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"
46 struct perf_tool tool;
47 struct record_opts opts;
49 struct perf_data_file file;
50 struct auxtrace_record *itr;
51 struct perf_evlist *evlist;
52 struct perf_session *session;
57 bool no_buildid_cache;
58 bool no_buildid_cache_set;
60 bool timestamp_filename;
62 unsigned long long samples;
65 static int record__write(struct record *rec, void *bf, size_t size)
67 if (perf_data_file__write(rec->session->file, bf, size) < 0) {
68 pr_err("failed to write perf data, error: %m\n");
72 rec->bytes_written += size;
76 static int process_synthesized_event(struct perf_tool *tool,
77 union perf_event *event,
78 struct perf_sample *sample __maybe_unused,
79 struct machine *machine __maybe_unused)
81 struct record *rec = container_of(tool, struct record, tool);
82 return record__write(rec, event, event->header.size);
85 static int record__mmap_read(struct record *rec, int idx)
87 struct perf_mmap *md = &rec->evlist->mmap[idx];
88 u64 head = perf_mmap__read_head(md);
90 unsigned char *data = md->base + page_size;
102 if ((old & md->mask) + size != (head & md->mask)) {
103 buf = &data[old & md->mask];
104 size = md->mask + 1 - (old & md->mask);
107 if (record__write(rec, buf, size) < 0) {
113 buf = &data[old & md->mask];
117 if (record__write(rec, buf, size) < 0) {
123 perf_evlist__mmap_consume(rec->evlist, idx);
128 static volatile int done;
129 static volatile int signr = -1;
130 static volatile int child_finished;
132 static volatile int auxtrace_record__snapshot_started;
133 static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
134 static DEFINE_TRIGGER(switch_output_trigger);
136 static void sig_handler(int sig)
146 static void record__sig_exit(void)
151 signal(signr, SIG_DFL);
155 #ifdef HAVE_AUXTRACE_SUPPORT
157 static int record__process_auxtrace(struct perf_tool *tool,
158 union perf_event *event, void *data1,
159 size_t len1, void *data2, size_t len2)
161 struct record *rec = container_of(tool, struct record, tool);
162 struct perf_data_file *file = &rec->file;
166 if (!perf_data_file__is_pipe(file)) {
168 int fd = perf_data_file__fd(file);
171 file_offset = lseek(fd, 0, SEEK_CUR);
172 if (file_offset == -1)
174 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
180 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
181 padding = (len1 + len2) & 7;
183 padding = 8 - padding;
185 record__write(rec, event, event->header.size);
186 record__write(rec, data1, len1);
188 record__write(rec, data2, len2);
189 record__write(rec, &pad, padding);
194 static int record__auxtrace_mmap_read(struct record *rec,
195 struct auxtrace_mmap *mm)
199 ret = auxtrace_mmap__read(mm, rec->itr, &rec->tool,
200 record__process_auxtrace);
210 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
211 struct auxtrace_mmap *mm)
215 ret = auxtrace_mmap__read_snapshot(mm, rec->itr, &rec->tool,
216 record__process_auxtrace,
217 rec->opts.auxtrace_snapshot_size);
227 static int record__auxtrace_read_snapshot_all(struct record *rec)
232 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
233 struct auxtrace_mmap *mm =
234 &rec->evlist->mmap[i].auxtrace_mmap;
239 if (record__auxtrace_mmap_read_snapshot(rec, mm) != 0) {
248 static void record__read_auxtrace_snapshot(struct record *rec)
250 pr_debug("Recording AUX area tracing snapshot\n");
251 if (record__auxtrace_read_snapshot_all(rec) < 0) {
252 trigger_error(&auxtrace_snapshot_trigger);
254 if (auxtrace_record__snapshot_finish(rec->itr))
255 trigger_error(&auxtrace_snapshot_trigger);
257 trigger_ready(&auxtrace_snapshot_trigger);
264 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
265 struct auxtrace_mmap *mm __maybe_unused)
271 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused)
276 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
283 static int record__open(struct record *rec)
