builtin-record.c 14.5 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 18
#include "util/header.h"

19
#include <unistd.h>
20 21
#include <sched.h>

22 23
#define ALIGN(x, a)		__ALIGN_MASK(x, (typeof(x))(a)-1)
#define __ALIGN_MASK(x, mask)	(((x)+(mask))&~(mask))
24

25
static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
26 27 28

static long			default_interval		= 100000;

29
static int			nr_cpus				= 0;
30
static unsigned int		page_size;
31
static unsigned int		mmap_pages			= 128;
I
Ingo Molnar 已提交
32
static int			freq				= 0;
33
static int			output;
34
static const char		*output_name			= "perf.data";
35
static int			group				= 0;
36 37
static unsigned int		realtime_prio			= 0;
static int			system_wide			= 0;
38
static pid_t			target_pid			= -1;
39
static int			inherit				= 1;
40
static int			force				= 0;
I
Ingo Molnar 已提交
41
static int			append_file			= 0;
42
static int			call_graph			= 0;
43 44
static int			inherit_stat			= 0;
static int			no_samples			= 0;
45
static int			sample_address			= 0;
46

47 48 49 50
static long			samples;
static struct timeval		last_read;
static struct timeval		this_read;

51
static u64			bytes_written;
52 53 54 55 56 57

static struct pollfd		event_array[MAX_NR_CPUS * MAX_COUNTERS];

static int			nr_poll;
static int			nr_cpu;

58
static int			file_new = 1;
59 60

struct perf_header		*header;
61

62
struct mmap_data {
63 64 65 66
	int			counter;
	void			*base;
	unsigned int		mask;
	unsigned int		prev;
67 68
};

69 70
static struct mmap_data		mmap_array[MAX_NR_CPUS][MAX_COUNTERS];

71
static unsigned long mmap_read_head(struct mmap_data *md)
72 73
{
	struct perf_counter_mmap_page *pc = md->base;
74
	long head;
75 76 77 78 79 80 81

	head = pc->data_head;
	rmb();

	return head;
}

82 83 84 85 86 87 88 89 90 91 92
static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
{
	struct perf_counter_mmap_page *pc = md->base;

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

93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
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;
	}
}

108 109 110 111 112 113 114 115 116 117 118 119 120
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
121
	 * the writer will bite our tail and mess up the samples under us.
122 123 124 125 126 127
	 *
	 * If we somehow ended up ahead of the head, we got messed up.
	 *
	 * In either case, truncate and restart at head.
	 */
	diff = head - old;
128
	if (diff < 0) {
129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146
		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)
147
		samples++;
148 149 150 151 152 153 154

	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;
155

156
		write_output(buf, size);
157 158 159 160 161
	}

	buf = &data[old & md->mask];
	size = head - old;
	old += size;
162

163
	write_output(buf, size);
164 165

	md->prev = old;
166
	mmap_write_tail(md, old);
167 168 169
}

static volatile int done = 0;
170
static volatile int signr = -1;
171

172
static void sig_handler(int sig)
173
{
174
	done = 1;
175 176 177 178 179 180 181 182 183 184
	signr = sig;
}

static void sig_atexit(void)
{
	if (signr == -1)
		return;

	signal(signr, SIG_DFL);
	kill(getpid(), signr);
185 186
}

187
static void pid_synthesize_comm_event(pid_t pid, int full)
188
{
189
	struct comm_event comm_ev;
190 191
	char filename[PATH_MAX];
	char bf[BUFSIZ];
192
	int fd;
193
	size_t size;
194
	char *field, *sep;
195 196
	DIR *tasks;
	struct dirent dirent, *next;
197 198 199 200 201

	snprintf(filename, sizeof(filename), "/proc/%d/stat", pid);

	fd = open(filename, O_RDONLY);
	if (fd < 0) {
202 203 204 205 206 207
		/*
		 * We raced with a task exiting - just return:
		 */
		if (verbose)
			fprintf(stderr, "couldn't open %s\n", filename);
		return;
208 209 210 211 212 213 214
	}
	if (read(fd, bf, sizeof(bf)) < 0) {
		fprintf(stderr, "couldn't read %s\n", filename);
		exit(EXIT_FAILURE);
	}
	close(fd);

