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.
12 #include "util/build-id.h"
13 #include "util/util.h"
14 #include <subcmd/parse-options.h>
15 #include "util/parse-events.h"
17 #include "util/callchain.h"
18 #include "util/cgroup.h"
19 #include "util/header.h"
20 #include "util/event.h"
21 #include "util/evlist.h"
22 #include "util/evsel.h"
23 #include "util/debug.h"
24 #include "util/session.h"
25 #include "util/tool.h"
26 #include "util/symbol.h"
27 #include "util/cpumap.h"
28 #include "util/thread_map.h"
29 #include "util/data.h"
30 #include "util/perf_regs.h"
31 #include "util/auxtrace.h"
33 #include "util/parse-branch-options.h"
34 #include "util/parse-regs-options.h"
35 #include "util/llvm-utils.h"
36 #include "util/bpf-loader.h"
37 #include "util/trigger.h"
47 struct perf_tool tool;
48 struct record_opts opts;
50 struct perf_data_file file;
51 struct auxtrace_record *itr;
52 struct perf_evlist *evlist;
53 struct perf_session *session;
58 bool no_buildid_cache;
59 bool no_buildid_cache_set;
61 bool timestamp_filename;
63 unsigned long long samples;
66 static int record__write(struct record *rec, void *bf, size_t size)
68 if (perf_data_file__write(rec->session->file, bf, size) < 0) {
69 pr_err("failed to write perf data, error: %m\n");
73 rec->bytes_written += size;
77 static int process_synthesized_event(struct perf_tool *tool,
78 union perf_event *event,
79 struct perf_sample *sample __maybe_unused,
80 struct machine *machine __maybe_unused)
82 struct record *rec = container_of(tool, struct record, tool);
83 return record__write(rec, event, event->header.size);
86 static int record__mmap_read(struct record *rec, int idx)
88 struct perf_mmap *md = &rec->evlist->mmap[idx];
89 u64 head = perf_mmap__read_head(md);
91 unsigned char *data = md->base + page_size;
102 if (size > (unsigned long)(md->mask) + 1) {
103 WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
106 perf_evlist__mmap_consume(rec->evlist, idx);
110 if ((old & md->mask) + size != (head & md->mask)) {
111 buf = &data[old & md->mask];
112 size = md->mask + 1 - (old & md->mask);
115 if (record__write(rec, buf, size) < 0) {
121 buf = &data[old & md->mask];
125 if (record__write(rec, buf, size) < 0) {
131 perf_evlist__mmap_consume(rec->evlist, idx);
136 static volatile int done;
137 static volatile int signr = -1;
138 static volatile int child_finished;
140 static volatile int auxtrace_record__snapshot_started;
141 static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
142 static DEFINE_TRIGGER(switch_output_trigger);
144 static void sig_handler(int sig)
154 static void record__sig_exit(void)
159 signal(signr, SIG_DFL);
163 #ifdef HAVE_AUXTRACE_SUPPORT
165 static int record__process_auxtrace(struct perf_tool *tool,
166 union perf_event *event, void *data1,
167 size_t len1, void *data2, size_t len2)
169 struct record *rec = container_of(tool, struct record, tool);
170 struct perf_data_file *file = &rec->file;
174 if (!perf_data_file__is_pipe(file)) {
176 int fd = perf_data_file__fd(file);
179 file_offset = lseek(fd, 0, SEEK_CUR);
180 if (file_offset == -1)
182 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
188 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
189 padding = (len1 + len2) & 7;
191 padding = 8 - padding;
193 record__write(rec, event, event->header.size);
194 record__write(rec, data1, len1);
196 record__write(rec, data2, len2);
197 record__write(rec, &pad, padding);
202 static int record__auxtrace_mmap_read(struct record *rec,
203 struct auxtrace_mmap *mm)
207 ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
208 record__process_auxtrace);
218 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
219 struct auxtrace_mmap *mm)
223 ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
224 record__process_auxtrace,
225 rec->opts.auxtrace_snapshot_size);
235 static int record__auxtrace_read_snapshot_all(struct record *rec)
240 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
241 struct auxtrace_mmap *mm =
242 &rec->evlist->mmap[i].auxtrace_mmap;
247 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
256 static void record__read_auxtrace_snapshot(struct record *rec)
