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
, struct perf_evlist
*evlist
, int idx
)
137 struct perf_mmap
*md
= &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(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(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(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__mmap_evlist(struct record
*rec
,
346 struct perf_evlist
*evlist
)
348 struct record_opts
*opts
= &rec
->opts
;
351 if (perf_evlist__mmap_ex(evlist
, opts
->mmap_pages
, false,
352 opts
->auxtrace_mmap_pages
,
353 opts
->auxtrace_snapshot_mode
) < 0) {
354 if (errno
== EPERM
) {
355 pr_err("Permission error mapping pages.\n"
356 "Consider increasing "
357 "/proc/sys/kernel/perf_event_mlock_kb,\n"
358 "or try again with a smaller value of -m/--mmap_pages.\n"
359 "(current value: %u,%u)\n",
360 opts
->mmap_pages
, opts
->auxtrace_mmap_pages
);
363 pr_err("failed to mmap with %d (%s)\n", errno
,
364 strerror_r(errno
, msg
, sizeof(msg
)));
374 static int record__mmap(struct record
*rec
)
376 return record__mmap_evlist(rec
, rec
->evlist
);
379 static int record__open(struct record
*rec
)
382 struct perf_evsel
*pos
;
383 struct perf_evlist
*evlist
= rec
->evlist
;
384 struct perf_session
*session
= rec
->session
;
385 struct record_opts
*opts
= &rec
->opts
;
388 perf_evlist__config(evlist
, opts
, &callchain_param
);
390 evlist__for_each_entry(evlist
, pos
) {
392 if (perf_evsel__open(pos
, pos
->cpus
, pos
->threads
) < 0) {
393 if (perf_evsel__fallback(pos
, errno
, msg
, sizeof(msg
))) {
395 ui__warning("%s\n", msg
);
400 perf_evsel__open_strerror(pos
, &opts
->target
,
401 errno
, msg
, sizeof(msg
));
402 ui__error("%s\n", msg
);
407 if (perf_evlist__apply_filters(evlist
, &pos
)) {
408 error("failed to set filter \"%s\" on event %s with %d (%s)\n",
409 pos
->filter
, perf_evsel__name(pos
), errno
,
410 strerror_r(errno
, msg
, sizeof(msg
)));
415 rc
= record__mmap(rec
);
419 session
->evlist
= evlist
;
420 perf_session__set_id_hdr_size(session
);
425 static int process_sample_event(struct perf_tool
*tool
,
426 union perf_event
*event
,
427 struct perf_sample
*sample
,
428 struct perf_evsel
*evsel
,
429 struct machine
*machine
)
431 struct record
*rec
= container_of(tool
, struct record
, tool
);
435 return build_id__mark_dso_hit(tool
, event
, sample
, evsel
, machine
);
438 static int process_buildids(struct record
*rec
)
440 struct perf_data_file
*file
= &rec
->file
;
441 struct perf_session
*session
= rec
->session
;
447 * During this process, it'll load kernel map and replace the
448 * dso->long_name to a real pathname it found. In this case
449 * we prefer the vmlinux path like
450 * /lib/modules/3.16.4/build/vmlinux
452 * rather than build-id path (in debug directory).
453 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
455 symbol_conf
.ignore_vmlinux_buildid
= true;
458 * If --buildid-all is given, it marks all DSO regardless of hits,
459 * so no need to process samples.
461 if (rec
->buildid_all
)
462 rec
->tool
.sample
= NULL
;
464 return perf_session__process_events(session
);
467 static void perf_event__synthesize_guest_os(struct machine
*machine
, void *data
)
470 struct perf_tool
*tool
= data
;
472 *As for guest kernel when processing subcommand record&report,
473 *we arrange module mmap prior to guest kernel mmap and trigger
474 *a preload dso because default guest module symbols are loaded
475 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
476 *method is used to avoid symbol missing when the first addr is
477 *in module instead of in guest kernel.
479 err
= perf_event__synthesize_modules(tool
, process_synthesized_event
,
482 pr_err("Couldn't record guest kernel [%d]'s reference"
483 " relocation symbol.\n", machine
->pid
);
486 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
487 * have no _text sometimes.
489 err
= perf_event__synthesize_kernel_mmap(tool
, process_synthesized_event
,
492 pr_err("Couldn't record guest kernel [%d]'s reference"
493 " relocation symbol.\n", machine
->pid
);
496 static struct perf_event_header finished_round_event
= {
497 .size
= sizeof(struct perf_event_header
),
498 .type
= PERF_RECORD_FINISHED_ROUND
,
501 static int record__mmap_read_evlist(struct record
*rec
, struct perf_evlist
*evlist
)
503 u64 bytes_written
= rec
->bytes_written
;
510 for (i
= 0; i
< evlist
->nr_mmaps
; i
++) {
511 struct auxtrace_mmap
*mm
= &evlist
->mmap
[i
].auxtrace_mmap
;
513 if (evlist
->mmap
[i
].base
) {
514 if (record__mmap_read(rec
, evlist
, i
) != 0) {
520 if (mm
->base
&& !rec
->opts
.auxtrace_snapshot_mode
&&
521 record__auxtrace_mmap_read(rec
, mm
) != 0) {
528 * Mark the round finished in case we wrote
529 * at least one event.
