Loading...
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/build-id.h"
13#include "util/util.h"
14#include <subcmd/parse-options.h>
15#include "util/parse-events.h"
16
17#include "util/callchain.h"
18#include "util/cgroup.h"
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/tool.h"
26#include "util/symbol.h"
27#include "util/cpumap.h"
28#include "util/thread_map.h"
29#include "util/data.h"
30#include "util/perf_regs.h"
31#include "util/auxtrace.h"
32#include "util/parse-branch-options.h"
33#include "util/parse-regs-options.h"
34#include "util/llvm-utils.h"
35#include "util/bpf-loader.h"
36#include "asm/bug.h"
37
38#include <unistd.h>
39#include <sched.h>
40#include <sys/mman.h>
41
42
43struct record {
44 struct perf_tool tool;
45 struct record_opts opts;
46 u64 bytes_written;
47 struct perf_data_file file;
48 struct auxtrace_record *itr;
49 struct perf_evlist *evlist;
50 struct perf_session *session;
51 const char *progname;
52 int realtime_prio;
53 bool no_buildid;
54 bool no_buildid_set;
55 bool no_buildid_cache;
56 bool no_buildid_cache_set;
57 bool buildid_all;
58 unsigned long long samples;
59};
60
61static int record__write(struct record *rec, void *bf, size_t size)
62{
63 if (perf_data_file__write(rec->session->file, bf, size) < 0) {
64 pr_err("failed to write perf data, error: %m\n");
65 return -1;
66 }
67
68 rec->bytes_written += size;
69 return 0;
70}
71
72static int process_synthesized_event(struct perf_tool *tool,
73 union perf_event *event,
74 struct perf_sample *sample __maybe_unused,
75 struct machine *machine __maybe_unused)
76{
77 struct record *rec = container_of(tool, struct record, tool);
78 return record__write(rec, event, event->header.size);
79}
80
81static int record__mmap_read(struct record *rec, int idx)
82{
83 struct perf_mmap *md = &rec->evlist->mmap[idx];
84 u64 head = perf_mmap__read_head(md);
85 u64 old = md->prev;
86 unsigned char *data = md->base + page_size;
87 unsigned long size;
88 void *buf;
89 int rc = 0;
90
91 if (old == head)
92 return 0;
93
94 rec->samples++;
95
96 size = head - old;
97
98 if ((old & md->mask) + size != (head & md->mask)) {
99 buf = &data[old & md->mask];
100 size = md->mask + 1 - (old & md->mask);
101 old += size;
102
103 if (record__write(rec, buf, size) < 0) {
104 rc = -1;
105 goto out;
106 }
107 }
108
109 buf = &data[old & md->mask];
110 size = head - old;
111 old += size;
112
113 if (record__write(rec, buf, size) < 0) {
114 rc = -1;
115 goto out;
116 }
117
118 md->prev = old;
119 perf_evlist__mmap_consume(rec->evlist, idx);
120out:
121 return rc;
122}
123
124static volatile int done;
125static volatile int signr = -1;
126static volatile int child_finished;
127static volatile int auxtrace_snapshot_enabled;
128static volatile int auxtrace_snapshot_err;
129static volatile int auxtrace_record__snapshot_started;
130
131static void sig_handler(int sig)
132{
133 if (sig == SIGCHLD)
134 child_finished = 1;
135 else
136 signr = sig;
137
138 done = 1;
139}
140
141static void record__sig_exit(void)
142{
143 if (signr == -1)
144 return;
145
146 signal(signr, SIG_DFL);
147 raise(signr);
148}
149
150#ifdef HAVE_AUXTRACE_SUPPORT
151
152static int record__process_auxtrace(struct perf_tool *tool,
153 union perf_event *event, void *data1,
154 size_t len1, void *data2, size_t len2)
155{
156 struct record *rec = container_of(tool, struct record, tool);
157 struct perf_data_file *file = &rec->file;
158 size_t padding;
159 u8 pad[8] = {0};
160
161 if (!perf_data_file__is_pipe(file)) {
162 off_t file_offset;
163 int fd = perf_data_file__fd(file);
164 int err;
165
166 file_offset = lseek(fd, 0, SEEK_CUR);
167 if (file_offset == -1)
168 return -1;
169 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
170 event, file_offset);
171 if (err)
172 return err;
173 }
174
175 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
176 padding = (len1 + len2) & 7;
177 if (padding)
178 padding = 8 - padding;
179
180 record__write(rec, event, event->header.size);
181 record__write(rec, data1, len1);
182 if (len2)
183 record__write(rec, data2, len2);
184 record__write(rec, &pad, padding);
185
186 return 0;
187}
188
189static int record__auxtrace_mmap_read(struct record *rec,
190 struct auxtrace_mmap *mm)
191{
192 int ret;
193
194 ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
195 record__process_auxtrace);
196 if (ret < 0)
197 return ret;
198
199 if (ret)
200 rec->samples++;
201
202 return 0;
203}
204
205static int record__auxtrace_mmap_read_snapshot(struct record *rec,
206 struct auxtrace_mmap *mm)
207{
208 int ret;
209
210 ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
211 record__process_auxtrace,
212 rec->opts.auxtrace_snapshot_size);
213 if (ret < 0)
214 return ret;
215
216 if (ret)
217 rec->samples++;
218
219 return 0;
220}
221
222static int record__auxtrace_read_snapshot_all(struct record *rec)
223{
224 int i;
225 int rc = 0;
226
227 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
228 struct auxtrace_mmap *mm =
229 &rec->evlist->mmap[i].auxtrace_mmap;
230
231 if (!mm->base)
232 continue;
233
234 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
235 rc = -1;
236 goto out;
237 }
238 }
239out:
240 return rc;
241}
242
243static void record__read_auxtrace_snapshot(struct record *rec)
244{
245 pr_debug("Recording AUX area tracing snapshot\n");
246 if (record__auxtrace_read_snapshot_all(rec) < 0) {
247 auxtrace_snapshot_err = -1;
248 } else {
249 auxtrace_snapshot_err = auxtrace_record__snapshot_finish(rec->itr);
250 if (!auxtrace_snapshot_err)
251 auxtrace_snapshot_enabled = 1;
252 }
253}
254
255#else
256
257static inline
258int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
259 struct auxtrace_mmap *mm __maybe_unused)
260{
261 return 0;
262}
263
264static inline
265void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
266{
267}
268
269static inline
270int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
271{
272 return 0;
273}
274
275#endif
276
277static int record__open(struct record *rec)
278{
279 char msg[512];
280 struct perf_evsel *pos;
281 struct perf_evlist *evlist = rec->evlist;
282 struct perf_session *session = rec->session;
283 struct record_opts *opts = &rec->opts;
284 int rc = 0;
285
286 perf_evlist__config(evlist, opts);
287
288 evlist__for_each(evlist, pos) {
289try_again:
290 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
291 if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
292 if (verbose)
293 ui__warning("%s\n", msg);
294 goto try_again;
295 }
296
297 rc = -errno;
298 perf_evsel__open_strerror(pos, &opts->target,
299 errno, msg, sizeof(msg));
300 ui__error("%s\n", msg);
301 goto out;
302 }
303 }
304
305 if (perf_evlist__apply_filters(evlist, &pos)) {
306 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
307 pos->filter, perf_evsel__name(pos), errno,
308 strerror_r(errno, msg, sizeof(msg)));
309 rc = -1;
310 goto out;
311 }
312
313 if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
314 opts->auxtrace_mmap_pages,
315 opts->auxtrace_snapshot_mode) < 0) {
316 if (errno == EPERM) {
317 pr_err("Permission error mapping pages.\n"
318 "Consider increasing "
319 "/proc/sys/kernel/perf_event_mlock_kb,\n"
320 "or try again with a smaller value of -m/--mmap_pages.\n"
321 "(current value: %u,%u)\n",
322 opts->mmap_pages, opts->auxtrace_mmap_pages);
323 rc = -errno;
324 } else {
325 pr_err("failed to mmap with %d (%s)\n", errno,
326 strerror_r(errno, msg, sizeof(msg)));
327 if (errno)
328 rc = -errno;
329 else
330 rc = -EINVAL;
331 }
332 goto out;
333 }
334
335 session->evlist = evlist;
336 perf_session__set_id_hdr_size(session);
337out:
338 return rc;
339}
340
341static int process_sample_event(struct perf_tool *tool,
342 union perf_event *event,
343 struct perf_sample *sample,
344 struct perf_evsel *evsel,
345 struct machine *machine)
346{
347 struct record *rec = container_of(tool, struct record, tool);
348
349 rec->samples++;
350
351 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
352}
353
354static int process_buildids(struct record *rec)
355{
356 struct perf_data_file *file = &rec->file;
357 struct perf_session *session = rec->session;
358
359 if (file->size == 0)
360 return 0;
361
362 /*
363 * During this process, it'll load kernel map and replace the
364 * dso->long_name to a real pathname it found. In this case
365 * we prefer the vmlinux path like
366 * /lib/modules/3.16.4/build/vmlinux
367 *
368 * rather than build-id path (in debug directory).
369 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
370 */
371 symbol_conf.ignore_vmlinux_buildid = true;
372
373 /*
374 * If --buildid-all is given, it marks all DSO regardless of hits,
375 * so no need to process samples.
376 */
377 if (rec->buildid_all)
378 rec->tool.sample = NULL;
379
380 return perf_session__process_events(session);
381}
382
383static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
384{
385 int err;
386 struct perf_tool *tool = data;
387 /*
388 *As for guest kernel when processing subcommand record&report,
389 *we arrange module mmap prior to guest kernel mmap and trigger
390 *a preload dso because default guest module symbols are loaded
391 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
392 *method is used to avoid symbol missing when the first addr is
393 *in module instead of in guest kernel.
394 */
395 err = perf_event__synthesize_modules(tool, process_synthesized_event,
396 machine);
397 if (err < 0)
398 pr_err("Couldn't record guest kernel [%d]'s reference"
399 " relocation symbol.\n", machine->pid);
400
401 /*
402 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
403 * have no _text sometimes.
404 */
405 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
406 machine);
407 if (err < 0)
408 pr_err("Couldn't record guest kernel [%d]'s reference"
409 " relocation symbol.\n", machine->pid);
410}
411
412static struct perf_event_header finished_round_event = {
413 .size = sizeof(struct perf_event_header),
414 .type = PERF_RECORD_FINISHED_ROUND,
415};
416
417static int record__mmap_read_all(struct record *rec)
418{
419 u64 bytes_written = rec->bytes_written;
420 int i;
421 int rc = 0;
422
423 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
424 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
425
426 if (rec->evlist->mmap[i].base) {
427 if (record__mmap_read(rec, i) != 0) {
428 rc = -1;
429 goto out;
430 }
431 }
432
433 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
434 record__auxtrace_mmap_read(rec, mm) != 0) {
435 rc = -1;
436 goto out;
437 }
438 }
439
440 /*
441 * Mark the round finished in case we wrote
442 * at least one event.
443 */
444 if (bytes_written != rec->bytes_written)
445 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
446
447out:
448 return rc;
449}
450
451static void record__init_features(struct record *rec)
452{
453 struct perf_session *session = rec->session;
454 int feat;
455
456 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
457 perf_header__set_feat(&session->header, feat);
458
459 if (rec->no_buildid)
460 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
461
462 if (!have_tracepoints(&rec->evlist->entries))
463 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
464
465 if (!rec->opts.branch_stack)
466 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
467
468 if (!rec->opts.full_auxtrace)
469 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
470
471 perf_header__clear_feat(&session->header, HEADER_STAT);
472}
473
474static void
475record__finish_output(struct record *rec)
476{
477 struct perf_data_file *file = &rec->file;
478 int fd = perf_data_file__fd(file);
479
480 if (file->is_pipe)
481 return;
482
483 rec->session->header.data_size += rec->bytes_written;
484 file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
485
486 if (!rec->no_buildid) {
487 process_buildids(rec);
488
489 if (rec->buildid_all)
490 dsos__hit_all(rec->session);
491 }
492 perf_session__write_header(rec->session, rec->evlist, fd, true);
493
494 return;
495}
496
497static volatile int workload_exec_errno;
498
499/*
500 * perf_evlist__prepare_workload will send a SIGUSR1
501 * if the fork fails, since we asked by setting its
502 * want_signal to true.
503 */
504static void workload_exec_failed_signal(int signo __maybe_unused,
505 siginfo_t *info,
506 void *ucontext __maybe_unused)
507{
508 workload_exec_errno = info->si_value.sival_int;
509 done = 1;
510 child_finished = 1;
511}
512
513static void snapshot_sig_handler(int sig);
514
515static int record__synthesize(struct record *rec)
516{
517 struct perf_session *session = rec->session;
518 struct machine *machine = &session->machines.host;
519 struct perf_data_file *file = &rec->file;
520 struct record_opts *opts = &rec->opts;
521 struct perf_tool *tool = &rec->tool;
522 int fd = perf_data_file__fd(file);
523 int err = 0;
524
525 if (file->is_pipe) {
526 err = perf_event__synthesize_attrs(tool, session,
527 process_synthesized_event);
528 if (err < 0) {
529 pr_err("Couldn't synthesize attrs.\n");
530 goto out;
531 }
532
533 if (have_tracepoints(&rec->evlist->entries)) {
534 /*
535 * FIXME err <= 0 here actually means that
536 * there were no tracepoints so its not really
537 * an error, just that we don't need to
538 * synthesize anything. We really have to
539 * return this more properly and also
540 * propagate errors that now are calling die()
541 */
542 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
543 process_synthesized_event);
544 if (err <= 0) {
545 pr_err("Couldn't record tracing data.\n");
546 goto out;
547 }
548 rec->bytes_written += err;
549 }
550 }
551
552 if (rec->opts.full_auxtrace) {
553 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
554 session, process_synthesized_event);
555 if (err)
556 goto out;
557 }
558
559 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
560 machine);
561 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
562 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
563 "Check /proc/kallsyms permission or run as root.\n");
564
565 err = perf_event__synthesize_modules(tool, process_synthesized_event,
566 machine);
567 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
568 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
569 "Check /proc/modules permission or run as root.\n");
570
571 if (perf_guest) {
572 machines__process_guests(&session->machines,
573 perf_event__synthesize_guest_os, tool);
574 }
575
576 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
577 process_synthesized_event, opts->sample_address,
578 opts->proc_map_timeout);
579out:
580 return err;
581}
582
583static int __cmd_record(struct record *rec, int argc, const char **argv)
584{
585 int err;
586 int status = 0;
587 unsigned long waking = 0;
588 const bool forks = argc > 0;
589 struct machine *machine;
590 struct perf_tool *tool = &rec->tool;
591 struct record_opts *opts = &rec->opts;
592 struct perf_data_file *file = &rec->file;
593 struct perf_session *session;
594 bool disabled = false, draining = false;
595 int fd;
596
597 rec->progname = argv[0];
598
599 atexit(record__sig_exit);
600 signal(SIGCHLD, sig_handler);
601 signal(SIGINT, sig_handler);
602 signal(SIGTERM, sig_handler);
603 if (rec->opts.auxtrace_snapshot_mode)
604 signal(SIGUSR2, snapshot_sig_handler);
605 else
606 signal(SIGUSR2, SIG_IGN);
607
608 session = perf_session__new(file, false, tool);
609 if (session == NULL) {
610 pr_err("Perf session creation failed.\n");
611 return -1;
612 }
613
614 fd = perf_data_file__fd(file);
615 rec->session = session;
616
617 record__init_features(rec);
618
619 if (forks) {
620 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
621 argv, file->is_pipe,
622 workload_exec_failed_signal);
623 if (err < 0) {
624 pr_err("Couldn't run the workload!\n");
625 status = err;
626 goto out_delete_session;
627 }
628 }
629
630 if (record__open(rec) != 0) {
631 err = -1;
632 goto out_child;
633 }
634
635 err = bpf__apply_obj_config();
636 if (err) {
637 char errbuf[BUFSIZ];
638
639 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
640 pr_err("ERROR: Apply config to BPF failed: %s\n",
641 errbuf);
642 goto out_child;
643 }
644
645 /*
646 * Normally perf_session__new would do this, but it doesn't have the
647 * evlist.