286 struct perf_evsel *pos;
287 struct perf_evlist *evlist = rec->evlist;
288 struct perf_session *session = rec->session;
289 struct record_opts *opts = &rec->opts;
292 perf_evlist__config(evlist, opts, &callchain_param);
294 evlist__for_each(evlist, pos) {
296 if (perf_evsel__open(pos, pos->cpus, pos->threads) < 0) {
297 if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
299 ui__warning("%s\n", msg);
304 perf_evsel__open_strerror(pos, &opts->target,
305 errno, msg, sizeof(msg));
306 ui__error("%s\n", msg);
311 if (perf_evlist__apply_filters(evlist, &pos)) {
312 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
313 pos->filter, perf_evsel__name(pos), errno,
314 strerror_r(errno, msg, sizeof(msg)));
319 if (perf_evlist__mmap_ex(evlist, opts->mmap_pages, false,
320 opts->auxtrace_mmap_pages,
321 opts->auxtrace_snapshot_mode) < 0) {
322 if (errno == EPERM) {
323 pr_err("Permission error mapping pages.\n"
324 "Consider increasing "
325 "/proc/sys/kernel/perf_event_mlock_kb,\n"
326 "or try again with a smaller value of -m/--mmap_pages.\n"
327 "(current value: %u,%u)\n",
328 opts->mmap_pages, opts->auxtrace_mmap_pages);
331 pr_err("failed to mmap with %d (%s)\n", errno,
332 strerror_r(errno, msg, sizeof(msg)));
341 session->evlist = evlist;
342 perf_session__set_id_hdr_size(session);
347 static int process_sample_event(struct perf_tool *tool,
348 union perf_event *event,
349 struct perf_sample *sample,
350 struct perf_evsel *evsel,
351 struct machine *machine)
353 struct record *rec = container_of(tool, struct record, tool);
357 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
360 static int process_buildids(struct record *rec)
362 struct perf_data_file *file = &rec->file;
363 struct perf_session *session = rec->session;
369 * During this process, it'll load kernel map and replace the
370 * dso->long_name to a real pathname it found. In this case
371 * we prefer the vmlinux path like
372 * /lib/modules/3.16.4/build/vmlinux
374 * rather than build-id path (in debug directory).
375 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
377 symbol_conf.ignore_vmlinux_buildid = true;
380 * If --buildid-all is given, it marks all DSO regardless of hits,
381 * so no need to process samples.
383 if (rec->buildid_all)
384 rec->tool.sample = NULL;
386 return perf_session__process_events(session);
389 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
392 struct perf_tool *tool = data;
394 *As for guest kernel when processing subcommand record&report,
395 *we arrange module mmap prior to guest kernel mmap and trigger
396 *a preload dso because default guest module symbols are loaded
397 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
398 *method is used to avoid symbol missing when the first addr is
399 *in module instead of in guest kernel.
401 err = perf_event__synthesize_modules(tool, process_synthesized_event,
404 pr_err("Couldn't record guest kernel [%d]'s reference"
405 " relocation symbol.\n", machine->pid);
408 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
409 * have no _text sometimes.
411 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
414 pr_err("Couldn't record guest kernel [%d]'s reference"
415 " relocation symbol.\n", machine->pid);
418 static struct perf_event_header finished_round_event = {
419 .size = sizeof(struct perf_event_header),
420 .type = PERF_RECORD_FINISHED_ROUND,
423 static int record__mmap_read_all(struct record *rec)
425 u64 bytes_written = rec->bytes_written;
429 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
430 struct auxtrace_mmap *mm = &rec->evlist->mmap[i].auxtrace_mmap;
432 if (rec->evlist->mmap[i].base) {
433 if (record__mmap_read(rec, i) != 0) {
439 if (mm->base && !rec->opts.auxtrace_snapshot_mode &&
440 record__auxtrace_mmap_read(rec, mm) != 0) {
447 * Mark the round finished in case we wrote
448 * at least one event.
