1 // SPDX-License-Identifier: GPL-2.0
5 * Builtin record command: Record the profile of a workload
6 * (or a CPU, or a PID) into the perf.data output file - for
7 * later analysis via perf report.
11 #include "util/build-id.h"
12 #include <subcmd/parse-options.h>
13 #include "util/parse-events.h"
14 #include "util/config.h"
16 #include "util/callchain.h"
17 #include "util/cgroup.h"
18 #include "util/header.h"
19 #include "util/event.h"
20 #include "util/evlist.h"
21 #include "util/evsel.h"
22 #include "util/debug.h"
23 #include "util/mmap.h"
24 #include "util/target.h"
25 #include "util/session.h"
26 #include "util/tool.h"
27 #include "util/symbol.h"
28 #include "util/record.h"
29 #include "util/cpumap.h"
30 #include "util/thread_map.h"
31 #include "util/data.h"
32 #include "util/perf_regs.h"
33 #include "util/auxtrace.h"
35 #include "util/parse-branch-options.h"
36 #include "util/parse-regs-options.h"
37 #include "util/perf_api_probe.h"
38 #include "util/llvm-utils.h"
39 #include "util/bpf-loader.h"
40 #include "util/trigger.h"
41 #include "util/perf-hooks.h"
42 #include "util/cpu-set-sched.h"
43 #include "util/synthetic-events.h"
44 #include "util/time-utils.h"
45 #include "util/units.h"
46 #include "util/bpf-event.h"
47 #include "util/util.h"
49 #include "util/clockid.h"
50 #include "util/pmu-hybrid.h"
51 #include "util/evlist-hybrid.h"
63 #ifdef HAVE_EVENTFD_SUPPORT
64 #include <sys/eventfd.h>
68 #include <sys/types.h>
71 #include <linux/err.h>
72 #include <linux/string.h>
73 #include <linux/time64.h>
74 #include <linux/zalloc.h>
75 #include <linux/bitmap.h>
78 struct switch_output {
91 struct perf_tool tool;
92 struct record_opts opts;
94 struct perf_data data;
95 struct auxtrace_record *itr;
96 struct evlist *evlist;
97 struct perf_session *session;
98 struct evlist *sb_evlist;
101 bool switch_output_event_set;
104 bool no_buildid_cache;
105 bool no_buildid_cache_set;
108 bool timestamp_filename;
109 bool timestamp_boundary;
110 struct switch_output switch_output;
111 unsigned long long samples;
112 struct mmap_cpu_mask affinity_mask;
113 unsigned long output_max_size; /* = 0: unlimited */
116 static volatile int done;
118 static volatile int auxtrace_record__snapshot_started;
119 static DEFINE_TRIGGER(auxtrace_snapshot_trigger);
120 static DEFINE_TRIGGER(switch_output_trigger);
122 static const char *affinity_tags[PERF_AFFINITY_MAX] = {
126 static bool switch_output_signal(struct record *rec)
128 return rec->switch_output.signal &&
129 trigger_is_ready(&switch_output_trigger);
132 static bool switch_output_size(struct record *rec)
134 return rec->switch_output.size &&
135 trigger_is_ready(&switch_output_trigger) &&
136 (rec->bytes_written >= rec->switch_output.size);
139 static bool switch_output_time(struct record *rec)
141 return rec->switch_output.time &&
142 trigger_is_ready(&switch_output_trigger);
145 static bool record__output_max_size_exceeded(struct record *rec)
147 return rec->output_max_size &&
148 (rec->bytes_written >= rec->output_max_size);
151 static int record__write(struct record *rec, struct mmap *map __maybe_unused,
152 void *bf, size_t size)
154 struct perf_data_file *file = &rec->session->data->file;
156 if (perf_data_file__write(file, bf, size) < 0) {
157 pr_err("failed to write perf data, error: %m\n");
161 rec->bytes_written += size;
163 if (record__output_max_size_exceeded(rec) && !done) {
164 fprintf(stderr, "[ perf record: perf size limit reached (%" PRIu64 " KB),"
165 " stopping session ]\n",
166 rec->bytes_written >> 10);
170 if (switch_output_size(rec))
171 trigger_hit(&switch_output_trigger);
176 static int record__aio_enabled(struct record *rec);
177 static int record__comp_enabled(struct record *rec);
178 static size_t zstd_compress(struct perf_session *session, void *dst, size_t dst_size,
179 void *src, size_t src_size);
181 #ifdef HAVE_AIO_SUPPORT
182 static int record__aio_write(struct aiocb *cblock, int trace_fd,
183 void *buf, size_t size, off_t off)
187 cblock->aio_fildes = trace_fd;
188 cblock->aio_buf = buf;
189 cblock->aio_nbytes = size;
190 cblock->aio_offset = off;
191 cblock->aio_sigevent.sigev_notify = SIGEV_NONE;
194 rc = aio_write(cblock);
197 } else if (errno != EAGAIN) {
198 cblock->aio_fildes = -1;
199 pr_err("failed to queue perf data, error: %m\n");
207 static int record__aio_complete(struct mmap *md, struct aiocb *cblock)
213 ssize_t aio_ret, written;
215 aio_errno = aio_error(cblock);
216 if (aio_errno == EINPROGRESS)
219 written = aio_ret = aio_return(cblock);
221 if (aio_errno != EINTR)
222 pr_err("failed to write perf data, error: %m\n");
226 rem_size = cblock->aio_nbytes - written;
229 cblock->aio_fildes = -1;
231 * md->refcount is incremented in record__aio_pushfn() for
232 * every aio write request started in record__aio_push() so
233 * decrement it because the request is now complete.
235 perf_mmap__put(&md->core);
239 * aio write request may require restart with the
240 * reminder if the kernel didn't write whole
243 rem_off = cblock->aio_offset + written;
244 rem_buf = (void *)(cblock->aio_buf + written);
245 record__aio_write(cblock, cblock->aio_fildes,
246 rem_buf, rem_size, rem_off);
253 static int record__aio_sync(struct mmap *md, bool sync_all)
255 struct aiocb **aiocb = md->aio.aiocb;
256 struct aiocb *cblocks = md->aio.cblocks;
257 struct timespec timeout = { 0, 1000 * 1000 * 1 }; /* 1ms */
262 for (i = 0; i < md->aio.nr_cblocks; ++i) {
263 if (cblocks[i].aio_fildes == -1 || record__aio_complete(md, &cblocks[i])) {
270 * Started aio write is not complete yet
271 * so it has to be waited before the
274 aiocb[i] = &cblocks[i];
281 while (aio_suspend((const struct aiocb **)aiocb, md->aio.nr_cblocks, &timeout)) {
282 if (!(errno == EAGAIN || errno == EINTR))
283 pr_err("failed to sync perf data, error: %m\n");
294 static int record__aio_pushfn(struct mmap *map, void *to, void *buf, size_t size)
296 struct record_aio *aio = to;
299 * map->core.base data pointed by buf is copied into free map->aio.data[] buffer
300 * to release space in the kernel buffer as fast as possible, calling
301 * perf_mmap__consume() from perf_mmap__push() function.
303 * That lets the kernel to proceed with storing more profiling data into
304 * the kernel buffer earlier than other per-cpu kernel buffers are handled.
306 * Coping can be done in two steps in case the chunk of profiling data
307 * crosses the upper bound of the kernel buffer. In this case we first move
308 * part of data from map->start till the upper bound and then the reminder
309 * from the beginning of the kernel buffer till the end of the data chunk.
312 if (record__comp_enabled(aio->rec)) {
313 size = zstd_compress(aio->rec->session, aio->data + aio->size,
314 mmap__mmap_len(map) - aio->size,
317 memcpy(aio->data + aio->size, buf, size);
322 * Increment map->refcount to guard map->aio.data[] buffer
323 * from premature deallocation because map object can be
324 * released earlier than aio write request started on
325 * map->aio.data[] buffer is complete.
327 * perf_mmap__put() is done at record__aio_complete()
328 * after started aio request completion or at record__aio_push()
329 * if the request failed to start.
331 perf_mmap__get(&map->core);
339 static int record__aio_push(struct record *rec, struct mmap *map, off_t *off)
342 int trace_fd = rec->session->data->file.fd;
343 struct record_aio aio = { .rec = rec, .size = 0 };
346 * Call record__aio_sync() to wait till map->aio.data[] buffer
347 * becomes available after previous aio write operation.
350 idx = record__aio_sync(map, false);
351 aio.data = map->aio.data[idx];
352 ret = perf_mmap__push(map, &aio, record__aio_pushfn);
353 if (ret != 0) /* ret > 0 - no data, ret < 0 - error */
357 ret = record__aio_write(&(map->aio.cblocks[idx]), trace_fd, aio.data, aio.size, *off);
360 rec->bytes_written += aio.size;
361 if (switch_output_size(rec))
362 trigger_hit(&switch_output_trigger);
365 * Decrement map->refcount incremented in record__aio_pushfn()
366 * back if record__aio_write() operation failed to start, otherwise
367 * map->refcount is decremented in record__aio_complete() after
368 * aio write operation finishes successfully.
