4 * Builtin record command: Record the profile of a workload
5 * (or a CPU, or a PID) into the perf.data output file - for
6 * later analysis via perf report.
12 #include "util/build-id.h"
13 #include "util/util.h"
14 #include "util/parse-options.h"
15 #include "util/parse-events.h"
17 #include "util/callchain.h"
18 #include "util/cgroup.h"
19 #include "util/header.h"
20 #include "util/event.h"
21 #include "util/evlist.h"
22 #include "util/evsel.h"
23 #include "util/debug.h"
24 #include "util/session.h"
25 #include "util/tool.h"
26 #include "util/symbol.h"
27 #include "util/cpumap.h"
28 #include "util/thread_map.h"
29 #include "util/data.h"
37 struct perf_tool tool;
38 struct record_opts opts;
40 struct perf_data_file file;
41 struct perf_evlist *evlist;
42 struct perf_session *session;
46 bool no_buildid_cache;
50 static int record__write(struct record *rec, void *bf, size_t size)
52 if (perf_data_file__write(rec->session->file, bf, size) < 0) {
53 pr_err("failed to write perf data, error: %m\n");
57 rec->bytes_written += size;
61 static int process_synthesized_event(struct perf_tool *tool,
62 union perf_event *event,
63 struct perf_sample *sample __maybe_unused,
64 struct machine *machine __maybe_unused)
66 struct record *rec = container_of(tool, struct record, tool);
67 return record__write(rec, event, event->header.size);
70 static int record__mmap_read(struct record *rec, int idx)
72 struct perf_mmap *md = &rec->evlist->mmap[idx];
73 unsigned int head = perf_mmap__read_head(md);
74 unsigned int old = md->prev;
75 unsigned char *data = md->base + page_size;
87 if ((old & md->mask) + size != (head & md->mask)) {
88 buf = &data[old & md->mask];
89 size = md->mask + 1 - (old & md->mask);
92 if (record__write(rec, buf, size) < 0) {
98 buf = &data[old & md->mask];
102 if (record__write(rec, buf, size) < 0) {
108 perf_evlist__mmap_consume(rec->evlist, idx);
113 static volatile int done = 0;
114 static volatile int signr = -1;
115 static volatile int child_finished = 0;
117 static void sig_handler(int sig)
127 static void record__sig_exit(void)
132 signal(signr, SIG_DFL);
136 static int record__open(struct record *rec)
139 struct perf_evsel *pos;
140 struct perf_evlist *evlist = rec->evlist;
141 struct perf_session *session = rec->session;
142 struct record_opts *opts = &rec->opts;
145 perf_evlist__config(evlist, opts);
147 evlist__for_each(evlist, pos) {
149 if (perf_evsel__open(pos, evlist->cpus, evlist->threads) < 0) {
150 if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
152 ui__warning("%s\n", msg);
157 perf_evsel__open_strerror(pos, &opts->target,
158 errno, msg, sizeof(msg));
159 ui__error("%s\n", msg);
164 if (perf_evlist__apply_filters(evlist)) {
165 error("failed to set filter with %d (%s)\n", errno,
166 strerror_r(errno, msg, sizeof(msg)));
171 if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
172 if (errno == EPERM) {
173 pr_err("Permission error mapping pages.\n"
174 "Consider increasing "
175 "/proc/sys/kernel/perf_event_mlock_kb,\n"
176 "or try again with a smaller value of -m/--mmap_pages.\n"
177 "(current value: %u)\n", opts->mmap_pages);
180 pr_err("failed to mmap with %d (%s)\n", errno,
181 strerror_r(errno, msg, sizeof(msg)));
187 session->evlist = evlist;
188 perf_session__set_id_hdr_size(session);
193 static int process_sample_event(struct perf_tool *tool,
194 union perf_event *event,
195 struct perf_sample *sample,
196 struct perf_evsel *evsel,
197 struct machine *machine)
199 struct record *rec = container_of(tool, struct record, tool);
203 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
206 static int process_buildids(struct record *rec)
208 struct perf_data_file *file = &rec->file;
209 struct perf_session *session = rec->session;
211 u64 size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
218 * During this process, it'll load kernel map and replace the
219 * dso->long_name to a real pathname it found. In this case
220 * we prefer the vmlinux path like
221 * /lib/modules/3.16.4/build/vmlinux
223 * rather than build-id path (in debug directory).
224 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
226 symbol_conf.ignore_vmlinux_buildid = true;
228 return perf_session__process_events(session, &rec->tool);
231 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
234 struct perf_tool *tool = data;
236 *As for guest kernel when processing subcommand record&report,
237 *we arrange module mmap prior to guest kernel mmap and trigger
238 *a preload dso because default guest module symbols are loaded
239 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
240 *method is used to avoid symbol missing when the first addr is
241 *in module instead of in guest kernel.
