xref: /openbmc/linux/tools/perf/builtin-record.c (revision 565d76cb)
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 #define _FILE_OFFSET_BITS 64
9 
10 #include "builtin.h"
11 
12 #include "perf.h"
13 
14 #include "util/build-id.h"
15 #include "util/util.h"
16 #include "util/parse-options.h"
17 #include "util/parse-events.h"
18 
19 #include "util/header.h"
20 #include "util/event.h"
21 #include "util/evlist.h"
22 #include "util/evsel.h"
23 #include "util/debug.h"
24 #include "util/session.h"
25 #include "util/symbol.h"
26 #include "util/cpumap.h"
27 #include "util/thread_map.h"
28 
29 #include <unistd.h>
30 #include <sched.h>
31 #include <sys/mman.h>
32 
33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
34 
35 enum write_mode_t {
36 	WRITE_FORCE,
37 	WRITE_APPEND
38 };
39 
40 static u64			user_interval			= ULLONG_MAX;
41 static u64			default_interval		=      0;
42 
43 static unsigned int		page_size;
44 static unsigned int		mmap_pages			=    128;
45 static unsigned int		user_freq 			= UINT_MAX;
46 static int			freq				=   1000;
47 static int			output;
48 static int			pipe_output			=      0;
49 static const char		*output_name			= NULL;
50 static int			group				=      0;
51 static int			realtime_prio			=      0;
52 static bool			nodelay				=  false;
53 static bool			raw_samples			=  false;
54 static bool			sample_id_all_avail		=   true;
55 static bool			system_wide			=  false;
56 static pid_t			target_pid			=     -1;
57 static pid_t			target_tid			=     -1;
58 static pid_t			child_pid			=     -1;
59 static bool			no_inherit			=  false;
60 static enum write_mode_t	write_mode			= WRITE_FORCE;
61 static bool			call_graph			=  false;
62 static bool			inherit_stat			=  false;
63 static bool			no_samples			=  false;
64 static bool			sample_address			=  false;
65 static bool			sample_time			=  false;
66 static bool			no_buildid			=  false;
67 static bool			no_buildid_cache		=  false;
68 static struct perf_evlist	*evsel_list;
69 
70 static long			samples				=      0;
71 static u64			bytes_written			=      0;
72 
73 static int			file_new			=      1;
74 static off_t			post_processing_offset;
75 
76 static struct perf_session	*session;
77 static const char		*cpu_list;
78 
79 static void advance_output(size_t size)
80 {
81 	bytes_written += size;
82 }
83 
84 static void write_output(void *buf, size_t size)
85 {
86 	while (size) {
87 		int ret = write(output, buf, size);
88 
89 		if (ret < 0)
90 			die("failed to write");
91 
92 		size -= ret;
93 		buf += ret;
94 
95 		bytes_written += ret;
96 	}
97 }
98 
99 static int process_synthesized_event(union perf_event *event,
100 				     struct perf_sample *sample __used,
101 				     struct perf_session *self __used)
102 {
103 	write_output(event, event->header.size);
104 	return 0;
105 }
106 
107 static void mmap_read(struct perf_mmap *md)
108 {
109 	unsigned int head = perf_mmap__read_head(md);
110 	unsigned int old = md->prev;
111 	unsigned char *data = md->base + page_size;
112 	unsigned long size;
113 	void *buf;
114 
115 	if (old == head)
116 		return;
117 
118 	samples++;
119 
120 	size = head - old;
121 
122 	if ((old & md->mask) + size != (head & md->mask)) {
123 		buf = &data[old & md->mask];
124 		size = md->mask + 1 - (old & md->mask);
125 		old += size;
126 
127 		write_output(buf, size);
128 	}
129 
130 	buf = &data[old & md->mask];
131 	size = head - old;
132 	old += size;
133 
134 	write_output(buf, size);
135 
136 	md->prev = old;
137 	perf_mmap__write_tail(md, old);
138 }
139 
140 static volatile int done = 0;
141 static volatile int signr = -1;
142 
143 static void sig_handler(int sig)
144 {
145 	done = 1;
146 	signr = sig;
147 }
148 
149 static void sig_atexit(void)
150 {
151 	if (child_pid > 0)
152 		kill(child_pid, SIGTERM);
153 
154 	if (signr == -1 || signr == SIGUSR1)
155 		return;
156 
157 	signal(signr, SIG_DFL);
158 	kill(getpid(), signr);
159 }
160 
161 static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
162 {
163 	struct perf_event_attr *attr = &evsel->attr;
164 	int track = !evsel->idx; /* only the first counter needs these */
165 
166 	attr->read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
167 				  PERF_FORMAT_TOTAL_TIME_RUNNING |
168 				  PERF_FORMAT_ID;
169 
170 	attr->sample_type	|= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
171 
172 	if (evlist->nr_entries > 1)
173 		attr->sample_type |= PERF_SAMPLE_ID;
174 
175 	/*
176 	 * We default some events to a 1 default interval. But keep
177 	 * it a weak assumption overridable by the user.
