4 * Builtin record command: Record the profile of a workload
5 * (or a CPU, or a PID) into the perf.data output file - for
6 * later analysis via perf report.
12 #include "util/build-id.h"
13 #include "util/util.h"
14 #include <subcmd/parse-options.h>
15 #include "util/parse-events.h"
16 #include "util/config.h"
18 #include "util/callchain.h"
19 #include "util/cgroup.h"
20 #include "util/header.h"
21 #include "util/event.h"
22 #include "util/evlist.h"
23 #include "util/evsel.h"
24 #include "util/debug.h"
25 #include "util/session.h"
26 #include "util/tool.h"
27 #include "util/symbol.h"
28 #include "util/cpumap.h"
29 #include "util/thread_map.h"
30 #include "util/data.h"
31 #include "util/perf_regs.h"
32 #include "util/auxtrace.h"
34 #include "util/parse-branch-options.h"
35 #include "util/parse-regs-options.h"
36 #include "util/llvm-utils.h"
37 #include "util/bpf-loader.h"
38 #include "util/trigger.h"
48 struct perf_tool tool
;
49 struct record_opts opts
;
51 struct perf_data_file file
;
52 struct auxtrace_record
*itr
;
53 struct perf_evlist
*evlist
;
54 struct perf_session
*session
;
59 bool no_buildid_cache
;
60 bool no_buildid_cache_set
;
62 bool timestamp_filename
;
64 unsigned long long samples
;
67 static int record__write(struct record
*rec
, void *bf
, size_t size
)
69 if (perf_data_file__write(rec
->session
->file
, bf
, size
) < 0) {
70 pr_err("failed to write perf data, error: %m\n");
74 rec
->bytes_written
+= size
;
78 static int process_synthesized_event(struct perf_tool
*tool
,
79 union perf_event
*event
,
80 struct perf_sample
*sample __maybe_unused
,
81 struct machine
*machine __maybe_unused
)
83 struct record
*rec
= container_of(tool
, struct record
, tool
);
84 return record__write(rec
, event
, event
->header
.size
);
88 backward_rb_find_range(void *buf
, int mask
, u64 head
, u64
*start
, u64
*end
)
90 struct perf_event_header
*pheader
;
94 pr_debug2("backward_rb_find_range: buf=%p, head=%"PRIx64
"\n", buf
, head
);
95 pheader
= (struct perf_event_header
*)(buf
+ (head
& mask
));
98 if (evt_head
- head
>= (unsigned int)size
) {
99 pr_debug("Finshed reading backward ring buffer: rewind\n");
100 if (evt_head
- head
> (unsigned int)size
)
101 evt_head
-= pheader
->size
;
106 pheader
= (struct perf_event_header
*)(buf
+ (evt_head
& mask
));
108 if (pheader
->size
== 0) {
109 pr_debug("Finshed reading backward ring buffer: get start\n");
114 evt_head
+= pheader
->size
;
115 pr_debug3("move evt_head: %"PRIx64
"\n", evt_head
);
117 WARN_ONCE(1, "Shouldn't get here\n");
122 rb_find_range(struct perf_evlist
*evlist
,
123 void *data
, int mask
, u64 head
, u64 old
,
124 u64
*start
, u64
*end
)
126 if (!evlist
->backward
) {
132 return backward_rb_find_range(data
, mask
, head
, start
, end
);
135 static int record__mmap_read(struct record
*rec
, int idx
)
137 struct perf_mmap
*md
= &rec
->evlist
->mmap
[idx
];
138 u64 head
= perf_mmap__read_head(md
);
140 u64 end
= head
, start
= old
;
141 unsigned char *data
= md
->base
+ page_size
;
146 if (rb_find_range(rec
->evlist
, data
, md
->mask
, head
,
156 if (size
> (unsigned long)(md
->mask
) + 1) {
157 WARN_ONCE(1, "failed to keep up with mmap data. (warn only once)\n");
160 perf_evlist__mmap_consume(rec
->evlist
, idx
);
164 if ((start
& md
->mask
) + size
!= (end
& md
->mask
)) {
165 buf
= &data
[start
& md
->mask
];
166 size
= md
->mask
+ 1 - (start
& md
->mask
);
169 if (record__write(rec
, buf
, size
) < 0) {
175 buf
= &data
[start
& md
->mask
];
179 if (record__write(rec
, buf
, size
) < 0) {
185 perf_evlist__mmap_consume(rec
->evlist
, idx
);
190 static volatile int done
;
191 static volatile int signr
= -1;
192 static volatile int child_finished
;
194 static volatile int auxtrace_record__snapshot_started
;
195 static DEFINE_TRIGGER(auxtrace_snapshot_trigger
);
196 static DEFINE_TRIGGER(switch_output_trigger
);
198 static void sig_handler(int sig
)
208 static void record__sig_exit(void)
213 signal(signr
, SIG_DFL
);
217 #ifdef HAVE_AUXTRACE_SUPPORT
219 static int record__process_auxtrace(struct perf_tool
*tool
,
220 union perf_event
*event
, void *data1
,
221 size_t len1
, void *data2
, size_t len2
)
223 struct record
*rec
= container_of(tool
, struct record
, tool
);
224 struct perf_data_file
*file
= &rec
->file
;
228 if (!perf_data_file__is_pipe(file
)) {
230 int fd
= perf_data_file__fd(file
);
233 file_offset
= lseek(fd
, 0, SEEK_CUR
);
234 if (file_offset
== -1)
236 err
= auxtrace_index__auxtrace_event(&rec
->session
->auxtrace_index
,
242 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
243 padding
= (len1
+ len2
) & 7;
245 padding
= 8 - padding
;
247 record__write(rec
, event
, event
->header
.size
);
248 record__write(rec
, data1
