Loading...
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 "perf.h"
12
13#include "util/build-id.h"
14#include "util/util.h"
15#include <subcmd/parse-options.h>
16#include "util/parse-events.h"
17#include "util/config.h"
18
19#include "util/callchain.h"
20#include "util/cgroup.h"
21#include "util/header.h"
22#include "util/event.h"
23#include "util/evlist.h"
24#include "util/evsel.h"
25#include "util/debug.h"
26#include "util/drv_configs.h"
27#include "util/session.h"
28#include "util/tool.h"
29#include "util/symbol.h"
30#include "util/cpumap.h"
31#include "util/thread_map.h"
32#include "util/data.h"
33#include "util/perf_regs.h"
34#include "util/auxtrace.h"
35#include "util/tsc.h"
36#include "util/parse-branch-options.h"
37#include "util/parse-regs-options.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/time-utils.h"
43#include "util/units.h"
44#include "asm/bug.h"
45
46#include <errno.h>
47#include <inttypes.h>
48#include <locale.h>
49#include <poll.h>
50#include <unistd.h>
51#include <sched.h>
52#include <signal.h>
53#include <sys/mman.h>
54#include <sys/wait.h>
55#include <linux/time64.h>
56
57struct switch_output {
58 bool enabled;
59 bool signal;
60 unsigned long size;
61 unsigned long time;
62 const char *str;
63 bool set;
64};
65
66struct record {
67 struct perf_tool tool;
68 struct record_opts opts;
69 u64 bytes_written;
70 struct perf_data data;
71 struct auxtrace_record *itr;
72 struct perf_evlist *evlist;
73 struct perf_session *session;
74 int realtime_prio;
75 bool no_buildid;
76 bool no_buildid_set;
77 bool no_buildid_cache;
78 bool no_buildid_cache_set;
79 bool buildid_all;
80 bool timestamp_filename;
81 bool timestamp_boundary;
82 struct switch_output switch_output;
83 unsigned long long samples;
84};
85
86static volatile int auxtrace_record__snapshot_started;
87static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
88static DEFINE_TRIGGER(switch_output_trigger);
89
90static bool switch_output_signal(struct record *rec)
91{
92 return rec->switch_output.signal &&
93 trigger_is_ready(&switch_output_trigger);
94}
95
96static bool switch_output_size(struct record *rec)
97{
98 return rec->switch_output.size &&
99 trigger_is_ready(&switch_output_trigger) &&
100 (rec->bytes_written >= rec->switch_output.size);
101}
102
103static bool switch_output_time(struct record *rec)
104{
105 return rec->switch_output.time &&
106 trigger_is_ready(&switch_output_trigger);
107}
108
109static int record__write(struct record *rec, void *bf, size_t size)
110{
111 if (perf_data__write(rec->session->data, bf, size) < 0) {
112 pr_err("failed to write perf data, error: %m\n");
113 return -1;
114 }
115
116 rec->bytes_written += size;
117
118 if (switch_output_size(rec))
119 trigger_hit(&switch_output_trigger);
120
121 return 0;
122}
123
124static int process_synthesized_event(struct perf_tool *tool,
125 union perf_event *event,
126 struct perf_sample *sample __maybe_unused,
127 struct machine *machine __maybe_unused)
128{
129 struct record *rec = container_of(tool, struct record, tool);
130 return record__write(rec, event, event->header.size);
131}
132
133static int record__pushfn(void *to, void *bf, size_t size)
134{
135 struct record *rec = to;
136
137 rec->samples++;
138 return record__write(rec, bf, size);
139}
140
141static volatile int done;
142static volatile int signr = -1;
143static volatile int child_finished;
144
145static void sig_handler(int sig)
146{
147 if (sig == SIGCHLD)
148 child_finished = 1;
149 else
150 signr = sig;
151
152 done = 1;
153}
154
155static void sigsegv_handler(int sig)
156{
157 perf_hooks__recover();
158 sighandler_dump_stack(sig);
159}
160
161static void record__sig_exit(void)
162{
163 if (signr == -1)
164 return;
165
166 signal(signr, SIG_DFL);
167 raise(signr);
168}
169
170#ifdef HAVE_AUXTRACE_SUPPORT
171
172static int record__process_auxtrace(struct perf_tool *tool,
173 union perf_event *event, void *data1,
174 size_t len1, void *data2, size_t len2)
175{
176 struct record *rec = container_of(tool, struct record, tool);
177 struct perf_data *data = &rec->data;
178 size_t padding;
179 u8 pad[8] = {0};
180
181 if (!perf_data__is_pipe(data)) {
182 off_t file_offset;
183 int fd = perf_data__fd(data);
184 int err;
185
186 file_offset = lseek(fd, 0, SEEK_CUR);
187 if (file_offset == -1)
188 return -1;
189 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
190 event, file_offset);
191 if (err)
192 return err;
193 }
194
195 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
196 padding = (len1 + len2) & 7;
197 if (padding)
198 padding = 8 - padding;
199
200 record__write(rec, event, event->header.size);
201 record__write(rec, data1, len1);
202 if (len2)
203 record__write(rec, data2, len2);
204 record__write(rec, &pad, padding);
205
206 return 0;
207}
208
209static int record__auxtrace_mmap_read(struct record *rec,
210 struct auxtrace_mmap *mm)
211{
212 int ret;
213
214 ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
215 record__process_auxtrace);
216 if (ret < 0)
217 return ret;
218
219 if (ret)
220 rec->samples++;
221
222 return 0;
223}
224
225static int record__auxtrace_mmap_read_snapshot(struct record *rec,
226 struct auxtrace_mmap *mm)
227{
228 int ret;
229
230 ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
231 record__process_auxtrace,
232 rec->opts.auxtrace_snapshot_size);
233 if (ret < 0)
234 return ret;
235
236 if (ret)
237 rec->samples++;
238
239 return 0;
240}
241
242static int record__auxtrace_read_snapshot_all(struct record *rec)
243{
244 int i;
245 int rc = 0;
246
247 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
248 struct auxtrace_mmap *mm =
249 &rec->evlist->mmap[i].auxtrace_mmap;
250
251 if (!mm->base)
252 continue;
253
254 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
255 rc = -1;
256 goto out;
257 }
258 }
259out:
260 return rc;
261}
262
263static void record__read_auxtrace_snapshot(struct record *rec)
264{
265 pr_debug("Recording AUX area tracing snapshot\n");
266 if (record__auxtrace_read_snapshot_all(rec) < 0) {
267 trigger_error(&auxtrace_snapshot_trigger);
268 } else {
269 if (auxtrace_record__snapshot_finish(rec->itr))
270 trigger_error(&auxtrace_snapshot_trigger);
271 else
272 trigger_ready(&auxtrace_snapshot_trigger);
273 }
274}
275
276static int record__auxtrace_init(struct record *rec)
277{
278 int err;
279
280 if (!rec->itr) {
281 rec->itr = auxtrace_record__init(rec->evlist, &err);
282 if (err)
283 return err;
284 }
285
286 err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
287 rec->opts.auxtrace_snapshot_opts);
288 if (err)
289 return err;
290
291 return auxtrace_parse_filters(rec->evlist);
292}
293
294#else
295
296static inline
297int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
298 struct auxtrace_mmap *mm __maybe_unused)
299{
300 return 0;
301}
302
303static inline
304void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
305{
306}
307
308static inline
309int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
310{
311 return 0;
312}
313
314static int record__auxtrace_init(struct record *rec __maybe_unused)
315{
316 return 0;
317}
318
319#endif
320
321static int record__mmap_evlist(struct record *rec,
322 struct perf_evlist *evlist)
323{
324 struct record_opts *opts = &rec->opts;
325 char msg[512];
326
327 if (perf_evlist__mmap_ex(evlist, opts->mmap_pages,
328 opts->auxtrace_mmap_pages,
329 opts->auxtrace_snapshot_mode) < 0) {
330 if (errno == EPERM) {
331 pr_err("Permission error mapping pages.\n"
332 "Consider increasing "
333 "/proc/sys/kernel/perf_event_mlock_kb,\n"
334 "or try again with a smaller value of -m/--mmap_pages.\n"
335 "(current value: %u,%u)\n",
336 opts->mmap_pages, opts->auxtrace_mmap_pages);
337 return -errno;
338 } else {
339 pr_err("failed to mmap with %d (%s)\n", errno,
340 str_error_r(errno, msg, sizeof(msg)));
341 if (errno)
342 return -errno;
343 else
344 return -EINVAL;
345 }
346 }
347 return 0;
348}
349
350static int record__mmap(struct record *rec)
351{
352 return record__mmap_evlist(rec, rec->evlist);
353}
354
355static int record__open(struct record *rec)
356{
357 char msg[BUFSIZ];
358 struct perf_evsel *pos;
359 struct perf_evlist *evlist = rec->evlist;
360 struct perf_session *session = rec->session;
361 struct record_opts *opts = &rec->opts;
362 struct perf_evsel_config_term *err_term;
363 int rc = 0;
364
365 /*
366 * For initial_delay we need to add a dummy event so that we can track
367 * PERF_RECORD_MMAP while we wait for the initial delay to enable the
368 * real events, the ones asked by the user.
369 */
370 if (opts->initial_delay) {
371 if (perf_evlist__add_dummy(evlist))
372 return -ENOMEM;
373
374 pos = perf_evlist__first(evlist);
375 pos->tracking = 0;
376 pos = perf_evlist__last(evlist);
377 pos->tracking = 1;
378 pos->attr.enable_on_exec = 1;
379 }
380
381 perf_evlist__config(evlist, opts, &callchain_param);
382
383 evlist__for_each_entry(evlist, pos) {
384try_again:
385 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
386 if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
387 if (verbose > 0)
388 ui__warning("%s\n", msg);
389 goto try_again;
390 }
391
392 rc = -errno;
393 perf_evsel__open_strerror(pos, &opts->target,
394 errno, msg, sizeof(msg));
395 ui__error("%s\n", msg);
396 goto out;
397 }
398
399 pos->supported = true;
400 }
401
402 if (perf_evlist__apply_filters(evlist, &pos)) {
403 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
404 pos->filter, perf_evsel__name(pos), errno,
405 str_error_r(errno, msg, sizeof(msg)));
406 rc = -1;
407 goto out;
408 }
409
410 if (perf_evlist__apply_drv_configs(evlist, &pos, &err_term)) {
411 pr_err("failed to set config \"%s\" on event %s with %d (%s)\n",
412 err_term->val.drv_cfg, perf_evsel__name(pos), errno,
413 str_error_r(errno, msg, sizeof(msg)));
414 rc = -1;
415 goto out;
416 }
417
418 rc = record__mmap(rec);
419 if (rc)
420 goto out;
421
422 session->evlist = evlist;
423 perf_session__set_id_hdr_size(session);
424out:
425 return rc;
426}
427
428static int process_sample_event(struct perf_tool *tool,
429 union perf_event *event,
430 struct perf_sample *sample,
431 struct perf_evsel *evsel,
432 struct machine *machine)
433{
434 struct record *rec = container_of(tool, struct record, tool);
435
436 if (rec->evlist->first_sample_time == 0)
437 rec->evlist->first_sample_time = sample->time;
438
439 rec->evlist->last_sample_time = sample->time;
440
441 if (rec->buildid_all)
442 return 0;
443
444 rec->samples++;
445 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
446}
447
448static int process_buildids(struct record *rec)
449{
450 struct perf_data *data = &rec->data;
451 struct perf_session *session = rec->session;
452
453 if (data->size == 0)
454 return 0;
455
456 /*
457 * During this process, it'll load kernel map and replace the
458 * dso->long_name to a real pathname it found. In this case
459 * we prefer the vmlinux path like
460 * /lib/modules/3.16.4/build/vmlinux
461 *
462 * rather than build-id path (in debug directory).
463 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
464 */
465 symbol_conf.ignore_vmlinux_buildid = true;
466
467 /*
468 * If --buildid-all is given, it marks all DSO regardless of hits,
469 * so no need to process samples. But if timestamp_boundary is enabled,
470 * it still needs to walk on all samples to get the timestamps of
471 * first/last samples.
472 */
473 if (rec->buildid_all && !rec->timestamp_boundary)
474 rec->tool.sample = NULL;
475
476 return perf_session__process_events(session);
477}
478
479static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
480{
481 int err;
482 struct perf_tool *tool = data;
483 /*
484 *As for guest kernel when processing subcommand record&report,
485 *we arrange module mmap prior to guest kernel mmap and trigger
486 *a preload dso because default guest module symbols are loaded
487 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
488 *method is used to avoid symbol missing when the first addr is
489 *in module instead of in guest kernel.
490 */
491 err = perf_event__synthesize_modules(tool, process_synthesized_event,
492 machine);
493 if (err < 0)
494 pr_err("Couldn't record guest kernel [%d]'s reference"
495 " relocation symbol.\n", machine->pid);
496
497 /*
498 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
499 * have no _text sometimes.
