1 // SPDX-License-Identifier: GPL-2.0
5 * Builtin record command: Record the profile of a workload
6 * (or a CPU, or a PID) into the perf.data output file - for
7 * later analysis via perf report.
13 #include "util/build-id.h"
14 #include "util/util.h"
15 #include <subcmd/parse-options.h>
16 #include "util/parse-events.h"
17 #include "util/config.h"
19 #include "util/callchain.h"
20 #include "util/cgroup.h"
21 #include "util/header.h"
22 #include "util/event.h"
23 #include "util/evlist.h"
24 #include "util/evsel.h"
25 #include "util/debug.h"
26 #include "util/session.h"
27 #include "util/tool.h"
28 #include "util/symbol.h"
29 #include "util/cpumap.h"
30 #include "util/thread_map.h"
31 #include "util/data.h"
32 #include "util/perf_regs.h"
33 #include "util/auxtrace.h"
35 #include "util/parse-branch-options.h"
36 #include "util/parse-regs-options.h"
37 #include "util/llvm-utils.h"
38 #include "util/bpf-loader.h"
39 #include "util/trigger.h"
40 #include "util/perf-hooks.h"
41 #include "util/cpu-set-sched.h"
42 #include "util/time-utils.h"
43 #include "util/units.h"
44 #include "util/bpf-event.h"
56 #include <linux/time64.h>
58 struct switch_output
{
71 struct perf_tool tool
;
72 struct record_opts opts
;
74 struct perf_data data
;
75 struct auxtrace_record
*itr
;
76 struct perf_evlist
*evlist
;
77 struct perf_session
*session
;
81 bool no_buildid_cache
;
82 bool no_buildid_cache_set
;
84 bool timestamp_filename
;
85 bool timestamp_boundary
;
86 struct switch_output switch_output
;
87 unsigned long long samples
;
88 cpu_set_t affinity_mask
;
91 static volatile int auxtrace_record__snapshot_started
;
92 static DEFINE_TRIGGER(auxtrace_snapshot_trigger
);
93 static DEFINE_TRIGGER(switch_output_trigger
);
95 static const char *affinity_tags
[PERF_AFFINITY_MAX
] = {
99 static bool switch_output_signal(struct record
*rec
)
101 return rec
->switch_output
.signal
&&
102 trigger_is_ready(&switch_output_trigger
);
105 static bool switch_output_size(struct record
*rec
)
107 return rec
->switch_output
.size
&&
108 trigger_is_ready(&switch_output_trigger
) &&
109 (rec
->bytes_written
>= rec
->switch_output
.size
);
112 static bool switch_output_time(struct record
*rec
)
114 return rec
->switch_output
.time
&&
115 trigger_is_ready(&switch_output_trigger
);
118 static int record__write(struct record
*rec
, struct perf_mmap
*map __maybe_unused
,
119 void *bf
, size_t size
)
121 struct perf_data_file
*file
= &rec
->session
->data
->file
;
123 if (perf_data_file__write(file
, bf
, size
) < 0) {
124 pr_err("failed to write perf data, error: %m\n");
128 rec
->bytes_written
+= size
;
130 if (switch_output_size(rec
))
131 trigger_hit(&switch_output_trigger
);
136 #ifdef HAVE_AIO_SUPPORT
137 static int record__aio_write(struct aiocb
*cblock
, int trace_fd
,
138 void *buf
, size_t size
, off_t off
)
142 cblock
->aio_fildes
= trace_fd
;
143 cblock
->aio_buf
= buf
;
144 cblock
->aio_nbytes
= size
;
145 cblock
->aio_offset
= off
;
146 cblock
->aio_sigevent
.sigev_notify
= SIGEV_NONE
;
149 rc
= aio_write(cblock
);
152 } else if (errno
!= EAGAIN
) {
153 cblock
->aio_fildes
= -1;
154 pr_err("failed to queue perf data, error: %m\n");
162 static int record__aio_complete(struct perf_mmap
*md
, struct aiocb
*cblock
)
168 ssize_t aio_ret
, written
;
170 aio_errno
= aio_error(cblock
);
171 if (aio_errno
== EINPROGRESS
)
174 written
= aio_ret
= aio_return(cblock
);
176 if (aio_errno
!= EINTR
)
177 pr_err("failed to write perf data, error: %m\n");
181 rem_size
= cblock
->aio_nbytes
- written
;
184 cblock
->aio_fildes
= -1;
186 * md->refcount is incremented in perf_mmap__push() for
187 * every enqueued aio write request so decrement it because
188 * the request is now complete.
194 * aio write request may require restart with the
195 * reminder if the kernel didn't write whole
198 rem_off
= cblock
->aio_offset
+ written
;
199 rem_buf
= (void *)(cblock
->aio_buf
+ written
);
200 record__aio_write(cblock
, cblock
->aio_fildes
,
201 rem_buf
, rem_size
, rem_off
);
208 static int record__aio_sync(struct perf_mmap
*md
, bool sync_all
)
210 struct aiocb
**aiocb
= md
->aio
.aiocb
;
211 struct aiocb
*cblocks
= md
->aio
.cblocks
;
212 struct timespec timeout
= { 0, 1000 * 1000 * 1 }; /* 1ms */
217 for (i
= 0; i
< md
->aio
.nr_cblocks
; ++i
) {
218 if (cblocks
[i
].aio_fildes
== -1 || record__aio_complete(md
, &cblocks
[i
])) {
225 * Started aio write is not complete yet
226 * so it has to be waited before the
229 aiocb
[i
] = &cblocks
[i
];
236 while (aio_suspend((const struct aiocb
**)aiocb
, md
->aio
.nr_cblocks
, &timeout
)) {
237 if (!(errno
== EAGAIN
|| errno
== EINTR
))
238 pr_err("failed to sync perf data, error: %m\n");
243 static int record__aio_pushfn(void *to
, struct aiocb
*cblock
, void *bf
, size_t size
, off_t off
)
245 struct record
*rec
= to
;
246 int ret
, trace_fd
= rec
->session
->data
->file
.fd
;
250 ret
= record__aio_write(cblock
, trace_fd
, bf
, size
, off
);
252 rec
->bytes_written
+= size
;
253 if (switch_output_size(rec
))
254 trigger_hit(&switch_output_trigger
);
260 static off_t
record__aio_get_pos(int trace_fd
)
262 return lseek(trace_fd
, 0, SEEK_CUR
);
265 static void record__aio_set_pos(int trace_fd
, off_t pos
)
267 lseek(trace_fd
, pos
, SEEK_SET
);
270 static void record__aio_mmap_read_sync(struct record
*rec
)
273 struct perf_evlist
*evlist
= rec
->evlist
;
274 struct perf_mmap
*maps
= evlist
->mmap
;
276 if (!rec
->opts
.nr_cblocks
)
279 for (i
= 0; i
< evlist
->nr_mmaps
; i
++) {
280 struct perf_mmap
*map
= &maps
[i
];
283 record__aio_sync(map
, true);
287 static int nr_cblocks_default
= 1;
288 static int nr_cblocks_max
= 4;
290 static int record__aio_parse(const struct option
*opt
,
294 struct record_opts
*opts
= (struct record_opts
*)opt
->value
;
297 opts
->nr_cblocks
= 0;
300 opts
->nr_cblocks
= strtol(str
, NULL
, 0);
301 if (!opts
->nr_cblocks
)
302 opts
->nr_cblocks
= nr_cblocks_default
;
307 #else /* HAVE_AIO_SUPPORT */
308 static int nr_cblocks_max
= 0;
310 static int record__aio_sync(struct perf_mmap
*md __maybe_unused
, bool sync_all __maybe_unused
)
315 static int record__aio_pushfn(void *to __maybe_unused
, struct aiocb
*cblock __maybe_unused
,
316 void *bf __maybe_unused
, size_t size __maybe_unused
, off_t off __maybe_unused
)
321 static off_t
record__aio_get_pos(int trace_fd __maybe_unused
)
326 static void record__aio_set_pos(int trace_fd __maybe_unused
, off_t pos __maybe_unused
)
330 static void record__aio_mmap_read_sync(struct record
*rec __maybe_unused
)
335 static int record__aio_enabled(struct record
*rec
)
337 return rec
->opts
.nr_cblocks
> 0;
340 static int process_synthesized_event(struct perf_tool
*tool
,
341 union perf_event
*event
,
342 struct perf_sample
*sample __maybe_unused
,
343 struct machine
*machine __maybe_unused
)
345 struct record
*rec
= container_of(tool
, struct record