, len1
);
250 record__write(rec
, data2
, len2
);
251 record__write(rec
, &pad
, padding
);
256 static int record__auxtrace_mmap_read(struct record
*rec
,
257 struct auxtrace_mmap
*mm
)
261 ret
= auxtrace_mmap__read(mm
, rec
->itr
, &rec
->tool
,
262 record__process_auxtrace
);
272 static int record__auxtrace_mmap_read_snapshot(struct record
*rec
,
273 struct auxtrace_mmap
*mm
)
277 ret
= auxtrace_mmap__read_snapshot(mm
, rec
->itr
, &rec
->tool
,
278 record__process_auxtrace
,
279 rec
->opts
.auxtrace_snapshot_size
);
289 static int record__auxtrace_read_snapshot_all(struct record
*rec
)
294 for (i
= 0; i
< rec
->evlist
->nr_mmaps
; i
++) {
295 struct auxtrace_mmap
*mm
=
296 &rec
->evlist
->mmap
[i
].auxtrace_mmap
;
301 if (record__auxtrace_mmap_read_snapshot(rec
, mm
) != 0) {
310 static void record__read_auxtrace_snapshot(struct record
*rec
)
312 pr_debug("Recording AUX area tracing snapshot\n");
313 if (record__auxtrace_read_snapshot_all(rec
) < 0) {
314 trigger_error(&auxtrace_snapshot_trigger
);
316 if (auxtrace_record__snapshot_finish(rec
->itr
))
317 trigger_error(&auxtrace_snapshot_trigger
);
319 trigger_ready(&auxtrace_snapshot_trigger
);
326 int record__auxtrace_mmap_read(struct record
*rec __maybe_unused
,
327 struct auxtrace_mmap
*mm __maybe_unused
)
333 void record__read_auxtrace_snapshot(struct record
*rec __maybe_unused
)
338 int auxtrace_record__snapshot_start(struct auxtrace_record
*itr __maybe_unused
)
345 static int record__open(struct record
*rec
)
348 struct perf_evsel
*pos
;
349 struct perf_evlist
*evlist
= rec
->evlist
;
350 struct perf_session
*session
= rec
->session
;
351 struct record_opts
*opts
= &rec
->opts
;
354 perf_evlist__config(evlist
, opts
, &callchain_param
);
356 evlist__for_each_entry(evlist
, pos
) {
358 if (perf_evsel__open(pos
, pos
->cpus
, pos
->threads
) < 0) {
359 if (perf_evsel__fallback(pos
, errno
, msg
, sizeof(msg
))) {
361 ui__warning("%s\n", msg
);
366 perf_evsel__open_strerror(pos
, &opts
->target
,
367 errno
, msg
, sizeof(msg
));
368 ui__error("%s\n", msg
);
373 if (perf_evlist__apply_filters(evlist
, &pos
)) {
374 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
375 pos
->filter
, perf_evsel__name(pos
), errno
,
376 strerror_r(errno
, msg
, sizeof(msg
)));
381 if (perf_evlist__mmap_ex(evlist
, opts
->mmap_pages
, false,
382 opts
->auxtrace_mmap_pages
,
383 opts
->auxtrace_snapshot_mode
) < 0) {
384 if (errno
== EPERM
) {
385 pr_err("Permission error mapping pages.\n"
386 "Consider increasing "
387 "/proc/sys/kernel/perf_event_mlock_kb,\n"
388 "or try again with a smaller value of -m/--mmap_pages.\n"
389 "(current value: %u,%u)\n",
390 opts
->mmap_pages
, opts
->auxtrace_mmap_pages
);
393 pr_err("failed to mmap with %d (%s)\n", errno
,
394 strerror_r(errno
, msg
, sizeof(msg
)));
403 session
->evlist
= evlist
;
404 perf_session__set_id_hdr_size(session
);
409 static int process_sample_event(struct perf_tool
*tool
,
410 union perf_event
*event
,
411 struct perf_sample
*sample
,
412 struct perf_evsel
*evsel
,
413 struct machine
*machine
)
415 struct record
*rec
= container_of(tool
, struct record
, tool
);
419 return build_id__mark_dso_hit(tool
, event
, sample
, evsel
, machine
);
422 static int process_buildids(struct record
*rec
)
424 struct perf_data_file
*file
= &rec
->file
;
425 struct perf_session
*session
= rec
->session
;
431 * During this process, it'll load kernel map and replace the
432 * dso->long_name to a real pathname it found. In this case
433 * we prefer the vmlinux path like
434 * /lib/modules/3.16.4/build/vmlinux
436 * rather than build-id path (in debug directory).
437 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
439 symbol_conf
.ignore_vmlinux_buildid
= true;
442 * If --buildid-all is given, it marks all DSO regardless of hits,
443 * so no need to process samples.
445 if (rec
->buildid_all
)
446 rec
->tool
.sample
= NULL
;
448 return perf_session__process_events(session
);
451 static void perf_event__synthesize_guest_os(struct machine
*machine
, void *data
)
454 struct perf_tool
*tool
= data
;
456 *As for guest kernel when processing subcommand record&report,
457 *we arrange module mmap prior to guest kernel mmap and trigger
458 *a preload dso because default guest module symbols are loaded
459 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
460 *method is used to avoid symbol missing when the first addr is
461 *in module instead of in guest kernel.
463 err
= perf_event__synthesize_modules(tool
, process_synthesized_event
,
466 pr_err("Couldn't record guest kernel [%d]'s reference"
467 " relocation symbol.\n", machine
->pid
);
470 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
471 * have no _text sometimes.
