projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
perf record: Move perf_mmap__write_tail to perf.h
[deliverable/linux.git] / tools / perf / builtin-record.c
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
28 #include <unistd.h>
29 #include <sched.h>
30 #include <sys/mman.h>
31
32 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
33
34 enum write_mode_t {
35 WRITE_FORCE,
36 WRITE_APPEND
37 };
38
39 static u64 user_interval = ULLONG_MAX;
40 static u64 default_interval = 0;
41 static u64 sample_type;
42
43 static struct cpu_map *cpus;
44 static unsigned int page_size;
45 static unsigned int mmap_pages = 128;
46 static unsigned int user_freq = UINT_MAX;
47 static int freq = 1000;
48 static int output;
49 static int pipe_output = 0;
50 static const char *output_name = "perf.data";
51 static int group = 0;
52 static int realtime_prio = 0;
53 static bool nodelay = false;
54 static bool raw_samples = false;
55 static bool sample_id_all_avail = true;
56 static bool system_wide = false;
57 static pid_t target_pid = -1;
58 static pid_t target_tid = -1;
59 static struct thread_map *threads;
60 static pid_t child_pid = -1;
61 static bool no_inherit = false;
62 static enum write_mode_t write_mode = WRITE_FORCE;
63 static bool call_graph = false;
64 static bool inherit_stat = false;
65 static bool no_samples = false;
66 static bool sample_address = false;
67 static bool sample_time = false;
68 static bool no_buildid = false;
69 static bool no_buildid_cache = false;
70 static struct perf_evlist *evsel_list;
71
72 static long samples = 0;
73 static u64 bytes_written = 0;
74
75 static int file_new = 1;
76 static off_t post_processing_offset;
77
78 static struct perf_session *session;
79 static const char *cpu_list;
80
81 static struct perf_mmap mmap_array[MAX_NR_CPUS];
82
83 static void advance_output(size_t size)
84 {
85 bytes_written += size;
86 }
87
88 static void write_output(void *buf, size_t size)
89 {
90 while (size) {
91 int ret = write(output, buf, size);
92
93 if (ret < 0)
94 die("failed to write");
95
96 size -= ret;
97 buf += ret;
98
99 bytes_written += ret;
100 }
101 }
102
103 static int process_synthesized_event(event_t *event,
104 struct sample_data *sample __used,
105 struct perf_session *self __used)
106 {
107 write_output(event, event->header.size);
108 return 0;
109 }
110
111 static void mmap_read(struct perf_mmap *md)
112 {
113 unsigned int head = perf_mmap__read_head(md);
114 unsigned int old = md->prev;
115 unsigned char *data = md->base + page_size;
116 unsigned long size;
117 void *buf;
118 int diff;
119
120 /*
121 * If we're further behind than half the buffer, there's a chance
122 * the writer will bite our tail and mess up the samples under us.
123 *
124 * If we somehow ended up ahead of the head, we got messed up.
125 *
126 * In either case, truncate and restart at head.
127 */
128 diff = head - old;
129 if (diff < 0) {
130 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
131 /*
132 * head points to a known good entry, start there.