243 err = perf_event__synthesize_modules(tool, process_synthesized_event,
246 pr_err("Couldn't record guest kernel [%d]'s reference"
247 " relocation symbol.\n", machine->pid);
250 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
251 * have no _text sometimes.
253 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
256 pr_err("Couldn't record guest kernel [%d]'s reference"
257 " relocation symbol.\n", machine->pid);
260 static struct perf_event_header finished_round_event = {
261 .size = sizeof(struct perf_event_header),
262 .type = PERF_RECORD_FINISHED_ROUND,
265 static int record__mmap_read_all(struct record *rec)
267 u64 bytes_written = rec->bytes_written;
271 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
272 if (rec->evlist->mmap[i].base) {
273 if (record__mmap_read(rec, i) != 0) {
281 * Mark the round finished in case we wrote
282 * at least one event.
284 if (bytes_written != rec->bytes_written)
285 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
291 static void record__init_features(struct record *rec)
293 struct perf_session *session = rec->session;
296 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
297 perf_header__set_feat(&session->header, feat);
300 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
302 if (!have_tracepoints(&rec->evlist->entries))
303 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
305 if (!rec->opts.branch_stack)
306 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
309 static volatile int workload_exec_errno;
312 * perf_evlist__prepare_workload will send a SIGUSR1
313 * if the fork fails, since we asked by setting its
314 * want_signal to true.
316 static void workload_exec_failed_signal(int signo __maybe_unused,
318 void *ucontext __maybe_unused)
320 workload_exec_errno = info->si_value.sival_int;
325 static int __cmd_record(struct record *rec, int argc, const char **argv)
329 unsigned long waking = 0;
330 const bool forks = argc > 0;
331 struct machine *machine;
332 struct perf_tool *tool = &rec->tool;
333 struct record_opts *opts = &rec->opts;
334 struct perf_data_file *file = &rec->file;
335 struct perf_session *session;
336 bool disabled = false, draining = false;
339 rec->progname = argv[0];
341 atexit(record__sig_exit);
342 signal(SIGCHLD, sig_handler);
343 signal(SIGINT, sig_handler);
344 signal(SIGTERM, sig_handler);
346 session = perf_session__new(file, false, NULL);
347 if (session == NULL) {
348 pr_err("Perf session creation failed.\n");
352 fd = perf_data_file__fd(file);
353 rec->session = session;
355 record__init_features(rec);
358 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
360 workload_exec_failed_signal);
362 pr_err("Couldn't run the workload!\n");
364 goto out_delete_session;
368 if (record__open(rec) != 0) {
373 if (!rec->evlist->nr_groups)
374 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
377 err = perf_header__write_pipe(fd);
381 err = perf_session__write_header(session, rec->evlist, fd, false);
387 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
388 pr_err("Couldn't generate buildids. "
389 "Use --no-buildid to profile anyway.\n");
394 machine = &session->machines.host;
397 err = perf_event__synthesize_attrs(tool, session,
398 process_synthesized_event);
400 pr_err("Couldn't synthesize attrs.\n");
404 if (have_tracepoints(&rec->evlist->entries)) {
406 * FIXME err <= 0 here actually means that
407 * there were no tracepoints so its not really
408 * an error, just that we don't need to
409 * synthesize anything. We really have to
410 * return this more properly and also
411 * propagate errors that now are calling die()
413 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
414 process_synthesized_event);
416 pr_err("Couldn't record tracing data.\n");
419 rec->bytes_written += err;
423 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
426 pr_err("Couldn't record kernel reference relocation symbol\n"
427 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
428 "Check /proc/kallsyms permission or run as root.\n");
430 err = perf_event__synthesize_modules(tool, process_synthesized_event,
433 pr_err("Couldn't record kernel module information.\n"
434 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
435 "Check /proc/modules permission or run as root.\n");
438 machines__process_guests(&session->machines,
439 perf_event__synthesize_guest_os, tool);
442 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
443 process_synthesized_event, opts->sample_address);
447 if (rec->realtime_prio) {
448 struct sched_param param;
450 param.sched_priority = rec->realtime_prio;
451 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
452 pr_err("Could not set realtime priority.\n");
459 * When perf is starting the traced process, all the events
460 * (apart from group members) have enable_on_exec=1 set,
461 * so don't spoil it by prematurely enabling them.
463 if (!target__none(&opts->target) && !opts->initial_delay)
464 perf_evlist__enable(rec->evlist);
470 perf_evlist__start_workload(rec->evlist);
472 if (opts->initial_delay) {
473 usleep(opts->initial_delay * 1000);
474 perf_evlist__enable(rec->evlist);
478 int hits = rec->samples;
480 if (record__mmap_read_all(rec) < 0) {
485 if (hits == rec->samples) {
486 if (done || draining)
488 err = perf_evlist__poll(rec->evlist, -1);
490 * Propagate error, only if there's any. Ignore positive
491 * number of returned events and interrupt error.