500 */
501 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
502 machine);
503 if (err < 0)
504 pr_err("Couldn't record guest kernel [%d]'s reference"
505 " relocation symbol.\n", machine->pid);
506}
507
508static struct perf_event_header finished_round_event = {
509 .size = sizeof(struct perf_event_header),
510 .type = PERF_RECORD_FINISHED_ROUND,
511};
512
513static int record__mmap_read_evlist(struct record *rec, struct perf_evlist *evlist,
514 bool overwrite)
515{
516 u64 bytes_written = rec->bytes_written;
517 int i;
518 int rc = 0;
519 struct perf_mmap *maps;
520
521 if (!evlist)
522 return 0;
523
524 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
525 if (!maps)
526 return 0;
527
528 if (overwrite && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
529 return 0;
530
531 for (i = 0; i < evlist->nr_mmaps; i++) {
532 struct auxtrace_mmap *mm = &maps[i].auxtrace_mmap;
533
534 if (maps[i].base) {
535 if (perf_mmap__push(&maps[i], rec, record__pushfn) != 0) {
536 rc = -1;
537 goto out;
538 }
539 }
540
541 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
542 record__auxtrace_mmap_read(rec, mm) != 0) {
543 rc = -1;
544 goto out;
545 }
546 }
547
548 /*
549 * Mark the round finished in case we wrote
550 * at least one event.
551 */
552 if (bytes_written != rec->bytes_written)
553 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
554
555 if (overwrite)
556 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
557out:
558 return rc;
559}
560
561static int record__mmap_read_all(struct record *rec)
562{
563 int err;
564
565 err = record__mmap_read_evlist(rec, rec->evlist, false);
566 if (err)
567 return err;
568
569 return record__mmap_read_evlist(rec, rec->evlist, true);
570}
571
572static void record__init_features(struct record *rec)
573{
574 struct perf_session *session = rec->session;
575 int feat;
576
577 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
578 perf_header__set_feat(&session->header, feat);
579
580 if (rec->no_buildid)
581 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
582
583 if (!have_tracepoints(&rec->evlist->entries))
584 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
585
586 if (!rec->opts.branch_stack)
587 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
588
589 if (!rec->opts.full_auxtrace)
590 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
591
592 perf_header__clear_feat(&session->header, HEADER_STAT);
593}
594
595static void
596record__finish_output(struct record *rec)
597{
598 struct perf_data *data = &rec->data;
599 int fd = perf_data__fd(data);
600
601 if (data->is_pipe)
602 return;
603
604 rec->session->header.data_size += rec->bytes_written;
605 data->size = lseek(perf_data__fd(data), 0, SEEK_CUR);
606
607 if (!rec->no_buildid) {
608 process_buildids(rec);
609
610 if (rec->buildid_all)
611 dsos__hit_all(rec->session);
612 }
613 perf_session__write_header(rec->session, rec->evlist, fd, true);
614
615 return;
616}
617
618static int record__synthesize_workload(struct record *rec, bool tail)
619{
620 int err;
621 struct thread_map *thread_map;
622
623 if (rec->opts.tail_synthesize != tail)
624 return 0;
625
626 thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
627 if (thread_map == NULL)
628 return -1;
629
630 err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
631 process_synthesized_event,
632 &rec->session->machines.host,
633 rec->opts.sample_address,
634 rec->opts.proc_map_timeout);
635 thread_map__put(thread_map);
636 return err;
637}
638
639static int record__synthesize(struct record *rec, bool tail);
640
641static int
642record__switch_output(struct record *rec, bool at_exit)
643{
644 struct perf_data *data = &rec->data;
645 int fd, err;
646
647 /* Same Size: "2015122520103046"*/
648 char timestamp[] = "InvalidTimestamp";
649
650 record__synthesize(rec, true);
651 if (target__none(&rec->opts.target))
652 record__synthesize_workload(rec, true);
653
654 rec->samples = 0;
655 record__finish_output(rec);
656 err = fetch_current_timestamp(timestamp, sizeof(timestamp));
657 if (err) {
658 pr_err("Failed to get current timestamp\n");
659 return -EINVAL;
660 }
661
662 fd = perf_data__switch(data, timestamp,
663 rec->session->header.data_offset,
664 at_exit);
665 if (fd >= 0 && !at_exit) {
666 rec->bytes_written = 0;
667 rec->session->header.data_size = 0;
668 }
669
670 if (!quiet)
671 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
672 data->file.path, timestamp);
673
674 /* Output tracking events */
675 if (!at_exit) {
676 record__synthesize(rec, false);
677
678 /*
679 * In 'perf record --switch-output' without -a,
680 * record__synthesize() in record__switch_output() won't
681 * generate tracking events because there's no thread_map
682 * in evlist. Which causes newly created perf.data doesn't
683 * contain map and comm information.
684 * Create a fake thread_map and directly call
685 * perf_event__synthesize_thread_map() for those events.
686 */
687 if (target__none(&rec->opts.target))
688 record__synthesize_workload(rec, false);
689 }
690 return fd;
691}
692
693static volatile int workload_exec_errno;
694
695/*
696 * perf_evlist__prepare_workload will send a SIGUSR1
697 * if the fork fails, since we asked by setting its
698 * want_signal to true.
699 */
700static void workload_exec_failed_signal(int signo __maybe_unused,
701 siginfo_t *info,
702 void *ucontext __maybe_unused)
703{
704 workload_exec_errno = info->si_value.sival_int;
705 done = 1;
706 child_finished = 1;
707}
708
709static void snapshot_sig_handler(int sig);
710static void alarm_sig_handler(int sig);
711
712int __weak
713perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
714 struct perf_tool *tool __maybe_unused,
715 perf_event__handler_t process __maybe_unused,
716 struct machine *machine __maybe_unused)
717{
718 return 0;
719}
720
721static const struct perf_event_mmap_page *
722perf_evlist__pick_pc(struct perf_evlist *evlist)
723{
724 if (evlist) {
725 if (evlist->mmap && evlist->mmap[0].base)
726 return evlist->mmap[0].base;
727 if (evlist->overwrite_mmap && evlist->overwrite_mmap[0].base)
728 return evlist->overwrite_mmap[0].base;
729 }
730 return NULL;
731}
732
733static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
734{
735 const struct perf_event_mmap_page *pc;
736
737 pc = perf_evlist__pick_pc(rec->evlist);
738 if (pc)
739 return pc;
740 return NULL;
741}
742
743static int record__synthesize(struct record *rec, bool tail)
744{
745 struct perf_session *session = rec->session;
746 struct machine *machine = &session->machines.host;
747 struct perf_data *data = &rec->data;
748 struct record_opts *opts = &rec->opts;
749 struct perf_tool *tool = &rec->tool;
750 int fd = perf_data__fd(data);
751 int err = 0;
752
753 if (rec->opts.tail_synthesize != tail)
754 return 0;
755
756 if (data->is_pipe) {
757 /*
758 * We need to synthesize events first, because some
759 * features works on top of them (on report side).
760 */
761 err = perf_event__synthesize_attrs(tool, session,
762 process_synthesized_event);
763 if (err < 0) {
764 pr_err("Couldn't synthesize attrs.\n");
765 goto out;
766 }
767
768 err = perf_event__synthesize_features(tool, session, rec->evlist,
769 process_synthesized_event);
770 if (err < 0) {
771 pr_err("Couldn't synthesize features.\n");
772 return err;
773 }
774
775 if (have_tracepoints(&rec->evlist->entries)) {
776 /*
777 * FIXME err <= 0 here actually means that
778 * there were no tracepoints so its not really
779 * an error, just that we don't need to
780 * synthesize anything. We really have to
781 * return this more properly and also
782 * propagate errors that now are calling die()
783 */
784 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
785 process_synthesized_event);
786 if (err <= 0) {
787 pr_err("Couldn't record tracing data.\n");
788 goto out;
789 }
790 rec->bytes_written += err;
791 }
792 }
793
794 err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
795 process_synthesized_event, machine);
796 if (err)
797 goto out;
798
799 if (rec->opts.full_auxtrace) {
800 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
801 session, process_synthesized_event);
802 if (err)
803 goto out;
804 }
805
806 if (!perf_evlist__exclude_kernel(rec->evlist)) {
807 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
808 machine);
809 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
810 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
811 "Check /proc/kallsyms permission or run as root.\n");
812
813 err = perf_event__synthesize_modules(tool, process_synthesized_event,
814 machine);
815 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
816 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
817 "Check /proc/modules permission or run as root.\n");
818 }
819
820 if (perf_guest) {
821 machines__process_guests(&session->machines,
822 perf_event__synthesize_guest_os, tool);
823 }
824
825 err = perf_event__synthesize_extra_attr(&rec->tool,
826 rec->evlist,
827 process_synthesized_event,
828 data->is_pipe);
829 if (err)
830 goto out;
831
832 err = perf_event__synthesize_thread_map2(&rec->tool, rec->evlist->threads,
833 process_synthesized_event,
834 NULL);
835 if (err < 0) {
836 pr_err("Couldn't synthesize thread map.\n");
837 return err;
838 }
839
840 err = perf_event__synthesize_cpu_map(&rec->tool, rec->evlist->cpus,
841 process_synthesized_event, NULL);
842 if (err < 0) {
843 pr_err("Couldn't synthesize cpu map.\n");
844 return err;
845 }
846
847 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
848 process_synthesized_event, opts->sample_address,
849 opts->proc_map_timeout, 1);
850out:
851 return err;
852}
853
854static int __cmd_record(struct record *rec, int argc, const char **argv)
855{
856 int err;
857 int status = 0;
858 unsigned long waking = 0;
859 const bool forks = argc > 0;
860 struct perf_tool *tool = &rec->tool;
861 struct record_opts *opts = &rec->opts;
862 struct perf_data *data = &rec->data;
863 struct perf_session *session;
864 bool disabled = false, draining = false;
865 int fd;
866
867 atexit(record__sig_exit);
868 signal(SIGCHLD, sig_handler);
869 signal(SIGINT, sig_handler);
870 signal(SIGTERM, sig_handler);
871 signal(SIGSEGV, sigsegv_handler);
872
873 if (rec->opts.record_namespaces)
874 tool->namespace_events = true;
875
876 if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
877 signal(SIGUSR2, snapshot_sig_handler);
878 if (rec->opts.auxtrace_snapshot_mode)
879 trigger_on(&auxtrace_snapshot_trigger);
880 if (rec->switch_output.enabled)
881 trigger_on(&switch_output_trigger);
882 } else {
883 signal(SIGUSR2, SIG_IGN);
884 }
885
886 session = perf_session__new(data, false, tool);
887 if (session == NULL) {
888 pr_err("Perf session creation failed.\n");
889 return -1;
890 }
891
892 fd = perf_data__fd(data);
893 rec->session = session;
894
895 record__init_features(rec);
896
897 if (forks) {
898 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
899 argv, data->is_pipe,
900 workload_exec_failed_signal);
901 if (err < 0) {
902 pr_err("Couldn't run the workload!\n");
903 status = err;
904 goto out_delete_session;
905 }
906 }
907
908 /*
909 * If we have just single event and are sending data
910 * through pipe, we need to force the ids allocation,
911 * because we synthesize event name through the pipe
912 * and need the id for that.
913 */
914 if (data->is_pipe && rec->evlist->nr_entries == 1)
915 rec->opts.sample_id = true;
916
917 if (record__open(rec) != 0) {
918 err = -1;
919 goto out_child;
920 }
921
922 err = bpf__apply_obj_config();
923 if (err) {
924 char errbuf[BUFSIZ];
925
926 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
927 pr_err("ERROR: Apply config to BPF failed: %s\n",
928 errbuf);
929 goto out_child;
930 }
931
932 /*
933 * Normally perf_session__new would do this, but it doesn't have the
934 * evlist.
935 */
936 if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
937 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
938 rec->tool.ordered_events = false;
939 }
940
941 if (!rec->evlist->nr_groups)
942 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
943
944 if (data->is_pipe) {
945 err = perf_header__write_pipe(fd);
946 if (err < 0)
947 goto out_child;
948 } else {
949 err = perf_session__write_header(session, rec->evlist, fd, false);
950 if (err < 0)
951 goto out_child;
952 }
953
954 if (!rec->no_buildid
955 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
956 pr_err("Couldn't generate buildids. "
957 "Use --no-buildid to profile anyway.\n");
958 err = -1;
959 goto out_child;
960 }
961
962 err = record__synthesize(rec, false);
963 if (err < 0)
964 goto out_child;
965
966 if (rec->realtime_prio) {
967 struct sched_param param;
968
969 param.sched_priority = rec->realtime_prio;
970 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
971 pr_err("Could not set realtime priority.\n");
972 err = -1;
973 goto out_child;
974 }
975 }
976
977 /*
978 * When perf is starting the traced process, all the events
979 * (apart from group members) have enable_on_exec=1 set,
980 * so don't spoil it by prematurely enabling them.