450 if (bytes_written != rec->bytes_written)
451 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
457 static void record__init_features(struct record *rec)
459 struct perf_session *session = rec->session;
462 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
463 perf_header__set_feat(&session->header, feat);
466 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
468 if (!have_tracepoints(&rec->evlist->entries))
469 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
471 if (!rec->opts.branch_stack)
472 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
474 if (!rec->opts.full_auxtrace)
475 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
477 perf_header__clear_feat(&session->header, HEADER_STAT);
481 record__finish_output(struct record *rec)
483 struct perf_data_file *file = &rec->file;
484 int fd = perf_data_file__fd(file);
489 rec->session->header.data_size += rec->bytes_written;
490 file->size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
492 if (!rec->no_buildid) {
493 process_buildids(rec);
495 if (rec->buildid_all)
496 dsos__hit_all(rec->session);
498 perf_session__write_header(rec->session, rec->evlist, fd, true);
503 static int record__synthesize_workload(struct record *rec)
506 struct thread_map map;
507 struct thread_map_data map_data;
510 thread_map.map.nr = 1;
511 thread_map.map.map[0].pid = rec->evlist->workload.pid;
512 thread_map.map.map[0].comm = NULL;
513 return perf_event__synthesize_thread_map(&rec->tool, &thread_map.map,
514 process_synthesized_event,
515 &rec->session->machines.host,
516 rec->opts.sample_address,
517 rec->opts.proc_map_timeout);
520 static int record__synthesize(struct record *rec);
523 record__switch_output(struct record *rec, bool at_exit)
525 struct perf_data_file *file = &rec->file;
528 /* Same Size: "2015122520103046"*/
529 char timestamp[] = "InvalidTimestamp";
532 record__finish_output(rec);
533 err = fetch_current_timestamp(timestamp, sizeof(timestamp));
535 pr_err("Failed to get current timestamp\n");
539 fd = perf_data_file__switch(file, timestamp,
540 rec->session->header.data_offset,
542 if (fd >= 0 && !at_exit) {
543 rec->bytes_written = 0;
544 rec->session->header.data_size = 0;
548 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
549 file->path, timestamp);
551 /* Output tracking events */
553 record__synthesize(rec);
556 * In 'perf record --switch-output' without -a,
557 * record__synthesize() in record__switch_output() won't
558 * generate tracking events because there's no thread_map
559 * in evlist. Which causes newly created perf.data doesn't
560 * contain map and comm information.
561 * Create a fake thread_map and directly call
562 * perf_event__synthesize_thread_map() for those events.
564 if (target__none(&rec->opts.target))
565 record__synthesize_workload(rec);
570 static volatile int workload_exec_errno;
573 * perf_evlist__prepare_workload will send a SIGUSR1
574 * if the fork fails, since we asked by setting its
575 * want_signal to true.
577 static void workload_exec_failed_signal(int signo __maybe_unused,
579 void *ucontext __maybe_unused)
581 workload_exec_errno = info->si_value.sival_int;
586 static void snapshot_sig_handler(int sig);
589 perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused,
590 struct perf_tool *tool __maybe_unused,
591 perf_event__handler_t process __maybe_unused,
592 struct machine *machine __maybe_unused)
597 static int record__synthesize(struct record *rec)
599 struct perf_session *session = rec->session;
600 struct machine *machine = &session->machines.host;
601 struct perf_data_file *file = &rec->file;
602 struct record_opts *opts = &rec->opts;
603 struct perf_tool *tool = &rec->tool;
604 int fd = perf_data_file__fd(file);
608 err = perf_event__synthesize_attrs(tool, session,
609 process_synthesized_event);
611 pr_err("Couldn't synthesize attrs.\n");
615 if (have_tracepoints(&rec->evlist->entries)) {
617 * FIXME err <= 0 here actually means that
618 * there were no tracepoints so its not really
619 * an error, just that we don't need to
620 * synthesize anything. We really have to
621 * return this more properly and also
622 * propagate errors that now are calling die()
624 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
625 process_synthesized_event);
627 pr_err("Couldn't record tracing data.\n");
630 rec->bytes_written += err;
634 err = perf_event__synth_time_conv(rec->evlist->mmap[0].base, tool,
635 process_synthesized_event, machine);
639 if (rec->opts.full_auxtrace) {
640 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
641 session, process_synthesized_event);
646 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
648 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
649 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
