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 "util/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"
32 #include "util/parse-branch-options.h"
33 #include "util/parse-regs-options.h"
41 struct perf_tool tool;
42 struct record_opts opts;
44 struct perf_data_file file;
45 struct auxtrace_record *itr;
46 struct perf_evlist *evlist;
47 struct perf_session *session;
51 bool no_buildid_cache;
52 unsigned long long samples;
55 static int record__write(struct record *rec, void *bf, size_t size)
57 if (perf_data_file__write(rec->session->file, bf, size) < 0) {
58 pr_err("failed to write perf data, error: %m\n");
62 rec->bytes_written += size;
66 static int process_synthesized_event(struct perf_tool *tool,
67 union perf_event *event,
68 struct perf_sample *sample __maybe_unused,
69 struct machine *machine __maybe_unused)
71 struct record *rec = container_of(tool, struct record, tool);
72 return record__write(rec, event, event->header.size);
75 static int record__mmap_read(struct record *rec, int idx)
77 struct perf_mmap *md = &rec->evlist->mmap[idx];
78 u64 head = perf_mmap__read_head(md);
80 unsigned char *data = md->base + page_size;
92 if ((old & md->mask) + size != (head & md->mask)) {
93 buf = &data[old & md->mask];
94 size = md->mask + 1 - (old & md->mask);
97 if (record__write(rec, buf, size) < 0) {
103 buf = &data[old & md->mask];
107 if (record__write(rec, buf, size) < 0) {
113 perf_evlist__mmap_consume(rec->evlist, idx);
118 static volatile int done;
119 static volatile int signr = -1;
120 static volatile int child_finished;
121 static volatile int auxtrace_snapshot_enabled;
122 static volatile int auxtrace_snapshot_err;
123 static volatile int auxtrace_record__snapshot_started;
125 static void sig_handler(int sig)
135 static void record__sig_exit(void)
140 signal(signr, SIG_DFL);
144 #ifdef HAVE_AUXTRACE_SUPPORT
146 static int record__process_auxtrace(struct perf_tool *tool,
147 union perf_event *event, void *data1,
148 size_t len1, void *data2, size_t len2)
150 struct record *rec = container_of(tool, struct record, tool);
151 struct perf_data_file *file = &rec->file;
155 if (!perf_data_file__is_pipe(file)) {
157 int fd = perf_data_file__fd(file);
160 file_offset = lseek(fd, 0, SEEK_CUR);
161 if (file_offset == -1)
163 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
169 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
170 padding = (len1 + len2) & 7;
172 padding = 8 - padding;
174 record__write(rec, event, event->header.size);
175 record__write(rec, data1, len1);
177 record__write(rec, data2, len2);
178 record__write(rec, &pad, padding);
183 static int record__auxtrace_mmap_read(struct record *rec,
184 struct auxtrace_mmap *mm)
188 ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
189 record__process_auxtrace);
199 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
200 struct auxtrace_mmap *mm)
204 ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
205 record__process_auxtrace,
206 rec->opts.auxtrace_snapshot_size);
216 static int record__auxtrace_read_snapshot_all(struct record *rec)
221 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
222 struct auxtrace_mmap *mm =
223 &rec->evlist->mmap[i].auxtrace_mmap;
228 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
237 static void record__read_auxtrace_snapshot(struct record *rec)
239 pr_debug("Recording AUX area tracing snapshot\n");
240 if (record__auxtrace_read_snapshot_all(rec) < 0) {
241 auxtrace_snapshot_err = -1;
243 auxtrace_snapshot_err = auxtrace_record__snapshot_finish(rec->itr);
244 if (!auxtrace_snapshot_err)
245 auxtrace_snapshot_enabled = 1;
252 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
253 struct auxtrace_mmap *mm __maybe_unused)
259 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
264 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
271 static int record__open(struct record *rec)
274 struct perf_evsel *pos;
275 struct perf_evlist *evlist = rec->evlist;
276 struct perf_session *session = rec->session;
277 struct record_opts *opts = &rec->opts;
280 perf_evlist__config(evlist, opts);
282 evlist__for_each(evlist, pos) {
284 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