531 if (bytes_written
!= rec
->bytes_written
)
532 rc
= record__write(rec
, &finished_round_event
, sizeof(finished_round_event
));
538 static int record__mmap_read_all(struct record
*rec
)
542 err
= record__mmap_read_evlist(rec
, rec
->evlist
);
549 static void record__init_features(struct record
*rec
)
551 struct perf_session
*session
= rec
->session
;
554 for (feat
= HEADER_FIRST_FEATURE
; feat
< HEADER_LAST_FEATURE
; feat
++)
555 perf_header__set_feat(&session
->header
, feat
);
558 perf_header__clear_feat(&session
->header
, HEADER_BUILD_ID
);
560 if (!have_tracepoints(&rec
->evlist
->entries
))
561 perf_header__clear_feat(&session
->header
, HEADER_TRACING_DATA
);
563 if (!rec
->opts
.branch_stack
)
564 perf_header__clear_feat(&session
->header
, HEADER_BRANCH_STACK
);
566 if (!rec
->opts
.full_auxtrace
)
567 perf_header__clear_feat(&session
->header
, HEADER_AUXTRACE
);
569 perf_header__clear_feat(&session
->header
, HEADER_STAT
);
573 record__finish_output(struct record
*rec
)
575 struct perf_data_file
*file
= &rec
->file
;
576 int fd
= perf_data_file__fd(file
);
581 rec
->session
->header
.data_size
+= rec
->bytes_written
;
582 file
->size
= lseek(perf_data_file__fd(file
), 0, SEEK_CUR
);
584 if (!rec
->no_buildid
) {
585 process_buildids(rec
);
587 if (rec
->buildid_all
)
588 dsos__hit_all(rec
->session
);
590 perf_session__write_header(rec
->session
, rec
->evlist
, fd
, true);
595 static int record__synthesize_workload(struct record
*rec
)
598 struct thread_map map
;
599 struct thread_map_data map_data
;
602 thread_map
.map
.nr
= 1;
603 thread_map
.map
.map
[0].pid
= rec
->evlist
->workload
.pid
;
604 thread_map
.map
.map
[0].comm
= NULL
;
605 return perf_event__synthesize_thread_map(&rec
->tool
, &thread_map
.map
,
606 process_synthesized_event
,
607 &rec
->session
->machines
.host
,
608 rec
->opts
.sample_address
,
609 rec
->opts
.proc_map_timeout
);
612 static int record__synthesize(struct record
*rec
);
615 record__switch_output(struct record
*rec
, bool at_exit
)
617 struct perf_data_file
*file
= &rec
->file
;
620 /* Same Size: "2015122520103046"*/
621 char timestamp
[] = "InvalidTimestamp";
624 record__finish_output(rec
);
625 err
= fetch_current_timestamp(timestamp
, sizeof(timestamp
));
627 pr_err("Failed to get current timestamp\n");
631 fd
= perf_data_file__switch(file
, timestamp
,
632 rec
->session
->header
.data_offset
,
634 if (fd
>= 0 && !at_exit
) {
635 rec
->bytes_written
= 0;
636 rec
->session
->header
.data_size
= 0;
640 fprintf(stderr
, "[ perf record: Dump %s.%s ]\n",
641 file
->path
, timestamp
);
643 /* Output tracking events */
645 record__synthesize(rec
);
648 * In 'perf record --switch-output' without -a,
649 * record__synthesize() in record__switch_output() won't
650 * generate tracking events because there's no thread_map
651 * in evlist. Which causes newly created perf.data doesn't
652 * contain map and comm information.
653 * Create a fake thread_map and directly call
654 * perf_event__synthesize_thread_map() for those events.
656 if (target__none(&rec
->opts
.target
))
657 record__synthesize_workload(rec
);
662 static volatile int workload_exec_errno
;
665 * perf_evlist__prepare_workload will send a SIGUSR1
666 * if the fork fails, since we asked by setting its
667 * want_signal to true.