650 "Check /proc/kallsyms permission or run as root.\n");
652 err = perf_event__synthesize_modules(tool, process_synthesized_event,
654 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
655 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
656 "Check /proc/modules permission or run as root.\n");
659 machines__process_guests(&session->machines,
660 perf_event__synthesize_guest_os, tool);
663 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
664 process_synthesized_event, opts->sample_address,
665 opts->proc_map_timeout);
670 static int __cmd_record(struct record *rec, int argc, const char **argv)
674 unsigned long waking = 0;
675 const bool forks = argc > 0;
676 struct machine *machine;
677 struct perf_tool *tool = &rec->tool;
678 struct record_opts *opts = &rec->opts;
679 struct perf_data_file *file = &rec->file;
680 struct perf_session *session;
681 bool disabled = false, draining = false;
684 rec->progname = argv[0];
686 atexit(record__sig_exit);
687 signal(SIGCHLD, sig_handler);
688 signal(SIGINT, sig_handler);
689 signal(SIGTERM, sig_handler);
691 if (rec->opts.auxtrace_snapshot_mode || rec->switch_output) {
692 signal(SIGUSR2, snapshot_sig_handler);
693 if (rec->opts.auxtrace_snapshot_mode)
694 trigger_on(&auxtrace_snapshot_trigger);
695 if (rec->switch_output)
696 trigger_on(&switch_output_trigger);
698 signal(SIGUSR2, SIG_IGN);
701 session = perf_session__new(file, false, tool);
702 if (session == NULL) {
703 pr_err("Perf session creation failed.\n");
707 fd = perf_data_file__fd(file);
708 rec->session = session;
710 record__init_features(rec);
713 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
715 workload_exec_failed_signal);
717 pr_err("Couldn't run the workload!\n");
719 goto out_delete_session;
723 if (record__open(rec) != 0) {
728 err = bpf__apply_obj_config();
732 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
733 pr_err("ERROR: Apply config to BPF failed: %s\n",
739 * Normally perf_session__new would do this, but it doesn't have the
742 if (rec->tool.ordered_events && !perf_evlist__sample_id_all(rec->evlist)) {
743 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
744 rec->tool.ordered_events = false;
747 if (!rec->evlist->nr_groups)
748 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
751 err = perf_header__write_pipe(fd);
755 err = perf_session__write_header(session, rec->evlist, fd, false);
761 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
762 pr_err("Couldn't generate buildids. "
763 "Use --no-buildid to profile anyway.\n");
768 machine = &session->machines.host;
770 err = record__synthesize(rec);
774 if (rec->realtime_prio) {
775 struct sched_param param;
777 param.sched_priority = rec->realtime_prio;
778 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
779 pr_err("Could not set realtime priority.\n");
786 * When perf is starting the traced process, all the events
787 * (apart from group members) have enable_on_exec=1 set,
788 * so don't spoil it by prematurely enabling them.
790 if (!target__none(&opts->target) && !opts->initial_delay)
791 perf_evlist__enable(rec->evlist);
797 union perf_event *event;
799 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
806 * Some H/W events are generated before COMM event
807 * which is emitted during exec(), so perf script
808 * cannot see a correct process name for those events.
809 * Synthesize COMM event to prevent it.
811 perf_event__synthesize_comm(tool, event,
812 rec->evlist->workload.pid,
813 process_synthesized_event,
817 perf_evlist__start_workload(rec->evlist);
820 if (opts->initial_delay) {
821 usleep(opts->initial_delay * 1000);
822 perf_evlist__enable(rec->evlist);
825 trigger_ready(&auxtrace_snapshot_trigger);
826 trigger_ready(&switch_output_trigger);
828 unsigned long long hits = rec->samples;
830 if (record__mmap_read_all(rec) < 0) {
831 trigger_error(&auxtrace_snapshot_trigger);
832 trigger_error(&switch_output_trigger);
837 if (auxtrace_record__snapshot_started) {
838 auxtrace_record__snapshot_started = 0;
839 if (!trigger_is_error(&auxtrace_snapshot_trigger))
840 record__read_auxtrace_snapshot(rec);
841 if (trigger_is_error(&auxtrace_snapshot_trigger)) {
842 pr_err("AUX area tracing snapshot failed\n");
848 if (trigger_is_hit(&switch_output_trigger)) {
849 trigger_ready(&switch_output_trigger);
852 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
855 fd = record__switch_output(rec, false);
857 pr_err("Failed to switch to new file\n");
858 trigger_error(&switch_output_trigger);
864 if (hits == rec->samples) {
865 if (done || draining)
867 err = perf_evlist__poll(rec->evlist, -1);
869 * Propagate error, only if there's any. Ignore positive
870 * number of returned events and interrupt error.
