builtin-stat.c 13.0 KB
Newer Older
1
/*
2 3 4 5 6 7
 * builtin-stat.c
 *
 * Builtin stat command: Give a precise performance counters summary
 * overview about any workload, CPU or specific PID.
 *
 * Sample output:
8

9 10
   $ perf stat ~/hackbench 10
   Time: 0.104
11

12
    Performance counter stats for '/home/mingo/hackbench':
13

14 15 16 17 18 19 20 21
       1255.538611  task clock ticks     #      10.143 CPU utilization factor
             54011  context switches     #       0.043 M/sec
               385  CPU migrations       #       0.000 M/sec
             17755  pagefaults           #       0.014 M/sec
        3808323185  CPU cycles           #    3033.219 M/sec
        1575111190  instructions         #    1254.530 M/sec
          17367895  cache references     #      13.833 M/sec
           7674421  cache misses         #       6.112 M/sec
22

23
    Wall-clock time elapsed:   123.786620 msecs
24

25 26 27 28 29 30 31 32 33 34
 *
 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
 *
 * Improvements and fixes by:
 *
 *   Arjan van de Ven <arjan@linux.intel.com>
 *   Yanmin Zhang <yanmin.zhang@intel.com>
 *   Wu Fengguang <fengguang.wu@intel.com>
 *   Mike Galbraith <efault@gmx.de>
 *   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 44
#include "util/parse-options.h"
#include "util/parse-events.h"
45 46
#include "util/event.h"
#include "util/debug.h"
47
#include "util/header.h"
48
#include "util/cpumap.h"
49 50

#include <sys/prctl.h>
51
#include <math.h>
52

53
static struct perf_event_attr default_attrs[] = {
54

55 56 57 58
  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_TASK_CLOCK		},
  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CONTEXT_SWITCHES	},
  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_CPU_MIGRATIONS		},
  { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS		},
59

60 61 62 63
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES		},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS		},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS	},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES		},
I
Ingo Molnar 已提交
64 65
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES	},
  { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES		},
66

67
};
68

69
static int			system_wide			=  0;
70
static unsigned int		nr_cpus				=  0;
71
static int			run_idx				=  0;
72

73 74
static int			run_count			=  1;
static int			inherit				=  1;
75
static int			scale				=  1;
76 77
static pid_t			target_pid			= -1;
static pid_t			child_pid			= -1;
78
static int			null_run			=  0;
79

80
static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
81

82
static int			event_scaled[MAX_COUNTERS];
83

84 85
static volatile int done = 0;

86 87
struct stats
{
88
	double n, mean, M2;
89
};
90

91 92
static void update_stats(struct stats *stats, u64 val)
{
93
	double delta;
94

95 96 97 98
	stats->n++;
	delta = val - stats->mean;
	stats->mean += delta / stats->n;
	stats->M2 += delta*(val - stats->mean);
99 100
}

101 102
static double avg_stats(struct stats *stats)
{
103
	return stats->mean;
104
}
105

106
/*
107 108
 * http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
 *
109 110 111
 *       (\Sum n_i^2) - ((\Sum n_i)^2)/n
 * s^2 = -------------------------------
 *                  n - 1
112 113 114 115 116 117 118 119 120
 *
 * http://en.wikipedia.org/wiki/Stddev
 *
 * The std dev of the mean is related to the std dev by:
 *
 *             s
 * s_mean = -------
 *          sqrt(n)
 *
121 122 123
 */
static double stddev_stats(struct stats *stats)
{
124 125
	double variance = stats->M2 / (stats->n - 1);
	double variance_mean = variance / stats->n;
126

127
	return sqrt(variance_mean);
128
}
129

130 131 132 133
struct stats			event_res_stats[MAX_COUNTERS][3];
struct stats			runtime_nsecs_stats;
struct stats			walltime_nsecs_stats;
struct stats			runtime_cycles_stats;
134
struct stats			runtime_branches_stats;
135

136 137 138 139
#define MATCH_EVENT(t, c, counter)			\
	(attrs[counter].type == PERF_TYPE_##t &&	\
	 attrs[counter].config == PERF_COUNT_##c)

140
#define ERR_PERF_OPEN \
141
"Error: counter %d, sys_perf_event_open() syscall returned with %d (%s)\n"
142

143
static void create_perf_stat_counter(int counter, int pid)
144
{
145
	struct perf_event_attr *attr = attrs + counter;
146

147
	if (scale)
148 149
		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
				    PERF_FORMAT_TOTAL_TIME_RUNNING;
150 151

	if (system_wide) {
152 153
		unsigned int cpu;