669 static void workload_exec_failed_signal(int signo __maybe_unused
,
671 void *ucontext __maybe_unused
)
673 workload_exec_errno
= info
->si_value
.sival_int
;
678 static void snapshot_sig_handler(int sig
);
681 perf_event__synth_time_conv(const struct perf_event_mmap_page
*pc __maybe_unused
,
682 struct perf_tool
*tool __maybe_unused
,
683 perf_event__handler_t process __maybe_unused
,
684 struct machine
*machine __maybe_unused
)
689 static const struct perf_event_mmap_page
*
690 perf_evlist__pick_pc(struct perf_evlist
*evlist
)
692 if (evlist
&& evlist
->mmap
&& evlist
->mmap
[0].base
)
693 return evlist
->mmap
[0].base
;
697 static const struct perf_event_mmap_page
*record__pick_pc(struct record
*rec
)
699 const struct perf_event_mmap_page
*pc
;
701 pc
= perf_evlist__pick_pc(rec
->evlist
);
707 static int record__synthesize(struct record
*rec
)
709 struct perf_session
*session
= rec
->session
;
710 struct machine
*machine
= &session
->machines
.host
;
711 struct perf_data_file
*file
= &rec
->file
;
712 struct record_opts
*opts
= &rec
->opts
;
713 struct perf_tool
*tool
= &rec
->tool
;
714 int fd
= perf_data_file__fd(file
);
718 err
= perf_event__synthesize_attrs(tool
, session
,
719 process_synthesized_event
);
721 pr_err("Couldn't synthesize attrs.\n");
725 if (have_tracepoints(&rec
->evlist
->entries
)) {
727 * FIXME err <= 0 here actually means that
728 * there were no tracepoints so its not really
729 * an error, just that we don't need to
730 * synthesize anything. We really have to
731 * return this more properly and also
732 * propagate errors that now are calling die()
734 err
= perf_event__synthesize_tracing_data(tool
, fd
, rec
->evlist
,
735 process_synthesized_event
);
737 pr_err("Couldn't record tracing data.\n");
740 rec
->bytes_written
+= err
;
744 err
= perf_event__synth_time_conv(record__pick_pc(rec
), tool
,
745 process_synthesized_event
, machine
);
749 if (rec
->opts
.full_auxtrace
) {
750 err
= perf_event__synthesize_auxtrace_info(rec
->itr
, tool
,
751 session
, process_synthesized_event
);
756 err
= perf_event__synthesize_kernel_mmap(tool
, process_synthesized_event
,
758 WARN_ONCE(err
< 0, "Couldn't record kernel reference relocation symbol\n"
759 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
760 "Check /proc/kallsyms permission or run as root.\n");
762 err
= perf_event__synthesize_modules(tool
, process_synthesized_event
,
764 WARN_ONCE(err
< 0, "Couldn't record kernel module information.\n"
765 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
766 "Check /proc/modules permission or run as root.\n");
769 machines__process_guests(&session
->machines
,
770 perf_event__synthesize_guest_os
, tool
);
773 err
= __machine__synthesize_threads(machine
, tool
, &opts
->target
, rec
->evlist
->threads
,
774 process_synthesized_event
, opts
->sample_address
,
775 opts
->proc_map_timeout
);
780 static int __cmd_record(struct record
*rec
, int argc
, const char **argv
)
784 unsigned long waking
= 0;
785 const bool forks
= argc
> 0;
786 struct machine
*machine
;
787 struct perf_tool
*tool
= &rec
->tool
;
788 struct record_opts
*opts
= &rec
->opts
;
789 struct perf_data_file
*file
= &rec
->file
;
790 struct perf_session
*session
;
791 bool disabled
= false, draining
= false;
794 rec
->progname
= argv
[0];
796 atexit(record__sig_exit
);
797 signal(SIGCHLD
, sig_handler
);
798 signal(SIGINT
, sig_handler
);
799 signal(SIGTERM
, sig_handler
);
801 if (rec
->opts
.auxtrace_snapshot_mode
|| rec
->switch_output
) {
802 signal(SIGUSR2
, snapshot_sig_handler
);
803 if (rec
->opts
.auxtrace_snapshot_mode
)
804 trigger_on(&auxtrace_snapshot_trigger
);
805 if (rec
->switch_output
)
806 trigger_on(&switch_output_trigger
);
808 signal(SIGUSR2
, SIG_IGN
);
811 session
= perf_session__new(file
, false, tool
);
812 if (session
== NULL
) {
813 pr_err("Perf session creation failed.\n");
817 fd
= perf_data_file__fd(file
);
818 rec
->session
= session
;
820 record__init_features(rec
);
823 err
= perf_evlist__prepare_workload(rec
->evlist
, &opts
->target
,
825 workload_exec_failed_signal
);
827 pr_err("Couldn't run the workload!\n");
829 goto out_delete_session
;
833 if (record__open(rec
) != 0) {
838 err
= bpf__apply_obj_config();
842 bpf__strerror_apply_obj_config(err
, errbuf
, sizeof(errbuf
));
843 pr_err("ERROR: Apply config to BPF failed: %s\n",
849 * Normally perf_session__new would do this, but it doesn't have the
852 if (rec
->tool
.ordered_events
&& !perf_evlist__sample_id_all(rec
->evlist
)) {
853 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
854 rec
->tool
.ordered_events
= false;
857 if (!rec
->evlist
->nr_groups
)
858 perf_header__clear_feat(&session
->header
, HEADER_GROUP_DESC
);
861 err
= perf_header__write_pipe(fd
);
865 err
= perf_session__write_header(session
, rec
->evlist
, fd
, false);
871 && !perf_header__has_feat(&session
->header
, HEADER_BUILD_ID
)) {
872 pr_err("Couldn't generate buildids. "
873 "Use --no-buildid to profile anyway.\n");
878 machine
= &session
->machines
.host
;
880 err
= record__synthesize(rec
);
884 if (rec
->realtime_prio
) {
885 struct sched_param param
;
887 param
.sched_priority
= rec
->realtime_prio
;
888 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
889 pr_err("Could not set realtime priority.\n");
896 * When perf is starting the traced process, all the events
897 * (apart from group members) have enable_on_exec=1 set,
898 * so don't spoil it by prematurely enabling them.