648 */
649 if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
650 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
651 rec->tool.ordered_events = false;
652 }
653
654 if (!rec->evlist->nr_groups)
655 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
656
657 if (file->is_pipe) {
658 err = perf_header__write_pipe(fd);
659 if (err < 0)
660 goto out_child;
661 } else {
662 err = perf_session__write_header(session, rec->evlist, fd, false);
663 if (err < 0)
664 goto out_child;
665 }
666
667 if (!rec->no_buildid
668 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
669 pr_err("Couldn't generate buildids. "
670 "Use --no-buildid to profile anyway.\n");
671 err = -1;
672 goto out_child;
673 }
674
675 machine = &session->machines.host;
676
677 err = record__synthesize(rec);
678 if (err < 0)
679 goto out_child;
680
681 if (rec->realtime_prio) {
682 struct sched_param param;
683
684 param.sched_priority = rec->realtime_prio;
685 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
686 pr_err("Could not set realtime priority.\n");
687 err = -1;
688 goto out_child;
689 }
690 }
691
692 /*
693 * When perf is starting the traced process, all the events
694 * (apart from group members) have enable_on_exec=1 set,
695 * so don't spoil it by prematurely enabling them.
696 */
697 if (!target__none(&opts->target) && !opts->initial_delay)
698 perf_evlist__enable(rec->evlist);
699
700 /*
701 * Let the child rip
702 */
703 if (forks) {
704 union perf_event *event;
705
706 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
707 if (event == NULL) {
708 err = -ENOMEM;
709 goto out_child;
710 }
711
712 /*
713 * Some H/W events are generated before COMM event
714 * which is emitted during exec(), so perf script
715 * cannot see a correct process name for those events.
716 * Synthesize COMM event to prevent it.
717 */
718 perf_event__synthesize_comm(tool, event,
719 rec->evlist->workload.pid,
720 process_synthesized_event,
721 machine);
722 free(event);
723
724 perf_evlist__start_workload(rec->evlist);
725 }
726
727 if (opts->initial_delay) {
728 usleep(opts->initial_delay * 1000);
729 perf_evlist__enable(rec->evlist);
730 }
731
732 auxtrace_snapshot_enabled = 1;
733 for (;;) {
734 unsigned long long hits = rec->samples;
735
736 if (record__mmap_read_all(rec) < 0) {
737 auxtrace_snapshot_enabled = 0;
738 err = -1;
739 goto out_child;
740 }
741
742 if (auxtrace_record__snapshot_started) {
743 auxtrace_record__snapshot_started = 0;
744 if (!auxtrace_snapshot_err)
745 record__read_auxtrace_snapshot(rec);
746 if (auxtrace_snapshot_err) {
747 pr_err("AUX area tracing snapshot failed\n");
748 err = -1;
749 goto out_child;
750 }
751 }
752
753 if (hits == rec->samples) {
754 if (done || draining)
755 break;
756 err = perf_evlist__poll(rec->evlist, -1);
757 /*
758 * Propagate error, only if there's any. Ignore positive
759 * number of returned events and interrupt error.
760 */
761 if (err > 0 || (err < 0 && errno == EINTR))
762 err = 0;
763 waking++;
764
765 if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
766 draining = true;
767 }
768
769 /*
770 * When perf is starting the traced process, at the end events
771 * die with the process and we wait for that. Thus no need to
772 * disable events in this case.
773 */
774 if (done && !disabled && !target__none(&opts->target)) {
775 auxtrace_snapshot_enabled = 0;
776 perf_evlist__disable(rec->evlist);
777 disabled = true;
778 }
779 }
780 auxtrace_snapshot_enabled = 0;
781
782 if (forks && workload_exec_errno) {
783 char msg[STRERR_BUFSIZE];
784 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
785 pr_err("Workload failed: %s\n", emsg);
786 err = -1;
787 goto out_child;
788 }
789
790 if (!quiet)
791 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
792
793out_child:
794 if (forks) {
795 int exit_status;
796
797 if (!child_finished)
798 kill(rec->evlist->workload.pid, SIGTERM);
799
800 wait(&exit_status);
801
802 if (err < 0)
803 status = err;
804 else if (WIFEXITED(exit_status))
805 status = WEXITSTATUS(exit_status);
806 else if (WIFSIGNALED(exit_status))
807 signr = WTERMSIG(exit_status);
808 } else
809 status = err;
810
811 /* this will be recalculated during process_buildids() */
812 rec->samples = 0;
813
814 if (!err)
815 record__finish_output(rec);
816
817 if (!err && !quiet) {
818 char samples[128];
819
820 if (rec->samples && !rec->opts.full_auxtrace)
821 scnprintf(samples, sizeof(samples),
822 " (%" PRIu64 " samples)", rec->samples);
823 else
824 samples[0] = '\0';
825
826 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
827 perf_data_file__size(file) / 1024.0 / 1024.0,
828 file->path, samples);
829 }
830
831out_delete_session:
832 perf_session__delete(session);
833 return status;
834}
835
836static void callchain_debug(void)
837{
838 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
839
840 pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
841
842 if (callchain_param.record_mode == CALLCHAIN_DWARF)
843 pr_debug("callchain: stack dump size %d\n",
844 callchain_param.dump_size);
845}
846
847int record_parse_callchain_opt(const struct option *opt,
848 const char *arg,
849 int unset)
850{
851 int ret;
852 struct record_opts *record = (struct record_opts *)opt->value;
853
854 record->callgraph_set = true;
855 callchain_param.enabled = !unset;
856
857 /* --no-call-graph */
858 if (unset) {
859 callchain_param.record_mode = CALLCHAIN_NONE;
860 pr_debug("callchain: disabled\n");
861 return 0;
862 }
863
864 ret = parse_callchain_record_opt(arg, &callchain_param);
865 if (!ret) {
866 /* Enable data address sampling for DWARF unwind. */
867 if (callchain_param.record_mode == CALLCHAIN_DWARF)
868 record->sample_address = true;
869 callchain_debug();
870 }
871
872 return ret;
873}
874
875int record_callchain_opt(const struct option *opt,
876 const char *arg __maybe_unused,
877 int unset __maybe_unused)
878{
879 struct record_opts *record = (struct record_opts *)opt->value;
880
881 record->callgraph_set = true;
882 callchain_param.enabled = true;
883
884 if (callchain_param.record_mode == CALLCHAIN_NONE)
885 callchain_param.record_mode = CALLCHAIN_FP;
886
887 callchain_debug();
888 return 0;
889}
890
891static int perf_record_config(const char *var, const char *value, void *cb)
892{
893 struct record *rec = cb;
894
895 if (!strcmp(var, "record.build-id")) {
896 if (!strcmp(value, "cache"))
897 rec->no_buildid_cache = false;
898 else if (!strcmp(value, "no-cache"))
899 rec->no_buildid_cache = true;
900 else if (!strcmp(value, "skip"))
901 rec->no_buildid = true;
902 else
903 return -1;
904 return 0;
905 }
906 if (!strcmp(var, "record.call-graph"))
907 var = "call-graph.record-mode"; /* fall-through */
908
909 return perf_default_config(var, value, cb);
910}
911
912struct clockid_map {
913 const char *name;
914 int clockid;
915};
916
917#define CLOCKID_MAP(n, c) \
918 { .name = n, .clockid = (c), }
919
920#define CLOCKID_END { .name = NULL, }
921
922
923/*
924 * Add the missing ones, we need to build on many distros...
925 */
926#ifndef CLOCK_MONOTONIC_RAW
927#define CLOCK_MONOTONIC_RAW 4
928#endif
929#ifndef CLOCK_BOOTTIME
930#define CLOCK_BOOTTIME 7
931#endif
932#ifndef CLOCK_TAI
933#define CLOCK_TAI 11
934#endif
935
936static const struct clockid_map clockids[] = {
937 /* available for all events, NMI safe */
938 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
939 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
940
941 /* available for some events */
942 CLOCKID_MAP("realtime", CLOCK_REALTIME),
943 CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
944 CLOCKID_MAP("tai", CLOCK_TAI),
945
946 /* available for the lazy */
947 CLOCKID_MAP("mono", CLOCK_MONOTONIC),
948 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
949 CLOCKID_MAP("real", CLOCK_REALTIME),
950 CLOCKID_MAP("boot", CLOCK_BOOTTIME),
951
952 CLOCKID_END,
953};
954
955static int parse_clockid(const struct option *opt, const char *str, int unset)
956{
957 struct record_opts *opts = (struct record_opts *)opt->value;
958 const struct clockid_map *cm;
959 const char *ostr = str;
960
961 if (unset) {
962 opts->use_clockid = 0;
963 return 0;
964 }
965
966 /* no arg passed */
967 if (!str)
968 return 0;
969
970 /* no setting it twice */
971 if (opts->use_clockid)
972 return -1;
973
974 opts->use_clockid = true;
975
976 /* if its a number, we're done */
977 if (sscanf(str, "%d", &opts->clockid) == 1)
978 return 0;
979
980 /* allow a "CLOCK_" prefix to the name */
981 if (!strncasecmp(str, "CLOCK_", 6))
982 str += 6;
983
984 for (cm = clockids; cm->name; cm++) {
985 if (!strcasecmp(str, cm->name)) {
986 opts->clockid = cm->clockid;
987 return 0;
988 }
989 }
990
991 opts->use_clockid = false;
992 ui__warning("unknown clockid %s, check man page\n", ostr);
993 return -1;
994}
995
996static int record__parse_mmap_pages(const struct option *opt,
997 const char *str,
998 int unset __maybe_unused)
999{
1000 struct record_opts *opts = opt->value;
1001 char *s, *p;
1002 unsigned int mmap_pages;
1003 int ret;
1004
1005 if (!str)
1006 return -EINVAL;
1007
1008 s = strdup(str);
1009 if (!s)
1010 return -ENOMEM;
1011
1012 p = strchr(s, ',');
1013 if (p)
1014 *p = '\0';
1015
1016 if (*s) {
1017 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1018 if (ret)
1019 goto out_free;
1020 opts->mmap_pages = mmap_pages;
1021 }
1022
1023 if (!p) {
1024 ret = 0;
1025 goto out_free;
1026 }
1027
1028 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1029 if (ret)
1030 goto out_free;
1031
1032 opts->auxtrace_mmap_pages = mmap_pages;
1033
1034out_free:
1035 free(s);
1036 return ret;
1037}
1038
1039static const char * const __record_usage[] = {
1040 "perf record [<options>] [<command>]",
1041 "perf record [<options>] -- <command> [<options>]",
1042 NULL
1043};
1044const char * const *record_usage = __record_usage;
1045
1046/*
1047 * XXX Ideally would be local to cmd_record() and passed to a record__new
1048 * because we need to have access to it in record__exit, that is called
1049 * after cmd_record() exits, but since record_options need to be accessible to
1050 * builtin-script, leave it here.
1051 *
1052 * At least we don't ouch it in all the other functions here directly.
1053 *
1054 * Just say no to tons of global variables, sigh.
1055 */
1056static struct record record = {
1057 .opts = {
1058 .sample_time = true,
1059 .mmap_pages = UINT_MAX,
1060 .user_freq = UINT_MAX,
1061 .user_interval = ULLONG_MAX,
1062 .freq = 4000,
1063 .target = {
1064 .uses_mmap = true,
1065 .default_per_cpu = true,
1066 },
1067 .proc_map_timeout = 500,
1068 },
1069 .tool = {
1070 .sample = process_sample_event,
1071 .fork = perf_event__process_fork,
1072 .exit = perf_event__process_exit,
1073 .comm = perf_event__process_comm,
1074 .mmap = perf_event__process_mmap,
1075 .mmap2 = perf_event__process_mmap2,
1076 .ordered_events = true,
1077 },
1078};
1079
1080const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1081 "\n\t\t\t\tDefault: fp";
1082
1083/*
1084 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1085 * with it and switch to use the library functions in perf_evlist that came
1086 * from builtin-record.c, i.e. use record_opts,
1087 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1088 * using pipes, etc.
1089 */
1090struct option __record_options[] = {
1091 OPT_CALLBACK('e', "event", &record.evlist, "event",
1092 "event selector. use 'perf list' to list available events",
1093 parse_events_option),
1094 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1095 "event filter", parse_filter),
1096 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1097 NULL, "don't record events from perf itself",
1098 exclude_perf),
1099 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1100 "record events on existing process id"),
1101 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1102 "record events on existing thread id"),
1103 OPT_INTEGER('r', "realtime", &record.realtime_prio,
1104 "collect data with this RT SCHED_FIFO priority"),
1105 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1106 "collect data without buffering"),
1107 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1108 "collect raw sample records from all opened counters"),
1109 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1110 "system-wide collection from all CPUs"),
1111 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1112 "list of cpus to monitor"),
1113 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1114 OPT_STRING('o', "output", &record.file.path, "file",
1115 "output file name"),
1116 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1117 &record.opts.no_inherit_set,
1118 "child tasks do not inherit counters"),
1119 OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1120 OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1121 "number of mmap data pages and AUX area tracing mmap pages",
1122 record__parse_mmap_pages),
1123 OPT_BOOLEAN(0, "group", &record.opts.group,
1124 "put the counters into a counter group"),
1125 OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
1126 NULL, "enables call-graph recording" ,
1127 &record_callchain_opt),
1128 OPT_CALLBACK(0, "call-graph", &record.opts,
1129 "record_mode[,record_size]", record_callchain_help,
1130 &record_parse_callchain_opt),
1131 OPT_INCR('v', "verbose", &verbose,
1132 "be more verbose (show counter open errors, etc)"),
1133 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1134 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1135 "per thread counts"),
1136 OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1137 OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1138 &record.opts.sample_time_set,
1139 "Record the sample timestamps"),
1140 OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1141 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1142 "don't sample"),
1143 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1144 &record.no_buildid_cache_set,
1145 "do not update the buildid cache"),
1146 OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1147 &record.no_buildid_set,
1148 "do not collect buildids in perf.data"),
1149 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1150 "monitor event in cgroup name only",
1151 parse_cgroups),
1152 OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1153 "ms to wait before starting measurement after program start"),
1154 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1155 "user to profile"),
1156
1157 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1158 "branch any", "sample any taken branches",
1159 parse_branch_stack),
1160
1161 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1162 "branch filter mask", "branch stack filter modes",
1163 parse_branch_stack),
1164 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1165 "sample by weight (on special events only)"),
1166 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1167 "sample transaction flags (special events only)"),
1168 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1169 "use per-thread mmaps"),
1170 OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1171 "sample selected machine registers on interrupt,"
1172 " use -I ? to list register names", parse_regs),
1173 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1174 "Record running/enabled time of read (:S) events"),
1175 OPT_CALLBACK('k', "clockid", &record.opts,
1176 "clockid", "clockid to use for events, see clock_gettime()",
1177 parse_clockid),
1178 OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1179 "opts", "AUX area tracing Snapshot Mode", ""),
1180 OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1181 "per thread proc mmap processing timeout in ms"),
1182 OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1183 "Record context switch events"),
1184 OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1185 "Configure all used events to run in kernel space.",
1186 PARSE_OPT_EXCLUSIVE),
1187 OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1188 "Configure all used events to run in user space.",
1189 PARSE_OPT_EXCLUSIVE),
1190 OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1191 "clang binary to use for compiling BPF scriptlets"),
1192 OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1193 "options passed to clang when compiling BPF scriptlets"),
1194 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1195 "file", "vmlinux pathname"),
1196 OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1197 "Record build-id of all DSOs regardless of hits"),
1198 OPT_END()
1199};
1200
1201struct option *record_options = __record_options;
1202
1203int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1204{
1205 int err;
1206 struct record *rec = &record;
1207 char errbuf[BUFSIZ];
1208
1209#ifndef HAVE_LIBBPF_SUPPORT
1210# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1211 set_nobuild('\0', "clang-path", true);
1212 set_nobuild('\0', "clang-opt", true);
1213# undef set_nobuild
1214#endif
1215
1216#ifndef HAVE_BPF_PROLOGUE
1217# if !defined (HAVE_DWARF_SUPPORT)
1218# define REASON "NO_DWARF=1"
1219# elif !defined (HAVE_LIBBPF_SUPPORT)
1220# define REASON "NO_LIBBPF=1"
1221# else
1222# define REASON "this architecture doesn't support BPF prologue"
1223# endif
1224# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1225 set_nobuild('\0', "vmlinux", true);
1226# undef set_nobuild
1227# undef REASON
1228#endif
1229
1230 rec->evlist = perf_evlist__new();
1231 if (rec->evlist == NULL)
1232 return -ENOMEM;
1233
1234 perf_config(perf_record_config, rec);
1235
1236 argc = parse_options(argc, argv, record_options, record_usage,
1237 PARSE_OPT_STOP_AT_NON_OPTION);
1238 if (!argc && target__none(&rec->opts.target))
1239 usage_with_options(record_usage, record_options);
1240
1241 if (nr_cgroups && !rec->opts.target.system_wide) {
1242 usage_with_options_msg(record_usage, record_options,
1243 "cgroup monitoring only available in system-wide mode");
1244
1245 }
1246 if (rec->opts.record_switch_events &&
1247 !perf_can_record_switch_events()) {
1248 ui__error("kernel does not support recording context switch events\n");
1249 parse_options_usage(record_usage, record_options, "switch-events", 0);
1250 return -EINVAL;
1251 }
1252
1253 if (!rec->itr) {
1254 rec->itr = auxtrace_record__init(rec->evlist, &err);
1255 if (err)
1256 return err;
1257 }
1258
1259 err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1260 rec->opts.auxtrace_snapshot_opts);
1261 if (err)
1262 return err;
1263
1264 err = -ENOMEM;
1265
1266 symbol__init(NULL);
1267
1268 if (symbol_conf.kptr_restrict)
1269 pr_warning(
1270"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1271"check /proc/sys/kernel/kptr_restrict.\n\n"
1272"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1273"file is not found in the buildid cache or in the vmlinux path.\n\n"
1274"Samples in kernel modules won't be resolved at all.\n\n"
1275"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1276"even with a suitable vmlinux or kallsyms file.\n\n");
1277
1278 if (rec->no_buildid_cache || rec->no_buildid)
1279 disable_buildid_cache();
1280
1281 if (rec->evlist->nr_entries == 0 &&
1282 perf_evlist__add_default(rec->evlist) < 0) {
1283 pr_err("Not enough memory for event selector list\n");
1284 goto out_symbol_exit;
1285 }
1286
1287 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1288 rec->opts.no_inherit = true;
1289
1290 err = target__validate(&rec->opts.target);
1291 if (err) {
1292 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1293 ui__warning("%s", errbuf);
1294 }
1295
1296 err = target__parse_uid(&rec->opts.target);
1297 if (err) {
1298 int saved_errno = errno;
1299
1300 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1301 ui__error("%s", errbuf);
1302
1303 err = -saved_errno;
1304 goto out_symbol_exit;
1305 }
1306
1307 err = -ENOMEM;
1308 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1309 usage_with_options(record_usage, record_options);
1310
1311 err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1312 if (err)
1313 goto out_symbol_exit;
1314
1315 /*
1316 * We take all buildids when the file contains
1317 * AUX area tracing data because we do not decode the
1318 * trace because it would take too long.