258 pr_debug("Recording AUX area tracing snapshot\n");
259 if (record__auxtrace_read_snapshot_all(rec) < 0) {
260 trigger_error(&auxtrace_snapshot_trigger);
262 if (auxtrace_record__snapshot_finish(rec->itr))
263 trigger_error(&auxtrace_snapshot_trigger);
265 trigger_ready(&auxtrace_snapshot_trigger);
272 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
273 struct auxtrace_mmap *mm __maybe_unused)
279 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
284 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
291 static int record__open(struct record *rec)
294 struct perf_evsel *pos;
295 struct perf_evlist *evlist = rec->evlist;
296 struct perf_session *session = rec->session;
297 struct record_opts *opts = &rec->opts;
300 perf_evlist__config(evlist, opts, &callchain_param);
302 evlist__for_each(evlist, pos) {
304 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
305 if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
307 ui__warning("%s\n", msg);
312 perf_evsel__open_strerror(pos, &opts->target,
313 errno, msg, sizeof(msg));
314 ui__error("%s\n", msg);
319 if (perf_evlist__apply_filters(evlist, &pos)) {
320 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
321 pos->filter, perf_evsel__name(pos), errno,
322 strerror_r(errno, msg, sizeof(msg)));
327 if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
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);
339 pr_err("failed to mmap with %d (%s)\n", errno,
340 strerror_r(errno, msg, sizeof(msg)));
349 session->evlist = evlist;
350 perf_session__set_id_hdr_size(session);
355 static int process_sample_event(struct perf_tool *tool,
356 union perf_event *event,
357 struct perf_sample *sample,
358 struct perf_evsel *evsel,
359 struct machine *machine)
361 struct record *rec = container_of(tool, struct record, tool);
365 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
368 static int process_buildids(struct record *rec)
370 struct perf_data_file *file = &rec->file;
371 struct perf_session *session = rec->session;
377 * During this process, it'll load kernel map and replace the
378 * dso->long_name to a real pathname it found. In this case
379 * we prefer the vmlinux path like
380 * /lib/modules/3.16.4/build/vmlinux
382 * rather than build-id path (in debug directory).
383 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
385 symbol_conf.ignore_vmlinux_buildid = true;
388 * If --buildid-all is given, it marks all DSO regardless of hits,
389 * so no need to process samples.
391 if (rec->buildid_all)
392 rec->tool.sample = NULL;
394 return perf_session__process_events(session);
397 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
400 struct perf_tool *tool = data;
402 *As for guest kernel when processing subcommand record&report,
403 *we arrange module mmap prior to guest kernel mmap and trigger
404 *a preload dso because default guest module symbols are loaded
405 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
406 *method is used to avoid symbol missing when the first addr is
407 *in module instead of in guest kernel.
409 err = perf_event__synthesize_modules(tool, process_synthesized_event,
412 pr_err("Couldn't record guest kernel [%d]'s reference"
413 " relocation symbol.\n", machine->pid);
416 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
417 * have no _text sometimes.
419 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
422 pr_err("Couldn't record guest kernel [%d]'s reference"
423 " relocation symbol.\n", machine->pid);
426 static struct perf_event_header finished_round_event = {
427 .size = sizeof(struct perf_event_header),
428 .type = PERF_RECORD_FINISHED_ROUND,
431 static int record__mmap_read_all(struct record *rec)
433 u64 bytes_written = rec->bytes_written;
437 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
438 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
440 if (rec->evlist->mmap[i].base) {
441 if (record__mmap_read(rec, i) != 0) {
447 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
448 record__auxtrace_mmap_read(rec, mm) != 0) {
455 * Mark the round finished in case we wrote
456 * at least one event.