154
		for (cpu = 0; cpu < nr_cpus; cpu++) {
155
			fd[cpu][counter] = sys_perf_event_open(attr, -1, cpumap[cpu], -1, 0);
156 157 158
			if (fd[cpu][counter] < 0 && verbose)
				fprintf(stderr, ERR_PERF_OPEN, counter,
					fd[cpu][counter], strerror(errno));
159 160
		}
	} else {
161 162 163
		attr->inherit	     = inherit;
		attr->disabled	     = 1;
		attr->enable_on_exec = 1;
164

165
		fd[0][counter] = sys_perf_event_open(attr, pid, -1, -1, 0);
166 167 168
		if (fd[0][counter] < 0 && verbose)
			fprintf(stderr, ERR_PERF_OPEN, counter,
				fd[0][counter], strerror(errno));
169 170 171
	}
}

172 173 174 175 176
/*
 * Does the counter have nsecs as a unit?
 */
static inline int nsec_counter(int counter)
{
177 178
	if (MATCH_EVENT(SOFTWARE, SW_CPU_CLOCK, counter) ||
	    MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
179 180 181 182 183 184
		return 1;

	return 0;
}

/*
185
 * Read out the results of a single counter:
186
 */
187
static void read_counter(int counter)
188
{
189
	u64 count[3], single_count[3];
190 191
	unsigned int cpu;
	size_t res, nv;
192
	int scaled;
193
	int i;
194 195

	count[0] = count[1] = count[2] = 0;
196

197
	nv = scale ? 3 : 1;
198
	for (cpu = 0; cpu < nr_cpus; cpu++) {
199 200 201
		if (fd[cpu][counter] < 0)
			continue;

202 203
		res = read(fd[cpu][counter], single_count, nv * sizeof(u64));
		assert(res == nv * sizeof(u64));
204

205 206
		close(fd[cpu][counter]);
		fd[cpu][counter] = -1;
207 208 209 210 211 212 213 214 215 216 217

		count[0] += single_count[0];
		if (scale) {
			count[1] += single_count[1];
			count[2] += single_count[2];
		}
	}

	scaled = 0;
	if (scale) {
		if (count[2] == 0) {
218
			event_scaled[counter] = -1;
219
			count[0] = 0;
220 221
			return;
		}
222

223
		if (count[2] < count[1]) {
224
			event_scaled[counter] = 1;
225 226 227 228
			count[0] = (unsigned long long)
				((double)count[0] * count[1] / count[2] + 0.5);
		}
	}
229 230 231 232 233 234 235 236 237

	for (i = 0; i < 3; i++)
		update_stats(&event_res_stats[counter][i], count[i]);

	if (verbose) {
		fprintf(stderr, "%s: %Ld %Ld %Ld\n", event_name(counter),
				count[0], count[1], count[2]);
	}

238 239 240
	/*
	 * Save the full runtime - to allow normalization during printout:
	 */
241
	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter))
242
		update_stats(&runtime_nsecs_stats, count[0]);
243
	if (MATCH_EVENT(HARDWARE, HW_CPU_CYCLES, counter))
244
		update_stats(&runtime_cycles_stats, count[0]);
245 246
	if (MATCH_EVENT(HARDWARE, HW_BRANCH_INSTRUCTIONS, counter))
		update_stats(&runtime_branches_stats, count[0]);
247 248
}

249
static int run_perf_stat(int argc __used, const char **argv)
250 251 252 253
{
	unsigned long long t0, t1;
	int status = 0;
	int counter;
254
	int pid = target_pid;
255
	int child_ready_pipe[2], go_pipe[2];
256
	const bool forks = (target_pid == -1 && argc > 0);
257
	char buf;
258 259 260 261

	if (!system_wide)
		nr_cpus = 1;

262
	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
263 264 265 266
		perror("failed to create pipes");
		exit(1);
	}

267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298
	if (forks) {
		if ((pid = fork()) < 0)
			perror("failed to fork");

		if (!pid) {
			close(child_ready_pipe[0]);
			close(go_pipe[1]);
			fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);

			/*
			 * Do a dummy execvp to get the PLT entry resolved,
			 * so we avoid the resolver overhead on the real
			 * execvp call.
			 */
			execvp("", (char **)argv);

			/*
			 * Tell the parent we're ready to go
			 */
			close(child_ready_pipe[1]);