178 	 */
179 	if (!attr->sample_period || (user_freq != UINT_MAX &&
180 				     user_interval != ULLONG_MAX)) {
181 		if (freq) {
182 			attr->sample_type	|= PERF_SAMPLE_PERIOD;
183 			attr->freq		= 1;
184 			attr->sample_freq	= freq;
185 		} else {
186 			attr->sample_period = default_interval;
187 		}
188 	}
189 
190 	if (no_samples)
191 		attr->sample_freq = 0;
192 
193 	if (inherit_stat)
194 		attr->inherit_stat = 1;
195 
196 	if (sample_address) {
197 		attr->sample_type	|= PERF_SAMPLE_ADDR;
198 		attr->mmap_data = track;
199 	}
200 
201 	if (call_graph)
202 		attr->sample_type	|= PERF_SAMPLE_CALLCHAIN;
203 
204 	if (system_wide)
205 		attr->sample_type	|= PERF_SAMPLE_CPU;
206 
207 	if (sample_id_all_avail &&
208 	    (sample_time || system_wide || !no_inherit || cpu_list))
209 		attr->sample_type	|= PERF_SAMPLE_TIME;
210 
211 	if (raw_samples) {
212 		attr->sample_type	|= PERF_SAMPLE_TIME;
213 		attr->sample_type	|= PERF_SAMPLE_RAW;
214 		attr->sample_type	|= PERF_SAMPLE_CPU;
215 	}
216 
217 	if (nodelay) {
218 		attr->watermark = 0;
219 		attr->wakeup_events = 1;
220 	}
221 
222 	attr->mmap		= track;
223 	attr->comm		= track;
224 
225 	if (target_pid == -1 && target_tid == -1 && !system_wide) {
226 		attr->disabled = 1;
227 		attr->enable_on_exec = 1;
228 	}
229 }
230 
231 static bool perf_evlist__equal(struct perf_evlist *evlist,
232 			       struct perf_evlist *other)
233 {
234 	struct perf_evsel *pos, *pair;
235 
236 	if (evlist->nr_entries != other->nr_entries)
237 		return false;
238 
239 	pair = list_entry(other->entries.next, struct perf_evsel, node);
240 
241 	list_for_each_entry(pos, &evlist->entries, node) {
242 		if (memcmp(&pos->attr, &pair->attr, sizeof(pos->attr) != 0))
243 			return false;
244 		pair = list_entry(pair->node.next, struct perf_evsel, node);
245 	}
246 
247 	return true;
248 }
249 
250 static void open_counters(struct perf_evlist *evlist)
251 {
252 	struct perf_evsel *pos;
253 
254 	list_for_each_entry(pos, &evlist->entries, node) {
255 		struct perf_event_attr *attr = &pos->attr;
256 		/*
257 		 * Check if parse_single_tracepoint_event has already asked for
258 		 * PERF_SAMPLE_TIME.
259 		 *
260 		 * XXX this is kludgy but short term fix for problems introduced by
261 		 * eac23d1c that broke 'perf script' by having different sample_types
262 		 * when using multiple tracepoint events when we use a perf binary
263 		 * that tries to use sample_id_all on an older kernel.