981 */
982 if (!target__none(&opts->target) && !opts->initial_delay)
983 perf_evlist__enable(rec->evlist);
984
985 /*
986 * Let the child rip
987 */
988 if (forks) {
989 struct machine *machine = &session->machines.host;
990 union perf_event *event;
991 pid_t tgid;
992
993 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
994 if (event == NULL) {
995 err = -ENOMEM;
996 goto out_child;
997 }
998
999 /*
1000 * Some H/W events are generated before COMM event
1001 * which is emitted during exec(), so perf script
1002 * cannot see a correct process name for those events.
1003 * Synthesize COMM event to prevent it.
1004 */
1005 tgid = perf_event__synthesize_comm(tool, event,
1006 rec->evlist->workload.pid,
1007 process_synthesized_event,
1008 machine);
1009 free(event);
1010
1011 if (tgid == -1)
1012 goto out_child;
1013
1014 event = malloc(sizeof(event->namespaces) +
1015 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
1016 machine->id_hdr_size);
1017 if (event == NULL) {
1018 err = -ENOMEM;
1019 goto out_child;
1020 }
1021
1022 /*
1023 * Synthesize NAMESPACES event for the command specified.
1024 */
1025 perf_event__synthesize_namespaces(tool, event,
1026 rec->evlist->workload.pid,
1027 tgid, process_synthesized_event,
1028 machine);
1029 free(event);
1030
1031 perf_evlist__start_workload(rec->evlist);
1032 }
1033
1034 if (opts->initial_delay) {
1035 usleep(opts->initial_delay * USEC_PER_MSEC);
1036 perf_evlist__enable(rec->evlist);
1037 }
1038
1039 trigger_ready(&auxtrace_snapshot_trigger);
1040 trigger_ready(&switch_output_trigger);
1041 perf_hooks__invoke_record_start();
1042 for (;;) {
1043 unsigned long long hits = rec->samples;
1044
1045 /*
1046 * rec->evlist->bkw_mmap_state is possible to be
1047 * BKW_MMAP_EMPTY here: when done == true and
1048 * hits != rec->samples in previous round.
1049 *
1050 * perf_evlist__toggle_bkw_mmap ensure we never
1051 * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
1052 */
1053 if (trigger_is_hit(&switch_output_trigger) || done || draining)
1054 perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
1055
1056 if (record__mmap_read_all(rec) < 0) {
1057 trigger_error(&auxtrace_snapshot_trigger);
1058 trigger_error(&switch_output_trigger);
1059 err = -1;
1060 goto out_child;
1061 }
1062
1063 if (auxtrace_record__snapshot_started) {
1064 auxtrace_record__snapshot_started = 0;
1065 if (!trigger_is_error(&auxtrace_snapshot_trigger))
1066 record__read_auxtrace_snapshot(rec);
1067 if (trigger_is_error(&auxtrace_snapshot_trigger)) {
1068 pr_err("AUX area tracing snapshot failed\n");
1069 err = -1;
1070 goto out_child;
1071 }
1072 }
1073
1074 if (trigger_is_hit(&switch_output_trigger)) {
1075 /*
1076 * If switch_output_trigger is hit, the data in
1077 * overwritable ring buffer should have been collected,
1078 * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1079 *
1080 * If SIGUSR2 raise after or during record__mmap_read_all(),
1081 * record__mmap_read_all() didn't collect data from
1082 * overwritable ring buffer. Read again.
1083 */
1084 if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1085 continue;
1086 trigger_ready(&switch_output_trigger);
1087
1088 /*
1089 * Reenable events in overwrite ring buffer after
1090 * record__mmap_read_all(): we should have collected
1091 * data from it.
1092 */
1093 perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1094
1095 if (!quiet)
1096 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1097 waking);
1098 waking = 0;
1099 fd = record__switch_output(rec, false);
1100 if (fd < 0) {
1101 pr_err("Failed to switch to new file\n");
1102 trigger_error(&switch_output_trigger);
1103 err = fd;
1104 goto out_child;
1105 }
1106
1107 /* re-arm the alarm */
1108 if (rec->switch_output.time)
1109 alarm(rec->switch_output.time);
1110 }
1111
1112 if (hits == rec->samples) {
1113 if (done || draining)
1114 break;
1115 err = perf_evlist__poll(rec->evlist, -1);
1116 /*
1117 * Propagate error, only if there's any. Ignore positive
1118 * number of returned events and interrupt error.
1119 */
1120 if (err > 0 || (err < 0 && errno == EINTR))
1121 err = 0;
1122 waking++;
1123
1124 if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1125 draining = true;
1126 }
1127
1128 /*
1129 * When perf is starting the traced process, at the end events
1130 * die with the process and we wait for that. Thus no need to
1131 * disable events in this case.
1132 */
1133 if (done && !disabled && !target__none(&opts->target)) {
1134 trigger_off(&auxtrace_snapshot_trigger);
1135 perf_evlist__disable(rec->evlist);
1136 disabled = true;
1137 }
1138 }
1139 trigger_off(&auxtrace_snapshot_trigger);
1140 trigger_off(&switch_output_trigger);
1141
1142 if (forks && workload_exec_errno) {
1143 char msg[STRERR_BUFSIZE];
1144 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
1145 pr_err("Workload failed: %s\n", emsg);
1146 err = -1;
1147 goto out_child;
1148 }
1149
1150 if (!quiet)
1151 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1152
1153 if (target__none(&rec->opts.target))
1154 record__synthesize_workload(rec, true);
1155
1156out_child:
1157 if (forks) {
1158 int exit_status;
1159
1160 if (!child_finished)
1161 kill(rec->evlist->workload.pid, SIGTERM);
1162
1163 wait(&exit_status);
1164
1165 if (err < 0)
1166 status = err;
1167 else if (WIFEXITED(exit_status))
1168 status = WEXITSTATUS(exit_status);
1169 else if (WIFSIGNALED(exit_status))
1170 signr = WTERMSIG(exit_status);
1171 } else
1172 status = err;
1173
1174 record__synthesize(rec, true);
1175 /* this will be recalculated during process_buildids() */
1176 rec->samples = 0;
1177
1178 if (!err) {
1179 if (!rec->timestamp_filename) {
1180 record__finish_output(rec);
1181 } else {
1182 fd = record__switch_output(rec, true);
1183 if (fd < 0) {
1184 status = fd;
1185 goto out_delete_session;
1186 }
1187 }
1188 }
1189
1190 perf_hooks__invoke_record_end();
1191
1192 if (!err && !quiet) {
1193 char samples[128];
1194 const char *postfix = rec->timestamp_filename ?
1195 ".<timestamp>" : "";
1196
1197 if (rec->samples && !rec->opts.full_auxtrace)
1198 scnprintf(samples, sizeof(samples),
1199 " (%" PRIu64 " samples)", rec->samples);
1200 else
1201 samples[0] = '\0';
1202
1203 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1204 perf_data__size(data) / 1024.0 / 1024.0,
1205 data->file.path, postfix, samples);
1206 }
1207
1208out_delete_session:
1209 perf_session__delete(session);
1210 return status;
1211}
1212
1213static void callchain_debug(struct callchain_param *callchain)
1214{
1215 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
1216
1217 pr_debug("callchain: type %s\n", str[callchain->record_mode]);
1218
1219 if (callchain->record_mode == CALLCHAIN_DWARF)
1220 pr_debug("callchain: stack dump size %d\n",
1221 callchain->dump_size);
1222}
1223
1224int record_opts__parse_callchain(struct record_opts *record,
1225 struct callchain_param *callchain,
1226 const char *arg, bool unset)
1227{
1228 int ret;
1229 callchain->enabled = !unset;
1230
1231 /* --no-call-graph */
1232 if (unset) {
1233 callchain->record_mode = CALLCHAIN_NONE;
1234 pr_debug("callchain: disabled\n");
1235 return 0;
1236 }
1237
1238 ret = parse_callchain_record_opt(arg, callchain);
1239 if (!ret) {
1240 /* Enable data address sampling for DWARF unwind. */
1241 if (callchain->record_mode == CALLCHAIN_DWARF)
1242 record->sample_address = true;
1243 callchain_debug(callchain);
1244 }
1245
1246 return ret;
1247}
1248
1249int record_parse_callchain_opt(const struct option *opt,
1250 const char *arg,
1251 int unset)
1252{
1253 return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1254}
1255
1256int record_callchain_opt(const struct option *opt,
1257 const char *arg __maybe_unused,
1258 int unset __maybe_unused)
1259{
1260 struct callchain_param *callchain = opt->value;
1261
1262 callchain->enabled = true;
1263
1264 if (callchain->record_mode == CALLCHAIN_NONE)
1265 callchain->record_mode = CALLCHAIN_FP;
1266
1267 callchain_debug(callchain);
1268 return 0;
1269}
1270
1271static int perf_record_config(const char *var, const char *value, void *cb)
1272{
1273 struct record *rec = cb;
1274
1275 if (!strcmp(var, "record.build-id")) {
1276 if (!strcmp(value, "cache"))
1277 rec->no_buildid_cache = false;
1278 else if (!strcmp(value, "no-cache"))
1279 rec->no_buildid_cache = true;
1280 else if (!strcmp(value, "skip"))
1281 rec->no_buildid = true;
1282 else
1283 return -1;
1284 return 0;
1285 }
1286 if (!strcmp(var, "record.call-graph")) {
1287 var = "call-graph.record-mode";
1288 return perf_default_config(var, value, cb);
1289 }
1290
1291 return 0;
1292}
1293
1294struct clockid_map {
1295 const char *name;
1296 int clockid;
1297};
1298
1299#define CLOCKID_MAP(n, c) \
1300 { .name = n, .clockid = (c), }
1301
1302#define CLOCKID_END { .name = NULL, }
1303
1304
1305/*
1306 * Add the missing ones, we need to build on many distros...