900 if (!target__none(&opts
->target
) && !opts
->initial_delay
)
901 perf_evlist__enable(rec
->evlist
);
907 union perf_event
*event
;
909 event
= malloc(sizeof(event
->comm
) + machine
->id_hdr_size
);
916 * Some H/W events are generated before COMM event
917 * which is emitted during exec(), so perf script
918 * cannot see a correct process name for those events.
919 * Synthesize COMM event to prevent it.
921 perf_event__synthesize_comm(tool
, event
,
922 rec
->evlist
->workload
.pid
,
923 process_synthesized_event
,
927 perf_evlist__start_workload(rec
->evlist
);
930 if (opts
->initial_delay
) {
931 usleep(opts
->initial_delay
* 1000);
932 perf_evlist__enable(rec
->evlist
);
935 trigger_ready(&auxtrace_snapshot_trigger
);
936 trigger_ready(&switch_output_trigger
);
938 unsigned long long hits
= rec
->samples
;
940 if (record__mmap_read_all(rec
) < 0) {
941 trigger_error(&auxtrace_snapshot_trigger
);
942 trigger_error(&switch_output_trigger
);
947 if (auxtrace_record__snapshot_started
) {
948 auxtrace_record__snapshot_started
= 0;
949 if (!trigger_is_error(&auxtrace_snapshot_trigger
))
950 record__read_auxtrace_snapshot(rec
);
951 if (trigger_is_error(&auxtrace_snapshot_trigger
)) {
952 pr_err("AUX area tracing snapshot failed\n");
958 if (trigger_is_hit(&switch_output_trigger
)) {
959 trigger_ready(&switch_output_trigger
);
962 fprintf(stderr
, "[ perf record: dump data: Woken up %ld times ]\n",
965 fd
= record__switch_output(rec
, false);
967 pr_err("Failed to switch to new file\n");
968 trigger_error(&switch_output_trigger
);
974 if (hits
== rec
->samples
) {
975 if (done
|| draining
)
977 err
= perf_evlist__poll(rec
->evlist
, -1);
979 * Propagate error, only if there's any. Ignore positive
980 * number of returned events and interrupt error.
982 if (err
> 0 || (err
< 0 && errno
== EINTR
))
986 if (perf_evlist__filter_pollfd(rec
->evlist
, POLLERR
| POLLHUP
) == 0)
991 * When perf is starting the traced process, at the end events
992 * die with the process and we wait for that. Thus no need to
993 * disable events in this case.
995 if (done
&& !disabled
&& !target__none(&opts
->target
)) {
996 trigger_off(&auxtrace_snapshot_trigger
);
997 perf_evlist__disable(rec
->evlist
);
1001 trigger_off(&auxtrace_snapshot_trigger
);
1002 trigger_off(&switch_output_trigger
);
1004 if (forks
&& workload_exec_errno
) {
1005 char msg
[STRERR_BUFSIZE
];
1006 const char *emsg
= strerror_r(workload_exec_errno
, msg
, sizeof(msg
));
1007 pr_err("Workload failed: %s\n", emsg
);
1013 fprintf(stderr
, "[ perf record: Woken up %ld times to write data ]\n", waking
);
1019 if (!child_finished
)
1020 kill(rec
->evlist
->workload
.pid
, SIGTERM
);
1026 else if (WIFEXITED(exit_status
))
1027 status
= WEXITSTATUS(exit_status
);
1028 else if (WIFSIGNALED(exit_status
))
1029 signr
= WTERMSIG(exit_status
);
1033 /* this will be recalculated during process_buildids() */
1037 if (!rec
->timestamp_filename
) {
1038 record__finish_output(rec
);
1040 fd
= record__switch_output(rec
, true);
1043 goto out_delete_session
;
1048 if (!err
&& !quiet
) {
1050 const char *postfix
= rec
->timestamp_filename
?
