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.
8 #define _FILE_OFFSET_BITS 64
14 #include "util/build-id.h"
15 #include "util/util.h"
16 #include "util/parse-options.h"
17 #include "util/parse-events.h"
19 #include "util/header.h"
20 #include "util/event.h"
21 #include "util/evlist.h"
22 #include "util/evsel.h"
23 #include "util/debug.h"
24 #include "util/session.h"
25 #include "util/symbol.h"
26 #include "util/cpumap.h"
27 #include "util/thread_map.h"
33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
34 #define SID(e, x, y) xyarray__entry(e->id, x, y)
41 static u64 user_interval
= ULLONG_MAX
;
42 static u64 default_interval
= 0;
43 static u64 sample_type
;
45 static unsigned int page_size
;
46 static unsigned int mmap_pages
= 128;
47 static unsigned int user_freq
= UINT_MAX
;
48 static int freq
= 1000;
50 static int pipe_output
= 0;
51 static const char *output_name
= NULL
;
53 static int realtime_prio
= 0;
54 static bool nodelay
= false;
55 static bool raw_samples
= false;
56 static bool sample_id_all_avail
= true;
57 static bool system_wide
= false;
58 static pid_t target_pid
= -1;
59 static pid_t target_tid
= -1;
60 static pid_t child_pid
= -1;
61 static bool no_inherit
= false;
62 static enum write_mode_t write_mode
= WRITE_FORCE
;
63 static bool call_graph
= false;
64 static bool inherit_stat
= false;
65 static bool no_samples
= false;
66 static bool sample_address
= false;
67 static bool sample_time
= false;
68 static bool no_buildid
= false;
69 static bool no_buildid_cache
= false;
70 static struct perf_evlist
*evsel_list
;
72 static long samples
= 0;
73 static u64 bytes_written
= 0;
75 static int file_new
= 1;
76 static off_t post_processing_offset
;
78 static struct perf_session
*session
;
79 static const char *cpu_list
;
81 static void advance_output(size_t size
)
83 bytes_written
+= size
;
86 static void write_output(void *buf
, size_t size
)
89 int ret
= write(output
, buf
, size
);
92 die("failed to write");
101 static int process_synthesized_event(union perf_event
*event
,
102 struct perf_sample
*sample __used
,
103 struct perf_session
*self __used
)
105 write_output(event
, event
->header
.size
);
109 static void mmap_read(struct perf_mmap
*md
)
111 unsigned int head
= perf_mmap__read_head(md
);
112 unsigned int old
= md
->prev
;
113 unsigned char *data
= md
->base
+ page_size
;
124 if ((old
& md
->mask
) + size
!= (head
& md
->mask
)) {
125 buf
= &data
[old
& md
->mask
];
126 size
= md
->mask
+ 1 - (old
& md
->mask
);
129 write_output(buf
, size
);
132 buf
= &data
[old
& md
->mask
];
136 write_output(buf
, size
);
139 perf_mmap__write_tail(md
, old
);
142 static volatile int done
= 0;
143 static volatile int signr
= -1;
145 static void sig_handler(int sig
)
151 static void sig_atexit(void)
154 kill(child_pid
, SIGTERM
);
156 if (signr
== -1 || signr
== SIGUSR1
)
159 signal(signr
, SIG_DFL
);
160 kill(getpid(), signr
);
163 static struct perf_header_attr
*get_header_attr(struct perf_event_attr
*a
, int nr
)
165 struct perf_header_attr
*h_attr
;
167 if (nr
< session
->header
.attrs
) {
168 h_attr
= session
->header
.attr
[nr
];
170 h_attr
= perf_header_attr__new(a
);
172 if (perf_header__add_attr(&session
->header
, h_attr
) < 0) {
173 perf_header_attr__delete(h_attr
);
181 static void create_counter(struct perf_evsel
*evsel
, int cpu
)
183 char *filter
= evsel
->filter
;
184 struct perf_event_attr
*attr
= &evsel
->attr
;
185 struct perf_header_attr
*h_attr
;
186 struct perf_sample_id
*sid
;
190 for (thread_index
= 0; thread_index
< evsel_list
->threads
->nr
; thread_index
++) {
191 h_attr
= get_header_attr(attr
, evsel
->idx
);
196 if (memcmp(&h_attr
->attr
, attr
, sizeof(*attr
))) {
197 fprintf(stderr
, "incompatible append\n");
202 sid
= SID(evsel
, cpu
, thread_index
);
203 if (perf_header_attr__add_id(h_attr
, sid
->id
) < 0) {
204 pr_warning("Not enough memory to add id\n");
208 if (filter
!= NULL
) {
209 ret
= ioctl(FD(evsel
, cpu
, thread_index
),
210 PERF_EVENT_IOC_SET_FILTER
, filter
);
212 error("failed to set filter with %d (%s)\n", errno
,
220 sample_type
= attr
->sample_type
;
223 static void config_attr(struct perf_evsel
*evsel
, struct perf_evlist
*evlist
)
225 struct perf_event_attr
*attr
= &evsel
->attr
;
226 int track
= !evsel
->idx
; /* only the first counter needs these */
228 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
229 PERF_FORMAT_TOTAL_TIME_RUNNING
|
232 attr
->sample_type
|= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
234 if (evlist
->nr_entries
> 1)
235 attr
->sample_type
|= PERF_SAMPLE_ID
;
238 * We default some events to a 1 default interval. But keep
239 * it a weak assumption overridable by the user.
