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, &pos)) {
165 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
166 pos->filter, perf_evsel__name(pos), errno,
167 strerror_r(errno, msg, sizeof(msg)));
172 if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
173 if (errno == EPERM) {
174 pr_err("Permission error mapping pages.\n"
175 "Consider increasing "
176 "/proc/sys/kernel/perf_event_mlock_kb,\n"
177 "or try again with a smaller value of -m/--mmap_pages.\n"
178 "(current value: %u)\n", opts->mmap_pages);
181 pr_err("failed to mmap with %d (%s)\n", errno,
182 strerror_r(errno, msg, sizeof(msg)));
188 session->evlist = evlist;
189 perf_session__set_id_hdr_size(session);
194 static int process_sample_event(struct perf_tool *tool,
195 union perf_event *event,
196 struct perf_sample *sample,
197 struct perf_evsel *evsel,
198 struct machine *machine)
200 struct record *rec = container_of(tool, struct record, tool);
204 return build_id__mark_dso_hit(tool, event, sample, evsel, machine);
207 static int process_buildids(struct record *rec)
209 struct perf_data_file *file = &rec->file;
210 struct perf_session *session = rec->session;
212 u64 size = lseek(perf_data_file__fd(file), 0, SEEK_CUR);
219 * During this process, it'll load kernel map and replace the
220 * dso->long_name to a real pathname it found. In this case
221 * we prefer the vmlinux path like
222 * /lib/modules/3.16.4/build/vmlinux
224 * rather than build-id path (in debug directory).
225 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
227 symbol_conf.ignore_vmlinux_buildid = true;
229 return perf_session__process_events(session);
232 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
235 struct perf_tool *tool = data;
237 *As for guest kernel when processing subcommand record&report,
238 *we arrange module mmap prior to guest kernel mmap and trigger
239 *a preload dso because default guest module symbols are loaded
240 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
241 *method is used to avoid symbol missing when the first addr is
242 *in module instead of in guest kernel.
244 err = perf_event__synthesize_modules(tool, process_synthesized_event,
247 pr_err("Couldn't record guest kernel [%d]'s reference"
248 " relocation symbol.\n", machine->pid);
251 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
252 * have no _text sometimes.
254 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
257 pr_err("Couldn't record guest kernel [%d]'s reference"
258 " relocation symbol.\n", machine->pid);
261 static struct perf_event_header finished_round_event = {
262 .size = sizeof(struct perf_event_header),
263 .type = PERF_RECORD_FINISHED_ROUND,
266 static int record__mmap_read_all(struct record *rec)
268 u64 bytes_written = rec->bytes_written;
272 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
273 if (rec->evlist->mmap[i].base) {
274 if (record__mmap_read(rec, i) != 0) {
282 * Mark the round finished in case we wrote
283 * at least one event.
285 if (bytes_written != rec->bytes_written)
286 rc = record__write(rec, &finished_round_event, sizeof(finished_round_event));
292 static void record__init_features(struct record *rec)
294 struct perf_session *session = rec->session;
297 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
298 perf_header__set_feat(&session->header, feat);
301 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
303 if (!have_tracepoints(&rec->evlist->entries))
304 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
306 if (!rec->opts.branch_stack)
307 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
310 static volatile int workload_exec_errno;
313 * perf_evlist__prepare_workload will send a SIGUSR1
314 * if the fork fails, since we asked by setting its
315 * want_signal to true.
