_TASK_CLOCK) {
269 
270 		if (walltime_nsecs_avg)
271 			fprintf(stderr, " # %10.3f CPUs ",
272 				(double)count[0] / (double)walltime_nsecs_avg);
273 	}
274 	print_noise(count, noise);
275 }
276 
277 static void abs_printout(int counter, u64 *count, u64 *noise)
278 {
279 	fprintf(stderr, " %14Ld  %-20s", count[0], event_name(counter));
280 
281 	if (runtime_cycles_avg &&
282 		attrs[counter].type == PERF_TYPE_HARDWARE &&
283 			attrs[counter].config == PERF_COUNT_HW_INSTRUCTIONS) {
284 
285 		fprintf(stderr, " # %10.3f IPC  ",
286 			(double)count[0] / (double)runtime_cycles_avg);
287 	} else {
288 		if (runtime_nsecs_avg) {
289 			fprintf(stderr, " # %10.3f M/sec",
290 				(double)count[0]/runtime_nsecs_avg*1000.0);
291 		}
292 	}
293 	print_noise(count, noise);
294 }
295 
296 /*
297  * Print out the results of a single counter:
298  */
299 static void print_counter(int counter)
300 {
301 	u64 *count, *noise;
302 	int scaled;
303 
304 	count = event_res_avg[counter];
305 	noise = event_res_noise[counter];
306 	scaled = event_scaled_avg[counter];
307 
308 	if (scaled == -1) {
309 		fprintf(stderr, " %14s  %-20s\n",
310 			"<not counted>", event_name(counter));
311 		return;
312 	}
313 
314 	if (nsec_counter(counter))
315 		nsec_printout(counter, count, noise);
316 	else
317 		abs_printout(counter, count, noise);
318 
319 	if (scaled)
320 		fprintf(stderr, "  (scaled from %.2f%%)",
321 			(double) count[2] / count[1] * 100);
322 
323 	fprintf(stderr, "\n");
324 }
325 
326 /*
327  * normalize_noise noise values down to stddev:
328  */
329 static void normalize_noise(u64 *val)
330 {
331 	double res;
332 
333 	res = (double)*val / (run_count * sqrt((double)run_count));
334 
335 	*val = (u64)res;
336 }
337 
338 static void update_avg(const char *name, int idx, u64 *avg, u64 *val)
339 {
340 	*avg += *val;
341 
342 	if (verbose > 1)
343 		fprintf(stderr, "debug: %20s[%d]: %Ld\n", name, idx, *val);
344 }
345 /*
346  * Calculate the averages and noises:
347  */
348 static void calc_avg(void)
349 {
350 	int i, j;
351 
352 	if (verbose > 1)
353 		fprintf(stderr, "\n");
354 
355 	for (i = 0; i < run_count; i++) {
356 		update_avg("runtime", 0, &runtime_nsecs_avg, runtime_nsecs + i);
357 		update_avg("walltime", 0, &walltime_nsecs_avg, walltime_nsecs + i);
358 		update_avg("runtime_cycles", 0, &runtime_cycles_avg, runtime_cycles + i);
359 
360 		for (j = 0; j < nr_counters; j++) {
361 			update_avg("counter/0", j,
362 				event_res_avg[j]+0, event_res[i][j]+0);
363 			update_avg("counter/1", j,
364 				event_res_avg[j]+1, event_res[i][j]+1);
365 			update_avg("counter/2", j,
366 				event_res_avg[j]+2, event_res[i][j]+2);
367 			update_avg("scaled", j,
368 				event_scaled_avg + j, event_scaled[i]+j);
369 		}
370 	}
371 	runtime_nsecs_avg /= run_count;
372 	walltime_nsecs_avg /= run_count;
373 	runtime_cycles_avg /= run_count;
374 
375 	for (j = 0; j < nr_counters; j++) {
376 		event_res_avg[j][0] /= run_count;
377 		event_res_avg[j][1] /= run_count;
378 		event_res_avg[j][2] /= run_count;
379 	}
380 
381 	for (i = 0; i < run_count; i++) {
382 		runtime_nsecs_noise +=
383 			abs((s64)(runtime_nsecs[i] - runtime_nsecs_avg));
384 		walltime_nsecs_noise +=
385 			abs((s64)(walltime_nsecs[i] - walltime_nsecs_avg));
386 		runtime_cycles_noise +=
387 			abs((s64)(runtime_cycles[i] - runtime_cycles_avg));
388 
389 		for (j = 0; j < nr_counters; j++) {
390 			event_res_noise[j][0] +=
391 				abs((s64)(event_res[i][j][0] - event_res_avg[j][0]));
392 			event_res_noise[j][1] +=
393 				abs((s64)(event_res[i][j][1] - event_res_avg[j][1]));
