xref: /openbmc/linux/tools/perf/builtin-record.c (revision e8e0929d)
1 /*
2  * builtin-record.c
3  *
4  * Builtin record command: Record the profile of a workload
5  * (or a CPU, or a PID) into the perf.data output file - for
6  * later analysis via perf report.
7  */
8 #include "builtin.h"
9 
10 #include "perf.h"
11 
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/trace-event.h"
21 
22 #include <unistd.h>
23 #include <sched.h>
24 
25 #define ALIGN(x, a)		__ALIGN_MASK(x, (typeof(x))(a)-1)
26 #define __ALIGN_MASK(x, mask)	(((x)+(mask))&~(mask))
27 
28 static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
29 
30 static long			default_interval		= 100000;
31 
32 static int			nr_cpus				= 0;
33 static unsigned int		page_size;
34 static unsigned int		mmap_pages			= 128;
35 static int			freq				= 0;
36 static int			output;
37 static const char		*output_name			= "perf.data";
38 static int			group				= 0;
39 static unsigned int		realtime_prio			= 0;
40 static int			raw_samples			= 0;
41 static int			system_wide			= 0;
42 static int			profile_cpu			= -1;
43 static pid_t			target_pid			= -1;
44 static int			inherit				= 1;
45 static int			force				= 0;
46 static int			append_file			= 0;
47 static int			call_graph			= 0;
48 static int			inherit_stat			= 0;
49 static int			no_samples			= 0;
50 static int			sample_address			= 0;
51 static int			multiplex			= 0;
52 static int			multiplex_fd			= -1;
53 
54 static long			samples;
55 static struct timeval		last_read;
56 static struct timeval		this_read;
57 
58 static u64			bytes_written;
59 
60 static struct pollfd		event_array[MAX_NR_CPUS * MAX_COUNTERS];
61 
62 static int			nr_poll;
63 static int			nr_cpu;
64 
65 static int			file_new = 1;
66 
67 struct perf_header		*header;
68 
69 struct mmap_data {
70 	int			counter;
71 	void			*base;
72 	unsigned int		mask;
73 	unsigned int		prev;
74 };
75 
76 static struct mmap_data		mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
77 
78 static unsigned long mmap_read_head(struct mmap_data *md)
79 {
80 	struct perf_event_mmap_page *pc = md->base;
81 	long head;
82 
83 	head = pc->data_head;
84 	rmb();
85 
86 	return head;
87 }
88 
89 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
90 {
91 	struct perf_event_mmap_page *pc = md->base;
92 
93 	/*
94 	 * ensure all reads are done before we write the tail out.
95 	 */
96 	/* mb(); */
97 	pc->data_tail = tail;
98 }
99 
100 static void write_output(void *buf, size_t size)
101 {
102 	while (size) {
103 		int ret = write(output, buf, size);
104 
105 		if (ret < 0)
106 			die("failed to write");
107 
108 		size -= ret;
109 		buf += ret;
110 
111 		bytes_written += ret;
112 	}
113 }
114 
115 static void mmap_read(struct mmap_data *md)
116 {
117 	unsigned int head = mmap_read_head(md);
118 	unsigned int old = md->prev;
119 	unsigned char *data = md->base + page_size;
120 	unsigned long size;
121 	void *buf;
122 	int diff;
123 
124 	gettimeofday(&this_read, NULL);
125 
126 	/*
127 	 * If we're further behind than half the buffer, there's a chance
128 	 * the writer will bite our tail and mess up the samples under us.
129 	 *
130 	 * If we somehow ended up ahead of the head, we got messed up.
131 	 *
132 	 * In either case, truncate and restart at head.
133 	 */
134 	diff = head - old;
135 	if (diff < 0) {
136 		struct timeval iv;
137 		unsigned long msecs;
138 
139 		timersub(&this_read, &last_read, &iv);
140 		msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
141 
142 		fprintf(stderr, "WARNING: failed to keep up with mmap data."
143 				"  Last read %lu msecs ago.\n", msecs);
144 
145 		/*
146 		 * head points to a known good entry, start there.