317 static void workload_exec_failed_signal(int signo __maybe_unused,
319 void *ucontext __maybe_unused)
321 workload_exec_errno = info->si_value.sival_int;
326 static int __cmd_record(struct record *rec, int argc, const char **argv)
330 unsigned long waking = 0;
331 const bool forks = argc > 0;
332 struct machine *machine;
333 struct perf_tool *tool = &rec->tool;
334 struct record_opts *opts = &rec->opts;
335 struct perf_data_file *file = &rec->file;
336 struct perf_session *session;
337 bool disabled = false, draining = false;
340 rec->progname = argv[0];
342 atexit(record__sig_exit);
343 signal(SIGCHLD, sig_handler);
344 signal(SIGINT, sig_handler);
345 signal(SIGTERM, sig_handler);
347 session = perf_session__new(file, false, tool);
348 if (session == NULL) {
349 pr_err("Perf session creation failed.\n");
353 fd = perf_data_file__fd(file);
354 rec->session = session;
356 record__init_features(rec);
359 err = perf_evlist__prepare_workload(rec->evlist, &opts->target,
361 workload_exec_failed_signal);
363 pr_err("Couldn't run the workload!\n");
365 goto out_delete_session;
369 if (record__open(rec) != 0) {
374 if (!rec->evlist->nr_groups)
375 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
378 err = perf_header__write_pipe(fd);
382 err = perf_session__write_header(session, rec->evlist, fd, false);
388 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
389 pr_err("Couldn't generate buildids. "
390 "Use --no-buildid to profile anyway.\n");
395 machine = &session->machines.host;
398 err = perf_event__synthesize_attrs(tool, session,
399 process_synthesized_event);
401 pr_err("Couldn't synthesize attrs.\n");
405 if (have_tracepoints(&rec->evlist->entries)) {
407 * FIXME err <= 0 here actually means that
408 * there were no tracepoints so its not really
409 * an error, just that we don't need to
410 * synthesize anything. We really have to
411 * return this more properly and also
412 * propagate errors that now are calling die()
414 err = perf_event__synthesize_tracing_data(tool, fd, rec->evlist,
415 process_synthesized_event);
417 pr_err("Couldn't record tracing data.\n");
420 rec->bytes_written += err;
424 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
427 pr_err("Couldn't record kernel reference relocation symbol\n"
428 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
429 "Check /proc/kallsyms permission or run as root.\n");
431 err = perf_event__synthesize_modules(tool, process_synthesized_event,
434 pr_err("Couldn't record kernel module information.\n"
435 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
436 "Check /proc/modules permission or run as root.\n");
439 machines__process_guests(&session->machines,
440 perf_event__synthesize_guest_os, tool);
443 err = __machine__synthesize_threads(machine, tool, &opts->target, rec->evlist->threads,
444 process_synthesized_event, opts->sample_address);
448 if (rec->realtime_prio) {
449 struct sched_param param;
451 param.sched_priority = rec->realtime_prio;
452 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
453 pr_err("Could not set realtime priority.\n");
460 * When perf is starting the traced process, all the events
461 * (apart from group members) have enable_on_exec=1 set,
462 * so don't spoil it by prematurely enabling them.
464 if (!target__none(&opts->target) && !opts->initial_delay)
465 perf_evlist__enable(rec->evlist);
471 perf_evlist__start_workload(rec->evlist);
473 if (opts->initial_delay) {
474 usleep(opts->initial_delay * 1000);
475 perf_evlist__enable(rec->evlist);
479 int hits = rec->samples;
481 if (record__mmap_read_all(rec) < 0) {
486 if (hits == rec->samples) {
487 if (done || draining)
489 err = perf_evlist__poll(rec->evlist, -1);
491 * Propagate error, only if there's any. Ignore positive
492 * number of returned events and interrupt error.
494 if (err > 0 || (err < 0 && errno == EINTR))
498 if (perf_evlist__filter_pollfd(rec->evlist, POLLERR | POLLHUP) == 0)
503 * When perf is starting the traced process, at the end events
504 * die with the process and we wait for that. Thus no need to
505 * disable events in this case.
507 if (done && !disabled && !target__none(&opts->target)) {
508 perf_evlist__disable(rec->evlist);
513 if (forks && workload_exec_errno) {
514 char msg[STRERR_BUFSIZE];
515 const char *emsg = strerror_r(workload_exec_errno, msg, sizeof(msg));
516 pr_err("Workload failed: %s\n", emsg);
522 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
529 kill(rec->evlist->workload.pid, SIGTERM);
535 else if (WIFEXITED(exit_status))
536 status = WEXITSTATUS(exit_status);
537 else if (WIFSIGNALED(exit_status))
538 signr = WTERMSIG(exit_status);
542 /* this will be recalculated during process_buildids() */
545 if (!err && !file->is_pipe) {
546 rec->session->header.data_size += rec->bytes_written;
548 if (!rec->no_buildid)
549 process_buildids(rec);
550 perf_session__write_header(rec->session, rec->evlist, fd, true);
553 if (!err && !quiet) {