, tool
);
346 return record__write(rec
, NULL
, event
, event
->header
.size
);
349 static int record__pushfn(struct perf_mmap
*map
, void *to
, void *bf
, size_t size
)
351 struct record
*rec
= to
;
354 return record__write(rec
, map
, bf
, size
);
357 static volatile int done
;
358 static volatile int signr
= -1;
359 static volatile int child_finished
;
361 static void sig_handler(int sig
)
371 static void sigsegv_handler(int sig
)
373 perf_hooks__recover();
374 sighandler_dump_stack(sig
);
377 static void record__sig_exit(void)
382 signal(signr
, SIG_DFL
);
386 #ifdef HAVE_AUXTRACE_SUPPORT
388 static int record__process_auxtrace(struct perf_tool
*tool
,
389 struct perf_mmap
*map
,
390 union perf_event
*event
, void *data1
,
391 size_t len1
, void *data2
, size_t len2
)
393 struct record
*rec
= container_of(tool
, struct record
, tool
);
394 struct perf_data
*data
= &rec
->data
;
398 if (!perf_data__is_pipe(data
) && !perf_data__is_dir(data
)) {
400 int fd
= perf_data__fd(data
);
403 file_offset
= lseek(fd
, 0, SEEK_CUR
);
404 if (file_offset
== -1)
406 err
= auxtrace_index__auxtrace_event(&rec
->session
->auxtrace_index
,
412 /* event.auxtrace.size includes padding, see __auxtrace_mmap__read() */
413 padding
= (len1
+ len2
) & 7;
415 padding
= 8 - padding
;
417 record__write(rec
, map
, event
, event
->header
.size
);
418 record__write(rec
, map
, data1
, len1
);
420 record__write(rec
, map
, data2
, len2
);
421 record__write(rec
, map
, &pad
, padding
);
426 static int record__auxtrace_mmap_read(struct record
*rec
,
427 struct perf_mmap
*map
)
431 ret
= auxtrace_mmap__read(map
, rec
->itr
, &rec
->tool
,
432 record__process_auxtrace
);
442 static int record__auxtrace_mmap_read_snapshot(struct record
*rec
,
443 struct perf_mmap
*map
)
447 ret
= auxtrace_mmap__read_snapshot(map
, rec
->itr
, &rec
->tool
,
448 record__process_auxtrace
,
449 rec
->opts
.auxtrace_snapshot_size
);
459 static int record__auxtrace_read_snapshot_all(struct record
*rec
)
464 for (i
= 0; i
< rec
->evlist
->nr_mmaps
; i
++) {
465 struct perf_mmap
*map
= &rec
->evlist
->mmap
[i
];
467 if (!map
->auxtrace_mmap
.base
)
470 if (record__auxtrace_mmap_read_snapshot(rec
, map
) != 0) {
479 static void record__read_auxtrace_snapshot(struct record
*rec
)
481 pr_debug("Recording AUX area tracing snapshot\n");
482 if (record__auxtrace_read_snapshot_all(rec
) < 0) {
483 trigger_error(&auxtrace_snapshot_trigger
);
485 if (auxtrace_record__snapshot_finish(rec
->itr
))
486 trigger_error(&auxtrace_snapshot_trigger
);
488 trigger_ready(&auxtrace_snapshot_trigger
);
492 static int record__auxtrace_init(struct record
*rec
)
497 rec
->itr
= auxtrace_record__init(rec
->evlist
, &err
);
502 err
= auxtrace_parse_snapshot_options(rec
->itr
, &rec
->opts
,
503 rec
->opts
.auxtrace_snapshot_opts
);
507 return auxtrace_parse_filters(rec
->evlist
);
513 int record__auxtrace_mmap_read(struct record
*rec __maybe_unused
,
514 struct perf_mmap
*map __maybe_unused
)
520 void record__read_auxtrace_snapshot(struct record
*rec __maybe_unused
)
525 int auxtrace_record__snapshot_start(struct auxtrace_record
*itr __maybe_unused
)
530 static int record__auxtrace_init(struct record
*rec __maybe_unused
)
537 static int record__mmap_evlist(struct record
*rec
,
538 struct perf_evlist
*evlist
)
540 struct record_opts
*opts
= &rec
->opts
;
543 if (opts
->affinity
!= PERF_AFFINITY_SYS
)
544 cpu__setup_cpunode_map();
546 if (perf_evlist__mmap_ex(evlist
, opts
->mmap_pages
,
547 opts
->auxtrace_mmap_pages
,
548 opts
->auxtrace_snapshot_mode
,
549 opts
->nr_cblocks
, opts
->affinity
) < 0) {
550 if (errno
== EPERM
) {
551 pr_err("Permission error mapping pages.\n"
552 "Consider increasing "
553 "/proc/sys/kernel/perf_event_mlock_kb,\n"
554 "or try again with a smaller value of -m/--mmap_pages.\n"
555 "(current value: %u,%u)\n",
556 opts
->mmap_pages
, opts
->auxtrace_mmap_pages
);
559 pr_err("failed to mmap with %d (%s)\n", errno
,
560 str_error_r(errno
, msg
, sizeof(msg
)));
570 static int record__mmap(struct record
*rec
)
572 return record__mmap_evlist(rec
, rec
->evlist
);
575 static int record__open(struct record
*rec
)
578 struct perf_evsel
*pos
;
579 struct perf_evlist
*evlist
= rec
->evlist
;
580 struct perf_session
*session
= rec
->session
;
581 struct record_opts
*opts
= &rec
->opts
;
585 * For initial_delay we need to add a dummy event so that we can track
586 * PERF_RECORD_MMAP while we wait for the initial delay to enable the
587 * real events, the ones asked by the user.
589 if (opts
->initial_delay
) {
590 if (perf_evlist__add_dummy(evlist
))
593 pos
= perf_evlist__first(evlist
);
595 pos
= perf_evlist__last(evlist
);
597 pos
->attr
.enable_on_exec
= 1;
600 perf_evlist__config(evlist
, opts
, &callchain_param
);
602 evlist__for_each_entry(evlist
, pos
) {
604 if (perf_evsel__open(pos
, pos
->cpus
, pos
->threads
) < 0) {
605 if (perf_evsel__fallback(pos
, errno
, msg
, sizeof(msg
))) {
607 ui__warning("%s\n", msg
);
610 if ((errno
== EINVAL
|| errno
== EBADF
) &&
611 pos
->leader
!= pos
&&
613 pos
= perf_evlist__reset_weak_group(evlist
, pos
);
617 perf_evsel__open_strerror(pos
, &opts
->target
,
618 errno
, msg
, sizeof(msg
));
619 ui__error("%s\n", msg
);
623 pos
->supported
= true;
626 if (perf_evlist__apply_filters(evlist
, &pos
)) {
627 pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
628 pos
->filter
, perf_evsel__name(pos
), errno
,
629 str_error_r(errno
, msg
, sizeof(msg
)));
634 rc
= record__mmap(rec
);
638 session
->evlist
= evlist
;
639 perf_session__set_id_hdr_size(session
);
644 static int process_sample_event(struct perf_tool
*tool
,
645 union perf_event
*event
,
646 struct perf_sample
*sample
,
647 struct perf_evsel
*evsel
,
648 struct machine
*machine
)
650 struct record
*rec
= container_of(tool
, struct record
, tool
);
652 if (rec
->evlist
->first_sample_time
== 0)
653 rec
->evlist
->first_sample_time
= sample
->time
;
655 rec
->evlist
->last_sample_time
= sample
->time
;
657 if (rec
->buildid_all
)
661 return build_id__mark_dso_hit(tool
, event
, sample
, evsel
, machine
);
664 static int process_buildids(struct record
*rec
)
666 struct perf_session
*session
= rec
->session
;
668 if (perf_data__size(&rec
->data
) == 0)
672 * During this process, it'll load kernel map and replace the
673 * dso->long_name to a real pathname it found. In this case
674 * we prefer the vmlinux path like
675 * /lib/modules/3.16.4/build/vmlinux
677 * rather than build-id path (in debug directory).
678 * $HOME/.debug/.build-id/f0/6e17aa50adf4d00b88925e03775de107611551
680 symbol_conf
.ignore_vmlinux_buildid
= true;
683 * If --buildid-all is given, it marks all DSO regardless of hits,
684 * so no need to process samples. But if timestamp_boundary is enabled,
685 * it still needs to walk on all samples to get the timestamps of
686 * first/last samples.