370 perf_mmap__put(&map->core);
376 static off_t record__aio_get_pos(int trace_fd)
378 return lseek(trace_fd, 0, SEEK_CUR);
381 static void record__aio_set_pos(int trace_fd, off_t pos)
383 lseek(trace_fd, pos, SEEK_SET);
386 static void record__aio_mmap_read_sync(struct record *rec)
389 struct evlist *evlist = rec->evlist;
390 struct mmap *maps = evlist->mmap;
392 if (!record__aio_enabled(rec))
395 for (i = 0; i < evlist->core.nr_mmaps; i++) {
396 struct mmap *map = &maps[i];
399 record__aio_sync(map, true);
403 static int nr_cblocks_default = 1;
404 static int nr_cblocks_max = 4;
406 static int record__aio_parse(const struct option *opt,
410 struct record_opts *opts = (struct record_opts *)opt->value;
413 opts->nr_cblocks = 0;
416 opts->nr_cblocks = strtol(str, NULL, 0);
417 if (!opts->nr_cblocks)
418 opts->nr_cblocks = nr_cblocks_default;
423 #else /* HAVE_AIO_SUPPORT */
424 static int nr_cblocks_max = 0;
426 static int record__aio_push(struct record *rec __maybe_unused, struct mmap *map __maybe_unused,
427 off_t *off __maybe_unused)
432 static off_t record__aio_get_pos(int trace_fd __maybe_unused)
437 static void record__aio_set_pos(int trace_fd __maybe_unused, off_t pos __maybe_unused)
441 static void record__aio_mmap_read_sync(struct record *rec __maybe_unused)
446 static int record__aio_enabled(struct record *rec)
448 return rec->opts.nr_cblocks > 0;
451 #define MMAP_FLUSH_DEFAULT 1
452 static int record__mmap_flush_parse(const struct option *opt,
457 struct record_opts *opts = (struct record_opts *)opt->value;
458 static struct parse_tag tags[] = {
459 { .tag = 'B', .mult = 1 },
460 { .tag = 'K', .mult = 1 << 10 },
461 { .tag = 'M', .mult = 1 << 20 },
462 { .tag = 'G', .mult = 1 << 30 },
470 opts->mmap_flush = parse_tag_value(str, tags);
471 if (opts->mmap_flush == (int)-1)
472 opts->mmap_flush = strtol(str, NULL, 0);
475 if (!opts->mmap_flush)
476 opts->mmap_flush = MMAP_FLUSH_DEFAULT;
478 flush_max = evlist__mmap_size(opts->mmap_pages);
480 if (opts->mmap_flush > flush_max)
481 opts->mmap_flush = flush_max;
486 #ifdef HAVE_ZSTD_SUPPORT
487 static unsigned int comp_level_default = 1;
489 static int record__parse_comp_level(const struct option *opt, const char *str, int unset)
491 struct record_opts *opts = opt->value;
494 opts->comp_level = 0;
497 opts->comp_level = strtol(str, NULL, 0);
498 if (!opts->comp_level)
499 opts->comp_level = comp_level_default;
505 static unsigned int comp_level_max = 22;
507 static int record__comp_enabled(struct record *rec)
509 return rec->opts.comp_level > 0;
512 static int process_synthesized_event(struct perf_tool *tool,
513 union perf_event *event,
514 struct perf_sample *sample __maybe_unused,
515 struct machine *machine __maybe_unused)
517 struct record *rec = container_of(tool, struct record, tool);
518 return record__write(rec, NULL, event, event->header.size);
521 static int process_locked_synthesized_event(struct perf_tool *tool,
522 union perf_event *event,
523 struct perf_sample *sample __maybe_unused,
524 struct machine *machine __maybe_unused)
526 static pthread_mutex_t synth_lock = PTHREAD_MUTEX_INITIALIZER;
529 pthread_mutex_lock(&synth_lock);
530 ret = process_synthesized_event(tool, event, sample, machine);
531 pthread_mutex_unlock(&synth_lock);
535 static int record__pushfn(struct mmap *map, void *to, void *bf, size_t size)
537 struct record *rec = to;
539 if (record__comp_enabled(rec)) {
540 size = zstd_compress(rec->session, map->data, mmap__mmap_len(map), bf, size);
545 return record__write(rec, map, bf, size);
548 static volatile int signr = -1;
549 static volatile int child_finished;
550 #ifdef HAVE_EVENTFD_SUPPORT
551 static int done_fd = -1;
554 static void sig_handler(int sig)
562 #ifdef HAVE_EVENTFD_SUPPORT
566 * It is possible for this signal handler to run after done is checked
567 * in the main loop, but before the perf counter fds are polled. If this
568 * happens, the poll() will continue to wait even though done is set,
569 * and will only break out if either another signal is received, or the
570 * counters are ready for read. To ensure the poll() doesn't sleep when
571 * done is set, use an eventfd (done_fd) to wake up the poll().
573 if (write(done_fd, &tmp, sizeof(tmp)) < 0)
574 pr_err("failed to signal wakeup fd, error: %m\n");
576 #endif // HAVE_EVENTFD_SUPPORT
579 static void sigsegv_handler(int sig)
581 perf_hooks__recover();
582 sighandler_dump_stack(sig);
585 static void record__sig_exit(void)
590 signal(signr, SIG_DFL);
594 #ifdef HAVE_AUXTRACE_SUPPORT
596 static int record__process_auxtrace(struct perf_tool *tool,
598 union perf_event *event, void *data1,
599 size_t len1, void *data2, size_t len2)
601 struct record *rec = container_of(tool, struct record, tool);
602 struct perf_data *data = &rec->data;
606 if (!perf_data__is_pipe(data) && perf_data__is_single_file(data)) {
608 int fd = perf_data__fd(data);
611 file_offset = lseek(fd, 0, SEEK_CUR);
612 if (file_offset == -1)
614 err = auxtrace_index__auxtrace_event(&rec->session->auxtrace_index,
620 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
621 padding = (len1 + len2) & 7;
623 padding = 8 - padding;
625 record__write(rec, map, event, event->header.size);
626 record__write(rec, map, data1, len1);
628 record__write(rec, map, data2, len2);
629 record__write(rec, map, &pad, padding);
634 static int record__auxtrace_mmap_read(struct record *rec,
639 ret = auxtrace_mmap__read(map, rec->itr, &rec->tool,
640 record__process_auxtrace);
650 static int record__auxtrace_mmap_read_snapshot(struct record *rec,
655 ret = auxtrace_mmap__read_snapshot(map, rec->itr, &rec->tool,
656 record__process_auxtrace,
657 rec->opts.auxtrace_snapshot_size);
667 static int record__auxtrace_read_snapshot_all(struct record *rec)
672 for (i = 0; i < rec->evlist->core.nr_mmaps; i++) {
673 struct mmap *map = &rec->evlist->mmap[i];
675 if (!map->auxtrace_mmap.base)
678 if (record__auxtrace_mmap_read_snapshot(rec, map) != 0) {
687 static void record__read_auxtrace_snapshot(struct record *rec, bool on_exit)
689 pr_debug("Recording AUX area tracing snapshot\n");
690 if (record__auxtrace_read_snapshot_all(rec) < 0) {
691 trigger_error(&auxtrace_snapshot_trigger);
693 if (auxtrace_record__snapshot_finish(rec->itr, on_exit))
694 trigger_error(&auxtrace_snapshot_trigger);
696 trigger_ready(&auxtrace_snapshot_trigger);
700 static int record__auxtrace_snapshot_exit(struct record *rec)
702 if (trigger_is_error(&auxtrace_snapshot_trigger))
705 if (!auxtrace_record__snapshot_started &&
706 auxtrace_record__snapshot_start(rec->itr))
709 record__read_auxtrace_snapshot(rec, true);
710 if (trigger_is_error(&auxtrace_snapshot_trigger))
716 static int record__auxtrace_init(struct record *rec)
721 rec->itr = auxtrace_record__init(rec->evlist, &err);
726 err = auxtrace_parse_snapshot_options(rec->itr, &rec->opts,
727 rec->opts.auxtrace_snapshot_opts);
731 err = auxtrace_parse_sample_options(rec->itr, rec->evlist, &rec->opts,
732 rec->opts.auxtrace_sample_opts);
736 auxtrace_regroup_aux_output(rec->evlist);
738 return auxtrace_parse_filters(rec->evlist);
744 int record__auxtrace_mmap_read(struct record *rec __maybe_unused,
745 struct mmap *map __maybe_unused)
751 void record__read_auxtrace_snapshot(struct record *rec __maybe_unused,
752 bool on_exit __maybe_unused)
757 int auxtrace_record__snapshot_start(struct auxtrace_record *itr __maybe_unused)
763 int record__auxtrace_snapshot_exit(struct record *rec __maybe_unused)
768 static int record__auxtrace_init(struct record *rec __maybe_unused)
775 static int record__config_text_poke(struct evlist *evlist)
780 /* Nothing to do if text poke is already configured */
781 evlist__for_each_entry(evlist, evsel) {
782 if (evsel->core.attr.text_poke)
786 err = parse_events(evlist, "dummy:u", NULL);
790 evsel = evlist__last(evlist);
792 evsel->core.attr.freq = 0;
793 evsel->core.attr.sample_period = 1;
794 evsel->core.attr.text_poke = 1;
795 evsel->core.attr.ksymbol = 1;
797 evsel->core.system_wide = true;
798 evsel->no_aux_samples = true;
799 evsel->immediate = true;
801 /* Text poke must be collected on all CPUs */
802 perf_cpu_map__put(evsel->core.own_cpus);
803 evsel->core.own_cpus = perf_cpu_map__new(NULL);
804 perf_cpu_map__put(evsel->core.cpus);
805 evsel->core.cpus = perf_cpu_map__get(evsel->core.own_cpus);
807 evsel__set_sample_bit(evsel, TIME);
812 static bool record__kcore_readable(struct machine *machine)
814 char kcore[PATH_MAX];
817 scnprintf(kcore, sizeof(kcore), "%s/proc/kcore", machine->root_dir);
819 fd = open(kcore, O_RDONLY);
828 static int record__kcore_copy(struct machine *machine, struct perf_data *data)
830 char from_dir[PATH_MAX];
831 char kcore_dir[PATH_MAX];
834 snprintf(from_dir, sizeof(from_dir), "%s/proc", machine->root_dir);
836 ret = perf_data__make_kcore_dir(data, kcore_dir, sizeof(kcore_dir));
840 return kcore_copy(from_dir, kcore_dir);
843 static int record__mmap_evlist(struct record *rec,
844 struct evlist *evlist)
846 struct record_opts *opts = &rec->opts;
847 bool auxtrace_overwrite = opts->auxtrace_snapshot_mode ||
848 opts->auxtrace_sample_mode;
851 if (opts->affinity != PERF_AFFINITY_SYS)
852 cpu__setup_cpunode_map();
854 if (evlist__mmap_ex(evlist, opts->mmap_pages,
855 opts->auxtrace_mmap_pages,
857 opts->nr_cblocks, opts->affinity,
858 opts->mmap_flush, opts->comp_level) < 0) {
859 if (errno == EPERM) {
860 pr_err("Permission error mapping pages.\n"
861 "Consider increasing "
862 "/proc/sys/kernel/perf_event_mlock_kb,\n"
863 "or try again with a smaller value of -m/--mmap_pages.\n"
864 "(current value: %u,%u)\n",
865 opts->mmap_pages, opts->auxtrace_mmap_pages);
868 pr_err("failed to mmap with %d (%s)\n", errno,
869 str_error_r(errno, msg, sizeof(msg)));
879 static int record__mmap(struct record *rec)
881 return record__mmap_evlist(rec, rec->evlist);
884 static int record__open(struct record *rec)
888 struct evlist *evlist = rec->evlist;
889 struct perf_session *session = rec->session;
890 struct record_opts *opts = &rec->opts;
894 * For initial_delay, system wide or a hybrid system, we need to add a
895 * dummy event so that we can track PERF_RECORD_MMAP to cover the delay
896 * of waiting or event synthesis.
