?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
fs/ext4/inode.c: use pr_warn_ratelimited()
[linux/fpc-iii.git] / tools / perf / builtin-record.c
blob7069bd3e90b308871d4f309b99335cba09b9fbbc
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/evsel.h"
22 #include "util/debug.h"
23 #include "util/session.h"
24 #include "util/symbol.h"
25 #include "util/cpumap.h"
26
27 #include <unistd.h>
28 #include <sched.h>
29 #include <sys/mman.h>
30
31 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
32
33 enum write_mode_t {
34 WRITE_FORCE,
35 WRITE_APPEND
36 };
37
38 static u64 user_interval = ULLONG_MAX;
39 static u64 default_interval = 0;
40 static u64 sample_type;
41
42 static struct cpu_map *cpus;
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 = "perf.data";
50 static int group = 0;
51 static int realtime_prio = 0;
52 static bool raw_samples = false;
53 static bool sample_id_all_avail = true;
54 static bool system_wide = false;
55 static pid_t target_pid = -1;
56 static pid_t target_tid = -1;
57 static struct thread_map *threads;
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
69 static long samples = 0;
70 static u64 bytes_written = 0;
71
72 static struct pollfd *event_array;
73
74 static int nr_poll = 0;
75 static int nr_cpu = 0;
76
77 static int file_new = 1;
78 static off_t post_processing_offset;
79
80 static struct perf_session *session;
81 static const char *cpu_list;
82
83 struct mmap_data {
84 void *base;
85 unsigned int mask;
86 unsigned int prev;
87 };
88
89 static struct mmap_data mmap_array[MAX_NR_CPUS];
90
91 static unsigned long mmap_read_head(struct mmap_data *md)
92 {
93 struct perf_event_mmap_page *pc = md->base;
94 long head;
95
96 head = pc->data_head;
97 rmb();
98
99 return head;
100 }
101
102 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
103 {
104 struct perf_event_mmap_page *pc = md->base;
105
106 /*
107 * ensure all reads are done before we write the tail out.
108 */
109 /* mb(); */
110 pc->data_tail = tail;
111 }
112
113 static void advance_output(size_t size)
114 {
115 bytes_written += size;
116 }
117
118 static void write_output(void *buf, size_t size)
119 {
120 while (size) {
121 int ret = write(output, buf, size);
122
123 if (ret < 0)
124 die("failed to write");
125
126 size -= ret;
127 buf += ret;
128
129 bytes_written += ret;
130 }
131 }
132
133 static int process_synthesized_event(event_t *event,
134 struct sample_data *sample __used,
135 struct perf_session *self __used)
136 {
137 write_output(event, event->header.size);
138 return 0;
139 }
140
141 static void mmap_read(struct mmap_data *md)
142 {
143 unsigned int head = mmap_read_head(md);
144 unsigned int old = md->prev;
145 unsigned char *data = md->base + page_size;
146 unsigned long size;
147 void *buf;
148 int diff;
149
150 /*
151 * If we're further behind than half the buffer, there's a chance
152 * the writer will bite our tail and mess up the samples under us.
153 *
154 * If we somehow ended up ahead of the head, we got messed up.
155 *
156 * In either case, truncate and restart at head.
157 */
158 diff = head - old;
159 if (diff < 0) {
160 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
161 /*
162 * head points to a known good entry, start there.