1051 ".<timestamp>" : "";
1053 if (rec
->samples
&& !rec
->opts
.full_auxtrace
)
1054 scnprintf(samples
, sizeof(samples
),
1055 " (%" PRIu64
" samples)", rec
->samples
);
1059 fprintf(stderr
, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1060 perf_data_file__size(file
) / 1024.0 / 1024.0,
1061 file
->path
, postfix
, samples
);
1065 perf_session__delete(session
);
1069 static void callchain_debug(struct callchain_param
*callchain
)
1071 static const char *str
[CALLCHAIN_MAX
] = { "NONE", "FP", "DWARF", "LBR" };
1073 pr_debug("callchain: type %s\n", str
[callchain
->record_mode
]);
1075 if (callchain
->record_mode
== CALLCHAIN_DWARF
)
1076 pr_debug("callchain: stack dump size %d\n",
1077 callchain
->dump_size
);
1080 int record_opts__parse_callchain(struct record_opts
*record
,
1081 struct callchain_param
*callchain
,
1082 const char *arg
, bool unset
)
1085 callchain
->enabled
= !unset
;
1087 /* --no-call-graph */
1089 callchain
->record_mode
= CALLCHAIN_NONE
;
1090 pr_debug("callchain: disabled\n");
1094 ret
= parse_callchain_record_opt(arg
, callchain
);
1096 /* Enable data address sampling for DWARF unwind. */
1097 if (callchain
->record_mode
== CALLCHAIN_DWARF
)
1098 record
->sample_address
= true;
1099 callchain_debug(callchain
);
1105 int record_parse_callchain_opt(const struct option
*opt
,
1109 return record_opts__parse_callchain(opt
->value
, &callchain_param
, arg
, unset
);
1112 int record_callchain_opt(const struct option
*opt
,
1113 const char *arg __maybe_unused
,
1114 int unset __maybe_unused
)
1116 struct callchain_param
*callchain
= opt
->value
;
1118 callchain
->enabled
= true;
1120 if (callchain
->record_mode
== CALLCHAIN_NONE
)
1121 callchain
->record_mode
= CALLCHAIN_FP
;
1123 callchain_debug(callchain
);
1127 static int perf_record_config(const char *var
, const char *value
, void *cb
)
1129 struct record
*rec
= cb
;
1131 if (!strcmp(var
, "record.build-id")) {
1132 if (!strcmp(value
, "cache"))
1133 rec
->no_buildid_cache
= false;
1134 else if (!strcmp(value
, "no-cache"))
1135 rec
->no_buildid_cache
= true;
1136 else if (!strcmp(value
, "skip"))
1137 rec
->no_buildid
= true;
1142 if (!strcmp(var
, "record.call-graph"))
1143 var
= "call-graph.record-mode"; /* fall-through */
1145 return perf_default_config(var
, value
, cb
);
1148 struct clockid_map
{
1153 #define CLOCKID_MAP(n, c) \
1154 { .name = n, .clockid = (c), }
1156 #define CLOCKID_END { .name = NULL, }
1160 * Add the missing ones, we need to build on many distros...
1162 #ifndef CLOCK_MONOTONIC_RAW
1163 #define CLOCK_MONOTONIC_RAW 4
1165 #ifndef CLOCK_BOOTTIME
1166 #define CLOCK_BOOTTIME 7
1169 #define CLOCK_TAI 11
1172 static const struct clockid_map clockids
[] = {
1173 /* available for all events, NMI safe */
1174 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC
),
1175 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW
),
1177 /* available for some events */
1178 CLOCKID_MAP("realtime", CLOCK_REALTIME
),
1179 CLOCKID_MAP("boottime", CLOCK_BOOTTIME
),
1180 CLOCKID_MAP("tai", CLOCK_TAI
),
1182 /* available for the lazy */
1183 CLOCKID_MAP("mono", CLOCK_MONOTONIC
),
1184 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW
),
1185 CLOCKID_MAP("real", CLOCK_REALTIME
),
1186 CLOCKID_MAP("boot", CLOCK_BOOTTIME
),
1191 static int parse_clockid(const struct option
*opt
, const char *str
, int unset
)
1193 struct record_opts
*opts
= (struct record_opts
*)opt
->value
;
1194 const struct clockid_map
*cm
;
1195 const char *ostr
= str
;
1198 opts
->use_clockid
= 0;
1206 /* no setting it twice */
1207 if (opts
->use_clockid
)
1210 opts
->use_clockid
= true;
1212 /* if its a number, we're done */
1213 if (sscanf(str
, "%d", &opts
->clockid
) == 1)
1216 /* allow a "CLOCK_" prefix to the name */
1217 if (!strncasecmp(str
, "CLOCK_", 6))
1220 for (cm
= clockids
; cm
->name
; cm
++) {
1221 if (!strcasecmp(str
, cm
->name
)) {
1222 opts
->clockid
= cm
->clockid
;
1227 opts
->use_clockid
= false;
1228 ui__warning("unknown clockid %s, check man page\n", ostr
);
1232 static int record__parse_mmap_pages(const struct option
*opt
,
1234 int unset __maybe_unused
)
1236 struct record_opts
*opts
= opt
->value
;
1238 unsigned int mmap_pages
;
1253 ret
= __perf_evlist__parse_mmap_pages(&mmap_pages
, s
);
1256 opts
->mmap_pages
= mmap_pages
;
1264 ret
= __perf_evlist__parse_mmap_pages(&mmap_pages
, p
+ 1);
1268 opts
->auxtrace_mmap_pages
= mmap_pages
;
1275 static const char * const __record_usage
[] = {
1276 "perf record [<options>] [<command>]",
1277 "perf record [<options>] -- <command> [<options>]",
1280 const char * const *record_usage
= __record_usage
;
1283 * XXX Ideally would be local to cmd_record() and passed to a record__new
1284 * because we need to have access to it in record__exit, that is called
1285 * after cmd_record() exits, but since record_options need to be accessible to
1286 * builtin-script, leave it here.