898 if (opts->initial_delay || target__has_cpu(&opts->target) ||
899 perf_pmu__has_hybrid()) {
900 pos = evlist__get_tracking_event(evlist);
901 if (!evsel__is_dummy_event(pos)) {
902 /* Set up dummy event. */
903 if (evlist__add_dummy(evlist))
905 pos = evlist__last(evlist);
906 evlist__set_tracking_event(evlist, pos);
910 * Enable the dummy event when the process is forked for
911 * initial_delay, immediately for system wide.
913 if (opts->initial_delay && !pos->immediate)
914 pos->core.attr.enable_on_exec = 1;
919 evlist__config(evlist, opts, &callchain_param);
921 evlist__for_each_entry(evlist, pos) {
923 if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
924 if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
926 ui__warning("%s\n", msg);
929 if ((errno == EINVAL || errno == EBADF) &&
930 pos->core.leader != &pos->core &&
932 pos = evlist__reset_weak_group(evlist, pos, true);
936 evsel__open_strerror(pos, &opts->target, errno, msg, sizeof(msg));
937 ui__error("%s\n", msg);
941 pos->supported = true;
944 if (symbol_conf.kptr_restrict && !evlist__exclude_kernel(evlist)) {
946 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
947 "check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
948 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
949 "file is not found in the buildid cache or in the vmlinux path.\n\n"
950 "Samples in kernel modules won't be resolved at all.\n\n"
951 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
952 "even with a suitable vmlinux or kallsyms file.\n\n");
955 if (evlist__apply_filters(evlist, &pos)) {
956 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
957 pos->filter, evsel__name(pos), errno,
958 str_error_r(errno, msg, sizeof(msg)));
963 rc = record__mmap(rec);
967 session->evlist = evlist;
968 perf_session__set_id_hdr_size(session);
973 static void set_timestamp_boundary(struct record *rec, u64 sample_time)
975 if (rec->evlist->first_sample_time == 0)
976 rec->evlist->first_sample_time = sample_time;
979 rec->evlist->last_sample_time = sample_time;
982 static int process_sample_event(struct perf_tool *tool,
983 union perf_event *event,
984 struct perf_sample *sample,
986 struct machine *machine)
988 struct record *rec = container_of(tool, struct record, tool);
990 set_timestamp_boundary(rec, sample->time);
992 if (rec->buildid_all)
996 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
999 static int process_buildids(struct record *rec)
1001 struct perf_session *session = rec->session;
1003 if (perf_data__size(&rec->data) == 0)
1007 * During this process, it'll load kernel map and replace the
1008 * dso->long_name to a real pathname it found. In this case
1009 * we prefer the vmlinux path like
1010 * /lib/modules/3.16.4/build/vmlinux
1012 * rather than build-id path (in debug directory).
1013 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
1015 symbol_conf.ignore_vmlinux_buildid = true;
1018 * If --buildid-all is given, it marks all DSO regardless of hits,
1019 * so no need to process samples. But if timestamp_boundary is enabled,
1020 * it still needs to walk on all samples to get the timestamps of
1021 * first/last samples.
1023 if (rec->buildid_all && !rec->timestamp_boundary)
1024 rec->tool.sample = NULL;
1026 return perf_session__process_events(session);
1029 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
1032 struct perf_tool *tool = data;
1034 *As for guest kernel when processing subcommand record&report,
1035 *we arrange module mmap prior to guest kernel mmap and trigger
1036 *a preload dso because default guest module symbols are loaded
1037 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
1038 *method is used to avoid symbol missing when the first addr is
1039 *in module instead of in guest kernel.
1041 err = perf_event__synthesize_modules(tool, process_synthesized_event,
1044 pr_err("Couldn't record guest kernel [%d]'s reference"
1045 " relocation symbol.\n", machine->pid);
1048 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
1049 * have no _text sometimes.
1051 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1054 pr_err("Couldn't record guest kernel [%d]'s reference"
1055 " relocation symbol.\n", machine->pid);
1058 static struct perf_event_header finished_round_event = {
1059 .size = sizeof(struct perf_event_header),
1060 .type = PERF_RECORD_FINISHED_ROUND,
1063 static void record__adjust_affinity(struct record *rec, struct mmap *map)
1065 if (rec->opts.affinity != PERF_AFFINITY_SYS &&
1066 !bitmap_equal(rec->affinity_mask.bits, map->affinity_mask.bits,
1067 rec->affinity_mask.nbits)) {
1068 bitmap_zero(rec->affinity_mask.bits, rec->affinity_mask.nbits);
1069 bitmap_or(rec->affinity_mask.bits, rec->affinity_mask.bits,
1070 map->affinity_mask.bits, rec->affinity_mask.nbits);
1071 sched_setaffinity(0, MMAP_CPU_MASK_BYTES(&rec->affinity_mask),
1072 (cpu_set_t *)rec->affinity_mask.bits);
1074 mmap_cpu_mask__scnprintf(&rec->affinity_mask, "thread");
1078 static size_t process_comp_header(void *record, size_t increment)
1080 struct perf_record_compressed *event = record;
1081 size_t size = sizeof(*event);
1084 event->header.size += increment;
1088 event->header.type = PERF_RECORD_COMPRESSED;
1089 event->header.size = size;
1094 static size_t zstd_compress(struct perf_session *session, void *dst, size_t dst_size,
1095 void *src, size_t src_size)
1098 size_t max_record_size = PERF_SAMPLE_MAX_SIZE - sizeof(struct perf_record_compressed) - 1;
1100 compressed = zstd_compress_stream_to_records(&session->zstd_data, dst, dst_size, src, src_size,
1101 max_record_size, process_comp_header);
1103 session->bytes_transferred += src_size;
1104 session->bytes_compressed += compressed;
1109 static int record__mmap_read_evlist(struct record *rec, struct evlist *evlist,
1110 bool overwrite, bool synch)
1112 u64 bytes_written = rec->bytes_written;
1116 int trace_fd = rec->data.file.fd;
1122 maps = overwrite ? evlist->overwrite_mmap : evlist->mmap;
1126 if (overwrite && evlist->bkw_mmap_state != BKW_MMAP_DATA_PENDING)
1129 if (record__aio_enabled(rec))
1130 off = record__aio_get_pos(trace_fd);
1132 for (i = 0; i < evlist->core.nr_mmaps; i++) {
1134 struct mmap *map = &maps[i];
1136 if (map->core.base) {
1137 record__adjust_affinity(rec, map);
1139 flush = map->core.flush;
1140 map->core.flush = 1;
1142 if (!record__aio_enabled(rec)) {
1143 if (perf_mmap__push(map, rec, record__pushfn) < 0) {
1145 map->core.flush = flush;
1150 if (record__aio_push(rec, map, &off) < 0) {
1151 record__aio_set_pos(trace_fd, off);
1153 map->core.flush = flush;
1159 map->core.flush = flush;
1162 if (map->auxtrace_mmap.base && !rec->opts.auxtrace_snapshot_mode &&
1163 !rec->opts.auxtrace_sample_mode &&
1164 record__auxtrace_mmap_read(rec, map) != 0) {
1170 if (record__aio_enabled(rec))
1171 record__aio_set_pos(trace_fd, off);
1174 * Mark the round finished in case we wrote
1175 * at least one event.