458 if (bytes_written != rec->bytes_written)
459 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
465 static void record__init_features(struct record *rec)
467 struct perf_session *session = rec->session;
470 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
471 perf_header__set_feat(&session->header, feat);
474 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
476 if (!have_tracepoints(&rec->evlist->entries))
477 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
479 if (!rec->opts.branch_stack)
480 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
482 if (!rec->opts.full_auxtrace)
483 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
485 perf_header__clear_feat(&session->header, HEADER_STAT);
489 record__finish_output(struct record *rec)
491 struct perf_data_file *file = &rec->file;
492 int fd = perf_data_file__fd(file);
497 rec->session->header.data_size += rec->bytes_written;
498 file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
500 if (!rec->no_buildid) {
501 process_buildids(rec);
503 if (rec->buildid_all)
504 dsos__hit_all(rec->session);
506 perf_session__write_header(rec->session, rec->evlist, fd, true);
511 static int record__synthesize_workload(struct record *rec)
514 struct thread_map map;
515 struct thread_map_data map_data;
518 thread_map.map.nr = 1;
519 thread_map.map.map[0].pid = rec->evlist->workload.pid;
520 thread_map.map.map[0].comm = NULL;
521 return perf_event__synthesize_thread_map(&rec->tool, &thread_map.map,
522 process_synthesized_event,
523 &rec->session->machines.host,
524 rec->opts.sample_address,
525 rec->opts.proc_map_timeout);
528 static int record__synthesize(struct record *rec);
531 record__switch_output(struct record *rec, bool at_exit)
533 struct perf_data_file *file = &rec->file;
536 /* Same Size: "2015122520103046"*/
537 char timestamp[] = "InvalidTimestamp";
540 record__finish_output(rec);
541 err = fetch_current_timestamp(timestamp, sizeof(timestamp));
543 pr_err("Failed to get current timestamp\n");
547 fd = perf_data_file__switch(file, timestamp,
548 rec->session->header.data_offset,
550 if (fd >= 0 && !at_exit) {
551 rec->bytes_written = 0;
552 rec->session->header.data_size = 0;
556 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
557 file->path, timestamp);
559 /* Output tracking events */
561 record__synthesize(rec);
564 * In 'perf record --switch-output' without -a,
565 * record__synthesize() in record__switch_output() won't
566 * generate tracking events because there's no thread_map
567 * in evlist. Which causes newly created perf.data doesn't
568 * contain map and comm information.
569 * Create a fake thread_map and directly call
570 * perf_event__synthesize_thread_map() for those events.
572 if (target__none(&rec->opts.target))
573 record__synthesize_workload(rec);
578 static volatile int workload_exec_errno;
581 * perf_evlist__prepare_workload will send a SIGUSR1
582 * if the fork fails, since we asked by setting its
583 * want_signal to true.
585 static void workload_exec_failed_signal(int signo __maybe_unused,
587 void *ucontext __maybe_unused)
589 workload_exec_errno = info->si_value.sival_int;
594 static void snapshot_sig_handler(int sig);
597 perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
598 struct perf_tool *tool __maybe_unused,
599 perf_event__handler_t process __maybe_unused,
600 struct machine *machine __maybe_unused)
605 static int record__synthesize(struct record *rec)
607 struct perf_session *session = rec->session;
608 struct machine *machine = &session->machines.host;
609 struct perf_data_file *file = &rec->file;
610 struct record_opts *opts = &rec->opts;
611 struct perf_tool *tool = &rec->tool;
612 int fd = perf_data_file__fd(file);
616 err = perf_event__synthesize_attrs(tool, session,
617 process_synthesized_event);
619 pr_err("Couldn't synthesize attrs.\n");
623 if (have_tracepoints(&rec->evlist->entries)) {
625 * FIXME err <= 0 here actually means that
626 * there were no tracepoints so its not really
627 * an error, just that we don't need to
628 * synthesize anything. We really have to