215
	/* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
216
	memset(&comm_ev, 0, sizeof(comm_ev));
217 218 219 220 221 222 223 224
	field = strchr(bf, '(');
	if (field == NULL)
		goto out_failure;
	sep = strchr(++field, ')');
	if (sep == NULL)
		goto out_failure;
	size = sep - field;
	memcpy(comm_ev.comm, field, size++);
225 226

	comm_ev.pid = pid;
227
	comm_ev.header.type = PERF_EVENT_COMM;
228
	size = ALIGN(size, sizeof(u64));
229
	comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
230

231 232 233
	if (!full) {
		comm_ev.tid = pid;

234
		write_output(&comm_ev, comm_ev.header.size);
235 236 237 238 239 240 241 242 243 244 245 246 247 248
		return;
	}

	snprintf(filename, sizeof(filename), "/proc/%d/task", pid);

	tasks = opendir(filename);
	while (!readdir_r(tasks, &dirent, &next) && next) {
		char *end;
		pid = strtol(dirent.d_name, &end, 10);
		if (*end)
			continue;

		comm_ev.tid = pid;

249
		write_output(&comm_ev, comm_ev.header.size);
250
	}
251 252 253
	closedir(tasks);
	return;

254 255 256 257
out_failure:
	fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
		filename);
	exit(EXIT_FAILURE);
258 259
}

260
static void pid_synthesize_mmap_samples(pid_t pid)
261 262 263 264 265 266 267 268
{
	char filename[PATH_MAX];
	FILE *fp;

	snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);

	fp = fopen(filename, "r");
	if (fp == NULL) {
269 270 271 272 273 274
		/*
		 * We raced with a task exiting - just return:
		 */
		if (verbose)
			fprintf(stderr, "couldn't open %s\n", filename);
		return;
275 276
	}
	while (1) {
277
		char bf[BUFSIZ], *pbf = bf;
278
		struct mmap_event mmap_ev = {
279
			.header = { .type = PERF_EVENT_MMAP },
280
		};
281
		int n;
282 283 284 285 286
		size_t size;
		if (fgets(bf, sizeof(bf), fp) == NULL)
			break;

		/* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
287 288 289 290 291 292 293 294 295
		n = hex2u64(pbf, &mmap_ev.start);
		if (n < 0)
			continue;
		pbf += n + 1;
		n = hex2u64(pbf, &mmap_ev.len);
		if (n < 0)
			continue;
		pbf += n + 3;
		if (*pbf == 'x') { /* vm_exec */
296
			char *execname = strchr(bf, '/');
297

298 299 300 301
			/* Catch VDSO */
			if (execname == NULL)
				execname = strstr(bf, "[vdso]");

302
			if (execname == NULL)
303 304 305 306 307
				continue;

			size = strlen(execname);
			execname[size - 1] = '\0'; /* Remove \n */
			memcpy(mmap_ev.filename, execname, size);
308
			size = ALIGN(size, sizeof(u64));
309 310 311
			mmap_ev.len -= mmap_ev.start;
			mmap_ev.header.size = (sizeof(mmap_ev) -
					       (sizeof(mmap_ev.filename) - size));
312
			mmap_ev.pid = pid;
313 314
			mmap_ev.tid = pid;

315
			write_output(&mmap_ev, mmap_ev.header.size);
316 317 318 319 320 321
		}
	}

	fclose(fp);
}

322
static void synthesize_all(void)
323 324 325 326 327 328 329 330 331 332 333 334 335 336 337
{
	DIR *proc;
	struct dirent dirent, *next;

	proc = opendir("/proc");

	while (!readdir_r(proc, &dirent, &next) && next) {
		char *end;
		pid_t pid;

		pid = strtol(dirent.d_name, &end, 10);
		if (*end) /* only interested in proper numerical dirents */
			continue;

		pid_synthesize_comm_event(pid, 1);
338
		pid_synthesize_mmap_samples(pid);
339 340 341 342 343
	}

	closedir(proc);
}

344 345
static int group_fd;

346 347 348 349 350 351 352 353 354 355 356 357 358 359
static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
{
	struct perf_header_attr *h_attr;

	if (nr < header->attrs) {
		h_attr = header->attr[nr];
	} else {
		h_attr = perf_header_attr__new(a);
		perf_header__add_attr(header, h_attr);
	}

	return h_attr;
}

360
static void create_counter(int counter, int cpu, pid_t pid)
361
{
362
	struct perf_counter_attr *attr = attrs + counter;
363 364 365 366 367 368 369 370 371 372 373 374
	struct perf_header_attr *h_attr;
	int track = !counter; /* only the first counter needs these */
	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;
375