264 		 *
265 		 * We need to move counter creation to perf_session, support
266 		 * different sample_types, etc.
267 		 */
268 		bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
269 
270 		config_attr(pos, evlist);
271 retry_sample_id:
272 		attr->sample_id_all = sample_id_all_avail ? 1 : 0;
273 try_again:
274 		if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group,
275 				     !no_inherit) < 0) {
276 			int err = errno;
277 
278 			if (err == EPERM || err == EACCES)
279 				die("Permission error - are you root?\n"
280 					"\t Consider tweaking"
281 					" /proc/sys/kernel/perf_event_paranoid.\n");
282 			else if (err ==  ENODEV && cpu_list) {
283 				die("No such device - did you specify"
284 					" an out-of-range profile CPU?\n");
285 			} else if (err == EINVAL && sample_id_all_avail) {
286 				/*
287 				 * Old kernel, no attr->sample_id_type_all field
288 				 */
289 				sample_id_all_avail = false;
290 				if (!sample_time && !raw_samples && !time_needed)
291 					attr->sample_type &= ~PERF_SAMPLE_TIME;
292 
293 				goto retry_sample_id;
294 			}
295 
296 			/*
297 			 * If it's cycles then fall back to hrtimer
298 			 * based cpu-clock-tick sw counter, which
299 			 * is always available even if no PMU support:
300 			 */
301 			if (attr->type == PERF_TYPE_HARDWARE
302 					&& attr->config == PERF_COUNT_HW_CPU_CYCLES) {
303 
304 				if (verbose)
305 					warning(" ... trying to fall back to cpu-clock-ticks\n");
306 				attr->type = PERF_TYPE_SOFTWARE;
307 				attr->config = PERF_COUNT_SW_CPU_CLOCK;
308 				goto try_again;
309 			}
310 			printf("\n");
311 			error("sys_perf_event_open() syscall returned with %d (%s).  /bin/dmesg may provide additional information.\n",
312 			      err, strerror(err));
313 
314 #if defined(__i386__) || defined(__x86_64__)
315 			if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
316 				die("No hardware sampling interrupt available."
317 				    " No APIC? If so then you can boot the kernel"
318 				    " with the \"lapic\" boot parameter to"
319 				    " force-enable it.\n");
320 #endif
321 
322 			die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
323 		}
324 	}
325 
326 	if (perf_evlist__set_filters(evlist)) {
327 		error("failed to set filter with %d (%s)\n", errno,
328 			strerror(errno));
329 		exit(-1);
330 	}
331 
332 	if (perf_evlist__mmap(evlist, mmap_pages, false) < 0)
333 		die("failed to mmap with %d (%s)\n", errno, strerror(errno));
334 
335 	if (file_new)
336 		session->evlist = evlist;
337 	else {
338 		if (!perf_evlist__equal(session->evlist, evlist)) {
339 			fprintf(stderr, "incompatible append\n");
340 			exit(-1);
341 		}
342  	}
343 
344 	perf_session__update_sample_type(session);
345 }
346 
347 static int process_buildids(void)
348 {
349 	u64 size = lseek(output, 0, SEEK_CUR);
350 
351 	if (size == 0)
352 		return 0;
353 
354 	session->fd = output;
355 	return __perf_session__process_events(session, post_processing_offset,
356 					      size - post_processing_offset,
357 					      size, &build_id__mark_dso_hit_ops);
358 }
359 
360 static void atexit_header(void)
361 {
362 	if (!pipe_output) {
363 		session->header.data_size += bytes_written;
364 
365 		if (!no_buildid)
366 			process_buildids();
367 		perf_session__write_header(session, evsel_list, output, true);
368 		perf_session__delete(session);
369 		perf_evlist__delete(evsel_list);
370 		symbol__exit();
371 	}
372 }
373 
374 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
375 {
376 	int err;
377 	struct perf_session *psession = data;
378 
379 	if (machine__is_host(machine))
380 		return;
381 
382 	/*
383 	 *As for guest kernel when processing subcommand record&report,
384 	 *we arrange module mmap prior to guest kernel mmap and trigger
385 	 *a preload dso because default guest module symbols are loaded
386 	 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
387 	 *method is used to avoid symbol missing when the first addr is
388 	 *in module instead of in guest kernel.