285 if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
287 ui__warning("%s\n", msg);
292 perf_evsel__open_strerror(pos, &opts->target,
293 errno, msg, sizeof(msg));
294 ui__error("%s\n", msg);
299 if (perf_evlist__apply_filters(evlist, &pos)) {
300 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
301 pos->filter, perf_evsel__name(pos), errno,
302 strerror_r(errno, msg, sizeof(msg)));
307 if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
308 opts->auxtrace_mmap_pages,
309 opts->auxtrace_snapshot_mode) < 0) {
310 if (errno == EPERM) {
311 pr_err("Permission error mapping pages.\n"
312 "Consider increasing "
313 "/proc/sys/kernel/perf_event_mlock_kb,\n"
314 "or try again with a smaller value of -m/--mmap_pages.\n"
315 "(current value: %u,%u)\n",
316 opts->mmap_pages, opts->auxtrace_mmap_pages);
319 pr_err("failed to mmap with %d (%s)\n", errno,
320 strerror_r(errno, msg, sizeof(msg)));
326 session->evlist = evlist;
327 perf_session__set_id_hdr_size(session);
332 static int process_sample_event(struct perf_tool *tool,
333 union perf_event *event,
334 struct perf_sample *sample,
335 struct perf_evsel *evsel,
336 struct machine *machine)
338 struct record *rec = container_of(tool, struct record, tool);
342 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
345 static int process_buildids(struct record *rec)
347 struct perf_data_file *file = &rec->file;
348 struct perf_session *session = rec->session;
354 * During this process, it'll load kernel map and replace the
355 * dso->long_name to a real pathname it found. In this case
356 * we prefer the vmlinux path like
357 * /lib/modules/3.16.4/build/vmlinux
359 * rather than build-id path (in debug directory).
360 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
362 symbol_conf.ignore_vmlinux_buildid = true;
364 return perf_session__process_events(session);
367 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
370 struct perf_tool *tool = data;
372 *As for guest kernel when processing subcommand record&report,
373 *we arrange module mmap prior to guest kernel mmap and trigger
374 *a preload dso because default guest module symbols are loaded
375 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
376 *method is used to avoid symbol missing when the first addr is
377 *in module instead of in guest kernel.
379 err = perf_event__synthesize_modules(tool, process_synthesized_event,
382 pr_err("Couldn't record guest kernel [%d]'s reference"
383 " relocation symbol.\n", machine->pid);
386 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
387 * have no _text sometimes.
389 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
392 pr_err("Couldn't record guest kernel [%d]'s reference"
393 " relocation symbol.\n", machine->pid);
396 static struct perf_event_header finished_round_event = {
397 .size = sizeof(struct perf_event_header),
398 .type = PERF_RECORD_FINISHED_ROUND,
401 static int record__mmap_read_all(struct record *rec)
403 u64 bytes_written = rec->bytes_written;
407 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
408 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
410 if (rec->evlist->mmap[i].base) {
411 if (record__mmap_read(rec, i) != 0) {
417 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
418 record__auxtrace_mmap_read(rec, mm) != 0) {
425 * Mark the round finished in case we wrote
426 * at least one event.
428 if (bytes_written != rec->bytes_written)
429 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
435 static void record__init_features(struct record *rec)
437 struct perf_session *session = rec->session;
440 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
441 perf_header__set_feat(&session->header, feat);
444 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
446 if (!have_tracepoints(&rec->evlist->entries))
447 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
449 if (!rec->opts.branch_stack)
450 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
452 if (!rec->opts.full_auxtrace)
453 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
456 static volatile int workload_exec_errno;
459 * perf_evlist__prepare_workload will send a SIGUSR1
460 * if the fork fails, since we asked by setting its
461 * want_signal to true.