376
	attr->sample_type	= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
377

378
	if (freq) {
379
		attr->sample_type	|= PERF_SAMPLE_PERIOD;
380 381
		attr->freq		= 1;
		attr->sample_freq	= freq;
382
	}
383

384 385 386 387 388 389
	if (no_samples)
		attr->sample_freq = 0;

	if (inherit_stat)
		attr->inherit_stat = 1;

390 391 392
	if (sample_address)
		attr->sample_type	|= PERF_SAMPLE_ADDR;

393 394 395
	if (call_graph)
		attr->sample_type	|= PERF_SAMPLE_CALLCHAIN;

396

397 398 399
	attr->mmap		= track;
	attr->comm		= track;
	attr->inherit		= (cpu < 0) && inherit;
400
	attr->disabled		= 1;
401

402
try_again:
403
	fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
404

405 406
	if (fd[nr_cpu][counter] < 0) {
		int err = errno;
407

408
		if (err == EPERM)
409 410 411 412 413 414 415 416
			die("Permission error - are you root?\n");

		/*
		 * 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
417
			&& attr->config == PERF_COUNT_HW_CPU_CYCLES) {
418 419 420 421

			if (verbose)
				warning(" ... trying to fall back to cpu-clock-ticks\n");
			attr->type = PERF_TYPE_SOFTWARE;
422
			attr->config = PERF_COUNT_SW_CPU_CLOCK;
423 424
			goto try_again;
		}
425 426 427 428
		printf("\n");
		error("perfcounter syscall returned with %d (%s)\n",
			fd[nr_cpu][counter], strerror(err));
		die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
429 430
		exit(-1);
	}
431

432 433 434 435 436 437 438 439 440
	h_attr = get_header_attr(attr, counter);

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

441 442 443 444
	if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
		perror("Unable to read perf file descriptor\n");
		exit(-1);
	}
445 446 447

	perf_header_attr__add_id(h_attr, read_data.id);

448 449
	assert(fd[nr_cpu][counter] >= 0);
	fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
450

451 452 453 454 455 456 457 458 459 460 461 462 463 464
	/*
	 * First counter acts as the group leader:
	 */
	if (group && group_fd == -1)
		group_fd = fd[nr_cpu][counter];

	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,
465
			PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
466 467 468 469
	if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
		error("failed to mmap with %d (%s)\n", errno, strerror(errno));
		exit(-1);
	}
470 471

	ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
472
}
473

474 475 476
static void open_counters(int cpu, pid_t pid)
{
	int counter;
477

478 479 480 481
	group_fd = -1;
	for (counter = 0; counter < nr_counters; counter++)
		create_counter(counter, cpu, pid);

482 483 484
	nr_cpu++;
}

485 486
static void atexit_header(void)
{
487
	header->data_size += bytes_written;
488

489
	perf_header__write(header, output);
490 491
}

492
static int __cmd_record(int argc, const char **argv)
493 494
{
	int i, counter;
I
Ingo Molnar 已提交
495
	struct stat st;
496
	pid_t pid = 0;
I
Ingo Molnar 已提交
497
	int flags;
498 499 500 501 502 503 504
	int ret;

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

505 506 507 508
	atexit(sig_atexit);
	signal(SIGCHLD, sig_handler);
	signal(SIGINT, sig_handler);

509 510 511 512 513 514 515 516
	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;
517 518
	}

I
Ingo Molnar 已提交
519 520
	flags = O_CREAT|O_RDWR;
	if (append_file)
521
		file_new = 0;
I
Ingo Molnar 已提交
522 523 524 525
	else
		flags |= O_TRUNC;

	output = open(output_name, flags, S_IRUSR|S_IWUSR);
526 527 528 529 530
	if (output < 0) {
		perror("failed to create output file");
		exit(-1);
	}

531 532 533 534
	if (!file_new)
		header = perf_header__read(output);
	else
		header = perf_header__new();
535 536 537

	atexit(atexit_header);

