builtin-record.c 14.6 KB
Newer Older
I
Ingo Molnar 已提交
1
/*
2 3 4 5 6
 * builtin-record.c
 *
 * Builtin record command: Record the profile of a workload
 * (or a CPU, or a PID) into the perf.data output file - for
 * later analysis via perf report.
I
Ingo Molnar 已提交
7
 */
8
#include "builtin.h"
9 10 11

#include "perf.h"

12
#include "util/util.h"
13
#include "util/parse-options.h"
14
#include "util/parse-events.h"
15
#include "util/string.h"
16

17
#include "util/header.h"
18
#include "util/event.h"
19
#include "util/debug.h"
20
#include "util/symbol.h"
21

22
#include <unistd.h>
23 24 25
#include <sched.h>

static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
26

27
static long			default_interval		=      0;
28

29
static int			nr_cpus				=      0;
30
static unsigned int		page_size;
31 32
static unsigned int		mmap_pages			=    128;
static int			freq				=   1000;
33
static int			output;
34
static const char		*output_name			= "perf.data";
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52
static int			group				=      0;
static unsigned int		realtime_prio			=      0;
static int			raw_samples			=      0;
static int			system_wide			=      0;
static int			profile_cpu			=     -1;
static pid_t			target_pid			=     -1;
static pid_t			child_pid			=     -1;
static int			inherit				=      1;
static int			force				=      0;
static int			append_file			=      0;
static int			call_graph			=      0;
static int			inherit_stat			=      0;
static int			no_samples			=      0;
static int			sample_address			=      0;
static int			multiplex			=      0;
static int			multiplex_fd			=     -1;

static long			samples				=      0;
53 54 55
static struct timeval		last_read;
static struct timeval		this_read;

56
static u64			bytes_written			=      0;
57 58 59

static struct pollfd		event_array[MAX_NR_CPUS * MAX_COUNTERS];

60 61
static int			nr_poll				=      0;
static int			nr_cpu				=      0;
62

63
static int			file_new			=      1;
64

65
struct perf_header		*header				=   NULL;
66

67
struct mmap_data {
68 69 70 71
	int			counter;
	void			*base;
	unsigned int		mask;
	unsigned int		prev;
72 73
};

74 75
static struct mmap_data		mmap_array[MAX_NR_CPUS][MAX_COUNTERS];

76
static unsigned long mmap_read_head(struct mmap_data *md)
77
{
78
	struct perf_event_mmap_page *pc = md->base;
79
	long head;
80 81 82 83 84 85 86

	head = pc->data_head;
	rmb();

	return head;
}

87 88
static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
{
89
	struct perf_event_mmap_page *pc = md->base;
90 91 92 93 94 95 96 97

	/*
	 * ensure all reads are done before we write the tail out.
	 */
	/* mb(); */
	pc->data_tail = tail;
}

98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
static void write_output(void *buf, size_t size)
{
	while (size) {
		int ret = write(output, buf, size);

		if (ret < 0)
			die("failed to write");

		size -= ret;
		buf += ret;

		bytes_written += ret;
	}
}

113 114 115 116 117 118 119 120 121 122 123 124
static void write_event(event_t *buf, size_t size)
{
	/*
	* Add it to the list of DSOs, so that when we finish this
	 * record session we can pick the available build-ids.
	 */
	if (buf->header.type == PERF_RECORD_MMAP)
		dsos__findnew(buf->mmap.filename);

	write_output(buf, size);
}

125 126
static int process_synthesized_event(event_t *event)
{
127
	write_event(event, event->header.size);
128 129 130
	return 0;
}

131 132 133 134 135 136 137 138 139 140 141 142 143
static void mmap_read(struct mmap_data *md)
{
	unsigned int head = mmap_read_head(md);
	unsigned int old = md->prev;
	unsigned char *data = md->base + page_size;
	unsigned long size;
	void *buf;
	int diff;

	gettimeofday(&this_read, NULL);