1177 if (bytes_written != rec->bytes_written)
1178 rc = record__write(rec, NULL, &finished_round_event, sizeof(finished_round_event));
1181 evlist__toggle_bkw_mmap(evlist, BKW_MMAP_EMPTY);
1186 static int record__mmap_read_all(struct record *rec, bool synch)
1190 err = record__mmap_read_evlist(rec, rec->evlist, false, synch);
1194 return record__mmap_read_evlist(rec, rec->evlist, true, synch);
1197 static void record__init_features(struct record *rec)
1199 struct perf_session *session = rec->session;
1202 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
1203 perf_header__set_feat(&session->header, feat);
1205 if (rec->no_buildid)
1206 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
1208 if (!have_tracepoints(&rec->evlist->core.entries))
1209 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
1211 if (!rec->opts.branch_stack)
1212 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
1214 if (!rec->opts.full_auxtrace)
1215 perf_header__clear_feat(&session->header, HEADER_AUXTRACE);
1217 if (!(rec->opts.use_clockid && rec->opts.clockid_res_ns))
1218 perf_header__clear_feat(&session->header, HEADER_CLOCKID);
1220 if (!rec->opts.use_clockid)
1221 perf_header__clear_feat(&session->header, HEADER_CLOCK_DATA);
1223 perf_header__clear_feat(&session->header, HEADER_DIR_FORMAT);
1224 if (!record__comp_enabled(rec))
1225 perf_header__clear_feat(&session->header, HEADER_COMPRESSED);
1227 perf_header__clear_feat(&session->header, HEADER_STAT);
1231 record__finish_output(struct record *rec)
1233 struct perf_data *data = &rec->data;
1234 int fd = perf_data__fd(data);
1239 rec->session->header.data_size += rec->bytes_written;
1240 data->file.size = lseek(perf_data__fd(data), 0, SEEK_CUR);
1242 if (!rec->no_buildid) {
1243 process_buildids(rec);
1245 if (rec->buildid_all)
1246 dsos__hit_all(rec->session);
1248 perf_session__write_header(rec->session, rec->evlist, fd, true);
1253 static int record__synthesize_workload(struct record *rec, bool tail)
1256 struct perf_thread_map *thread_map;
1258 if (rec->opts.tail_synthesize != tail)
1261 thread_map = thread_map__new_by_tid(rec->evlist->workload.pid);
1262 if (thread_map == NULL)
1265 err = perf_event__synthesize_thread_map(&rec->tool, thread_map,
1266 process_synthesized_event,
1267 &rec->session->machines.host,
1268 rec->opts.sample_address);
1269 perf_thread_map__put(thread_map);
1273 static int record__synthesize(struct record *rec, bool tail);
1276 record__switch_output(struct record *rec, bool at_exit)
1278 struct perf_data *data = &rec->data;
1282 /* Same Size: "2015122520103046"*/
1283 char timestamp[] = "InvalidTimestamp";
1285 record__aio_mmap_read_sync(rec);
1287 record__synthesize(rec, true);
1288 if (target__none(&rec->opts.target))
1289 record__synthesize_workload(rec, true);
1292 record__finish_output(rec);
1293 err = fetch_current_timestamp(timestamp, sizeof(timestamp));
1295 pr_err("Failed to get current timestamp\n");
1299 fd = perf_data__switch(data, timestamp,
1300 rec->session->header.data_offset,
1301 at_exit, &new_filename);
1302 if (fd >= 0 && !at_exit) {
1303 rec->bytes_written = 0;
1304 rec->session->header.data_size = 0;
1308 fprintf(stderr, "[ perf record: Dump %s.%s ]\n",
1309 data->path, timestamp);
1311 if (rec->switch_output.num_files) {
1312 int n = rec->switch_output.cur_file + 1;
1314 if (n >= rec->switch_output.num_files)
1316 rec->switch_output.cur_file = n;
1317 if (rec->switch_output.filenames[n]) {
1318 remove(rec->switch_output.filenames[n]);
1319 zfree(&rec->switch_output.filenames[n]);
1321 rec->switch_output.filenames[n] = new_filename;
1326 /* Output tracking events */
1328 record__synthesize(rec, false);
1331 * In 'perf record --switch-output' without -a,
1332 * record__synthesize() in record__switch_output() won't
1333 * generate tracking events because there's no thread_map
1334 * in evlist. Which causes newly created perf.data doesn't
1335 * contain map and comm information.
1336 * Create a fake thread_map and directly call
1337 * perf_event__synthesize_thread_map() for those events.
1339 if (target__none(&rec->opts.target))
1340 record__synthesize_workload(rec, false);
1345 static volatile int workload_exec_errno;
1348 * evlist__prepare_workload will send a SIGUSR1
1349 * if the fork fails, since we asked by setting its
1350 * want_signal to true.
1352 static void workload_exec_failed_signal(int signo __maybe_unused,
1354 void *ucontext __maybe_unused)
1356 workload_exec_errno = info->si_value.sival_int;
1361 static void snapshot_sig_handler(int sig);
1362 static void alarm_sig_handler(int sig);
1364 static const struct perf_event_mmap_page *evlist__pick_pc(struct evlist *evlist)
1367 if (evlist->mmap && evlist->mmap[0].core.base)
1368 return evlist->mmap[0].core.base;
1369 if (evlist->overwrite_mmap && evlist->overwrite_mmap[0].core.base)
1370 return evlist->overwrite_mmap[0].core.base;
1375 static const struct perf_event_mmap_page *record__pick_pc(struct record *rec)
1377 const struct perf_event_mmap_page *pc = evlist__pick_pc(rec->evlist);
1383 static int record__synthesize(struct record *rec, bool tail)
1385 struct perf_session *session = rec->session;
1386 struct machine *machine = &session->machines.host;
1387 struct perf_data *data = &rec->data;
1388 struct record_opts *opts = &rec->opts;
1389 struct perf_tool *tool = &rec->tool;
1390 int fd = perf_data__fd(data);
1392 event_op f = process_synthesized_event;
1394 if (rec->opts.tail_synthesize != tail)
1397 if (data->is_pipe) {
1399 * We need to synthesize events first, because some
1400 * features works on top of them (on report side).
1402 err = perf_event__synthesize_attrs(tool, rec->evlist,
1403 process_synthesized_event);
1405 pr_err("Couldn't synthesize attrs.\n");
1409 err = perf_event__synthesize_features(tool, session, rec->evlist,
1410 process_synthesized_event);
1412 pr_err("Couldn't synthesize features.\n");
1416 if (have_tracepoints(&rec->evlist->core.entries)) {
1418 * FIXME err <= 0 here actually means that
1419 * there were no tracepoints so its not really
1420 * an error, just that we don't need to
1421 * synthesize anything. We really have to
1422 * return this more properly and also
1423 * propagate errors that now are calling die()
1425 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
1426 process_synthesized_event);
1428 pr_err("Couldn't record tracing data.\n");
1431 rec->bytes_written += err;
1435 err = perf_event__synth_time_conv(record__pick_pc(rec), tool,
1436 process_synthesized_event, machine);
1440 /* Synthesize id_index before auxtrace_info */
1441 if (rec->opts.auxtrace_sample_mode) {
1442 err = perf_event__synthesize_id_index(tool,
1443 process_synthesized_event,
1444 session->evlist, machine);
1449 if (rec->opts.full_auxtrace) {
1450 err = perf_event__synthesize_auxtrace_info(rec->itr, tool,
1451 session, process_synthesized_event);
1456 if (!evlist__exclude_kernel(rec->evlist)) {
1457 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
1459 WARN_ONCE(err < 0, "Couldn't record kernel reference relocation symbol\n"
1460 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
1461 "Check /proc/kallsyms permission or run as root.\n");
1463 err = perf_event__synthesize_modules(tool, process_synthesized_event,
1465 WARN_ONCE(err < 0, "Couldn't record kernel module information.\n"
1466 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
1467 "Check /proc/modules permission or run as root.\n");
1471 machines__process_guests(&session->machines,
1472 perf_event__synthesize_guest_os, tool);
1475 err = perf_event__synthesize_extra_attr(&rec->tool,
1477 process_synthesized_event,
1482 err = perf_event__synthesize_thread_map2(&rec->tool, rec->evlist->core.threads,
1483 process_synthesized_event,
1486 pr_err("Couldn't synthesize thread map.\n");
1490 err = perf_event__synthesize_cpu_map(&rec->tool, rec->evlist->core.cpus,
1491 process_synthesized_event, NULL);
1493 pr_err("Couldn't synthesize cpu map.\n");
1497 err = perf_event__synthesize_bpf_events(session, process_synthesized_event,
1500 pr_warning("Couldn't synthesize bpf events.\n");
1502 err = perf_event__synthesize_cgroups(tool, process_synthesized_event,
1505 pr_warning("Couldn't synthesize cgroup events.\n");
1507 if (rec->opts.nr_threads_synthesize > 1) {
1508 perf_set_multithreaded();
1509 f = process_locked_synthesized_event;
1512 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->core.threads,
1513 f, opts->sample_address,
1514 rec->opts.nr_threads_synthesize);
1516 if (rec->opts.nr_threads_synthesize > 1)
1517 perf_set_singlethreaded();
1523 static int record__process_signal_event(union perf_event *event __maybe_unused, void *data)
1525 struct record *rec = data;
1526 pthread_kill(rec->thread_id, SIGUSR2);
1530 static int record__setup_sb_evlist(struct record *rec)
1532 struct record_opts *opts = &rec->opts;
1534 if (rec->sb_evlist != NULL) {
1536 * We get here if --switch-output-event populated the
1537 * sb_evlist, so associate a callback that will send a SIGUSR2
1538 * to the main thread.