1288 * At least we don't ouch it in all the other functions here directly.
1290 * Just say no to tons of global variables, sigh.
1292 static struct record record
= {
1294 .sample_time
= true,
1295 .mmap_pages
= UINT_MAX
,
1296 .user_freq
= UINT_MAX
,
1297 .user_interval
= ULLONG_MAX
,
1301 .default_per_cpu
= true,
1303 .proc_map_timeout
= 500,
1306 .sample
= process_sample_event
,
1307 .fork
= perf_event__process_fork
,
1308 .exit
= perf_event__process_exit
,
1309 .comm
= perf_event__process_comm
,
1310 .mmap
= perf_event__process_mmap
,
1311 .mmap2
= perf_event__process_mmap2
,
1312 .ordered_events
= true,
1316 const char record_callchain_help
[] = CALLCHAIN_RECORD_HELP
1317 "\n\t\t\t\tDefault: fp";
1319 static bool dry_run
;
1322 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1323 * with it and switch to use the library functions in perf_evlist that came
1324 * from builtin-record.c, i.e. use record_opts,
1325 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1328 struct option __record_options
[] = {
1329 OPT_CALLBACK('e', "event", &record
.evlist
, "event",
1330 "event selector. use 'perf list' to list available events",
1331 parse_events_option
),
1332 OPT_CALLBACK(0, "filter", &record
.evlist
, "filter",
1333 "event filter", parse_filter
),
1334 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record
.evlist
,
1335 NULL
, "don't record events from perf itself",
1337 OPT_STRING('p', "pid", &record
.opts
.target
.pid
, "pid",
1338 "record events on existing process id"),
1339 OPT_STRING('t', "tid", &record
.opts
.target
.tid
, "tid",
1340 "record events on existing thread id"),
1341 OPT_INTEGER('r', "realtime", &record
.realtime_prio
,
1342 "collect data with this RT SCHED_FIFO priority"),
1343 OPT_BOOLEAN(0, "no-buffering", &record
.opts
.no_buffering
,
1344 "collect data without buffering"),
1345 OPT_BOOLEAN('R', "raw-samples", &record
.opts
.raw_samples
,
1346 "collect raw sample records from all opened counters"),
1347 OPT_BOOLEAN('a', "all-cpus", &record
.opts
.target
.system_wide
,
1348 "system-wide collection from all CPUs"),
1349 OPT_STRING('C', "cpu", &record
.opts
.target
.cpu_list
, "cpu",
1350 "list of cpus to monitor"),
1351 OPT_U64('c', "count", &record
.opts
.user_interval
, "event period to sample"),
1352 OPT_STRING('o', "output", &record
.file
.path
, "file",
1353 "output file name"),
1354 OPT_BOOLEAN_SET('i', "no-inherit", &record
.opts
.no_inherit
,
1355 &record
.opts
.no_inherit_set
,
1356 "child tasks do not inherit counters"),
1357 OPT_UINTEGER('F', "freq", &record
.opts
.user_freq
, "profile at this frequency"),
1358 OPT_CALLBACK('m', "mmap-pages", &record
.opts
, "pages[,pages]",
1359 "number of mmap data pages and AUX area tracing mmap pages",
1360 record__parse_mmap_pages
),
1361 OPT_BOOLEAN(0, "group", &record
.opts
.group
,
1362 "put the counters into a counter group"),
1363 OPT_CALLBACK_NOOPT('g', NULL
, &callchain_param
,
1364 NULL
, "enables call-graph recording" ,
1365 &record_callchain_opt
),
1366 OPT_CALLBACK(0, "call-graph", &record
.opts
,
1367 "record_mode[,record_size]", record_callchain_help
,
1368 &record_parse_callchain_opt
),
1369 OPT_INCR('v', "verbose", &verbose
,
1370 "be more verbose (show counter open errors, etc)"),
1371 OPT_BOOLEAN('q', "quiet", &quiet
, "don't print any message"),
1372 OPT_BOOLEAN('s', "stat", &record
.opts
.inherit_stat
,
1373 "per thread counts"),
1374 OPT_BOOLEAN('d', "data", &record
.opts
.sample_address
, "Record the sample addresses"),
1375 OPT_BOOLEAN_SET('T', "timestamp", &record
.opts
.sample_time
,
1376 &record
.opts
.sample_time_set
,
1377 "Record the sample timestamps"),
1378 OPT_BOOLEAN('P', "period", &record
.opts
.period
, "Record the sample period"),
1379 OPT_BOOLEAN('n', "no-samples", &record
.opts
.no_samples
,
1381 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record
.no_buildid_cache
,
1382 &record
.no_buildid_cache_set
,
1383 "do not update the buildid cache"),