473 err
= perf_event__synthesize_kernel_mmap(tool
, process_synthesized_event
,
476 pr_err("Couldn't record guest kernel [%d]'s reference"
477 " relocation symbol.\n", machine
->pid
);
480 static struct perf_event_header finished_round_event
= {
481 .size
= sizeof(struct perf_event_header
),
482 .type
= PERF_RECORD_FINISHED_ROUND
,
485 static int record__mmap_read_all(struct record
*rec
)
487 u64 bytes_written
= rec
->bytes_written
;
491 for (i
= 0; i
< rec
->evlist
->nr_mmaps
; i
++) {
492 struct auxtrace_mmap
*mm
= &rec
->evlist
->mmap
[i
].auxtrace_mmap
;
494 if (rec
->evlist
->mmap
[i
].base
) {
495 if (record__mmap_read(rec
, i
) != 0) {
501 if (mm
->base
&& !rec
->opts
.auxtrace_snapshot_mode
&&
502 record__auxtrace_mmap_read(rec
, mm
) != 0) {
509 * Mark the round finished in case we wrote
510 * at least one event.
512 if (bytes_written
!= rec
->bytes_written
)
513 rc
= record__write(rec
, &finished_round_event
, sizeof(finished_round_event
));
519 static void record__init_features(struct record
*rec
)
521 struct perf_session
*session
= rec
->session
;
524 for (feat
= HEADER_FIRST_FEATURE
; feat
< HEADER_LAST_FEATURE
; feat
++)
525 perf_header__set_feat(&session
->header
, feat
);
528 perf_header__clear_feat(&session
->header
, HEADER_BUILD_ID
);
530 if (!have_tracepoints(&rec
->evlist
->entries
))
531 perf_header__clear_feat(&session
->header
, HEADER_TRACING_DATA
);
533 if (!rec
->opts
.branch_stack
)
534 perf_header__clear_feat(&session
->header
, HEADER_BRANCH_STACK
);
536 if (!rec
->opts
.full_auxtrace
)
537 perf_header__clear_feat(&session
->header
, HEADER_AUXTRACE
);
539 perf_header__clear_feat(&session
->header
, HEADER_STAT
);
543 record__finish_output(struct record
*rec
)
545 struct perf_data_file
*file
= &rec
->file
;
546 int fd
= perf_data_file__fd(file
);
551 rec
->session
->header
.data_size
+= rec
->bytes_written
;
552 file
->size
= lseek(perf_data_file__fd(file
), 0, SEEK_CUR
);
554 if (!rec
->no_buildid
) {
555 process_buildids(rec
);
557 if (rec
->buildid_all
)
558 dsos__hit_all(rec
->session
);
560 perf_session__write_header(rec
->session
, rec
->evlist
, fd
, true);
565 static int record__synthesize_workload(struct record
*rec
)
568 struct thread_map map
;
569 struct thread_map_data map_data
;
572 thread_map
.map
.nr
= 1;
573 thread_map
.map
.map
[0].pid
= rec
->evlist
->workload
.pid
;
574 thread_map
.map
.map
[0].comm
= NULL
;
575 return perf_event__synthesize_thread_map(&rec
->tool
, &thread_map
.map
,
576 process_synthesized_event
,
577 &rec
->session
->machines
.host
,
578 rec
->opts
.sample_address
,
579 rec
->opts
.proc_map_timeout
);
582 static int record__synthesize(struct record
*rec
);
585 record__switch_output(struct record
*rec
, bool at_exit
)
587 struct perf_data_file
*file
= &rec
->file
;
590 /* Same Size: "2015122520103046"*/
591 char timestamp
[] = "InvalidTimestamp";
594 record__finish_output(rec
);
595 err
= fetch_current_timestamp(timestamp
, sizeof(timestamp
));
597 pr_err("Failed to get current timestamp\n");
601 fd
= perf_data_file__switch(file
, timestamp
,
602 rec
->session
->header
.data_offset
,
604 if (fd
>= 0 && !at_exit
) {
605 rec
->bytes_written
= 0;
606 rec
->session
->header
.data_size
= 0;
610 fprintf(stderr
, "[ perf record: Dump %s.%s ]\n",
611 file
->path
, timestamp
);
613 /* Output tracking events */
615 record__synthesize(rec
);
618 * In 'perf record --switch-output' without -a,
619 * record__synthesize() in record__switch_output() won't
620 * generate tracking events because there's no thread_map
621 * in evlist. Which causes newly created perf.data doesn't
622 * contain map and comm information.
623 * Create a fake thread_map and directly call
624 * perf_event__synthesize_thread_map() for those events.
626 if (target__none(&rec
->opts
.target
))
627 record__synthesize_workload(rec
);
632 static volatile int workload_exec_errno
;
635 * perf_evlist__prepare_workload will send a SIGUSR1
636 * if the fork fails, since we asked by setting its
637 * want_signal to true.