1540 evlist__set_cb(rec->sb_evlist, record__process_signal_event, rec);
1541 rec->thread_id = pthread_self();
1543 #ifdef HAVE_LIBBPF_SUPPORT
1544 if (!opts->no_bpf_event) {
1545 if (rec->sb_evlist == NULL) {
1546 rec->sb_evlist = evlist__new();
1548 if (rec->sb_evlist == NULL) {
1549 pr_err("Couldn't create side band evlist.\n.");
1554 if (evlist__add_bpf_sb_event(rec->sb_evlist, &rec->session->header.env)) {
1555 pr_err("Couldn't ask for PERF_RECORD_BPF_EVENT side band events.\n.");
1560 if (evlist__start_sb_thread(rec->sb_evlist, &rec->opts.target)) {
1561 pr_debug("Couldn't start the BPF side band thread:\nBPF programs starting from now on won't be annotatable\n");
1562 opts->no_bpf_event = true;
1568 static int record__init_clock(struct record *rec)
1570 struct perf_session *session = rec->session;
1571 struct timespec ref_clockid;
1572 struct timeval ref_tod;
1575 if (!rec->opts.use_clockid)
1578 if (rec->opts.use_clockid && rec->opts.clockid_res_ns)
1579 session->header.env.clock.clockid_res_ns = rec->opts.clockid_res_ns;
1581 session->header.env.clock.clockid = rec->opts.clockid;
1583 if (gettimeofday(&ref_tod, NULL) != 0) {
1584 pr_err("gettimeofday failed, cannot set reference time.\n");
1588 if (clock_gettime(rec->opts.clockid, &ref_clockid)) {
1589 pr_err("clock_gettime failed, cannot set reference time.\n");
1593 ref = (u64) ref_tod.tv_sec * NSEC_PER_SEC +
1594 (u64) ref_tod.tv_usec * NSEC_PER_USEC;
1596 session->header.env.clock.tod_ns = ref;
1598 ref = (u64) ref_clockid.tv_sec * NSEC_PER_SEC +
1599 (u64) ref_clockid.tv_nsec;
1601 session->header.env.clock.clockid_ns = ref;
1605 static void hit_auxtrace_snapshot_trigger(struct record *rec)
1607 if (trigger_is_ready(&auxtrace_snapshot_trigger)) {
1608 trigger_hit(&auxtrace_snapshot_trigger);
1609 auxtrace_record__snapshot_started = 1;
1610 if (auxtrace_record__snapshot_start(rec->itr))
1611 trigger_error(&auxtrace_snapshot_trigger);
1615 static void record__uniquify_name(struct record *rec)
1618 struct evlist *evlist = rec->evlist;
1622 if (!perf_pmu__has_hybrid())
1625 evlist__for_each_entry(evlist, pos) {
1626 if (!evsel__is_hybrid(pos))
1629 if (strchr(pos->name, '/'))
1632 ret = asprintf(&new_name, "%s/%s/",
1633 pos->pmu_name, pos->name);
1636 pos->name = new_name;
1641 static int __cmd_record(struct record *rec, int argc, const char **argv)
1645 unsigned long waking = 0;
1646 const bool forks = argc > 0;
1647 struct perf_tool *tool = &rec->tool;
1648 struct record_opts *opts = &rec->opts;
1649 struct perf_data *data = &rec->data;
1650 struct perf_session *session;
1651 bool disabled = false, draining = false;
1654 enum evlist_ctl_cmd cmd = EVLIST_CTL_CMD_UNSUPPORTED;
1656 atexit(record__sig_exit);
1657 signal(SIGCHLD, sig_handler);
1658 signal(SIGINT, sig_handler);
1659 signal(SIGTERM, sig_handler);
1660 signal(SIGSEGV, sigsegv_handler);
1662 if (rec->opts.record_namespaces)
1663 tool->namespace_events = true;
1665 if (rec->opts.record_cgroup) {
1666 #ifdef HAVE_FILE_HANDLE
1667 tool->cgroup_events = true;
1669 pr_err("cgroup tracking is not supported\n");
1674 if (rec->opts.auxtrace_snapshot_mode || rec->switch_output.enabled) {
1675 signal(SIGUSR2, snapshot_sig_handler);
1676 if (rec->opts.auxtrace_snapshot_mode)
1677 trigger_on(&auxtrace_snapshot_trigger);
1678 if (rec->switch_output.enabled)
1679 trigger_on(&switch_output_trigger);
1681 signal(SIGUSR2, SIG_IGN);
1684 session = perf_session__new(data, false, tool);
1685 if (IS_ERR(session)) {
1686 pr_err("Perf session creation failed.\n");
1687 return PTR_ERR(session);
1690 fd = perf_data__fd(data);
1691 rec->session = session;
1693 if (zstd_init(&session->zstd_data, rec->opts.comp_level) < 0) {
1694 pr_err("Compression initialization failed.\n");
1697 #ifdef HAVE_EVENTFD_SUPPORT
1698 done_fd = eventfd(0, EFD_NONBLOCK);
1700 pr_err("Failed to create wakeup eventfd, error: %m\n");
1702 goto out_delete_session;
1704 err = evlist__add_wakeup_eventfd(rec->evlist, done_fd);
1706 pr_err("Failed to add wakeup eventfd to poll list\n");
1708 goto out_delete_session;
1710 #endif // HAVE_EVENTFD_SUPPORT
1712 session->header.env.comp_type = PERF_COMP_ZSTD;
1713 session->header.env.comp_level = rec->opts.comp_level;
1715 if (rec->opts.kcore &&
1716 !record__kcore_readable(&session->machines.host)) {
1717 pr_err("ERROR: kcore is not readable.\n");
1721 if (record__init_clock(rec))
1724 record__init_features(rec);
1727 err = evlist__prepare_workload(rec->evlist, &opts->target, argv, data->is_pipe,
1728 workload_exec_failed_signal);
1730 pr_err("Couldn't run the workload!\n");
1732 goto out_delete_session;
1737 * If we have just single event and are sending data
1738 * through pipe, we need to force the ids allocation,
1739 * because we synthesize event name through the pipe
1740 * and need the id for that.
1742 if (data->is_pipe && rec->evlist->core.nr_entries == 1)
1743 rec->opts.sample_id = true;
1745 record__uniquify_name(rec);
1747 if (record__open(rec) != 0) {
1751 session->header.env.comp_mmap_len = session->evlist->core.mmap_len;
1753 if (rec->opts.kcore) {
1754 err = record__kcore_copy(&session->machines.host, data);
1756 pr_err("ERROR: Failed to copy kcore\n");
1761 err = bpf__apply_obj_config();
1763 char errbuf[BUFSIZ];
1765 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
1766 pr_err("ERROR: Apply config to BPF failed: %s\n",
1772 * Normally perf_session__new would do this, but it doesn't have the
1775 if (rec->tool.ordered_events && !evlist__sample_id_all(rec->evlist)) {
1776 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
1777 rec->tool.ordered_events = false;
1780 if (!rec->evlist->core.nr_groups)
1781 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
1783 if (data->is_pipe) {
1784 err = perf_header__write_pipe(fd);
1788 err = perf_session__write_header(session, rec->evlist, fd, false);
1794 if (!rec->no_buildid
1795 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
1796 pr_err("Couldn't generate buildids. "
1797 "Use --no-buildid to profile anyway.\n");
1801 err = record__setup_sb_evlist(rec);
1805 err = record__synthesize(rec, false);
1809 if (rec->realtime_prio) {
1810 struct sched_param param;
1812 param.sched_priority = rec->realtime_prio;
1813 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
1814 pr_err("Could not set realtime priority.\n");
1821 * When perf is starting the traced process, all the events
1822 * (apart from group members) have enable_on_exec=1 set,
1823 * so don't spoil it by prematurely enabling them.
1825 if (!target__none(&opts->target) && !opts->initial_delay)
1826 evlist__enable(rec->evlist);
1832 struct machine *machine = &session->machines.host;
1833 union perf_event *event;
1836 event = malloc(sizeof(event->comm) + machine->id_hdr_size);
1837 if (event == NULL) {
1843 * Some H/W events are generated before COMM event
1844 * which is emitted during exec(), so perf script
1845 * cannot see a correct process name for those events.
1846 * Synthesize COMM event to prevent it.
1848 tgid = perf_event__synthesize_comm(tool, event,
1849 rec->evlist->workload.pid,
1850 process_synthesized_event,
1857 event = malloc(sizeof(event->namespaces) +
1858 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) +
1859 machine->id_hdr_size);
1860 if (event == NULL) {
1866 * Synthesize NAMESPACES event for the command specified.
1868 perf_event__synthesize_namespaces(tool, event,
1869 rec->evlist->workload.pid,
1870 tgid, process_synthesized_event,
1874 evlist__start_workload(rec->evlist);
1877 if (evlist__initialize_ctlfd(rec->evlist, opts->ctl_fd, opts->ctl_fd_ack))
1880 if (opts->initial_delay) {
1881 pr_info(EVLIST_DISABLED_MSG);
1882 if (opts->initial_delay > 0) {
1883 usleep(opts->initial_delay * USEC_PER_MSEC);
1884 evlist__enable(rec->evlist);
1885 pr_info(EVLIST_ENABLED_MSG);
1889 trigger_ready(&auxtrace_snapshot_trigger);
1890 trigger_ready(&switch_output_trigger);
1891 perf_hooks__invoke_record_start();
1893 unsigned long long hits = rec->samples;
1896 * rec->evlist->bkw_mmap_state is possible to be
1897 * BKW_MMAP_EMPTY here: when done == true and
1898 * hits != rec->samples in previous round.
1900 * evlist__toggle_bkw_mmap ensure we never
1901 * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
1903 if (trigger_is_hit(&switch_output_trigger) || done || draining)
1904 evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_DATA_PENDING);
1906 if (record__mmap_read_all(rec, false) < 0) {
1907 trigger_error(&auxtrace_snapshot_trigger);
1908 trigger_error(&switch_output_trigger);
1913 if (auxtrace_record__snapshot_started) {
1914 auxtrace_record__snapshot_started = 0;
1915 if (!trigger_is_error(&auxtrace_snapshot_trigger))
1916 record__read_auxtrace_snapshot(rec, false);
1917 if (trigger_is_error(&auxtrace_snapshot_trigger)) {
1918 pr_err("AUX area tracing snapshot failed\n");
1924 if (trigger_is_hit(&switch_output_trigger)) {
1926 * If switch_output_trigger is hit, the data in
1927 * overwritable ring buffer should have been collected,
1928 * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1930 * If SIGUSR2 raise after or during record__mmap_read_all(),
1931 * record__mmap_read_all() didn't collect data from
1932 * overwritable ring buffer. Read again.
1934 if (rec->evlist->bkw_mmap_state == BKW_MMAP_RUNNING)
1936 trigger_ready(&switch_output_trigger);
1939 * Reenable events in overwrite ring buffer after
1940 * record__mmap_read_all(): we should have collected
1943 evlist__toggle_bkw_mmap(rec->evlist, BKW_MMAP_RUNNING);
1946 fprintf(stderr, "[ perf record: dump data: Woken up %ld times ]\n",
1949 fd = record__switch_output(rec, false);
1951 pr_err("Failed to switch to new file\n");
1952 trigger_error(&switch_output_trigger);
1957 /* re-arm the alarm */
1958 if (rec->switch_output.time)
1959 alarm(rec->switch_output.time);
1962 if (hits == rec->samples) {
1963 if (done || draining)
1965 err = evlist__poll(rec->evlist, -1);
1967 * Propagate error, only if there's any. Ignore positive
1968 * number of returned events and interrupt error.
1970 if (err > 0 || (err < 0 && errno == EINTR))
1974 if (evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
1978 if (evlist__ctlfd_process(rec->evlist, &cmd) > 0) {
1980 case EVLIST_CTL_CMD_SNAPSHOT:
1981 hit_auxtrace_snapshot_trigger(rec);
1982 evlist__ctlfd_ack(rec->evlist);
1984 case EVLIST_CTL_CMD_STOP:
1987 case EVLIST_CTL_CMD_ACK:
1988 case EVLIST_CTL_CMD_UNSUPPORTED:
1989 case EVLIST_CTL_CMD_ENABLE:
1990 case EVLIST_CTL_CMD_DISABLE:
1991 case EVLIST_CTL_CMD_EVLIST:
1992 case EVLIST_CTL_CMD_PING:
1999 * When perf is starting the traced process, at the end events
2000 * die with the process and we wait for that. Thus no need to
2001 * disable events in this case.