163 */
164 old = head;
165 }
166
167 if (old != head)
168 samples++;
169
170 size = head - old;
171
172 if ((old & md->mask) + size != (head & md->mask)) {
173 buf = &data[old & md->mask];
174 size = md->mask + 1 - (old & md->mask);
175 old += size;
176
177 write_output(buf, size);
178 }
179
180 buf = &data[old & md->mask];
181 size = head - old;
182 old += size;
183
184 write_output(buf, size);
185
186 md->prev = old;
187 mmap_write_tail(md, old);
188 }
189
190 static volatile int done = 0;
191 static volatile int signr = -1;
192
193 static void sig_handler(int sig)
194 {
195 done = 1;
196 signr = sig;
197 }
198
199 static void sig_atexit(void)
200 {
201 if (child_pid > 0)
202 kill(child_pid, SIGTERM);
203
204 if (signr == -1 || signr == SIGUSR1)
205 return;
206
207 signal(signr, SIG_DFL);
208 kill(getpid(), signr);
209 }
210
211 static int group_fd;
212
213 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
214 {
215 struct perf_header_attr *h_attr;
216
217 if (nr < session->header.attrs) {
218 h_attr = session->header.attr[nr];
219 } else {
220 h_attr = perf_header_attr__new(a);
221 if (h_attr != NULL)
222 if (perf_header__add_attr(&session->header, h_attr) < 0) {
223 perf_header_attr__delete(h_attr);
224 h_attr = NULL;
225 }
226 }
227
228 return h_attr;
229 }
230
231 static void create_counter(struct perf_evsel *evsel, int cpu)
232 {
233 char *filter = evsel->filter;
234 struct perf_event_attr *attr = &evsel->attr;
235 struct perf_header_attr *h_attr;
236 int track = !evsel->idx; /* only the first counter needs these */
237 int thread_index;
238 int ret;
239 struct {
240 u64 count;
241 u64 time_enabled;
242 u64 time_running;
243 u64 id;
244 } read_data;
245 /*
246 * Check if parse_single_tracepoint_event has already asked for
247 * PERF_SAMPLE_TIME.
248 *
249 * XXX this is kludgy but short term fix for problems introduced by
250 * eac23d1c that broke 'perf script' by having different sample_types
251 * when using multiple tracepoint events when we use a perf binary
252 * that tries to use sample_id_all on an older kernel.
253 *
254 * We need to move counter creation to perf_session, support
255 * different sample_types, etc.
256 */
257 bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
258
259 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
260 PERF_FORMAT_TOTAL_TIME_RUNNING |
261 PERF_FORMAT_ID;
262
263 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
264
265 if (nr_counters > 1)
266 attr->sample_type |= PERF_SAMPLE_ID;
267
268 /*
269 * We default some events to a 1 default interval. But keep
270 * it a weak assumption overridable by the user.
271 */
272 if (!attr->sample_period || (user_freq != UINT_MAX &&
273 user_interval != ULLONG_MAX)) {
274 if (freq) {
275 attr->sample_type |= PERF_SAMPLE_PERIOD;
276 attr->freq = 1;
277 attr->sample_freq = freq;
278 } else {
279 attr->sample_period = default_interval;
280 }
281 }
282
283 if (no_samples)
284 attr->sample_freq = 0;
285
286 if (inherit_stat)
287 attr->inherit_stat = 1;
288
289 if (sample_address) {
290 attr->sample_type |= PERF_SAMPLE_ADDR;
291 attr->mmap_data = track;
292 }
293
294 if (call_graph)
295 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
296
297 if (system_wide)
298 attr->sample_type |= PERF_SAMPLE_CPU;
299
300 if (sample_id_all_avail &&
301 (sample_time || system_wide || !no_inherit || cpu_list))
302 attr->sample_type |= PERF_SAMPLE_TIME;
303
304 if (raw_samples) {
305 attr->sample_type |= PERF_SAMPLE_TIME;
306 attr->sample_type |= PERF_SAMPLE_RAW;
307 attr->sample_type |= PERF_SAMPLE_CPU;
308 }
309
310 attr->mmap = track;
311 attr->comm = track;
312 attr->inherit = !no_inherit;