133 */
134 old = head;
135 }
136
137 if (old != head)
138 samples++;
139
140 size = head - old;
141
142 if ((old & md->mask) + size != (head & md->mask)) {
143 buf = &data[old & md->mask];
144 size = md->mask + 1 - (old & md->mask);
145 old += size;
146
147 write_output(buf, size);
148 }
149
150 buf = &data[old & md->mask];
151 size = head - old;
152 old += size;
153
154 write_output(buf, size);
155
156 md->prev = old;
157 perf_mmap__write_tail(md, old);
158 }
159
160 static volatile int done = 0;
161 static volatile int signr = -1;
162
163 static void sig_handler(int sig)
164 {
165 done = 1;
166 signr = sig;
167 }
168
169 static void sig_atexit(void)
170 {
171 if (child_pid > 0)
172 kill(child_pid, SIGTERM);
173
174 if (signr == -1 || signr == SIGUSR1)
175 return;
176
177 signal(signr, SIG_DFL);
178 kill(getpid(), signr);
179 }
180
181 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
182 {
183 struct perf_header_attr *h_attr;
184
185 if (nr < session->header.attrs) {
186 h_attr = session->header.attr[nr];
187 } else {
188 h_attr = perf_header_attr__new(a);
189 if (h_attr != NULL)
190 if (perf_header__add_attr(&session->header, h_attr) < 0) {
191 perf_header_attr__delete(h_attr);
192 h_attr = NULL;
193 }
194 }
195
196 return h_attr;
197 }
198
199 static void create_counter(struct perf_evlist *evlist,
200 struct perf_evsel *evsel, int cpu)
201 {
202 char *filter = evsel->filter;
203 struct perf_event_attr *attr = &evsel->attr;
204 struct perf_header_attr *h_attr;
205 int thread_index;
206 int ret;
207 struct {
208 u64 count;
209 u64 time_enabled;
210 u64 time_running;
211 u64 id;
212 } read_data;
213
214 for (thread_index = 0; thread_index < threads->nr; thread_index++) {
215 h_attr = get_header_attr(attr, evsel->idx);
216 if (h_attr == NULL)
217 die("nomem\n");
218
219 if (!file_new) {
220 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
221 fprintf(stderr, "incompatible append\n");
222 exit(-1);
223 }
224 }
225
226 if (read(FD(evsel, cpu, thread_index), &read_data, sizeof(read_data)) == -1) {
227 perror("Unable to read perf file descriptor");
228 exit(-1);
229 }
230
231 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
232 pr_warning("Not enough memory to add id\n");
233 exit(-1);
234 }
235
236 assert(FD(evsel, cpu, thread_index) >= 0);
237 fcntl(FD(evsel, cpu, thread_index), F_SETFL, O_NONBLOCK);
238
239 if (evsel->idx || thread_index) {
240 struct perf_evsel *first;
241 first = list_entry(evlist->entries.next, struct perf_evsel, node);
242 ret = ioctl(FD(evsel, cpu, thread_index),
243 PERF_EVENT_IOC_SET_OUTPUT,
244 FD(first, cpu, 0));
245 if (ret) {
246 error("failed to set output: %d (%s)\n", errno,
247 strerror(errno));
248 exit(-1);
249 }
250 } else {
251 mmap_array[cpu].prev = 0;
252 mmap_array[cpu].mask = mmap_pages*page_size - 1;
253 mmap_array[cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
254 PROT_READ | PROT_WRITE, MAP_SHARED, FD(evsel, cpu, thread_index), 0);
255 if (mmap_array[cpu].base == MAP_FAILED) {
256 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
257 exit(-1);
258 }
259
260 evlist->pollfd[evlist->nr_fds].fd = FD(evsel, cpu, thread_index);
261 evlist->pollfd[evlist->nr_fds].events = POLLIN;
262 evlist->nr_fds++;
263 }
264
265 if (filter != NULL) {
266 ret = ioctl(FD(evsel, cpu, thread_index),
267 PERF_EVENT_IOC_SET_FILTER, filter);
268 if (ret) {
269 error("failed to set filter with %d (%s)\n", errno,
270 strerror(errno));
271 exit(-1);
272 }
273 }
274 }
275
276 if (!sample_type)
277 sample_type = attr->sample_type;
278 }
279
280 static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
281 {
282 struct perf_event_attr *attr = &evsel->attr;
283 int track = !evsel->idx; /* only the first counter needs these */
284
285 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
286 PERF_FORMAT_TOTAL_TIME_RUNNING |
287 PERF_FORMAT_ID;
288
289 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
290
291 if (evlist->nr_entries > 1)
292 attr->sample_type |= PERF_SAMPLE_ID;
293
294 /*
295 * We default some events to a 1 default interval. But keep
296 * it a weak assumption overridable by the user.