147 		 */
148 		old = head;
149 	}
150 
151 	last_read = this_read;
152 
153 	if (old != head)
154 		samples++;
155 
156 	size = head - old;
157 
158 	if ((old & md->mask) + size != (head & md->mask)) {
159 		buf = &data[old & md->mask];
160 		size = md->mask + 1 - (old & md->mask);
161 		old += size;
162 
163 		write_output(buf, size);
164 	}
165 
166 	buf = &data[old & md->mask];
167 	size = head - old;
168 	old += size;
169 
170 	write_output(buf, size);
171 
172 	md->prev = old;
173 	mmap_write_tail(md, old);
174 }
175 
176 static volatile int done = 0;
177 static volatile int signr = -1;
178 
179 static void sig_handler(int sig)
180 {
181 	done = 1;
182 	signr = sig;
183 }
184 
185 static void sig_atexit(void)
186 {
187 	if (signr == -1)
188 		return;
189 
190 	signal(signr, SIG_DFL);
191 	kill(getpid(), signr);
192 }
193 
194 static pid_t pid_synthesize_comm_event(pid_t pid, int full)
195 {
196 	struct comm_event comm_ev;
197 	char filename[PATH_MAX];
198 	char bf[BUFSIZ];
199 	FILE *fp;
200 	size_t size = 0;
201 	DIR *tasks;
202 	struct dirent dirent, *next;
203 	pid_t tgid = 0;
204 
205 	snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
206 
207 	fp = fopen(filename, "r");
208 	if (fp == NULL) {
209 		/*
210 		 * We raced with a task exiting - just return:
211 		 */
212 		if (verbose)
213 			fprintf(stderr, "couldn't open %s\n", filename);
214 		return 0;
215 	}
216 
217 	memset(&comm_ev, 0, sizeof(comm_ev));
218 	while (!comm_ev.comm[0] || !comm_ev.pid) {
219 		if (fgets(bf, sizeof(bf), fp) == NULL)
220 			goto out_failure;
221 
222 		if (memcmp(bf, "Name:", 5) == 0) {
223 			char *name = bf + 5;
224 			while (*name && isspace(*name))
225 				++name;
226 			size = strlen(name) - 1;
227 			memcpy(comm_ev.comm, name, size++);
228 		} else if (memcmp(bf, "Tgid:", 5) == 0) {
229 			char *tgids = bf + 5;
230 			while (*tgids && isspace(*tgids))
231 				++tgids;
232 			tgid = comm_ev.pid = atoi(tgids);
233 		}
234 	}
235 
236 	comm_ev.header.type = PERF_RECORD_COMM;
237 	size = ALIGN(size, sizeof(u64));
238 	comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
239 
240 	if (!full) {
241 		comm_ev.tid = pid;
242 
243 		write_output(&comm_ev, comm_ev.header.size);
244 		goto out_fclose;
245 	}
246 
247 	snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
248 
249 	tasks = opendir(filename);
250 	while (!readdir_r(tasks, &dirent, &next) && next) {
251 		char *end;
252 		pid = strtol(dirent.d_name, &end, 10);
253 		if (*end)
254 			continue;
255 
256 		comm_ev.tid = pid;
257 
258 		write_output(&comm_ev, comm_ev.header.size);
259 	}
260 	closedir(tasks);
261 
262 out_fclose:
263 	fclose(fp);
264 	return tgid;
265 
266 out_failure:
267 	fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
268 		filename);
269 	exit(EXIT_FAILURE);
270 }
271 
272 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
273 {
274 	char filename[PATH_MAX];
275 	FILE *fp;
276 
277 	snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
278 
279 	fp = fopen(filename, "r");
280 	if (fp == NULL) {
281 		/*
282 		 * We raced with a task exiting - just return:
283 		 */
284 		if (verbose)
285 			fprintf(stderr, "couldn't open %s\n", filename);
286 		return;
287 	}
288 	while (1) {
289 		char bf[BUFSIZ], *pbf = bf;
290 		struct mmap_event mmap_ev = {
291 			.header = { .type = PERF_RECORD_MMAP },
292 		};
293 		int n;
294 		size_t size;
295 		if (fgets(bf, sizeof(bf), fp) == NULL)
296 			break;
297 
298 		/* 00400000-0040c000 r-xp 00000000 fd:01 41038  /bin/cat */
299 		n = hex2u64(pbf, &mmap_ev.start);
300 		if (n < 0)
301 			continue;
302 		pbf += n + 1;
303 		n = hex2u64(pbf, &mmap_ev.len);
304 		if (n < 0)
305 			continue;
306 		pbf += n + 3;
307 		if (*pbf == 'x') { /* vm_exec */
308 			char *execname = strchr(bf, '/');
309 
310 			/* Catch VDSO */
311 			if (execname == NULL)
312 				execname = strstr(bf, "[vdso]");
313 
314 			if (execname == NULL)
315 				continue;
316 
317 			size = strlen(execname);
318 			execname[size - 1] = '\0'; /* Remove \n */
319 			memcpy(mmap_ev.filename, execname, size);
320 			size = ALIGN(size, sizeof(u64));
321 			mmap_ev.len -= mmap_ev.start;
322 			mmap_ev.header.size = (sizeof(mmap_ev) -
323 					       (sizeof(mmap_ev.filename) - size));