1319 */
1320 if (rec->opts.full_auxtrace)
1321 rec->buildid_all = true;
1322
1323 if (record_opts__config(&rec->opts)) {
1324 err = -EINVAL;
1325 goto out_symbol_exit;
1326 }
1327
1328 err = __cmd_record(&record, argc, argv);
1329out_symbol_exit:
1330 perf_evlist__delete(rec->evlist);
1331 symbol__exit();
1332 auxtrace_record__free(rec->itr);
1333 return err;
1334}
1335
1336static void snapshot_sig_handler(int sig __maybe_unused)
1337{
1338 if (!auxtrace_snapshot_enabled)
1339 return;
1340 auxtrace_snapshot_enabled = 0;
1341 auxtrace_snapshot_err = auxtrace_record__snapshot_start(record.itr);
1342 auxtrace_record__snapshot_started = 1;
1343}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * builtin-record.c
4 *
5 * Builtin record command: Record the profile of a workload
6 * (or a CPU, or a PID) into the perf.data output file - for
7 * later analysis via perf report.
8 */
9#include "builtin.h"
10
11#include "util/build-id.h"
12#include <subcmd/parse-options.h>
13#include "util/parse-events.h"
14#include "util/config.h"
15
16#include "util/callchain.h"
17#include "util/cgroup.h"
18#include "util/header.h"
19#include "util/event.h"
20#include "util/evlist.h"
21#include "util/evsel.h"
22#include "util/debug.h"
23#include "util/mmap.h"
24#include "util/target.h"
25#include "util/session.h"
26#include "util/tool.h"
27#include "util/symbol.h"
28#include "util/record.h"
29#include "util/cpumap.h"
30#include "util/thread_map.h"
31#include "util/data.h"
32#include "util/perf_regs.h"
33#include "util/auxtrace.h"
34#include "util/tsc.h"
35#include "util/parse-branch-options.h"
36#include "util/parse-regs-options.h"
37#include "util/perf_api_probe.h"
38#include "util/llvm-utils.h"
39#include "util/bpf-loader.h"
40#include "util/trigger.h"
41#include "util/perf-hooks.h"
42#include "util/cpu-set-sched.h"
43#include "util/synthetic-events.h"
44#include "util/time-utils.h"
45#include "util/units.h"
46#include "util/bpf-event.h"
47#include "util/util.h"
48#include "util/pfm.h"
49#include "util/clockid.h"
50#include "asm/bug.h"
51#include "perf.h"
52
53#include <errno.h>
54#include <inttypes.h>
55#include <locale.h>
56#include <poll.h>
57#include <pthread.h>
58#include <unistd.h>
59#include <sched.h>
60#include <signal.h>
61#ifdef HAVE_EVENTFD_SUPPORT
62#include <sys/eventfd.h>
63#endif
64#include <sys/mman.h>
65#include <sys/wait.h>
66#include <sys/types.h>
67#include <sys/stat.h>
68#include <fcntl.h>
69#include <linux/err.h>
70#include <linux/string.h>
71#include <linux/time64.h>
72#include <linux/zalloc.h>
73#include <linux/bitmap.h>
74#include <sys/time.h>
75
76struct switch_output {
77 bool enabled;
78 bool signal;
79 unsigned long size;
80 unsigned long time;
81 const char *str;
82 bool set;
83 char **filenames;
84 int num_files;
85 int cur_file;
86};
87
88struct record {
89 struct perf_tool tool;
90 struct record_opts opts;
91 u64 bytes_written;
92 struct perf_data data;
93 struct auxtrace_record *itr;
94 struct evlist *evlist;
95 struct perf_session *session;
96 struct evlist *sb_evlist;
97 pthread_t thread_id;
98 int realtime_prio;
99 bool switch_output_event_set;
100 bool no_buildid;
101 bool no_buildid_set;
102 bool no_buildid_cache;
103 bool no_buildid_cache_set;
104 bool buildid_all;
105 bool timestamp_filename;
106 bool timestamp_boundary;
107 struct switch_output switch_output;
108 unsigned long long samples;
109 struct mmap_cpu_mask affinity_mask;
110 unsigned long output_max_size; /* = 0: unlimited */
111};
112
113static volatile int done;
114
115static volatile int auxtrace_record__snapshot_started;
116static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
117static DEFINE_TRIGGER(switch_output_trigger);
118
119static const char *affinity_tags[PERF_AFFINITY_MAX] = {
120 "SYS", "NODE", "CPU"
121};
122
123static bool switch_output_signal(struct record *rec)
124{
125 return rec->switch_output.signal &&
126 trigger_is_ready(&switch_output_trigger);
127}
128
129static bool switch_output_size(struct record *rec)
130{
131 return rec->switch_output.size &&
132 trigger_is_ready(&switch_output_trigger) &&
133 (rec->bytes_written >= rec->switch_output.size);
134}
135
136static bool switch_output_time(struct record *rec)
137{
138 return rec->switch_output.time &&
139 trigger_is_ready(&switch_output_trigger);
140}
141
142static bool record__output_max_size_exceeded(struct record *rec)
143{
144 return rec->output_max_size &&
145 (rec->bytes_written >= rec->output_max_size);
146}
147
148static int record__write(struct record *rec, struct mmap *map __maybe_unused,
149 void *bf, size_t size)
150{
151 struct perf_data_file *file = &rec->session->data->file;
152
153 if (perf_data_file__write(file, bf, size) < 0) {
154 pr_err("failed to write perf data, error: %m\n");
155 return -1;
156 }
157
158 rec->bytes_written += size;
159
160 if (record__output_max_size_exceeded(rec) && !done) {
161 fprintf(stderr, "[ perf record: perf size limit reached (%" PRIu64 " KB),"
162 " stopping session ]\n",
163 rec->bytes_written >> 10);
164 done = 1;
165 }
166
167 if (switch_output_size(rec))
168 trigger_hit(&switch_output_trigger);
169
170 return 0;
171}
172
173static int record__aio_enabled(struct record *rec);
174static int record__comp_enabled(struct record *rec);
175static size_t zstd_compress(struct perf_session *session, void *dst, size_t dst_size,
176 void *src, size_t src_size);
177
178#ifdef HAVE_AIO_SUPPORT
179static int record__aio_write(struct aiocb *cblock, int trace_fd,
180 void *buf, size_t size, off_t off)
181{
182 int rc;
183
184 cblock->aio_fildes = trace_fd;
185 cblock->aio_buf = buf;
186 cblock->aio_nbytes = size;
187 cblock->aio_offset = off;
188 cblock->aio_sigevent.sigev_notify = SIGEV_NONE;
189
190 do {
191 rc = aio_write(cblock);
192 if (rc == 0) {
193 break;
194 } else if (errno != EAGAIN) {
195 cblock->aio_fildes = -1;
196 pr_err("failed to queue perf data, error: %m\n");
197 break;
198 }
199 } while (1);
200
201 return rc;
202}
203
204static int record__aio_complete(struct mmap *md, struct aiocb *cblock)
205{
206 void *rem_buf;
207 off_t rem_off;
208 size_t rem_size;
209 int rc, aio_errno;
210 ssize_t aio_ret, written;
211
212 aio_errno = aio_error(cblock);
213 if (aio_errno == EINPROGRESS)
214 return 0;
215
216 written = aio_ret = aio_return(cblock);
217 if (aio_ret < 0) {
218 if (aio_errno != EINTR)
219 pr_err("failed to write perf data, error: %m\n");
220 written = 0;
221 }
222
223 rem_size = cblock->aio_nbytes - written;
224
225 if (rem_size == 0) {
226 cblock->aio_fildes = -1;
227 /*
228 * md->refcount is incremented in record__aio_pushfn() for
229 * every aio write request started in record__aio_push() so
230 * decrement it because the request is now complete.
231 */
232 perf_mmap__put(&md->core);
233 rc = 1;
234 } else {
235 /*
236 * aio write request may require restart with the
237 * reminder if the kernel didn't write whole
238 * chunk at once.
239 */
240 rem_off = cblock->aio_offset + written;
241 rem_buf = (void *)(cblock->aio_buf + written);
242 record__aio_write(cblock, cblock->aio_fildes,
243 rem_buf, rem_size, rem_off);
244 rc = 0;
245 }
246
247 return rc;
248}
249
250static int record__aio_sync(struct mmap *md, bool sync_all)
251{
252 struct aiocb **aiocb = md->aio.aiocb;
253 struct aiocb *cblocks = md->aio.cblocks;
254 struct timespec timeout = { 0, 1000 * 1000 * 1 }; /* 1ms */
255 int i, do_suspend;
256
257 do {
258 do_suspend = 0;
259 for (i = 0; i < md->aio.nr_cblocks; ++i) {
260 if (cblocks[i].aio_fildes == -1 || record__aio_complete(md, &cblocks[i])) {
261 if (sync_all)
262 aiocb[i] = NULL;
263 else
264 return i;
265 } else {
266 /*
267 * Started aio write is not complete yet
268 * so it has to be waited before the
269 * next allocation.
270 */
271 aiocb[i] = &cblocks[i];
272 do_suspend = 1;
273 }
274 }
275 if (!do_suspend)
276 return -1;
277
278 while (aio_suspend((const struct aiocb **)aiocb, md->aio.nr_cblocks, &timeout)) {
279 if (!(errno == EAGAIN || errno == EINTR))
280 pr_err("failed to sync perf data, error: %m\n");
281 }
282 } while (1);
283}
284
285struct record_aio {
286 struct record *rec;
287 void *data;
288 size_t size;
289};
290
291static int record__aio_pushfn(struct mmap *map, void *to, void *buf, size_t size)
292{
293 struct record_aio *aio = to;
294
295 /*
296 * map->core.base data pointed by buf is copied into free map->aio.data[] buffer
297 * to release space in the kernel buffer as fast as possible, calling
298 * perf_mmap__consume() from perf_mmap__push() function.
299 *
300 * That lets the kernel to proceed with storing more profiling data into
301 * the kernel buffer earlier than other per-cpu kernel buffers are handled.
302 *
303 * Coping can be done in two steps in case the chunk of profiling data
304 * crosses the upper bound of the kernel buffer. In this case we first move
305 * part of data from map->start till the upper bound and then the reminder
306 * from the beginning of the kernel buffer till the end of the data chunk.
307 */
308
309 if (record__comp_enabled(aio->rec)) {
310 size = zstd_compress(aio->rec->session, aio->data + aio->size,
311 mmap__mmap_len(map) - aio->size,
312 buf, size);
313 } else {
314 memcpy(aio->data + aio->size, buf, size);
315 }
316
317 if (!aio->size) {
318 /*
319 * Increment map->refcount to guard map->aio.data[] buffer
320 * from premature deallocation because map object can be
321 * released earlier than aio write request started on
322 * map->aio.data[] buffer is complete.
323 *
324 * perf_mmap__put() is done at record__aio_complete()
325 * after started aio request completion or at record__aio_push()
326 * if the request failed to start.
327 */
328 perf_mmap__get(&map->core);
329 }
330
331 aio->size += size;
332
333 return size;
334}
335
336static int record__aio_push(struct record *rec, struct mmap *map, off_t *off)
337{
338 int ret, idx;
339 int trace_fd = rec->session->data->file.fd;
340 struct record_aio aio = { .rec = rec, .size = 0 };
341
342 /*
343 * Call record__aio_sync() to wait till map->aio.data[] buffer
344 * becomes available after previous aio write operation.
345 */
346
347 idx = record__aio_sync(map, false);
348 aio.data = map->aio.data[idx];
349 ret = perf_mmap__push(map, &aio, record__aio_pushfn);
350 if (ret != 0) /* ret > 0 - no data, ret < 0 - error */
351 return ret;
352
353 rec->samples++;
354 ret = record__aio_write(&(map->aio.cblocks[idx]), trace_fd, aio.data, aio.size, *off);
355 if (!ret) {
356 *off += aio.size;
357 rec->bytes_written += aio.size;
358 if (switch_output_size(rec))
359 trigger_hit(&switch_output_trigger);
360 } else {
361 /*
362 * Decrement map->refcount incremented in record__aio_pushfn()
363 * back if record__aio_write() operation failed to start, otherwise
364 * map->refcount is decremented in record__aio_complete() after
365 * aio write operation finishes successfully.