297 */
298 if (!attr->sample_period || (user_freq != UINT_MAX &&
299 user_interval != ULLONG_MAX)) {
300 if (freq) {
301 attr->sample_type |= PERF_SAMPLE_PERIOD;
302 attr->freq = 1;
303 attr->sample_freq = freq;
304 } else {
305 attr->sample_period = default_interval;
306 }
307 }
308
309 if (no_samples)
310 attr->sample_freq = 0;
311
312 if (inherit_stat)
313 attr->inherit_stat = 1;
314
315 if (sample_address) {
316 attr->sample_type |= PERF_SAMPLE_ADDR;
317 attr->mmap_data = track;
318 }
319
320 if (call_graph)
321 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
322
323 if (system_wide)
324 attr->sample_type |= PERF_SAMPLE_CPU;
325
326 if (sample_id_all_avail &&
327 (sample_time || system_wide || !no_inherit || cpu_list))
328 attr->sample_type |= PERF_SAMPLE_TIME;
329
330 if (raw_samples) {
331 attr->sample_type |= PERF_SAMPLE_TIME;
332 attr->sample_type |= PERF_SAMPLE_RAW;
333 attr->sample_type |= PERF_SAMPLE_CPU;
334 }
335
336 if (nodelay) {
337 attr->watermark = 0;
338 attr->wakeup_events = 1;
339 }
340
341 attr->mmap = track;
342 attr->comm = track;
343
344 if (target_pid == -1 && target_tid == -1 && !system_wide) {
345 attr->disabled = 1;
346 attr->enable_on_exec = 1;
347 }
348 }
349
350 static void open_counters(struct perf_evlist *evlist)
351 {
352 struct perf_evsel *pos;
353 int cpu;
354
355 list_for_each_entry(pos, &evlist->entries, node) {
356 struct perf_event_attr *attr = &pos->attr;
357 /*
358 * Check if parse_single_tracepoint_event has already asked for
359 * PERF_SAMPLE_TIME.
360 *
361 * XXX this is kludgy but short term fix for problems introduced by
362 * eac23d1c that broke 'perf script' by having different sample_types
363 * when using multiple tracepoint events when we use a perf binary
364 * that tries to use sample_id_all on an older kernel.
365 *
366 * We need to move counter creation to perf_session, support
367 * different sample_types, etc.
368 */
369 bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
370
371 config_attr(pos, evlist);
372 retry_sample_id:
373 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
374 try_again:
375 if (perf_evsel__open(pos, cpus, threads, group, !no_inherit) < 0) {
376 int err = errno;
377
378 if (err == EPERM || err == EACCES)
379 die("Permission error - are you root?\n"
380 "\t Consider tweaking"
381 " /proc/sys/kernel/perf_event_paranoid.\n");
382 else if (err == ENODEV && cpu_list) {
383 die("No such device - did you specify"
384 " an out-of-range profile CPU?\n");
385 } else if (err == EINVAL && sample_id_all_avail) {
386 /*
387 * Old kernel, no attr->sample_id_type_all field
388 */
389 sample_id_all_avail = false;
390 if (!sample_time && !raw_samples && !time_needed)
391 attr->sample_type &= ~PERF_SAMPLE_TIME;
392
393 goto retry_sample_id;
394 }
395
396 /*
397 * If it's cycles then fall back to hrtimer
398 * based cpu-clock-tick sw counter, which
399 * is always available even if no PMU support:
400 */
401 if (attr->type == PERF_TYPE_HARDWARE
402 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
403
404 if (verbose)
405 warning(" ... trying to fall back to cpu-clock-ticks\n");
406 attr->type = PERF_TYPE_SOFTWARE;
407 attr->config = PERF_COUNT_SW_CPU_CLOCK;
408 goto try_again;
409 }
410 printf("\n");
411 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
412 err, strerror(err));
413
414 #if defined(__i386__) || defined(__x86_64__)
415 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
416 die("No hardware sampling interrupt available."