688 if (rec
->buildid_all
&& !rec
->timestamp_boundary
)
689 rec
->tool
.sample
= NULL
;
691 return perf_session__process_events(session
);
694 static void perf_event__synthesize_guest_os(struct machine
*machine
, void *data
)
697 struct perf_tool
*tool
= data
;
699 *As for guest kernel when processing subcommand record&report,
700 *we arrange module mmap prior to guest kernel mmap and trigger
701 *a preload dso because default guest module symbols are loaded
702 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
703 *method is used to avoid symbol missing when the first addr is
704 *in module instead of in guest kernel.
706 err
= perf_event__synthesize_modules(tool
, process_synthesized_event
,
709 pr_err("Couldn't record guest kernel [%d]'s reference"
710 " relocation symbol.\n", machine
->pid
);
713 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
714 * have no _text sometimes.
716 err
= perf_event__synthesize_kernel_mmap(tool
, process_synthesized_event
,
719 pr_err("Couldn't record guest kernel [%d]'s reference"
720 " relocation symbol.\n", machine
->pid
);
723 static struct perf_event_header finished_round_event
= {
724 .size
= sizeof(struct perf_event_header
),
725 .type
= PERF_RECORD_FINISHED_ROUND
,
728 static void record__adjust_affinity(struct record
*rec
, struct perf_mmap
*map
)
730 if (rec
->opts
.affinity
!= PERF_AFFINITY_SYS
&&
731 !CPU_EQUAL(&rec
->affinity_mask
, &map
->affinity_mask
)) {
732 CPU_ZERO(&rec
->affinity_mask
);
733 CPU_OR(&rec
->affinity_mask
, &rec
->affinity_mask
, &map
->affinity_mask
);
734 sched_setaffinity(0, sizeof(rec
->affinity_mask
), &rec
->affinity_mask
);
738 static int record__mmap_read_evlist(struct record
*rec
, struct perf_evlist
*evlist
,
741 u64 bytes_written
= rec
->bytes_written
;
744 struct perf_mmap
*maps
;
745 int trace_fd
= rec
->data
.file
.fd
;
751 maps
= overwrite
? evlist
->overwrite_mmap
: evlist
->mmap
;
755 if (overwrite
&& evlist
->bkw_mmap_state
!= BKW_MMAP_DATA_PENDING
)
758 if (record__aio_enabled(rec
))
759 off
= record__aio_get_pos(trace_fd
);
761 for (i
= 0; i
< evlist
->nr_mmaps
; i
++) {
762 struct perf_mmap
*map
= &maps
[i
];
765 record__adjust_affinity(rec
, map
);
766 if (!record__aio_enabled(rec
)) {
767 if (perf_mmap__push(map
, rec
, record__pushfn
) != 0) {
774 * Call record__aio_sync() to wait till map->data buffer
775 * becomes available after previous aio write request.
777 idx
= record__aio_sync(map
, false);
778 if (perf_mmap__aio_push(map
, rec
, idx
, record__aio_pushfn
, &off
) != 0) {
779 record__aio_set_pos(trace_fd
, off
);
786 if (map
->auxtrace_mmap
.base
&& !rec
->opts
.auxtrace_snapshot_mode
&&
787 record__auxtrace_mmap_read(rec
, map
) != 0) {
793 if (record__aio_enabled(rec
))
794 record__aio_set_pos(trace_fd
, off
);
797 * Mark the round finished in case we wrote
798 * at least one event.
800 if (bytes_written
!= rec
->bytes_written
)
801 rc
= record__write(rec
, NULL
, &finished_round_event
, sizeof(finished_round_event
));
804 perf_evlist__toggle_bkw_mmap(evlist
, BKW_MMAP_EMPTY
);
809 static int record__mmap_read_all(struct record
*rec
)
813 err
= record__mmap_read_evlist(rec
, rec
->evlist
, false);
817 return record__mmap_read_evlist(rec
, rec
->evlist
, true);
820 static void record__init_features(struct record
*rec
)
822 struct perf_session
*session
= rec
->session
;
825 for (feat
= HEADER_FIRST_FEATURE
; feat
< HEADER_LAST_FEATURE
; feat
++)
826 perf_header__set_feat(&session
->header
, feat
);
829 perf_header__clear_feat(&session
->header
, HEADER_BUILD_ID
);
831 if (!have_tracepoints(&rec
->evlist
->entries
))
832 perf_header__clear_feat(&session
->header
, HEADER_TRACING_DATA
);
834 if (!rec
->opts
.branch_stack
)
835 perf_header__clear_feat(&session
->header
, HEADER_BRANCH_STACK
);
837 if (!rec
->opts
.full_auxtrace
)
838 perf_header__clear_feat(&session
->header
, HEADER_AUXTRACE
);
840 if (!(rec
->opts
.use_clockid
&& rec
->opts
.clockid_res_ns
))
841 perf_header__clear_feat(&session
->header
, HEADER_CLOCKID
);
843 perf_header__clear_feat(&session
->header
, HEADER_DIR_FORMAT
);
845 perf_header__clear_feat(&session
->header
, HEADER_STAT
);
849 record__finish_output(struct record
*rec
)
851 struct perf_data
*data
= &rec
->data
;
852 int fd
= perf_data__fd(data
);
857 rec
->session
->header
.data_size
+= rec
->bytes_written
;
858 data
->file
.size
= lseek(perf_data__fd(data
), 0, SEEK_CUR
);
860 if (!rec
->no_buildid
) {
861 process_buildids(rec
);
863 if (rec
->buildid_all
)
864 dsos__hit_all(rec
->session
);
866 perf_session__write_header(rec
->session
, rec
->evlist
, fd
, true);
871 static int record__synthesize_workload(struct record
*rec
, bool tail
)
874 struct thread_map
*thread_map
;
876 if (rec
->opts
.tail_synthesize
!= tail
)
879 thread_map
= thread_map__new_by_tid(rec
->evlist
->workload
.pid
);
880 if (thread_map
== NULL
)
883 err
= perf_event__synthesize_thread_map(&rec
->tool
, thread_map
,
884 process_synthesized_event
,
885 &rec
->session
->machines
.host
,
886 rec
->opts
.sample_address
);
887 thread_map__put(thread_map
);
891 static int record__synthesize(struct record
*rec
, bool tail
);
894 record__switch_output(struct record
*rec
, bool at_exit
)
896 struct perf_data
*data
= &rec
->data
;
900 /* Same Size: "2015122520103046"*/
901 char timestamp
[] = "InvalidTimestamp";
903 record__aio_mmap_read_sync(rec
);
905 record__synthesize(rec
, true);
906 if (target__none(&rec
->opts
.target
))
907 record__synthesize_workload(rec
, true);
910 record__finish_output(rec
);
911 err
= fetch_current_timestamp(timestamp
, sizeof(timestamp
));
913 pr_err("Failed to get current timestamp\n");
917 fd
= perf_data__switch(data
, timestamp
,
918 rec
->session
->header
.data_offset
,
919 at_exit
, &new_filename
);
920 if (fd
>= 0 && !at_exit
) {
921 rec
->bytes_written
= 0;
922 rec
->session
->header
.data_size
= 0;
926 fprintf(stderr
, "[ perf record: Dump %s.%s ]\n",
927 data
->path
, timestamp
);
929 if (rec
->switch_output
.num_files
) {
930 int n
= rec
->switch_output
.cur_file
+ 1;
932 if (n
>= rec
->switch_output
.num_files
)
934 rec
->switch_output
.cur_file
= n
;
935 if (rec
->switch_output
.filenames
[n
]) {
936 remove(rec
->switch_output
.filenames
[n
]);
937 free(rec
->switch_output
.filenames
[n
]);
939 rec
->switch_output
.filenames
[n
] = new_filename
;
944 /* Output tracking events */
946 record__synthesize(rec
, false);
949 * In 'perf record --switch-output' without -a,
950 * record__synthesize() in record__switch_output() won't
951 * generate tracking events because there's no thread_map
952 * in evlist. Which causes newly created perf.data doesn't
953 * contain map and comm information.
954 * Create a fake thread_map and directly call
955 * perf_event__synthesize_thread_map() for those events.
957 if (target__none(&rec
->opts
.target
))
958 record__synthesize_workload(rec
, false);
963 static volatile int workload_exec_errno
;
966 * perf_evlist__prepare_workload will send a SIGUSR1
967 * if the fork fails, since we asked by setting its
968 * want_signal to true.