629 * return this more properly and also
630 * propagate errors that now are calling die()
632 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
633 process_synthesized_event);
635 pr_err("Couldn't record tracing data.\n");
638 rec->bytes_written += err;
642 err = perf_event__synth_time_conv(rec->evlist->mmap[0].base, tool,
643 process_synthesized_event, machine);
647 if (rec->opts.full_auxtrace) {
648 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
649 session, process_synthesized_event);
654 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
656 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
657 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
658 "Check /proc/kallsyms permission or run as root.\n");
660 err = perf_event__synthesize_modules(tool, process_synthesized_event,
662 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
663 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
664 "Check /proc/modules permission or run as root.\n");
667 machines__process_guests(&session->machines,
668 perf_event__synthesize_guest_os, tool);
671 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
672 process_synthesized_event, opts->sample_address,
673 opts->proc_map_timeout);
678 static int __cmd_record(struct record *rec, int argc, const char **argv)
682 unsigned long waking = 0;
683 const bool forks = argc > 0;
684 struct machine *machine;
685 struct perf_tool *tool = &rec->tool;
686 struct record_opts *opts = &rec->opts;
687 struct perf_data_file *file = &rec->file;
688 struct perf_session *session;
689 bool disabled = false, draining = false;
692 rec->progname = argv[0];
694 atexit(record__sig_exit);
695 signal(SIGCHLD, sig_handler);
696 signal(SIGINT, sig_handler);
697 signal(SIGTERM, sig_handler);
699 if (rec->opts.auxtrace_snapshot_mode || rec->switch_output) {
700 signal(SIGUSR2, snapshot_sig_handler);
701 if (rec->opts.auxtrace_snapshot_mode)
702 trigger_on(&auxtrace_snapshot_trigger);
703 if (rec->switch_output)
704 trigger_on(&switch_output_trigger);
706 signal(SIGUSR2, SIG_IGN);
709 session = perf_session__new(file, false, tool);
710 if (session == NULL) {
711 pr_err("Perf session creation failed.\n");
715 fd = perf_data_file__fd(file);
716 rec->session = session;
718 record__init_features(rec);
721 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
723 workload_exec_failed_signal);
725 pr_err("Couldn't run the workload!\n");
727 goto out_delete_session;
731 if (record__open(rec) != 0) {
736 err = bpf__apply_obj_config();
740 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
741 pr_err("ERROR: Apply config to BPF failed: %s\n",
747 * Normally perf_session__new would do this, but it doesn't have the
750 if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
751 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
752 rec->tool.ordered_events = false;
755 if (!rec->evlist->nr_groups)
756 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
759 err = perf_header__write_pipe(fd);
763 err = perf_session__write_header(session, rec->evlist, fd, false);
769 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
770 pr_err("Couldn't generate buildids. "
771 "Use --no-buildid to profile anyway.\n");
776 machine = &session->machines.host;
778 err = record__synthesize(rec);
782 if (rec->realtime_prio) {
783 struct sched_param param;
785 param.sched_priority = rec->realtime_prio;
786 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
787 pr_err("Could not set realtime priority.\n");
794 * When perf is starting the traced process, all the events
795 * (apart from group members) have enable_on_exec=1 set,
796 * so don't spoil it by prematurely enabling them.
798 if (!target__none(&opts->target) && !opts->initial_delay)
799 perf_evlist__enable(rec->evlist);
805 union perf_event *event;
807 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
814 * Some H/W events are generated before COMM event
815 * which is emitted during exec(), so perf script
816 * cannot see a correct process name for those events.
817 * Synthesize COMM event to prevent it.