417 " No APIC? If so then you can boot the kernel"
418 " with the \"lapic\" boot parameter to"
419 " force-enable it.\n");
420 #endif
421
422 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
423 }
424 }
425
426 for (cpu = 0; cpu < cpus->nr; ++cpu) {
427 list_for_each_entry(pos, &evlist->entries, node)
428 create_counter(evlist, pos, cpu);
429 }
430 }
431
432 static int process_buildids(void)
433 {
434 u64 size = lseek(output, 0, SEEK_CUR);
435
436 if (size == 0)
437 return 0;
438
439 session->fd = output;
440 return __perf_session__process_events(session, post_processing_offset,
441 size - post_processing_offset,
442 size, &build_id__mark_dso_hit_ops);
443 }
444
445 static void atexit_header(void)
446 {
447 if (!pipe_output) {
448 session->header.data_size += bytes_written;
449
450 if (!no_buildid)
451 process_buildids();
452 perf_header__write(&session->header, evsel_list, output, true);
453 perf_session__delete(session);
454 perf_evlist__delete(evsel_list);
455 symbol__exit();
456 }
457 }
458
459 static void event__synthesize_guest_os(struct machine *machine, void *data)
460 {
461 int err;
462 struct perf_session *psession = data;
463
464 if (machine__is_host(machine))
465 return;
466
467 /*
468 *As for guest kernel when processing subcommand record&report,
469 *we arrange module mmap prior to guest kernel mmap and trigger
470 *a preload dso because default guest module symbols are loaded
471 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
472 *method is used to avoid symbol missing when the first addr is
473 *in module instead of in guest kernel.
474 */
475 err = event__synthesize_modules(process_synthesized_event,
476 psession, machine);
477 if (err < 0)
478 pr_err("Couldn't record guest kernel [%d]'s reference"
479 " relocation symbol.\n", machine->pid);
480
481 /*
482 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
483 * have no _text sometimes.
484 */
485 err = event__synthesize_kernel_mmap(process_synthesized_event,
486 psession, machine, "_text");
487 if (err < 0)
488 err = event__synthesize_kernel_mmap(process_synthesized_event,
489 psession, machine, "_stext");
490 if (err < 0)
491 pr_err("Couldn't record guest kernel [%d]'s reference"
492 " relocation symbol.\n", machine->pid);
493 }
494
495 static struct perf_event_header finished_round_event = {
496 .size = sizeof(struct perf_event_header),
497 .type = PERF_RECORD_FINISHED_ROUND,
498 };
499
500 static void mmap_read_all(void)
501 {
502 int i;
503
504 for (i = 0; i < cpus->nr; i++) {
505 if (mmap_array[i].base)
506 mmap_read(&mmap_array[i]);
507 }
508
509 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
510 write_output(&finished_round_event, sizeof(finished_round_event));
511 }
512
513 static int __cmd_record(int argc, const char **argv)
514 {
515 int i;
516 struct stat st;
517 int flags;
518 int err;
519 unsigned long waking = 0;
520 int child_ready_pipe[2], go_pipe[2];
521 const bool forks = argc > 0;
522 char buf;
523 struct machine *machine;
524
525 page_size = sysconf(_SC_PAGE_SIZE);
526
527 atexit(sig_atexit);
528 signal(SIGCHLD, sig_handler);
529 signal(SIGINT, sig_handler);
530 signal(SIGUSR1, sig_handler);
531
532 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
533 perror("failed to create pipes");
534 exit(-1);
535 }
536
537 if (!strcmp(output_name, "-"))
538 pipe_output = 1;
539 else if (!stat(output_name, &st) && st.st_size) {
540 if (write_mode == WRITE_FORCE) {
541 char oldname[PATH_MAX];
542 snprintf(oldname, sizeof(oldname), "%s.old",
543 output_name);
544 unlink(oldname);
545 rename(output_name, oldname);
546 }
547 } else if (write_mode == WRITE_APPEND) {
548 write_mode = WRITE_FORCE;