2003 if (done && !disabled && !target__none(&opts->target)) {
2004 trigger_off(&auxtrace_snapshot_trigger);
2005 evlist__disable(rec->evlist);
2010 trigger_off(&auxtrace_snapshot_trigger);
2011 trigger_off(&switch_output_trigger);
2013 if (opts->auxtrace_snapshot_on_exit)
2014 record__auxtrace_snapshot_exit(rec);
2016 if (forks && workload_exec_errno) {
2017 char msg[STRERR_BUFSIZE], strevsels[2048];
2018 const char *emsg = str_error_r(workload_exec_errno, msg, sizeof(msg));
2020 evlist__scnprintf_evsels(rec->evlist, sizeof(strevsels), strevsels);
2022 pr_err("Failed to collect '%s' for the '%s' workload: %s\n",
2023 strevsels, argv[0], emsg);
2029 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
2031 if (target__none(&rec->opts.target))
2032 record__synthesize_workload(rec, true);
2035 evlist__finalize_ctlfd(rec->evlist);
2036 record__mmap_read_all(rec, true);
2037 record__aio_mmap_read_sync(rec);
2039 if (rec->session->bytes_transferred && rec->session->bytes_compressed) {
2040 ratio = (float)rec->session->bytes_transferred/(float)rec->session->bytes_compressed;
2041 session->header.env.comp_ratio = ratio + 0.5;
2047 if (!child_finished)
2048 kill(rec->evlist->workload.pid, SIGTERM);
2054 else if (WIFEXITED(exit_status))
2055 status = WEXITSTATUS(exit_status);
2056 else if (WIFSIGNALED(exit_status))
2057 signr = WTERMSIG(exit_status);
2061 record__synthesize(rec, true);
2062 /* this will be recalculated during process_buildids() */
2066 if (!rec->timestamp_filename) {
2067 record__finish_output(rec);
2069 fd = record__switch_output(rec, true);
2072 goto out_delete_session;
2077 perf_hooks__invoke_record_end();
2079 if (!err && !quiet) {
2081 const char *postfix = rec->timestamp_filename ?
2082 ".<timestamp>" : "";
2084 if (rec->samples && !rec->opts.full_auxtrace)
2085 scnprintf(samples, sizeof(samples),
2086 " (%" PRIu64 " samples)", rec->samples);
2090 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s%s",
2091 perf_data__size(data) / 1024.0 / 1024.0,
2092 data->path, postfix, samples);
2094 fprintf(stderr, ", compressed (original %.3f MB, ratio is %.3f)",
2095 rec->session->bytes_transferred / 1024.0 / 1024.0,
2098 fprintf(stderr, " ]\n");
2102 #ifdef HAVE_EVENTFD_SUPPORT
2106 zstd_fini(&session->zstd_data);
2107 perf_session__delete(session);
2109 if (!opts->no_bpf_event)
2110 evlist__stop_sb_thread(rec->sb_evlist);
2114 static void callchain_debug(struct callchain_param *callchain)
2116 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
2118 pr_debug("callchain: type %s\n", str[callchain->record_mode]);
2120 if (callchain->record_mode == CALLCHAIN_DWARF)
2121 pr_debug("callchain: stack dump size %d\n",
2122 callchain->dump_size);
2125 int record_opts__parse_callchain(struct record_opts *record,
2126 struct callchain_param *callchain,
2127 const char *arg, bool unset)
2130 callchain->enabled = !unset;
2132 /* --no-call-graph */
2134 callchain->record_mode = CALLCHAIN_NONE;
2135 pr_debug("callchain: disabled\n");
2139 ret = parse_callchain_record_opt(arg, callchain);
2141 /* Enable data address sampling for DWARF unwind. */
2142 if (callchain->record_mode == CALLCHAIN_DWARF)
2143 record->sample_address = true;
2144 callchain_debug(callchain);
2150 int record_parse_callchain_opt(const struct option *opt,
2154 return record_opts__parse_callchain(opt->value, &callchain_param, arg, unset);
2157 int record_callchain_opt(const struct option *opt,
2158 const char *arg __maybe_unused,
2159 int unset __maybe_unused)
2161 struct callchain_param *callchain = opt->value;
2163 callchain->enabled = true;
2165 if (callchain->record_mode == CALLCHAIN_NONE)
2166 callchain->record_mode = CALLCHAIN_FP;
2168 callchain_debug(callchain);
2172 static int perf_record_config(const char *var, const char *value, void *cb)
2174 struct record *rec = cb;
2176 if (!strcmp(var, "record.build-id")) {
2177 if (!strcmp(value, "cache"))
2178 rec->no_buildid_cache = false;
2179 else if (!strcmp(value, "no-cache"))
2180 rec->no_buildid_cache = true;
2181 else if (!strcmp(value, "skip"))
2182 rec->no_buildid = true;
2183 else if (!strcmp(value, "mmap"))
2184 rec->buildid_mmap = true;
2189 if (!strcmp(var, "record.call-graph")) {
2190 var = "call-graph.record-mode";
2191 return perf_default_config(var, value, cb);
2193 #ifdef HAVE_AIO_SUPPORT
2194 if (!strcmp(var, "record.aio")) {
2195 rec->opts.nr_cblocks = strtol(value, NULL, 0);
2196 if (!rec->opts.nr_cblocks)
2197 rec->opts.nr_cblocks = nr_cblocks_default;
2205 static int record__parse_affinity(const struct option *opt, const char *str, int unset)
2207 struct record_opts *opts = (struct record_opts *)opt->value;
2212 if (!strcasecmp(str, "node"))
2213 opts->affinity = PERF_AFFINITY_NODE;
2214 else if (!strcasecmp(str, "cpu"))
2215 opts->affinity = PERF_AFFINITY_CPU;
2220 static int parse_output_max_size(const struct option *opt,
2221 const char *str, int unset)
2223 unsigned long *s = (unsigned long *)opt->value;
2224 static struct parse_tag tags_size[] = {
2225 { .tag = 'B', .mult = 1 },
2226 { .tag = 'K', .mult = 1 << 10 },
2227 { .tag = 'M', .mult = 1 << 20 },
2228 { .tag = 'G', .mult = 1 << 30 },
2238 val = parse_tag_value(str, tags_size);
2239 if (val != (unsigned long) -1) {
2247 static int record__parse_mmap_pages(const struct option *opt,
2249 int unset __maybe_unused)
2251 struct record_opts *opts = opt->value;
2253 unsigned int mmap_pages;
2268 ret = __evlist__parse_mmap_pages(&mmap_pages, s);
2271 opts->mmap_pages = mmap_pages;
2279 ret = __evlist__parse_mmap_pages(&mmap_pages, p + 1);
2283 opts->auxtrace_mmap_pages = mmap_pages;
2290 static int parse_control_option(const struct option *opt,
2292 int unset __maybe_unused)
2294 struct record_opts *opts = opt->value;
2296 return evlist__parse_control(str, &opts->ctl_fd, &opts->ctl_fd_ack, &opts->ctl_fd_close);
2299 static void switch_output_size_warn(struct record *rec)
2301 u64 wakeup_size = evlist__mmap_size(rec->opts.mmap_pages);
2302 struct switch_output *s = &rec->switch_output;
2306 if (s->size < wakeup_size) {
2309 unit_number__scnprintf(buf, sizeof(buf), wakeup_size);
2310 pr_warning("WARNING: switch-output data size lower than "
2311 "wakeup kernel buffer size (%s) "
2312 "expect bigger perf.data sizes\n", buf);
2316 static int switch_output_setup(struct record *rec)
2318 struct switch_output *s = &rec->switch_output;
2319 static struct parse_tag tags_size[] = {
2320 { .tag = 'B', .mult = 1 },
2321 { .tag = 'K', .mult = 1 << 10 },
2322 { .tag = 'M', .mult = 1 << 20 },
2323 { .tag = 'G', .mult = 1 << 30 },
2326 static struct parse_tag tags_time[] = {
2327 { .tag = 's', .mult = 1 },
2328 { .tag = 'm', .mult = 60 },
2329 { .tag = 'h', .mult = 60*60 },
2330 { .tag = 'd', .mult = 60*60*24 },
2336 * If we're using --switch-output-events, then we imply its
2337 * --switch-output=signal, as we'll send a SIGUSR2 from the side band
2338 * thread to its parent.
2340 if (rec->switch_output_event_set)
2346 if (!strcmp(s->str, "signal")) {
2349 pr_debug("switch-output with SIGUSR2 signal\n");
2353 val = parse_tag_value(s->str, tags_size);
2354 if (val != (unsigned long) -1) {
2356 pr_debug("switch-output with %s size threshold\n", s->str);
2360 val = parse_tag_value(s->str, tags_time);
2361 if (val != (unsigned long) -1) {
2363 pr_debug("switch-output with %s time threshold (%lu seconds)\n",
2371 rec->timestamp_filename = true;
2374 if (s->size && !rec->opts.no_buffering)
2375 switch_output_size_warn(rec);
2380 static const char * const __record_usage[] = {
2381 "perf record [<options>] [<command>]",
2382 "perf record [<options>] -- <command> [<options>]",
2385 const char * const *record_usage = __record_usage;
2387 static int build_id__process_mmap(struct perf_tool *tool, union perf_event *event,
2388 struct perf_sample *sample, struct machine *machine)
2391 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
2392 * no need to add them twice.
2394 if (!(event->header.misc & PERF_RECORD_MISC_USER))
2396 return perf_event__process_mmap(tool, event, sample, machine);
2399 static int build_id__process_mmap2(struct perf_tool *tool, union perf_event *event,
2400 struct perf_sample *sample, struct machine *machine)
2403 * We already have the kernel maps, put in place via perf_session__create_kernel_maps()
2404 * no need to add them twice.