639 static void workload_exec_failed_signal(int signo __maybe_unused
,
641 void *ucontext __maybe_unused
)
643 workload_exec_errno
= info
->si_value
.sival_int
;
648 static void snapshot_sig_handler(int sig
);
651 perf_event__synth_time_conv(const struct perf_event_mmap_page
*pc __maybe_unused
,
652 struct perf_tool
*tool __maybe_unused
,
653 perf_event__handler_t process __maybe_unused
,
654 struct machine
*machine __maybe_unused
)
659 static const struct perf_event_mmap_page
*record__pick_pc(struct record
*rec
)
661 if (rec
->evlist
&& rec
->evlist
->mmap
&& rec
->evlist
->mmap
[0].base
)
662 return rec
->evlist
->mmap
[0].base
;
666 static int record__synthesize(struct record
*rec
)
668 struct perf_session
*session
= rec
->session
;
669 struct machine
*machine
= &session
->machines
.host
;
670 struct perf_data_file
*file
= &rec
->file
;
671 struct record_opts
*opts
= &rec
->opts
;
672 struct perf_tool
*tool
= &rec
->tool
;
673 int fd
= perf_data_file__fd(file
);
677 err
= perf_event__synthesize_attrs(tool
, session
,
678 process_synthesized_event
);
680 pr_err("Couldn't synthesize attrs.\n");
684 if (have_tracepoints(&rec
->evlist
->entries
)) {
686 * FIXME err <= 0 here actually means that
687 * there were no tracepoints so its not really
688 * an error, just that we don't need to
689 * synthesize anything. We really have to
690 * return this more properly and also
691 * propagate errors that now are calling die()
693 err
= perf_event__synthesize_tracing_data(tool
, fd
, rec
->evlist
,
694 process_synthesized_event
);
696 pr_err("Couldn't record tracing data.\n");
699 rec
->bytes_written
+= err
;
703 err
= perf_event__synth_time_conv(record__pick_pc(rec
), tool
,
704 process_synthesized_event
, machine
);
708 if (rec
->opts
.full_auxtrace
) {
709 err
= perf_event__synthesize_auxtrace_info(rec
->itr
, tool
,
710 session
, process_synthesized_event
);
715 err
= perf_event__synthesize_kernel_mmap(tool
, process_synthesized_event
,
717 WARN_ONCE(err
< 0, "Couldn't record kernel reference relocation symbol\n"
718 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
719 "Check /proc/kallsyms permission or run as root.\n");
721 err
= perf_event__synthesize_modules(tool
, process_synthesized_event
,
723 WARN_ONCE(err
< 0, "Couldn't record kernel module information.\n"
724 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
725 "Check /proc/modules permission or run as root.\n");
728 machines__process_guests(&session
->machines
,
729 perf_event__synthesize_guest_os
, tool
);
732 err
= __machine__synthesize_threads(machine
, tool
, &opts
->target
, rec
->evlist
->threads
,
733 process_synthesized_event
, opts
->sample_address
,
734 opts
->proc_map_timeout
);
739 static int __cmd_record(struct record
*rec
, int argc
, const char **argv
)
743 unsigned long waking
= 0;
744 const bool forks
= argc
> 0;
745 struct machine
*machine
;
746 struct perf_tool
*tool
= &rec
->tool
;
747 struct record_opts
*opts
= &rec
->opts
;
748 struct perf_data_file
*file
= &rec
->file
;
749 struct perf_session
*session
;
750 bool disabled
= false, draining
= false;
753 rec
->progname
= argv
[0];
755 atexit(record__sig_exit
);
756 signal(SIGCHLD
, sig_handler
);
757 signal(SIGINT
, sig_handler
);
758 signal(SIGTERM
, sig_handler
);
760 if (rec
->opts
.auxtrace_snapshot_mode
|| rec
->switch_output
) {
761 signal(SIGUSR2
, snapshot_sig_handler
);
762 if (rec
->opts
.auxtrace_snapshot_mode
)
763 trigger_on(&auxtrace_snapshot_trigger
);
764 if (rec
->switch_output
)
765 trigger_on(&switch_output_trigger
);
767 signal(SIGUSR2
, SIG_IGN
);
770 session
= perf_session__new(file
, false, tool
);
771 if (session
== NULL
) {
772 pr_err("Perf session creation failed.\n");
776 fd
= perf_data_file__fd(file
);
777 rec
->session
= session
;
779 record__init_features(rec
);
782 err
= perf_evlist__prepare_workload(rec
->evlist
, &opts
->target
,
784 workload_exec_failed_signal
);
786 pr_err("Couldn't run the workload!\n");
788 goto out_delete_session
;
792 if (record__open(rec
) != 0) {
797 err
= bpf__apply_obj_config();
801 bpf__strerror_apply_obj_config(err
, errbuf
, sizeof(errbuf
));
802 pr_err("ERROR: Apply config to BPF failed: %s\n",
808 * Normally perf_session__new would do this, but it doesn't have the
811 if (rec
->tool
.ordered_events
&& !perf_evlist__sample_id_all(rec
->evlist
)) {
812 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
813 rec
->tool
.ordered_events
= false;
816 if (!rec
->evlist
->nr_groups
)
817 perf_header__clear_feat(&session
->header
, HEADER_GROUP_DESC
);
820 err
= perf_header__write_pipe(fd
);
824 err
= perf_session__write_header(session
, rec
->evlist
, fd
, false);
830 && !perf_header__has_feat(&session
->header
, HEADER_BUILD_ID
)) {
831 pr_err("Couldn't generate buildids. "
832 "Use --no-buildid to profile anyway.\n");
837 machine
= &session
->machines
.host
;
839 err
= record__synthesize(rec
);
843 if (rec
->realtime_prio
) {
844 struct sched_param param
;
846 param
.sched_priority
= rec
->realtime_prio
;
847 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
848 pr_err("Could not set realtime priority.\n");
855 * When perf is starting the traced process, all the events
856 * (apart from group members) have enable_on_exec=1 set,
857 * so don't spoil it by prematurely enabling them.
859 if (!target__none(&opts
->target
) && !opts
->initial_delay
)
860 perf_evlist__enable(rec
->evlist
);
866 union perf_event
*event
;
868 event
= malloc(sizeof(event
->comm
) + machine
->id_hdr_size
);
875 * Some H/W events are generated before COMM event
876 * which is emitted during exec(), so perf script
877 * cannot see a correct process name for those events.
878 * Synthesize COMM event to prevent it.