313 if (target_pid == -1 && target_tid == -1 && !system_wide) {
314 attr->disabled = 1;
315 attr->enable_on_exec = 1;
316 }
317 retry_sample_id:
318 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
319
320 for (thread_index = 0; thread_index < threads->nr; thread_index++) {
321 try_again:
322 FD(evsel, nr_cpu, thread_index) = sys_perf_event_open(attr, threads->map[thread_index], cpu, group_fd, 0);
323
324 if (FD(evsel, nr_cpu, thread_index) < 0) {
325 int err = errno;
326
327 if (err == EPERM || err == EACCES)
328 die("Permission error - are you root?\n"
329 "\t Consider tweaking"
330 " /proc/sys/kernel/perf_event_paranoid.\n");
331 else if (err == ENODEV && cpu_list) {
332 die("No such device - did you specify"
333 " an out-of-range profile CPU?\n");
334 } else if (err == ENOENT) {
335 die("%s event is not supported. ",
336 event_name(evsel));
337 } else if (err == EINVAL && sample_id_all_avail) {
338 /*
339 * Old kernel, no attr->sample_id_type_all field
340 */
341 sample_id_all_avail = false;
342 if (!sample_time && !raw_samples && !time_needed)
343 attr->sample_type &= ~PERF_SAMPLE_TIME;
344
345 goto retry_sample_id;
346 }
347
348 /*
349 * If it's cycles then fall back to hrtimer
350 * based cpu-clock-tick sw counter, which
351 * is always available even if no PMU support:
352 */
353 if (attr->type == PERF_TYPE_HARDWARE
354 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
355
356 if (verbose)
357 warning(" ... trying to fall back to cpu-clock-ticks\n");
358 attr->type = PERF_TYPE_SOFTWARE;
359 attr->config = PERF_COUNT_SW_CPU_CLOCK;
360 goto try_again;
361 }
362 printf("\n");
363 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
364 FD(evsel, nr_cpu, thread_index), strerror(err));
365
366 #if defined(__i386__) || defined(__x86_64__)
367 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
368 die("No hardware sampling interrupt available."
369 " No APIC? If so then you can boot the kernel"
370 " with the \"lapic\" boot parameter to"
371 " force-enable it.\n");
372 #endif
373
374 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
375 exit(-1);
376 }
377
378 h_attr = get_header_attr(attr, evsel->idx);
379 if (h_attr == NULL)
380 die("nomem\n");
381
382 if (!file_new) {
383 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
384 fprintf(stderr, "incompatible append\n");
385 exit(-1);
386 }
387 }
388
389 if (read(FD(evsel, nr_cpu, thread_index), &read_data, sizeof(read_data)) == -1) {
390 perror("Unable to read perf file descriptor");
391 exit(-1);
392 }
393
394 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
395 pr_warning("Not enough memory to add id\n");
396 exit(-1);
397 }
398
399 assert(FD(evsel, nr_cpu, thread_index) >= 0);
400 fcntl(FD(evsel, nr_cpu, thread_index), F_SETFL, O_NONBLOCK);
401
402 /*
403 * First counter acts as the group leader:
404 */
405 if (group && group_fd == -1)
406 group_fd = FD(evsel, nr_cpu, thread_index);
407
408 if (evsel->idx || thread_index) {
409 struct perf_evsel *first;
410 first = list_entry(evsel_list.next, struct perf_evsel, node);
411 ret = ioctl(FD(evsel, nr_cpu, thread_index),
412 PERF_EVENT_IOC_SET_OUTPUT,
413 FD(first, nr_cpu, 0));
414 if (ret) {
415 error("failed to set output: %d (%s)\n", errno,
416 strerror(errno));
417 exit(-1);
418 }
419 } else {
420 mmap_array[nr_cpu].prev = 0;
421 mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
422 mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
423 PROT_READ | PROT_WRITE, MAP_SHARED, FD(evsel, nr_cpu, thread_index), 0);
424 if (mmap_array[nr_cpu].base == MAP_FAILED) {
425 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
426 exit(-1);
427 }
428