	/*
	 * If we're further behind than half the buffer, there's a chance
144
	 * the writer will bite our tail and mess up the samples under us.
145 146 147 148 149 150
	 *
	 * If we somehow ended up ahead of the head, we got messed up.
	 *
	 * In either case, truncate and restart at head.
	 */
	diff = head - old;
151
	if (diff < 0) {
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169
		struct timeval iv;
		unsigned long msecs;

		timersub(&this_read, &last_read, &iv);
		msecs = iv.tv_sec*1000 + iv.tv_usec/1000;

		fprintf(stderr, "WARNING: failed to keep up with mmap data."
				"  Last read %lu msecs ago.\n", msecs);

		/*
		 * head points to a known good entry, start there.
		 */
		old = head;
	}

	last_read = this_read;

	if (old != head)
170
		samples++;
171 172 173 174 175 176 177

	size = head - old;

	if ((old & md->mask) + size != (head & md->mask)) {
		buf = &data[old & md->mask];
		size = md->mask + 1 - (old & md->mask);
		old += size;
178

179
		write_event(buf, size);
180 181 182 183 184
	}

	buf = &data[old & md->mask];
	size = head - old;
	old += size;
185

186
	write_event(buf, size);
187 188

	md->prev = old;
189
	mmap_write_tail(md, old);
190 191 192
}

static volatile int done = 0;
193
static volatile int signr = -1;
194

195
static void sig_handler(int sig)
196
{
197
	done = 1;
198 199 200 201 202
	signr = sig;
}

static void sig_atexit(void)
{
203 204 205
	if (child_pid != -1)
		kill(child_pid, SIGTERM);

206 207 208 209 210
	if (signr == -1)
		return;

	signal(signr, SIG_DFL);
	kill(getpid(), signr);
211 212
}

213 214
static int group_fd;

215
static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
216 217 218 219 220 221 222
{
	struct perf_header_attr *h_attr;

	if (nr < header->attrs) {
		h_attr = header->attr[nr];
	} else {
		h_attr = perf_header_attr__new(a);
223
		if (h_attr != NULL)
224 225 226 227
			if (perf_header__add_attr(header, h_attr) < 0) {
				perf_header_attr__delete(h_attr);
				h_attr = NULL;
			}
228 229 230 231 232
	}

	return h_attr;
}

233
static void create_counter(int counter, int cpu, pid_t pid)
234
{
L
Li Zefan 已提交
235
	char *filter = filters[counter];
236
	struct perf_event_attr *attr = attrs + counter;
237 238
	struct perf_header_attr *h_attr;
	int track = !counter; /* only the first counter needs these */
L
Li Zefan 已提交
239
	int ret;
240 241 242 243 244 245 246 247 248 249
	struct {
		u64 count;
		u64 time_enabled;
		u64 time_running;
		u64 id;
	} read_data;

	attr->read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
				  PERF_FORMAT_TOTAL_TIME_RUNNING |
				  PERF_FORMAT_ID;
250

251
	attr->sample_type	|= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
252

253
	if (freq) {
254
		attr->sample_type	|= PERF_SAMPLE_PERIOD;
255 256
		attr->freq		= 1;
		attr->sample_freq	= freq;
257
	}
258

259 260 261 262 263 264
	if (no_samples)
		attr->sample_freq = 0;

	if (inherit_stat)
		attr->inherit_stat = 1;

265 266 267
	if (sample_address)
		attr->sample_type	|= PERF_SAMPLE_ADDR;

268 269 270
	if (call_graph)
		attr->sample_type	|= PERF_SAMPLE_CALLCHAIN;

I
Ingo Molnar 已提交
271
	if (raw_samples) {
272
		attr->sample_type	|= PERF_SAMPLE_TIME;
273
		attr->sample_type	|= PERF_SAMPLE_RAW;
I
Ingo Molnar 已提交
274 275
		attr->sample_type	|= PERF_SAMPLE_CPU;
	}
276

277 278 279
	attr->mmap		= track;
	attr->comm		= track;
	attr->inherit		= (cpu < 0) && inherit;
280
	attr->disabled		= 1;
281