389 	 */
390 	err = perf_event__synthesize_modules(process_synthesized_event,
391 					     psession, machine);
392 	if (err < 0)
393 		pr_err("Couldn't record guest kernel [%d]'s reference"
394 		       " relocation symbol.\n", machine->pid);
395 
396 	/*
397 	 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
398 	 * have no _text sometimes.
399 	 */
400 	err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
401 						 psession, machine, "_text");
402 	if (err < 0)
403 		err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
404 							 psession, machine,
405 							 "_stext");
406 	if (err < 0)
407 		pr_err("Couldn't record guest kernel [%d]'s reference"
408 		       " relocation symbol.\n", machine->pid);
409 }
410 
411 static struct perf_event_header finished_round_event = {
412 	.size = sizeof(struct perf_event_header),
413 	.type = PERF_RECORD_FINISHED_ROUND,
414 };
415 
416 static void mmap_read_all(void)
417 {
418 	int i;
419 
420 	for (i = 0; i < evsel_list->cpus->nr; i++) {
421 		if (evsel_list->mmap[i].base)
422 			mmap_read(&evsel_list->mmap[i]);
423 	}
424 
425 	if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
426 		write_output(&finished_round_event, sizeof(finished_round_event));
427 }
428 
429 static int __cmd_record(int argc, const char **argv)
430 {
431 	int i;
432 	struct stat st;
433 	int flags;
434 	int err;
435 	unsigned long waking = 0;
436 	int child_ready_pipe[2], go_pipe[2];
437 	const bool forks = argc > 0;
438 	char buf;
439 	struct machine *machine;
440 
441 	page_size = sysconf(_SC_PAGE_SIZE);
442 
443 	atexit(sig_atexit);
444 	signal(SIGCHLD, sig_handler);
445 	signal(SIGINT, sig_handler);
446 	signal(SIGUSR1, sig_handler);
447 
448 	if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
449 		perror("failed to create pipes");
450 		exit(-1);
451 	}
452 
453 	if (!output_name) {
454 		if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
455 			pipe_output = 1;
456 		else
457 			output_name = "perf.data";
458 	}
459 	if (output_name) {
460 		if (!strcmp(output_name, "-"))
461 			pipe_output = 1;
462 		else if (!stat(output_name, &st) && st.st_size) {
463 			if (write_mode == WRITE_FORCE) {
464 				char oldname[PATH_MAX];
465 				snprintf(oldname, sizeof(oldname), "%s.old",
466 					 output_name);
467 				unlink(oldname);
468 				rename(output_name, oldname);
469 			}
470 		} else if (write_mode == WRITE_APPEND) {
471 			write_mode = WRITE_FORCE;
472 		}
473 	}
474 
475 	flags = O_CREAT|O_RDWR;
476 	if (write_mode == WRITE_APPEND)
477 		file_new = 0;
478 	else
479 		flags |= O_TRUNC;
480 
481 	if (pipe_output)
482 		output = STDOUT_FILENO;
483 	else
484 		output = open(output_name, flags, S_IRUSR | S_IWUSR);
485 	if (output < 0) {
486 		perror("failed to create output file");
487 		exit(-1);
488 	}
489 
490 	session = perf_session__new(output_name, O_WRONLY,
491 				    write_mode == WRITE_FORCE, false, NULL);
492 	if (session == NULL) {
493 		pr_err("Not enough memory for reading perf file header\n");
494 		return -1;
495 	}
496 
497 	if (!no_buildid)
498 		perf_header__set_feat(&session->header, HEADER_BUILD_ID);
499 
500 	if (!file_new) {
501 		err = perf_session__read_header(session, output);
502 		if (err < 0)
503 			goto out_delete_session;
504 	}
505 
506 	if (have_tracepoints(&evsel_list->entries))
507 		perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
508 
509 	if (forks) {
510 		child_pid = fork();
511 		if (child_pid < 0) {
512 			perror("failed to fork");
513 			exit(-1);
514 		}
515 
516 		if (!child_pid) {
517 			if (pipe_output)
518 				dup2(2, 1);
519 			close(child_ready_pipe[0]);
520 			close(go_pipe[1]);
521 			fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
522 
523 			/*
524 			 * Do a dummy execvp to get the PLT entry resolved,
525 			 * so we avoid the resolver overhead on the real
526 			 * execvp call.