2406 if (!(event->header.misc & PERF_RECORD_MISC_USER))
2409 return perf_event__process_mmap2(tool, event, sample, machine);
2412 static int process_timestamp_boundary(struct perf_tool *tool,
2413 union perf_event *event __maybe_unused,
2414 struct perf_sample *sample,
2415 struct machine *machine __maybe_unused)
2417 struct record *rec = container_of(tool, struct record, tool);
2419 set_timestamp_boundary(rec, sample->time);
2424 * XXX Ideally would be local to cmd_record() and passed to a record__new
2425 * because we need to have access to it in record__exit, that is called
2426 * after cmd_record() exits, but since record_options need to be accessible to
2427 * builtin-script, leave it here.
2429 * At least we don't ouch it in all the other functions here directly.
2431 * Just say no to tons of global variables, sigh.
2433 static struct record record = {
2435 .sample_time = true,
2436 .mmap_pages = UINT_MAX,
2437 .user_freq = UINT_MAX,
2438 .user_interval = ULLONG_MAX,
2442 .default_per_cpu = true,
2444 .mmap_flush = MMAP_FLUSH_DEFAULT,
2445 .nr_threads_synthesize = 1,
2450 .sample = process_sample_event,
2451 .fork = perf_event__process_fork,
2452 .exit = perf_event__process_exit,
2453 .comm = perf_event__process_comm,
2454 .namespaces = perf_event__process_namespaces,
2455 .mmap = build_id__process_mmap,
2456 .mmap2 = build_id__process_mmap2,
2457 .itrace_start = process_timestamp_boundary,
2458 .aux = process_timestamp_boundary,
2459 .ordered_events = true,
2463 const char record_callchain_help[] = CALLCHAIN_RECORD_HELP
2464 "\n\t\t\t\tDefault: fp";
2466 static bool dry_run;
2469 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
2470 * with it and switch to use the library functions in perf_evlist that came
2471 * from builtin-record.c, i.e. use record_opts,
2472 * evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
2475 static struct option __record_options[] = {
2476 OPT_CALLBACK('e', "event", &record.evlist, "event",
2477 "event selector. use 'perf list' to list available events",
2478 parse_events_option),
2479 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
2480 "event filter", parse_filter),
2481 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record.evlist,
2482 NULL, "don't record events from perf itself",
2484 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
2485 "record events on existing process id"),
2486 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
2487 "record events on existing thread id"),
2488 OPT_INTEGER('r', "realtime", &record.realtime_prio,
2489 "collect data with this RT SCHED_FIFO priority"),
2490 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
2491 "collect data without buffering"),
2492 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
2493 "collect raw sample records from all opened counters"),
2494 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
2495 "system-wide collection from all CPUs"),
2496 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
2497 "list of cpus to monitor"),
2498 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
2499 OPT_STRING('o', "output", &record.data.path, "file",
2500 "output file name"),
2501 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
2502 &record.opts.no_inherit_set,
2503 "child tasks do not inherit counters"),
2504 OPT_BOOLEAN(0, "tail-synthesize", &record.opts.tail_synthesize,
2505 "synthesize non-sample events at the end of output"),
2506 OPT_BOOLEAN(0, "overwrite", &record.opts.overwrite, "use overwrite mode"),
2507 OPT_BOOLEAN(0, "no-bpf-event", &record.opts.no_bpf_event, "do not record bpf events"),
2508 OPT_BOOLEAN(0, "strict-freq", &record.opts.strict_freq,
2509 "Fail if the specified frequency can't be used"),
2510 OPT_CALLBACK('F', "freq", &record.opts, "freq or 'max'",
2511 "profile at this frequency",
2512 record__parse_freq),
2513 OPT_CALLBACK('m', "mmap-pages", &record.opts, "pages[,pages]",
2514 "number of mmap data pages and AUX area tracing mmap pages",
2515 record__parse_mmap_pages),
2516 OPT_CALLBACK(0, "mmap-flush", &record.opts, "number",
2517 "Minimal number of bytes that is extracted from mmap data pages (default: 1)",
2518 record__mmap_flush_parse),
2519 OPT_BOOLEAN(0, "group", &record.opts.group,
2520 "put the counters into a counter group"),
2521 OPT_CALLBACK_NOOPT('g', NULL, &callchain_param,
2522 NULL, "enables call-graph recording" ,
2523 &record_callchain_opt),
2524 OPT_CALLBACK(0, "call-graph", &record.opts,
2525 "record_mode[,record_size]", record_callchain_help,
2526 &record_parse_callchain_opt),
2527 OPT_INCR('v', "verbose", &verbose,
2528 "be more verbose (show counter open errors, etc)"),
2529 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
2530 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
2531 "per thread counts"),
2532 OPT_BOOLEAN('d', "data", &record.opts.sample_address, "Record the sample addresses"),
2533 OPT_BOOLEAN(0, "phys-data", &record.opts.sample_phys_addr,
2534 "Record the sample physical addresses"),
2535 OPT_BOOLEAN(0, "data-page-size", &record.opts.sample_data_page_size,
2536 "Record the sampled data address data page size"),
2537 OPT_BOOLEAN(0, "code-page-size", &record.opts.sample_code_page_size,
2538 "Record the sampled code address (ip) page size"),
2539 OPT_BOOLEAN(0, "sample-cpu", &record.opts.sample_cpu, "Record the sample cpu"),
2540 OPT_BOOLEAN_SET('T', "timestamp", &record.opts.sample_time,
2541 &record.opts.sample_time_set,
2542 "Record the sample timestamps"),
2543 OPT_BOOLEAN_SET('P', "period", &record.opts.period, &record.opts.period_set,
2544 "Record the sample period"),
2545 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
2547 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record.no_buildid_cache,
2548 &record.no_buildid_cache_set,
2549 "do not update the buildid cache"),
2550 OPT_BOOLEAN_SET('B', "no-buildid", &record.no_buildid,
2551 &record.no_buildid_set,
2552 "do not collect buildids in perf.data"),
2553 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
2554 "monitor event in cgroup name only",
2556 OPT_INTEGER('D', "delay", &record.opts.initial_delay,
2557 "ms to wait before starting measurement after program start (-1: start with events disabled)"),
2558 OPT_BOOLEAN(0, "kcore", &record.opts.kcore, "copy /proc/kcore"),
2559 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
2562 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
2563 "branch any", "sample any taken branches",
2564 parse_branch_stack),
2566 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
2567 "branch filter mask", "branch stack filter modes",
2568 parse_branch_stack),
2569 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
2570 "sample by weight (on special events only)"),
2571 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
2572 "sample transaction flags (special events only)"),
2573 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
2574 "use per-thread mmaps"),
2575 OPT_CALLBACK_OPTARG('I', "intr-regs", &record.opts.sample_intr_regs, NULL, "any register",
2576 "sample selected machine registers on interrupt,"
2577 " use '-I?' to list register names", parse_intr_regs),
2578 OPT_CALLBACK_OPTARG(0, "user-regs", &record.opts.sample_user_regs, NULL, "any register",
2579 "sample selected machine registers on interrupt,"
2580 " use '--user-regs=?' to list register names", parse_user_regs),
2581 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
2582 "Record running/enabled time of read (:S) events"),
2583 OPT_CALLBACK('k', "clockid", &record.opts,
2584 "clockid", "clockid to use for events, see clock_gettime()",
2586 OPT_STRING_OPTARG('S', "snapshot", &record.opts.auxtrace_snapshot_opts,
2587 "opts", "AUX area tracing Snapshot Mode", ""),
2588 OPT_STRING_OPTARG(0, "aux-sample", &record.opts.auxtrace_sample_opts,
2589 "opts", "sample AUX area", ""),
2590 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
2591 "per thread proc mmap processing timeout in ms"),
2592 OPT_BOOLEAN(0, "namespaces", &record.opts.record_namespaces,
2593 "Record namespaces events"),
2594 OPT_BOOLEAN(0, "all-cgroups", &record.opts.record_cgroup,
2595 "Record cgroup events"),
2596 OPT_BOOLEAN_SET(0, "switch-events", &record.opts.record_switch_events,
2597 &record.opts.record_switch_events_set,
2598 "Record context switch events"),
2599 OPT_BOOLEAN_FLAG(0, "all-kernel", &record.opts.all_kernel,
2600 "Configure all used events to run in kernel space.",
2601 PARSE_OPT_EXCLUSIVE),
2602 OPT_BOOLEAN_FLAG(0, "all-user", &record.opts.all_user,
2603 "Configure all used events to run in user space.",
2604 PARSE_OPT_EXCLUSIVE),
2605 OPT_BOOLEAN(0, "kernel-callchains", &record.opts.kernel_callchains,
2606 "collect kernel callchains"),
2607 OPT_BOOLEAN(0, "user-callchains", &record.opts.user_callchains,
2608 "collect user callchains"),
2609 OPT_STRING(0, "clang-path", &llvm_param.clang_path, "clang path",
2610 "clang binary to use for compiling BPF scriptlets"),
2611 OPT_STRING(0, "clang-opt", &llvm_param.clang_opt, "clang options",
2612 "options passed to clang when compiling BPF scriptlets"),
2613 OPT_STRING(0, "vmlinux", &symbol_conf.vmlinux_name,
2614 "file", "vmlinux pathname"),
2615 OPT_BOOLEAN(0, "buildid-all", &record.buildid_all,
2616 "Record build-id of all DSOs regardless of hits"),
2617 OPT_BOOLEAN(0, "buildid-mmap", &record.buildid_mmap,
2618 "Record build-id in map events"),