394 			event_res_noise[j][2] +=
395 				abs((s64)(event_res[i][j][2] - event_res_avg[j][2]));
396 		}
397 	}
398 
399 	normalize_noise(&runtime_nsecs_noise);
400 	normalize_noise(&walltime_nsecs_noise);
401 	normalize_noise(&runtime_cycles_noise);
402 
403 	for (j = 0; j < nr_counters; j++) {
404 		normalize_noise(&event_res_noise[j][0]);
405 		normalize_noise(&event_res_noise[j][1]);
406 		normalize_noise(&event_res_noise[j][2]);
407 	}
408 }
409 
410 static void print_stat(int argc, const char **argv)
411 {
412 	int i, counter;
413 
414 	calc_avg();
415 
416 	fflush(stdout);
417 
418 	fprintf(stderr, "\n");
419 	fprintf(stderr, " Performance counter stats for \'%s", argv[0]);
420 
421 	for (i = 1; i < argc; i++)
422 		fprintf(stderr, " %s", argv[i]);
423 
424 	fprintf(stderr, "\'");
425 	if (run_count > 1)
426 		fprintf(stderr, " (%d runs)", run_count);
427 	fprintf(stderr, ":\n\n");
428 
429 	for (counter = 0; counter < nr_counters; counter++)
430 		print_counter(counter);
431 
432 
433 	fprintf(stderr, "\n");
434 	fprintf(stderr, " %14.9f  seconds time elapsed.\n",
435 			(double)walltime_nsecs_avg/1e9);
436 	fprintf(stderr, "\n");
437 }
438 
439 static volatile int signr = -1;
440 
441 static void skip_signal(int signo)
442 {
443 	signr = signo;
444 }
445 
446 static void sig_atexit(void)
447 {
448 	if (signr == -1)
449 		return;
450 
451 	signal(signr, SIG_DFL);
452 	kill(getpid(), signr);
453 }
454 
455 static const char * const stat_usage[] = {
456 	"perf stat [<options>] <command>",
457 	NULL
458 };
459 
460 static const struct option options[] = {
461 	OPT_CALLBACK('e', "event", NULL, "event",
462 		     "event selector. use 'perf list' to list available events",
463 		     parse_events),
464 	OPT_BOOLEAN('i', "inherit", &inherit,
465 		    "child tasks inherit counters"),
466 	OPT_INTEGER('p', "pid", &target_pid,
467 		    "stat events on existing pid"),
468 	OPT_BOOLEAN('a', "all-cpus", &system_wide,
469 			    "system-wide collection from all CPUs"),
470 	OPT_BOOLEAN('S', "scale", &scale,
471 			    "scale/normalize counters"),
472 	OPT_BOOLEAN('v', "verbose", &verbose,
473 		    "be more verbose (show counter open errors, etc)"),
474 	OPT_INTEGER('r', "repeat", &run_count,
475 		    "repeat command and print average + stddev (max: 100)"),
476 	OPT_END()
477 };
478 
479 int cmd_stat(int argc, const char **argv, const char *prefix)
480 {
481 	int status;
482 
483 	page_size = sysconf(_SC_PAGE_SIZE);
484 
485 	memcpy(attrs, default_attrs, sizeof(attrs));
486 
487 	argc = parse_options(argc, argv, options, stat_usage, 0);
488 	if (!argc)
489 		usage_with_options(stat_usage, options);
490 	if (run_count <= 0 || run_count > MAX_RUN)
491 		usage_with_options(stat_usage, options);
492 
493 	if (!nr_counters)
494 		nr_counters = 8;
495 
496 	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
497 	assert(nr_cpus <= MAX_NR_CPUS);
498 	assert(nr_cpus >= 0);
499 
500 	/*
501 	 * We dont want to block the signals - that would cause
502 	 * child tasks to inherit that and Ctrl-C would not work.
503 	 * What we want is for Ctrl-C to work in the exec()-ed
504 	 * task, but being ignored by perf stat itself:
505 	 */
506 	atexit(sig_atexit);
507 	signal(SIGINT,  skip_signal);
508 	signal(SIGALRM, skip_signal);
509 	signal(SIGABRT, skip_signal);
510 
511 	status = 0;
512 	for (run_idx = 0; run_idx < run_count; run_idx++) {
513 		if (run_count != 1 && verbose)
514 			fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx+1);
515 		status = run_perf_stat(argc, argv);
516 	}
517 
518 	print_stat(argc, argv);
519 
520 	return status;
521 }
522