527 			 */
528 			execvp("", (char **)argv);
529 
530 			/*
531 			 * Tell the parent we're ready to go
532 			 */
533 			close(child_ready_pipe[1]);
534 
535 			/*
536 			 * Wait until the parent tells us to go.
537 			 */
538 			if (read(go_pipe[0], &buf, 1) == -1)
539 				perror("unable to read pipe");
540 
541 			execvp(argv[0], (char **)argv);
542 
543 			perror(argv[0]);
544 			kill(getppid(), SIGUSR1);
545 			exit(-1);
546 		}
547 
548 		if (!system_wide && target_tid == -1 && target_pid == -1)
549 			evsel_list->threads->map[0] = child_pid;
550 
551 		close(child_ready_pipe[1]);
552 		close(go_pipe[0]);
553 		/*
554 		 * wait for child to settle
555 		 */
556 		if (read(child_ready_pipe[0], &buf, 1) == -1) {
557 			perror("unable to read pipe");
558 			exit(-1);
559 		}
560 		close(child_ready_pipe[0]);
561 	}
562 
563 	open_counters(evsel_list);
564 
565 	/*
566 	 * perf_session__delete(session) will be called at atexit_header()
567 	 */
568 	atexit(atexit_header);
569 
570 	if (pipe_output) {
571 		err = perf_header__write_pipe(output);
572 		if (err < 0)
573 			return err;
574 	} else if (file_new) {
575 		err = perf_session__write_header(session, evsel_list,
576 						 output, false);
577 		if (err < 0)
578 			return err;
579 	}
580 
581 	post_processing_offset = lseek(output, 0, SEEK_CUR);
582 
583 	if (pipe_output) {
584 		err = perf_session__synthesize_attrs(session,
585 						     process_synthesized_event);
586 		if (err < 0) {
587 			pr_err("Couldn't synthesize attrs.\n");
588 			return err;
589 		}
590 
591 		err = perf_event__synthesize_event_types(process_synthesized_event,
592 							 session);
593 		if (err < 0) {
594 			pr_err("Couldn't synthesize event_types.\n");
595 			return err;
596 		}
597 
598 		if (have_tracepoints(&evsel_list->entries)) {
599 			/*
600 			 * FIXME err <= 0 here actually means that
601 			 * there were no tracepoints so its not really
602 			 * an error, just that we don't need to
603 			 * synthesize anything.  We really have to
604 			 * return this more properly and also
605 			 * propagate errors that now are calling die()
606 			 */
607 			err = perf_event__synthesize_tracing_data(output, evsel_list,
608 								  process_synthesized_event,
609 								  session);
610 			if (err <= 0) {
611 				pr_err("Couldn't record tracing data.\n");
612 				return err;
613 			}
614 			advance_output(err);
615 		}
616 	}
617 
618 	machine = perf_session__find_host_machine(session);
619 	if (!machine) {
620 		pr_err("Couldn't find native kernel information.\n");
621 		return -1;
622 	}
623 
624 	err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
625 						 session, machine, "_text");
626 	if (err < 0)
627 		err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
628 							 session, machine, "_stext");
629 	if (err < 0)
630 		pr_err("Couldn't record kernel reference relocation symbol\n"
631 		       "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
632 		       "Check /proc/kallsyms permission or run as root.\n");
633 
634 	err = perf_event__synthesize_modules(process_synthesized_event,
635 					     session, machine);
636 	if (err < 0)
637 		pr_err("Couldn't record kernel module information.\n"
638 		       "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
639 		       "Check /proc/modules permission or run as root.\n");
640 
641 	if (perf_guest)
642 		perf_session__process_machines(session,
643 					       perf_event__synthesize_guest_os);