1307 */
1308#ifndef CLOCK_MONOTONIC_RAW
1309#define CLOCK_MONOTONIC_RAW 4
1310#endif
1311#ifndef CLOCK_BOOTTIME
1312#define CLOCK_BOOTTIME 7
1313#endif
1314#ifndef CLOCK_TAI
1315#define CLOCK_TAI 11
1316#endif
1317
1318static const struct clockid_map clockids[] = {
1319 /* available for all events, NMI safe */
1320 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1321 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1322
1323 /* available for some events */
1324 CLOCKID_MAP("realtime", CLOCK_REALTIME),
1325 CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1326 CLOCKID_MAP("tai", CLOCK_TAI),
1327
1328 /* available for the lazy */
1329 CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1330 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1331 CLOCKID_MAP("real", CLOCK_REALTIME),
1332 CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1333
1334 CLOCKID_END,
1335};
1336
1337static int parse_clockid(const struct option *opt, const char *str, int unset)
1338{
1339 struct record_opts *opts = (struct record_opts *)opt->value;
1340 const struct clockid_map *cm;
1341 const char *ostr = str;
1342
1343 if (unset) {
1344 opts->use_clockid = 0;
1345 return 0;
1346 }
1347
1348 /* no arg passed */
1349 if (!str)
1350 return 0;
1351
1352 /* no setting it twice */
1353 if (opts->use_clockid)
1354 return -1;
1355
1356 opts->use_clockid = true;
1357
1358 /* if its a number, we're done */
1359 if (sscanf(str, "%d", &opts->clockid) == 1)
1360 return 0;
1361
1362 /* allow a "CLOCK_" prefix to the name */
1363 if (!strncasecmp(str, "CLOCK_", 6))
1364 str += 6;
1365
1366 for (cm = clockids; cm->name; cm++) {
1367 if (!strcasecmp(str, cm->name)) {
1368 opts->clockid = cm->clockid;
1369 return 0;
1370 }
1371 }
1372
1373 opts->use_clockid = false;
1374 ui__warning("unknown clockid %s, check man page\n", ostr);
1375 return -1;
1376}
1377
1378static int record__parse_mmap_pages(const struct option *opt,
1379 const char *str,
1380 int unset __maybe_unused)
1381{
1382 struct record_opts *opts = opt->value;
1383 char *s, *p;
1384 unsigned int mmap_pages;
1385 int ret;
1386
1387 if (!str)
1388 return -EINVAL;
1389
1390 s = strdup(str);
1391 if (!s)
1392 return -ENOMEM;
1393
1394 p = strchr(s, ',');
1395 if (p)
1396 *p = '\0';
1397
1398 if (*s) {
1399 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1400 if (ret)
1401 goto out_free;
1402 opts->mmap_pages = mmap_pages;
1403 }
1404
1405 if (!p) {
1406 ret = 0;
1407 goto out_free;
1408 }
1409
1410 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1411 if (ret)
1412 goto out_free;
1413
1414 opts->auxtrace_mmap_pages = mmap_pages;
1415
1416out_free:
1417 free(s);
1418 return ret;
1419}
1420
1421static void switch_output_size_warn(struct record *rec)
1422{
1423 u64 wakeup_size = perf_evlist__mmap_size(rec->opts.mmap_pages);
1424 struct switch_output *s = &rec->switch_output;
1425
1426 wakeup_size /= 2;
1427
1428 if (s->size < wakeup_size) {
1429 char buf[100];
1430
1431 unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
1432 pr_warning("WARNING: switch-output data size lower than "
1433 "wakeup kernel buffer size (%s) "
1434 "expect bigger perf.data sizes\n", buf);
1435 }
1436}
1437
1438static int switch_output_setup(struct record *rec)
1439{
1440 struct switch_output *s = &rec->switch_output;
1441 static struct parse_tag tags_size[] = {
1442 { .tag = 'B', .mult = 1 },
1443 { .tag = 'K', .mult = 1 << 10 },
1444 { .tag = 'M', .mult = 1 << 20 },
1445 { .tag = 'G', .mult = 1 << 30 },
1446 { .tag = 0 },
1447 };
1448 static struct parse_tag tags_time[] = {
1449 { .tag = 's', .mult = 1 },
1450 { .tag = 'm', .mult = 60 },
1451 { .tag = 'h', .mult = 60*60 },
1452 { .tag = 'd', .mult = 60*60*24 },
1453 { .tag = 0 },
1454 };
1455 unsigned long val;
1456
1457 if (!s->set)
1458 return 0;
1459
1460 if (!strcmp(s->str, "signal")) {
1461 s->signal = true;
1462 pr_debug("switch-output with SIGUSR2 signal\n");
1463 goto enabled;
1464 }
1465
1466 val = parse_tag_value(s->str, tags_size);
1467 if (val != (unsigned long) -1) {
1468 s->size = val;
1469 pr_debug("switch-output with %s size threshold\n", s->str);
1470 goto enabled;
1471 }
1472
1473 val = parse_tag_value(s->str, tags_time);
1474 if (val != (unsigned long) -1) {
1475 s->time = val;
1476 pr_debug("switch-output with %s time threshold (%lu seconds)\n",
1477 s->str, s->time);
1478 goto enabled;
1479 }
1480
1481 return -1;
1482
1483enabled:
1484 rec->timestamp_filename = true;
1485 s->enabled = true;
1486
1487 if (s->size && !rec->opts.no_buffering)
1488 switch_output_size_warn(rec);
1489
1490 return 0;
1491}
1492
1493static const char * const __record_usage[] = {
1494 "perf record [<options>] [<command>]",
1495 "perf record [<options>] -- <command> [<options>]",
1496 NULL
1497};
1498const char * const *record_usage = __record_usage;
1499
1500/*
1501 * XXX Ideally would be local to cmd_record() and passed to a record__new
1502 * because we need to have access to it in record__exit, that is called
1503 * after cmd_record() exits, but since record_options need to be accessible to
1504 * builtin-script, leave it here.
1505 *
1506 * At least we don't ouch it in all the other functions here directly.
1507 *
1508 * Just say no to tons of global variables, sigh.
1509 */
1510static struct record record = {
1511 .opts = {
1512 .sample_time = true,
1513 .mmap_pages = UINT_MAX,
1514 .user_freq = UINT_MAX,
1515 .user_interval = ULLONG_MAX,
1516 .freq = 4000,
1517 .target = {
1518 .uses_mmap = true,
1519 .default_per_cpu = true,
1520 },
1521 .proc_map_timeout = 500,
1522 },
1523 .tool = {
1524 .sample = process_sample_event,
1525 .fork = perf_event__process_fork,
1526 .exit = perf_event__process_exit,
1527 .comm = perf_event__process_comm,
1528 .namespaces = perf_event__process_namespaces,
1529 .mmap = perf_event__process_mmap,
1530 .mmap2 = perf_event__process_mmap2,
1531 .ordered_events = true,
1532 },
1533};
1534
1535const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1536 "\n\t\t\t\tDefault: fp";
1537
1538static bool dry_run;
1539
1540/*
1541 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1542 * with it and switch to use the library functions in perf_evlist that came
1543 * from builtin-record.c, i.e. use record_opts,
1544 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1545 * using pipes, etc.
1546 */
1547static struct option __record_options[] = {
1548 OPT_CALLBACK('e', "event", &record.evlist, "event",
1549 "event selector. use 'perf list' to list available events",
1550 parse_events_option),
1551 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1552 "event filter", parse_filter),
1553 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1554 NULL, "don't record events from perf itself",
1555 exclude_perf),
1556 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1557 "record events on existing process id"),
1558 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1559 "record events on existing thread id"),
1560 OPT_INTEGER('r', "realtime", &record.realtime_prio,
1561 "collect data with this RT SCHED_FIFO priority"),
1562 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1563 "collect data without buffering"),
1564 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1565 "collect raw sample records from all opened counters"),
1566 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1567 "system-wide collection from all CPUs"),
1568 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1569 "list of cpus to monitor"),
1570 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1571 OPT_STRING('o', "output", &record.data.file.path, "file",
1572 "output file name"),
1573 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1574 &record.opts.no_inherit_set,
1575 "child tasks do not inherit counters"),
1576 OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
1577 "synthesize non-sample events at the end of output"),
1578 OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
1579 OPT_BOOLEAN(0, "strict-freq", &record.opts.strict_freq,
1580 "Fail if the specified frequency can't be used"),
1581 OPT_CALLBACK('F', "freq", &record.opts, "freq or 'max'",
1582 "profile at this frequency",
1583 record__parse_freq),
1584 OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1585 "number of mmap data pages and AUX area tracing mmap pages",
1586 record__parse_mmap_pages),
1587 OPT_BOOLEAN(0, "group", &record.opts.group,
1588 "put the counters into a counter group"),
1589 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1590 NULL, "enables call-graph recording" ,
1591 &record_callchain_opt),
1592 OPT_CALLBACK(0, "call-graph", &record.opts,
1593 "record_mode[,record_size]", record_callchain_help,
1594 &record_parse_callchain_opt),
1595 OPT_INCR('v', "verbose", &verbose,
1596 "be more verbose (show counter open errors, etc)"),
1597 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1598 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1599 "per thread counts"),
1600 OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1601 OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
1602 "Record the sample physical addresses"),
1603 OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
1604 OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1605 &record.opts.sample_time_set,
1606 "Record the sample timestamps"),
1607 OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
1608 "Record the sample period"),
1609 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1610 "don't sample"),
1611 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1612 &record.no_buildid_cache_set,
1613 "do not update the buildid cache"),
1614 OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1615 &record.no_buildid_set,
1616 "do not collect buildids in perf.data"),
1617 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1618 "monitor event in cgroup name only",
1619 parse_cgroups),
1620 OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1621 "ms to wait before starting measurement after program start"),
1622 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1623 "user to profile"),
1624
1625 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1626 "branch any", "sample any taken branches",
1627 parse_branch_stack),
1628
1629 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1630 "branch filter mask", "branch stack filter modes",
1631 parse_branch_stack),
1632 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1633 "sample by weight (on special events only)"),
1634 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1635 "sample transaction flags (special events only)"),
1636 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1637 "use per-thread mmaps"),
1638 OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1639 "sample selected machine registers on interrupt,"
1640 " use -I ? to list register names", parse_regs),
1641 OPT_CALLBACK_OPTARG(0, "user-regs", &record.opts.sample_user_regs, NULL, "any register",
1642 "sample selected machine registers on interrupt,"
1643 " use -I ? to list register names", parse_regs),
1644 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1645 "Record running/enabled time of read (:S) events"),
1646 OPT_CALLBACK('k', "clockid", &record.opts,
1647 "clockid", "clockid to use for events, see clock_gettime()",
1648 parse_clockid),
1649 OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1650 "opts", "AUX area tracing Snapshot Mode", ""),
1651 OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1652 "per thread proc mmap processing timeout in ms"),
1653 OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
1654 "Record namespaces events"),
1655 OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1656 "Record context switch events"),
1657 OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1658 "Configure all used events to run in kernel space.",
1659 PARSE_OPT_EXCLUSIVE),
1660 OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1661 "Configure all used events to run in user space.",
1662 PARSE_OPT_EXCLUSIVE),
1663 OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1664 "clang binary to use for compiling BPF scriptlets"),
1665 OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1666 "options passed to clang when compiling BPF scriptlets"),
1667 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1668 "file", "vmlinux pathname"),
1669 OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1670 "Record build-id of all DSOs regardless of hits"),
1671 OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1672 "append timestamp to output filename"),
1673 OPT_BOOLEAN(0, "timestamp-boundary", &record.timestamp_boundary,
1674 "Record timestamp boundary (time of first/last samples)"),
1675 OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
1676 &record.switch_output.set, "signal,size,time",
1677 "Switch output when receive SIGUSR2 or cross size,time threshold",
1678 "signal"),
1679 OPT_BOOLEAN(0, "dry-run", &dry_run,
1680 "Parse options then exit"),
1681 OPT_END()
1682};
1683
1684struct option *record_options = __record_options;
1685
1686int cmd_record(int argc, const char **argv)
1687{
1688 int err;
1689 struct record *rec = &record;
1690 char errbuf[BUFSIZ];
1691
1692 setlocale(LC_ALL, "");
1693
1694#ifndef HAVE_LIBBPF_SUPPORT
1695# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1696 set_nobuild('\0', "clang-path", true);
1697 set_nobuild('\0', "clang-opt", true);
1698# undef set_nobuild
1699#endif
1700
1701#ifndef HAVE_BPF_PROLOGUE
1702# if !defined (HAVE_DWARF_SUPPORT)
1703# define REASON "NO_DWARF=1"
1704# elif !defined (HAVE_LIBBPF_SUPPORT)
1705# define REASON "NO_LIBBPF=1"
1706# else
1707# define REASON "this architecture doesn't support BPF prologue"
1708# endif
1709# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1710 set_nobuild('\0', "vmlinux", true);
1711# undef set_nobuild
1712# undef REASON
1713#endif
1714
1715 rec->evlist = perf_evlist__new();
1716 if (rec->evlist == NULL)
1717 return -ENOMEM;
1718
1719 err = perf_config(perf_record_config, rec);
1720 if (err)
1721 return err;
1722
1723 argc = parse_options(argc, argv, record_options, record_usage,
1724 PARSE_OPT_STOP_AT_NON_OPTION);
1725 if (quiet)
1726 perf_quiet_option();
1727
1728 /* Make system wide (-a) the default target. */
1729 if (!argc && target__none(&rec->opts.target))
1730 rec->opts.target.system_wide = true;
1731
1732 if (nr_cgroups && !rec->opts.target.system_wide) {
1733 usage_with_options_msg(record_usage, record_options,
1734 "cgroup monitoring only available in system-wide mode");
1735
1736 }
1737 if (rec->opts.record_switch_events &&
1738 !perf_can_record_switch_events()) {
1739 ui__error("kernel does not support recording context switch events\n");
1740 parse_options_usage(record_usage, record_options, "switch-events", 0);
1741 return -EINVAL;
1742 }
1743
1744 if (switch_output_setup(rec)) {
1745 parse_options_usage(record_usage, record_options, "switch-output", 0);
1746 return -EINVAL;
1747 }
1748
1749 if (rec->switch_output.time) {
1750 signal(SIGALRM, alarm_sig_handler);
1751 alarm(rec->switch_output.time);
1752 }
1753
1754 /*
1755 * Allow aliases to facilitate the lookup of symbols for address
1756 * filters. Refer to auxtrace_parse_filters().