872 if (err > 0 || (err < 0 && errno == EINTR))
876 if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
881 * When perf is starting the traced process, at the end events
882 * die with the process and we wait for that. Thus no need to
883 * disable events in this case.
885 if (done && !disabled && !target__none(&opts->target)) {
886 trigger_off(&auxtrace_snapshot_trigger);
887 perf_evlist__disable(rec->evlist);
891 trigger_off(&auxtrace_snapshot_trigger);
892 trigger_off(&switch_output_trigger);
894 if (forks && workload_exec_errno) {
895 char msg[STRERR_BUFSIZE];
896 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
897 pr_err("Workload failed: %s\n", emsg);
903 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
910 kill(rec->evlist->workload.pid, SIGTERM);
916 else if (WIFEXITED(exit_status))
917 status = WEXITSTATUS(exit_status);
918 else if (WIFSIGNALED(exit_status))
919 signr = WTERMSIG(exit_status);
923 /* this will be recalculated during process_buildids() */
927 if (!rec->timestamp_filename) {
928 record__finish_output(rec);
930 fd = record__switch_output(rec, true);
933 goto out_delete_session;
938 if (!err && !quiet) {
940 const char *postfix = rec->timestamp_filename ?
943 if (rec->samples && !rec->opts.full_auxtrace)
944 scnprintf(samples, sizeof(samples),
945 " (%" PRIu64 " samples)", rec->samples);
949 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
950 perf_data_file__size(file) / 1024.0 / 1024.0,
951 file->path, postfix, samples);
955 perf_session__delete(session);
959 static void callchain_debug(struct callchain_param *callchain)
961 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
963 pr_debug("callchain: type %s\n", str[callchain->record_mode]);
965 if (callchain->record_mode == CALLCHAIN_DWARF)
966 pr_debug("callchain: stack dump size %d\n",
967 callchain->dump_size);
970 int record_opts__parse_callchain(struct record_opts *record,
971 struct callchain_param *callchain,
972 const char *arg, bool unset)
975 callchain->enabled = !unset;
977 /* --no-call-graph */
979 callchain->record_mode = CALLCHAIN_NONE;
980 pr_debug("callchain: disabled\n");
984 ret = parse_callchain_record_opt(arg, callchain);
986 /* Enable data address sampling for DWARF unwind. */
987 if (callchain->record_mode == CALLCHAIN_DWARF)
988 record->sample_address = true;
989 callchain_debug(callchain);
995 int record_parse_callchain_opt(const struct option *opt,
999 return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
1002 int record_callchain_opt(const struct option *opt,
1003 const char *arg __maybe_unused,
1004 int unset __maybe_unused)
1006 struct callchain_param *callchain = opt->value;
1008 callchain->enabled = true;
1010 if (callchain->record_mode == CALLCHAIN_NONE)
1011 callchain->record_mode = CALLCHAIN_FP;
1013 callchain_debug(callchain);
1017 static int perf_record_config(const char *var, const char *value, void *cb)
1019 struct record *rec = cb;
1021 if (!strcmp(var, "record.build-id")) {
1022 if (!strcmp(value, "cache"))
1023 rec->no_buildid_cache = false;
1024 else if (!strcmp(value, "no-cache"))
1025 rec->no_buildid_cache = true;
1026 else if (!strcmp(value, "skip"))
1027 rec->no_buildid = true;
1032 if (!strcmp(var, "record.call-graph"))
1033 var = "call-graph.record-mode"; /* fall-through */
1035 return perf_default_config(var, value, cb);
1038 struct clockid_map {
1043 #define CLOCKID_MAP(n, c) \
1044 { .name = n, .clockid = (c), }
1046 #define CLOCKID_END { .name = NULL, }
1050 * Add the missing ones, we need to build on many distros...