549 }
550
551 flags = O_CREAT|O_RDWR;
552 if (write_mode == WRITE_APPEND)
553 file_new = 0;
554 else
555 flags |= O_TRUNC;
556
557 if (pipe_output)
558 output = STDOUT_FILENO;
559 else
560 output = open(output_name, flags, S_IRUSR | S_IWUSR);
561 if (output < 0) {
562 perror("failed to create output file");
563 exit(-1);
564 }
565
566 session = perf_session__new(output_name, O_WRONLY,
567 write_mode == WRITE_FORCE, false, NULL);
568 if (session == NULL) {
569 pr_err("Not enough memory for reading perf file header\n");
570 return -1;
571 }
572
573 if (!no_buildid)
574 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
575
576 if (!file_new) {
577 err = perf_header__read(session, output);
578 if (err < 0)
579 goto out_delete_session;
580 }
581
582 if (have_tracepoints(&evsel_list->entries))
583 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
584
585 /*
586 * perf_session__delete(session) will be called at atexit_header()
587 */
588 atexit(atexit_header);
589
590 if (forks) {
591 child_pid = fork();
592 if (child_pid < 0) {
593 perror("failed to fork");
594 exit(-1);
595 }
596
597 if (!child_pid) {
598 if (pipe_output)
599 dup2(2, 1);
600 close(child_ready_pipe[0]);
601 close(go_pipe[1]);
602 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
603
604 /*
605 * Do a dummy execvp to get the PLT entry resolved,
606 * so we avoid the resolver overhead on the real
607 * execvp call.
608 */
609 execvp("", (char **)argv);
610
611 /*
612 * Tell the parent we're ready to go
613 */
614 close(child_ready_pipe[1]);
615
616 /*
617 * Wait until the parent tells us to go.
618 */
619 if (read(go_pipe[0], &buf, 1) == -1)
620 perror("unable to read pipe");
621
622 execvp(argv[0], (char **)argv);
623
624 perror(argv[0]);
625 kill(getppid(), SIGUSR1);
626 exit(-1);
627 }
628
629 if (!system_wide && target_tid == -1 && target_pid == -1)
630 threads->map[0] = child_pid;
631
632 close(child_ready_pipe[1]);
633 close(go_pipe[0]);
634 /*
635 * wait for child to settle
636 */
637 if (read(child_ready_pipe[0], &buf, 1) == -1) {
638 perror("unable to read pipe");
639 exit(-1);
640 }
641 close(child_ready_pipe[0]);
642 }
643
644 open_counters(evsel_list);
645
646 perf_session__set_sample_type(session, sample_type);
647
648 if (pipe_output) {
649 err = perf_header__write_pipe(output);
650 if (err < 0)
651 return err;
652 } else if (file_new) {
653 err = perf_header__write(&session->header, evsel_list,
654 output, false);
655 if (err < 0)
656 return err;
657 }
658
659 post_processing_offset = lseek(output, 0, SEEK_CUR);
660
661 perf_session__set_sample_id_all(session, sample_id_all_avail);
662
663 if (pipe_output) {
664 err = event__synthesize_attrs(&session->header,
665 process_synthesized_event,
666 session);
667 if (err < 0) {
668 pr_err("Couldn't synthesize attrs.\n");
669 return err;
670 }
671
672 err = event__synthesize_event_types(process_synthesized_event,
673 session);
674 if (err < 0) {
675 pr_err("Couldn't synthesize event_types.\n");
676 return err;
677 }
678
679 if (have_tracepoints(&evsel_list->entries)) {
680 /*
681 * FIXME err <= 0 here actually means that
682 * there were no tracepoints so its not really
683 * an error, just that we don't need to
684 * synthesize anything. We really have to
685 * return this more properly and also
686 * propagate errors that now are calling die()
687 */
688 err = event__synthesize_tracing_data(output, evsel_list,
689 process_synthesized_event,
690 session);
691 if (err <= 0) {
692 pr_err("Couldn't record tracing data.\n");
693 return err;
694 }
695 advance_output(err);
696 }
697 }
698
699 machine = perf_session__find_host_machine(session);
700 if (!machine) {