819 perf_event__synthesize_comm(tool, event,
820 rec->evlist->workload.pid,
821 process_synthesized_event,
825 perf_evlist__start_workload(rec->evlist);
828 if (opts->initial_delay) {
829 usleep(opts->initial_delay * 1000);
830 perf_evlist__enable(rec->evlist);
833 trigger_ready(&auxtrace_snapshot_trigger);
834 trigger_ready(&switch_output_trigger);
836 unsigned long long hits = rec->samples;
838 if (record__mmap_read_all(rec) < 0) {
839 trigger_error(&auxtrace_snapshot_trigger);
840 trigger_error(&switch_output_trigger);
845 if (auxtrace_record__snapshot_started) {
846 auxtrace_record__snapshot_started = 0;
847 if (!trigger_is_error(&auxtrace_snapshot_trigger))
848 record__read_auxtrace_snapshot(rec);
849 if (trigger_is_error(&auxtrace_snapshot_trigger)) {
850 pr_err("AUX area tracing snapshot failed\n");
856 if (trigger_is_hit(&switch_output_trigger)) {
857 trigger_ready(&switch_output_trigger);
860 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
863 fd = record__switch_output(rec, false);
865 pr_err("Failed to switch to new file\n");
866 trigger_error(&switch_output_trigger);
872 if (hits == rec->samples) {
873 if (done || draining)
875 err = perf_evlist__poll(rec->evlist, -1);
877 * Propagate error, only if there's any. Ignore positive
878 * number of returned events and interrupt error.
880 if (err > 0 || (err < 0 && errno == EINTR))
884 if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
889 * When perf is starting the traced process, at the end events
890 * die with the process and we wait for that. Thus no need to
891 * disable events in this case.
893 if (done && !disabled && !target__none(&opts->target)) {
894 trigger_off(&auxtrace_snapshot_trigger);
895 perf_evlist__disable(rec->evlist);
899 trigger_off(&auxtrace_snapshot_trigger);
900 trigger_off(&switch_output_trigger);
902 if (forks && workload_exec_errno) {
903 char msg[STRERR_BUFSIZE];
904 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
905 pr_err("Workload failed: %s\n", emsg);
911 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
918 kill(rec->evlist->workload.pid, SIGTERM);
924 else if (WIFEXITED(exit_status))
925 status = WEXITSTATUS(exit_status);
926 else if (WIFSIGNALED(exit_status))
927 signr = WTERMSIG(exit_status);
931 /* this will be recalculated during process_buildids() */
935 if (!rec->timestamp_filename) {
936 record__finish_output(rec);
938 fd = record__switch_output(rec, true);
941 goto out_delete_session;
946 if (!err && !quiet) {
948 const char *postfix = rec->timestamp_filename ?
951 if (rec->samples && !rec->opts.full_auxtrace)
952 scnprintf(samples, sizeof(samples),
953 " (%" PRIu64 " samples)", rec->samples);
957 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
958 perf_data_file__size(file) / 1024.0 / 1024.0,
959 file->path, postfix, samples);
963 perf_session__delete(session);
967 static void callchain_debug(struct callchain_param *callchain)
969 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
971 pr_debug("callchain: type %s\n", str[callchain->record_mode]);
973 if (callchain->record_mode == CALLCHAIN_DWARF)
974 pr_debug("callchain: stack dump size %d\n",
975 callchain->dump_size);
978 int record_opts__parse_callchain(struct record_opts *record,
979 struct callchain_param *callchain,
980 const char *arg, bool unset)
983 callchain->enabled = !unset;
985 /* --no-call-graph */
987 callchain->record_mode = CALLCHAIN_NONE;
988 pr_debug("callchain: disabled\n");
992 ret = parse_callchain_record_opt(arg, callchain);
994 /* Enable data address sampling for DWARF unwind. */
995 if (callchain->record_mode == CALLCHAIN_DWARF)
996 record->sample_address = true;
997 callchain_debug(callchain);
1003 int record_parse_callchain_opt(const struct option *opt,
1007 return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1010 int record_callchain_opt(const struct option *opt,
1011 const char *arg __maybe_unused,
1012 int unset __maybe_unused)
1014 struct callchain_param *callchain = opt->value;
1016 callchain->enabled = true;
1018 if (callchain->record_mode == CALLCHAIN_NONE)
1019 callchain->record_mode = CALLCHAIN_FP;
1021 callchain_debug(callchain);
1025 static int perf_record_config(const char *var, const char *value, void *cb)
1027 struct record *rec = cb;
1029 if (!strcmp(var, "record.build-id")) {
1030 if (!strcmp(value, "cache"))
1031 rec->no_buildid_cache = false;
1032 else if (!strcmp(value, "no-cache"))
1033 rec->no_buildid_cache = true;
1034 else if (!strcmp(value, "skip"))
1035 rec->no_buildid = true;
1040 if (!strcmp(var, "record.call-graph"))
1041 var = "call-graph.record-mode"; /* fall-through */
1043 return perf_default_config(var, value, cb);
1046 struct clockid_map {
1051 #define CLOCKID_MAP(n, c) \
1052 { .name = n, .clockid = (c), }
1054 #define CLOCKID_END { .name = NULL, }
1058 * Add the missing ones, we need to build on many distros...