366 */
367 perf_mmap__put(&map->core);
368 }
369
370 return ret;
371}
372
373static off_t record__aio_get_pos(int trace_fd)
374{
375 return lseek(trace_fd, 0, SEEK_CUR);
376}
377
378static void record__aio_set_pos(int trace_fd, off_t pos)
379{
380 lseek(trace_fd, pos, SEEK_SET);
381}
382
383static void record__aio_mmap_read_sync(struct record *rec)
384{
385 int i;
386 struct evlist *evlist = rec->evlist;
387 struct mmap *maps = evlist->mmap;
388
389 if (!record__aio_enabled(rec))
390 return;
391
392 for (i = 0; i < evlist->core.nr_mmaps; i++) {
393 struct mmap *map = &maps[i];
394
395 if (map->core.base)
396 record__aio_sync(map, true);
397 }
398}
399
400static int nr_cblocks_default = 1;
401static int nr_cblocks_max = 4;
402
403static int record__aio_parse(const struct option *opt,
404 const char *str,
405 int unset)
406{
407 struct record_opts *opts = (struct record_opts *)opt->value;
408
409 if (unset) {
410 opts->nr_cblocks = 0;
411 } else {
412 if (str)
413 opts->nr_cblocks = strtol(str, NULL, 0);
414 if (!opts->nr_cblocks)
415 opts->nr_cblocks = nr_cblocks_default;
416 }
417
418 return 0;
419}
420#else /* HAVE_AIO_SUPPORT */
421static int nr_cblocks_max = 0;
422
423static int record__aio_push(struct record *rec __maybe_unused, struct mmap *map __maybe_unused,
424 off_t *off __maybe_unused)
425{
426 return -1;
427}
428
429static off_t record__aio_get_pos(int trace_fd __maybe_unused)
430{
431 return -1;
432}
433
434static void record__aio_set_pos(int trace_fd __maybe_unused, off_t pos __maybe_unused)
435{
436}
437
438static void record__aio_mmap_read_sync(struct record *rec __maybe_unused)
439{
440}
441#endif
442
443static int record__aio_enabled(struct record *rec)
444{
445 return rec->opts.nr_cblocks > 0;
446}
447
448#define MMAP_FLUSH_DEFAULT 1
449static int record__mmap_flush_parse(const struct option *opt,
450 const char *str,
451 int unset)
452{
453 int flush_max;
454 struct record_opts *opts = (struct record_opts *)opt->value;
455 static struct parse_tag tags[] = {
456 { .tag = 'B', .mult = 1 },
457 { .tag = 'K', .mult = 1 << 10 },
458 { .tag = 'M', .mult = 1 << 20 },
459 { .tag = 'G', .mult = 1 << 30 },
460 { .tag = 0 },
461 };
462
463 if (unset)
464 return 0;
465
466 if (str) {
467 opts->mmap_flush = parse_tag_value(str, tags);
468 if (opts->mmap_flush == (int)-1)
469 opts->mmap_flush = strtol(str, NULL, 0);
470 }
471
472 if (!opts->mmap_flush)
473 opts->mmap_flush = MMAP_FLUSH_DEFAULT;
474
475 flush_max = evlist__mmap_size(opts->mmap_pages);
476 flush_max /= 4;
477 if (opts->mmap_flush > flush_max)
478 opts->mmap_flush = flush_max;
479
480 return 0;
481}
482
483#ifdef HAVE_ZSTD_SUPPORT
484static unsigned int comp_level_default = 1;
485
486static int record__parse_comp_level(const struct option *opt, const char *str, int unset)
487{
488 struct record_opts *opts = opt->value;
489
490 if (unset) {
491 opts->comp_level = 0;
492 } else {
493 if (str)
494 opts->comp_level = strtol(str, NULL, 0);
495 if (!opts->comp_level)
496 opts->comp_level = comp_level_default;
497 }
498
499 return 0;
500}
501#endif
502static unsigned int comp_level_max = 22;
503
504static int record__comp_enabled(struct record *rec)
505{
506 return rec->opts.comp_level > 0;
507}
508
509static int process_synthesized_event(struct perf_tool *tool,
510 union perf_event *event,
511 struct perf_sample *sample __maybe_unused,
512 struct machine *machine __maybe_unused)
513{
514 struct record *rec = container_of(tool, struct record, tool);
515 return record__write(rec, NULL, event, event->header.size);
516}
517
518static int process_locked_synthesized_event(struct perf_tool *tool,
519 union perf_event *event,
520 struct perf_sample *sample __maybe_unused,
521 struct machine *machine __maybe_unused)
522{
523 static pthread_mutex_t synth_lock = PTHREAD_MUTEX_INITIALIZER;
524 int ret;
525
526 pthread_mutex_lock(&synth_lock);
527 ret = process_synthesized_event(tool, event, sample, machine);
528 pthread_mutex_unlock(&synth_lock);
529 return ret;
530}
531
532static int record__pushfn(struct mmap *map, void *to, void *bf, size_t size)
533{
534 struct record *rec = to;
535
536 if (record__comp_enabled(rec)) {
537 size = zstd_compress(rec->session, map->data, mmap__mmap_len(map), bf, size);
538 bf = map->data;
539 }
540
541 rec->samples++;
542 return record__write(rec, map, bf, size);
543}
544
545static volatile int signr = -1;
546static volatile int child_finished;
547#ifdef HAVE_EVENTFD_SUPPORT
548static int done_fd = -1;
549#endif
550
551static void sig_handler(int sig)
552{
553 if (sig == SIGCHLD)
554 child_finished = 1;
555 else
556 signr = sig;
557
558 done = 1;
559#ifdef HAVE_EVENTFD_SUPPORT
560{
561 u64 tmp = 1;
562 /*
563 * It is possible for this signal handler to run after done is checked
564 * in the main loop, but before the perf counter fds are polled. If this
565 * happens, the poll() will continue to wait even though done is set,
566 * and will only break out if either another signal is received, or the
567 * counters are ready for read. To ensure the poll() doesn't sleep when
568 * done is set, use an eventfd (done_fd) to wake up the poll().
569 */
570 if (write(done_fd, &tmp, sizeof(tmp)) < 0)
571 pr_err("failed to signal wakeup fd, error: %m\n");
572}
573#endif // HAVE_EVENTFD_SUPPORT
574}
575
576static void sigsegv_handler(int sig)
577{
578 perf_hooks__recover();
579 sighandler_dump_stack(sig);
580}
581
582static void record__sig_exit(void)
583{
584 if (signr == -1)
585 return;
586
587 signal(signr, SIG_DFL);
588 raise(signr);
589}
590
591#ifdef HAVE_AUXTRACE_SUPPORT
592
593static int record__process_auxtrace(struct perf_tool *tool,
594 struct mmap *map,
595 union perf_event *event, void *data1,
596 size_t len1, void *data2, size_t len2)
597{
598 struct record *rec = container_of(tool, struct record, tool);
599 struct perf_data *data = &rec->data;
600 size_t padding;
601 u8 pad[8] = {0};
602
603 if (!perf_data__is_pipe(data) && perf_data__is_single_file(data)) {
604 off_t file_offset;
605 int fd = perf_data__fd(data);
606 int err;
607
608 file_offset = lseek(fd, 0, SEEK_CUR);
609 if (file_offset == -1)
610 return -1;
611 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
612 event, file_offset);
613 if (err)
614 return err;
615 }
616
617 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
618 padding = (len1 + len2) & 7;
619 if (padding)
620 padding = 8 - padding;
621
622 record__write(rec, map, event, event->header.size);
623 record__write(rec, map, data1, len1);
624 if (len2)
625 record__write(rec, map, data2, len2);
626 record__write(rec, map, &pad, padding);
627
628 return 0;
629}
630
631static int record__auxtrace_mmap_read(struct record *rec,
632 struct mmap *map)
633{
634 int ret;
635
636 ret = auxtrace_mmap__read(map, rec->itr, &rec->tool,
637 record__process_auxtrace);
638 if (ret < 0)
639 return ret;
640
641 if (ret)
642 rec->samples++;
643
644 return 0;
645}
646
647static int record__auxtrace_mmap_read_snapshot(struct record *rec,
648 struct mmap *map)
649{
650 int ret;
651
652 ret = auxtrace_mmap__read_snapshot(map, rec->itr, &rec->tool,
653 record__process_auxtrace,
654 rec->opts.auxtrace_snapshot_size);
655 if (ret < 0)
656 return ret;
657
658 if (ret)
659 rec->samples++;
660
661 return 0;
662}
663
664static int record__auxtrace_read_snapshot_all(struct record *rec)
665{
666 int i;
667 int rc = 0;
668
669 for (i = 0; i < rec->evlist->core.nr_mmaps; i++) {
670 struct mmap *map = &rec->evlist->mmap[i];
671
672 if (!map->auxtrace_mmap.base)
673 continue;
674
675 if (record__auxtrace_mmap_read_snapshot(rec, map) != 0) {
676 rc = -1;
677 goto out;
678 }
679 }
680out:
681 return rc;
682}
683
684static void record__read_auxtrace_snapshot(struct record *rec, bool on_exit)
685{
686 pr_debug("Recording AUX area tracing snapshot\n");
687 if (record__auxtrace_read_snapshot_all(rec) < 0) {
688 trigger_error(&auxtrace_snapshot_trigger);
689 } else {
690 if (auxtrace_record__snapshot_finish(rec->itr, on_exit))
691 trigger_error(&auxtrace_snapshot_trigger);
692 else
693 trigger_ready(&auxtrace_snapshot_trigger);
694 }
695}
696
697static int record__auxtrace_snapshot_exit(struct record *rec)
698{
699 if (trigger_is_error(&auxtrace_snapshot_trigger))
700 return 0;
701
702 if (!auxtrace_record__snapshot_started &&
703 auxtrace_record__snapshot_start(rec->itr))
704 return -1;
705
706 record__read_auxtrace_snapshot(rec, true);
707 if (trigger_is_error(&auxtrace_snapshot_trigger))
708 return -1;
709
710 return 0;
711}
712
713static int record__auxtrace_init(struct record *rec)
714{
715 int err;
716
717 if (!rec->itr) {
718 rec->itr = auxtrace_record__init(rec->evlist, &err);
719 if (err)
720 return err;
721 }
722
723 err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
724 rec->opts.auxtrace_snapshot_opts);
725 if (err)
726 return err;
727
728 err = auxtrace_parse_sample_options(rec->itr, rec->evlist, &rec->opts,
729 rec->opts.auxtrace_sample_opts);
730 if (err)
731 return err;
732
733 return auxtrace_parse_filters(rec->evlist);
734}
735
736#else
737
738static inline
739int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
740 struct mmap *map __maybe_unused)
741{
742 return 0;
743}
744
745static inline
746void record__read_auxtrace_snapshot(struct record *rec __maybe_unused,
747 bool on_exit __maybe_unused)
748{
749}
750
751static inline
752int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
753{
754 return 0;
755}
756
757static inline
758int record__auxtrace_snapshot_exit(struct record *rec __maybe_unused)
759{
760 return 0;
761}
762
763static int record__auxtrace_init(struct record *rec __maybe_unused)
764{
765 return 0;
766}
767
768#endif
769
770static int record__config_text_poke(struct evlist *evlist)
771{
772 struct evsel *evsel;
773 int err;
774
775 /* Nothing to do if text poke is already configured */
776 evlist__for_each_entry(evlist, evsel) {
777 if (evsel->core.attr.text_poke)
778 return 0;
779 }
780
781 err = parse_events(evlist, "dummy:u", NULL);
782 if (err)
783 return err;
784
785 evsel = evlist__last(evlist);
786
787 evsel->core.attr.freq = 0;
788 evsel->core.attr.sample_period = 1;
789 evsel->core.attr.text_poke = 1;
790 evsel->core.attr.ksymbol = 1;
791
792 evsel->core.system_wide = true;
793 evsel->no_aux_samples = true;
794 evsel->immediate = true;
795
796 /* Text poke must be collected on all CPUs */
797 perf_cpu_map__put(evsel->core.own_cpus);
798 evsel->core.own_cpus = perf_cpu_map__new(NULL);
799 perf_cpu_map__put(evsel->core.cpus);
800 evsel->core.cpus = perf_cpu_map__get(evsel->core.own_cpus);
801
802 evsel__set_sample_bit(evsel, TIME);
803
804 return 0;
805}
806
807static bool record__kcore_readable(struct machine *machine)
808{
809 char kcore[PATH_MAX];
810 int fd;
811
812 scnprintf(kcore, sizeof(kcore), "%s/proc/kcore", machine->root_dir);
813
814 fd = open(kcore, O_RDONLY);
815 if (fd < 0)
816 return false;
817
818 close(fd);
819
820 return true;
821}
822
823static int record__kcore_copy(struct machine *machine, struct perf_data *data)
824{
825 char from_dir[PATH_MAX];
826 char kcore_dir[PATH_MAX];
827 int ret;
828
829 snprintf(from_dir, sizeof(from_dir), "%s/proc", machine->root_dir);
830
831 ret = perf_data__make_kcore_dir(data, kcore_dir, sizeof(kcore_dir));
832 if (ret)
833 return ret;
834
835 return kcore_copy(from_dir, kcore_dir);
836}
837
838static int record__mmap_evlist(struct record *rec,
839 struct evlist *evlist)
840{
841 struct record_opts *opts = &rec->opts;
842 bool auxtrace_overwrite = opts->auxtrace_snapshot_mode ||
843 opts->auxtrace_sample_mode;
844 char msg[512];
845
846 if (opts->affinity != PERF_AFFINITY_SYS)
847 cpu__setup_cpunode_map();
848
849 if (evlist__mmap_ex(evlist, opts->mmap_pages,
850 opts->auxtrace_mmap_pages,
851 auxtrace_overwrite,
852 opts->nr_cblocks, opts->affinity,
853 opts->mmap_flush, opts->comp_level) < 0) {
854 if (errno == EPERM) {
855 pr_err("Permission error mapping pages.\n"
856 "Consider increasing "
857 "/proc/sys/kernel/perf_event_mlock_kb,\n"
858 "or try again with a smaller value of -m/--mmap_pages.\n"
859 "(current value: %u,%u)\n",
860 opts->mmap_pages, opts->auxtrace_mmap_pages);
861 return -errno;
862 } else {
863 pr_err("failed to mmap with %d (%s)\n", errno,
864 str_error_r(errno, msg, sizeof(msg)));
865 if (errno)
866 return -errno;
867 else
868 return -EINVAL;
869 }
870 }
871 return 0;
872}
873
874static int record__mmap(struct record *rec)
875{
876 return record__mmap_evlist(rec, rec->evlist);
877}
878
879static int record__open(struct record *rec)
880{
881 char msg[BUFSIZ];
882 struct evsel *pos;
883 struct evlist *evlist = rec->evlist;
884 struct perf_session *session = rec->session;
885 struct record_opts *opts = &rec->opts;
886 int rc = 0;
887
888 /*
889 * For initial_delay or system wide, we need to add a dummy event so
890 * that we can track PERF_RECORD_MMAP to cover the delay of waiting or
891 * event synthesis.
892 */
893 if (opts->initial_delay || target__has_cpu(&opts->target)) {
894 pos = perf_evlist__get_tracking_event(evlist);
895 if (!evsel__is_dummy_event(pos)) {
896 /* Set up dummy event. */
897 if (evlist__add_dummy(evlist))
898 return -ENOMEM;
899 pos = evlist__last(evlist);
900 perf_evlist__set_tracking_event(evlist, pos);
901 }
902
903 /*
904 * Enable the dummy event when the process is forked for
905 * initial_delay, immediately for system wide.