644 
645 	if (!system_wide)
646 		perf_event__synthesize_thread_map(evsel_list->threads,
647 						  process_synthesized_event,
648 						  session);
649 	else
650 		perf_event__synthesize_threads(process_synthesized_event,
651 					       session);
652 
653 	if (realtime_prio) {
654 		struct sched_param param;
655 
656 		param.sched_priority = realtime_prio;
657 		if (sched_setscheduler(0, SCHED_FIFO, &param)) {
658 			pr_err("Could not set realtime priority.\n");
659 			exit(-1);
660 		}
661 	}
662 
663 	/*
664 	 * Let the child rip
665 	 */
666 	if (forks)
667 		close(go_pipe[1]);
668 
669 	for (;;) {
670 		int hits = samples;
671 		int thread;
672 
673 		mmap_read_all();
674 
675 		if (hits == samples) {
676 			if (done)
677 				break;
678 			err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
679 			waking++;
680 		}
681 
682 		if (done) {
683 			for (i = 0; i < evsel_list->cpus->nr; i++) {
684 				struct perf_evsel *pos;
685 
686 				list_for_each_entry(pos, &evsel_list->entries, node) {
687 					for (thread = 0;
688 						thread < evsel_list->threads->nr;
689 						thread++)
690 						ioctl(FD(pos, i, thread),
691 							PERF_EVENT_IOC_DISABLE);
692 				}
693 			}
694 		}
695 	}
696 
697 	if (quiet || signr == SIGUSR1)
698 		return 0;
699 
700 	fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
701 
702 	/*
703 	 * Approximate RIP event size: 24 bytes.
704 	 */
705 	fprintf(stderr,
706 		"[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
707 		(double)bytes_written / 1024.0 / 1024.0,
708 		output_name,
709 		bytes_written / 24);
710 
711 	return 0;
712 
713 out_delete_session:
714 	perf_session__delete(session);
715 	return err;
716 }
717 
718 static const char * const record_usage[] = {
719 	"perf record [<options>] [<command>]",
720 	"perf record [<options>] -- <command> [<options>]",
721 	NULL
722 };
723 
724 static bool force, append_file;
725 
726 const struct option record_options[] = {
727 	OPT_CALLBACK('e', "event", &evsel_list, "event",
728 		     "event selector. use 'perf list' to list available events",
729 		     parse_events),
730 	OPT_CALLBACK(0, "filter", &evsel_list, "filter",
731 		     "event filter", parse_filter),
732 	OPT_INTEGER('p', "pid", &target_pid,
733 		    "record events on existing process id"),
734 	OPT_INTEGER('t', "tid", &target_tid,
735 		    "record events on existing thread id"),
736 	OPT_INTEGER('r', "realtime", &realtime_prio,
737 		    "collect data with this RT SCHED_FIFO priority"),
738 	OPT_BOOLEAN('D', "no-delay", &nodelay,
739 		    "collect data without buffering"),
740 	OPT_BOOLEAN('R', "raw-samples", &raw_samples,
741 		    "collect raw sample records from all opened counters"),
742 	OPT_BOOLEAN('a', "all-cpus", &system_wide,
743 			    "system-wide collection from all CPUs"),
744 	OPT_BOOLEAN('A', "append", &append_file,
745 			    "append to the output file to do incremental profiling"),
746 	OPT_STRING('C', "cpu", &cpu_list, "cpu",
747 		    "list of cpus to monitor"),
748 	OPT_BOOLEAN('f', "force", &force,
749 			"overwrite existing data file (deprecated)"),
750 	OPT_U64('c', "count", &user_interval, "event period to sample"),
751 	OPT_STRING('o', "output", &output_name, "file",
752 		    "output file name"),
753 	OPT_BOOLEAN('i', "no-inherit", &no_inherit,
754 		    "child tasks do not inherit counters"),
755 	OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
756 	OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