241 if (!attr
->sample_period
|| (user_freq
!= UINT_MAX
&&
242 user_interval
!= ULLONG_MAX
)) {
244 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
246 attr
->sample_freq
= freq
;
248 attr
->sample_period
= default_interval
;
253 attr
->sample_freq
= 0;
256 attr
->inherit_stat
= 1;
258 if (sample_address
) {
259 attr
->sample_type
|= PERF_SAMPLE_ADDR
;
260 attr
->mmap_data
= track
;
264 attr
->sample_type
|= PERF_SAMPLE_CALLCHAIN
;
267 attr
->sample_type
|= PERF_SAMPLE_CPU
;
269 if (sample_id_all_avail
&&
270 (sample_time
|| system_wide
|| !no_inherit
|| cpu_list
))
271 attr
->sample_type
|= PERF_SAMPLE_TIME
;
274 attr
->sample_type
|= PERF_SAMPLE_TIME
;
275 attr
->sample_type
|= PERF_SAMPLE_RAW
;
276 attr
->sample_type
|= PERF_SAMPLE_CPU
;
281 attr
->wakeup_events
= 1;
287 if (target_pid
== -1 && target_tid
== -1 && !system_wide
) {
289 attr
->enable_on_exec
= 1;
293 static void open_counters(struct perf_evlist
*evlist
)
295 struct perf_evsel
*pos
;
298 list_for_each_entry(pos
, &evlist
->entries
, node
) {
299 struct perf_event_attr
*attr
= &pos
->attr
;
301 * Check if parse_single_tracepoint_event has already asked for
304 * XXX this is kludgy but short term fix for problems introduced by
305 * eac23d1c that broke 'perf script' by having different sample_types
306 * when using multiple tracepoint events when we use a perf binary
307 * that tries to use sample_id_all on an older kernel.
309 * We need to move counter creation to perf_session, support
310 * different sample_types, etc.
312 bool time_needed
= attr
->sample_type
& PERF_SAMPLE_TIME
;
314 config_attr(pos
, evlist
);
316 attr
->sample_id_all
= sample_id_all_avail
? 1 : 0;
318 if (perf_evsel__open(pos
, evlist
->cpus
, evlist
->threads
, group
,
322 if (err
== EPERM
|| err
== EACCES
)
323 die("Permission error - are you root?\n"
324 "\t Consider tweaking"
325 " /proc/sys/kernel/perf_event_paranoid.\n");
326 else if (err
== ENODEV
&& cpu_list
) {
327 die("No such device - did you specify"
328 " an out-of-range profile CPU?\n");
329 } else if (err
== EINVAL
&& sample_id_all_avail
) {
331 * Old kernel, no attr->sample_id_type_all field
333 sample_id_all_avail
= false;
334 if (!sample_time
&& !raw_samples
&& !time_needed
)
335 attr
->sample_type
&= ~PERF_SAMPLE_TIME
;
337 goto retry_sample_id
;
341 * If it's cycles then fall back to hrtimer
342 * based cpu-clock-tick sw counter, which
343 * is always available even if no PMU support:
345 if (attr
->type
== PERF_TYPE_HARDWARE
346 && attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
349 warning(" ... trying to fall back to cpu-clock-ticks\n");
350 attr
->type
= PERF_TYPE_SOFTWARE
;
351 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
355 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
358 #if defined(__i386__) || defined(__x86_64__)
359 if (attr
->type
== PERF_TYPE_HARDWARE
&& err
== EOPNOTSUPP
)
360 die("No hardware sampling interrupt available."