1757 */
1758 symbol_conf.allow_aliases = true;
1759
1760 symbol__init(NULL);
1761
1762 err = record__auxtrace_init(rec);
1763 if (err)
1764 goto out;
1765
1766 if (dry_run)
1767 goto out;
1768
1769 err = bpf__setup_stdout(rec->evlist);
1770 if (err) {
1771 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1772 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1773 errbuf);
1774 goto out;
1775 }
1776
1777 err = -ENOMEM;
1778
1779 if (symbol_conf.kptr_restrict && !perf_evlist__exclude_kernel(rec->evlist))
1780 pr_warning(
1781"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1782"check /proc/sys/kernel/kptr_restrict.\n\n"
1783"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1784"file is not found in the buildid cache or in the vmlinux path.\n\n"
1785"Samples in kernel modules won't be resolved at all.\n\n"
1786"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1787"even with a suitable vmlinux or kallsyms file.\n\n");
1788
1789 if (rec->no_buildid_cache || rec->no_buildid) {
1790 disable_buildid_cache();
1791 } else if (rec->switch_output.enabled) {
1792 /*
1793 * In 'perf record --switch-output', disable buildid
1794 * generation by default to reduce data file switching
1795 * overhead. Still generate buildid if they are required
1796 * explicitly using
1797 *
1798 * perf record --switch-output --no-no-buildid \
1799 * --no-no-buildid-cache
1800 *
1801 * Following code equals to:
1802 *
1803 * if ((rec->no_buildid || !rec->no_buildid_set) &&
1804 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1805 * disable_buildid_cache();
1806 */
1807 bool disable = true;
1808
1809 if (rec->no_buildid_set && !rec->no_buildid)
1810 disable = false;
1811 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1812 disable = false;
1813 if (disable) {
1814 rec->no_buildid = true;
1815 rec->no_buildid_cache = true;
1816 disable_buildid_cache();
1817 }
1818 }
1819
1820 if (record.opts.overwrite)
1821 record.opts.tail_synthesize = true;
1822
1823 if (rec->evlist->nr_entries == 0 &&
1824 __perf_evlist__add_default(rec->evlist, !record.opts.no_samples) < 0) {
1825 pr_err("Not enough memory for event selector list\n");
1826 goto out;
1827 }
1828
1829 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1830 rec->opts.no_inherit = true;
1831
1832 err = target__validate(&rec->opts.target);
1833 if (err) {
1834 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1835 ui__warning("%s\n", errbuf);
1836 }
1837
1838 err = target__parse_uid(&rec->opts.target);
1839 if (err) {
1840 int saved_errno = errno;
1841
1842 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1843 ui__error("%s", errbuf);
1844
1845 err = -saved_errno;
1846 goto out;
1847 }
1848
1849 /* Enable ignoring missing threads when -u/-p option is defined. */
1850 rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
1851
1852 err = -ENOMEM;
1853 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1854 usage_with_options(record_usage, record_options);
1855
1856 err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1857 if (err)
1858 goto out;
1859
1860 /*
1861 * We take all buildids when the file contains
1862 * AUX area tracing data because we do not decode the
1863 * trace because it would take too long.
1864 */
1865 if (rec->opts.full_auxtrace)
1866 rec->buildid_all = true;
1867
1868 if (record_opts__config(&rec->opts)) {
1869 err = -EINVAL;
1870 goto out;
1871 }
1872
1873 err = __cmd_record(&record, argc, argv);
1874out:
1875 perf_evlist__delete(rec->evlist);
1876 symbol__exit();
1877 auxtrace_record__free(rec->itr);
1878 return err;
1879}
1880
1881static void snapshot_sig_handler(int sig __maybe_unused)
1882{
1883 struct record *rec = &record;
1884
1885 if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1886 trigger_hit(&auxtrace_snapshot_trigger);
1887 auxtrace_record__snapshot_started = 1;
1888 if (auxtrace_record__snapshot_start(record.itr))
1889 trigger_error(&auxtrace_snapshot_trigger);
1890 }
1891
1892 if (switch_output_signal(rec))
1893 trigger_hit(&switch_output_trigger);
1894}
1895
1896static void alarm_sig_handler(int sig __maybe_unused)
1897{
1898 struct record *rec = &record;
1899
1900 if (switch_output_time(rec))
1901 trigger_hit(&switch_output_trigger);
1902}
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#include "util/config.h"
17
18#include "util/callchain.h"
19#include "util/cgroup.h"
20#include "util/header.h"
21#include "util/event.h"
22#include "util/evlist.h"
23#include "util/evsel.h"
24#include "util/debug.h"
25#include "util/drv_configs.h"
26#include "util/session.h"
27#include "util/tool.h"
28#include "util/symbol.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/llvm-utils.h"
38#include "util/bpf-loader.h"
39#include "util/trigger.h"
40#include "util/perf-hooks.h"
41#include "asm/bug.h"
42
43#include <unistd.h>
44#include <sched.h>
45#include <sys/mman.h>
46#include <asm/bug.h>
47#include <linux/time64.h>
48
49struct record {
50 struct perf_tool tool;
51 struct record_opts opts;
52 u64 bytes_written;
53 struct perf_data_file file;
54 struct auxtrace_record *itr;
55 struct perf_evlist *evlist;
56 struct perf_session *session;
57 const char *progname;
58 int realtime_prio;
59 bool no_buildid;
60 bool no_buildid_set;
61 bool no_buildid_cache;
62 bool no_buildid_cache_set;
63 bool buildid_all;
64 bool timestamp_filename;
65 bool switch_output;
66 unsigned long long samples;
67};
68
69static int record__write(struct record *rec, void *bf, size_t size)
70{
71 if (perf_data_file__write(rec->session->file, bf, size) < 0) {
72 pr_err("failed to write perf data, error: %m\n");
73 return -1;
74 }
75
76 rec->bytes_written += size;
77 return 0;
78}
79
80static int process_synthesized_event(struct perf_tool *tool,
81 union perf_event *event,
82 struct perf_sample *sample __maybe_unused,
83 struct machine *machine __maybe_unused)
84{
85 struct record *rec = container_of(tool, struct record, tool);
86 return record__write(rec, event, event->header.size);
87}
88
89static int
90backward_rb_find_range(void *buf, int mask, u64 head, u64 *start, u64 *end)
91{
92 struct perf_event_header *pheader;
93 u64 evt_head = head;
94 int size = mask + 1;
95
96 pr_debug2("backward_rb_find_range: buf=%p, head=%"PRIx64"\n", buf, head);
97 pheader = (struct perf_event_header *)(buf + (head & mask));
98 *start = head;
99 while (true) {
100 if (evt_head - head >= (unsigned int)size) {
101 pr_debug("Finished reading backward ring buffer: rewind\n");
102 if (evt_head - head > (unsigned int)size)
103 evt_head -= pheader->size;
104 *end = evt_head;
105 return 0;
106 }
107
108 pheader = (struct perf_event_header *)(buf + (evt_head & mask));
109
110 if (pheader->size == 0) {
111 pr_debug("Finished reading backward ring buffer: get start\n");
112 *end = evt_head;
113 return 0;
114 }
115
116 evt_head += pheader->size;
117 pr_debug3("move evt_head: %"PRIx64"\n", evt_head);
118 }
119 WARN_ONCE(1, "Shouldn't get here\n");
120 return -1;
121}
122
123static int
124rb_find_range(void *data, int mask, u64 head, u64 old,
125 u64 *start, u64 *end, bool backward)
126{
127 if (!backward) {
128 *start = old;
129 *end = head;
130 return 0;
131 }
132
133 return backward_rb_find_range(data, mask, head, start, end);
134}
135
136static int
137record__mmap_read(struct record *rec, struct perf_mmap *md,
138 bool overwrite, bool backward)
139{
140 u64 head = perf_mmap__read_head(md);
141 u64 old = md->prev;
142 u64 end = head, start = old;
143 unsigned char *data = md->base + page_size;
144 unsigned long size;
145 void *buf;
146 int rc = 0;
147
148 if (rb_find_range(data, md->mask, head,
149 old, &start, &end, backward))
150 return -1;
151
152 if (start == end)
153 return 0;
154
155 rec->samples++;
156
157 size = end - start;
158 if (size > (unsigned long)(md->mask) + 1) {
159 WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
160
161 md->prev = head;
162 perf_mmap__consume(md, overwrite || backward);
163 return 0;
164 }
165
166 if ((start & md->mask) + size != (end & md->mask)) {
167 buf = &data[start & md->mask];
168 size = md->mask + 1 - (start & md->mask);
169 start += size;
170
171 if (record__write(rec, buf, size) < 0) {
172 rc = -1;
173 goto out;
174 }
175 }
176
177 buf = &data[start & md->mask];
178 size = end - start;
179 start += size;
180
181 if (record__write(rec, buf, size) < 0) {
182 rc = -1;
183 goto out;
184 }
185
186 md->prev = head;
187 perf_mmap__consume(md, overwrite || backward);
188out:
189 return rc;
190}
191
192static volatile int done;
193static volatile int signr = -1;
194static volatile int child_finished;
195
196static volatile int auxtrace_record__snapshot_started;
197static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
198static DEFINE_TRIGGER(switch_output_trigger);
199
200static void sig_handler(int sig)
201{
202 if (sig == SIGCHLD)
203 child_finished = 1;
204 else
205 signr = sig;
206
207 done = 1;
208}
209
210static void sigsegv_handler(int sig)
211{
212 perf_hooks__recover();
213 sighandler_dump_stack(sig);
214}
215
216static void record__sig_exit(void)
217{
218 if (signr == -1)
219 return;
220
221 signal(signr, SIG_DFL);
222 raise(signr);
223}
224
225#ifdef HAVE_AUXTRACE_SUPPORT
226
227static int record__process_auxtrace(struct perf_tool *tool,
228 union perf_event *event, void *data1,
229 size_t len1, void *data2, size_t len2)
230{
231 struct record *rec = container_of(tool, struct record, tool);
232 struct perf_data_file *file = &rec->file;
233 size_t padding;
234 u8 pad[8] = {0};
235
236 if (!perf_data_file__is_pipe(file)) {
237 off_t file_offset;
238 int fd = perf_data_file__fd(file);
239 int err;
240
241 file_offset = lseek(fd, 0, SEEK_CUR);
242 if (file_offset == -1)
243 return -1;
244 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
245 event, file_offset);
246 if (err)
247 return err;
248 }
249
250 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
251 padding = (len1 + len2) & 7;
252 if (padding)
253 padding = 8 - padding;
254
255 record__write(rec, event, event->header.size);
256 record__write(rec, data1, len1);
257 if (len2)
258 record__write(rec, data2, len2);
259 record__write(rec, &pad, padding);
260
261 return 0;
262}
263
264static int record__auxtrace_mmap_read(struct record *rec,
265 struct auxtrace_mmap *mm)
266{
267 int ret;
268
269 ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
270 record__process_auxtrace);
271 if (ret < 0)
272 return ret;
273
274 if (ret)
275 rec->samples++;
276
277 return 0;
278}
279
280static int record__auxtrace_mmap_read_snapshot(struct record *rec,
281 struct auxtrace_mmap *mm)
282{
283 int ret;
284
285 ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
286 record__process_auxtrace,
287 rec->opts.auxtrace_snapshot_size);
288 if (ret < 0)
289 return ret;
290
291 if (ret)
292 rec->samples++;
293
294 return 0;
295}
296
297static int record__auxtrace_read_snapshot_all(struct record *rec)
298{
299 int i;
300 int rc = 0;
301
302 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
303 struct auxtrace_mmap *mm =
304 &rec->evlist->mmap[i].auxtrace_mmap;
305
306 if (!mm->base)
307 continue;
308
309 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
310 rc = -1;
311 goto out;
312 }
313 }
314out:
315 return rc;
316}
317
318static void record__read_auxtrace_snapshot(struct record *rec)
319{
320 pr_debug("Recording AUX area tracing snapshot\n");
321 if (record__auxtrace_read_snapshot_all(rec) < 0) {
322 trigger_error(&auxtrace_snapshot_trigger);
323 } else {
324 if (auxtrace_record__snapshot_finish(rec->itr))
325 trigger_error(&auxtrace_snapshot_trigger);
326 else
327 trigger_ready(&auxtrace_snapshot_trigger);
328 }
329}
330
331#else
332
333static inline
334int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
335 struct auxtrace_mmap *mm __maybe_unused)
336{
337 return 0;
338}
339
340static inline
341void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
342{
343}
344
345static inline
346int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
347{
348 return 0;
349}
350
351#endif
352
353static int record__mmap_evlist(struct record *rec,
354 struct perf_evlist *evlist)
355{
356 struct record_opts *opts = &rec->opts;
357 char msg[512];
358
359 if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
360 opts->auxtrace_mmap_pages,
361 opts->auxtrace_snapshot_mode) < 0) {
362 if (errno == EPERM) {
363 pr_err("Permission error mapping pages.\n"
364 "Consider increasing "
365 "/proc/sys/kernel/perf_event_mlock_kb,\n"
366 "or try again with a smaller value of -m/--mmap_pages.\n"
367 "(current value: %u,%u)\n",
368 opts->mmap_pages, opts->auxtrace_mmap_pages);
369 return -errno;
370 } else {
371 pr_err("failed to mmap with %d (%s)\n", errno,
372 str_error_r(errno, msg, sizeof(msg)));
373 if (errno)
374 return -errno;
375 else
376 return -EINVAL;
377 }
378 }
379 return 0;
380}
381
382static int record__mmap(struct record *rec)
383{
384 return record__mmap_evlist(rec, rec->evlist);
385}
386
387static int record__open(struct record *rec)
388{
389 char msg[512];
390 struct perf_evsel *pos;
391 struct perf_evlist *evlist = rec->evlist;
392 struct perf_session *session = rec->session;
393 struct record_opts *opts = &rec->opts;
394 struct perf_evsel_config_term *err_term;
395 int rc = 0;
396
397 perf_evlist__config(evlist, opts, &callchain_param);
398
399 evlist__for_each_entry(evlist, pos) {
400try_again:
401 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
402 if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
403 if (verbose)
404 ui__warning("%s\n", msg);
405 goto try_again;
406 }
407
408 rc = -errno;
409 perf_evsel__open_strerror(pos, &opts->target,
410 errno, msg, sizeof(msg));
411 ui__error("%s\n", msg);
412 goto out;
413 }
414 }
415
416 if (perf_evlist__apply_filters(evlist, &pos)) {
417 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
418 pos->filter, perf_evsel__name(pos), errno,
419 str_error_r(errno, msg, sizeof(msg)));
420 rc = -1;
421 goto out;
422 }
423
424 if (perf_evlist__apply_drv_configs(evlist, &pos, &err_term)) {
425 error("failed to set config \"%s\" on event %s with %d (%s)\n",
426 err_term->val.drv_cfg, perf_evsel__name(pos), errno,
427 str_error_r(errno, msg, sizeof(msg)));
428 rc = -1;
429 goto out;
430 }
431
432 rc = record__mmap(rec);
433 if (rc)
434 goto out;
435
436 session->evlist = evlist;
437 perf_session__set_id_hdr_size(session);
438out:
439 return rc;
440}
441
442static int process_sample_event(struct perf_tool *tool,
443 union perf_event *event,
444 struct perf_sample *sample,
445 struct perf_evsel *evsel,
446 struct machine *machine)
447{
448 struct record *rec = container_of(tool, struct record, tool);
449
450 rec->samples++;
451
452 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
453}
454
455static int process_buildids(struct record *rec)
456{
457 struct perf_data_file *file = &rec->file;
458 struct perf_session *session = rec->session;
459
460 if (file->size == 0)
461 return 0;
462
463 /*
464 * During this process, it'll load kernel map and replace the
465 * dso->long_name to a real pathname it found. In this case
466 * we prefer the vmlinux path like
467 * /lib/modules/3.16.4/build/vmlinux
468 *
469 * rather than build-id path (in debug directory).