757 	OPT_BOOLEAN('g', "call-graph", &call_graph,
758 		    "do call-graph (stack chain/backtrace) recording"),
759 	OPT_INCR('v', "verbose", &verbose,
760 		    "be more verbose (show counter open errors, etc)"),
761 	OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
762 	OPT_BOOLEAN('s', "stat", &inherit_stat,
763 		    "per thread counts"),
764 	OPT_BOOLEAN('d', "data", &sample_address,
765 		    "Sample addresses"),
766 	OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
767 	OPT_BOOLEAN('n', "no-samples", &no_samples,
768 		    "don't sample"),
769 	OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
770 		    "do not update the buildid cache"),
771 	OPT_BOOLEAN('B', "no-buildid", &no_buildid,
772 		    "do not collect buildids in perf.data"),
773 	OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
774 		     "monitor event in cgroup name only",
775 		     parse_cgroups),
776 	OPT_END()
777 };
778 
779 int cmd_record(int argc, const char **argv, const char *prefix __used)
780 {
781 	int err = -ENOMEM;
782 	struct perf_evsel *pos;
783 
784 	evsel_list = perf_evlist__new(NULL, NULL);
785 	if (evsel_list == NULL)
786 		return -ENOMEM;
787 
788 	argc = parse_options(argc, argv, record_options, record_usage,
789 			    PARSE_OPT_STOP_AT_NON_OPTION);
790 	if (!argc && target_pid == -1 && target_tid == -1 &&
791 		!system_wide && !cpu_list)
792 		usage_with_options(record_usage, record_options);
793 
794 	if (force && append_file) {
795 		fprintf(stderr, "Can't overwrite and append at the same time."
796 				" You need to choose between -f and -A");
797 		usage_with_options(record_usage, record_options);
798 	} else if (append_file) {
799 		write_mode = WRITE_APPEND;
800 	} else {
801 		write_mode = WRITE_FORCE;
802 	}
803 
804 	if (nr_cgroups && !system_wide) {
805 		fprintf(stderr, "cgroup monitoring only available in"
806 			" system-wide mode\n");
807 		usage_with_options(record_usage, record_options);
808 	}
809 
810 	symbol__init();
811 
812 	if (no_buildid_cache || no_buildid)
813 		disable_buildid_cache();
814 
815 	if (evsel_list->nr_entries == 0 &&
816 	    perf_evlist__add_default(evsel_list) < 0) {
817 		pr_err("Not enough memory for event selector list\n");
818 		goto out_symbol_exit;
819 	}
820 
821 	if (target_pid != -1)
822 		target_tid = target_pid;
823 
824 	if (perf_evlist__create_maps(evsel_list, target_pid,
825 				     target_tid, cpu_list) < 0)
826 		usage_with_options(record_usage, record_options);
827 
828 	list_for_each_entry(pos, &evsel_list->entries, node) {
829 		if (perf_evsel__alloc_fd(pos, evsel_list->cpus->nr,
830 					 evsel_list->threads->nr) < 0)
831 			goto out_free_fd;
832 		if (perf_header__push_event(pos->attr.config, event_name(pos)))
833 			goto out_free_fd;
834 	}
835 
836 	if (perf_evlist__alloc_pollfd(evsel_list) < 0)
837 		goto out_free_fd;
838 
839 	if (user_interval != ULLONG_MAX)
840 		default_interval = user_interval;
841 	if (user_freq != UINT_MAX)
842 		freq = user_freq;
843 
844 	/*
845 	 * User specified count overrides default frequency.
846 	 */
847 	if (default_interval)
848 		freq = 0;
849 	else if (freq) {
850 		default_interval = freq;
851 	} else {
852 		fprintf(stderr, "frequency and count are zero, aborting\n");
853 		err = -EINVAL;
854 		goto out_free_fd;
855 	}
856 
857 	err = __cmd_record(argc, argv);
858 out_free_fd:
859 	perf_evlist__delete_maps(evsel_list);
860 out_symbol_exit:
861 	symbol__exit();
862 	return err;
863 }
864