429 event_array[nr_poll].fd = FD(evsel, nr_cpu, thread_index);
430 event_array[nr_poll].events = POLLIN;
431 nr_poll++;
432 }
433
434 if (filter != NULL) {
435 ret = ioctl(FD(evsel, nr_cpu, thread_index),
436 PERF_EVENT_IOC_SET_FILTER, filter);
437 if (ret) {
438 error("failed to set filter with %d (%s)\n", errno,
439 strerror(errno));
440 exit(-1);
441 }
442 }
443 }
444
445 if (!sample_type)
446 sample_type = attr->sample_type;
447 }
448
449 static void open_counters(int cpu)
450 {
451 struct perf_evsel *pos;
452
453 group_fd = -1;
454
455 list_for_each_entry(pos, &evsel_list, node)
456 create_counter(pos, cpu);
457
458 nr_cpu++;
459 }
460
461 static int process_buildids(void)
462 {
463 u64 size = lseek(output, 0, SEEK_CUR);
464
465 if (size == 0)
466 return 0;
467
468 session->fd = output;
469 return __perf_session__process_events(session, post_processing_offset,
470 size - post_processing_offset,
471 size, &build_id__mark_dso_hit_ops);
472 }
473
474 static void atexit_header(void)
475 {
476 if (!pipe_output) {
477 session->header.data_size += bytes_written;
478
479 if (!no_buildid)
480 process_buildids();
481 perf_header__write(&session->header, output, true);
482 perf_session__delete(session);
483 symbol__exit();
484 }
485 }
486
487 static void event__synthesize_guest_os(struct machine *machine, void *data)
488 {
489 int err;
490 struct perf_session *psession = data;
491
492 if (machine__is_host(machine))
493 return;
494
495 /*
496 *As for guest kernel when processing subcommand record&report,
497 *we arrange module mmap prior to guest kernel mmap and trigger
498 *a preload dso because default guest module symbols are loaded
499 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
500 *method is used to avoid symbol missing when the first addr is
501 *in module instead of in guest kernel.
502 */
503 err = event__synthesize_modules(process_synthesized_event,
504 psession, machine);
505 if (err < 0)
506 pr_err("Couldn't record guest kernel [%d]'s reference"
507 " relocation symbol.\n", machine->pid);
508
509 /*
510 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
511 * have no _text sometimes.
512 */
513 err = event__synthesize_kernel_mmap(process_synthesized_event,
514 psession, machine, "_text");
515 if (err < 0)
516 err = event__synthesize_kernel_mmap(process_synthesized_event,
517 psession, machine, "_stext");
518 if (err < 0)
519 pr_err("Couldn't record guest kernel [%d]'s reference"
520 " relocation symbol.\n", machine->pid);
521 }
522
523 static struct perf_event_header finished_round_event = {
524 .size = sizeof(struct perf_event_header),
525 .type = PERF_RECORD_FINISHED_ROUND,
526 };
527
528 static void mmap_read_all(void)
529 {
530 int i;
531
532 for (i = 0; i < nr_cpu; i++) {
533 if (mmap_array[i].base)
534 mmap_read(&mmap_array[i]);
535 }
536
537 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
538 write_output(&finished_round_event, sizeof(finished_round_event));
539 }
540
541 static int __cmd_record(int argc, const char **argv)
542 {
543 int i;
544 struct stat st;
545 int flags;
546 int err;
547 unsigned long waking = 0;
548 int child_ready_pipe[2], go_pipe[2];
549 const bool forks = argc > 0;
550 char buf;
551 struct machine *machine;
552
553 page_size = sysconf(_SC_PAGE_SIZE);
554
555 atexit(sig_atexit);
556 signal(SIGCHLD, sig_handler);
557 signal(SIGINT, sig_handler);
558 signal(SIGUSR1, sig_handler);
559
560 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
561 perror("failed to create pipes");
562 exit(-1);
563 }
564
565 if (!strcmp(output_name, "-"))
566 pipe_output = 1;
567 else if (!stat(output_name, &st) && st.st_size) {
568 if (write_mode == WRITE_FORCE) {