			/*
			 * Wait until the parent tells us to go.
			 */
			if (read(go_pipe[0], &buf, 1) == -1)
				perror("unable to read pipe");

			execvp(argv[0], (char **)argv);

			perror(argv[0]);
			exit(-1);
		}
299

300
		child_pid = pid;
301 302

		/*
303
		 * Wait for the child to be ready to exec.
304 305
		 */
		close(child_ready_pipe[1]);
306 307
		close(go_pipe[0]);
		if (read(child_ready_pipe[0], &buf, 1) == -1)
308
			perror("unable to read pipe");
309
		close(child_ready_pipe[0]);
310 311
	}

312
	for (counter = 0; counter < nr_counters; counter++)
313
		create_perf_stat_counter(counter, pid);
314 315 316 317 318 319

	/*
	 * Enable counters and exec the command:
	 */
	t0 = rdclock();

320 321 322 323 324 325
	if (forks) {
		close(go_pipe[1]);
		wait(&status);
	} else {
		while(!done);
	}
326 327 328

	t1 = rdclock();

329
	update_stats(&walltime_nsecs_stats, t1 - t0);
330 331 332 333 334 335 336

	for (counter = 0; counter < nr_counters; counter++)
		read_counter(counter);

	return WEXITSTATUS(status);
}

337
static void print_noise(int counter, double avg)
338
{
339 340 341 342 343
	if (run_count == 1)
		return;

	fprintf(stderr, "   ( +- %7.3f%% )",
			100 * stddev_stats(&event_res_stats[counter][0]) / avg);
344 345
}

346
static void nsec_printout(int counter, double avg)
I
Ingo Molnar 已提交
347
{
348
	double msecs = avg / 1e6;
I
Ingo Molnar 已提交
349

350
	fprintf(stderr, " %14.6f  %-24s", msecs, event_name(counter));
I
Ingo Molnar 已提交
351

352
	if (MATCH_EVENT(SOFTWARE, SW_TASK_CLOCK, counter)) {
353 354
		fprintf(stderr, " # %10.3f CPUs ",
				avg / avg_stats(&walltime_nsecs_stats));
I
Ingo Molnar 已提交
355 356 357
	}
}

358
static void abs_printout(int counter, double avg)
I
Ingo Molnar 已提交
359
{
360 361
	double total, ratio = 0.0;

362
	fprintf(stderr, " %14.0f  %-24s", avg, event_name(counter));
I
Ingo Molnar 已提交
363

364
	if (MATCH_EVENT(HARDWARE, HW_INSTRUCTIONS, counter)) {
365 366 367 368 369 370
		total = avg_stats(&runtime_cycles_stats);

		if (total)
			ratio = avg / total;

		fprintf(stderr, " # %10.3f IPC  ", ratio);
371 372
	} else if (MATCH_EVENT(HARDWARE, HW_BRANCH_MISSES, counter) &&
			runtime_branches_stats.n != 0) {
373 374 375 376 377
		total = avg_stats(&runtime_branches_stats);

		if (total)
			ratio = avg * 100 / total;

I
Ingo Molnar 已提交
378
		fprintf(stderr, " # %10.3f %%    ", ratio);
379

380
	} else if (runtime_nsecs_stats.n != 0) {
381 382 383 384 385 386
		total = avg_stats(&runtime_nsecs_stats);

		if (total)
			ratio = 1000.0 * avg / total;

		fprintf(stderr, " # %10.3f M/sec", ratio);
I
Ingo Molnar 已提交
387 388 389
	}
}

390 391 392 393 394
/*
 * Print out the results of a single counter:
 */
static void print_counter(int counter)
{
395
	double avg = avg_stats(&event_res_stats[counter][0]);
396
	int scaled = event_scaled[counter];
397 398

	if (scaled == -1) {
399
		fprintf(stderr, " %14s  %-24s\n",
400 401 402
			"<not counted>", event_name(counter));
		return;
	}
403

I
Ingo Molnar 已提交
404
	if (nsec_counter(counter))
405
		nsec_printout(counter, avg);
I
Ingo Molnar 已提交
406
	else
407 408 409
		abs_printout(counter, avg);

	print_noise(counter, avg);
410 411 412 413 414 415

	if (scaled) {
		double avg_enabled, avg_running;

		avg_enabled = avg_stats(&event_res_stats[counter][1]);
		avg_running = avg_stats(&event_res_stats[counter][2]);
416

417
		fprintf(stderr, "  (scaled from %.2f%%)",
418 419
				100 * avg_running / avg_enabled);
	}
I
Ingo Molnar 已提交
420

421 422 423
	fprintf(stderr, "\n");
}

424 425 426 427
static void print_stat(int argc, const char **argv)
{
	int i, counter;