282
try_again:
283
	fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
284

285 286
	if (fd[nr_cpu][counter] < 0) {
		int err = errno;
287

P
Pekka Enberg 已提交
288
		if (err == EPERM || err == EACCES)
289
			die("Permission error - are you root?\n");
290 291
		else if (err ==  ENODEV && profile_cpu != -1)
			die("No such device - did you specify an out-of-range profile CPU?\n");
292 293 294 295 296 297 298

		/*
		 * If it's cycles then fall back to hrtimer
		 * based cpu-clock-tick sw counter, which
		 * is always available even if no PMU support:
		 */
		if (attr->type == PERF_TYPE_HARDWARE
299
			&& attr->config == PERF_COUNT_HW_CPU_CYCLES) {
300 301 302 303

			if (verbose)
				warning(" ... trying to fall back to cpu-clock-ticks\n");
			attr->type = PERF_TYPE_SOFTWARE;
304
			attr->config = PERF_COUNT_SW_CPU_CLOCK;
305 306
			goto try_again;
		}
307 308 309
		printf("\n");
		error("perfcounter syscall returned with %d (%s)\n",
			fd[nr_cpu][counter], strerror(err));
310
		die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
311 312
		exit(-1);
	}
313

314
	h_attr = get_header_attr(attr, counter);
315 316
	if (h_attr == NULL)
		die("nomem\n");
317 318 319 320 321 322 323 324

	if (!file_new) {
		if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
			fprintf(stderr, "incompatible append\n");
			exit(-1);
		}
	}

325 326 327 328
	if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
		perror("Unable to read perf file descriptor\n");
		exit(-1);
	}
329

330 331 332 333
	if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
		pr_warning("Not enough memory to add id\n");
		exit(-1);
	}
334

335 336
	assert(fd[nr_cpu][counter] >= 0);
	fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
337

338 339 340 341 342
	/*
	 * First counter acts as the group leader:
	 */
	if (group && group_fd == -1)
		group_fd = fd[nr_cpu][counter];
343 344
	if (multiplex && multiplex_fd == -1)
		multiplex_fd = fd[nr_cpu][counter];
345

346
	if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
347

348
		ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364
		assert(ret != -1);
	} else {
		event_array[nr_poll].fd = fd[nr_cpu][counter];
		event_array[nr_poll].events = POLLIN;
		nr_poll++;

		mmap_array[nr_cpu][counter].counter = counter;
		mmap_array[nr_cpu][counter].prev = 0;
		mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
		mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
				PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
		if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
			error("failed to mmap with %d (%s)\n", errno, strerror(errno));
			exit(-1);
		}
	}
365

L
Li Zefan 已提交
366 367 368 369 370 371 372 373 374 375
	if (filter != NULL) {
		ret = ioctl(fd[nr_cpu][counter],
			    PERF_EVENT_IOC_SET_FILTER, filter);
		if (ret) {
			error("failed to set filter with %d (%s)\n", errno,
			      strerror(errno));
			exit(-1);
		}
	}

376
	ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
377
}
378

379 380 381
static void open_counters(int cpu, pid_t pid)
{
	int counter;
382

383 384 385 386
	group_fd = -1;
	for (counter = 0; counter < nr_counters; counter++)
		create_counter(counter, cpu, pid);

387 388 389
	nr_cpu++;
}

390 391
static void atexit_header(void)
{
392
	header->data_size += bytes_written;
393

394
	perf_header__write(header, output, true);
395 396
}

397
static int __cmd_record(int argc, const char **argv)
398 399
{
	int i, counter;
I
Ingo Molnar 已提交
400
	struct stat st;
401
	pid_t pid = 0;
I
Ingo Molnar 已提交
402
	int flags;
403
	int ret;
404
	unsigned long waking = 0;
405 406 407 408 409 410

	page_size = sysconf(_SC_PAGE_SIZE);
	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
	assert(nr_cpus <= MAX_NR_CPUS);
	assert(nr_cpus >= 0);