970 static void workload_exec_failed_signal(int signo __maybe_unused
,
972 void *ucontext __maybe_unused
)
974 workload_exec_errno
= info
->si_value
.sival_int
;
979 static void snapshot_sig_handler(int sig
);
980 static void alarm_sig_handler(int sig
);
983 perf_event__synth_time_conv(const struct perf_event_mmap_page
*pc __maybe_unused
,
984 struct perf_tool
*tool __maybe_unused
,
985 perf_event__handler_t process __maybe_unused
,
986 struct machine
*machine __maybe_unused
)
991 static const struct perf_event_mmap_page
*
992 perf_evlist__pick_pc(struct perf_evlist
*evlist
)
995 if (evlist
->mmap
&& evlist
->mmap
[0].base
)
996 return evlist
->mmap
[0].base
;
997 if (evlist
->overwrite_mmap
&& evlist
->overwrite_mmap
[0].base
)
998 return evlist
->overwrite_mmap
[0].base
;
1003 static const struct perf_event_mmap_page
*record__pick_pc(struct record
*rec
)
1005 const struct perf_event_mmap_page
*pc
;
1007 pc
= perf_evlist__pick_pc(rec
->evlist
);
1013 static int record__synthesize(struct record
*rec
, bool tail
)
1015 struct perf_session
*session
= rec
->session
;
1016 struct machine
*machine
= &session
->machines
.host
;
1017 struct perf_data
*data
= &rec
->data
;
1018 struct record_opts
*opts
= &rec
->opts
;
1019 struct perf_tool
*tool
= &rec
->tool
;
1020 int fd
= perf_data__fd(data
);
1023 if (rec
->opts
.tail_synthesize
!= tail
)
1026 if (data
->is_pipe
) {
1028 * We need to synthesize events first, because some
1029 * features works on top of them (on report side).
1031 err
= perf_event__synthesize_attrs(tool
, rec
->evlist
,
1032 process_synthesized_event
);
1034 pr_err("Couldn't synthesize attrs.\n");
1038 err
= perf_event__synthesize_features(tool
, session
, rec
->evlist
,
1039 process_synthesized_event
);
1041 pr_err("Couldn't synthesize features.\n");
1045 if (have_tracepoints(&rec
->evlist
->entries
)) {
1047 * FIXME err <= 0 here actually means that
1048 * there were no tracepoints so its not really
1049 * an error, just that we don't need to
1050 * synthesize anything. We really have to
1051 * return this more properly and also
1052 * propagate errors that now are calling die()
1054 err
= perf_event__synthesize_tracing_data(tool
, fd
, rec
->evlist
,
1055 process_synthesized_event
);
1057 pr_err("Couldn't record tracing data.\n");
1060 rec
->bytes_written
+= err
;
1064 err
= perf_event__synth_time_conv(record__pick_pc(rec
), tool
,
1065 process_synthesized_event
, machine
);
1069 if (rec
->opts
.full_auxtrace
) {
1070 err
= perf_event__synthesize_auxtrace_info(rec
->itr
, tool
,
1071 session
, process_synthesized_event
);
1076 if (!perf_evlist__exclude_kernel(rec
->evlist
)) {
1077 err
= perf_event__synthesize_kernel_mmap(tool
, process_synthesized_event
,
1079 WARN_ONCE(err
< 0, "Couldn't record kernel reference relocation symbol\n"
1080 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
1081 "Check /proc/kallsyms permission or run as root.\n");
1083 err
= perf_event__synthesize_modules(tool
, process_synthesized_event
,
1085 WARN_ONCE(err
< 0, "Couldn't record kernel module information.\n"
1086 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
1087 "Check /proc/modules permission or run as root.\n");
1091 machines__process_guests(&session
->machines
,
1092 perf_event__synthesize_guest_os
, tool
);
1095 err
= perf_event__synthesize_extra_attr(&rec
->tool
,
1097 process_synthesized_event
,
1102 err
= perf_event__synthesize_thread_map2(&rec
->tool
, rec
->evlist
->threads
,
1103 process_synthesized_event
,
1106 pr_err("Couldn't synthesize thread map.\n");
1110 err
= perf_event__synthesize_cpu_map(&rec
->tool
, rec
->evlist
->cpus
,
1111 process_synthesized_event
, NULL
);
1113 pr_err("Couldn't synthesize cpu map.\n");
1117 err
= perf_event__synthesize_bpf_events(session
, process_synthesized_event
,
1120 pr_warning("Couldn't synthesize bpf events.\n");
1122 err
= __machine__synthesize_threads(machine
, tool
, &opts
->target
, rec
->evlist
->threads
,
1123 process_synthesized_event
, opts
->sample_address
,
1129 static int __cmd_record(struct record
*rec
, int argc
, const char **argv
)
1133 unsigned long waking
= 0;
1134 const bool forks
= argc
> 0;
1135 struct perf_tool
*tool
= &rec
->tool
;
1136 struct record_opts
*opts
= &rec
->opts
;
1137 struct perf_data
*data
= &rec
->data
;
1138 struct perf_session
*session
;
1139 bool disabled
= false, draining
= false;
1142 atexit(record__sig_exit
);
1143 signal(SIGCHLD
, sig_handler
);
1144 signal(SIGINT
, sig_handler
);
1145 signal(SIGTERM
, sig_handler
);
1146 signal(SIGSEGV
, sigsegv_handler
);
1148 if (rec
->opts
.record_namespaces
)
1149 tool
->namespace_events
= true;
1151 if (rec
->opts
.auxtrace_snapshot_mode
|| rec
->switch_output
.enabled
) {
1152 signal(SIGUSR2
, snapshot_sig_handler
);
1153 if (rec
->opts
.auxtrace_snapshot_mode
)
1154 trigger_on(&auxtrace_snapshot_trigger
);
1155 if (rec
->switch_output
.enabled
)
1156 trigger_on(&switch_output_trigger
);
1158 signal(SIGUSR2
, SIG_IGN
);
1161 session
= perf_session__new(data
, false, tool
);
1162 if (session
== NULL
) {
1163 pr_err("Perf session creation failed.\n");
1167 fd
= perf_data__fd(data
);
1168 rec
->session
= session
;
1170 record__init_features(rec
);
1172 if (rec
->opts
.use_clockid
&& rec
->opts
.clockid_res_ns
)
1173 session
->header
.env
.clockid_res_ns
= rec
->opts
.clockid_res_ns
;
1176 err
= perf_evlist__prepare_workload(rec
->evlist
, &opts
->target
,
1177 argv
, data
->is_pipe
,
1178 workload_exec_failed_signal
);
1180 pr_err("Couldn't run the workload!\n");
1182 goto out_delete_session
;
1187 * If we have just single event and are sending data
1188 * through pipe, we need to force the ids allocation,
1189 * because we synthesize event name through the pipe
1190 * and need the id for that.
1192 if (data
->is_pipe
&& rec
->evlist
->nr_entries
== 1)
1193 rec
->opts
.sample_id
= true;
1195 if (record__open(rec
) != 0) {
1200 err
= bpf__apply_obj_config();
1202 char errbuf
[BUFSIZ
];
1204 bpf__strerror_apply_obj_config(err
, errbuf
, sizeof(errbuf
));
1205 pr_err("ERROR: Apply config to BPF failed: %s\n",
1211 * Normally perf_session__new would do this, but it doesn't have the
1214 if (rec
->tool
.ordered_events
&& !perf_evlist__sample_id_all(rec
->evlist
)) {
1215 pr_warning("WARNING: No sample_id_all support, falling back to unordered processing\n");
1216 rec
->tool
.ordered_events
= false;
1219 if (!rec
->evlist
->nr_groups
)
1220 perf_header__clear_feat(&session
->header
, HEADER_GROUP_DESC
);
1222 if (data
->is_pipe
) {
1223 err
= perf_header__write_pipe(fd
);
1227 err
= perf_session__write_header(session
, rec
->evlist
, fd
, false);
1232 if (!rec
->no_buildid
1233 && !perf_header__has_feat(&session
->header
, HEADER_BUILD_ID
)) {
1234 pr_err("Couldn't generate buildids. "
1235 "Use --no-buildid to profile anyway.\n");
1240 err
= record__synthesize(rec
, false);
1244 if (rec
->realtime_prio
) {
1245 struct sched_param param
;
1247 param
.sched_priority
= rec
->realtime_prio
;
1248 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
1249 pr_err("Could not set realtime priority.\n");
1256 * When perf is starting the traced process, all the events
1257 * (apart from group members) have enable_on_exec=1 set,
1258 * so don't spoil it by prematurely enabling them.