1384 OPT_BOOLEAN_SET('B', "no-buildid", &record
.no_buildid
,
1385 &record
.no_buildid_set
,
1386 "do not collect buildids in perf.data"),
1387 OPT_CALLBACK('G', "cgroup", &record
.evlist
, "name",
1388 "monitor event in cgroup name only",
1390 OPT_UINTEGER('D', "delay", &record
.opts
.initial_delay
,
1391 "ms to wait before starting measurement after program start"),
1392 OPT_STRING('u', "uid", &record
.opts
.target
.uid_str
, "user",
1395 OPT_CALLBACK_NOOPT('b', "branch-any", &record
.opts
.branch_stack
,
1396 "branch any", "sample any taken branches",
1397 parse_branch_stack
),
1399 OPT_CALLBACK('j', "branch-filter", &record
.opts
.branch_stack
,
1400 "branch filter mask", "branch stack filter modes",
1401 parse_branch_stack
),
1402 OPT_BOOLEAN('W', "weight", &record
.opts
.sample_weight
,
1403 "sample by weight (on special events only)"),
1404 OPT_BOOLEAN(0, "transaction", &record
.opts
.sample_transaction
,
1405 "sample transaction flags (special events only)"),
1406 OPT_BOOLEAN(0, "per-thread", &record
.opts
.target
.per_thread
,
1407 "use per-thread mmaps"),
1408 OPT_CALLBACK_OPTARG('I', "intr-regs", &record
.opts
.sample_intr_regs
, NULL
, "any register",
1409 "sample selected machine registers on interrupt,"
1410 " use -I ? to list register names", parse_regs
),
1411 OPT_BOOLEAN(0, "running-time", &record
.opts
.running_time
,
1412 "Record running/enabled time of read (:S) events"),
1413 OPT_CALLBACK('k', "clockid", &record
.opts
,
1414 "clockid", "clockid to use for events, see clock_gettime()",
1416 OPT_STRING_OPTARG('S', "snapshot", &record
.opts
.auxtrace_snapshot_opts
,
1417 "opts", "AUX area tracing Snapshot Mode", ""),
1418 OPT_UINTEGER(0, "proc-map-timeout", &record
.opts
.proc_map_timeout
,
1419 "per thread proc mmap processing timeout in ms"),
1420 OPT_BOOLEAN(0, "switch-events", &record
.opts
.record_switch_events
,
1421 "Record context switch events"),
1422 OPT_BOOLEAN_FLAG(0, "all-kernel", &record
.opts
.all_kernel
,
1423 "Configure all used events to run in kernel space.",
1424 PARSE_OPT_EXCLUSIVE
),
1425 OPT_BOOLEAN_FLAG(0, "all-user", &record
.opts
.all_user
,
1426 "Configure all used events to run in user space.",
1427 PARSE_OPT_EXCLUSIVE
),
1428 OPT_STRING(0, "clang-path", &llvm_param
.clang_path
, "clang path",
1429 "clang binary to use for compiling BPF scriptlets"),
1430 OPT_STRING(0, "clang-opt", &llvm_param
.clang_opt
, "clang options",
1431 "options passed to clang when compiling BPF scriptlets"),
1432 OPT_STRING(0, "vmlinux", &symbol_conf
.vmlinux_name
,
1433 "file", "vmlinux pathname"),
1434 OPT_BOOLEAN(0, "buildid-all", &record
.buildid_all
,
1435 "Record build-id of all DSOs regardless of hits"),
1436 OPT_BOOLEAN(0, "timestamp-filename", &record
.timestamp_filename
,
1437 "append timestamp to output filename"),
1438 OPT_BOOLEAN(0, "switch-output", &record
.switch_output
,
1439 "Switch output when receive SIGUSR2"),
1440 OPT_BOOLEAN(0, "dry-run", &dry_run
,
1441 "Parse options then exit"),
1445 struct option
*record_options
= __record_options
;
1447 int cmd_record(int argc
, const char **argv
, const char *prefix __maybe_unused
)
1450 struct record
*rec
= &record
;
1451 char errbuf
[BUFSIZ
];
1453 #ifndef HAVE_LIBBPF_SUPPORT
1454 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
1455 set_nobuild('\0', "clang-path", true);
1456 set_nobuild('\0', "clang-opt", true);
1460 #ifndef HAVE_BPF_PROLOGUE
1461 # if !defined (HAVE_DWARF_SUPPORT)
1462 # define REASON "NO_DWARF=1"
1463 # elif !defined (HAVE_LIBBPF_SUPPORT)
1464 # define REASON "NO_LIBBPF=1"
1466 # define REASON "this architecture doesn't support BPF prologue"
1468 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
1469 set_nobuild('\0', "vmlinux", true);
1474 rec
->evlist
= perf_evlist__new();
1475 if (rec
->evlist
== NULL
)
1478 perf_config(perf_record_config
, rec
);
1480 argc
= parse_options(argc
, argv
, record_options
, record_usage
,
1481 PARSE_OPT_STOP_AT_NON_OPTION