2619 OPT_BOOLEAN(0, "timestamp-filename", &record.timestamp_filename,
2620 "append timestamp to output filename"),
2621 OPT_BOOLEAN(0, "timestamp-boundary", &record.timestamp_boundary,
2622 "Record timestamp boundary (time of first/last samples)"),
2623 OPT_STRING_OPTARG_SET(0, "switch-output", &record.switch_output.str,
2624 &record.switch_output.set, "signal or size[BKMG] or time[smhd]",
2625 "Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold",
2627 OPT_CALLBACK_SET(0, "switch-output-event", &record.sb_evlist, &record.switch_output_event_set, "switch output event",
2628 "switch output event selector. use 'perf list' to list available events",
2629 parse_events_option_new_evlist),
2630 OPT_INTEGER(0, "switch-max-files", &record.switch_output.num_files,
2631 "Limit number of switch output generated files"),
2632 OPT_BOOLEAN(0, "dry-run", &dry_run,
2633 "Parse options then exit"),
2634 #ifdef HAVE_AIO_SUPPORT
2635 OPT_CALLBACK_OPTARG(0, "aio", &record.opts,
2636 &nr_cblocks_default, "n", "Use <n> control blocks in asynchronous trace writing mode (default: 1, max: 4)",
2639 OPT_CALLBACK(0, "affinity", &record.opts, "node|cpu",
2640 "Set affinity mask of trace reading thread to NUMA node cpu mask or cpu of processed mmap buffer",
2641 record__parse_affinity),
2642 #ifdef HAVE_ZSTD_SUPPORT
2643 OPT_CALLBACK_OPTARG('z', "compression-level", &record.opts, &comp_level_default,
2644 "n", "Compressed records using specified level (default: 1 - fastest compression, 22 - greatest compression)",
2645 record__parse_comp_level),
2647 OPT_CALLBACK(0, "max-size", &record.output_max_size,
2648 "size", "Limit the maximum size of the output file", parse_output_max_size),
2649 OPT_UINTEGER(0, "num-thread-synthesize",
2650 &record.opts.nr_threads_synthesize,
2651 "number of threads to run for event synthesis"),
2653 OPT_CALLBACK(0, "pfm-events", &record.evlist, "event",
2654 "libpfm4 event selector. use 'perf list' to list available events",
2655 parse_libpfm_events_option),
2657 OPT_CALLBACK(0, "control", &record.opts, "fd:ctl-fd[,ack-fd] or fifo:ctl-fifo[,ack-fifo]",
2658 "Listen on ctl-fd descriptor for command to control measurement ('enable': enable events, 'disable': disable events,\n"
2659 "\t\t\t 'snapshot': AUX area tracing snapshot).\n"
2660 "\t\t\t Optionally send control command completion ('ack\\n') to ack-fd descriptor.\n"
2661 "\t\t\t Alternatively, ctl-fifo / ack-fifo will be opened and used as ctl-fd / ack-fd.",
2662 parse_control_option),
2666 struct option *record_options = __record_options;
2668 int cmd_record(int argc, const char **argv)
2671 struct record *rec = &record;
2672 char errbuf[BUFSIZ];
2674 setlocale(LC_ALL, "");
2676 #ifndef HAVE_LIBBPF_SUPPORT
2677 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
2678 set_nobuild('\0', "clang-path", true);
2679 set_nobuild('\0', "clang-opt", true);
2683 #ifndef HAVE_BPF_PROLOGUE
2684 # if !defined (HAVE_DWARF_SUPPORT)
2685 # define REASON "NO_DWARF=1"
2686 # elif !defined (HAVE_LIBBPF_SUPPORT)
2687 # define REASON "NO_LIBBPF=1"
2689 # define REASON "this architecture doesn't support BPF prologue"
2691 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
2692 set_nobuild('\0', "vmlinux", true);
2697 rec->opts.affinity = PERF_AFFINITY_SYS;
2699 rec->evlist = evlist__new();
2700 if (rec->evlist == NULL)
2703 err = perf_config(perf_record_config, rec);
2707 argc = parse_options(argc, argv, record_options, record_usage,
2708 PARSE_OPT_STOP_AT_NON_OPTION);
2710 perf_quiet_option();
2712 /* Make system wide (-a) the default target. */
2713 if (!argc && target__none(&rec->opts.target))
2714 rec->opts.target.system_wide = true;
2716 if (nr_cgroups && !rec->opts.target.system_wide) {
2717 usage_with_options_msg(record_usage, record_options,
2718 "cgroup monitoring only available in system-wide mode");
2722 if (rec->buildid_mmap) {
2723 if (!perf_can_record_build_id()) {
2724 pr_err("Failed: no support to record build id in mmap events, update your kernel.\n");
2728 pr_debug("Enabling build id in mmap2 events.\n");
2729 /* Enable mmap build id synthesizing. */
2730 symbol_conf.buildid_mmap2 = true;
2731 /* Enable perf_event_attr::build_id bit. */
2732 rec->opts.build_id = true;
2733 /* Disable build id cache. */
2734 rec->no_buildid = true;
2737 if (rec->opts.record_cgroup && !perf_can_record_cgroup()) {
2738 pr_err("Kernel has no cgroup sampling support.\n");
2743 if (rec->opts.kcore)
2744 rec->data.is_dir = true;
2746 if (rec->opts.comp_level != 0) {
2747 pr_debug("Compression enabled, disabling build id collection at the end of the session.\n");
2748 rec->no_buildid = true;
2751 if (rec->opts.record_switch_events &&
2752 !perf_can_record_switch_events()) {
2753 ui__error("kernel does not support recording context switch events\n");
2754 parse_options_usage(record_usage, record_options, "switch-events", 0);
2759 if (switch_output_setup(rec)) {
2760 parse_options_usage(record_usage, record_options, "switch-output", 0);
2765 if (rec->switch_output.time) {
2766 signal(SIGALRM, alarm_sig_handler);
2767 alarm(rec->switch_output.time);
2770 if (rec->switch_output.num_files) {
2771 rec->switch_output.filenames = calloc(sizeof(char *),
2772 rec->switch_output.num_files);
2773 if (!rec->switch_output.filenames) {
2780 * Allow aliases to facilitate the lookup of symbols for address
2781 * filters. Refer to auxtrace_parse_filters().
2783 symbol_conf.allow_aliases = true;
2787 if (rec->opts.affinity != PERF_AFFINITY_SYS) {
2788 rec->affinity_mask.nbits = cpu__max_cpu();
2789 rec->affinity_mask.bits = bitmap_alloc(rec->affinity_mask.nbits);
2790 if (!rec->affinity_mask.bits) {
2791 pr_err("Failed to allocate thread mask for %zd cpus\n", rec->affinity_mask.nbits);
2795 pr_debug2("thread mask[%zd]: empty\n", rec->affinity_mask.nbits);
2798 err = record__auxtrace_init(rec);
2805 err = bpf__setup_stdout(rec->evlist);
2807 bpf__strerror_setup_stdout(rec->evlist, err, errbuf, sizeof(errbuf));
2808 pr_err("ERROR: Setup BPF stdout failed: %s\n",
2815 if (rec->no_buildid_cache || rec->no_buildid) {
2816 disable_buildid_cache();
2817 } else if (rec->switch_output.enabled) {
2819 * In 'perf record --switch-output', disable buildid
2820 * generation by default to reduce data file switching
2821 * overhead. Still generate buildid if they are required
2824 * perf record --switch-output --no-no-buildid \
2825 * --no-no-buildid-cache
2827 * Following code equals to:
2829 * if ((rec->no_buildid || !rec->no_buildid_set) &&
2830 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
2831 * disable_buildid_cache();
2833 bool disable = true;
2835 if (rec->no_buildid_set && !rec->no_buildid)
2837 if (rec->no_buildid_cache_set && !rec->no_buildid_cache)
2840 rec->no_buildid = true;
2841 rec->no_buildid_cache = true;
2842 disable_buildid_cache();
2846 if (record.opts.overwrite)
2847 record.opts.tail_synthesize = true;
2849 if (rec->evlist->core.nr_entries == 0) {
2850 if (perf_pmu__has_hybrid()) {
2851 err = evlist__add_default_hybrid(rec->evlist,
2852 !record.opts.no_samples);
2854 err = __evlist__add_default(rec->evlist,
2855 !record.opts.no_samples);
2859 pr_err("Not enough memory for event selector list\n");
2864 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
2865 rec->opts.no_inherit = true;
2867 err = target__validate(&rec->opts.target);
2869 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
2870 ui__warning("%s\n", errbuf);
2873 err = target__parse_uid(&rec->opts.target);
2875 int saved_errno = errno;
2877 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
2878 ui__error("%s", errbuf);
2884 /* Enable ignoring missing threads when -u/-p option is defined. */
2885 rec->opts.ignore_missing_thread = rec->opts.target.uid != UINT_MAX || rec->opts.target.pid;
2888 if (evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
2889 usage_with_options(record_usage, record_options);
2891 err = auxtrace_record__options(rec->itr, rec->evlist, &rec->opts);
2896 * We take all buildids when the file contains
2897 * AUX area tracing data because we do not decode the
2898 * trace because it would take too long.
2900 if (rec->opts.full_auxtrace)
2901 rec->buildid_all = true;
2903 if (rec->opts.text_poke) {
2904 err = record__config_text_poke(rec->evlist);
2906 pr_err("record__config_text_poke failed, error %d\n", err);
2911 if (record_opts__config(&rec->opts)) {
2916 if (rec->opts.nr_cblocks > nr_cblocks_max)
2917 rec->opts.nr_cblocks = nr_cblocks_max;
2918 pr_debug("nr_cblocks: %d\n", rec->opts.nr_cblocks);
2920 pr_debug("affinity: %s\n", affinity_tags[rec->opts.affinity]);
2921 pr_debug("mmap flush: %d\n", rec->opts.mmap_flush);
2923 if (rec->opts.comp_level > comp_level_max)
2924 rec->opts.comp_level = comp_level_max;
2925 pr_debug("comp level: %d\n", rec->opts.comp_level);
2927 err = __cmd_record(&record, argc, argv);
2929 bitmap_free(rec->affinity_mask.bits);
2930 evlist__delete(rec->evlist);
2932 auxtrace_record__free(rec->itr);
2934 evlist__close_control(rec->opts.ctl_fd, rec->opts.ctl_fd_ack, &rec->opts.ctl_fd_close);
2938 static void snapshot_sig_handler(int sig __maybe_unused)
2940 struct record *rec = &record;
2942 hit_auxtrace_snapshot_trigger(rec);
2944 if (switch_output_signal(rec))
2945 trigger_hit(&switch_output_trigger);
2948 static void alarm_sig_handler(int sig __maybe_unused)
2950 struct record *rec = &record;
2952 if (switch_output_time(rec))
2953 trigger_hit(&switch_output_trigger);