1260 if (!target__none(&opts
->target
) && !opts
->initial_delay
)
1261 perf_evlist__enable(rec
->evlist
);
1267 struct machine
*machine
= &session
->machines
.host
;
1268 union perf_event
*event
;
1271 event
= malloc(sizeof(event
->comm
) + machine
->id_hdr_size
);
1272 if (event
== NULL
) {
1278 * Some H/W events are generated before COMM event
1279 * which is emitted during exec(), so perf script
1280 * cannot see a correct process name for those events.
1281 * Synthesize COMM event to prevent it.
1283 tgid
= perf_event__synthesize_comm(tool
, event
,
1284 rec
->evlist
->workload
.pid
,
1285 process_synthesized_event
,
1292 event
= malloc(sizeof(event
->namespaces
) +
1293 (NR_NAMESPACES
* sizeof(struct perf_ns_link_info
)) +
1294 machine
->id_hdr_size
);
1295 if (event
== NULL
) {
1301 * Synthesize NAMESPACES event for the command specified.
1303 perf_event__synthesize_namespaces(tool
, event
,
1304 rec
->evlist
->workload
.pid
,
1305 tgid
, process_synthesized_event
,
1309 perf_evlist__start_workload(rec
->evlist
);
1312 if (opts
->initial_delay
) {
1313 usleep(opts
->initial_delay
* USEC_PER_MSEC
);
1314 perf_evlist__enable(rec
->evlist
);
1317 trigger_ready(&auxtrace_snapshot_trigger
);
1318 trigger_ready(&switch_output_trigger
);
1319 perf_hooks__invoke_record_start();
1321 unsigned long long hits
= rec
->samples
;
1324 * rec->evlist->bkw_mmap_state is possible to be
1325 * BKW_MMAP_EMPTY here: when done == true and
1326 * hits != rec->samples in previous round.
1328 * perf_evlist__toggle_bkw_mmap ensure we never
1329 * convert BKW_MMAP_EMPTY to BKW_MMAP_DATA_PENDING.
1331 if (trigger_is_hit(&switch_output_trigger
) || done
|| draining
)
1332 perf_evlist__toggle_bkw_mmap(rec
->evlist
, BKW_MMAP_DATA_PENDING
);
1334 if (record__mmap_read_all(rec
) < 0) {
1335 trigger_error(&auxtrace_snapshot_trigger
);
1336 trigger_error(&switch_output_trigger
);
1341 if (auxtrace_record__snapshot_started
) {
1342 auxtrace_record__snapshot_started
= 0;
1343 if (!trigger_is_error(&auxtrace_snapshot_trigger
))
1344 record__read_auxtrace_snapshot(rec
);
1345 if (trigger_is_error(&auxtrace_snapshot_trigger
)) {
1346 pr_err("AUX area tracing snapshot failed\n");
1352 if (trigger_is_hit(&switch_output_trigger
)) {
1354 * If switch_output_trigger is hit, the data in
1355 * overwritable ring buffer should have been collected,
1356 * so bkw_mmap_state should be set to BKW_MMAP_EMPTY.
1358 * If SIGUSR2 raise after or during record__mmap_read_all(),
1359 * record__mmap_read_all() didn't collect data from
1360 * overwritable ring buffer. Read again.
1362 if (rec
->evlist
->bkw_mmap_state
== BKW_MMAP_RUNNING
)
1364 trigger_ready(&switch_output_trigger
);
1367 * Reenable events in overwrite ring buffer after
1368 * record__mmap_read_all(): we should have collected
1371 perf_evlist__toggle_bkw_mmap(rec
->evlist
, BKW_MMAP_RUNNING
);
1374 fprintf(stderr
, "[ perf record: dump data: Woken up %ld times ]\n",
1377 fd
= record__switch_output(rec
, false);
1379 pr_err("Failed to switch to new file\n");
1380 trigger_error(&switch_output_trigger
);
1385 /* re-arm the alarm */
1386 if (rec
->switch_output
.time
)
1387 alarm(rec
->switch_output
.time
);
1390 if (hits
== rec
->samples
) {
1391 if (done
|| draining
)
1393 err
= perf_evlist__poll(rec
->evlist
, -1);
1395 * Propagate error, only if there's any. Ignore positive
1396 * number of returned events and interrupt error.
1398 if (err
> 0 || (err
< 0 && errno
== EINTR
))
1402 if (perf_evlist__filter_pollfd(rec
->evlist
, POLLERR
| POLLHUP
) == 0)
1407 * When perf is starting the traced process, at the end events
1408 * die with the process and we wait for that. Thus no need to
1409 * disable events in this case.
1411 if (done
&& !disabled
&& !target__none(&opts
->target
)) {
1412 trigger_off(&auxtrace_snapshot_trigger
);
1413 perf_evlist__disable(rec
->evlist
);
1417 trigger_off(&auxtrace_snapshot_trigger
);
1418 trigger_off(&switch_output_trigger
);
1420 if (forks
&& workload_exec_errno
) {
1421 char msg
[STRERR_BUFSIZE
];
1422 const char *emsg
= str_error_r(workload_exec_errno
, msg
, sizeof(msg
));
1423 pr_err("Workload failed: %s\n", emsg
);
1429 fprintf(stderr
, "[ perf record: Woken up %ld times to write data ]\n", waking
);
1431 if (target__none(&rec
->opts
.target
))
1432 record__synthesize_workload(rec
, true);
1435 record__aio_mmap_read_sync(rec
);
1440 if (!child_finished
)
1441 kill(rec
->evlist
->workload
.pid
, SIGTERM
);
1447 else if (WIFEXITED(exit_status
))
1448 status
= WEXITSTATUS(exit_status
);
1449 else if (WIFSIGNALED(exit_status
))
1450 signr
= WTERMSIG(exit_status
);
1454 record__synthesize(rec
, true);
1455 /* this will be recalculated during process_buildids() */
1459 if (!rec
->timestamp_filename
) {
1460 record__finish_output(rec
);
1462 fd
= record__switch_output(rec
, true);
1465 goto out_delete_session
;
1470 perf_hooks__invoke_record_end();
1472 if (!err
&& !quiet
) {
1474 const char *postfix
= rec
->timestamp_filename
?
1475 ".<timestamp>" : "";
1477 if (rec
->samples
&& !rec
->opts
.full_auxtrace
)
1478 scnprintf(samples
, sizeof(samples
),
1479 " (%" PRIu64
" samples)", rec
->samples
);
1483 fprintf(stderr
, "[ perf record: Captured and wrote %.3f MB %s%s%s ]\n",
1484 perf_data__size(data
) / 1024.0 / 1024.0,
1485 data
->path
, postfix
, samples
);
1489 perf_session__delete(session
);
1493 static void callchain_debug(struct callchain_param
*callchain
)
1495 static const char *str
[CALLCHAIN_MAX
] = { "NONE", "FP", "DWARF", "LBR" };
1497 pr_debug("callchain: type %s\n", str
[callchain
->record_mode
]);
1499 if (callchain
->record_mode
== CALLCHAIN_DWARF
)
1500 pr_debug("callchain: stack dump size %d\n",
1501 callchain
->dump_size
);
1504 int record_opts__parse_callchain(struct record_opts
*record
,
1505 struct callchain_param
*callchain
,
1506 const char *arg
, bool unset
)
1509 callchain
->enabled
= !unset
;
1511 /* --no-call-graph */
1513 callchain
->record_mode
= CALLCHAIN_NONE
;
1514 pr_debug("callchain: disabled\n");
1518 ret
= parse_callchain_record_opt(arg
, callchain
);
1520 /* Enable data address sampling for DWARF unwind. */
1521 if (callchain
->record_mode
== CALLCHAIN_DWARF
)
1522 record
->sample_address
= true;
1523 callchain_debug(callchain
);
1529 int record_parse_callchain_opt(const struct option
*opt
,
1533 return record_opts__parse_callchain(opt
->value
, &callchain_param
, arg
, unset
);
1536 int record_callchain_opt(const struct option
*opt
,
1537 const char *arg __maybe_unused
,
1538 int unset __maybe_unused
)
1540 struct callchain_param
*callchain
= opt
->value
;
1542 callchain
->enabled
= true;
1544 if (callchain
->record_mode
== CALLCHAIN_NONE
)
1545 callchain
->record_mode
= CALLCHAIN_FP
;
1547 callchain_debug(callchain
);
1551 static int perf_record_config(const char *var
, const char *value
, void *cb
)
1553 struct record
*rec
= cb
;
1555 if (!strcmp(var
, "record.build-id")) {
1556 if (!strcmp(value
, "cache"))
1557 rec
->no_buildid_cache
= false;
1558 else if (!strcmp(value
, "no-cache"))
1559 rec
->no_buildid_cache
= true;
1560 else if (!strcmp(value
, "skip"))
1561 rec
->no_buildid
= true;
1566 if (!strcmp(var
, "record.call-graph")) {
1567 var
= "call-graph.record-mode";
1568 return perf_default_config(var
, value
, cb
);
1570 #ifdef HAVE_AIO_SUPPORT
1571 if (!strcmp(var
, "record.aio")) {
1572 rec
->opts
.nr_cblocks
= strtol(value
, NULL
, 0);
1573 if (!rec
->opts
.nr_cblocks
)
1574 rec
->opts
.nr_cblocks
= nr_cblocks_default
;
1581 struct clockid_map
{
1586 #define CLOCKID_MAP(n, c) \
1587 { .name = n, .clockid = (c), }
1589 #define CLOCKID_END { .name = NULL, }
1593 * Add the missing ones, we need to build on many distros...