1052 #ifndef CLOCK_MONOTONIC_RAW
1053 #define CLOCK_MONOTONIC_RAW 4
1055 #ifndef CLOCK_BOOTTIME
1056 #define CLOCK_BOOTTIME 7
1059 #define CLOCK_TAI 11
1062 static const struct clockid_map clockids[] = {
1063 /* available for all events, NMI safe */
1064 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC),
1065 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW),
1067 /* available for some events */
1068 CLOCKID_MAP("realtime", CLOCK_REALTIME),
1069 CLOCKID_MAP("boottime", CLOCK_BOOTTIME),
1070 CLOCKID_MAP("tai", CLOCK_TAI),
1072 /* available for the lazy */
1073 CLOCKID_MAP("mono", CLOCK_MONOTONIC),
1074 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW),
1075 CLOCKID_MAP("real", CLOCK_REALTIME),
1076 CLOCKID_MAP("boot", CLOCK_BOOTTIME),
1081 static int parse_clockid(const struct option *opt, const char *str, int unset)
1083 struct record_opts *opts = (struct record_opts *)opt->value;
1084 const struct clockid_map *cm;
1085 const char *ostr = str;
1088 opts->use_clockid = 0;
1096 /* no setting it twice */
1097 if (opts->use_clockid)
1100 opts->use_clockid = true;
1102 /* if its a number, we're done */
1103 if (sscanf(str, "%d", &opts->clockid) == 1)
1106 /* allow a "CLOCK_" prefix to the name */
1107 if (!strncasecmp(str, "CLOCK_", 6))
1110 for (cm = clockids; cm->name; cm++) {
1111 if (!strcasecmp(str, cm->name)) {
1112 opts->clockid = cm->clockid;
1117 opts->use_clockid = false;
1118 ui__warning("unknown clockid %s, check man page\n", ostr);
1122 static int record__parse_mmap_pages(const struct option *opt,
1124 int unset __maybe_unused)
1126 struct record_opts *opts = opt->value;
1128 unsigned int mmap_pages;
1143 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, s);
1146 opts->mmap_pages = mmap_pages;
1154 ret = __perf_evlist__parse_mmap_pages(&mmap_pages, p + 1);
1158 opts->auxtrace_mmap_pages = mmap_pages;
1165 static const char * const __record_usage[] = {
1166 "perf record [<options>] [<command>]",
1167 "perf record [<options>] -- <command> [<options>]",
1170 const char * const *record_usage = __record_usage;
1173 * XXX Ideally would be local to cmd_record() and passed to a record__new
1174 * because we need to have access to it in record__exit, that is called
1175 * after cmd_record() exits, but since record_options need to be accessible to
1176 * builtin-script, leave it here.
1178 * At least we don't ouch it in all the other functions here directly.
1180 * Just say no to tons of global variables, sigh.