470 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
471 */
472 symbol_conf.ignore_vmlinux_buildid = true;
473
474 /*
475 * If --buildid-all is given, it marks all DSO regardless of hits,
476 * so no need to process samples.
477 */
478 if (rec->buildid_all)
479 rec->tool.sample = NULL;
480
481 return perf_session__process_events(session);
482}
483
484static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
485{
486 int err;
487 struct perf_tool *tool = data;
488 /*
489 *As for guest kernel when processing subcommand record&report,
490 *we arrange module mmap prior to guest kernel mmap and trigger
491 *a preload dso because default guest module symbols are loaded
492 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
493 *method is used to avoid symbol missing when the first addr is
494 *in module instead of in guest kernel.
495 */
496 err = perf_event__synthesize_modules(tool, process_synthesized_event,
497 machine);
498 if (err < 0)
499 pr_err("Couldn't record guest kernel [%d]'s reference"
500 " relocation symbol.\n", machine->pid);
501
502 /*
503 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
504 * have no _text sometimes.
505 */
506 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
507 machine);
508 if (err < 0)
509 pr_err("Couldn't record guest kernel [%d]'s reference"
510 " relocation symbol.\n", machine->pid);
511}
512
513static struct perf_event_header finished_round_event = {
514 .size = sizeof(struct perf_event_header),
515 .type = PERF_RECORD_FINISHED_ROUND,
516};
517
518static int record__mmap_read_evlist(struct record *rec, struct perf_evlist *evlist,
519 bool backward)
520{
521 u64 bytes_written = rec->bytes_written;
522 int i;
523 int rc = 0;
524 struct perf_mmap *maps;
525
526 if (!evlist)
527 return 0;
528
529 maps = backward ? evlist->backward_mmap : evlist->mmap;
530 if (!maps)
531 return 0;
532
533 if (backward && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
534 return 0;
535
536 for (i = 0; i < evlist->nr_mmaps; i++) {
537 struct auxtrace_mmap *mm = &maps[i].auxtrace_mmap;
538
539 if (maps[i].base) {
540 if (record__mmap_read(rec, &maps[i],
541 evlist->overwrite, backward) != 0) {
542 rc = -1;
543 goto out;
544 }
545 }
546
547 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
548 record__auxtrace_mmap_read(rec, mm) != 0) {
549 rc = -1;
550 goto out;
551 }
552 }
553
554 /*
555 * Mark the round finished in case we wrote
556 * at least one event.
557 */
558 if (bytes_written != rec->bytes_written)
559 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
560
561 if (backward)
562 perf_evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
563out:
564 return rc;
565}
566
567static int record__mmap_read_all(struct record *rec)
568{
569 int err;
570
571 err = record__mmap_read_evlist(rec, rec->evlist, false);
572 if (err)
573 return err;
574
575 return record__mmap_read_evlist(rec, rec->evlist, true);
576}
577
578static void record__init_features(struct record *rec)
579{
580 struct perf_session *session = rec->session;
581 int feat;
582
583 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
584 perf_header__set_feat(&session->header, feat);
585
586 if (rec->no_buildid)
587 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
588
589 if (!have_tracepoints(&rec->evlist->entries))
590 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
591
592 if (!rec->opts.branch_stack)
593 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
594
595 if (!rec->opts.full_auxtrace)
596 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
597
598 perf_header__clear_feat(&session->header, HEADER_STAT);
599}
600
601static void
602record__finish_output(struct record *rec)
603{
604 struct perf_data_file *file = &rec->file;
605 int fd = perf_data_file__fd(file);
606
607 if (file->is_pipe)
608 return;
609
610 rec->session->header.data_size += rec->bytes_written;
611 file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
612
613 if (!rec->no_buildid) {
614 process_buildids(rec);
615
616 if (rec->buildid_all)
617 dsos__hit_all(rec->session);
618 }
619 perf_session__write_header(rec->session, rec->evlist, fd, true);
620
621 return;
622}
623
624static int record__synthesize_workload(struct record *rec, bool tail)
625{
626 struct {
627 struct thread_map map;
628 struct thread_map_data map_data;
629 } thread_map;
630
631 if (rec->opts.tail_synthesize != tail)
632 return 0;
633
634 thread_map.map.nr = 1;
635 thread_map.map.map[0].pid = rec->evlist->workload.pid;
636 thread_map.map.map[0].comm = NULL;
637 return perf_event__synthesize_thread_map(&rec->tool, &thread_map.map,
638 process_synthesized_event,
639 &rec->session->machines.host,
640 rec->opts.sample_address,
641 rec->opts.proc_map_timeout);
642}
643
644static int record__synthesize(struct record *rec, bool tail);
645
646static int
647record__switch_output(struct record *rec, bool at_exit)
648{
649 struct perf_data_file *file = &rec->file;
650 int fd, err;
651
652 /* Same Size: "2015122520103046"*/
653 char timestamp[] = "InvalidTimestamp";
654
655 record__synthesize(rec, true);
656 if (target__none(&rec->opts.target))
657 record__synthesize_workload(rec, true);
658
659 rec->samples = 0;
660 record__finish_output(rec);
661 err = fetch_current_timestamp(timestamp, sizeof(timestamp));
662 if (err) {
663 pr_err("Failed to get current timestamp\n");
664 return -EINVAL;
665 }
666
667 fd = perf_data_file__switch(file, timestamp,
668 rec->session->header.data_offset,
669 at_exit);
670 if (fd >= 0 && !at_exit) {
671 rec->bytes_written = 0;
672 rec->session->header.data_size = 0;
673 }
674
675 if (!quiet)
676 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
677 file->path, timestamp);
678
679 /* Output tracking events */
680 if (!at_exit) {
681 record__synthesize(rec, false);
682
683 /*
684 * In 'perf record --switch-output' without -a,
685 * record__synthesize() in record__switch_output() won't
686 * generate tracking events because there's no thread_map
687 * in evlist. Which causes newly created perf.data doesn't
688 * contain map and comm information.
689 * Create a fake thread_map and directly call
690 * perf_event__synthesize_thread_map() for those events.
691 */
692 if (target__none(&rec->opts.target))
693 record__synthesize_workload(rec, false);
694 }
695 return fd;
696}
697
698static volatile int workload_exec_errno;
699
700/*
701 * perf_evlist__prepare_workload will send a SIGUSR1
702 * if the fork fails, since we asked by setting its
703 * want_signal to true.
704 */
705static void workload_exec_failed_signal(int signo __maybe_unused,
706 siginfo_t *info,
707 void *ucontext __maybe_unused)
708{
709 workload_exec_errno = info->si_value.sival_int;
710 done = 1;
711 child_finished = 1;
712}
713
714static void snapshot_sig_handler(int sig);
715
716int __weak
717perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
718 struct perf_tool *tool __maybe_unused,
719 perf_event__handler_t process __maybe_unused,
720 struct machine *machine __maybe_unused)
721{
722 return 0;
723}
724
725static const struct perf_event_mmap_page *
726perf_evlist__pick_pc(struct perf_evlist *evlist)
727{
728 if (evlist) {
729 if (evlist->mmap && evlist->mmap[0].base)
730 return evlist->mmap[0].base;
731 if (evlist->backward_mmap && evlist->backward_mmap[0].base)
732 return evlist->backward_mmap[0].base;
733 }
734 return NULL;
735}
736
737static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
738{
739 const struct perf_event_mmap_page *pc;
740
741 pc = perf_evlist__pick_pc(rec->evlist);
742 if (pc)
743 return pc;
744 return NULL;
745}
746
747static int record__synthesize(struct record *rec, bool tail)
748{
749 struct perf_session *session = rec->session;
750 struct machine *machine = &session->machines.host;
751 struct perf_data_file *file = &rec->file;
752 struct record_opts *opts = &rec->opts;
753 struct perf_tool *tool = &rec->tool;
754 int fd = perf_data_file__fd(file);
755 int err = 0;
756
757 if (rec->opts.tail_synthesize != tail)
758 return 0;
759
760 if (file->is_pipe) {
761 err = perf_event__synthesize_attrs(tool, session,
762 process_synthesized_event);
763 if (err < 0) {
764 pr_err("Couldn't synthesize attrs.\n");
765 goto out;
766 }
767
768 if (have_tracepoints(&rec->evlist->entries)) {
769 /*
770 * FIXME err <= 0 here actually means that
771 * there were no tracepoints so its not really
772 * an error, just that we don't need to
773 * synthesize anything. We really have to
774 * return this more properly and also
775 * propagate errors that now are calling die()
776 */
777 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
778 process_synthesized_event);
779 if (err <= 0) {
780 pr_err("Couldn't record tracing data.\n");
781 goto out;
782 }
783 rec->bytes_written += err;
784 }
785 }
786
787 err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
788 process_synthesized_event, machine);
789 if (err)
790 goto out;
791
792 if (rec->opts.full_auxtrace) {
793 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
794 session, process_synthesized_event);
795 if (err)
796 goto out;
797 }
798
799 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
800 machine);
801 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
802 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
803 "Check /proc/kallsyms permission or run as root.\n");
804
805 err = perf_event__synthesize_modules(tool, process_synthesized_event,
806 machine);
807 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
808 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
809 "Check /proc/modules permission or run as root.\n");
810
811 if (perf_guest) {
812 machines__process_guests(&session->machines,
813 perf_event__synthesize_guest_os, tool);
814 }
815
816 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
817 process_synthesized_event, opts->sample_address,
818 opts->proc_map_timeout);
819out:
820 return err;
821}
822
823static int __cmd_record(struct record *rec, int argc, const char **argv)
824{
825 int err;
826 int status = 0;
827 unsigned long waking = 0;
828 const bool forks = argc > 0;
829 struct machine *machine;
830 struct perf_tool *tool = &rec->tool;
831 struct record_opts *opts = &rec->opts;
832 struct perf_data_file *file = &rec->file;
833 struct perf_session *session;
834 bool disabled = false, draining = false;
835 int fd;
836
837 rec->progname = argv[0];
838
839 atexit(record__sig_exit);
840 signal(SIGCHLD, sig_handler);
841 signal(SIGINT, sig_handler);
842 signal(SIGTERM, sig_handler);
843 signal(SIGSEGV, sigsegv_handler);
844
845 if (rec->opts.auxtrace_snapshot_mode || rec->switch_output) {
846 signal(SIGUSR2, snapshot_sig_handler);
847 if (rec->opts.auxtrace_snapshot_mode)
848 trigger_on(&auxtrace_snapshot_trigger);
849 if (rec->switch_output)
850 trigger_on(&switch_output_trigger);
851 } else {
852 signal(SIGUSR2, SIG_IGN);
853 }
854
855 session = perf_session__new(file, false, tool);
856 if (session == NULL) {
857 pr_err("Perf session creation failed.\n");
858 return -1;
859 }
860
861 fd = perf_data_file__fd(file);
862 rec->session = session;
863
864 record__init_features(rec);
865
866 if (forks) {
867 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
868 argv, file->is_pipe,
869 workload_exec_failed_signal);
870 if (err < 0) {
871 pr_err("Couldn't run the workload!\n");
872 status = err;
873 goto out_delete_session;
874 }
875 }
876
877 if (record__open(rec) != 0) {
878 err = -1;
879 goto out_child;
880 }
881
882 err = bpf__apply_obj_config();
883 if (err) {
884 char errbuf[BUFSIZ];
885
886 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
887 pr_err("ERROR: Apply config to BPF failed: %s\n",
888 errbuf);
889 goto out_child;
890 }
891
892 /*
893 * Normally perf_session__new would do this, but it doesn't have the
894 * evlist.