361 " No APIC? If so then you can boot the kernel"
362 " with the \"lapic\" boot parameter to"
363 " force-enable it.\n");
366 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
370 if (perf_evlist__mmap(evlist
, mmap_pages
, false) < 0)
371 die("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
373 for (cpu
= 0; cpu
< evsel_list
->cpus
->nr
; ++cpu
) {
374 list_for_each_entry(pos
, &evlist
->entries
, node
)
375 create_counter(pos
, cpu
);
379 static int process_buildids(void)
381 u64 size
= lseek(output
, 0, SEEK_CUR
);
386 session
->fd
= output
;
387 return __perf_session__process_events(session
, post_processing_offset
,
388 size
- post_processing_offset
,
389 size
, &build_id__mark_dso_hit_ops
);
392 static void atexit_header(void)
395 session
->header
.data_size
+= bytes_written
;
399 perf_header__write(&session
->header
, evsel_list
, output
, true);
400 perf_session__delete(session
);
401 perf_evlist__delete(evsel_list
);
406 static void perf_event__synthesize_guest_os(struct machine
*machine
, void *data
)
409 struct perf_session
*psession
= data
;
411 if (machine__is_host(machine
))
415 *As for guest kernel when processing subcommand record&report,
416 *we arrange module mmap prior to guest kernel mmap and trigger
417 *a preload dso because default guest module symbols are loaded
418 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
419 *method is used to avoid symbol missing when the first addr is
420 *in module instead of in guest kernel.
422 err
= perf_event__synthesize_modules(process_synthesized_event
,
425 pr_err("Couldn't record guest kernel [%d]'s reference"
426 " relocation symbol.\n", machine
->pid
);
429 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
430 * have no _text sometimes.
432 err
= perf_event__synthesize_kernel_mmap(process_synthesized_event
,
433 psession
, machine
, "_text");
435 err
= perf_event__synthesize_kernel_mmap(process_synthesized_event
,
439 pr_err("Couldn't record guest kernel [%d]'s reference"
440 " relocation symbol.\n", machine
->pid
);
443 static struct perf_event_header finished_round_event
= {
444 .size
= sizeof(struct perf_event_header
),
445 .type
= PERF_RECORD_FINISHED_ROUND
,
448 static void mmap_read_all(void)
452 for (i
= 0; i
< evsel_list
->cpus
->nr
; i
++) {
453 if (evsel_list
->mmap
[i
].base
)
454 mmap_read(&evsel_list
->mmap
[i
]);
457 if (perf_header__has_feat(&session
->header
, HEADER_TRACE_INFO
))
458 write_output(&finished_round_event
, sizeof(finished_round_event
));
461 static int __cmd_record(int argc
, const char **argv
)
467 unsigned long waking
= 0;
468 int child_ready_pipe
[2], go_pipe
[2];
469 const bool forks
= argc
> 0;
471 struct machine
*machine
;
473 page_size
= sysconf(_SC_PAGE_SIZE
);
476 signal(SIGCHLD
, sig_handler
);
477 signal(SIGINT
, sig_handler
);
478 signal(SIGUSR1
, sig_handler
);
480 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
481 perror("failed to create pipes");
486 if (!fstat(STDOUT_FILENO
, &st
) && S_ISFIFO(st
.st_mode
))
489 output_name
= "perf.data";
492 if (!strcmp(output_name
, "-"))
494 else if (!stat(output_name
, &st
) && st
.st_size
) {
495 if (write_mode
== WRITE_FORCE
) {
496 char oldname
[PATH_MAX
];
497 snprintf(oldname
, sizeof(oldname
), "%s.old",
500 rename(output_name
, oldname
);
502 } else if (write_mode
== WRITE_APPEND
) {
503 write_mode
= WRITE_FORCE
;
507 flags
= O_CREAT
|O_RDWR
;
508 if (write_mode
== WRITE_APPEND
)
514 output
= STDOUT_FILENO
;
516 output
= open(output_name
, flags
, S_IRUSR
| S_IWUSR
);
518 perror("failed to create output file");
522 session
= perf_session__new(output_name
, O_WRONLY
,
523 write_mode
== WRITE_FORCE
, false, NULL
);
524 if (session
== NULL
) {
525 pr_err("Not enough memory for reading perf file header\n");
530 perf_header__set_feat(&session
->header
, HEADER_BUILD_ID
);
533 err
= perf_header__read(session
, output
);
535 goto out_delete_session
;
538 if (have_tracepoints(&evsel_list
->entries
))
539 perf_header__set_feat(&session
->header
, HEADER_TRACE_INFO
);
542 * perf_session__delete(session) will be called at atexit_header()
544 atexit(atexit_header
);
549 perror("failed to fork");
556 close(child_ready_pipe
[0]);
558 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
561 * Do a dummy execvp to get the PLT entry resolved,
562 * so we avoid the resolver overhead on the real
565 execvp("", (char **)argv
);
568 * Tell the parent we're ready to go
570 close(child_ready_pipe
[1]);
573 * Wait until the parent tells us to go.