701 pr_err("Couldn't find native kernel information.\n");
702 return -1;
703 }
704
705 err = event__synthesize_kernel_mmap(process_synthesized_event,
706 session, machine, "_text");
707 if (err < 0)
708 err = event__synthesize_kernel_mmap(process_synthesized_event,
709 session, machine, "_stext");
710 if (err < 0)
711 pr_err("Couldn't record kernel reference relocation symbol\n"
712 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
713 "Check /proc/kallsyms permission or run as root.\n");
714
715 err = event__synthesize_modules(process_synthesized_event,
716 session, machine);
717 if (err < 0)
718 pr_err("Couldn't record kernel module information.\n"
719 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
720 "Check /proc/modules permission or run as root.\n");
721
722 if (perf_guest)
723 perf_session__process_machines(session, event__synthesize_guest_os);
724
725 if (!system_wide)
726 event__synthesize_thread(target_tid, process_synthesized_event,
727 session);
728 else
729 event__synthesize_threads(process_synthesized_event, session);
730
731 if (realtime_prio) {
732 struct sched_param param;
733
734 param.sched_priority = realtime_prio;
735 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
736 pr_err("Could not set realtime priority.\n");
737 exit(-1);
738 }
739 }
740
741 /*
742 * Let the child rip
743 */
744 if (forks)
745 close(go_pipe[1]);
746
747 for (;;) {
748 int hits = samples;
749 int thread;
750
751 mmap_read_all();
752
753 if (hits == samples) {
754 if (done)
755 break;
756 err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
757 waking++;
758 }
759
760 if (done) {
761 for (i = 0; i < cpus->nr; i++) {
762 struct perf_evsel *pos;
763
764 list_for_each_entry(pos, &evsel_list->entries, node) {
765 for (thread = 0;
766 thread < threads->nr;
767 thread++)
768 ioctl(FD(pos, i, thread),
769 PERF_EVENT_IOC_DISABLE);
770 }
771 }
772 }
773 }
774
775 if (quiet || signr == SIGUSR1)
776 return 0;
777
778 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
779
780 /*
781 * Approximate RIP event size: 24 bytes.
782 */
783 fprintf(stderr,
784 "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
785 (double)bytes_written / 1024.0 / 1024.0,
786 output_name,
787 bytes_written / 24);
788
789 return 0;
790
791 out_delete_session:
792 perf_session__delete(session);
793 return err;
794 }
795
796 static const char * const record_usage[] = {
797 "perf record [<options>] [<command>]",
798 "perf record [<options>] -- <command> [<options>]",
799 NULL
800 };
801
802 static bool force, append_file;
803
804 const struct option record_options[] = {
805 OPT_CALLBACK('e', "event", &evsel_list, "event",
806 "event selector. use 'perf list' to list available events",
807 parse_events),
808 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
809 "event filter", parse_filter),
810 OPT_INTEGER('p', "pid", &target_pid,
811 "record events on existing process id"),
812 OPT_INTEGER('t', "tid", &target_tid,
813 "record events on existing thread id"),
814 OPT_INTEGER('r', "realtime", &realtime_prio,
815 "collect data with this RT SCHED_FIFO priority"),
816 OPT_BOOLEAN('D', "no-delay", &nodelay,
817 "collect data without buffering"),
818 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
819 "collect raw sample records from all opened counters"),
820 OPT_BOOLEAN('a', "all-cpus", &system_wide,
821 "system-wide collection from all CPUs"),
822 OPT_BOOLEAN('A', "append", &append_file,
823 "append to the output file to do incremental profiling"),
824 OPT_STRING('C', "cpu", &cpu_list, "cpu",
825 "list of cpus to monitor"),
826 OPT_BOOLEAN('f', "force", &force,
827 "overwrite existing data file (deprecated)"),