880 perf_event__synthesize_comm(tool
, event
,
881 rec
->evlist
->workload
.pid
,
882 process_synthesized_event
,
886 perf_evlist__start_workload(rec
->evlist
);
889 if (opts
->initial_delay
) {
890 usleep(opts
->initial_delay
* 1000);
891 perf_evlist__enable(rec
->evlist
);
894 trigger_ready(&auxtrace_snapshot_trigger
);
895 trigger_ready(&switch_output_trigger
);
897 unsigned long long hits
= rec
->samples
;
899 if (record__mmap_read_all(rec
) < 0) {
900 trigger_error(&auxtrace_snapshot_trigger
);
901 trigger_error(&switch_output_trigger
);
906 if (auxtrace_record__snapshot_started
) {
907 auxtrace_record__snapshot_started
= 0;
908 if (!trigger_is_error(&auxtrace_snapshot_trigger
))
909 record__read_auxtrace_snapshot(rec
);
910 if (trigger_is_error(&auxtrace_snapshot_trigger
)) {
911 pr_err("AUX area tracing snapshot failed\n");
917 if (trigger_is_hit(&switch_output_trigger
)) {
918 trigger_ready(&switch_output_trigger
);
921 fprintf(stderr
, "[ perf record: dump data: Woken up %ld times ]\n",
924 fd
= record__switch_output(rec
, false);
926 pr_err("Failed to switch to new file\n");
927 trigger_error(&switch_output_trigger
);
933 if (hits
== rec
->samples
) {
934 if (done
|| draining
)
936 err
= perf_evlist__poll(rec
->evlist
, -1);
938 * Propagate error, only if there's any. Ignore positive
939 * number of returned events and interrupt error.
941 if (err
> 0 || (err
< 0 && errno
== EINTR
))
945 if (perf_evlist__filter_pollfd(rec
->evlist
, POLLERR
| POLLHUP
) == 0)
950 * When perf is starting the traced process, at the end events
951 * die with the process and we wait for that. Thus no need to
952 * disable events in this case.
954 if (done
&& !disabled
&& !target__none(&opts
->target
)) {
955 trigger_off(&auxtrace_snapshot_trigger
);
956 perf_evlist__disable(rec
->evlist
);
960 trigger_off(&auxtrace_snapshot_trigger
);
961 trigger_off(&switch_output_trigger
);
963 if (forks
&& workload_exec_errno
) {
964 char msg
[STRERR_BUFSIZE
];
965 const char *emsg
= strerror_r(workload_exec_errno
, msg
, sizeof(msg
));
966 pr_err("Workload failed: %s\n", emsg
);
972 fprintf(stderr
, "[ perf record: Woken up %ld times to write data ]\n", waking
);
979 kill(rec
->evlist
->workload
.pid
, SIGTERM
);
985 else if (WIFEXITED(exit_status
))
986 status
= WEXITSTATUS(exit_status
);
987 else if (WIFSIGNALED(exit_status
))
988 signr
= WTERMSIG(exit_status
);
992 /* this will be recalculated during process_buildids() */
996 if (!rec
->timestamp_filename
) {
997 record__finish_output(rec
);
999 fd
= record__switch_output(rec
, true);
1002 goto out_delete_session
;
1007 if (!err
&& !quiet
) {
1009 const char *postfix
= rec
->timestamp_filename
?
1010 ".<timestamp>" : "";
1012 if (rec
->samples
&& !rec
->opts
.full_auxtrace
)
1013 scnprintf(samples
, sizeof(samples
),
1014 " (%" PRIu64
" samples)", rec
->samples
);
1018 fprintf(stderr
, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1019 perf_data_file__size(file
) / 1024.0 / 1024.0,
1020 file
->path
, postfix
, samples
);
1024 perf_session__delete(session
);
1028 static void callchain_debug(struct callchain_param
*callchain
)
1030 static const char *str
[CALLCHAIN_MAX
] = { "NONE", "FP", "DWARF", "LBR" };
1032 pr_debug("callchain: type %s\n", str
[callchain
->record_mode
]);
1034 if (callchain
->record_mode
== CALLCHAIN_DWARF
)
1035 pr_debug("callchain: stack dump size %d\n",
1036 callchain
->dump_size
);
1039 int record_opts__parse_callchain(struct record_opts
*record
,
1040 struct callchain_param
*callchain
,
1041 const char *arg
, bool unset
)
1044 callchain
->enabled
= !unset
;
1046 /* --no-call-graph */
1048 callchain
->record_mode
= CALLCHAIN_NONE
;
1049 pr_debug("callchain: disabled\n");
1053 ret
= parse_callchain_record_opt(arg
, callchain
);
1055 /* Enable data address sampling for DWARF unwind. */
1056 if (callchain
->record_mode
== CALLCHAIN_DWARF
)
1057 record
->sample_address
= true;
1058 callchain_debug(callchain
);
1064 int record_parse_callchain_opt(const struct option
*opt
,
1068 return record_opts__parse_callchain(opt
->value
, &callchain_param
, arg
, unset
);
1071 int record_callchain_opt(const struct option
*opt
,
1072 const char *arg __maybe_unused
,
1073 int unset __maybe_unused
)
1075 struct callchain_param
*callchain
= opt
->value
;
1077 callchain
->enabled
= true;
1079 if (callchain
->record_mode
== CALLCHAIN_NONE
)
1080 callchain
->record_mode
= CALLCHAIN_FP
;
1082 callchain_debug(callchain
);
1086 static int perf_record_config(const char *var
, const char *value
, void *cb
)
1088 struct record
*rec
= cb
;
1090 if (!strcmp(var
, "record.build-id")) {
1091 if (!strcmp(value
, "cache"))
1092 rec
->no_buildid_cache
= false;
1093 else if (!strcmp(value
, "no-cache"))
1094 rec
->no_buildid_cache
= true;
1095 else if (!strcmp(value
, "skip"))
1096 rec
->no_buildid
= true;
1101 if (!strcmp(var
, "record.call-graph"))
1102 var
= "call-graph.record-mode"; /* fall-through */
1104 return perf_default_config(var
, value
, cb
);
1107 struct clockid_map
{
1112 #define CLOCKID_MAP(n, c) \
1113 { .name = n, .clockid = (c), }
1115 #define CLOCKID_END { .name = NULL, }
1119 * Add the missing ones, we need to build on many distros...