906 */
907 if (opts->initial_delay && !pos->immediate)
908 pos->core.attr.enable_on_exec = 1;
909 else
910 pos->immediate = 1;
911 }
912
913 perf_evlist__config(evlist, opts, &callchain_param);
914
915 evlist__for_each_entry(evlist, pos) {
916try_again:
917 if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
918 if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
919 if (verbose > 0)
920 ui__warning("%s\n", msg);
921 goto try_again;
922 }
923 if ((errno == EINVAL || errno == EBADF) &&
924 pos->leader != pos &&
925 pos->weak_group) {
926 pos = perf_evlist__reset_weak_group(evlist, pos, true);
927 goto try_again;
928 }
929 rc = -errno;
930 evsel__open_strerror(pos, &opts->target, errno, msg, sizeof(msg));
931 ui__error("%s\n", msg);
932 goto out;
933 }
934
935 pos->supported = true;
936 }
937
938 if (symbol_conf.kptr_restrict && !perf_evlist__exclude_kernel(evlist)) {
939 pr_warning(
940"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
941"check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
942"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
943"file is not found in the buildid cache or in the vmlinux path.\n\n"
944"Samples in kernel modules won't be resolved at all.\n\n"
945"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
946"even with a suitable vmlinux or kallsyms file.\n\n");
947 }
948
949 if (perf_evlist__apply_filters(evlist, &pos)) {
950 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
951 pos->filter, evsel__name(pos), errno,
952 str_error_r(errno, msg, sizeof(msg)));
953 rc = -1;
954 goto out;
955 }
956
957 rc = record__mmap(rec);
958 if (rc)
959 goto out;
960
961 session->evlist = evlist;
962 perf_session__set_id_hdr_size(session);
963out:
964 return rc;
965}
966
967static int process_sample_event(struct perf_tool *tool,
968 union perf_event *event,
969 struct perf_sample *sample,
970 struct evsel *evsel,
971 struct machine *machine)
972{
973 struct record *rec = container_of(tool, struct record, tool);
974
975 if (rec->evlist->first_sample_time == 0)
976 rec->evlist->first_sample_time = sample->time;
977
978 rec->evlist->last_sample_time = sample->time;
979
980 if (rec->buildid_all)
981 return 0;
982
983 rec->samples++;
984 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
985}
986
987static int process_buildids(struct record *rec)
988{
989 struct perf_session *session = rec->session;
990
991 if (perf_data__size(&rec->data) == 0)
992 return 0;
993
994 /*
995 * During this process, it'll load kernel map and replace the
996 * dso->long_name to a real pathname it found. In this case
997 * we prefer the vmlinux path like
998 * /lib/modules/3.16.4/build/vmlinux
999 *
1000 * rather than build-id path (in debug directory).
1001 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
1002 */
1003 symbol_conf.ignore_vmlinux_buildid = true;
1004
1005 /*
1006 * If --buildid-all is given, it marks all DSO regardless of hits,
1007 * so no need to process samples. But if timestamp_boundary is enabled,
1008 * it still needs to walk on all samples to get the timestamps of
1009 * first/last samples.
1010 */
1011 if (rec->buildid_all && !rec->timestamp_boundary)
1012 rec->tool.sample = NULL;
1013
1014 return perf_session__process_events(session);
1015}
1016
1017static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
1018{
1019 int err;
1020 struct perf_tool *tool = data;
1021 /*
1022 *As for guest kernel when processing subcommand record&report,
1023 *we arrange module mmap prior to guest kernel mmap and trigger
1024 *a preload dso because default guest module symbols are loaded
1025 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
1026 *method is used to avoid symbol missing when the first addr is
1027 *in module instead of in guest kernel.
1028 */
1029 err = perf_event__synthesize_modules(tool, process_synthesized_event,
1030 machine);
1031 if (err < 0)
1032 pr_err("Couldn't record guest kernel [%d]'s reference"
1033 " relocation symbol.\n", machine->pid);
1034
1035 /*
1036 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
1037 * have no _text sometimes.
1038 */
1039 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1040 machine);
1041 if (err < 0)
1042 pr_err("Couldn't record guest kernel [%d]'s reference"
1043 " relocation symbol.\n", machine->pid);
1044}
1045
1046static struct perf_event_header finished_round_event = {
1047 .size = sizeof(struct perf_event_header),
1048 .type = PERF_RECORD_FINISHED_ROUND,
1049};
1050
1051static void record__adjust_affinity(struct record *rec, struct mmap *map)
1052{
1053 if (rec->opts.affinity != PERF_AFFINITY_SYS &&
1054 !bitmap_equal(rec->affinity_mask.bits, map->affinity_mask.bits,
1055 rec->affinity_mask.nbits)) {
1056 bitmap_zero(rec->affinity_mask.bits, rec->affinity_mask.nbits);
1057 bitmap_or(rec->affinity_mask.bits, rec->affinity_mask.bits,
1058 map->affinity_mask.bits, rec->affinity_mask.nbits);
1059 sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&rec->affinity_mask),
1060 (cpu_set_t *)rec->affinity_mask.bits);
1061 if (verbose == 2)
1062 mmap_cpu_mask__scnprintf(&rec->affinity_mask, "thread");
1063 }
1064}
1065
1066static size_t process_comp_header(void *record, size_t increment)
1067{
1068 struct perf_record_compressed *event = record;
1069 size_t size = sizeof(*event);
1070
1071 if (increment) {
1072 event->header.size += increment;
1073 return increment;
1074 }
1075
1076 event->header.type = PERF_RECORD_COMPRESSED;
1077 event->header.size = size;
1078
1079 return size;
1080}
1081
1082static size_t zstd_compress(struct perf_session *session, void *dst, size_t dst_size,
1083 void *src, size_t src_size)
1084{
1085 size_t compressed;
1086 size_t max_record_size = PERF_SAMPLE_MAX_SIZE - sizeof(struct perf_record_compressed) - 1;
1087
1088 compressed = zstd_compress_stream_to_records(&session->zstd_data, dst, dst_size, src, src_size,
1089 max_record_size, process_comp_header);
1090
1091 session->bytes_transferred += src_size;
1092 session->bytes_compressed += compressed;
1093
1094 return compressed;
1095}
1096
1097static int record__mmap_read_evlist(struct record *rec, struct evlist *evlist,
1098 bool overwrite, bool synch)
1099{
1100 u64 bytes_written = rec->bytes_written;
1101 int i;
1102 int rc = 0;
1103 struct mmap *maps;
1104 int trace_fd = rec->data.file.fd;
1105 off_t off = 0;
1106
1107 if (!evlist)
1108 return 0;
1109
1110 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
1111 if (!maps)
1112 return 0;
1113
1114 if (overwrite && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
1115 return 0;
1116
1117 if (record__aio_enabled(rec))
1118 off = record__aio_get_pos(trace_fd);
1119
1120 for (i = 0; i < evlist->core.nr_mmaps; i++) {
1121 u64 flush = 0;
1122 struct mmap *map = &maps[i];
1123
1124 if (map->core.base) {
1125 record__adjust_affinity(rec, map);
1126 if (synch) {
1127 flush = map->core.flush;
1128 map->core.flush = 1;
1129 }
1130 if (!record__aio_enabled(rec)) {
1131 if (perf_mmap__push(map, rec, record__pushfn) < 0) {
1132 if (synch)
1133 map->core.flush = flush;
1134 rc = -1;
1135 goto out;
1136 }
1137 } else {
1138 if (record__aio_push(rec, map, &off) < 0) {
1139 record__aio_set_pos(trace_fd, off);
1140 if (synch)
1141 map->core.flush = flush;
1142 rc = -1;
1143 goto out;
1144 }
1145 }
1146 if (synch)
1147 map->core.flush = flush;
1148 }
1149
1150 if (map->auxtrace_mmap.base && !rec->opts.auxtrace_snapshot_mode &&
1151 !rec->opts.auxtrace_sample_mode &&
1152 record__auxtrace_mmap_read(rec, map) != 0) {
1153 rc = -1;
1154 goto out;
1155 }
1156 }
1157
1158 if (record__aio_enabled(rec))
1159 record__aio_set_pos(trace_fd, off);
1160
1161 /*
1162 * Mark the round finished in case we wrote
1163 * at least one event.
1164 */
1165 if (bytes_written != rec->bytes_written)
1166 rc = record__write(rec, NULL, &finished_round_event, sizeof(finished_round_event));
1167
1168 if (overwrite)
1169 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
1170out:
1171 return rc;
1172}
1173
1174static int record__mmap_read_all(struct record *rec, bool synch)
1175{
1176 int err;
1177
1178 err = record__mmap_read_evlist(rec, rec->evlist, false, synch);
1179 if (err)
1180 return err;
1181
1182 return record__mmap_read_evlist(rec, rec->evlist, true, synch);
1183}
1184
1185static void record__init_features(struct record *rec)
1186{
1187 struct perf_session *session = rec->session;
1188 int feat;
1189
1190 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1191 perf_header__set_feat(&session->header, feat);
1192
1193 if (rec->no_buildid)
1194 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1195
1196 if (!have_tracepoints(&rec->evlist->core.entries))
1197 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1198
1199 if (!rec->opts.branch_stack)
1200 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1201
1202 if (!rec->opts.full_auxtrace)
1203 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1204
1205 if (!(rec->opts.use_clockid && rec->opts.clockid_res_ns))
1206 perf_header__clear_feat(&session->header, HEADER_CLOCKID);
1207
1208 if (!rec->opts.use_clockid)
1209 perf_header__clear_feat(&session->header, HEADER_CLOCK_DATA);
1210
1211 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1212 if (!record__comp_enabled(rec))
1213 perf_header__clear_feat(&session->header, HEADER_COMPRESSED);
1214
1215 perf_header__clear_feat(&session->header, HEADER_STAT);
1216}
1217
1218static void
1219record__finish_output(struct record *rec)
1220{
1221 struct perf_data *data = &rec->data;
1222 int fd = perf_data__fd(data);
1223
1224 if (data->is_pipe)
1225 return;
1226
1227 rec->session->header.data_size += rec->bytes_written;
1228 data->file.size = lseek(perf_data__fd(data), 0, SEEK_CUR);
1229
1230 if (!rec->no_buildid) {
1231 process_buildids(rec);
1232
1233 if (rec->buildid_all)
1234 dsos__hit_all(rec->session);
1235 }
1236 perf_session__write_header(rec->session, rec->evlist, fd, true);
1237
1238 return;
1239}
1240
1241static int record__synthesize_workload(struct record *rec, bool tail)
1242{
1243 int err;
1244 struct perf_thread_map *thread_map;
1245
1246 if (rec->opts.tail_synthesize != tail)
1247 return 0;
1248
1249 thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
1250 if (thread_map == NULL)
1251 return -1;
1252
1253 err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
1254 process_synthesized_event,
1255 &rec->session->machines.host,
1256 rec->opts.sample_address);
1257 perf_thread_map__put(thread_map);
1258 return err;
1259}
1260
1261static int record__synthesize(struct record *rec, bool tail);
1262
1263static int
1264record__switch_output(struct record *rec, bool at_exit)
1265{
1266 struct perf_data *data = &rec->data;
1267 int fd, err;
1268 char *new_filename;
1269
1270 /* Same Size: "2015122520103046"*/
1271 char timestamp[] = "InvalidTimestamp";
1272
1273 record__aio_mmap_read_sync(rec);
1274
1275 record__synthesize(rec, true);
1276 if (target__none(&rec->opts.target))
1277 record__synthesize_workload(rec, true);
1278
1279 rec->samples = 0;
1280 record__finish_output(rec);
1281 err = fetch_current_timestamp(timestamp, sizeof(timestamp));
1282 if (err) {
1283 pr_err("Failed to get current timestamp\n");
1284 return -EINVAL;
1285 }
1286
1287 fd = perf_data__switch(data, timestamp,
1288 rec->session->header.data_offset,
1289 at_exit, &new_filename);
1290 if (fd >= 0 && !at_exit) {
1291 rec->bytes_written = 0;
1292 rec->session->header.data_size = 0;
1293 }
1294
1295 if (!quiet)
1296 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
1297 data->path, timestamp);
1298
1299 if (rec->switch_output.num_files) {
1300 int n = rec->switch_output.cur_file + 1;
1301
1302 if (n >= rec->switch_output.num_files)
1303 n = 0;
1304 rec->switch_output.cur_file = n;
1305 if (rec->switch_output.filenames[n]) {
1306 remove(rec->switch_output.filenames[n]);
1307 zfree(&rec->switch_output.filenames[n]);
1308 }
1309 rec->switch_output.filenames[n] = new_filename;
1310 } else {
1311 free(new_filename);
1312 }
1313
1314 /* Output tracking events */
1315 if (!at_exit) {
1316 record__synthesize(rec, false);
1317
1318 /*
1319 * In 'perf record --switch-output' without -a,
1320 * record__synthesize() in record__switch_output() won't
1321 * generate tracking events because there's no thread_map
1322 * in evlist. Which causes newly created perf.data doesn't
1323 * contain map and comm information.
1324 * Create a fake thread_map and directly call
1325 * perf_event__synthesize_thread_map() for those events.
1326 */
1327 if (target__none(&rec->opts.target))
1328 record__synthesize_workload(rec, false);
1329 }
1330 return fd;
1331}
1332
1333static volatile int workload_exec_errno;
1334
1335/*
1336 * perf_evlist__prepare_workload will send a SIGUSR1
1337 * if the fork fails, since we asked by setting its
1338 * want_signal to true.
1339 */
1340static void workload_exec_failed_signal(int signo __maybe_unused,
1341 siginfo_t *info,
1342 void *ucontext __maybe_unused)
1343{
1344 workload_exec_errno = info->si_value.sival_int;
1345 done = 1;
1346 child_finished = 1;
1347}
1348
1349static void snapshot_sig_handler(int sig);
1350static void alarm_sig_handler(int sig);
1351
1352static const struct perf_event_mmap_page *
1353perf_evlist__pick_pc(struct evlist *evlist)
1354{
1355 if (evlist) {
1356 if (evlist->mmap && evlist->mmap[0].core.base)
1357 return evlist->mmap[0].core.base;
1358 if (evlist->overwrite_mmap && evlist->overwrite_mmap[0].core.base)
1359 return evlist->overwrite_mmap[0].core.base;
1360 }
1361 return NULL;
1362}
1363
1364static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
1365{
1366 const struct perf_event_mmap_page *pc;
1367
1368 pc = perf_evlist__pick_pc(rec->evlist);
1369 if (pc)
1370 return pc;
1371 return NULL;
1372}
1373
1374static int record__synthesize(struct record *rec, bool tail)
1375{
1376 struct perf_session *session = rec->session;
1377 struct machine *machine = &session->machines.host;
1378 struct perf_data *data = &rec->data;
1379 struct record_opts *opts = &rec->opts;
1380 struct perf_tool *tool = &rec->tool;
1381 int fd = perf_data__fd(data);
1382 int err = 0;
1383 event_op f = process_synthesized_event;
1384
1385 if (rec->opts.tail_synthesize != tail)
1386 return 0;
1387
1388 if (data->is_pipe) {
1389 /*
1390 * We need to synthesize events first, because some
1391 * features works on top of them (on report side).