324 			mmap_ev.pid = tgid;
325 			mmap_ev.tid = pid;
326 
327 			write_output(&mmap_ev, mmap_ev.header.size);
328 		}
329 	}
330 
331 	fclose(fp);
332 }
333 
334 static void synthesize_all(void)
335 {
336 	DIR *proc;
337 	struct dirent dirent, *next;
338 
339 	proc = opendir("/proc");
340 
341 	while (!readdir_r(proc, &dirent, &next) && next) {
342 		char *end;
343 		pid_t pid, tgid;
344 
345 		pid = strtol(dirent.d_name, &end, 10);
346 		if (*end) /* only interested in proper numerical dirents */
347 			continue;
348 
349 		tgid = pid_synthesize_comm_event(pid, 1);
350 		pid_synthesize_mmap_samples(pid, tgid);
351 	}
352 
353 	closedir(proc);
354 }
355 
356 static int group_fd;
357 
358 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
359 {
360 	struct perf_header_attr *h_attr;
361 
362 	if (nr < header->attrs) {
363 		h_attr = header->attr[nr];
364 	} else {
365 		h_attr = perf_header_attr__new(a);
366 		perf_header__add_attr(header, h_attr);
367 	}
368 
369 	return h_attr;
370 }
371 
372 static void create_counter(int counter, int cpu, pid_t pid)
373 {
374 	struct perf_event_attr *attr = attrs + counter;
375 	struct perf_header_attr *h_attr;
376 	int track = !counter; /* only the first counter needs these */
377 	struct {
378 		u64 count;
379 		u64 time_enabled;
380 		u64 time_running;
381 		u64 id;
382 	} read_data;
383 
384 	attr->read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
385 				  PERF_FORMAT_TOTAL_TIME_RUNNING |
386 				  PERF_FORMAT_ID;
387 
388 	attr->sample_type	|= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
389 
390 	if (freq) {
391 		attr->sample_type	|= PERF_SAMPLE_PERIOD;
392 		attr->freq		= 1;
393 		attr->sample_freq	= freq;
394 	}
395 
396 	if (no_samples)
397 		attr->sample_freq = 0;
398 
399 	if (inherit_stat)
400 		attr->inherit_stat = 1;
401 
402 	if (sample_address)
403 		attr->sample_type	|= PERF_SAMPLE_ADDR;
404 
405 	if (call_graph)
406 		attr->sample_type	|= PERF_SAMPLE_CALLCHAIN;
407 
408 	if (raw_samples) {
409 		attr->sample_type	|= PERF_SAMPLE_TIME;
410 		attr->sample_type	|= PERF_SAMPLE_RAW;
411 		attr->sample_type	|= PERF_SAMPLE_CPU;
412 	}
413 
414 	attr->mmap		= track;
415 	attr->comm		= track;
416 	attr->inherit		= (cpu < 0) && inherit;
417 	attr->disabled		= 1;
418 
419 try_again:
420 	fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
421 
422 	if (fd[nr_cpu][counter] < 0) {
423 		int err = errno;
424 
425 		if (err == EPERM)
426 			die("Permission error - are you root?\n");
427 		else if (err ==  ENODEV && profile_cpu != -1)
428 			die("No such device - did you specify an out-of-range profile CPU?\n");
429 
430 		/*
431 		 * If it's cycles then fall back to hrtimer
432 		 * based cpu-clock-tick sw counter, which
433 		 * is always available even if no PMU support:
434 		 */
435 		if (attr->type == PERF_TYPE_HARDWARE
436 			&& attr->config == PERF_COUNT_HW_CPU_CYCLES) {
437 
438 			if (verbose)
439 				warning(" ... trying to fall back to cpu-clock-ticks\n");
440 			attr->type = PERF_TYPE_SOFTWARE;
441 			attr->config = PERF_COUNT_SW_CPU_CLOCK;
442 			goto try_again;
443 		}
444 		printf("\n");
445 		error("perfcounter syscall returned with %d (%s)\n",
446 			fd[nr_cpu][counter], strerror(err));
447 		die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
448 		exit(-1);
449 	}
450 
451 	h_attr = get_header_attr(attr, counter);
452 
453 	if (!file_new) {
454 		if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
455 			fprintf(stderr, "incompatible append\n");
456 			exit(-1);
457 		}
458 	}
459 
460 	if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
461 		perror("Unable to read perf file descriptor\n");
462 		exit(-1);
463 	}
464 
465 	perf_header_attr__add_id(h_attr, read_data.id);
466 
467 	assert(fd[nr_cpu][counter] >= 0);
468 	fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
469 
470 	/*
471 	 * First counter acts as the group leader:
472 	 */
473 	if (group && group_fd == -1)
474 		group_fd = fd[nr_cpu][counter];
475 	if (multiplex && multiplex_fd == -1)
476 		multiplex_fd = fd[nr_cpu][counter];
477 
478 	if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
479 		int ret;
480 
481 		ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