569 char oldname[PATH_MAX];
570 snprintf(oldname, sizeof(oldname), "%s.old",
571 output_name);
572 unlink(oldname);
573 rename(output_name, oldname);
574 }
575 } else if (write_mode == WRITE_APPEND) {
576 write_mode = WRITE_FORCE;
577 }
578
579 flags = O_CREAT|O_RDWR;
580 if (write_mode == WRITE_APPEND)
581 file_new = 0;
582 else
583 flags |= O_TRUNC;
584
585 if (pipe_output)
586 output = STDOUT_FILENO;
587 else
588 output = open(output_name, flags, S_IRUSR | S_IWUSR);
589 if (output < 0) {
590 perror("failed to create output file");
591 exit(-1);
592 }
593
594 session = perf_session__new(output_name, O_WRONLY,
595 write_mode == WRITE_FORCE, false, NULL);
596 if (session == NULL) {
597 pr_err("Not enough memory for reading perf file header\n");
598 return -1;
599 }
600
601 if (!no_buildid)
602 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
603
604 if (!file_new) {
605 err = perf_header__read(session, output);
606 if (err < 0)
607 goto out_delete_session;
608 }
609
610 if (have_tracepoints(&evsel_list))
611 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
612
613 /*
614 * perf_session__delete(session) will be called at atexit_header()
615 */
616 atexit(atexit_header);
617
618 if (forks) {
619 child_pid = fork();
620 if (child_pid < 0) {
621 perror("failed to fork");
622 exit(-1);
623 }
624
625 if (!child_pid) {
626 if (pipe_output)
627 dup2(2, 1);
628 close(child_ready_pipe[0]);
629 close(go_pipe[1]);
630 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
631
632 /*
633 * Do a dummy execvp to get the PLT entry resolved,
634 * so we avoid the resolver overhead on the real
635 * execvp call.
636 */
637 execvp("", (char **)argv);
638
639 /*
640 * Tell the parent we're ready to go
641 */
642 close(child_ready_pipe[1]);
643
644 /*
645 * Wait until the parent tells us to go.
646 */
647 if (read(go_pipe[0], &buf, 1) == -1)
648 perror("unable to read pipe");
649
650 execvp(argv[0], (char **)argv);
651
652 perror(argv[0]);
653 kill(getppid(), SIGUSR1);
654 exit(-1);
655 }
656
657 if (!system_wide && target_tid == -1 && target_pid == -1)
658 threads->map[0] = child_pid;
659
660 close(child_ready_pipe[1]);
661 close(go_pipe[0]);
662 /*
663 * wait for child to settle
664 */
665 if (read(child_ready_pipe[0], &buf, 1) == -1) {
666 perror("unable to read pipe");
667 exit(-1);
668 }
669 close(child_ready_pipe[0]);
670 }
671
672 if (!system_wide && no_inherit && !cpu_list) {
673 open_counters(-1);
674 } else {
675 for (i = 0; i < cpus->nr; i++)
676 open_counters(cpus->map[i]);
677 }
678
679 perf_session__set_sample_type(session, sample_type);
680
681 if (pipe_output) {
682 err = perf_header__write_pipe(output);
683 if (err < 0)
684 return err;
685 } else if (file_new) {
686 err = perf_header__write(&session->header, output, false);
687 if (err < 0)
688 return err;
689 }
690
691 post_processing_offset = lseek(output, 0, SEEK_CUR);
692
693 perf_session__set_sample_id_all(session, sample_id_all_avail);
694
695 if (pipe_output) {
696 err = event__synthesize_attrs(&session->header,
697 process_synthesized_event,
698 session);
699 if (err < 0) {
700 pr_err("Couldn't synthesize attrs.\n");
701 return err;
702 }
703
704 err = event__synthesize_event_types(process_synthesized_event,
705 session);
706 if (err < 0) {
707 pr_err("Couldn't synthesize event_types.\n");
708 return err;
709 }
710
711 if (have_tracepoints(&evsel_list)) {
712 /*
713 * FIXME err <= 0 here actually means that
714 * there were no tracepoints so its not really
715 * an error, just that we don't need to
716 * synthesize anything. We really have to
717 * return this more properly and also
718 * propagate errors that now are calling die()