828 OPT_U64('c', "count", &user_interval, "event period to sample"),
829 OPT_STRING('o', "output", &output_name, "file",
830 "output file name"),
831 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
832 "child tasks do not inherit counters"),
833 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
834 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
835 OPT_BOOLEAN('g', "call-graph", &call_graph,
836 "do call-graph (stack chain/backtrace) recording"),
837 OPT_INCR('v', "verbose", &verbose,
838 "be more verbose (show counter open errors, etc)"),
839 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
840 OPT_BOOLEAN('s', "stat", &inherit_stat,
841 "per thread counts"),
842 OPT_BOOLEAN('d', "data", &sample_address,
843 "Sample addresses"),
844 OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
845 OPT_BOOLEAN('n', "no-samples", &no_samples,
846 "don't sample"),
847 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
848 "do not update the buildid cache"),
849 OPT_BOOLEAN('B', "no-buildid", &no_buildid,
850 "do not collect buildids in perf.data"),
851 OPT_END()
852 };
853
854 int cmd_record(int argc, const char **argv, const char *prefix __used)
855 {
856 int err = -ENOMEM;
857 struct perf_evsel *pos;
858
859 evsel_list = perf_evlist__new();
860 if (evsel_list == NULL)
861 return -ENOMEM;
862
863 argc = parse_options(argc, argv, record_options, record_usage,
864 PARSE_OPT_STOP_AT_NON_OPTION);
865 if (!argc && target_pid == -1 && target_tid == -1 &&
866 !system_wide && !cpu_list)
867 usage_with_options(record_usage, record_options);
868
869 if (force && append_file) {
870 fprintf(stderr, "Can't overwrite and append at the same time."
871 " You need to choose between -f and -A");
872 usage_with_options(record_usage, record_options);
873 } else if (append_file) {
874 write_mode = WRITE_APPEND;
875 } else {
876 write_mode = WRITE_FORCE;
877 }
878
879 symbol__init();
880
881 if (no_buildid_cache || no_buildid)
882 disable_buildid_cache();
883
884 if (evsel_list->nr_entries == 0 &&
885 perf_evlist__add_default(evsel_list) < 0) {
886 pr_err("Not enough memory for event selector list\n");
887 goto out_symbol_exit;
888 }
889
890 if (target_pid != -1)
891 target_tid = target_pid;
892
893 threads = thread_map__new(target_pid, target_tid);
894 if (threads == NULL) {
895 pr_err("Problems finding threads of monitor\n");
896 usage_with_options(record_usage, record_options);
897 }
898
899 if (target_tid != -1)
900 cpus = cpu_map__dummy_new();
901 else
902 cpus = cpu_map__new(cpu_list);
903
904 if (cpus == NULL)
905 usage_with_options(record_usage, record_options);
906
907 list_for_each_entry(pos, &evsel_list->entries, node) {
908 if (perf_evsel__alloc_fd(pos, cpus->nr, threads->nr) < 0)
909 goto out_free_fd;
910 if (perf_header__push_event(pos->attr.config, event_name(pos)))
911 goto out_free_fd;
912 }
913
914 if (perf_evlist__alloc_pollfd(evsel_list, cpus->nr, threads->nr) < 0)
915 goto out_free_fd;
916
917 if (user_interval != ULLONG_MAX)
918 default_interval = user_interval;
919 if (user_freq != UINT_MAX)
920 freq = user_freq;
921
922 /*
923 * User specified count overrides default frequency.
924 */
925 if (default_interval)
926 freq = 0;
927 else if (freq) {
928 default_interval = freq;
929 } else {
930 fprintf(stderr, "frequency and count are zero, aborting\n");
931 err = -EINVAL;
932 goto out_free_fd;
933 }
934
935 err = __cmd_record(argc, argv);
936
937 out_free_fd:
938 thread_map__delete(threads);
939 threads = NULL;
940 out_symbol_exit:
941 symbol__exit();
942 return err;
943 }
Linux kernel - Deliverables
This page took 0.050441 seconds and 5 git commands to generate.