1595 #ifndef CLOCK_MONOTONIC_RAW
1596 #define CLOCK_MONOTONIC_RAW 4
1598 #ifndef CLOCK_BOOTTIME
1599 #define CLOCK_BOOTTIME 7
1602 #define CLOCK_TAI 11
1605 static const struct clockid_map clockids
[] = {
1606 /* available for all events, NMI safe */
1607 CLOCKID_MAP("monotonic", CLOCK_MONOTONIC
),
1608 CLOCKID_MAP("monotonic_raw", CLOCK_MONOTONIC_RAW
),
1610 /* available for some events */
1611 CLOCKID_MAP("realtime", CLOCK_REALTIME
),
1612 CLOCKID_MAP("boottime", CLOCK_BOOTTIME
),
1613 CLOCKID_MAP("tai", CLOCK_TAI
),
1615 /* available for the lazy */
1616 CLOCKID_MAP("mono", CLOCK_MONOTONIC
),
1617 CLOCKID_MAP("raw", CLOCK_MONOTONIC_RAW
),
1618 CLOCKID_MAP("real", CLOCK_REALTIME
),
1619 CLOCKID_MAP("boot", CLOCK_BOOTTIME
),
1624 static int get_clockid_res(clockid_t clk_id
, u64
*res_ns
)
1626 struct timespec res
;
1629 if (!clock_getres(clk_id
, &res
))
1630 *res_ns
= res
.tv_nsec
+ res
.tv_sec
* NSEC_PER_SEC
;
1632 pr_warning("WARNING: Failed to determine specified clock resolution.\n");
1637 static int parse_clockid(const struct option
*opt
, const char *str
, int unset
)
1639 struct record_opts
*opts
= (struct record_opts
*)opt
->value
;
1640 const struct clockid_map
*cm
;
1641 const char *ostr
= str
;
1644 opts
->use_clockid
= 0;
1652 /* no setting it twice */
1653 if (opts
->use_clockid
)
1656 opts
->use_clockid
= true;
1658 /* if its a number, we're done */
1659 if (sscanf(str
, "%d", &opts
->clockid
) == 1)
1660 return get_clockid_res(opts
->clockid
, &opts
->clockid_res_ns
);
1662 /* allow a "CLOCK_" prefix to the name */
1663 if (!strncasecmp(str
, "CLOCK_", 6))
1666 for (cm
= clockids
; cm
->name
; cm
++) {
1667 if (!strcasecmp(str
, cm
->name
)) {
1668 opts
->clockid
= cm
->clockid
;
1669 return get_clockid_res(opts
->clockid
,
1670 &opts
->clockid_res_ns
);
1674 opts
->use_clockid
= false;
1675 ui__warning("unknown clockid %s, check man page\n", ostr
);
1679 static int record__parse_affinity(const struct option
*opt
, const char *str
, int unset
)
1681 struct record_opts
*opts
= (struct record_opts
*)opt
->value
;
1686 if (!strcasecmp(str
, "node"))
1687 opts
->affinity
= PERF_AFFINITY_NODE
;
1688 else if (!strcasecmp(str
, "cpu"))
1689 opts
->affinity
= PERF_AFFINITY_CPU
;
1694 static int record__parse_mmap_pages(const struct option
*opt
,
1696 int unset __maybe_unused
)
1698 struct record_opts
*opts
= opt
->value
;
1700 unsigned int mmap_pages
;
1715 ret
= __perf_evlist__parse_mmap_pages(&mmap_pages
, s
);
1718 opts
->mmap_pages
= mmap_pages
;
1726 ret
= __perf_evlist__parse_mmap_pages(&mmap_pages
, p
+ 1);
1730 opts
->auxtrace_mmap_pages
= mmap_pages
;
1737 static void switch_output_size_warn(struct record
*rec
)
1739 u64 wakeup_size
= perf_evlist__mmap_size(rec
->opts
.mmap_pages
);
1740 struct switch_output
*s
= &rec
->switch_output
;
1744 if (s
->size
< wakeup_size
) {
1747 unit_number__scnprintf(buf
, sizeof(buf
), wakeup_size
);
1748 pr_warning("WARNING: switch-output data size lower than "
1749 "wakeup kernel buffer size (%s) "
1750 "expect bigger perf.data sizes\n", buf
);
1754 static int switch_output_setup(struct record
*rec
)
1756 struct switch_output
*s
= &rec
->switch_output
;
1757 static struct parse_tag tags_size
[] = {
1758 { .tag
= 'B', .mult
= 1 },
1759 { .tag
= 'K', .mult
= 1 << 10 },
1760 { .tag
= 'M', .mult
= 1 << 20 },
1761 { .tag
= 'G', .mult
= 1 << 30 },
1764 static struct parse_tag tags_time
[] = {
1765 { .tag
= 's', .mult
= 1 },
1766 { .tag
= 'm', .mult
= 60 },
1767 { .tag
= 'h', .mult
= 60*60 },
1768 { .tag
= 'd', .mult
= 60*60*24 },
1776 if (!strcmp(s
->str
, "signal")) {
1778 pr_debug("switch-output with SIGUSR2 signal\n");
1782 val
= parse_tag_value(s
->str
, tags_size
);
1783 if (val
!= (unsigned long) -1) {
1785 pr_debug("switch-output with %s size threshold\n", s
->str
);
1789 val
= parse_tag_value(s
->str
, tags_time
);
1790 if (val
!= (unsigned long) -1) {
1792 pr_debug("switch-output with %s time threshold (%lu seconds)\n",
1800 rec
->timestamp_filename
= true;
1803 if (s
->size
&& !rec
->opts
.no_buffering
)
1804 switch_output_size_warn(rec
);
1809 static const char * const __record_usage
[] = {
1810 "perf record [<options>] [<command>]",
1811 "perf record [<options>] -- <command> [<options>]",
1814 const char * const *record_usage
= __record_usage
;
1817 * XXX Ideally would be local to cmd_record() and passed to a record__new
1818 * because we need to have access to it in record__exit, that is called
1819 * after cmd_record() exits, but since record_options need to be accessible to
1820 * builtin-script, leave it here.
1822 * At least we don't ouch it in all the other functions here directly.
1824 * Just say no to tons of global variables, sigh.