1060 #ifndef CLOCK_MONOTONIC_RAW
1061 #define CLOCK_MONOTONIC_RAW 4
1063 #ifndef CLOCK_BOOTTIME
1064 #define CLOCK_BOOTTIME 7
1067 #define CLOCK_TAI 11
1070 static const struct clockid_map clockids[] = {
1071 /* available for all events, NMI safe */
1072 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1073 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1075 /* available for some events */
1076 CLOCKID_MAP("realtime", CLOCK_REALTIME),
1077 CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1078 CLOCKID_MAP("tai", CLOCK_TAI),
1080 /* available for the lazy */
1081 CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1082 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1083 CLOCKID_MAP("real", CLOCK_REALTIME),
1084 CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1089 static int parse_clockid(const struct option *opt, const char *str, int unset)
1091 struct record_opts *opts = (struct record_opts *)opt->value;
1092 const struct clockid_map *cm;
1093 const char *ostr = str;
1096 opts->use_clockid = 0;
1104 /* no setting it twice */
1105 if (opts->use_clockid)
1108 opts->use_clockid = true;
1110 /* if its a number, we're done */
1111 if (sscanf(str, "%d", &opts->clockid) == 1)
1114 /* allow a "CLOCK_" prefix to the name */
1115 if (!strncasecmp(str, "CLOCK_", 6))
1118 for (cm = clockids; cm->name; cm++) {
1119 if (!strcasecmp(str, cm->name)) {
1120 opts->clockid = cm->clockid;
1125 opts->use_clockid = false;
1126 ui__warning("unknown clockid %s, check man page\n", ostr);
1130 static int record__parse_mmap_pages(const struct option *opt,
1132 int unset __maybe_unused)
1134 struct record_opts *opts = opt->value;
1136 unsigned int mmap_pages;
1151 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1154 opts->mmap_pages = mmap_pages;
1162 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1166 opts->auxtrace_mmap_pages = mmap_pages;
1173 static const char * const __record_usage[] = {
1174 "perf record [<options>] [<command>]",
1175 "perf record [<options>] -- <command> [<options>]",
1178 const char * const *record_usage = __record_usage;
1181 * XXX Ideally would be local to cmd_record() and passed to a record__new
1182 * because we need to have access to it in record__exit, that is called
1183 * after cmd_record() exits, but since record_options need to be accessible to
1184 * builtin-script, leave it here.
1186 * At least we don't ouch it in all the other functions here directly.
1188 * Just say no to tons of global variables, sigh.