538
	if (!system_wide) {
539 540 541 542 543
		pid = target_pid;
		if (pid == -1)
			pid = getpid();

		open_counters(-1, pid);
544 545
	} else for (i = 0; i < nr_cpus; i++)
		open_counters(i, target_pid);
546

547 548 549 550 551 552 553 554 555
	if (file_new)
		perf_header__write(header, output);

	if (!system_wide) {
		pid_synthesize_comm_event(pid, 0);
		pid_synthesize_mmap_samples(pid);
	} else
		synthesize_all();

556
	if (target_pid == -1 && argc) {
557 558 559
		pid = fork();
		if (pid < 0)
			perror("failed to fork");
560

561
		if (!pid) {
562
			if (execvp(argv[0], (char **)argv)) {
563 564 565
				perror(argv[0]);
				exit(-1);
			}
566 567 568 569 570 571 572 573 574 575 576 577 578
		}
	}

	if (realtime_prio) {
		struct sched_param param;

		param.sched_priority = realtime_prio;
		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
			printf("Could not set realtime priority.\n");
			exit(-1);
		}
	}

579
	for (;;) {
580
		int hits = samples;
581

582
		for (i = 0; i < nr_cpu; i++) {
583 584 585 586
			for (counter = 0; counter < nr_counters; counter++)
				mmap_read(&mmap_array[i][counter]);
		}

587 588 589
		if (hits == samples) {
			if (done)
				break;
590
			ret = poll(event_array, nr_poll, 100);
591
		}
592 593
	}

594 595 596 597
	/*
	 * Approximate RIP event size: 24 bytes.
	 */
	fprintf(stderr,
598
		"[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
599 600 601
		(double)bytes_written / 1024.0 / 1024.0,
		output_name,
		bytes_written / 24);
602

603 604
	return 0;
}
605 606

static const char * const record_usage[] = {
607 608
	"perf record [<options>] [<command>]",
	"perf record [<options>] -- <command> [<options>]",
609 610 611
	NULL
};

612
static const struct option options[] = {
613
	OPT_CALLBACK('e', "event", NULL, "event",
614 615
		     "event selector. use 'perf list' to list available events",
		     parse_events),
616 617 618 619 620 621
	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"),
	OPT_BOOLEAN('a', "all-cpus", &system_wide,
			    "system-wide collection from all CPUs"),
I
Ingo Molnar 已提交
622 623
	OPT_BOOLEAN('A', "append", &append_file,
			    "append to the output file to do incremental profiling"),
624 625
	OPT_BOOLEAN('f', "force", &force,
			"overwrite existing data file"),
626
	OPT_LONG('c', "count", &default_interval,
I
Ingo Molnar 已提交
627 628 629 630 631
		    "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
632 633
	OPT_INTEGER('F', "freq", &freq,
		    "profile at this frequency"),
I
Ingo Molnar 已提交
634 635
	OPT_INTEGER('m', "mmap-pages", &mmap_pages,
		    "number of mmap data pages"),
636 637
	OPT_BOOLEAN('g', "call-graph", &call_graph,
		    "do call-graph (stack chain/backtrace) recording"),
638 639
	OPT_BOOLEAN('v', "verbose", &verbose,
		    "be more verbose (show counter open errors, etc)"),
640 641
	OPT_BOOLEAN('s', "stat", &inherit_stat,
		    "per thread counts"),
642 643
	OPT_BOOLEAN('d', "data", &sample_address,
		    "Sample addresses"),
644 645
	OPT_BOOLEAN('n', "no-samples", &no_samples,
		    "don't sample"),
646 647 648
	OPT_END()
};

649
int cmd_record(int argc, const char **argv, const char *prefix __used)
650 651 652
{
	int counter;

653 654
	argc = parse_options(argc, argv, options, record_usage,
		PARSE_OPT_STOP_AT_NON_OPTION);
655
	if (!argc && target_pid == -1 && !system_wide)
656 657
		usage_with_options(record_usage, options);

658 659 660 661 662
	if (!nr_counters) {
		nr_counters	= 1;
		attrs[0].type	= PERF_TYPE_HARDWARE;
		attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
	}
663 664

	for (counter = 0; counter < nr_counters; counter++) {
665
		if (attrs[counter].sample_period)
666 667
			continue;

668
		attrs[counter].sample_period = default_interval;
669 670 671 672
	}

	return __cmd_record(argc, argv);
}