428 429 430
	fflush(stdout);

	fprintf(stderr, "\n");
431 432 433 434 435 436 437
	fprintf(stderr, " Performance counter stats for ");
	if(target_pid == -1) {
		fprintf(stderr, "\'%s", argv[0]);
		for (i = 1; i < argc; i++)
			fprintf(stderr, " %s", argv[i]);
	}else
		fprintf(stderr, "task pid \'%d", target_pid);
I
Ingo Molnar 已提交
438

439 440 441 442
	fprintf(stderr, "\'");
	if (run_count > 1)
		fprintf(stderr, " (%d runs)", run_count);
	fprintf(stderr, ":\n\n");
443

444 445
	for (counter = 0; counter < nr_counters; counter++)
		print_counter(counter);
446 447

	fprintf(stderr, "\n");
I
Ingo Molnar 已提交
448
	fprintf(stderr, " %14.9f  seconds time elapsed",
449
			avg_stats(&walltime_nsecs_stats)/1e9);
I
Ingo Molnar 已提交
450 451
	if (run_count > 1) {
		fprintf(stderr, "   ( +- %7.3f%% )",
452 453
				100*stddev_stats(&walltime_nsecs_stats) /
				avg_stats(&walltime_nsecs_stats));
I
Ingo Molnar 已提交
454 455
	}
	fprintf(stderr, "\n\n");
456 457
}

458 459
static volatile int signr = -1;

460
static void skip_signal(int signo)
461
{
462 463 464
	if(target_pid != -1)
		done = 1;

465 466 467 468 469
	signr = signo;
}

static void sig_atexit(void)
{
470 471 472
	if (child_pid != -1)
		kill(child_pid, SIGTERM);

473 474 475 476 477
	if (signr == -1)
		return;

	signal(signr, SIG_DFL);
	kill(getpid(), signr);
478 479 480
}

static const char * const stat_usage[] = {
481
	"perf stat [<options>] [<command>]",
482 483 484 485 486
	NULL
};

static const struct option options[] = {
	OPT_CALLBACK('e', "event", NULL, "event",
487 488
		     "event selector. use 'perf list' to list available events",
		     parse_events),
489 490 491 492 493
	OPT_BOOLEAN('i', "inherit", &inherit,
		    "child tasks inherit counters"),
	OPT_INTEGER('p', "pid", &target_pid,
		    "stat events on existing pid"),
	OPT_BOOLEAN('a', "all-cpus", &system_wide,
494
		    "system-wide collection from all CPUs"),
495
	OPT_BOOLEAN('c', "scale", &scale,
496
		    "scale/normalize counters"),
497 498
	OPT_BOOLEAN('v', "verbose", &verbose,
		    "be more verbose (show counter open errors, etc)"),
499 500
	OPT_INTEGER('r', "repeat", &run_count,
		    "repeat command and print average + stddev (max: 100)"),
501 502
	OPT_BOOLEAN('n', "null", &null_run,
		    "null run - dont start any counters"),
503 504 505
	OPT_END()
};

506
int cmd_stat(int argc, const char **argv, const char *prefix __used)
507
{
508 509
	int status;

510 511
	argc = parse_options(argc, argv, options, stat_usage,
		PARSE_OPT_STOP_AT_NON_OPTION);
512
	if (!argc && target_pid == -1)
513
		usage_with_options(stat_usage, options);
514
	if (run_count <= 0)
515
		usage_with_options(stat_usage, options);
516

517 518 519 520 521
	/* Set attrs and nr_counters if no event is selected and !null_run */
	if (!null_run && !nr_counters) {
		memcpy(attrs, default_attrs, sizeof(default_attrs));
		nr_counters = ARRAY_SIZE(default_attrs);
	}
522

523 524 525 526
	if (system_wide)
		nr_cpus = read_cpu_map();
	else
		nr_cpus = 1;
527

I
Ingo Molnar 已提交
528 529 530 531 532 533
	/*
	 * We dont want to block the signals - that would cause
	 * child tasks to inherit that and Ctrl-C would not work.
	 * What we want is for Ctrl-C to work in the exec()-ed
	 * task, but being ignored by perf stat itself:
	 */
534
	atexit(sig_atexit);
I
Ingo Molnar 已提交
535 536 537 538
	signal(SIGINT,  skip_signal);
	signal(SIGALRM, skip_signal);
	signal(SIGABRT, skip_signal);

539 540 541
	status = 0;
	for (run_idx = 0; run_idx < run_count; run_idx++) {
		if (run_count != 1 && verbose)
542
			fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
543 544 545 546 547 548
		status = run_perf_stat(argc, argv);
	}

	print_stat(argc, argv);

	return status;
549
}