1121 #ifndef CLOCK_MONOTONIC_RAW
1122 #define CLOCK_MONOTONIC_RAW 4
1124 #ifndef CLOCK_BOOTTIME
1125 #define CLOCK_BOOTTIME 7
1128 #define CLOCK_TAI 11
1131 static const struct clockid_map clockids
[] = {
1132 /* available for all events, NMI safe */
1133 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC
),
1134 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW
),
1136 /* available for some events */
1137 CLOCKID_MAP("realtime", CLOCK_REALTIME
),
1138 CLOCKID_MAP("boottime", CLOCK_BOOTTIME
),
1139 CLOCKID_MAP("tai", CLOCK_TAI
),
1141 /* available for the lazy */
1142 CLOCKID_MAP("mono", CLOCK_MONOTONIC
),
1143 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW
),
1144 CLOCKID_MAP("real", CLOCK_REALTIME
),
1145 CLOCKID_MAP("boot", CLOCK_BOOTTIME
),
1150 static int parse_clockid(const struct option
*opt
, const char *str
, int unset
)
1152 struct record_opts
*opts
= (struct record_opts
*)opt
->value
;
1153 const struct clockid_map
*cm
;
1154 const char *ostr
= str
;
1157 opts
->use_clockid
= 0;
1165 /* no setting it twice */
1166 if (opts
->use_clockid
)
1169 opts
->use_clockid
= true;
1171 /* if its a number, we're done */
1172 if (sscanf(str
, "%d", &opts
->clockid
) == 1)
1175 /* allow a "CLOCK_" prefix to the name */
1176 if (!strncasecmp(str
, "CLOCK_", 6))
1179 for (cm
= clockids
; cm
->name
; cm
++) {
1180 if (!strcasecmp(str
, cm
->name
)) {
1181 opts
->clockid
= cm
->clockid
;
1186 opts
->use_clockid
= false;
1187 ui__warning("unknown clockid %s, check man page\n", ostr
);
1191 static int record__parse_mmap_pages(const struct option
*opt
,
1193 int unset __maybe_unused
)
1195 struct record_opts
*opts
= opt
->value
;
1197 unsigned int mmap_pages
;
1212 ret
= __perf_evlist__parse_mmap_pages(&mmap_pages
, s
);
1215 opts
->mmap_pages
= mmap_pages
;
1223 ret
= __perf_evlist__parse_mmap_pages(&mmap_pages
, p
+ 1);
1227 opts
->auxtrace_mmap_pages
= mmap_pages
;
1234 static const char * const __record_usage
[] = {
1235 "perf record [<options>] [<command>]",
1236 "perf record [<options>] -- <command> [<options>]",
1239 const char * const *record_usage
= __record_usage
;
1242 * XXX Ideally would be local to cmd_record() and passed to a record__new
1243 * because we need to have access to it in record__exit, that is called
1244 * after cmd_record() exits, but since record_options need to be accessible to
1245 * builtin-script, leave it here.
1247 * At least we don't ouch it in all the other functions here directly.
1249 * Just say no to tons of global variables, sigh.
1251 static struct record record
= {
1253 .sample_time
= true,
1254 .mmap_pages
= UINT_MAX
,
1255 .user_freq
= UINT_MAX
,
1256 .user_interval
= ULLONG_MAX
,
1260 .default_per_cpu
= true,
1262 .proc_map_timeout
= 500,
1265 .sample
= process_sample_event
,
1266 .fork
= perf_event__process_fork
,
1267 .exit
= perf_event__process_exit
,
1268 .comm
= perf_event__process_comm
,
1269 .mmap
= perf_event__process_mmap
,
1270 .mmap2
= perf_event__process_mmap2
,
1271 .ordered_events
= true,
1275 const char record_callchain_help
[] = CALLCHAIN_RECORD_HELP
1276 "\n\t\t\t\tDefault: fp";
1278 static bool dry_run
;
1281 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1282 * with it and switch to use the library functions in perf_evlist that came
1283 * from builtin-record.c, i.e. use record_opts,
1284 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1287 struct option __record_options
[] = {
1288 OPT_CALLBACK('e', "event", &record
.evlist
, "event",
1289 "event selector. use 'perf list' to list available events",
1290 parse_events_option
),
1291 OPT_CALLBACK(0, "filter", &record
.evlist
, "filter",
1292 "event filter", parse_filter
),
1293 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record
.evlist
,
1294 NULL
, "don't record events from perf itself",
1296 OPT_STRING('p', "pid", &record
.opts
.target
.pid
, "pid",
1297 "record events on existing process id"),
1298 OPT_STRING('t', "tid", &record
.opts
.target
.tid
, "tid",
1299 "record events on existing thread id"),
1300 OPT_INTEGER('r', "realtime", &record
.realtime_prio
,
1301 "collect data with this RT SCHED_FIFO priority"),
1302 OPT_BOOLEAN(0, "no-buffering", &record
.opts
.no_buffering
,
1303 "collect data without buffering"),
1304 OPT_BOOLEAN('R', "raw-samples", &record
.opts
.raw_samples
,
1305 "collect raw sample records from all opened counters"),
1306 OPT_BOOLEAN('a', "all-cpus", &record
.opts
.target
.system_wide
,
1307 "system-wide collection from all CPUs"),
1308 OPT_STRING('C', "cpu", &record
.opts
.target
.cpu_list
, "cpu",
1309 "list of cpus to monitor"),
1310 OPT_U64('c', "count", &record
.opts
.user_interval
, "event period to sample"),
1311 OPT_STRING('o', "output", &record
.file
.path
, "file",