719 */
720 err = event__synthesize_tracing_data(output, &evsel_list,
721 process_synthesized_event,
722 session);
723 if (err <= 0) {
724 pr_err("Couldn't record tracing data.\n");
725 return err;
726 }
727 advance_output(err);
728 }
729 }
730
731 machine = perf_session__find_host_machine(session);
732 if (!machine) {
733 pr_err("Couldn't find native kernel information.\n");
734 return -1;
735 }
736
737 err = event__synthesize_kernel_mmap(process_synthesized_event,
738 session, machine, "_text");
739 if (err < 0)
740 err = event__synthesize_kernel_mmap(process_synthesized_event,
741 session, machine, "_stext");
742 if (err < 0)
743 pr_err("Couldn't record kernel reference relocation symbol\n"
744 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
745 "Check /proc/kallsyms permission or run as root.\n");
746
747 err = event__synthesize_modules(process_synthesized_event,
748 session, machine);
749 if (err < 0)
750 pr_err("Couldn't record kernel module information.\n"
751 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
752 "Check /proc/modules permission or run as root.\n");
753
754 if (perf_guest)
755 perf_session__process_machines(session, event__synthesize_guest_os);
756
757 if (!system_wide)
758 event__synthesize_thread(target_tid, process_synthesized_event,
759 session);
760 else
761 event__synthesize_threads(process_synthesized_event, session);
762
763 if (realtime_prio) {
764 struct sched_param param;
765
766 param.sched_priority = realtime_prio;
767 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
768 pr_err("Could not set realtime priority.\n");
769 exit(-1);
770 }
771 }
772
773 /*
774 * Let the child rip
775 */
776 if (forks)
777 close(go_pipe[1]);
778
779 for (;;) {
780 int hits = samples;
781 int thread;
782
783 mmap_read_all();
784
785 if (hits == samples) {
786 if (done)
787 break;
788 err = poll(event_array, nr_poll, -1);
789 waking++;
790 }
791
792 if (done) {
793 for (i = 0; i < nr_cpu; i++) {
794 struct perf_evsel *pos;
795
796 list_for_each_entry(pos, &evsel_list, node) {
797 for (thread = 0;
798 thread < threads->nr;
799 thread++)
800 ioctl(FD(pos, i, thread),
801 PERF_EVENT_IOC_DISABLE);
802 }
803 }
804 }
805 }
806
807 if (quiet || signr == SIGUSR1)
808 return 0;
809
810 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
811
812 /*
813 * Approximate RIP event size: 24 bytes.
814 */
815 fprintf(stderr,
816 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
817 (double)bytes_written / 1024.0 / 1024.0,
818 output_name,
819 bytes_written / 24);
820
821 return 0;
822
823 out_delete_session:
824 perf_session__delete(session);
825 return err;
826 }
827
828 static const char * const record_usage[] = {
829 "perf record [<options>] [<command>]",
830 "perf record [<options>] -- <command> [<options>]",
831 NULL
832 };
833
834 static bool force, append_file;
835
836 const struct option record_options[] = {
837 OPT_CALLBACK('e', "event", NULL, "event",
838 "event selector. use 'perf list' to list available events",
839 parse_events),
840 OPT_CALLBACK(0, "filter", NULL, "filter",
841 "event filter", parse_filter),
842 OPT_INTEGER('p', "pid", &target_pid,
843 "record events on existing process id"),
844 OPT_INTEGER('t', "tid", &target_tid,
845 "record events on existing thread id"),
846 OPT_INTEGER('r', "realtime", &realtime_prio,
847 "collect data with this RT SCHED_FIFO priority"),
848 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
849 "collect raw sample records from all opened counters"),
850 OPT_BOOLEAN('a', "all-cpus", &system_wide,
851 "system-wide collection from all CPUs"),
852 OPT_BOOLEAN('A', "append", &append_file,