895 */
896 if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
897 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
898 rec->tool.ordered_events = false;
899 }
900
901 if (!rec->evlist->nr_groups)
902 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
903
904 if (file->is_pipe) {
905 err = perf_header__write_pipe(fd);
906 if (err < 0)
907 goto out_child;
908 } else {
909 err = perf_session__write_header(session, rec->evlist, fd, false);
910 if (err < 0)
911 goto out_child;
912 }
913
914 if (!rec->no_buildid
915 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
916 pr_err("Couldn't generate buildids. "
917 "Use --no-buildid to profile anyway.\n");
918 err = -1;
919 goto out_child;
920 }
921
922 machine = &session->machines.host;
923
924 err = record__synthesize(rec, false);
925 if (err < 0)
926 goto out_child;
927
928 if (rec->realtime_prio) {
929 struct sched_param param;
930
931 param.sched_priority = rec->realtime_prio;
932 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
933 pr_err("Could not set realtime priority.\n");
934 err = -1;
935 goto out_child;
936 }
937 }
938
939 /*
940 * When perf is starting the traced process, all the events
941 * (apart from group members) have enable_on_exec=1 set,
942 * so don't spoil it by prematurely enabling them.
943 */
944 if (!target__none(&opts->target) && !opts->initial_delay)
945 perf_evlist__enable(rec->evlist);
946
947 /*
948 * Let the child rip
949 */
950 if (forks) {
951 union perf_event *event;
952
953 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
954 if (event == NULL) {
955 err = -ENOMEM;
956 goto out_child;
957 }
958
959 /*
960 * Some H/W events are generated before COMM event
961 * which is emitted during exec(), so perf script
962 * cannot see a correct process name for those events.
963 * Synthesize COMM event to prevent it.
964 */
965 perf_event__synthesize_comm(tool, event,
966 rec->evlist->workload.pid,
967 process_synthesized_event,
968 machine);
969 free(event);
970
971 perf_evlist__start_workload(rec->evlist);
972 }
973
974 if (opts->initial_delay) {
975 usleep(opts->initial_delay * USEC_PER_MSEC);
976 perf_evlist__enable(rec->evlist);
977 }
978
979 trigger_ready(&auxtrace_snapshot_trigger);
980 trigger_ready(&switch_output_trigger);
981 perf_hooks__invoke_record_start();
982 for (;;) {
983 unsigned long long hits = rec->samples;
984
985 /*
986 * rec->evlist->bkw_mmap_state is possible to be
987 * BKW_MMAP_EMPTY here: when done == true and
988 * hits != rec->samples in previous round.
989 *
990 * perf_evlist__toggle_bkw_mmap ensure we never
991 * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
992 */
993 if (trigger_is_hit(&switch_output_trigger) || done || draining)
994 perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
995
996 if (record__mmap_read_all(rec) < 0) {
997 trigger_error(&auxtrace_snapshot_trigger);
998 trigger_error(&switch_output_trigger);
999 err = -1;
1000 goto out_child;
1001 }
1002
1003 if (auxtrace_record__snapshot_started) {
1004 auxtrace_record__snapshot_started = 0;
1005 if (!trigger_is_error(&auxtrace_snapshot_trigger))
1006 record__read_auxtrace_snapshot(rec);
1007 if (trigger_is_error(&auxtrace_snapshot_trigger)) {
1008 pr_err("AUX area tracing snapshot failed\n");
1009 err = -1;
1010 goto out_child;
1011 }
1012 }
1013
1014 if (trigger_is_hit(&switch_output_trigger)) {
1015 /*
1016 * If switch_output_trigger is hit, the data in
1017 * overwritable ring buffer should have been collected,
1018 * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1019 *
1020 * If SIGUSR2 raise after or during record__mmap_read_all(),
1021 * record__mmap_read_all() didn't collect data from
1022 * overwritable ring buffer. Read again.
1023 */
1024 if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1025 continue;
1026 trigger_ready(&switch_output_trigger);
1027
1028 /*
1029 * Reenable events in overwrite ring buffer after
1030 * record__mmap_read_all(): we should have collected
1031 * data from it.
1032 */
1033 perf_evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1034
1035 if (!quiet)
1036 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1037 waking);
1038 waking = 0;
1039 fd = record__switch_output(rec, false);
1040 if (fd < 0) {
1041 pr_err("Failed to switch to new file\n");
1042 trigger_error(&switch_output_trigger);
1043 err = fd;
1044 goto out_child;
1045 }
1046 }
1047
1048 if (hits == rec->samples) {
1049 if (done || draining)
1050 break;
1051 err = perf_evlist__poll(rec->evlist, -1);
1052 /*
1053 * Propagate error, only if there's any. Ignore positive
1054 * number of returned events and interrupt error.
1055 */
1056 if (err > 0 || (err < 0 && errno == EINTR))
1057 err = 0;
1058 waking++;
1059
1060 if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1061 draining = true;
1062 }
1063
1064 /*
1065 * When perf is starting the traced process, at the end events
1066 * die with the process and we wait for that. Thus no need to
1067 * disable events in this case.
1068 */
1069 if (done && !disabled && !target__none(&opts->target)) {
1070 trigger_off(&auxtrace_snapshot_trigger);
1071 perf_evlist__disable(rec->evlist);
1072 disabled = true;
1073 }
1074 }
1075 trigger_off(&auxtrace_snapshot_trigger);
1076 trigger_off(&switch_output_trigger);
1077
1078 if (forks && workload_exec_errno) {
1079 char msg[STRERR_BUFSIZE];
1080 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
1081 pr_err("Workload failed: %s\n", emsg);
1082 err = -1;
1083 goto out_child;
1084 }
1085
1086 if (!quiet)
1087 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
1088
1089 if (target__none(&rec->opts.target))
1090 record__synthesize_workload(rec, true);
1091
1092out_child:
1093 if (forks) {
1094 int exit_status;
1095
1096 if (!child_finished)
1097 kill(rec->evlist->workload.pid, SIGTERM);
1098
1099 wait(&exit_status);
1100
1101 if (err < 0)
1102 status = err;
1103 else if (WIFEXITED(exit_status))
1104 status = WEXITSTATUS(exit_status);
1105 else if (WIFSIGNALED(exit_status))
1106 signr = WTERMSIG(exit_status);
1107 } else
1108 status = err;
1109
1110 record__synthesize(rec, true);
1111 /* this will be recalculated during process_buildids() */
1112 rec->samples = 0;
1113
1114 if (!err) {
1115 if (!rec->timestamp_filename) {
1116 record__finish_output(rec);
1117 } else {
1118 fd = record__switch_output(rec, true);
1119 if (fd < 0) {
1120 status = fd;
1121 goto out_delete_session;
1122 }
1123 }
1124 }
1125
1126 perf_hooks__invoke_record_end();
1127
1128 if (!err && !quiet) {
1129 char samples[128];
1130 const char *postfix = rec->timestamp_filename ?
1131 ".<timestamp>" : "";
1132
1133 if (rec->samples && !rec->opts.full_auxtrace)
1134 scnprintf(samples, sizeof(samples),
1135 " (%" PRIu64 " samples)", rec->samples);
1136 else
1137 samples[0] = '\0';
1138
1139 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1140 perf_data_file__size(file) / 1024.0 / 1024.0,
1141 file->path, postfix, samples);
1142 }
1143
1144out_delete_session:
1145 perf_session__delete(session);
1146 return status;
1147}
1148
1149static void callchain_debug(struct callchain_param *callchain)
1150{
1151 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
1152
1153 pr_debug("callchain: type %s\n", str[callchain->record_mode]);
1154
1155 if (callchain->record_mode == CALLCHAIN_DWARF)
1156 pr_debug("callchain: stack dump size %d\n",
1157 callchain->dump_size);
1158}
1159
1160int record_opts__parse_callchain(struct record_opts *record,
1161 struct callchain_param *callchain,
1162 const char *arg, bool unset)
1163{
1164 int ret;
1165 callchain->enabled = !unset;
1166
1167 /* --no-call-graph */
1168 if (unset) {
1169 callchain->record_mode = CALLCHAIN_NONE;
1170 pr_debug("callchain: disabled\n");
1171 return 0;
1172 }
1173
1174 ret = parse_callchain_record_opt(arg, callchain);
1175 if (!ret) {
1176 /* Enable data address sampling for DWARF unwind. */
1177 if (callchain->record_mode == CALLCHAIN_DWARF)
1178 record->sample_address = true;
1179 callchain_debug(callchain);
1180 }
1181
1182 return ret;
1183}
1184
1185int record_parse_callchain_opt(const struct option *opt,
1186 const char *arg,
1187 int unset)
1188{
1189 return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1190}
1191
1192int record_callchain_opt(const struct option *opt,
1193 const char *arg __maybe_unused,
1194 int unset __maybe_unused)
1195{
1196 struct callchain_param *callchain = opt->value;
1197
1198 callchain->enabled = true;
1199
1200 if (callchain->record_mode == CALLCHAIN_NONE)
1201 callchain->record_mode = CALLCHAIN_FP;
1202
1203 callchain_debug(callchain);
1204 return 0;
1205}
1206
1207static int perf_record_config(const char *var, const char *value, void *cb)
1208{
1209 struct record *rec = cb;
1210
1211 if (!strcmp(var, "record.build-id")) {
1212 if (!strcmp(value, "cache"))
1213 rec->no_buildid_cache = false;
1214 else if (!strcmp(value, "no-cache"))
1215 rec->no_buildid_cache = true;
1216 else if (!strcmp(value, "skip"))
1217 rec->no_buildid = true;
1218 else
1219 return -1;
1220 return 0;
1221 }
1222 if (!strcmp(var, "record.call-graph"))
1223 var = "call-graph.record-mode"; /* fall-through */
1224
1225 return perf_default_config(var, value, cb);
1226}
1227
1228struct clockid_map {
1229 const char *name;
1230 int clockid;
1231};
1232
1233#define CLOCKID_MAP(n, c) \
1234 { .name = n, .clockid = (c), }
1235
1236#define CLOCKID_END { .name = NULL, }
1237
1238
1239/*
1240 * Add the missing ones, we need to build on many distros...
1241 */
1242#ifndef CLOCK_MONOTONIC_RAW
1243#define CLOCK_MONOTONIC_RAW 4
1244#endif
1245#ifndef CLOCK_BOOTTIME
1246#define CLOCK_BOOTTIME 7
1247#endif
1248#ifndef CLOCK_TAI
1249#define CLOCK_TAI 11
1250#endif
1251
1252static const struct clockid_map clockids[] = {
1253 /* available for all events, NMI safe */
1254 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1255 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1256
1257 /* available for some events */
1258 CLOCKID_MAP("realtime", CLOCK_REALTIME),
1259 CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1260 CLOCKID_MAP("tai", CLOCK_TAI),
1261
1262 /* available for the lazy */
1263 CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1264 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1265 CLOCKID_MAP("real", CLOCK_REALTIME),
1266 CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1267
1268 CLOCKID_END,
1269};
1270
1271static int parse_clockid(const struct option *opt, const char *str, int unset)
1272{
1273 struct record_opts *opts = (struct record_opts *)opt->value;
1274 const struct clockid_map *cm;
1275 const char *ostr = str;
1276
1277 if (unset) {
1278 opts->use_clockid = 0;
1279 return 0;
1280 }
1281
1282 /* no arg passed */
1283 if (!str)
1284 return 0;
1285
1286 /* no setting it twice */
1287 if (opts->use_clockid)
1288 return -1;
1289
1290 opts->use_clockid = true;
1291
1292 /* if its a number, we're done */
1293 if (sscanf(str, "%d", &opts->clockid) == 1)
1294 return 0;
1295
1296 /* allow a "CLOCK_" prefix to the name */
1297 if (!strncasecmp(str, "CLOCK_", 6))
1298 str += 6;
1299
1300 for (cm = clockids; cm->name; cm++) {
1301 if (!strcasecmp(str, cm->name)) {
1302 opts->clockid = cm->clockid;
1303 return 0;
1304 }
1305 }
1306
1307 opts->use_clockid = false;
1308 ui__warning("unknown clockid %s, check man page\n", ostr);
1309 return -1;
1310}
1311
1312static int record__parse_mmap_pages(const struct option *opt,
1313 const char *str,
1314 int unset __maybe_unused)
1315{
1316 struct record_opts *opts = opt->value;
1317 char *s, *p;
1318 unsigned int mmap_pages;
1319 int ret;
1320
1321 if (!str)
1322 return -EINVAL;
1323
1324 s = strdup(str);
1325 if (!s)
1326 return -ENOMEM;
1327
1328 p = strchr(s, ',');
1329 if (p)
1330 *p = '\0';
1331
1332 if (*s) {
1333 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1334 if (ret)
1335 goto out_free;
1336 opts->mmap_pages = mmap_pages;
1337 }
1338
1339 if (!p) {
1340 ret = 0;
1341 goto out_free;
1342 }
1343
1344 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1345 if (ret)
1346 goto out_free;
1347
1348 opts->auxtrace_mmap_pages = mmap_pages;
1349
1350out_free:
1351 free(s);
1352 return ret;
1353}
1354
1355static const char * const __record_usage[] = {
1356 "perf record [<options>] [<command>]",
1357 "perf record [<options>] -- <command> [<options>]",
1358 NULL
1359};
1360const char * const *record_usage = __record_usage;
1361
1362/*
1363 * XXX Ideally would be local to cmd_record() and passed to a record__new
1364 * because we need to have access to it in record__exit, that is called
1365 * after cmd_record() exits, but since record_options need to be accessible to
1366 * builtin-script, leave it here.