1392 */
1393 err = perf_event__synthesize_attrs(tool, rec->evlist,
1394 process_synthesized_event);
1395 if (err < 0) {
1396 pr_err("Couldn't synthesize attrs.\n");
1397 goto out;
1398 }
1399
1400 err = perf_event__synthesize_features(tool, session, rec->evlist,
1401 process_synthesized_event);
1402 if (err < 0) {
1403 pr_err("Couldn't synthesize features.\n");
1404 return err;
1405 }
1406
1407 if (have_tracepoints(&rec->evlist->core.entries)) {
1408 /*
1409 * FIXME err <= 0 here actually means that
1410 * there were no tracepoints so its not really
1411 * an error, just that we don't need to
1412 * synthesize anything. We really have to
1413 * return this more properly and also
1414 * propagate errors that now are calling die()
1415 */
1416 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
1417 process_synthesized_event);
1418 if (err <= 0) {
1419 pr_err("Couldn't record tracing data.\n");
1420 goto out;
1421 }
1422 rec->bytes_written += err;
1423 }
1424 }
1425
1426 err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
1427 process_synthesized_event, machine);
1428 if (err)
1429 goto out;
1430
1431 /* Synthesize id_index before auxtrace_info */
1432 if (rec->opts.auxtrace_sample_mode) {
1433 err = perf_event__synthesize_id_index(tool,
1434 process_synthesized_event,
1435 session->evlist, machine);
1436 if (err)
1437 goto out;
1438 }
1439
1440 if (rec->opts.full_auxtrace) {
1441 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
1442 session, process_synthesized_event);
1443 if (err)
1444 goto out;
1445 }
1446
1447 if (!perf_evlist__exclude_kernel(rec->evlist)) {
1448 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1449 machine);
1450 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
1451 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
1452 "Check /proc/kallsyms permission or run as root.\n");
1453
1454 err = perf_event__synthesize_modules(tool, process_synthesized_event,
1455 machine);
1456 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
1457 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
1458 "Check /proc/modules permission or run as root.\n");
1459 }
1460
1461 if (perf_guest) {
1462 machines__process_guests(&session->machines,
1463 perf_event__synthesize_guest_os, tool);
1464 }
1465
1466 err = perf_event__synthesize_extra_attr(&rec->tool,
1467 rec->evlist,
1468 process_synthesized_event,
1469 data->is_pipe);
1470 if (err)
1471 goto out;
1472
1473 err = perf_event__synthesize_thread_map2(&rec->tool, rec->evlist->core.threads,
1474 process_synthesized_event,
1475 NULL);
1476 if (err < 0) {
1477 pr_err("Couldn't synthesize thread map.\n");
1478 return err;
1479 }
1480
1481 err = perf_event__synthesize_cpu_map(&rec->tool, rec->evlist->core.cpus,
1482 process_synthesized_event, NULL);
1483 if (err < 0) {
1484 pr_err("Couldn't synthesize cpu map.\n");
1485 return err;
1486 }
1487
1488 err = perf_event__synthesize_bpf_events(session, process_synthesized_event,
1489 machine, opts);
1490 if (err < 0)
1491 pr_warning("Couldn't synthesize bpf events.\n");
1492
1493 err = perf_event__synthesize_cgroups(tool, process_synthesized_event,
1494 machine);
1495 if (err < 0)
1496 pr_warning("Couldn't synthesize cgroup events.\n");
1497
1498 if (rec->opts.nr_threads_synthesize > 1) {
1499 perf_set_multithreaded();
1500 f = process_locked_synthesized_event;
1501 }
1502
1503 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->core.threads,
1504 f, opts->sample_address,
1505 rec->opts.nr_threads_synthesize);
1506
1507 if (rec->opts.nr_threads_synthesize > 1)
1508 perf_set_singlethreaded();
1509
1510out:
1511 return err;
1512}
1513
1514static int record__process_signal_event(union perf_event *event __maybe_unused, void *data)
1515{
1516 struct record *rec = data;
1517 pthread_kill(rec->thread_id, SIGUSR2);
1518 return 0;
1519}
1520
1521static int record__setup_sb_evlist(struct record *rec)
1522{
1523 struct record_opts *opts = &rec->opts;
1524
1525 if (rec->sb_evlist != NULL) {
1526 /*
1527 * We get here if --switch-output-event populated the
1528 * sb_evlist, so associate a callback that will send a SIGUSR2
1529 * to the main thread.
1530 */
1531 evlist__set_cb(rec->sb_evlist, record__process_signal_event, rec);
1532 rec->thread_id = pthread_self();
1533 }
1534#ifdef HAVE_LIBBPF_SUPPORT
1535 if (!opts->no_bpf_event) {
1536 if (rec->sb_evlist == NULL) {
1537 rec->sb_evlist = evlist__new();
1538
1539 if (rec->sb_evlist == NULL) {
1540 pr_err("Couldn't create side band evlist.\n.");
1541 return -1;
1542 }
1543 }
1544
1545 if (evlist__add_bpf_sb_event(rec->sb_evlist, &rec->session->header.env)) {
1546 pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1547 return -1;
1548 }
1549 }
1550#endif
1551 if (perf_evlist__start_sb_thread(rec->sb_evlist, &rec->opts.target)) {
1552 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1553 opts->no_bpf_event = true;
1554 }
1555
1556 return 0;
1557}
1558
1559static int record__init_clock(struct record *rec)
1560{
1561 struct perf_session *session = rec->session;
1562 struct timespec ref_clockid;
1563 struct timeval ref_tod;
1564 u64 ref;
1565
1566 if (!rec->opts.use_clockid)
1567 return 0;
1568
1569 if (rec->opts.use_clockid && rec->opts.clockid_res_ns)
1570 session->header.env.clock.clockid_res_ns = rec->opts.clockid_res_ns;
1571
1572 session->header.env.clock.clockid = rec->opts.clockid;
1573
1574 if (gettimeofday(&ref_tod, NULL) != 0) {
1575 pr_err("gettimeofday failed, cannot set reference time.\n");
1576 return -1;
1577 }
1578
1579 if (clock_gettime(rec->opts.clockid, &ref_clockid)) {
1580 pr_err("clock_gettime failed, cannot set reference time.\n");
1581 return -1;
1582 }
1583
1584 ref = (u64) ref_tod.tv_sec * NSEC_PER_SEC +
1585 (u64) ref_tod.tv_usec * NSEC_PER_USEC;
1586
1587 session->header.env.clock.tod_ns = ref;
1588
1589 ref = (u64) ref_clockid.tv_sec * NSEC_PER_SEC +
1590 (u64) ref_clockid.tv_nsec;
1591
1592 session->header.env.clock.clockid_ns = ref;
1593 return 0;
1594}
1595
1596static int __cmd_record(struct record *rec, int argc, const char **argv)
1597{
1598 int err;
1599 int status = 0;
1600 unsigned long waking = 0;
1601 const bool forks = argc > 0;
1602 struct perf_tool *tool = &rec->tool;
1603 struct record_opts *opts = &rec->opts;
1604 struct perf_data *data = &rec->data;
1605 struct perf_session *session;
1606 bool disabled = false, draining = false;
1607 int fd;
1608 float ratio = 0;
1609 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1610
1611 atexit(record__sig_exit);
1612 signal(SIGCHLD, sig_handler);
1613 signal(SIGINT, sig_handler);
1614 signal(SIGTERM, sig_handler);
1615 signal(SIGSEGV, sigsegv_handler);
1616
1617 if (rec->opts.record_namespaces)
1618 tool->namespace_events = true;
1619
1620 if (rec->opts.record_cgroup) {
1621#ifdef HAVE_FILE_HANDLE
1622 tool->cgroup_events = true;
1623#else
1624 pr_err("cgroup tracking is not supported\n");
1625 return -1;
1626#endif
1627 }
1628
1629 if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
1630 signal(SIGUSR2, snapshot_sig_handler);
1631 if (rec->opts.auxtrace_snapshot_mode)
1632 trigger_on(&auxtrace_snapshot_trigger);
1633 if (rec->switch_output.enabled)
1634 trigger_on(&switch_output_trigger);
1635 } else {
1636 signal(SIGUSR2, SIG_IGN);
1637 }
1638
1639 session = perf_session__new(data, false, tool);
1640 if (IS_ERR(session)) {
1641 pr_err("Perf session creation failed.\n");
1642 return PTR_ERR(session);
1643 }
1644
1645 fd = perf_data__fd(data);
1646 rec->session = session;
1647
1648 if (zstd_init(&session->zstd_data, rec->opts.comp_level) < 0) {
1649 pr_err("Compression initialization failed.\n");
1650 return -1;
1651 }
1652#ifdef HAVE_EVENTFD_SUPPORT
1653 done_fd = eventfd(0, EFD_NONBLOCK);
1654 if (done_fd < 0) {
1655 pr_err("Failed to create wakeup eventfd, error: %m\n");
1656 status = -1;
1657 goto out_delete_session;
1658 }
1659 err = evlist__add_pollfd(rec->evlist, done_fd);
1660 if (err < 0) {
1661 pr_err("Failed to add wakeup eventfd to poll list\n");
1662 status = err;
1663 goto out_delete_session;
1664 }
1665#endif // HAVE_EVENTFD_SUPPORT
1666
1667 session->header.env.comp_type = PERF_COMP_ZSTD;
1668 session->header.env.comp_level = rec->opts.comp_level;
1669
1670 if (rec->opts.kcore &&
1671 !record__kcore_readable(&session->machines.host)) {
1672 pr_err("ERROR: kcore is not readable.\n");
1673 return -1;
1674 }
1675
1676 if (record__init_clock(rec))
1677 return -1;
1678
1679 record__init_features(rec);
1680
1681 if (forks) {
1682 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
1683 argv, data->is_pipe,
1684 workload_exec_failed_signal);
1685 if (err < 0) {
1686 pr_err("Couldn't run the workload!\n");
1687 status = err;
1688 goto out_delete_session;
1689 }
1690 }
1691
1692 /*
1693 * If we have just single event and are sending data
1694 * through pipe, we need to force the ids allocation,
1695 * because we synthesize event name through the pipe
1696 * and need the id for that.
1697 */
1698 if (data->is_pipe && rec->evlist->core.nr_entries == 1)
1699 rec->opts.sample_id = true;
1700
1701 if (record__open(rec) != 0) {
1702 err = -1;
1703 goto out_child;
1704 }
1705 session->header.env.comp_mmap_len = session->evlist->core.mmap_len;
1706
1707 if (rec->opts.kcore) {
1708 err = record__kcore_copy(&session->machines.host, data);
1709 if (err) {
1710 pr_err("ERROR: Failed to copy kcore\n");
1711 goto out_child;
1712 }
1713 }
1714
1715 err = bpf__apply_obj_config();
1716 if (err) {
1717 char errbuf[BUFSIZ];
1718
1719 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
1720 pr_err("ERROR: Apply config to BPF failed: %s\n",
1721 errbuf);
1722 goto out_child;
1723 }
1724
1725 /*
1726 * Normally perf_session__new would do this, but it doesn't have the
1727 * evlist.
1728 */
1729 if (rec->tool.ordered_events && !evlist__sample_id_all(rec->evlist)) {
1730 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
1731 rec->tool.ordered_events = false;
1732 }
1733
1734 if (!rec->evlist->nr_groups)
1735 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
1736
1737 if (data->is_pipe) {
1738 err = perf_header__write_pipe(fd);
1739 if (err < 0)
1740 goto out_child;
1741 } else {
1742 err = perf_session__write_header(session, rec->evlist, fd, false);
1743 if (err < 0)
1744 goto out_child;
1745 }
1746
1747 err = -1;
1748 if (!rec->no_buildid
1749 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
1750 pr_err("Couldn't generate buildids. "
1751 "Use --no-buildid to profile anyway.\n");
1752 goto out_child;
1753 }
1754
1755 err = record__setup_sb_evlist(rec);
1756 if (err)
1757 goto out_child;
1758
1759 err = record__synthesize(rec, false);
1760 if (err < 0)
1761 goto out_child;
1762
1763 if (rec->realtime_prio) {
1764 struct sched_param param;
1765
1766 param.sched_priority = rec->realtime_prio;
1767 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1768 pr_err("Could not set realtime priority.\n");
1769 err = -1;
1770 goto out_child;
1771 }
1772 }
1773
1774 /*
1775 * When perf is starting the traced process, all the events
1776 * (apart from group members) have enable_on_exec=1 set,
1777 * so don't spoil it by prematurely enabling them.
1778 */
1779 if (!target__none(&opts->target) && !opts->initial_delay)
1780 evlist__enable(rec->evlist);
1781
1782 /*
1783 * Let the child rip
1784 */
1785 if (forks) {
1786 struct machine *machine = &session->machines.host;
1787 union perf_event *event;
1788 pid_t tgid;
1789
1790 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
1791 if (event == NULL) {
1792 err = -ENOMEM;
1793 goto out_child;
1794 }
1795
1796 /*
1797 * Some H/W events are generated before COMM event
1798 * which is emitted during exec(), so perf script
1799 * cannot see a correct process name for those events.
1800 * Synthesize COMM event to prevent it.
1801 */
1802 tgid = perf_event__synthesize_comm(tool, event,
1803 rec->evlist->workload.pid,
1804 process_synthesized_event,
1805 machine);
1806 free(event);
1807
1808 if (tgid == -1)
1809 goto out_child;
1810
1811 event = malloc(sizeof(event->namespaces) +
1812 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
1813 machine->id_hdr_size);
1814 if (event == NULL) {
1815 err = -ENOMEM;
1816 goto out_child;
1817 }
1818
1819 /*
1820 * Synthesize NAMESPACES event for the command specified.
1821 */
1822 perf_event__synthesize_namespaces(tool, event,
1823 rec->evlist->workload.pid,
1824 tgid, process_synthesized_event,
1825 machine);
1826 free(event);
1827
1828 perf_evlist__start_workload(rec->evlist);
1829 }
1830
1831 if (evlist__initialize_ctlfd(rec->evlist, opts->ctl_fd, opts->ctl_fd_ack))
1832 goto out_child;
1833
1834 if (opts->initial_delay) {
1835 pr_info(EVLIST_DISABLED_MSG);
1836 if (opts->initial_delay > 0) {
1837 usleep(opts->initial_delay * USEC_PER_MSEC);
1838 evlist__enable(rec->evlist);
1839 pr_info(EVLIST_ENABLED_MSG);
1840 }
1841 }
1842
1843 trigger_ready(&auxtrace_snapshot_trigger);
1844 trigger_ready(&switch_output_trigger);
1845 perf_hooks__invoke_record_start();
1846 for (;;) {
1847 unsigned long long hits = rec->samples;
1848
1849 /*
1850 * rec->evlist->bkw_mmap_state is possible to be
1851 * BKW_MMAP_EMPTY here: when done == true and
1852 * hits != rec->samples in previous round.
1853 *
1854 * perf_evlist__toggle_bkw_mmap ensure we never
1855 * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
1856 */
1857 if (trigger_is_hit(&switch_output_trigger) || done || draining)
1858 perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
1859
1860 if (record__mmap_read_all(rec, false) < 0) {
1861 trigger_error(&auxtrace_snapshot_trigger);
1862 trigger_error(&switch_output_trigger);
1863 err = -1;
1864 goto out_child;
1865 }
1866
1867 if (auxtrace_record__snapshot_started) {
1868 auxtrace_record__snapshot_started = 0;
1869 if (!trigger_is_error(&auxtrace_snapshot_trigger))
1870 record__read_auxtrace_snapshot(rec, false);
1871 if (trigger_is_error(&auxtrace_snapshot_trigger)) {
1872 pr_err("AUX area tracing snapshot failed\n");
1873 err = -1;
1874 goto out_child;
1875 }
1876 }
1877
1878 if (trigger_is_hit(&switch_output_trigger)) {
1879 /*
1880 * If switch_output_trigger is hit, the data in
1881 * overwritable ring buffer should have been collected,
1882 * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1883 *
1884 * If SIGUSR2 raise after or during record__mmap_read_all(),
1885 * record__mmap_read_all() didn't collect data from
1886 * overwritable ring buffer. Read again.
1887 */
1888 if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1889 continue;
1890 trigger_ready(&switch_output_trigger);
1891
1892 /*
1893 * Reenable events in overwrite ring buffer after
1894 * record__mmap_read_all(): we should have collected
1895 * data from it.
1896 */
1897 perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1898
1899 if (!quiet)
1900 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1901 waking);
1902 waking = 0;
1903 fd = record__switch_output(rec, false);
1904 if (fd < 0) {
1905 pr_err("Failed to switch to new file\n");
1906 trigger_error(&switch_output_trigger);
1907 err = fd;
1908 goto out_child;
1909 }
1910
1911 /* re-arm the alarm */
1912 if (rec->switch_output.time)
1913 alarm(rec->switch_output.time);
1914 }
1915
1916 if (hits == rec->samples) {
1917 if (done || draining)
1918 break;
1919 err = evlist__poll(rec->evlist, -1);
1920 /*
1921 * Propagate error, only if there's any. Ignore positive
1922 * number of returned events and interrupt error.