1182 static struct record record = {
1184 .sample_time = true,
1185 .mmap_pages = UINT_MAX,
1186 .user_freq = UINT_MAX,
1187 .user_interval = ULLONG_MAX,
1191 .default_per_cpu = true,
1193 .proc_map_timeout = 500,
1196 .sample = process_sample_event,
1197 .fork = perf_event__process_fork,
1198 .exit = perf_event__process_exit,
1199 .comm = perf_event__process_comm,
1200 .mmap = perf_event__process_mmap,
1201 .mmap2 = perf_event__process_mmap2,
1202 .ordered_events = true,
1206 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
1207 "\n\t\t\t\tDefault: fp";
1210 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1211 * with it and switch to use the library functions in perf_evlist that came
1212 * from builtin-record.c, i.e. use record_opts,
1213 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1216 struct option __record_options[] = {
1217 OPT_CALLBACK('e', "event", &record.evlist, "event",
1218 "event selector. use 'perf list' to list available events",
1219 parse_events_option),
1220 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
1221 "event filter", parse_filter),
1222 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
1223 NULL, "don't record events from perf itself",
1225 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
1226 "record events on existing process id"),
1227 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
1228 "record events on existing thread id"),
1229 OPT_INTEGER('r', "realtime", &record.realtime_prio,
1230 "collect data with this RT SCHED_FIFO priority"),
1231 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
1232 "collect data without buffering"),
1233 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
1234 "collect raw sample records from all opened counters"),
1235 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
1236 "system-wide collection from all CPUs"),
1237 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
1238 "list of cpus to monitor"),
1239 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
1240 OPT_STRING('o', "output", &record.file.path, "file",
1241 "output file name"),
1242 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
1243 &record.opts.no_inherit_set,
1244 "child tasks do not inherit counters"),
1245 OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
1246 OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
1247 "number of mmap data pages and AUX area tracing mmap pages",
1248 record__parse_mmap_pages),
1249 OPT_BOOLEAN(0, "group", &record.opts.group,
1250 "put the counters into a counter group"),
1251 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
1252 NULL, "enables call-graph recording" ,
1253 &record_callchain_opt),
1254 OPT_CALLBACK(0, "call-graph", &record.opts,
1255 "record_mode[,record_size]", record_callchain_help,
1256 &record_parse_callchain_opt),
1257 OPT_INCR('v', "verbose", &verbose,
1258 "be more verbose (show counter open errors, etc)"),
1259 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
1260 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
1261 "per thread counts"),
1262 OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
1263 OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
1264 &record.opts.sample_time_set,
1265 "Record the sample timestamps"),
1266 OPT_BOOLEAN('P', "period", &record.opts.period, "Record the sample period"),
1267 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
1269 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
1270 &record.no_buildid_cache_set,
1271 "do not update the buildid cache"),
1272 OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
1273 &record.no_buildid_set,
1274 "do not collect buildids in perf.data"),
1275 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
1276 "monitor event in cgroup name only",
1278 OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
1279 "ms to wait before starting measurement after program start"),
1280 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
1283 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
1284 "branch any", "sample any taken branches",
1285 parse_branch_stack),
1287 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
1288 "branch filter mask", "branch stack filter modes",
1289 parse_branch_stack),
1290 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
1291 "sample by weight (on special events only)"),
1292 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
1293 "sample transaction flags (special events only)"),
1294 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
1295 "use per-thread mmaps"),
1296 OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
1297 "sample selected machine registers on interrupt,"
1298 " use -I ? to list register names", parse_regs),
1299 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
1300 "Record running/enabled time of read (:S) events"),
1301 OPT_CALLBACK('k', "clockid", &record.opts,
1302 "clockid", "clockid to use for events, see clock_gettime()",
1304 OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
1305 "opts", "AUX area tracing Snapshot Mode", ""),
1306 OPT_UINTEGER(0, "proc-map-timeout", &record.opts.proc_map_timeout,
1307 "per thread proc mmap processing timeout in ms"),
1308 OPT_BOOLEAN(0, "switch-events", &record.opts.record_switch_events,
1309 "Record context switch events"),
1310 OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
1311 "Configure all used events to run in kernel space.",
1312 PARSE_OPT_EXCLUSIVE),