493 if (err > 0 || (err < 0 && errno == EINTR))
497 if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
502 * When perf is starting the traced process, at the end events
503 * die with the process and we wait for that. Thus no need to
504 * disable events in this case.
506 if (done && !disabled && !target__none(&opts->target)) {
507 perf_evlist__disable(rec->evlist);
512 if (forks && workload_exec_errno) {
513 char msg[STRERR_BUFSIZE];
514 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
515 pr_err("Workload failed: %s\n", emsg);
521 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
528 kill(rec->evlist->workload.pid, SIGTERM);
534 else if (WIFEXITED(exit_status))
535 status = WEXITSTATUS(exit_status);
536 else if (WIFSIGNALED(exit_status))
537 signr = WTERMSIG(exit_status);
541 /* this will be recalculated during process_buildids() */
544 if (!err && !file->is_pipe) {
545 rec->session->header.data_size += rec->bytes_written;
547 if (!rec->no_buildid)
548 process_buildids(rec);
549 perf_session__write_header(rec->session, rec->evlist, fd, true);
552 if (!err && !quiet) {
556 scnprintf(samples, sizeof(samples),
557 " (%" PRIu64 " samples)", rec->samples);
561 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
562 perf_data_file__size(file) / 1024.0 / 1024.0,
563 file->path, samples);
567 perf_session__delete(session);
571 #define BRANCH_OPT(n, m) \
572 { .name = n, .mode = (m) }
574 #define BRANCH_END { .name = NULL }
581 static const struct branch_mode branch_modes[] = {
582 BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
583 BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
584 BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
585 BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
586 BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
587 BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
588 BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
589 BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
590 BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
591 BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
592 BRANCH_OPT("cond", PERF_SAMPLE_BRANCH_COND),
597 parse_branch_stack(const struct option *opt, const char *str, int unset)
600 (PERF_SAMPLE_BRANCH_USER |\
601 PERF_SAMPLE_BRANCH_KERNEL |\
602 PERF_SAMPLE_BRANCH_HV)
604 uint64_t *mode = (uint64_t *)opt->value;
605 const struct branch_mode *br;
606 char *s, *os = NULL, *p;
613 * cannot set it twice, -b + --branch-filter for instance
618 /* str may be NULL in case no arg is passed to -b */
620 /* because str is read-only */
621 s = os = strdup(str);
630 for (br = branch_modes; br->name; br++) {
631 if (!strcasecmp(s, br->name))
635 ui__warning("unknown branch filter %s,"
636 " check man page\n", s);
650 /* default to any branch */
651 if ((*mode & ~ONLY_PLM) == 0) {
652 *mode = PERF_SAMPLE_BRANCH_ANY;
659 static void callchain_debug(void)
661 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
663 pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
665 if (callchain_param.record_mode == CALLCHAIN_DWARF)
666 pr_debug("callchain: stack dump size %d\n",
667 callchain_param.dump_size);
670 int record_parse_callchain_opt(const struct option *opt __maybe_unused,
676 callchain_param.enabled = !unset;
678 /* --no-call-graph */
680 callchain_param.record_mode = CALLCHAIN_NONE;
681 pr_debug("callchain: disabled\n");
685 ret = parse_callchain_record_opt(arg);
692 int record_callchain_opt(const struct option *opt __maybe_unused,
693 const char *arg __maybe_unused,
694 int unset __maybe_unused)
696 callchain_param.enabled = true;
698 if (callchain_param.record_mode == CALLCHAIN_NONE)
699 callchain_param.record_mode = CALLCHAIN_FP;
705 static int perf_record_config(const char *var, const char *value, void *cb)
707 if (!strcmp(var, "record.call-graph"))
708 var = "call-graph.record-mode"; /* fall-through */
710 return perf_default_config(var, value, cb);
713 static const char * const __record_usage[] = {
714 "perf record [<options>] [<command>]",
715 "perf record [<options>] -- <command> [<options>]",
718 const char * const *record_usage = __record_usage;
721 * XXX Ideally would be local to cmd_record() and passed to a record__new
722 * because we need to have access to it in record__exit, that is called
723 * after cmd_record() exits, but since record_options need to be accessible to
724 * builtin-script, leave it here.
726 * At least we don't ouch it in all the other functions here directly.
728 * Just say no to tons of global variables, sigh.