1190 static struct record record = {
1192 .sample_time = true,
1193 .mmap_pages = UINT_MAX,
1194 .user_freq = UINT_MAX,
1195 .user_interval = ULLONG_MAX,
1199 .default_per_cpu = true,
1201 .proc_map_timeout = 500,
1204 .sample = process_sample_event,
1205 .fork = perf_event__process_fork,
1206 .exit = perf_event__process_exit,
1207 .comm = perf_event__process_comm,
1208 .mmap = perf_event__process_mmap,
1209 .mmap2 = perf_event__process_mmap2,
1210 .ordered_events = true,
1214 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1215 "\n\t\t\t\tDefault: fp";
1218 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1219 * with it and switch to use the library functions in perf_evlist that came
1220 * from builtin-record.c, i.e. use record_opts,
1221 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1224 struct option __record_options[] = {
1225 OPT_CALLBACK('e', "event", &record.evlist, "event",
1226 "event selector. use 'perf list' to list available events",
1227 parse_events_option),
1228 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1229 "event filter", parse_filter),
1230 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1231 NULL, "don't record events from perf itself",
1233 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1234 "record events on existing process id"),
1235 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1236 "record events on existing thread id"),
1237 OPT_INTEGER('r', "realtime", &record.realtime_prio,
1238 "collect data with this RT SCHED_FIFO priority"),
1239 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1240 "collect data without buffering"),
1241 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1242 "collect raw sample records from all opened counters"),
1243 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1244 "system-wide collection from all CPUs"),
1245 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1246 "list of cpus to monitor"),
1247 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1248 OPT_STRING('o', "output", &record.file.path, "file",
1249 "output file name"),
1250 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1251 &record.opts.no_inherit_set,
1252 "child tasks do not inherit counters"),
1253 OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1254 OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1255 "number of mmap data pages and AUX area tracing mmap pages",
1256 record__parse_mmap_pages),
1257 OPT_BOOLEAN(0, "group", &record.opts.group,
1258 "put the counters into a counter group"),
1259 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1260 NULL, "enables call-graph recording" ,
1261 &record_callchain_opt),
1262 OPT_CALLBACK(0, "call-graph", &record.opts,
1263 "record_mode[,record_size]", record_callchain_help,
1264 &record_parse_callchain_opt),
1265 OPT_INCR('v', "verbose", &verbose,
1266 "be more verbose (show counter open errors, etc)"),
1267 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1268 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1269 "per thread counts"),
1270 OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1271 OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1272 &record.opts.sample_time_set,
1273 "Record the sample timestamps"),
1274 OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1275 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1277 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1278 &record.no_buildid_cache_set,
1279 "do not update the buildid cache"),
1280 OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1281 &record.no_buildid_set,
1282 "do not collect buildids in perf.data"),
1283 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1284 "monitor event in cgroup name only",
1286 OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1287 "ms to wait before starting measurement after program start"),
1288 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1291 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1292 "branch any", "sample any taken branches",
1293 parse_branch_stack),
1295 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1296 "branch filter mask", "branch stack filter modes",
1297 parse_branch_stack),
1298 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1299 "sample by weight (on special events only)"),
1300 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1301 "sample transaction flags (special events only)"),
1302 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1303 "use per-thread mmaps"),
1304 OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1305 "sample selected machine registers on interrupt,"
1306 " use -I ? to list register names", parse_regs),
1307 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1308 "Record running/enabled time of read (:S) events"),
1309 OPT_CALLBACK('k', "clockid", &record.opts,
1310 "clockid", "clockid to use for events, see clock_gettime()",
1312 OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1313 "opts", "AUX area tracing Snapshot Mode", ""),
1314 OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1315 "per thread proc mmap processing timeout in ms"),
1316 OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1317 "Record context switch events"),
1318 OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1319 "Configure all used events to run in kernel space.",
1320 PARSE_OPT_EXCLUSIVE),
1321 OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1322 "Configure all used events to run in user space.",