1313 OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
1314 "Configure all used events to run in user space.",
1315 PARSE_OPT_EXCLUSIVE),
1316 OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
1317 "clang binary to use for compiling BPF scriptlets"),
1318 OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
1319 "options passed to clang when compiling BPF scriptlets"),
1320 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
1321 "file", "vmlinux pathname"),
1322 OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
1323 "Record build-id of all DSOs regardless of hits"),
1324 OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
1325 "append timestamp to output filename"),
1326 OPT_BOOLEAN(0, "switch-output", &record.switch_output,
1327 "Switch output when receive SIGUSR2"),
1331 struct option *record_options = __record_options;
1333 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
1336 struct record *rec = &record;
1337 char errbuf[BUFSIZ];
1339 #ifndef HAVE_LIBBPF_SUPPORT
1340 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1341 set_nobuild('\0', "clang-path", true);
1342 set_nobuild('\0', "clang-opt", true);
1346 #ifndef HAVE_BPF_PROLOGUE
1347 # if !defined (HAVE_DWARF_SUPPORT)
1348 # define REASON "NO_DWARF=1"
1349 # elif !defined (HAVE_LIBBPF_SUPPORT)
1350 # define REASON "NO_LIBBPF=1"
1352 # define REASON "this architecture doesn't support BPF prologue"
1354 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1355 set_nobuild('\0', "vmlinux", true);
1360 rec->evlist = perf_evlist__new();
1361 if (rec->evlist == NULL)
1364 perf_config(perf_record_config, rec);
1366 argc = parse_options(argc, argv, record_options, record_usage,
1367 PARSE_OPT_STOP_AT_NON_OPTION);
1368 if (!argc && target__none(&rec->opts.target))
1369 usage_with_options(record_usage, record_options);
1371 if (nr_cgroups && !rec->opts.target.system_wide) {
1372 usage_with_options_msg(record_usage, record_options,
1373 "cgroup monitoring only available in system-wide mode");
1376 if (rec->opts.record_switch_events &&
1377 !perf_can_record_switch_events()) {
1378 ui__error("kernel does not support recording context switch events\n");
1379 parse_options_usage(record_usage, record_options, "switch-events", 0);
1383 if (rec->switch_output)
1384 rec->timestamp_filename = true;
1387 rec->itr = auxtrace_record__init(rec->evlist, &err);
1392 err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
1393 rec->opts.auxtrace_snapshot_opts);
1397 err = bpf__setup_stdout(rec->evlist);
1399 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
1400 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1409 if (symbol_conf.kptr_restrict)
1411 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1412 "check /proc/sys/kernel/kptr_restrict.\n\n"
1413 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1414 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1415 "Samples in kernel modules won't be resolved at all.\n\n"
1416 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1417 "even with a suitable vmlinux or kallsyms file.\n\n");
1419 if (rec->no_buildid_cache || rec->no_buildid) {
1420 disable_buildid_cache();
1421 } else if (rec->switch_output) {
1423 * In 'perf record --switch-output', disable buildid
1424 * generation by default to reduce data file switching
1425 * overhead. Still generate buildid if they are required
1428 * perf record --signal-trigger --no-no-buildid \
1429 * --no-no-buildid-cache
1431 * Following code equals to:
1433 * if ((rec->no_buildid || !rec->no_buildid_set) &&
1434 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1435 * disable_buildid_cache();
1437 bool disable = true;
1439 if (rec->no_buildid_set && !rec->no_buildid)
1441 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
1444 rec->no_buildid = true;
1445 rec->no_buildid_cache = true;
1446 disable_buildid_cache();
1450 if (rec->evlist->nr_entries == 0 &&
1451 perf_evlist__add_default(rec->evlist) < 0) {
1452 pr_err("Not enough memory for event selector list\n");
1453 goto out_symbol_exit;
1456 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
1457 rec->opts.no_inherit = true;
1459 err = target__validate(&rec->opts.target);
1461 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1462 ui__warning("%s", errbuf);
1465 err = target__parse_uid(&rec->opts.target);
1467 int saved_errno = errno;
1469 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1470 ui__error("%s", errbuf);
1473 goto out_symbol_exit;
1477 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
1478 usage_with_options(record_usage, record_options);
1480 err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
1482 goto out_symbol_exit;
1485 * We take all buildids when the file contains
1486 * AUX area tracing data because we do not decode the
1487 * trace because it would take too long.
1489 if (rec->opts.full_auxtrace)
1490 rec->buildid_all = true;
1492 if (record_opts__config(&rec->opts)) {
1494 goto out_symbol_exit;
1497 err = __cmd_record(&record, argc, argv);
1499 perf_evlist__delete(rec->evlist);
1501 auxtrace_record__free(rec->itr);
1505 static void snapshot_sig_handler(int sig __maybe_unused)
1507 if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1508 trigger_hit(&auxtrace_snapshot_trigger);
1509 auxtrace_record__snapshot_started = 1;
1510 if (auxtrace_record__snapshot_start(record.itr))
1511 trigger_error(&auxtrace_snapshot_trigger);
1514 if (trigger_is_ready(&switch_output_trigger))
1515 trigger_hit(&switch_output_trigger);