411 412 413 414
	atexit(sig_atexit);
	signal(SIGCHLD, sig_handler);
	signal(SIGINT, sig_handler);

415 416 417 418 419 420 421 422
	if (!stat(output_name, &st) && st.st_size) {
		if (!force && !append_file) {
			fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
					output_name);
			exit(-1);
		}
	} else {
		append_file = 0;
423 424
	}

I
Ingo Molnar 已提交
425 426
	flags = O_CREAT|O_RDWR;
	if (append_file)
427
		file_new = 0;
I
Ingo Molnar 已提交
428 429 430 431
	else
		flags |= O_TRUNC;

	output = open(output_name, flags, S_IRUSR|S_IWUSR);
432 433 434 435 436
	if (output < 0) {
		perror("failed to create output file");
		exit(-1);
	}

437 438 439 440
	if (!file_new)
		header = perf_header__read(output);
	else
		header = perf_header__new();
441

442 443 444 445 446
	if (header == NULL) {
		pr_err("Not enough memory for reading perf file header\n");
		return -1;
	}

447
	if (raw_samples) {
448
		perf_header__set_feat(header, HEADER_TRACE_INFO);
449 450 451
	} else {
		for (i = 0; i < nr_counters; i++) {
			if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
452
				perf_header__set_feat(header, HEADER_TRACE_INFO);
453 454 455 456
				break;
			}
		}
	}
457

458 459
	atexit(atexit_header);

460
	if (!system_wide) {
461 462 463 464
		pid = target_pid;
		if (pid == -1)
			pid = getpid();

465 466 467 468 469 470 471 472 473
		open_counters(profile_cpu, pid);
	} else {
		if (profile_cpu != -1) {
			open_counters(profile_cpu, target_pid);
		} else {
			for (i = 0; i < nr_cpus; i++)
				open_counters(i, target_pid);
		}
	}
474

475
	if (file_new)
476
		perf_header__write(header, output, false);
477

478 479 480 481
	if (!system_wide)
		event__synthesize_thread(pid, process_synthesized_event);
	else
		event__synthesize_threads(process_synthesized_event);
482

483
	if (target_pid == -1 && argc) {
484 485
		pid = fork();
		if (pid < 0)
486
			die("failed to fork");
487

488
		if (!pid) {
489
			if (execvp(argv[0], (char **)argv)) {
490 491 492
				perror(argv[0]);
				exit(-1);
			}
493 494 495 496 497 498 499 500 501
		} else {
			/*
			 * Wait a bit for the execv'ed child to appear
			 * and be updated in /proc
			 * FIXME: Do you know a less heuristical solution?
			 */
			usleep(1000);
			event__synthesize_thread(pid,
						 process_synthesized_event);
502
		}
503 504

		child_pid = pid;
505 506 507 508 509 510 511
	}

	if (realtime_prio) {
		struct sched_param param;

		param.sched_priority = realtime_prio;
		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
512
			pr_err("Could not set realtime priority.\n");
513 514 515 516
			exit(-1);
		}
	}

517
	for (;;) {
518
		int hits = samples;
519

520
		for (i = 0; i < nr_cpu; i++) {
521 522 523 524
			for (counter = 0; counter < nr_counters; counter++) {
				if (mmap_array[i][counter].base)
					mmap_read(&mmap_array[i][counter]);
			}
525 526
		}

527 528 529
		if (hits == samples) {
			if (done)
				break;
530 531 532 533 534 535 536
			ret = poll(event_array, nr_poll, -1);
			waking++;
		}

		if (done) {
			for (i = 0; i < nr_cpu; i++) {
				for (counter = 0; counter < nr_counters; counter++)
537
					ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
538
			}
539
		}
540 541
	}

542 543
	fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);

544 545 546 547
	/*
	 * Approximate RIP event size: 24 bytes.
	 */
	fprintf(stderr,
548
		"[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
549 550 551
		(double)bytes_written / 1024.0 / 1024.0,
		output_name,
		bytes_written / 24);
552