1323 PARSE_OPT_EXCLUSIVE),
1324 OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1325 "clang binary to use for compiling BPF scriptlets"),
1326 OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1327 "options passed to clang when compiling BPF scriptlets"),
1328 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1329 "file", "vmlinux pathname"),
1330 OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1331 "Record build-id of all DSOs regardless of hits"),
1332 OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1333 "append timestamp to output filename"),
1334 OPT_BOOLEAN(0, "switch-output", &record.switch_output,
1335 "Switch output when receive SIGUSR2"),
1339 struct option *record_options = __record_options;
1341 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1344 struct record *rec = &record;
1345 char errbuf[BUFSIZ];
1347 #ifndef HAVE_LIBBPF_SUPPORT
1348 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1349 set_nobuild('\0', "clang-path", true);
1350 set_nobuild('\0', "clang-opt", true);
1354 #ifndef HAVE_BPF_PROLOGUE
1355 # if !defined (HAVE_DWARF_SUPPORT)
1356 # define REASON "NO_DWARF=1"
1357 # elif !defined (HAVE_LIBBPF_SUPPORT)
1358 # define REASON "NO_LIBBPF=1"
1360 # define REASON "this architecture doesn't support BPF prologue"
1362 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1363 set_nobuild('\0', "vmlinux", true);
1368 rec->evlist = perf_evlist__new();
1369 if (rec->evlist == NULL)
1372 perf_config(perf_record_config, rec);
1374 argc = parse_options(argc, argv, record_options, record_usage,
1375 PARSE_OPT_STOP_AT_NON_OPTION);
1376 if (!argc && target__none(&rec->opts.target))
1377 usage_with_options(record_usage, record_options);
1379 if (nr_cgroups && !rec->opts.target.system_wide) {
1380 usage_with_options_msg(record_usage, record_options,
1381 "cgroup monitoring only available in system-wide mode");
1384 if (rec->opts.record_switch_events &&
1385 !perf_can_record_switch_events()) {
1386 ui__error("kernel does not support recording context switch events\n");
1387 parse_options_usage(record_usage, record_options, "switch-events", 0);
1391 if (rec->switch_output)
1392 rec->timestamp_filename = true;
1395 rec->itr = auxtrace_record__init(rec->evlist, &err);
1400 err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1401 rec->opts.auxtrace_snapshot_opts);
1405 err = bpf__setup_stdout(rec->evlist);
1407 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1408 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1417 if (symbol_conf.kptr_restrict)
1419 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1420 "check /proc/sys/kernel/kptr_restrict.\n\n"
1421 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1422 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1423 "Samples in kernel modules won't be resolved at all.\n\n"
1424 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1425 "even with a suitable vmlinux or kallsyms file.\n\n");
1427 if (rec->no_buildid_cache || rec->no_buildid) {
1428 disable_buildid_cache();
1429 } else if (rec->switch_output) {
1431 * In 'perf record --switch-output', disable buildid
1432 * generation by default to reduce data file switching
1433 * overhead. Still generate buildid if they are required
1436 * perf record --signal-trigger --no-no-buildid \
1437 * --no-no-buildid-cache
1439 * Following code equals to:
1441 * if ((rec->no_buildid || !rec->no_buildid_set) &&
1442 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1443 * disable_buildid_cache();
1445 bool disable = true;
1447 if (rec->no_buildid_set && !rec->no_buildid)
1449 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1452 rec->no_buildid = true;
1453 rec->no_buildid_cache = true;
1454 disable_buildid_cache();
1458 if (rec->evlist->nr_entries == 0 &&
1459 perf_evlist__add_default(rec->evlist) < 0) {
1460 pr_err("Not enough memory for event selector list\n");
1461 goto out_symbol_exit;
1464 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1465 rec->opts.no_inherit = true;
1467 err = target__validate(&rec->opts.target);
1469 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1470 ui__warning("%s", errbuf);
1473 err = target__parse_uid(&rec->opts.target);
1475 int saved_errno = errno;
1477 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1478 ui__error("%s", errbuf);
1481 goto out_symbol_exit;
1485 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1486 usage_with_options(record_usage, record_options);
1488 err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1490 goto out_symbol_exit;
1493 * We take all buildids when the file contains
1494 * AUX area tracing data because we do not decode the
1495 * trace because it would take too long.
1497 if (rec->opts.full_auxtrace)
1498 rec->buildid_all = true;
1500 if (record_opts__config(&rec->opts)) {
1502 goto out_symbol_exit;
1505 err = __cmd_record(&record, argc, argv);
1507 perf_evlist__delete(rec->evlist);
1509 auxtrace_record__free(rec->itr);
1513 static void snapshot_sig_handler(int sig __maybe_unused)
1515 if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1516 trigger_hit(&auxtrace_snapshot_trigger);
1517 auxtrace_record__snapshot_started = 1;
1518 if (auxtrace_record__snapshot_start(record.itr))
1519 trigger_error(&auxtrace_snapshot_trigger);
1522 if (trigger_is_ready(&switch_output_trigger))
1523 trigger_hit(&switch_output_trigger);