1826 static struct record record
= {
1828 .sample_time
= true,
1829 .mmap_pages
= UINT_MAX
,
1830 .user_freq
= UINT_MAX
,
1831 .user_interval
= ULLONG_MAX
,
1835 .default_per_cpu
= true,
1839 .sample
= process_sample_event
,
1840 .fork
= perf_event__process_fork
,
1841 .exit
= perf_event__process_exit
,
1842 .comm
= perf_event__process_comm
,
1843 .namespaces
= perf_event__process_namespaces
,
1844 .mmap
= perf_event__process_mmap
,
1845 .mmap2
= perf_event__process_mmap2
,
1846 .ordered_events
= true,
1850 const char record_callchain_help
[] = CALLCHAIN_RECORD_HELP
1851 "\n\t\t\t\tDefault: fp";
1853 static bool dry_run
;
1856 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
1857 * with it and switch to use the library functions in perf_evlist that came
1858 * from builtin-record.c, i.e. use record_opts,
1859 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
1862 static struct option __record_options
[] = {
1863 OPT_CALLBACK('e', "event", &record
.evlist
, "event",
1864 "event selector. use 'perf list' to list available events",
1865 parse_events_option
),
1866 OPT_CALLBACK(0, "filter", &record
.evlist
, "filter",
1867 "event filter", parse_filter
),
1868 OPT_CALLBACK_NOOPT(0, "exclude-perf", &record
.evlist
,
1869 NULL
, "don't record events from perf itself",
1871 OPT_STRING('p', "pid", &record
.opts
.target
.pid
, "pid",
1872 "record events on existing process id"),
1873 OPT_STRING('t', "tid", &record
.opts
.target
.tid
, "tid",
1874 "record events on existing thread id"),
1875 OPT_INTEGER('r', "realtime", &record
.realtime_prio
,
1876 "collect data with this RT SCHED_FIFO priority"),
1877 OPT_BOOLEAN(0, "no-buffering", &record
.opts
.no_buffering
,
1878 "collect data without buffering"),
1879 OPT_BOOLEAN('R', "raw-samples", &record
.opts
.raw_samples
,
1880 "collect raw sample records from all opened counters"),
1881 OPT_BOOLEAN('a', "all-cpus", &record
.opts
.target
.system_wide
,
1882 "system-wide collection from all CPUs"),
1883 OPT_STRING('C', "cpu", &record
.opts
.target
.cpu_list
, "cpu",
1884 "list of cpus to monitor"),
1885 OPT_U64('c', "count", &record
.opts
.user_interval
, "event period to sample"),
1886 OPT_STRING('o', "output", &record
.data
.path
, "file",
1887 "output file name"),
1888 OPT_BOOLEAN_SET('i', "no-inherit", &record
.opts
.no_inherit
,
1889 &record
.opts
.no_inherit_set
,
1890 "child tasks do not inherit counters"),
1891 OPT_BOOLEAN(0, "tail-synthesize", &record
.opts
.tail_synthesize
,
1892 "synthesize non-sample events at the end of output"),
1893 OPT_BOOLEAN(0, "overwrite", &record
.opts
.overwrite
, "use overwrite mode"),
1894 OPT_BOOLEAN(0, "no-bpf-event", &record
.opts
.no_bpf_event
, "record bpf events"),
1895 OPT_BOOLEAN(0, "strict-freq", &record
.opts
.strict_freq
,
1896 "Fail if the specified frequency can't be used"),
1897 OPT_CALLBACK('F', "freq", &record
.opts
, "freq or 'max'",
1898 "profile at this frequency",
1899 record__parse_freq
),
1900 OPT_CALLBACK('m', "mmap-pages", &record
.opts
, "pages[,pages]",
1901 "number of mmap data pages and AUX area tracing mmap pages",
1902 record__parse_mmap_pages
),
1903 OPT_BOOLEAN(0, "group", &record
.opts
.group
,
1904 "put the counters into a counter group"),
1905 OPT_CALLBACK_NOOPT('g', NULL
, &callchain_param
,
1906 NULL
, "enables call-graph recording" ,
1907 &record_callchain_opt
),
1908 OPT_CALLBACK(0, "call-graph", &record
.opts
,
1909 "record_mode[,record_size]", record_callchain_help
,
1910 &record_parse_callchain_opt
),
1911 OPT_INCR('v', "verbose", &verbose
,
1912 "be more verbose (show counter open errors, etc)"),
1913 OPT_BOOLEAN('q', "quiet", &quiet
, "don't print any message"),
1914 OPT_BOOLEAN('s', "stat", &record
.opts
.inherit_stat
,
1915 "per thread counts"),
1916 OPT_BOOLEAN('d', "data", &record
.opts
.sample_address
, "Record the sample addresses"),
1917 OPT_BOOLEAN(0, "phys-data", &record
.opts
.sample_phys_addr
,
1918 "Record the sample physical addresses"),
1919 OPT_BOOLEAN(0, "sample-cpu", &record
.opts
.sample_cpu
, "Record the sample cpu"),
1920 OPT_BOOLEAN_SET('T', "timestamp", &record
.opts
.sample_time
,
1921 &record
.opts
.sample_time_set
,
1922 "Record the sample timestamps"),
1923 OPT_BOOLEAN_SET('P', "period", &record
.opts
.period
, &record
.opts
.period_set
,
1924 "Record the sample period"),
1925 OPT_BOOLEAN('n', "no-samples", &record
.opts
.no_samples
,
1927 OPT_BOOLEAN_SET('N', "no-buildid-cache", &record
.no_buildid_cache
,
1928 &record
.no_buildid_cache_set
,
1929 "do not update the buildid cache"),
1930 OPT_BOOLEAN_SET('B', "no-buildid", &record
.no_buildid
,
1931 &record
.no_buildid_set
,
1932 "do not collect buildids in perf.data"),
1933 OPT_CALLBACK('G', "cgroup", &record
.evlist
, "name",
1934 "monitor event in cgroup name only",
1936 OPT_UINTEGER('D', "delay", &record
.opts
.initial_delay
,
1937 "ms to wait before starting measurement after program start"),
1938 OPT_STRING('u', "uid", &record
.opts
.target
.uid_str
, "user",
1941 OPT_CALLBACK_NOOPT('b', "branch-any", &record
.opts
.branch_stack
,
1942 "branch any", "sample any taken branches",
1943 parse_branch_stack
),
1945 OPT_CALLBACK('j', "branch-filter", &record
.opts
.branch_stack
,
1946 "branch filter mask", "branch stack filter modes",
1947 parse_branch_stack
),
1948 OPT_BOOLEAN('W', "weight", &record
.opts
.sample_weight
,
1949 "sample by weight (on special events only)"),
1950 OPT_BOOLEAN(0, "transaction", &record
.opts
.sample_transaction
,
1951 "sample transaction flags (special events only)"),
1952 OPT_BOOLEAN(0, "per-thread", &record
.opts
.target
.per_thread
,
1953 "use per-thread mmaps"),
1954 OPT_CALLBACK_OPTARG('I', "intr-regs", &record
.opts
.sample_intr_regs
, NULL
, "any register",
1955 "sample selected machine registers on interrupt,"
1956 " use -I ? to list register names", parse_regs
),
1957 OPT_CALLBACK_OPTARG(0, "user-regs", &record
.opts
.sample_user_regs
, NULL
, "any register",
1958 "sample selected machine registers on interrupt,"
1959 " use -I ? to list register names", parse_regs
),
1960 OPT_BOOLEAN(0, "running-time", &record
.opts
.running_time
,
1961 "Record running/enabled time of read (:S) events"),
1962 OPT_CALLBACK('k', "clockid", &record
.opts
,
1963 "clockid", "clockid to use for events, see clock_gettime()",
1965 OPT_STRING_OPTARG('S', "snapshot", &record
.opts
.auxtrace_snapshot_opts
,
1966 "opts", "AUX area tracing Snapshot Mode", ""),
1967 OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout
,
1968 "per thread proc mmap processing timeout in ms"),
1969 OPT_BOOLEAN(0, "namespaces", &record
.opts
.record_namespaces
,
1970 "Record namespaces events"),
1971 OPT_BOOLEAN(0, "switch-events", &record
.opts
.record_switch_events
,
1972 "Record context switch events"),
1973 OPT_BOOLEAN_FLAG(0, "all-kernel", &record
.opts
.all_kernel
,
1974 "Configure all used events to run in kernel space.",
1975 PARSE_OPT_EXCLUSIVE
),
1976 OPT_BOOLEAN_FLAG(0, "all-user", &record
.opts
.all_user
,
1977 "Configure all used events to run in user space.",
1978 PARSE_OPT_EXCLUSIVE
),
1979 OPT_STRING(0, "clang-path", &llvm_param
.clang_path
, "clang path",
1980 "clang binary to use for compiling BPF scriptlets"),
1981 OPT_STRING(0, "clang-opt", &llvm_param
.clang_opt
, "clang options",
1982 "options passed to clang when compiling BPF scriptlets"),
1983 OPT_STRING(0, "vmlinux", &symbol_conf
.vmlinux_name
,