463 static void workload_exec_failed_signal(int signo __maybe_unused,
465 void *ucontext __maybe_unused)
467 workload_exec_errno = info->si_value.sival_int;
472 static void snapshot_sig_handler(int sig);
474 static int __cmd_record(struct record *rec, int argc, const char **argv)
478 unsigned long waking = 0;
479 const bool forks = argc > 0;
480 struct machine *machine;
481 struct perf_tool *tool = &rec->tool;
482 struct record_opts *opts = &rec->opts;
483 struct perf_data_file *file = &rec->file;
484 struct perf_session *session;
485 bool disabled = false, draining = false;
488 rec->progname = argv[0];
490 atexit(record__sig_exit);
491 signal(SIGCHLD, sig_handler);
492 signal(SIGINT, sig_handler);
493 signal(SIGTERM, sig_handler);
494 if (rec->opts.auxtrace_snapshot_mode)
495 signal(SIGUSR2, snapshot_sig_handler);
497 signal(SIGUSR2, SIG_IGN);
499 session = perf_session__new(file, false, tool);
500 if (session == NULL) {
501 pr_err("Perf session creation failed.\n");
505 fd = perf_data_file__fd(file);
506 rec->session = session;
508 record__init_features(rec);
511 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
513 workload_exec_failed_signal);
515 pr_err("Couldn't run the workload!\n");
517 goto out_delete_session;
521 if (record__open(rec) != 0) {
527 * Normally perf_session__new would do this, but it doesn't have the
530 if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
531 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
532 rec->tool.ordered_events = false;
535 if (!rec->evlist->nr_groups)
536 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
539 err = perf_header__write_pipe(fd);
543 err = perf_session__write_header(session, rec->evlist, fd, false);
549 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
550 pr_err("Couldn't generate buildids. "
551 "Use --no-buildid to profile anyway.\n");
556 machine = &session->machines.host;
559 err = perf_event__synthesize_attrs(tool, session,
560 process_synthesized_event);
562 pr_err("Couldn't synthesize attrs.\n");
566 if (have_tracepoints(&rec->evlist->entries)) {
568 * FIXME err <= 0 here actually means that
569 * there were no tracepoints so its not really
570 * an error, just that we don't need to
571 * synthesize anything. We really have to
572 * return this more properly and also
573 * propagate errors that now are calling die()
575 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
576 process_synthesized_event);
578 pr_err("Couldn't record tracing data.\n");
581 rec->bytes_written += err;
585 if (rec->opts.full_auxtrace) {
586 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
587 session, process_synthesized_event);
589 goto out_delete_session;
592 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
595 pr_err("Couldn't record kernel reference relocation symbol\n"
596 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
597 "Check /proc/kallsyms permission or run as root.\n");
599 err = perf_event__synthesize_modules(tool, process_synthesized_event,
602 pr_err("Couldn't record kernel module information.\n"
603 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
604 "Check /proc/modules permission or run as root.\n");
607 machines__process_guests(&session->machines,
608 perf_event__synthesize_guest_os, tool);
611 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
612 process_synthesized_event, opts->sample_address,
613 opts->proc_map_timeout);
617 if (rec->realtime_prio) {
618 struct sched_param param;
620 param.sched_priority = rec->realtime_prio;
621 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
622 pr_err("Could not set realtime priority.\n");
629 * When perf is starting the traced process, all the events
630 * (apart from group members) have enable_on_exec=1 set,
631 * so don't spoil it by prematurely enabling them.
633 if (!target__none(&opts->target) && !opts->initial_delay)
634 perf_evlist__enable(rec->evlist);
640 union perf_event *event;
642 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
649 * Some H/W events are generated before COMM event
650 * which is emitted during exec(), so perf script
651 * cannot see a correct process name for those events.
652 * Synthesize COMM event to prevent it.
654 perf_event__synthesize_comm(tool, event,
655 rec->evlist->workload.pid,
656 process_synthesized_event,
660 perf_evlist__start_workload(rec->evlist);
663 if (opts->initial_delay) {
664 usleep(opts->initial_delay * 1000);
665 perf_evlist__enable(rec->evlist);
668 auxtrace_snapshot_enabled = 1;
670 unsigned long long hits = rec->samples;
672 if (record__mmap_read_all(rec) < 0) {
673 auxtrace_snapshot_enabled = 0;
678 if (auxtrace_record__snapshot_started) {
679 auxtrace_record__snapshot_started = 0;
680 if (!auxtrace_snapshot_err)
681 record__read_auxtrace_snapshot(rec);
682 if (auxtrace_snapshot_err) {
683 pr_err("AUX area tracing snapshot failed\n");
689 if (hits == rec->samples) {
690 if (done || draining)
692 err = perf_evlist__poll(rec->evlist, -1);
694 * Propagate error, only if there's any. Ignore positive
695 * number of returned events and interrupt error.