);
1482 if (!argc
&& target__none(&rec
->opts
.target
))
1483 usage_with_options(record_usage
, record_options
);
1485 if (nr_cgroups
&& !rec
->opts
.target
.system_wide
) {
1486 usage_with_options_msg(record_usage
, record_options
,
1487 "cgroup monitoring only available in system-wide mode");
1490 if (rec
->opts
.record_switch_events
&&
1491 !perf_can_record_switch_events()) {
1492 ui__error("kernel does not support recording context switch events\n");
1493 parse_options_usage(record_usage
, record_options
, "switch-events", 0);
1497 if (rec
->switch_output
)
1498 rec
->timestamp_filename
= true;
1501 rec
->itr
= auxtrace_record__init(rec
->evlist
, &err
);
1506 err
= auxtrace_parse_snapshot_options(rec
->itr
, &rec
->opts
,
1507 rec
->opts
.auxtrace_snapshot_opts
);
1514 err
= bpf__setup_stdout(rec
->evlist
);
1516 bpf__strerror_setup_stdout(rec
->evlist
, err
, errbuf
, sizeof(errbuf
));
1517 pr_err("ERROR: Setup BPF stdout failed: %s\n",
1526 if (symbol_conf
.kptr_restrict
)
1528 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
1529 "check /proc/sys/kernel/kptr_restrict.\n\n"
1530 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1531 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1532 "Samples in kernel modules won't be resolved at all.\n\n"
1533 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1534 "even with a suitable vmlinux or kallsyms file.\n\n");
1536 if (rec
->no_buildid_cache
|| rec
->no_buildid
) {
1537 disable_buildid_cache();
1538 } else if (rec
->switch_output
) {
1540 * In 'perf record --switch-output', disable buildid
1541 * generation by default to reduce data file switching
1542 * overhead. Still generate buildid if they are required
1545 * perf record --signal-trigger --no-no-buildid \
1546 * --no-no-buildid-cache
1548 * Following code equals to:
1550 * if ((rec->no_buildid || !rec->no_buildid_set) &&
1551 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
1552 * disable_buildid_cache();
1554 bool disable
= true;
1556 if (rec
->no_buildid_set
&& !rec
->no_buildid
)
1558 if (rec
->no_buildid_cache_set
&& !rec
->no_buildid_cache
)
1561 rec
->no_buildid
= true;
1562 rec
->no_buildid_cache
= true;
1563 disable_buildid_cache();
1567 if (rec
->evlist
->nr_entries
== 0 &&
1568 perf_evlist__add_default(rec
->evlist
) < 0) {
1569 pr_err("Not enough memory for event selector list\n");
1570 goto out_symbol_exit
;
1573 if (rec
->opts
.target
.tid
&& !rec
->opts
.no_inherit_set
)
1574 rec
->opts
.no_inherit
= true;
1576 err
= target__validate(&rec
->opts
.target
);
1578 target__strerror(&rec
->opts
.target
, err
, errbuf
, BUFSIZ
);
1579 ui__warning("%s", errbuf
);
1582 err
= target__parse_uid(&rec
->opts
.target
);
1584 int saved_errno
= errno
;
1586 target__strerror(&rec
->opts
.target
, err
, errbuf
, BUFSIZ
);
1587 ui__error("%s", errbuf
);
1590 goto out_symbol_exit
;
1594 if (perf_evlist__create_maps(rec
->evlist
, &rec
->opts
.target
) < 0)
1595 usage_with_options(record_usage
, record_options
);
1597 err
= auxtrace_record__options(rec
->itr
, rec
->evlist
, &rec
->opts
);
1599 goto out_symbol_exit
;
1602 * We take all buildids when the file contains
1603 * AUX area tracing data because we do not decode the
1604 * trace because it would take too long.
1606 if (rec
->opts
.full_auxtrace
)
1607 rec
->buildid_all
= true;
1609 if (record_opts__config(&rec
->opts
)) {
1611 goto out_symbol_exit
;
1614 err
= __cmd_record(&record
, argc
, argv
);
1616 perf_evlist__delete(rec
->evlist
);
1618 auxtrace_record__free(rec
->itr
);
1622 static void snapshot_sig_handler(int sig __maybe_unused
)
1624 if (trigger_is_ready(&auxtrace_snapshot_trigger
)) {
1625 trigger_hit(&auxtrace_snapshot_trigger
);
1626 auxtrace_record__snapshot_started
= 1;
1627 if (auxtrace_record__snapshot_start(record
.itr
))
1628 trigger_error(&auxtrace_snapshot_trigger
);
1631 if (trigger_is_ready(&switch_output_trigger
))
1632 trigger_hit(&switch_output_trigger
);