482 		assert(ret != -1);
483 	} else {
484 		event_array[nr_poll].fd = fd[nr_cpu][counter];
485 		event_array[nr_poll].events = POLLIN;
486 		nr_poll++;
487 
488 		mmap_array[nr_cpu][counter].counter = counter;
489 		mmap_array[nr_cpu][counter].prev = 0;
490 		mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
491 		mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
492 				PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
493 		if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
494 			error("failed to mmap with %d (%s)\n", errno, strerror(errno));
495 			exit(-1);
496 		}
497 	}
498 
499 	ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
500 }
501 
502 static void open_counters(int cpu, pid_t pid)
503 {
504 	int counter;
505 
506 	group_fd = -1;
507 	for (counter = 0; counter < nr_counters; counter++)
508 		create_counter(counter, cpu, pid);
509 
510 	nr_cpu++;
511 }
512 
513 static void atexit_header(void)
514 {
515 	header->data_size += bytes_written;
516 
517 	perf_header__write(header, output);
518 }
519 
520 static int __cmd_record(int argc, const char **argv)
521 {
522 	int i, counter;
523 	struct stat st;
524 	pid_t pid = 0;
525 	int flags;
526 	int ret;
527 	unsigned long waking = 0;
528 
529 	page_size = sysconf(_SC_PAGE_SIZE);
530 	nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
531 	assert(nr_cpus <= MAX_NR_CPUS);
532 	assert(nr_cpus >= 0);
533 
534 	atexit(sig_atexit);
535 	signal(SIGCHLD, sig_handler);
536 	signal(SIGINT, sig_handler);
537 
538 	if (!stat(output_name, &st) && st.st_size) {
539 		if (!force && !append_file) {
540 			fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
541 					output_name);
542 			exit(-1);
543 		}
544 	} else {
545 		append_file = 0;
546 	}
547 
548 	flags = O_CREAT|O_RDWR;
549 	if (append_file)
550 		file_new = 0;
551 	else
552 		flags |= O_TRUNC;
553 
554 	output = open(output_name, flags, S_IRUSR|S_IWUSR);
555 	if (output < 0) {
556 		perror("failed to create output file");
557 		exit(-1);
558 	}
559 
560 	if (!file_new)
561 		header = perf_header__read(output);
562 	else
563 		header = perf_header__new();
564 
565 
566 	if (raw_samples) {
567 		read_tracing_data(attrs, nr_counters);
568 	} else {
569 		for (i = 0; i < nr_counters; i++) {
570 			if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
571 				read_tracing_data(attrs, nr_counters);
572 				break;
573 			}
574 		}
575 	}
576 	atexit(atexit_header);
577 
578 	if (!system_wide) {
579 		pid = target_pid;
580 		if (pid == -1)
581 			pid = getpid();
582 
583 		open_counters(profile_cpu, pid);
584 	} else {
585 		if (profile_cpu != -1) {
586 			open_counters(profile_cpu, target_pid);
587 		} else {
588 			for (i = 0; i < nr_cpus; i++)
589 				open_counters(i, target_pid);
590 		}
591 	}
592 
593 	if (file_new)
594 		perf_header__write(header, output);
595 
596 	if (!system_wide) {
597 		pid_t tgid = pid_synthesize_comm_event(pid, 0);
598 		pid_synthesize_mmap_samples(pid, tgid);
599 	} else
600 		synthesize_all();
601 
602 	if (target_pid == -1 && argc) {
603 		pid = fork();
604 		if (pid < 0)
605 			perror("failed to fork");
606 
607 		if (!pid) {
608 			if (execvp(argv[0], (char **)argv)) {
609 				perror(argv[0]);
610 				exit(-1);
611 			}
612 		}
613 	}
614 
615 	if (realtime_prio) {
616 		struct sched_param param;
617 
618 		param.sched_priority = realtime_prio;
619 		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
620 			printf("Could not set realtime priority.\n");
621 			exit(-1);
622 		}
623 	}
624 
625 	for (;;) {
626 		int hits = samples;
627 
628 		for (i = 0; i < nr_cpu; i++) {
629 			for (counter = 0; counter < nr_counters; counter++) {
630 				if (mmap_array[i][counter].base)
631 					mmap_read(&mmap_array[i][counter]);
632 			}
633 		}
634 
635 		if (hits == samples) {
636 			if (done)
637 				break;
638 			ret = poll(event_array, nr_poll, -1);
639 			waking++;
640 		}
641 
642 		if (done) {
643 			for (i = 0; i < nr_cpu; i++) {
644 				for (counter = 0; counter < nr_counters; counter++)
645 					ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
646 			}
647 		}
648 	}
649 
650 	fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
651 
652 	/*
653 	 * Approximate RIP event size: 24 bytes.