575 if (read(go_pipe
[0], &buf
, 1) == -1)
576 perror("unable to read pipe");
578 execvp(argv
[0], (char **)argv
);
581 kill(getppid(), SIGUSR1
);
585 if (!system_wide
&& target_tid
== -1 && target_pid
== -1)
586 evsel_list
->threads
->map
[0] = child_pid
;
588 close(child_ready_pipe
[1]);
591 * wait for child to settle
593 if (read(child_ready_pipe
[0], &buf
, 1) == -1) {
594 perror("unable to read pipe");
597 close(child_ready_pipe
[0]);
600 open_counters(evsel_list
);
602 perf_session__set_sample_type(session
, sample_type
);
605 err
= perf_header__write_pipe(output
);
608 } else if (file_new
) {
609 err
= perf_header__write(&session
->header
, evsel_list
,
615 post_processing_offset
= lseek(output
, 0, SEEK_CUR
);
617 perf_session__set_sample_id_all(session
, sample_id_all_avail
);
620 err
= perf_event__synthesize_attrs(&session
->header
,
621 process_synthesized_event
,
624 pr_err("Couldn't synthesize attrs.\n");
628 err
= perf_event__synthesize_event_types(process_synthesized_event
,
631 pr_err("Couldn't synthesize event_types.\n");
635 if (have_tracepoints(&evsel_list
->entries
)) {
637 * FIXME err <= 0 here actually means that
638 * there were no tracepoints so its not really
639 * an error, just that we don't need to
640 * synthesize anything. We really have to
641 * return this more properly and also
642 * propagate errors that now are calling die()
644 err
= perf_event__synthesize_tracing_data(output
, evsel_list
,
645 process_synthesized_event
,
648 pr_err("Couldn't record tracing data.\n");
655 machine
= perf_session__find_host_machine(session
);
657 pr_err("Couldn't find native kernel information.\n");
661 err
= perf_event__synthesize_kernel_mmap(process_synthesized_event
,
662 session
, machine
, "_text");
664 err
= perf_event__synthesize_kernel_mmap(process_synthesized_event
,
665 session
, machine
, "_stext");
667 pr_err("Couldn't record kernel reference relocation symbol\n"
668 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
669 "Check /proc/kallsyms permission or run as root.\n");
671 err
= perf_event__synthesize_modules(process_synthesized_event
,
674 pr_err("Couldn't record kernel module information.\n"
675 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
676 "Check /proc/modules permission or run as root.\n");
679 perf_session__process_machines(session
,
680 perf_event__synthesize_guest_os
);
683 perf_event__synthesize_thread_map(evsel_list
->threads
,
684 process_synthesized_event
,
687 perf_event__synthesize_threads(process_synthesized_event
,
691 struct sched_param param
;
693 param
.sched_priority
= realtime_prio
;
694 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
695 pr_err("Could not set realtime priority.\n");
712 if (hits
== samples
) {
715 err
= poll(evsel_list
->pollfd
, evsel_list
->nr_fds
, -1);
720 for (i
= 0; i
< evsel_list
->cpus
->nr
; i
++) {
721 struct perf_evsel
*pos
;
723 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
725 thread
< evsel_list
->threads
->nr
;
727 ioctl(FD(pos
, i
, thread
),
728 PERF_EVENT_IOC_DISABLE
);
734 if (quiet
|| signr
== SIGUSR1
)
737 fprintf(stderr
, "[ perf record: Woken up %ld times to write data ]\n", waking
);
740 * Approximate RIP event size: 24 bytes.