557 scnprintf(samples, sizeof(samples),
558 " (%" PRIu64 " samples)", rec->samples);
562 fprintf(stderr, "[ perf record: Captured and wrote %.3f MB %s%s ]\n",
563 perf_data_file__size(file) / 1024.0 / 1024.0,
564 file->path, samples);
568 perf_session__delete(session);
572 #define BRANCH_OPT(n, m) \
573 { .name = n, .mode = (m) }
575 #define BRANCH_END { .name = NULL }
582 static const struct branch_mode branch_modes[] = {
583 BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
584 BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
585 BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
586 BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
587 BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
588 BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
589 BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
590 BRANCH_OPT("abort_tx", PERF_SAMPLE_BRANCH_ABORT_TX),
591 BRANCH_OPT("in_tx", PERF_SAMPLE_BRANCH_IN_TX),
592 BRANCH_OPT("no_tx", PERF_SAMPLE_BRANCH_NO_TX),
593 BRANCH_OPT("cond", PERF_SAMPLE_BRANCH_COND),
598 parse_branch_stack(const struct option *opt, const char *str, int unset)
601 (PERF_SAMPLE_BRANCH_USER |\
602 PERF_SAMPLE_BRANCH_KERNEL |\
603 PERF_SAMPLE_BRANCH_HV)
605 uint64_t *mode = (uint64_t *)opt->value;
606 const struct branch_mode *br;
607 char *s, *os = NULL, *p;
614 * cannot set it twice, -b + --branch-filter for instance
619 /* str may be NULL in case no arg is passed to -b */
621 /* because str is read-only */
622 s = os = strdup(str);
631 for (br = branch_modes; br->name; br++) {
632 if (!strcasecmp(s, br->name))
636 ui__warning("unknown branch filter %s,"
637 " check man page\n", s);
651 /* default to any branch */
652 if ((*mode & ~ONLY_PLM) == 0) {
653 *mode = PERF_SAMPLE_BRANCH_ANY;
660 static void callchain_debug(void)
662 static const char *str[CALLCHAIN_MAX] = { "NONE", "FP", "DWARF", "LBR" };
664 pr_debug("callchain: type %s\n", str[callchain_param.record_mode]);
666 if (callchain_param.record_mode == CALLCHAIN_DWARF)
667 pr_debug("callchain: stack dump size %d\n",
668 callchain_param.dump_size);
671 int record_parse_callchain_opt(const struct option *opt __maybe_unused,
677 callchain_param.enabled = !unset;
679 /* --no-call-graph */
681 callchain_param.record_mode = CALLCHAIN_NONE;
682 pr_debug("callchain: disabled\n");
686 ret = parse_callchain_record_opt(arg);
693 int record_callchain_opt(const struct option *opt __maybe_unused,
694 const char *arg __maybe_unused,
695 int unset __maybe_unused)
697 callchain_param.enabled = true;
699 if (callchain_param.record_mode == CALLCHAIN_NONE)
700 callchain_param.record_mode = CALLCHAIN_FP;
706 static int perf_record_config(const char *var, const char *value, void *cb)
708 if (!strcmp(var, "record.call-graph"))
709 var = "call-graph.record-mode"; /* fall-through */
711 return perf_default_config(var, value, cb);
714 static const char * const __record_usage[] = {
715 "perf record [<options>] [<command>]",
716 "perf record [<options>] -- <command> [<options>]",
719 const char * const *record_usage = __record_usage;
722 * XXX Ideally would be local to cmd_record() and passed to a record__new
723 * because we need to have access to it in record__exit, that is called
724 * after cmd_record() exits, but since record_options need to be accessible to
725 * builtin-script, leave it here.
727 * At least we don't ouch it in all the other functions here directly.
729 * Just say no to tons of global variables, sigh.
731 static struct record record = {
734 .mmap_pages = UINT_MAX,
735 .user_freq = UINT_MAX,
736 .user_interval = ULLONG_MAX,
740 .default_per_cpu = true,
744 .sample = process_sample_event,
745 .fork = perf_event__process_fork,
746 .comm = perf_event__process_comm,
747 .mmap = perf_event__process_mmap,
748 .mmap2 = perf_event__process_mmap2,
752 #define CALLCHAIN_HELP "setup and enables call-graph (stack chain/backtrace) recording: "
754 #ifdef HAVE_DWARF_UNWIND_SUPPORT
755 const char record_callchain_help[] = CALLCHAIN_HELP "fp dwarf lbr";
757 const char record_callchain_help[] = CALLCHAIN_HELP "fp lbr";
761 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
762 * with it and switch to use the library functions in perf_evlist that came
763 * from builtin-record.c, i.e. use record_opts,
764 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
767 struct option __record_options[] = {
768 OPT_CALLBACK('e', "event", &record.evlist, "event",
769 "event selector. use 'perf list' to list available events",
770 parse_events_option),
771 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
772 "event filter", parse_filter),
773 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
774 "record events on existing process id"),
775 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
776 "record events on existing thread id"),
777 OPT_INTEGER('r', "realtime", &record.realtime_prio,
778 "collect data with this RT SCHED_FIFO priority"),
779 OPT_BOOLEAN(0, "no-buffering", &record.opts.no_buffering,