553 554
	return 0;
}
555 556

static const char * const record_usage[] = {
557 558
	"perf record [<options>] [<command>]",
	"perf record [<options>] -- <command> [<options>]",
559 560 561
	NULL
};

562
static const struct option options[] = {
563
	OPT_CALLBACK('e', "event", NULL, "event",
564 565
		     "event selector. use 'perf list' to list available events",
		     parse_events),
L
Li Zefan 已提交
566 567
	OPT_CALLBACK(0, "filter", NULL, "filter",
		     "event filter", parse_filter),
568 569 570 571
	OPT_INTEGER('p', "pid", &target_pid,
		    "record events on existing pid"),
	OPT_INTEGER('r', "realtime", &realtime_prio,
		    "collect data with this RT SCHED_FIFO priority"),
572 573
	OPT_BOOLEAN('R', "raw-samples", &raw_samples,
		    "collect raw sample records from all opened counters"),
574 575
	OPT_BOOLEAN('a', "all-cpus", &system_wide,
			    "system-wide collection from all CPUs"),
I
Ingo Molnar 已提交
576 577
	OPT_BOOLEAN('A', "append", &append_file,
			    "append to the output file to do incremental profiling"),
578 579
	OPT_INTEGER('C', "profile_cpu", &profile_cpu,
			    "CPU to profile on"),
580 581
	OPT_BOOLEAN('f', "force", &force,
			"overwrite existing data file"),
582
	OPT_LONG('c', "count", &default_interval,
I
Ingo Molnar 已提交
583 584 585 586 587
		    "event period to sample"),
	OPT_STRING('o', "output", &output_name, "file",
		    "output file name"),
	OPT_BOOLEAN('i', "inherit", &inherit,
		    "child tasks inherit counters"),
I
Ingo Molnar 已提交
588 589
	OPT_INTEGER('F', "freq", &freq,
		    "profile at this frequency"),
I
Ingo Molnar 已提交
590 591
	OPT_INTEGER('m', "mmap-pages", &mmap_pages,
		    "number of mmap data pages"),
592 593
	OPT_BOOLEAN('g', "call-graph", &call_graph,
		    "do call-graph (stack chain/backtrace) recording"),
594 595
	OPT_BOOLEAN('v', "verbose", &verbose,
		    "be more verbose (show counter open errors, etc)"),
596 597
	OPT_BOOLEAN('s', "stat", &inherit_stat,
		    "per thread counts"),
598 599
	OPT_BOOLEAN('d', "data", &sample_address,
		    "Sample addresses"),
600 601
	OPT_BOOLEAN('n', "no-samples", &no_samples,
		    "don't sample"),
602 603
	OPT_BOOLEAN('M', "multiplex", &multiplex,
		    "multiplex counter output in a single channel"),
604 605 606
	OPT_END()
};

607
int cmd_record(int argc, const char **argv, const char *prefix __used)
608 609 610
{
	int counter;

611 612
	symbol__init(0);

613 614
	argc = parse_options(argc, argv, options, record_usage,
		PARSE_OPT_STOP_AT_NON_OPTION);
615
	if (!argc && target_pid == -1 && !system_wide)
616 617
		usage_with_options(record_usage, options);

618 619 620 621 622
	if (!nr_counters) {
		nr_counters	= 1;
		attrs[0].type	= PERF_TYPE_HARDWARE;
		attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
	}
623

624 625 626 627 628 629 630 631 632 633 634 635
	/*
	 * User specified count overrides default frequency.
	 */
	if (default_interval)
		freq = 0;
	else if (freq) {
		default_interval = freq;
	} else {
		fprintf(stderr, "frequency and count are zero, aborting\n");
		exit(EXIT_FAILURE);
	}

636
	for (counter = 0; counter < nr_counters; counter++) {
637
		if (attrs[counter].sample_period)
638 639
			continue;

640
		attrs[counter].sample_period = default_interval;
641 642 643 644
	}

	return __cmd_record(argc, argv);
}