1923 */
1924 if (err > 0 || (err < 0 && errno == EINTR))
1925 err = 0;
1926 waking++;
1927
1928 if (evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1929 draining = true;
1930 }
1931
1932 if (evlist__ctlfd_process(rec->evlist, &cmd) > 0) {
1933 switch (cmd) {
1934 case EVLIST_CTL_CMD_ENABLE:
1935 pr_info(EVLIST_ENABLED_MSG);
1936 break;
1937 case EVLIST_CTL_CMD_DISABLE:
1938 pr_info(EVLIST_DISABLED_MSG);
1939 break;
1940 case EVLIST_CTL_CMD_ACK:
1941 case EVLIST_CTL_CMD_UNSUPPORTED:
1942 default:
1943 break;
1944 }
1945 }
1946
1947 /*
1948 * When perf is starting the traced process, at the end events
1949 * die with the process and we wait for that. Thus no need to
1950 * disable events in this case.
1951 */
1952 if (done && !disabled && !target__none(&opts->target)) {
1953 trigger_off(&auxtrace_snapshot_trigger);
1954 evlist__disable(rec->evlist);
1955 disabled = true;
1956 }
1957 }
1958
1959 trigger_off(&auxtrace_snapshot_trigger);
1960 trigger_off(&switch_output_trigger);
1961
1962 if (opts->auxtrace_snapshot_on_exit)
1963 record__auxtrace_snapshot_exit(rec);
1964
1965 if (forks && workload_exec_errno) {
1966 char msg[STRERR_BUFSIZE];
1967 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
1968 pr_err("Workload failed: %s\n", emsg);
1969 err = -1;
1970 goto out_child;
1971 }
1972
1973 if (!quiet)
1974 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1975
1976 if (target__none(&rec->opts.target))
1977 record__synthesize_workload(rec, true);
1978
1979out_child:
1980 evlist__finalize_ctlfd(rec->evlist);
1981 record__mmap_read_all(rec, true);
1982 record__aio_mmap_read_sync(rec);
1983
1984 if (rec->session->bytes_transferred && rec->session->bytes_compressed) {
1985 ratio = (float)rec->session->bytes_transferred/(float)rec->session->bytes_compressed;
1986 session->header.env.comp_ratio = ratio + 0.5;
1987 }
1988
1989 if (forks) {
1990 int exit_status;
1991
1992 if (!child_finished)
1993 kill(rec->evlist->workload.pid, SIGTERM);
1994
1995 wait(&exit_status);
1996
1997 if (err < 0)
1998 status = err;
1999 else if (WIFEXITED(exit_status))
2000 status = WEXITSTATUS(exit_status);
2001 else if (WIFSIGNALED(exit_status))
2002 signr = WTERMSIG(exit_status);
2003 } else
2004 status = err;
2005
2006 record__synthesize(rec, true);
2007 /* this will be recalculated during process_buildids() */
2008 rec->samples = 0;
2009
2010 if (!err) {
2011 if (!rec->timestamp_filename) {
2012 record__finish_output(rec);
2013 } else {
2014 fd = record__switch_output(rec, true);
2015 if (fd < 0) {
2016 status = fd;
2017 goto out_delete_session;
2018 }
2019 }
2020 }
2021
2022 perf_hooks__invoke_record_end();
2023
2024 if (!err && !quiet) {
2025 char samples[128];
2026 const char *postfix = rec->timestamp_filename ?
2027 ".<timestamp>" : "";
2028
2029 if (rec->samples && !rec->opts.full_auxtrace)
2030 scnprintf(samples, sizeof(samples),
2031 " (%" PRIu64 " samples)", rec->samples);
2032 else
2033 samples[0] = '\0';
2034
2035 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s",
2036 perf_data__size(data) / 1024.0 / 1024.0,
2037 data->path, postfix, samples);
2038 if (ratio) {
2039 fprintf(stderr, ", compressed (original %.3f MB, ratio is %.3f)",
2040 rec->session->bytes_transferred / 1024.0 / 1024.0,
2041 ratio);
2042 }
2043 fprintf(stderr, " ]\n");
2044 }
2045
2046out_delete_session:
2047#ifdef HAVE_EVENTFD_SUPPORT
2048 if (done_fd >= 0)
2049 close(done_fd);
2050#endif
2051 zstd_fini(&session->zstd_data);
2052 perf_session__delete(session);
2053
2054 if (!opts->no_bpf_event)
2055 perf_evlist__stop_sb_thread(rec->sb_evlist);
2056 return status;
2057}
2058
2059static void callchain_debug(struct callchain_param *callchain)
2060{
2061 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
2062
2063 pr_debug("callchain: type %s\n", str[callchain->record_mode]);
2064
2065 if (callchain->record_mode == CALLCHAIN_DWARF)
2066 pr_debug("callchain: stack dump size %d\n",
2067 callchain->dump_size);
2068}
2069
2070int record_opts__parse_callchain(struct record_opts *record,
2071 struct callchain_param *callchain,
2072 const char *arg, bool unset)
2073{
2074 int ret;
2075 callchain->enabled = !unset;
2076
2077 /* --no-call-graph */
2078 if (unset) {
2079 callchain->record_mode = CALLCHAIN_NONE;
2080 pr_debug("callchain: disabled\n");
2081 return 0;
2082 }
2083
2084 ret = parse_callchain_record_opt(arg, callchain);
2085 if (!ret) {
2086 /* Enable data address sampling for DWARF unwind. */
2087 if (callchain->record_mode == CALLCHAIN_DWARF)
2088 record->sample_address = true;
2089 callchain_debug(callchain);
2090 }
2091
2092 return ret;
2093}
2094
2095int record_parse_callchain_opt(const struct option *opt,
2096 const char *arg,
2097 int unset)
2098{
2099 return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
2100}
2101
2102int record_callchain_opt(const struct option *opt,
2103 const char *arg __maybe_unused,
2104 int unset __maybe_unused)
2105{
2106 struct callchain_param *callchain = opt->value;
2107
2108 callchain->enabled = true;
2109
2110 if (callchain->record_mode == CALLCHAIN_NONE)
2111 callchain->record_mode = CALLCHAIN_FP;
2112
2113 callchain_debug(callchain);
2114 return 0;
2115}
2116
2117static int perf_record_config(const char *var, const char *value, void *cb)
2118{
2119 struct record *rec = cb;
2120
2121 if (!strcmp(var, "record.build-id")) {
2122 if (!strcmp(value, "cache"))
2123 rec->no_buildid_cache = false;
2124 else if (!strcmp(value, "no-cache"))
2125 rec->no_buildid_cache = true;
2126 else if (!strcmp(value, "skip"))
2127 rec->no_buildid = true;
2128 else
2129 return -1;
2130 return 0;
2131 }
2132 if (!strcmp(var, "record.call-graph")) {
2133 var = "call-graph.record-mode";
2134 return perf_default_config(var, value, cb);
2135 }
2136#ifdef HAVE_AIO_SUPPORT
2137 if (!strcmp(var, "record.aio")) {
2138 rec->opts.nr_cblocks = strtol(value, NULL, 0);
2139 if (!rec->opts.nr_cblocks)
2140 rec->opts.nr_cblocks = nr_cblocks_default;
2141 }
2142#endif
2143
2144 return 0;
2145}
2146
2147
2148static int record__parse_affinity(const struct option *opt, const char *str, int unset)
2149{
2150 struct record_opts *opts = (struct record_opts *)opt->value;
2151
2152 if (unset || !str)
2153 return 0;
2154
2155 if (!strcasecmp(str, "node"))
2156 opts->affinity = PERF_AFFINITY_NODE;
2157 else if (!strcasecmp(str, "cpu"))
2158 opts->affinity = PERF_AFFINITY_CPU;
2159
2160 return 0;
2161}
2162
2163static int parse_output_max_size(const struct option *opt,
2164 const char *str, int unset)
2165{
2166 unsigned long *s = (unsigned long *)opt->value;
2167 static struct parse_tag tags_size[] = {
2168 { .tag = 'B', .mult = 1 },
2169 { .tag = 'K', .mult = 1 << 10 },
2170 { .tag = 'M', .mult = 1 << 20 },
2171 { .tag = 'G', .mult = 1 << 30 },
2172 { .tag = 0 },
2173 };
2174 unsigned long val;
2175
2176 if (unset) {
2177 *s = 0;
2178 return 0;
2179 }
2180
2181 val = parse_tag_value(str, tags_size);
2182 if (val != (unsigned long) -1) {
2183 *s = val;
2184 return 0;
2185 }
2186
2187 return -1;
2188}
2189
2190static int record__parse_mmap_pages(const struct option *opt,
2191 const char *str,
2192 int unset __maybe_unused)
2193{
2194 struct record_opts *opts = opt->value;
2195 char *s, *p;
2196 unsigned int mmap_pages;
2197 int ret;
2198
2199 if (!str)
2200 return -EINVAL;
2201
2202 s = strdup(str);
2203 if (!s)
2204 return -ENOMEM;
2205
2206 p = strchr(s, ',');
2207 if (p)
2208 *p = '\0';
2209
2210 if (*s) {
2211 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
2212 if (ret)
2213 goto out_free;
2214 opts->mmap_pages = mmap_pages;
2215 }
2216
2217 if (!p) {
2218 ret = 0;
2219 goto out_free;
2220 }
2221
2222 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
2223 if (ret)
2224 goto out_free;
2225
2226 opts->auxtrace_mmap_pages = mmap_pages;
2227
2228out_free:
2229 free(s);
2230 return ret;
2231}
2232
2233static int parse_control_option(const struct option *opt,
2234 const char *str,
2235 int unset __maybe_unused)
2236{
2237 char *comma = NULL, *endptr = NULL;
2238 struct record_opts *config = (struct record_opts *)opt->value;
2239
2240 if (strncmp(str, "fd:", 3))
2241 return -EINVAL;
2242
2243 config->ctl_fd = strtoul(&str[3], &endptr, 0);
2244 if (endptr == &str[3])
2245 return -EINVAL;
2246
2247 comma = strchr(str, ',');
2248 if (comma) {
2249 if (endptr != comma)
2250 return -EINVAL;
2251
2252 config->ctl_fd_ack = strtoul(comma + 1, &endptr, 0);
2253 if (endptr == comma + 1 || *endptr != '\0')
2254 return -EINVAL;
2255 }
2256
2257 return 0;
2258}
2259
2260static void switch_output_size_warn(struct record *rec)
2261{
2262 u64 wakeup_size = evlist__mmap_size(rec->opts.mmap_pages);
2263 struct switch_output *s = &rec->switch_output;
2264
2265 wakeup_size /= 2;
2266
2267 if (s->size < wakeup_size) {
2268 char buf[100];
2269
2270 unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
2271 pr_warning("WARNING: switch-output data size lower than "
2272 "wakeup kernel buffer size (%s) "
2273 "expect bigger perf.data sizes\n", buf);
2274 }
2275}
2276
2277static int switch_output_setup(struct record *rec)
2278{
2279 struct switch_output *s = &rec->switch_output;
2280 static struct parse_tag tags_size[] = {
2281 { .tag = 'B', .mult = 1 },
2282 { .tag = 'K', .mult = 1 << 10 },
2283 { .tag = 'M', .mult = 1 << 20 },
2284 { .tag = 'G', .mult = 1 << 30 },
2285 { .tag = 0 },
2286 };
2287 static struct parse_tag tags_time[] = {
2288 { .tag = 's', .mult = 1 },
2289 { .tag = 'm', .mult = 60 },
2290 { .tag = 'h', .mult = 60*60 },
2291 { .tag = 'd', .mult = 60*60*24 },
2292 { .tag = 0 },
2293 };
2294 unsigned long val;
2295
2296 /*
2297 * If we're using --switch-output-events, then we imply its
2298 * --switch-output=signal, as we'll send a SIGUSR2 from the side band
2299 * thread to its parent.
2300 */
2301 if (rec->switch_output_event_set)
2302 goto do_signal;
2303
2304 if (!s->set)
2305 return 0;
2306
2307 if (!strcmp(s->str, "signal")) {
2308do_signal:
2309 s->signal = true;
2310 pr_debug("switch-output with SIGUSR2 signal\n");
2311 goto enabled;
2312 }
2313
2314 val = parse_tag_value(s->str, tags_size);
2315 if (val != (unsigned long) -1) {
2316 s->size = val;
2317 pr_debug("switch-output with %s size threshold\n", s->str);
2318 goto enabled;
2319 }
2320
2321 val = parse_tag_value(s->str, tags_time);
2322 if (val != (unsigned long) -1) {
2323 s->time = val;
2324 pr_debug("switch-output with %s time threshold (%lu seconds)\n",
2325 s->str, s->time);
2326 goto enabled;
2327 }
2328
2329 return -1;
2330
2331enabled:
2332 rec->timestamp_filename = true;
2333 s->enabled = true;
2334
2335 if (s->size && !rec->opts.no_buffering)
2336 switch_output_size_warn(rec);
2337
2338 return 0;
2339}
2340
2341static const char * const __record_usage[] = {
2342 "perf record [<options>] [<command>]",
2343 "perf record [<options>] -- <command> [<options>]",
2344 NULL
2345};
2346const char * const *record_usage = __record_usage;
2347
2348static int build_id__process_mmap(struct perf_tool *tool, union perf_event *event,
2349 struct perf_sample *sample, struct machine *machine)
2350{
2351 /*
2352 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
2353 * no need to add them twice.
2354 */
2355 if (!(event->header.misc & PERF_RECORD_MISC_USER))
2356 return 0;
2357 return perf_event__process_mmap(tool, event, sample, machine);
2358}
2359
2360static int build_id__process_mmap2(struct perf_tool *tool, union perf_event *event,
2361 struct perf_sample *sample, struct machine *machine)
2362{
2363 /*
2364 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
2365 * no need to add them twice.
2366 */
2367 if (!(event->header.misc & PERF_RECORD_MISC_USER))
2368 return 0;
2369
2370 return perf_event__process_mmap2(tool, event, sample, machine);
2371}
2372
2373/*
2374 * XXX Ideally would be local to cmd_record() and passed to a record__new
2375 * because we need to have access to it in record__exit, that is called
2376 * after cmd_record() exits, but since record_options need to be accessible to
2377 * builtin-script, leave it here.
2378 *
2379 * At least we don't ouch it in all the other functions here directly.
2380 *
2381 * Just say no to tons of global variables, sigh.
2382 */
2383static struct record record = {
2384 .opts = {
2385 .sample_time = true,
2386 .mmap_pages = UINT_MAX,
2387 .user_freq = UINT_MAX,
2388 .user_interval = ULLONG_MAX,
2389 .freq = 4000,
2390 .target = {
2391 .uses_mmap = true,
2392 .default_per_cpu = true,
2393 },
2394 .mmap_flush = MMAP_FLUSH_DEFAULT,
2395 .nr_threads_synthesize = 1,
2396 .ctl_fd = -1,
2397 .ctl_fd_ack = -1,
2398 },
2399 .tool = {
2400 .sample = process_sample_event,
2401 .fork = perf_event__process_fork,
2402 .exit = perf_event__process_exit,
2403 .comm = perf_event__process_comm,
2404 .namespaces = perf_event__process_namespaces,
2405 .mmap = build_id__process_mmap,
2406 .mmap2 = build_id__process_mmap2,
2407 .ordered_events = true,
2408 },
2409};
2410
2411const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
2412 "\n\t\t\t\tDefault: fp";
2413
2414static bool dry_run;
2415
2416/*
2417 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
2418 * with it and switch to use the library functions in perf_evlist that came
2419 * from builtin-record.c, i.e. use record_opts,
2420 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
2421 * using pipes, etc.