1367 *
1368 * At least we don't ouch it in all the other functions here directly.
1369 *
1370 * Just say no to tons of global variables, sigh.
1371 */
1372static struct record record = {
1373 .opts = {
1374 .sample_time = true,
1375 .mmap_pages = UINT_MAX,
1376 .user_freq = UINT_MAX,
1377 .user_interval = ULLONG_MAX,
1378 .freq = 4000,
1379 .target = {
1380 .uses_mmap = true,
1381 .default_per_cpu = true,
1382 },
1383 .proc_map_timeout = 500,
1384 },
1385 .tool = {
1386 .sample = process_sample_event,
1387 .fork = perf_event__process_fork,
1388 .exit = perf_event__process_exit,
1389 .comm = perf_event__process_comm,
1390 .mmap = perf_event__process_mmap,
1391 .mmap2 = perf_event__process_mmap2,
1392 .ordered_events = true,
1393 },
1394};
1395
1396const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1397 "\n\t\t\t\tDefault: fp";
1398
1399static bool dry_run;
1400
1401/*
1402 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1403 * with it and switch to use the library functions in perf_evlist that came
1404 * from builtin-record.c, i.e. use record_opts,
1405 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1406 * using pipes, etc.
1407 */
1408static struct option __record_options[] = {
1409 OPT_CALLBACK('e', "event", &record.evlist, "event",
1410 "event selector. use 'perf list' to list available events",
1411 parse_events_option),
1412 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1413 "event filter", parse_filter),
1414 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1415 NULL, "don't record events from perf itself",
1416 exclude_perf),
1417 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1418 "record events on existing process id"),
1419 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1420 "record events on existing thread id"),
1421 OPT_INTEGER('r', "realtime", &record.realtime_prio,
1422 "collect data with this RT SCHED_FIFO priority"),
1423 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1424 "collect data without buffering"),
1425 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1426 "collect raw sample records from all opened counters"),
1427 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1428 "system-wide collection from all CPUs"),
1429 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1430 "list of cpus to monitor"),
1431 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1432 OPT_STRING('o', "output", &record.file.path, "file",
1433 "output file name"),
1434 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1435 &record.opts.no_inherit_set,
1436 "child tasks do not inherit counters"),
1437 OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
1438 "synthesize non-sample events at the end of output"),
1439 OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
1440 OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1441 OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1442 "number of mmap data pages and AUX area tracing mmap pages",
1443 record__parse_mmap_pages),
1444 OPT_BOOLEAN(0, "group", &record.opts.group,
1445 "put the counters into a counter group"),
1446 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1447 NULL, "enables call-graph recording" ,
1448 &record_callchain_opt),
1449 OPT_CALLBACK(0, "call-graph", &record.opts,
1450 "record_mode[,record_size]", record_callchain_help,
1451 &record_parse_callchain_opt),
1452 OPT_INCR('v', "verbose", &verbose,
1453 "be more verbose (show counter open errors, etc)"),
1454 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1455 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1456 "per thread counts"),
1457 OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1458 OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
1459 OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1460 &record.opts.sample_time_set,
1461 "Record the sample timestamps"),
1462 OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1463 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1464 "don't sample"),
1465 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1466 &record.no_buildid_cache_set,
1467 "do not update the buildid cache"),
1468 OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1469 &record.no_buildid_set,
1470 "do not collect buildids in perf.data"),
1471 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1472 "monitor event in cgroup name only",
1473 parse_cgroups),
1474 OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1475 "ms to wait before starting measurement after program start"),
1476 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1477 "user to profile"),
1478
1479 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1480 "branch any", "sample any taken branches",
1481 parse_branch_stack),
1482
1483 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1484 "branch filter mask", "branch stack filter modes",
1485 parse_branch_stack),
1486 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1487 "sample by weight (on special events only)"),
1488 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1489 "sample transaction flags (special events only)"),
1490 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1491 "use per-thread mmaps"),
1492 OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1493 "sample selected machine registers on interrupt,"
1494 " use -I ? to list register names", parse_regs),
1495 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1496 "Record running/enabled time of read (:S) events"),
1497 OPT_CALLBACK('k', "clockid", &record.opts,
1498 "clockid", "clockid to use for events, see clock_gettime()",
1499 parse_clockid),
1500 OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1501 "opts", "AUX area tracing Snapshot Mode", ""),
1502 OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1503 "per thread proc mmap processing timeout in ms"),
1504 OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1505 "Record context switch events"),
1506 OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1507 "Configure all used events to run in kernel space.",
1508 PARSE_OPT_EXCLUSIVE),
1509 OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1510 "Configure all used events to run in user space.",
1511 PARSE_OPT_EXCLUSIVE),
1512 OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1513 "clang binary to use for compiling BPF scriptlets"),
1514 OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1515 "options passed to clang when compiling BPF scriptlets"),
1516 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1517 "file", "vmlinux pathname"),
1518 OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1519 "Record build-id of all DSOs regardless of hits"),
1520 OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1521 "append timestamp to output filename"),
1522 OPT_BOOLEAN(0, "switch-output", &record.switch_output,
1523 "Switch output when receive SIGUSR2"),
1524 OPT_BOOLEAN(0, "dry-run", &dry_run,
1525 "Parse options then exit"),
1526 OPT_END()
1527};
1528
1529struct option *record_options = __record_options;
1530
1531int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1532{
1533 int err;
1534 struct record *rec = &record;
1535 char errbuf[BUFSIZ];
1536
1537#ifndef HAVE_LIBBPF_SUPPORT
1538# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1539 set_nobuild('\0', "clang-path", true);
1540 set_nobuild('\0', "clang-opt", true);
1541# undef set_nobuild
1542#endif
1543
1544#ifndef HAVE_BPF_PROLOGUE
1545# if !defined (HAVE_DWARF_SUPPORT)
1546# define REASON "NO_DWARF=1"
1547# elif !defined (HAVE_LIBBPF_SUPPORT)
1548# define REASON "NO_LIBBPF=1"
1549# else
1550# define REASON "this architecture doesn't support BPF prologue"
1551# endif
1552# define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1553 set_nobuild('\0', "vmlinux", true);
1554# undef set_nobuild
1555# undef REASON
1556#endif
1557
1558 rec->evlist = perf_evlist__new();
1559 if (rec->evlist == NULL)
1560 return -ENOMEM;
1561
1562 perf_config(perf_record_config, rec);
1563
1564 argc = parse_options(argc, argv, record_options, record_usage,
1565 PARSE_OPT_STOP_AT_NON_OPTION);
1566 if (!argc && target__none(&rec->opts.target))
1567 usage_with_options(record_usage, record_options);
1568
1569 if (nr_cgroups && !rec->opts.target.system_wide) {
1570 usage_with_options_msg(record_usage, record_options,
1571 "cgroup monitoring only available in system-wide mode");
1572
1573 }
1574 if (rec->opts.record_switch_events &&
1575 !perf_can_record_switch_events()) {
1576 ui__error("kernel does not support recording context switch events\n");
1577 parse_options_usage(record_usage, record_options, "switch-events", 0);
1578 return -EINVAL;
1579 }
1580
1581 if (rec->switch_output)
1582 rec->timestamp_filename = true;
1583
1584 if (!rec->itr) {
1585 rec->itr = auxtrace_record__init(rec->evlist, &err);
1586 if (err)
1587 goto out;
1588 }
1589
1590 err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1591 rec->opts.auxtrace_snapshot_opts);
1592 if (err)
1593 goto out;
1594
1595 /*
1596 * Allow aliases to facilitate the lookup of symbols for address
1597 * filters. Refer to auxtrace_parse_filters().
1598 */
1599 symbol_conf.allow_aliases = true;
1600
1601 symbol__init(NULL);
1602
1603 err = auxtrace_parse_filters(rec->evlist);
1604 if (err)
1605 goto out;
1606
1607 if (dry_run)
1608 goto out;
1609
1610 err = bpf__setup_stdout(rec->evlist);
1611 if (err) {
1612 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1613 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1614 errbuf);
1615 goto out;
1616 }
1617
1618 err = -ENOMEM;
1619
1620 if (symbol_conf.kptr_restrict)
1621 pr_warning(
1622"WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1623"check /proc/sys/kernel/kptr_restrict.\n\n"
1624"Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1625"file is not found in the buildid cache or in the vmlinux path.\n\n"
1626"Samples in kernel modules won't be resolved at all.\n\n"
1627"If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1628"even with a suitable vmlinux or kallsyms file.\n\n");
1629
1630 if (rec->no_buildid_cache || rec->no_buildid) {
1631 disable_buildid_cache();
1632 } else if (rec->switch_output) {
1633 /*
1634 * In 'perf record --switch-output', disable buildid
1635 * generation by default to reduce data file switching
1636 * overhead. Still generate buildid if they are required
1637 * explicitly using
1638 *
1639 * perf record --switch-output --no-no-buildid \
1640 * --no-no-buildid-cache
1641 *
1642 * Following code equals to:
1643 *
1644 * if ((rec->no_buildid || !rec->no_buildid_set) &&
1645 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1646 * disable_buildid_cache();
1647 */
1648 bool disable = true;
1649
1650 if (rec->no_buildid_set && !rec->no_buildid)
1651 disable = false;
1652 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1653 disable = false;
1654 if (disable) {
1655 rec->no_buildid = true;
1656 rec->no_buildid_cache = true;
1657 disable_buildid_cache();
1658 }
1659 }
1660
1661 if (record.opts.overwrite)
1662 record.opts.tail_synthesize = true;
1663
1664 if (rec->evlist->nr_entries == 0 &&
1665 perf_evlist__add_default(rec->evlist) < 0) {
1666 pr_err("Not enough memory for event selector list\n");
1667 goto out;
1668 }
1669
1670 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1671 rec->opts.no_inherit = true;
1672
1673 err = target__validate(&rec->opts.target);
1674 if (err) {
1675 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1676 ui__warning("%s", errbuf);
1677 }
1678
1679 err = target__parse_uid(&rec->opts.target);
1680 if (err) {
1681 int saved_errno = errno;
1682
1683 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1684 ui__error("%s", errbuf);
1685
1686 err = -saved_errno;
1687 goto out;
1688 }
1689
1690 /* Enable ignoring missing threads when -u option is defined. */
1691 rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX;
1692
1693 err = -ENOMEM;
1694 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1695 usage_with_options(record_usage, record_options);
1696
1697 err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1698 if (err)
1699 goto out;
1700
1701 /*
1702 * We take all buildids when the file contains
1703 * AUX area tracing data because we do not decode the
1704 * trace because it would take too long.
1705 */
1706 if (rec->opts.full_auxtrace)
1707 rec->buildid_all = true;
1708
1709 if (record_opts__config(&rec->opts)) {
1710 err = -EINVAL;
1711 goto out;
1712 }
1713
1714 err = __cmd_record(&record, argc, argv);
1715out:
1716 perf_evlist__delete(rec->evlist);
1717 symbol__exit();
1718 auxtrace_record__free(rec->itr);
1719 return err;
1720}
1721
1722static void snapshot_sig_handler(int sig __maybe_unused)
1723{
1724 if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1725 trigger_hit(&auxtrace_snapshot_trigger);
1726 auxtrace_record__snapshot_started = 1;
1727 if (auxtrace_record__snapshot_start(record.itr))
1728 trigger_error(&auxtrace_snapshot_trigger);
1729 }
1730
1731 if (trigger_is_ready(&switch_output_trigger))
1732 trigger_hit(&switch_output_trigger);
1733}