697 if (err > 0 || (err < 0 && errno == EINTR))
701 if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
706 * When perf is starting the traced process, at the end events
707 * die with the process and we wait for that. Thus no need to
708 * disable events in this case.
710 if (done && !disabled && !target__none(&opts->target)) {
711 auxtrace_snapshot_enabled = 0;
712 perf_evlist__disable(rec->evlist);
716 auxtrace_snapshot_enabled = 0;
718 if (forks && workload_exec_errno) {
719 char msg[STRERR_BUFSIZE];
720 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
721 pr_err("Workload failed: %s\n", emsg);
727 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
734 kill(rec->evlist->workload.pid, SIGTERM);
740 else if (WIFEXITED(exit_status))
741 status = WEXITSTATUS(exit_status);
742 else if (WIFSIGNALED(exit_status))
743 signr = WTERMSIG(exit_status);
747 /* this will be recalculated during process_buildids() */
750 if (!err && !file->is_pipe) {
751 rec->session->header.data_size += rec->bytes_written;
752 file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
754 if (!rec->no_buildid) {
755 process_buildids(rec);
757 * We take all buildids when the file contains
758 * AUX area tracing data because we do not decode the
759 * trace because it would take too long.
761 if (rec->opts.full_auxtrace)
762 dsos__hit_all(rec->session);
764 perf_session__write_header(rec->session, rec->evlist, fd, true);
767 if (!err && !quiet) {
770 if (rec->samples && !rec->opts.full_auxtrace)
771 scnprintf(samples, sizeof(samples),
772 " (%" PRIu64 " samples)", rec->samples);
776 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
777 perf_data_file__size(file) / 1024.0 / 1024.0,
778 file->path, samples);
782 perf_session__delete(session);
786 static void callchain_debug(void)
788 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
790 pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
792 if (callchain_param.record_mode == CALLCHAIN_DWARF)
793 pr_debug("callchain: stack dump size %d\n",
794 callchain_param.dump_size);
797 int record_parse_callchain_opt(const struct option *opt,
802 struct record_opts *record = (struct record_opts *)opt->value;
804 record->callgraph_set = true;
805 callchain_param.enabled = !unset;
807 /* --no-call-graph */
809 callchain_param.record_mode = CALLCHAIN_NONE;
810 pr_debug("callchain: disabled\n");
814 ret = parse_callchain_record_opt(arg, &callchain_param);
821 int record_callchain_opt(const struct option *opt,
822 const char *arg __maybe_unused,
823 int unset __maybe_unused)
825 struct record_opts *record = (struct record_opts *)opt->value;
827 record->callgraph_set = true;
828 callchain_param.enabled = true;
830 if (callchain_param.record_mode == CALLCHAIN_NONE)
831 callchain_param.record_mode = CALLCHAIN_FP;
837 static int perf_record_config(const char *var, const char *value, void *cb)
839 if (!strcmp(var, "record.call-graph"))
840 var = "call-graph.record-mode"; /* fall-through */
842 return perf_default_config(var, value, cb);
850 #define CLOCKID_MAP(n, c) \
851 { .name = n, .clockid = (c), }
853 #define CLOCKID_END { .name = NULL, }
857 * Add the missing ones, we need to build on many distros...