730 static struct record record = {
733 .mmap_pages = UINT_MAX,
734 .user_freq = UINT_MAX,
735 .user_interval = ULLONG_MAX,
739 .default_per_cpu = true,
743 .sample = process_sample_event,
744 .fork = perf_event__process_fork,
745 .comm = perf_event__process_comm,
746 .mmap = perf_event__process_mmap,
747 .mmap2 = perf_event__process_mmap2,
751 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
753 #ifdef HAVE_DWARF_UNWIND_SUPPORT
754 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf lbr";
756 const char record_callchain_help[] = CALLCHAIN_HELP "fp lbr";
760 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
761 * with it and switch to use the library functions in perf_evlist that came
762 * from builtin-record.c, i.e. use record_opts,
763 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
766 struct option __record_options[] = {
767 OPT_CALLBACK('e', "event", &record.evlist, "event",
768 "event selector. use 'perf list' to list available events",
769 parse_events_option),
770 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
771 "event filter", parse_filter),
772 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
773 "record events on existing process id"),
774 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
775 "record events on existing thread id"),
776 OPT_INTEGER('r', "realtime", &record.realtime_prio,
777 "collect data with this RT SCHED_FIFO priority"),
778 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
779 "collect data without buffering"),
780 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
781 "collect raw sample records from all opened counters"),
782 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
783 "system-wide collection from all CPUs"),
784 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
785 "list of cpus to monitor"),
786 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
787 OPT_STRING('o', "output", &record.file.path, "file",
789 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
790 &record.opts.no_inherit_set,
791 "child tasks do not inherit counters"),
792 OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
793 OPT_CALLBACK('m', "mmap-pages", &record.opts.mmap_pages, "pages",
794 "number of mmap data pages",
795 perf_evlist__parse_mmap_pages),
796 OPT_BOOLEAN(0, "group", &record.opts.group,
797 "put the counters into a counter group"),
798 OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
799 NULL, "enables call-graph recording" ,
800 &record_callchain_opt),
801 OPT_CALLBACK(0, "call-graph", &record.opts,
802 "mode[,dump_size]", record_callchain_help,
803 &record_parse_callchain_opt),
804 OPT_INCR('v', "verbose", &verbose,
805 "be more verbose (show counter open errors, etc)"),
806 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
807 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
808 "per thread counts"),
809 OPT_BOOLEAN('d', "data", &record.opts.sample_address,
811 OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
812 OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
813 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
815 OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
816 "do not update the buildid cache"),
817 OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
818 "do not collect buildids in perf.data"),
819 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
820 "monitor event in cgroup name only",
822 OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
823 "ms to wait before starting measurement after program start"),
824 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
827 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
828 "branch any", "sample any taken branches",
831 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
832 "branch filter mask", "branch stack filter modes",
834 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
835 "sample by weight (on special events only)"),
836 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
837 "sample transaction flags (special events only)"),
838 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
839 "use per-thread mmaps"),
840 OPT_BOOLEAN('I', "intr-regs", &record.opts.sample_intr_regs,
841 "Sample machine registers on interrupt"),
842 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
843 "Record running/enabled time of read (:S) events"),
847 struct option *record_options = __record_options;
849 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
852 struct record *rec = &record;
855 rec->evlist = perf_evlist__new();
856 if (rec->evlist == NULL)
859 perf_config(perf_record_config, rec);
861 argc = parse_options(argc, argv, record_options, record_usage,
862 PARSE_OPT_STOP_AT_NON_OPTION);
863 if (!argc && target__none(&rec->opts.target))
864 usage_with_options(record_usage, record_options);
866 if (nr_cgroups && !rec->opts.target.system_wide) {
867 ui__error("cgroup monitoring only available in"
868 " system-wide mode\n");
869 usage_with_options(record_usage, record_options);
874 if (symbol_conf.kptr_restrict)
876 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
877 "check /proc/sys/kernel/kptr_restrict.\n\n"
878 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
879 "file is not found in the buildid cache or in the vmlinux path.\n\n"
880 "Samples in kernel modules won't be resolved at all.\n\n"
881 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
882 "even with a suitable vmlinux or kallsyms file.\n\n");
884 if (rec->no_buildid_cache || rec->no_buildid)
885 disable_buildid_cache();
887 if (rec->evlist->nr_entries == 0 &&
888 perf_evlist__add_default(rec->evlist) < 0) {
889 pr_err("Not enough memory for event selector list\n");
890 goto out_symbol_exit;
893 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
894 rec->opts.no_inherit = true;
896 err = target__validate(&rec->opts.target);
898 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
899 ui__warning("%s", errbuf);
902 err = target__parse_uid(&rec->opts.target);
904 int saved_errno = errno;
906 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
907 ui__error("%s", errbuf);
910 goto out_symbol_exit;
914 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
915 usage_with_options(record_usage, record_options);
917 if (record_opts__config(&rec->opts)) {
919 goto out_symbol_exit;
922 err = __cmd_record(&record, argc, argv);
924 perf_evlist__delete(rec->evlist);