654 	 */
655 	fprintf(stderr,
656 		"[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
657 		(double)bytes_written / 1024.0 / 1024.0,
658 		output_name,
659 		bytes_written / 24);
660 
661 	return 0;
662 }
663 
664 static const char * const record_usage[] = {
665 	"perf record [<options>] [<command>]",
666 	"perf record [<options>] -- <command> [<options>]",
667 	NULL
668 };
669 
670 static const struct option options[] = {
671 	OPT_CALLBACK('e', "event", NULL, "event",
672 		     "event selector. use 'perf list' to list available events",
673 		     parse_events),
674 	OPT_INTEGER('p', "pid", &target_pid,
675 		    "record events on existing pid"),
676 	OPT_INTEGER('r', "realtime", &realtime_prio,
677 		    "collect data with this RT SCHED_FIFO priority"),
678 	OPT_BOOLEAN('R', "raw-samples", &raw_samples,
679 		    "collect raw sample records from all opened counters"),
680 	OPT_BOOLEAN('a', "all-cpus", &system_wide,
681 			    "system-wide collection from all CPUs"),
682 	OPT_BOOLEAN('A', "append", &append_file,
683 			    "append to the output file to do incremental profiling"),
684 	OPT_INTEGER('C', "profile_cpu", &profile_cpu,
685 			    "CPU to profile on"),
686 	OPT_BOOLEAN('f', "force", &force,
687 			"overwrite existing data file"),
688 	OPT_LONG('c', "count", &default_interval,
689 		    "event period to sample"),
690 	OPT_STRING('o', "output", &output_name, "file",
691 		    "output file name"),
692 	OPT_BOOLEAN('i', "inherit", &inherit,
693 		    "child tasks inherit counters"),
694 	OPT_INTEGER('F', "freq", &freq,
695 		    "profile at this frequency"),
696 	OPT_INTEGER('m', "mmap-pages", &mmap_pages,
697 		    "number of mmap data pages"),
698 	OPT_BOOLEAN('g', "call-graph", &call_graph,
699 		    "do call-graph (stack chain/backtrace) recording"),
700 	OPT_BOOLEAN('v', "verbose", &verbose,
701 		    "be more verbose (show counter open errors, etc)"),
702 	OPT_BOOLEAN('s', "stat", &inherit_stat,
703 		    "per thread counts"),
704 	OPT_BOOLEAN('d', "data", &sample_address,
705 		    "Sample addresses"),
706 	OPT_BOOLEAN('n', "no-samples", &no_samples,
707 		    "don't sample"),
708 	OPT_BOOLEAN('M', "multiplex", &multiplex,
709 		    "multiplex counter output in a single channel"),
710 	OPT_END()
711 };
712 
713 int cmd_record(int argc, const char **argv, const char *prefix __used)
714 {
715 	int counter;
716 
717 	argc = parse_options(argc, argv, options, record_usage,
718 		PARSE_OPT_STOP_AT_NON_OPTION);
719 	if (!argc && target_pid == -1 && !system_wide)
720 		usage_with_options(record_usage, options);
721 
722 	if (!nr_counters) {
723 		nr_counters	= 1;
724 		attrs[0].type	= PERF_TYPE_HARDWARE;
725 		attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
726 	}
727 
728 	for (counter = 0; counter < nr_counters; counter++) {
729 		if (attrs[counter].sample_period)
730 			continue;
731 
732 		attrs[counter].sample_period = default_interval;
733 	}
734 
735 	return __cmd_record(argc, argv);
736 }
737