859 #ifndef CLOCK_MONOTONIC_RAW
860 #define CLOCK_MONOTONIC_RAW 4
862 #ifndef CLOCK_BOOTTIME
863 #define CLOCK_BOOTTIME 7
869 static const struct clockid_map clockids[] = {
870 /* available for all events, NMI safe */
871 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
872 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
874 /* available for some events */
875 CLOCKID_MAP("realtime", CLOCK_REALTIME),
876 CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
877 CLOCKID_MAP("tai", CLOCK_TAI),
879 /* available for the lazy */
880 CLOCKID_MAP("mono", CLOCK_MONOTONIC),
881 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
882 CLOCKID_MAP("real", CLOCK_REALTIME),
883 CLOCKID_MAP("boot", CLOCK_BOOTTIME),
888 static int parse_clockid(const struct option *opt, const char *str, int unset)
890 struct record_opts *opts = (struct record_opts *)opt->value;
891 const struct clockid_map *cm;
892 const char *ostr = str;
895 opts->use_clockid = 0;
903 /* no setting it twice */
904 if (opts->use_clockid)
907 opts->use_clockid = true;
909 /* if its a number, we're done */
910 if (sscanf(str, "%d", &opts->clockid) == 1)
913 /* allow a "CLOCK_" prefix to the name */
914 if (!strncasecmp(str, "CLOCK_", 6))
917 for (cm = clockids; cm->name; cm++) {
918 if (!strcasecmp(str, cm->name)) {
919 opts->clockid = cm->clockid;
924 opts->use_clockid = false;
925 ui__warning("unknown clockid %s, check man page\n", ostr);
929 static int record__parse_mmap_pages(const struct option *opt,
931 int unset __maybe_unused)
933 struct record_opts *opts = opt->value;
935 unsigned int mmap_pages;
950 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
953 opts->mmap_pages = mmap_pages;
961 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
965 opts->auxtrace_mmap_pages = mmap_pages;
972 static const char * const __record_usage[] = {
973 "perf record [<options>] [<command>]",
974 "perf record [<options>] -- <command> [<options>]",
977 const char * const *record_usage = __record_usage;
980 * XXX Ideally would be local to cmd_record() and passed to a record__new
981 * because we need to have access to it in record__exit, that is called
982 * after cmd_record() exits, but since record_options need to be accessible to
983 * builtin-script, leave it here.
985 * At least we don't ouch it in all the other functions here directly.
987 * Just say no to tons of global variables, sigh.
989 static struct record record = {
992 .mmap_pages = UINT_MAX,
993 .user_freq = UINT_MAX,
994 .user_interval = ULLONG_MAX,
998 .default_per_cpu = true,
1000 .proc_map_timeout = 500,
1003 .sample = process_sample_event,
1004 .fork = perf_event__process_fork,
1005 .exit = perf_event__process_exit,
1006 .comm = perf_event__process_comm,
1007 .mmap = perf_event__process_mmap,
1008 .mmap2 = perf_event__process_mmap2,
1009 .ordered_events = true,
1013 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
1015 #ifdef HAVE_DWARF_UNWIND_SUPPORT
1016 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf lbr";
1018 const char record_callchain_help[] = CALLCHAIN_HELP "fp lbr";
1022 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1023 * with it and switch to use the library functions in perf_evlist that came
1024 * from builtin-record.c, i.e. use record_opts,
1025 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1028 struct option __record_options[] = {
1029 OPT_CALLBACK('e', "event", &record.evlist, "event",
1030 "event selector. use 'perf list' to list available events",
1031 parse_events_option),
1032 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1033 "event filter", parse_filter),
1034 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1035 NULL, "don't record events from perf itself",
1037 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1038 "record events on existing process id"),
1039 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1040 "record events on existing thread id"),
1041 OPT_INTEGER('r', "realtime", &record.realtime_prio,
1042 "collect data with this RT SCHED_FIFO priority"),
1043 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1044 "collect data without buffering"),
1045 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1046 "collect raw sample records from all opened counters"),
1047 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1048 "system-wide collection from all CPUs"),
1049 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1050 "list of cpus to monitor"),
1051 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1052 OPT_STRING('o', "output", &record.file.path, "file",
1053 "output file name"),
1054 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1055 &record.opts.no_inherit_set,
1056 "child tasks do not inherit counters"),
1057 OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1058 OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1059 "number of mmap data pages and AUX area tracing mmap pages",
1060 record__parse_mmap_pages),
1061 OPT_BOOLEAN(0, "group", &record.opts.group,
1062 "put the counters into a counter group"),
1063 OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
1064 NULL, "enables call-graph recording" ,
1065 &record_callchain_opt),
1066 OPT_CALLBACK(0, "call-graph", &record.opts,
1067 "mode[,dump_size]", record_callchain_help,