1312 "output file name"),
1313 OPT_BOOLEAN_SET('i', "no-inherit", &record
.opts
.no_inherit
,
1314 &record
.opts
.no_inherit_set
,
1315 "child tasks do not inherit counters"),
1316 OPT_UINTEGER('F', "freq", &record
.opts
.user_freq
, "profile at this frequency"),
1317 OPT_CALLBACK('m', "mmap-pages", &record
.opts
, "pages[,pages]",
1318 "number of mmap data pages and AUX area tracing mmap pages",
1319 record__parse_mmap_pages
),
1320 OPT_BOOLEAN(0, "group", &record
.opts
.group
,
1321 "put the counters into a counter group"),
1322 OPT_CALLBACK_NOOPT('g', NULL
, &callchain_param
,
1323 NULL
, "enables call-graph recording" ,
1324 &record_callchain_opt
),
1325 OPT_CALLBACK(0, "call-graph", &record
.opts
,
1326 "record_mode[,record_size]", record_callchain_help
,
1327 &record_parse_callchain_opt
),
1328 OPT_INCR('v', "verbose", &verbose
,
1329 "be more verbose (show counter open errors, etc)"),
1330 OPT_BOOLEAN('q', "quiet", &quiet
, "don't print any message"),
1331 OPT_BOOLEAN('s', "stat", &record
.opts
.inherit_stat
,
1332 "per thread counts"),
1333 OPT_BOOLEAN('d', "data", &record
.opts
.sample_address
, "Record the sample addresses"),
1334 OPT_BOOLEAN_SET('T', "timestamp", &record
.opts
.sample_time
,
1335 &record
.opts
.sample_time_set
,
1336 "Record the sample timestamps"),
1337 OPT_BOOLEAN('P', "period", &record
.opts
.period
, "Record the sample period"),
1338 OPT_BOOLEAN('n', "no-samples", &record
.opts
.no_samples
,
1340 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record
.no_buildid_cache
,
1341 &record
.no_buildid_cache_set
,
1342 "do not update the buildid cache"),
1343 OPT_BOOLEAN_SET('B', "no-buildid", &record
.no_buildid
,
1344 &record
.no_buildid_set
,
1345 "do not collect buildids in perf.data"),
1346 OPT_CALLBACK('G', "cgroup", &record
.evlist
, "name",
1347 "monitor event in cgroup name only",
1349 OPT_UINTEGER('D', "delay", &record
.opts
.initial_delay
,
1350 "ms to wait before starting measurement after program start"),
1351 OPT_STRING('u', "uid", &record
.opts
.target
.uid_str
, "user",
1354 OPT_CALLBACK_NOOPT('b', "branch-any", &record
.opts
.branch_stack
,
1355 "branch any", "sample any taken branches",
1356 parse_branch_stack
),
1358 OPT_CALLBACK('j', "branch-filter", &record
.opts
.branch_stack
,
1359 "branch filter mask", "branch stack filter modes",
1360 parse_branch_stack
),
1361 OPT_BOOLEAN('W', "weight", &record
.opts
.sample_weight
,
1362 "sample by weight (on special events only)"),
1363 OPT_BOOLEAN(0, "transaction", &record
.opts
.sample_transaction
,
1364 "sample transaction flags (special events only)"),
1365 OPT_BOOLEAN(0, "per-thread", &record
.opts
.target
.per_thread
,
1366 "use per-thread mmaps"),
1367 OPT_CALLBACK_OPTARG('I', "intr-regs", &record
.opts
.sample_intr_regs
, NULL
, "any register",
1368 "sample selected machine registers on interrupt,"
1369 " use -I ? to list register names", parse_regs
),
1370 OPT_BOOLEAN(0, "running-time", &record
.opts
.running_time
,
1371 "Record running/enabled time of read (:S) events"),
1372 OPT_CALLBACK('k', "clockid", &record
.opts
,
1373 "clockid", "clockid to use for events, see clock_gettime()",
1375 OPT_STRING_OPTARG('S', "snapshot", &record
.opts
.auxtrace_snapshot_opts
,
1376 "opts", "AUX area tracing Snapshot Mode", ""),
1377 OPT_UINTEGER(0, "proc-map-timeout", &record
.opts
.proc_map_timeout
,
1378 "per thread proc mmap processing timeout in ms"),
1379 OPT_BOOLEAN(0, "switch-events", &record
.opts
.record_switch_events
,
1380 "Record context switch events"),
1381 OPT_BOOLEAN_FLAG(0, "all-kernel", &record
.opts
.all_kernel
,
1382 "Configure all used events to run in kernel space.",
1383 PARSE_OPT_EXCLUSIVE
),
1384 OPT_BOOLEAN_FLAG(0, "all-user", &record
.opts
.all_user
,
1385 "Configure all used events to run in user space.",
1386 PARSE_OPT_EXCLUSIVE
),
1387 OPT_STRING(0, "clang-path", &llvm_param
.clang_path
, "clang path",
1388 "clang binary to use for compiling BPF scriptlets"),
1389 OPT_STRING(0, "clang-opt", &llvm_param
.clang_opt
, "clang options",
1390 "options passed to clang when compiling BPF scriptlets"),
1391 OPT_STRING(0, "vmlinux", &symbol_conf
.vmlinux_name
,
1392 "file", "vmlinux pathname"),
1393 OPT_BOOLEAN(0, "buildid-all", &record
.buildid_all
,
1394 "Record build-id of all DSOs regardless of hits"),
1395 OPT_BOOLEAN(0, "timestamp-filename", &record
.timestamp_filename
,
1396 "append timestamp to output filename"),
1397 OPT_BOOLEAN(0, "switch-output", &record
.switch_output
,
1398 "Switch output when receive SIGUSR2"),
1399 OPT_BOOLEAN(0, "dry-run", &dry_run
,
1400 "Parse options then exit"),
1404 struct option
*record_options
= __record_options
;
1406 int cmd_record(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1409 struct record