2422 */
2423static struct option __record_options[] = {
2424 OPT_CALLBACK('e', "event", &record.evlist, "event",
2425 "event selector. use 'perf list' to list available events",
2426 parse_events_option),
2427 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
2428 "event filter", parse_filter),
2429 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
2430 NULL, "don't record events from perf itself",
2431 exclude_perf),
2432 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
2433 "record events on existing process id"),
2434 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
2435 "record events on existing thread id"),
2436 OPT_INTEGER('r', "realtime", &record.realtime_prio,
2437 "collect data with this RT SCHED_FIFO priority"),
2438 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
2439 "collect data without buffering"),
2440 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
2441 "collect raw sample records from all opened counters"),
2442 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
2443 "system-wide collection from all CPUs"),
2444 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
2445 "list of cpus to monitor"),
2446 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
2447 OPT_STRING('o', "output", &record.data.path, "file",
2448 "output file name"),
2449 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
2450 &record.opts.no_inherit_set,
2451 "child tasks do not inherit counters"),
2452 OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
2453 "synthesize non-sample events at the end of output"),
2454 OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
2455 OPT_BOOLEAN(0, "no-bpf-event", &record.opts.no_bpf_event, "do not record bpf events"),
2456 OPT_BOOLEAN(0, "strict-freq", &record.opts.strict_freq,
2457 "Fail if the specified frequency can't be used"),
2458 OPT_CALLBACK('F', "freq", &record.opts, "freq or 'max'",
2459 "profile at this frequency",
2460 record__parse_freq),
2461 OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
2462 "number of mmap data pages and AUX area tracing mmap pages",
2463 record__parse_mmap_pages),
2464 OPT_CALLBACK(0, "mmap-flush", &record.opts, "number",
2465 "Minimal number of bytes that is extracted from mmap data pages (default: 1)",
2466 record__mmap_flush_parse),
2467 OPT_BOOLEAN(0, "group", &record.opts.group,
2468 "put the counters into a counter group"),
2469 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
2470 NULL, "enables call-graph recording" ,
2471 &record_callchain_opt),
2472 OPT_CALLBACK(0, "call-graph", &record.opts,
2473 "record_mode[,record_size]", record_callchain_help,
2474 &record_parse_callchain_opt),
2475 OPT_INCR('v', "verbose", &verbose,
2476 "be more verbose (show counter open errors, etc)"),
2477 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
2478 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
2479 "per thread counts"),
2480 OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
2481 OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
2482 "Record the sample physical addresses"),
2483 OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
2484 OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
2485 &record.opts.sample_time_set,
2486 "Record the sample timestamps"),
2487 OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
2488 "Record the sample period"),
2489 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
2490 "don't sample"),
2491 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
2492 &record.no_buildid_cache_set,
2493 "do not update the buildid cache"),
2494 OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
2495 &record.no_buildid_set,
2496 "do not collect buildids in perf.data"),
2497 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
2498 "monitor event in cgroup name only",
2499 parse_cgroups),
2500 OPT_INTEGER('D', "delay", &record.opts.initial_delay,
2501 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
2502 OPT_BOOLEAN(0, "kcore", &record.opts.kcore, "copy /proc/kcore"),
2503 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
2504 "user to profile"),
2505
2506 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
2507 "branch any", "sample any taken branches",
2508 parse_branch_stack),
2509
2510 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
2511 "branch filter mask", "branch stack filter modes",
2512 parse_branch_stack),
2513 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
2514 "sample by weight (on special events only)"),
2515 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
2516 "sample transaction flags (special events only)"),
2517 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
2518 "use per-thread mmaps"),
2519 OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
2520 "sample selected machine registers on interrupt,"
2521 " use '-I?' to list register names", parse_intr_regs),
2522 OPT_CALLBACK_OPTARG(0, "user-regs", &record.opts.sample_user_regs, NULL, "any register",
2523 "sample selected machine registers on interrupt,"
2524 " use '--user-regs=?' to list register names", parse_user_regs),
2525 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
2526 "Record running/enabled time of read (:S) events"),
2527 OPT_CALLBACK('k', "clockid", &record.opts,
2528 "clockid", "clockid to use for events, see clock_gettime()",
2529 parse_clockid),
2530 OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
2531 "opts", "AUX area tracing Snapshot Mode", ""),
2532 OPT_STRING_OPTARG(0, "aux-sample", &record.opts.auxtrace_sample_opts,
2533 "opts", "sample AUX area", ""),
2534 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
2535 "per thread proc mmap processing timeout in ms"),
2536 OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
2537 "Record namespaces events"),
2538 OPT_BOOLEAN(0, "all-cgroups", &record.opts.record_cgroup,
2539 "Record cgroup events"),
2540 OPT_BOOLEAN_SET(0, "switch-events", &record.opts.record_switch_events,
2541 &record.opts.record_switch_events_set,
2542 "Record context switch events"),
2543 OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
2544 "Configure all used events to run in kernel space.",
2545 PARSE_OPT_EXCLUSIVE),
2546 OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
2547 "Configure all used events to run in user space.",
2548 PARSE_OPT_EXCLUSIVE),
2549 OPT_BOOLEAN(0, "kernel-callchains", &record.opts.kernel_callchains,
2550 "collect kernel callchains"),
2551 OPT_BOOLEAN(0, "user-callchains", &record.opts.user_callchains,
2552 "collect user callchains"),
2553 OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
2554 "clang binary to use for compiling BPF scriptlets"),
2555 OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
2556 "options passed to clang when compiling BPF scriptlets"),
2557 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
2558 "file", "vmlinux pathname"),
2559 OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
2560 "Record build-id of all DSOs regardless of hits"),
2561 OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
2562 "append timestamp to output filename"),
2563 OPT_BOOLEAN(0, "timestamp-boundary", &record.timestamp_boundary,
2564 "Record timestamp boundary (time of first/last samples)"),
2565 OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
2566 &record.switch_output.set, "signal or size[BKMG] or time[smhd]",
2567 "Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold",
2568 "signal"),
2569 OPT_CALLBACK_SET(0, "switch-output-event", &record.sb_evlist, &record.switch_output_event_set, "switch output event",
2570 "switch output event selector. use 'perf list' to list available events",
2571 parse_events_option_new_evlist),
2572 OPT_INTEGER(0, "switch-max-files", &record.switch_output.num_files,
2573 "Limit number of switch output generated files"),
2574 OPT_BOOLEAN(0, "dry-run", &dry_run,
2575 "Parse options then exit"),
2576#ifdef HAVE_AIO_SUPPORT
2577 OPT_CALLBACK_OPTARG(0, "aio", &record.opts,
2578 &nr_cblocks_default, "n", "Use <n> control blocks in asynchronous trace writing mode (default: 1, max: 4)",
2579 record__aio_parse),
2580#endif
2581 OPT_CALLBACK(0, "affinity", &record.opts, "node|cpu",
2582 "Set affinity mask of trace reading thread to NUMA node cpu mask or cpu of processed mmap buffer",
2583 record__parse_affinity),
2584#ifdef HAVE_ZSTD_SUPPORT
2585 OPT_CALLBACK_OPTARG('z', "compression-level", &record.opts, &comp_level_default,
2586 "n", "Compressed records using specified level (default: 1 - fastest compression, 22 - greatest compression)",
2587 record__parse_comp_level),
2588#endif
2589 OPT_CALLBACK(0, "max-size", &record.output_max_size,
2590 "size", "Limit the maximum size of the output file", parse_output_max_size),
2591 OPT_UINTEGER(0, "num-thread-synthesize",
2592 &record.opts.nr_threads_synthesize,
2593 "number of threads to run for event synthesis"),
2594#ifdef HAVE_LIBPFM
2595 OPT_CALLBACK(0, "pfm-events", &record.evlist, "event",
2596 "libpfm4 event selector. use 'perf list' to list available events",
2597 parse_libpfm_events_option),
2598#endif
2599 OPT_CALLBACK(0, "control", &record.opts, "fd:ctl-fd[,ack-fd]",
2600 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events).\n"
2601 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.",
2602 parse_control_option),
2603 OPT_END()
2604};
2605
2606struct option *record_options = __record_options;
2607
2608int cmd_record(int argc, const char **argv)
2609{
2610 int err;
2611 struct record *rec = &record;
2612 char errbuf[BUFSIZ];
2613
2614 setlocale(LC_ALL, "");
2615
2616#ifndef HAVE_LIBBPF_SUPPORT
2617# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
2618 set_nobuild('\0', "clang-path", true);
2619 set_nobuild('\0', "clang-opt", true);
2620# undef set_nobuild
2621#endif
2622
2623#ifndef HAVE_BPF_PROLOGUE
2624# if !defined (HAVE_DWARF_SUPPORT)
2625# define REASON "NO_DWARF=1"
2626# elif !defined (HAVE_LIBBPF_SUPPORT)
2627# define REASON "NO_LIBBPF=1"
2628# else
2629# define REASON "this architecture doesn't support BPF prologue"
2630# endif
2631# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
2632 set_nobuild('\0', "vmlinux", true);
2633# undef set_nobuild
2634# undef REASON
2635#endif
2636
2637 rec->opts.affinity = PERF_AFFINITY_SYS;
2638
2639 rec->evlist = evlist__new();
2640 if (rec->evlist == NULL)
2641 return -ENOMEM;
2642
2643 err = perf_config(perf_record_config, rec);
2644 if (err)
2645 return err;
2646
2647 argc = parse_options(argc, argv, record_options, record_usage,
2648 PARSE_OPT_STOP_AT_NON_OPTION);
2649 if (quiet)
2650 perf_quiet_option();
2651
2652 /* Make system wide (-a) the default target. */
2653 if (!argc && target__none(&rec->opts.target))
2654 rec->opts.target.system_wide = true;
2655
2656 if (nr_cgroups && !rec->opts.target.system_wide) {
2657 usage_with_options_msg(record_usage, record_options,
2658 "cgroup monitoring only available in system-wide mode");
2659
2660 }
2661
2662 if (rec->opts.kcore)
2663 rec->data.is_dir = true;
2664
2665 if (rec->opts.comp_level != 0) {
2666 pr_debug("Compression enabled, disabling build id collection at the end of the session.\n");
2667 rec->no_buildid = true;
2668 }
2669
2670 if (rec->opts.record_switch_events &&
2671 !perf_can_record_switch_events()) {
2672 ui__error("kernel does not support recording context switch events\n");
2673 parse_options_usage(record_usage, record_options, "switch-events", 0);
2674 return -EINVAL;
2675 }
2676
2677 if (switch_output_setup(rec)) {
2678 parse_options_usage(record_usage, record_options, "switch-output", 0);
2679 return -EINVAL;
2680 }
2681
2682 if (rec->switch_output.time) {
2683 signal(SIGALRM, alarm_sig_handler);
2684 alarm(rec->switch_output.time);
2685 }
2686
2687 if (rec->switch_output.num_files) {
2688 rec->switch_output.filenames = calloc(sizeof(char *),
2689 rec->switch_output.num_files);
2690 if (!rec->switch_output.filenames)
2691 return -EINVAL;
2692 }
2693
2694 /*
2695 * Allow aliases to facilitate the lookup of symbols for address
2696 * filters. Refer to auxtrace_parse_filters().
2697 */
2698 symbol_conf.allow_aliases = true;
2699
2700 symbol__init(NULL);
2701
2702 if (rec->opts.affinity != PERF_AFFINITY_SYS) {
2703 rec->affinity_mask.nbits = cpu__max_cpu();
2704 rec->affinity_mask.bits = bitmap_alloc(rec->affinity_mask.nbits);
2705 if (!rec->affinity_mask.bits) {
2706 pr_err("Failed to allocate thread mask for %zd cpus\n", rec->affinity_mask.nbits);
2707 return -ENOMEM;
2708 }
2709 pr_debug2("thread mask[%zd]: empty\n", rec->affinity_mask.nbits);
2710 }
2711
2712 err = record__auxtrace_init(rec);
2713 if (err)
2714 goto out;
2715
2716 if (dry_run)
2717 goto out;
2718
2719 err = bpf__setup_stdout(rec->evlist);
2720 if (err) {
2721 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
2722 pr_err("ERROR: Setup BPF stdout failed: %s\n",
2723 errbuf);
2724 goto out;
2725 }
2726
2727 err = -ENOMEM;
2728
2729 if (rec->no_buildid_cache || rec->no_buildid) {
2730 disable_buildid_cache();
2731 } else if (rec->switch_output.enabled) {
2732 /*
2733 * In 'perf record --switch-output', disable buildid
2734 * generation by default to reduce data file switching
2735 * overhead. Still generate buildid if they are required
2736 * explicitly using
2737 *
2738 * perf record --switch-output --no-no-buildid \
2739 * --no-no-buildid-cache
2740 *
2741 * Following code equals to:
2742 *
2743 * if ((rec->no_buildid || !rec->no_buildid_set) &&
2744 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
2745 * disable_buildid_cache();
2746 */
2747 bool disable = true;
2748
2749 if (rec->no_buildid_set && !rec->no_buildid)
2750 disable = false;
2751 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
2752 disable = false;
2753 if (disable) {
2754 rec->no_buildid = true;
2755 rec->no_buildid_cache = true;
2756 disable_buildid_cache();
2757 }
2758 }
2759
2760 if (record.opts.overwrite)
2761 record.opts.tail_synthesize = true;
2762
2763 if (rec->evlist->core.nr_entries == 0 &&
2764 __evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) {
2765 pr_err("Not enough memory for event selector list\n");
2766 goto out;
2767 }
2768
2769 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
2770 rec->opts.no_inherit = true;
2771
2772 err = target__validate(&rec->opts.target);
2773 if (err) {
2774 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
2775 ui__warning("%s\n", errbuf);
2776 }
2777
2778 err = target__parse_uid(&rec->opts.target);
2779 if (err) {
2780 int saved_errno = errno;
2781
2782 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
2783 ui__error("%s", errbuf);
2784
2785 err = -saved_errno;
2786 goto out;
2787 }
2788
2789 /* Enable ignoring missing threads when -u/-p option is defined. */
2790 rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
2791
2792 err = -ENOMEM;
2793 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
2794 usage_with_options(record_usage, record_options);
2795
2796 err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
2797 if (err)
2798 goto out;
2799
2800 /*
2801 * We take all buildids when the file contains
2802 * AUX area tracing data because we do not decode the
2803 * trace because it would take too long.
2804 */
2805 if (rec->opts.full_auxtrace)
2806 rec->buildid_all = true;
2807
2808 if (rec->opts.text_poke) {
2809 err = record__config_text_poke(rec->evlist);
2810 if (err) {
2811 pr_err("record__config_text_poke failed, error %d\n", err);
2812 goto out;
2813 }
2814 }
2815
2816 if (record_opts__config(&rec->opts)) {
2817 err = -EINVAL;
2818 goto out;
2819 }
2820
2821 if (rec->opts.nr_cblocks > nr_cblocks_max)
2822 rec->opts.nr_cblocks = nr_cblocks_max;
2823 pr_debug("nr_cblocks: %d\n", rec->opts.nr_cblocks);
2824
2825 pr_debug("affinity: %s\n", affinity_tags[rec->opts.affinity]);
2826 pr_debug("mmap flush: %d\n", rec->opts.mmap_flush);
2827
2828 if (rec->opts.comp_level > comp_level_max)
2829 rec->opts.comp_level = comp_level_max;
2830 pr_debug("comp level: %d\n", rec->opts.comp_level);
2831
2832 err = __cmd_record(&record, argc, argv);
2833out:
2834 bitmap_free(rec->affinity_mask.bits);
2835 evlist__delete(rec->evlist);
2836 symbol__exit();
2837 auxtrace_record__free(rec->itr);
2838 return err;
2839}
2840
2841static void snapshot_sig_handler(int sig __maybe_unused)
2842{
2843 struct record *rec = &record;
2844
2845 if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
2846 trigger_hit(&auxtrace_snapshot_trigger);
2847 auxtrace_record__snapshot_started = 1;
2848 if (auxtrace_record__snapshot_start(record.itr))
2849 trigger_error(&auxtrace_snapshot_trigger);
2850 }
2851
2852 if (switch_output_signal(rec))
2853 trigger_hit(&switch_output_trigger);
2854}
2855
2856static void alarm_sig_handler(int sig __maybe_unused)
2857{
2858 struct record *rec = &record;
2859
2860 if (switch_output_time(rec))
2861 trigger_hit(&switch_output_trigger);
2862}