1984 "file", "vmlinux pathname"),
1985 OPT_BOOLEAN(0, "buildid-all", &record
.buildid_all
,
1986 "Record build-id of all DSOs regardless of hits"),
1987 OPT_BOOLEAN(0, "timestamp-filename", &record
.timestamp_filename
,
1988 "append timestamp to output filename"),
1989 OPT_BOOLEAN(0, "timestamp-boundary", &record
.timestamp_boundary
,
1990 "Record timestamp boundary (time of first/last samples)"),
1991 OPT_STRING_OPTARG_SET(0, "switch-output", &record
.switch_output
.str
,
1992 &record
.switch_output
.set
, "signal or size[BKMG] or time[smhd]",
1993 "Switch output when receiving SIGUSR2 (signal) or cross a size or time threshold",
1995 OPT_INTEGER(0, "switch-max-files", &record
.switch_output
.num_files
,
1996 "Limit number of switch output generated files"),
1997 OPT_BOOLEAN(0, "dry-run", &dry_run
,
1998 "Parse options then exit"),
1999 #ifdef HAVE_AIO_SUPPORT
2000 OPT_CALLBACK_OPTARG(0, "aio", &record
.opts
,
2001 &nr_cblocks_default
, "n", "Use <n> control blocks in asynchronous trace writing mode (default: 1, max: 4)",
2004 OPT_CALLBACK(0, "affinity", &record
.opts
, "node|cpu",
2005 "Set affinity mask of trace reading thread to NUMA node cpu mask or cpu of processed mmap buffer",
2006 record__parse_affinity
),
2010 struct option
*record_options
= __record_options
;
2012 int cmd_record(int argc
, const char **argv
)
2015 struct record
*rec
= &record
;
2016 char errbuf
[BUFSIZ
];
2018 setlocale(LC_ALL
, "");
2020 #ifndef HAVE_LIBBPF_SUPPORT
2021 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, "NO_LIBBPF=1", c)
2022 set_nobuild('\0', "clang-path", true);
2023 set_nobuild('\0', "clang-opt", true);
2027 #ifndef HAVE_BPF_PROLOGUE
2028 # if !defined (HAVE_DWARF_SUPPORT)
2029 # define REASON "NO_DWARF=1"
2030 # elif !defined (HAVE_LIBBPF_SUPPORT)
2031 # define REASON "NO_LIBBPF=1"
2033 # define REASON "this architecture doesn't support BPF prologue"
2035 # define set_nobuild(s, l, c) set_option_nobuild(record_options, s, l, REASON, c)
2036 set_nobuild('\0', "vmlinux", true);
2041 CPU_ZERO(&rec
->affinity_mask
);
2042 rec
->opts
.affinity
= PERF_AFFINITY_SYS
;
2044 rec
->evlist
= perf_evlist__new();
2045 if (rec
->evlist
== NULL
)
2048 err
= perf_config(perf_record_config
, rec
);
2052 argc
= parse_options(argc
, argv
, record_options
, record_usage
,
2053 PARSE_OPT_STOP_AT_NON_OPTION
);
2055 perf_quiet_option();
2057 /* Make system wide (-a) the default target. */
2058 if (!argc
&& target__none(&rec
->opts
.target
))
2059 rec
->opts
.target
.system_wide
= true;
2061 if (nr_cgroups
&& !rec
->opts
.target
.system_wide
) {
2062 usage_with_options_msg(record_usage
, record_options
,
2063 "cgroup monitoring only available in system-wide mode");
2066 if (rec
->opts
.record_switch_events
&&
2067 !perf_can_record_switch_events()) {
2068 ui__error("kernel does not support recording context switch events\n");
2069 parse_options_usage(record_usage
, record_options
, "switch-events", 0);
2073 if (switch_output_setup(rec
)) {
2074 parse_options_usage(record_usage
, record_options
, "switch-output", 0);
2078 if (rec
->switch_output
.time
) {
2079 signal(SIGALRM
, alarm_sig_handler
);
2080 alarm(rec
->switch_output
.time
);
2083 if (rec
->switch_output
.num_files
) {
2084 rec
->switch_output
.filenames
= calloc(sizeof(char *),
2085 rec
->switch_output
.num_files
);
2086 if (!rec
->switch_output
.filenames
)
2091 * Allow aliases to facilitate the lookup of symbols for address
2092 * filters. Refer to auxtrace_parse_filters().
2094 symbol_conf
.allow_aliases
= true;
2098 err
= record__auxtrace_init(rec
);
2105 err
= bpf__setup_stdout(rec
->evlist
);
2107 bpf__strerror_setup_stdout(rec
->evlist
, err
, errbuf
, sizeof(errbuf
));
2108 pr_err("ERROR: Setup BPF stdout failed: %s\n",
2115 if (symbol_conf
.kptr_restrict
&& !perf_evlist__exclude_kernel(rec
->evlist
))
2117 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
2118 "check /proc/sys/kernel/kptr_restrict.\n\n"
2119 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
2120 "file is not found in the buildid cache or in the vmlinux path.\n\n"
2121 "Samples in kernel modules won't be resolved at all.\n\n"
2122 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
2123 "even with a suitable vmlinux or kallsyms file.\n\n");
2125 if (rec
->no_buildid_cache
|| rec
->no_buildid
) {
2126 disable_buildid_cache();
2127 } else if (rec
->switch_output
.enabled
) {
2129 * In 'perf record --switch-output', disable buildid
2130 * generation by default to reduce data file switching
2131 * overhead. Still generate buildid if they are required
2134 * perf record --switch-output --no-no-buildid \
2135 * --no-no-buildid-cache
2137 * Following code equals to:
2139 * if ((rec->no_buildid || !rec->no_buildid_set) &&
2140 * (rec->no_buildid_cache || !rec->no_buildid_cache_set))
2141 * disable_buildid_cache();
2143 bool disable
= true;
2145 if (rec
->no_buildid_set
&& !rec
->no_buildid
)
2147 if (rec
->no_buildid_cache_set
&& !rec
->no_buildid_cache
)
2150 rec
->no_buildid
= true;
2151 rec
->no_buildid_cache
= true;
2152 disable_buildid_cache();
2156 if (record
.opts
.overwrite
)
2157 record
.opts
.tail_synthesize
= true;
2159 if (rec
->evlist
->nr_entries
== 0 &&
2160 __perf_evlist__add_default(rec
->evlist
, !record
.opts
.no_samples
) < 0) {
2161 pr_err("Not enough memory for event selector list\n");
2165 if (rec
->opts
.target
.tid
&& !rec
->opts
.no_inherit_set
)
2166 rec
->opts
.no_inherit
= true;
2168 err
= target__validate(&rec
->opts
.target
);
2170 target__strerror(&rec
->opts
.target
, err
, errbuf
, BUFSIZ
);
2171 ui__warning("%s\n", errbuf
);
2174 err
= target__parse_uid(&rec
->opts
.target
);
2176 int saved_errno
= errno
;
2178 target__strerror(&rec
->opts
.target
, err
, errbuf
, BUFSIZ
);
2179 ui__error("%s", errbuf
);
2185 /* Enable ignoring missing threads when -u/-p option is defined. */
2186 rec
->opts
.ignore_missing_thread
= rec
->opts
.target
.uid
!= UINT_MAX
|| rec
->opts
.target
.pid
;
2189 if (perf_evlist__create_maps(rec
->evlist
, &rec
->opts
.target
) < 0)
2190 usage_with_options(record_usage
, record_options
);
2192 err
= auxtrace_record__options(rec
->itr
, rec
->evlist
, &rec
->opts
);
2197 * We take all buildids when the file contains
2198 * AUX area tracing data because we do not decode the
2199 * trace because it would take too long.
2201 if (rec
->opts
.full_auxtrace
)
2202 rec
->buildid_all
= true;
2204 if (record_opts__config(&rec
->opts
)) {
2209 if (rec
->opts
.nr_cblocks
> nr_cblocks_max
)
2210 rec
->opts
.nr_cblocks
= nr_cblocks_max
;
2212 pr_info("nr_cblocks: %d\n", rec
->opts
.nr_cblocks
);
2214 pr_debug("affinity: %s\n", affinity_tags
[rec
->opts
.affinity
]);
2216 err
= __cmd_record(&record
, argc
, argv
);
2218 perf_evlist__delete(rec
->evlist
);
2220 auxtrace_record__free(rec
->itr
);
2224 static void snapshot_sig_handler(int sig __maybe_unused
)
2226 struct record
*rec
= &record
;
2228 if (trigger_is_ready(&auxtrace_snapshot_trigger
)) {
2229 trigger_hit(&auxtrace_snapshot_trigger
);
2230 auxtrace_record__snapshot_started
= 1;
2231 if (auxtrace_record__snapshot_start(record
.itr
))
2232 trigger_error(&auxtrace_snapshot_trigger
);
2235 if (switch_output_signal(rec
))
2236 trigger_hit(&switch_output_trigger
);
2239 static void alarm_sig_handler(int sig __maybe_unused
)
2241 struct record
*rec
= &record
;
2243 if (switch_output_time(rec
))
2244 trigger_hit(&switch_output_trigger
);