*rec
= &record
;
1410 char errbuf
[BUFSIZ
];
1412 #ifndef HAVE_LIBBPF_SUPPORT
1413 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1414 set_nobuild('\0', "clang-path", true);
1415 set_nobuild('\0', "clang-opt", true);
1419 #ifndef HAVE_BPF_PROLOGUE
1420 # if !defined (HAVE_DWARF_SUPPORT)
1421 # define REASON "NO_DWARF=1"
1422 # elif !defined (HAVE_LIBBPF_SUPPORT)
1423 # define REASON "NO_LIBBPF=1"
1425 # define REASON "this architecture doesn't support BPF prologue"
1427 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1428 set_nobuild('\0', "vmlinux", true);
1433 rec
->evlist
= perf_evlist__new();
1434 if (rec
->evlist
== NULL
)
1437 perf_config(perf_record_config
, rec
);
1439 argc
= parse_options(argc
, argv
, record_options
, record_usage
,
1440 PARSE_OPT_STOP_AT_NON_OPTION
);
1441 if (!argc
&& target__none(&rec
->opts
.target
))
1442 usage_with_options(record_usage
, record_options
);
1444 if (nr_cgroups
&& !rec
->opts
.target
.system_wide
) {
1445 usage_with_options_msg(record_usage
, record_options
,
1446 "cgroup monitoring only available in system-wide mode");
1449 if (rec
->opts
.record_switch_events
&&
1450 !perf_can_record_switch_events()) {
1451 ui__error("kernel does not support recording context switch events\n");
1452 parse_options_usage(record_usage
, record_options
, "switch-events", 0);
1456 if (rec
->switch_output
)
1457 rec
->timestamp_filename
= true;
1460 rec
->itr
= auxtrace_record__init(rec
->evlist
, &err
);
1465 err
= auxtrace_parse_snapshot_options(rec
->itr
, &rec
->opts
,
1466 rec
->opts
.auxtrace_snapshot_opts
);
1473 err
= bpf__setup_stdout(rec
->evlist
);
1475 bpf__strerror_setup_stdout(rec
->evlist
, err
, errbuf
, sizeof(errbuf
));
1476 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1485 if (symbol_conf
.kptr_restrict
)
1487 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1488 "check /proc/sys/kernel/kptr_restrict.\n\n"
1489 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1490 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1491 "Samples in kernel modules won't be resolved at all.\n\n"
1492 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1493 "even with a suitable vmlinux or kallsyms file.\n\n");
1495 if (rec
->no_buildid_cache
|| rec
->no_buildid
) {
1496 disable_buildid_cache();
1497 } else if (rec
->switch_output
) {
1499 * In 'perf record --switch-output', disable buildid
1500 * generation by default to reduce data file switching
1501 * overhead. Still generate buildid if they are required
1504 * perf record --signal-trigger --no-no-buildid \
1505 * --no-no-buildid-cache
1507 * Following code equals to:
1509 * if ((rec->no_buildid || !rec->no_buildid_set) &&
1510 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1511 * disable_buildid_cache();
1513 bool disable
= true;
1515 if (rec
->no_buildid_set
&& !rec
->no_buildid
)
1517 if (rec
->no_buildid_cache_set
&& !rec
->no_buildid_cache
)
1520 rec
->no_buildid
= true;
1521 rec
->no_buildid_cache
= true;
1522 disable_buildid_cache();
1526 if (rec
->evlist
->nr_entries
== 0 &&
1527 perf_evlist__add_default(rec
->evlist
) < 0) {
1528 pr_err("Not enough memory for event selector list\n");
1529 goto out_symbol_exit
;
1532 if (rec
->opts
.target
.tid
&& !rec
->opts
.no_inherit_set
)
1533 rec
->opts
.no_inherit
= true;
1535 err
= target__validate(&rec
->opts
.target
);
1537 target__strerror(&rec
->opts
.target
, err
, errbuf
, BUFSIZ
);
1538 ui__warning("%s", errbuf
);
1541 err
= target__parse_uid(&rec
->opts
.target
);
1543 int saved_errno
= errno
;
1545 target__strerror(&rec
->opts
.target
, err
, errbuf
, BUFSIZ
);
1546 ui__error("%s", errbuf
);
1549 goto out_symbol_exit
;
1553 if (perf_evlist__create_maps(rec
->evlist
, &rec
->opts
.target
) < 0)
1554 usage_with_options(record_usage
, record_options
);
1556 err
= auxtrace_record__options(rec
->itr
, rec
->evlist
, &rec
->opts
);
1558 goto out_symbol_exit
;
1561 * We take all buildids when the file contains
1562 * AUX area tracing data because we do not decode the
1563 * trace because it would take too long.
1565 if (rec
->opts
.full_auxtrace
)
1566 rec
->buildid_all
= true;
1568 if (record_opts__config(&rec
->opts
)) {
1570 goto out_symbol_exit
;
1573 err
= __cmd_record(&record
, argc
, argv
);
1575 perf_evlist__delete(rec
->evlist
);
1577 auxtrace_record__free(rec
->itr
);
1581 static void snapshot_sig_handler(int sig __maybe_unused
)
1583 if (trigger_is_ready(&auxtrace_snapshot_trigger
)) {
1584 trigger_hit(&auxtrace_snapshot_trigger
);
1585 auxtrace_record__snapshot_started
= 1;
1586 if (auxtrace_record__snapshot_start(record
.itr
))
1587 trigger_error(&auxtrace_snapshot_trigger
);
1590 if (trigger_is_ready(&switch_output_trigger
))
1591 trigger_hit(&switch_output_trigger
);