1068 &record_parse_callchain_opt),
1069 OPT_INCR('v', "verbose", &verbose,
1070 "be more verbose (show counter open errors, etc)"),
1071 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1072 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1073 "per thread counts"),
1074 OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1075 OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1076 &record.opts.sample_time_set,
1077 "Record the sample timestamps"),
1078 OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1079 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1081 OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
1082 "do not update the buildid cache"),
1083 OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
1084 "do not collect buildids in perf.data"),
1085 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1086 "monitor event in cgroup name only",
1088 OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1089 "ms to wait before starting measurement after program start"),
1090 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1093 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1094 "branch any", "sample any taken branches",
1095 parse_branch_stack),
1097 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1098 "branch filter mask", "branch stack filter modes",
1099 parse_branch_stack),
1100 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1101 "sample by weight (on special events only)"),
1102 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1103 "sample transaction flags (special events only)"),
1104 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1105 "use per-thread mmaps"),
1106 OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1107 "sample selected machine registers on interrupt,"
1108 " use -I ? to list register names", parse_regs),
1109 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1110 "Record running/enabled time of read (:S) events"),
1111 OPT_CALLBACK('k', "clockid", &record.opts,
1112 "clockid", "clockid to use for events, see clock_gettime()",
1114 OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1115 "opts", "AUX area tracing Snapshot Mode", ""),
1116 OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1117 "per thread proc mmap processing timeout in ms"),
1118 OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1119 "Record context switch events"),
1123 struct option *record_options = __record_options;
1125 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1128 struct record *rec = &record;
1129 char errbuf[BUFSIZ];
1131 rec->evlist = perf_evlist__new();
1132 if (rec->evlist == NULL)
1135 perf_config(perf_record_config, rec);
1137 argc = parse_options(argc, argv, record_options, record_usage,
1138 PARSE_OPT_STOP_AT_NON_OPTION);
1139 if (!argc && target__none(&rec->opts.target))
1140 usage_with_options(record_usage, record_options);
1142 if (nr_cgroups && !rec->opts.target.system_wide) {
1143 ui__error("cgroup monitoring only available in"
1144 " system-wide mode\n");
1145 usage_with_options(record_usage, record_options);
1147 if (rec->opts.record_switch_events &&
1148 !perf_can_record_switch_events()) {
1149 ui__error("kernel does not support recording context switch events (--switch-events option)\n");
1150 usage_with_options(record_usage, record_options);
1154 rec->itr = auxtrace_record__init(rec->evlist, &err);
1159 err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1160 rec->opts.auxtrace_snapshot_opts);
1168 if (symbol_conf.kptr_restrict)
1170 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1171 "check /proc/sys/kernel/kptr_restrict.\n\n"
1172 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1173 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1174 "Samples in kernel modules won't be resolved at all.\n\n"
1175 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1176 "even with a suitable vmlinux or kallsyms file.\n\n");
1178 if (rec->no_buildid_cache || rec->no_buildid)
1179 disable_buildid_cache();
1181 if (rec->evlist->nr_entries == 0 &&
1182 perf_evlist__add_default(rec->evlist) < 0) {
1183 pr_err("Not enough memory for event selector list\n");
1184 goto out_symbol_exit;
1187 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1188 rec->opts.no_inherit = true;
1190 err = target__validate(&rec->opts.target);
1192 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1193 ui__warning("%s", errbuf);
1196 err = target__parse_uid(&rec->opts.target);
1198 int saved_errno = errno;
1200 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1201 ui__error("%s", errbuf);
1204 goto out_symbol_exit;
1208 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1209 usage_with_options(record_usage, record_options);
1211 err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1213 goto out_symbol_exit;
1215 if (record_opts__config(&rec->opts)) {
1217 goto out_symbol_exit;
1220 err = __cmd_record(&record, argc, argv);
1222 perf_evlist__delete(rec->evlist);
1224 auxtrace_record__free(rec->itr);
1228 static void snapshot_sig_handler(int sig __maybe_unused)
1230 if (!auxtrace_snapshot_enabled)
1232 auxtrace_snapshot_enabled = 0;
1233 auxtrace_snapshot_err = auxtrace_record__snapshot_start(record.itr);
1234 auxtrace_record__snapshot_started = 1;