780 "collect data without buffering"),
781 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
782 "collect raw sample records from all opened counters"),
783 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
784 "system-wide collection from all CPUs"),
785 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
786 "list of cpus to monitor"),
787 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
788 OPT_STRING('o', "output", &record.file.path, "file",
790 OPT_BOOLEAN_SET('i', "no-inherit", &record.opts.no_inherit,
791 &record.opts.no_inherit_set,
792 "child tasks do not inherit counters"),
793 OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
794 OPT_CALLBACK('m', "mmap-pages", &record.opts.mmap_pages, "pages",
795 "number of mmap data pages",
796 perf_evlist__parse_mmap_pages),
797 OPT_BOOLEAN(0, "group", &record.opts.group,
798 "put the counters into a counter group"),
799 OPT_CALLBACK_NOOPT('g', NULL, &record.opts,
800 NULL, "enables call-graph recording" ,
801 &record_callchain_opt),
802 OPT_CALLBACK(0, "call-graph", &record.opts,
803 "mode[,dump_size]", record_callchain_help,
804 &record_parse_callchain_opt),
805 OPT_INCR('v', "verbose", &verbose,
806 "be more verbose (show counter open errors, etc)"),
807 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
808 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
809 "per thread counts"),
810 OPT_BOOLEAN('d', "data", &record.opts.sample_address,
812 OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
813 OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
814 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
816 OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
817 "do not update the buildid cache"),
818 OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
819 "do not collect buildids in perf.data"),
820 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
821 "monitor event in cgroup name only",
823 OPT_UINTEGER('D', "delay", &record.opts.initial_delay,
824 "ms to wait before starting measurement after program start"),
825 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
828 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
829 "branch any", "sample any taken branches",
832 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
833 "branch filter mask", "branch stack filter modes",
835 OPT_BOOLEAN('W', "weight", &record.opts.sample_weight,
836 "sample by weight (on special events only)"),
837 OPT_BOOLEAN(0, "transaction", &record.opts.sample_transaction,
838 "sample transaction flags (special events only)"),
839 OPT_BOOLEAN(0, "per-thread", &record.opts.target.per_thread,
840 "use per-thread mmaps"),
841 OPT_BOOLEAN('I', "intr-regs", &record.opts.sample_intr_regs,
842 "Sample machine registers on interrupt"),
843 OPT_BOOLEAN(0, "running-time", &record.opts.running_time,
844 "Record running/enabled time of read (:S) events"),
848 struct option *record_options = __record_options;
850 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
853 struct record *rec = &record;
856 rec->evlist = perf_evlist__new();
857 if (rec->evlist == NULL)
860 perf_config(perf_record_config, rec);
862 argc = parse_options(argc, argv, record_options, record_usage,
863 PARSE_OPT_STOP_AT_NON_OPTION);
864 if (!argc && target__none(&rec->opts.target))
865 usage_with_options(record_usage, record_options);
867 if (nr_cgroups && !rec->opts.target.system_wide) {
868 ui__error("cgroup monitoring only available in"
869 " system-wide mode\n");
870 usage_with_options(record_usage, record_options);
875 if (symbol_conf.kptr_restrict)
877 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
878 "check /proc/sys/kernel/kptr_restrict.\n\n"
879 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
880 "file is not found in the buildid cache or in the vmlinux path.\n\n"
881 "Samples in kernel modules won't be resolved at all.\n\n"
882 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
883 "even with a suitable vmlinux or kallsyms file.\n\n");
885 if (rec->no_buildid_cache || rec->no_buildid)
886 disable_buildid_cache();
888 if (rec->evlist->nr_entries == 0 &&
889 perf_evlist__add_default(rec->evlist) < 0) {
890 pr_err("Not enough memory for event selector list\n");
891 goto out_symbol_exit;
894 if (rec->opts.target.tid && !rec->opts.no_inherit_set)
895 rec->opts.no_inherit = true;
897 err = target__validate(&rec->opts.target);
899 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
900 ui__warning("%s", errbuf);
903 err = target__parse_uid(&rec->opts.target);
905 int saved_errno = errno;
907 target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
908 ui__error("%s", errbuf);
911 goto out_symbol_exit;
915 if (perf_evlist__create_maps(rec->evlist, &rec->opts.target) < 0)
916 usage_with_options(record_usage, record_options);
918 if (record_opts__config(&rec->opts)) {
920 goto out_symbol_exit;
923 err = __cmd_record(&record, argc, argv);
925 perf_evlist__delete(rec->evlist);