853 "append to the output file to do incremental profiling"),
854 OPT_STRING('C', "cpu", &cpu_list, "cpu",
855 "list of cpus to monitor"),
856 OPT_BOOLEAN('f', "force", &force,
857 "overwrite existing data file (deprecated)"),
858 OPT_U64('c', "count", &user_interval, "event period to sample"),
859 OPT_STRING('o', "output", &output_name, "file",
860 "output file name"),
861 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
862 "child tasks do not inherit counters"),
863 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
864 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
865 OPT_BOOLEAN('g', "call-graph", &call_graph,
866 "do call-graph (stack chain/backtrace) recording"),
867 OPT_INCR('v', "verbose", &verbose,
868 "be more verbose (show counter open errors, etc)"),
869 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
870 OPT_BOOLEAN('s', "stat", &inherit_stat,
871 "per thread counts"),
872 OPT_BOOLEAN('d', "data", &sample_address,
873 "Sample addresses"),
874 OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
875 OPT_BOOLEAN('n', "no-samples", &no_samples,
876 "don't sample"),
877 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
878 "do not update the buildid cache"),
879 OPT_BOOLEAN('B', "no-buildid", &no_buildid,
880 "do not collect buildids in perf.data"),
881 OPT_END()
882 };
883
884 int cmd_record(int argc, const char **argv, const char *prefix __used)
885 {
886 int err = -ENOMEM;
887 struct perf_evsel *pos;
888
889 argc = parse_options(argc, argv, record_options, record_usage,
890 PARSE_OPT_STOP_AT_NON_OPTION);
891 if (!argc && target_pid == -1 && target_tid == -1 &&
892 !system_wide && !cpu_list)
893 usage_with_options(record_usage, record_options);
894
895 if (force && append_file) {
896 fprintf(stderr, "Can't overwrite and append at the same time."
897 " You need to choose between -f and -A");
898 usage_with_options(record_usage, record_options);
899 } else if (append_file) {
900 write_mode = WRITE_APPEND;
901 } else {
902 write_mode = WRITE_FORCE;
903 }
904
905 symbol__init();
906
907 if (no_buildid_cache || no_buildid)
908 disable_buildid_cache();
909
910 if (list_empty(&evsel_list) && perf_evsel_list__create_default() < 0) {
911 pr_err("Not enough memory for event selector list\n");
912 goto out_symbol_exit;
913 }
914
915 if (target_pid != -1)
916 target_tid = target_pid;
917
918 threads = thread_map__new(target_pid, target_tid);
919 if (threads == NULL) {
920 pr_err("Problems finding threads of monitor\n");
921 usage_with_options(record_usage, record_options);
922 }
923
924 cpus = cpu_map__new(cpu_list);
925 if (cpus == NULL) {
926 perror("failed to parse CPUs map");
927 return -1;
928 }
929
930 list_for_each_entry(pos, &evsel_list, node) {
931 if (perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0)
932 goto out_free_fd;
933 }
934 event_array = malloc((sizeof(struct pollfd) * MAX_NR_CPUS *
935 MAX_COUNTERS * threads->nr));
936 if (!event_array)
937 goto out_free_fd;
938
939 if (user_interval != ULLONG_MAX)
940 default_interval = user_interval;
941 if (user_freq != UINT_MAX)
942 freq = user_freq;
943
944 /*
945 * User specified count overrides default frequency.
946 */
947 if (default_interval)
948 freq = 0;
949 else if (freq) {
950 default_interval = freq;
951 } else {
952 fprintf(stderr, "frequency and count are zero, aborting\n");
953 err = -EINVAL;
954 goto out_free_event_array;
955 }
956
957 err = __cmd_record(argc, argv);
958
959 out_free_event_array:
960 free(event_array);
961 out_free_fd:
962 thread_map__delete(threads);
963 threads = NULL;
964 out_symbol_exit:
965 symbol__exit();
966 return err;
967 }
Linux with added FPC-III support