743 "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64
" samples) ]\n",
744 (double)bytes_written
/ 1024.0 / 1024.0,
751 perf_session__delete(session
);
755 static const char * const record_usage
[] = {
756 "perf record [<options>] [<command>]",
757 "perf record [<options>] -- <command> [<options>]",
761 static bool force
, append_file
;
763 const struct option record_options
[] = {
764 OPT_CALLBACK('e', "event", &evsel_list
, "event",
765 "event selector. use 'perf list' to list available events",
767 OPT_CALLBACK(0, "filter", &evsel_list
, "filter",
768 "event filter", parse_filter
),
769 OPT_INTEGER('p', "pid", &target_pid
,
770 "record events on existing process id"),
771 OPT_INTEGER('t', "tid", &target_tid
,
772 "record events on existing thread id"),
773 OPT_INTEGER('r', "realtime", &realtime_prio
,
774 "collect data with this RT SCHED_FIFO priority"),
775 OPT_BOOLEAN('D', "no-delay", &nodelay
,
776 "collect data without buffering"),
777 OPT_BOOLEAN('R', "raw-samples", &raw_samples
,
778 "collect raw sample records from all opened counters"),
779 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
780 "system-wide collection from all CPUs"),
781 OPT_BOOLEAN('A', "append", &append_file
,
782 "append to the output file to do incremental profiling"),
783 OPT_STRING('C', "cpu", &cpu_list
, "cpu",
784 "list of cpus to monitor"),
785 OPT_BOOLEAN('f', "force", &force
,
786 "overwrite existing data file (deprecated)"),
787 OPT_U64('c', "count", &user_interval
, "event period to sample"),
788 OPT_STRING('o', "output", &output_name
, "file",
790 OPT_BOOLEAN('i', "no-inherit", &no_inherit
,
791 "child tasks do not inherit counters"),
792 OPT_UINTEGER('F', "freq", &user_freq
, "profile at this frequency"),
793 OPT_UINTEGER('m', "mmap-pages", &mmap_pages
, "number of mmap data pages"),
794 OPT_BOOLEAN('g', "call-graph", &call_graph
,
795 "do call-graph (stack chain/backtrace) recording"),
796 OPT_INCR('v', "verbose", &verbose
,
797 "be more verbose (show counter open errors, etc)"),
798 OPT_BOOLEAN('q', "quiet", &quiet
, "don't print any message"),
799 OPT_BOOLEAN('s', "stat", &inherit_stat
,
800 "per thread counts"),
801 OPT_BOOLEAN('d', "data", &sample_address
,
803 OPT_BOOLEAN('T', "timestamp", &sample_time
, "Sample timestamps"),
804 OPT_BOOLEAN('n', "no-samples", &no_samples
,
806 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache
,
807 "do not update the buildid cache"),
808 OPT_BOOLEAN('B', "no-buildid", &no_buildid
,
809 "do not collect buildids in perf.data"),
810 OPT_CALLBACK('G', "cgroup", &evsel_list
, "name",
811 "monitor event in cgroup name only",
816 int cmd_record(int argc
, const char **argv
, const char *prefix __used
)
819 struct perf_evsel
*pos
;
821 evsel_list
= perf_evlist__new(NULL
, NULL
);
822 if (evsel_list
== NULL
)
825 argc
= parse_options(argc
, argv
, record_options
, record_usage
,
826 PARSE_OPT_STOP_AT_NON_OPTION
);
827 if (!argc
&& target_pid
== -1 && target_tid
== -1 &&
828 !system_wide
&& !cpu_list
)
829 usage_with_options(record_usage
, record_options
);
831 if (force
&& append_file
) {
832 fprintf(stderr
, "Can't overwrite and append at the same time."
833 " You need to choose between -f and -A");
834 usage_with_options(record_usage
, record_options
);
835 } else if (append_file
) {
836 write_mode
= WRITE_APPEND
;
838 write_mode
= WRITE_FORCE
;
841 if (nr_cgroups
&& !system_wide
) {
842 fprintf(stderr
, "cgroup monitoring only available in"
843 " system-wide mode\n");
844 usage_with_options(record_usage
, record_options
);
849 if (no_buildid_cache
|| no_buildid
)
850 disable_buildid_cache();
852 if (evsel_list
->nr_entries
== 0 &&
853 perf_evlist__add_default(evsel_list
) < 0) {
854 pr_err("Not enough memory for event selector list\n");
855 goto out_symbol_exit
;
858 if (target_pid
!= -1)
859 target_tid
= target_pid
;
861 if (perf_evlist__create_maps(evsel_list
, target_pid
,
862 target_tid
, cpu_list
) < 0)
863 usage_with_options(record_usage
, record_options
);
865 list_for_each_entry(pos
, &evsel_list
->entries
, node
) {
866 if (perf_evsel__alloc_fd(pos
, evsel_list
->cpus
->nr
,
867 evsel_list
->threads
->nr
) < 0)
869 if (perf_header__push_event(pos
->attr
.config
, event_name(pos
)))
873 if (perf_evlist__alloc_pollfd(evsel_list
) < 0)
876 if (user_interval
!= ULLONG_MAX
)
877 default_interval
= user_interval
;
878 if (user_freq
!= UINT_MAX
)
882 * User specified count overrides default frequency.
884 if (default_interval
)
887 default_interval
= freq
;
889 fprintf(stderr
, "frequency and count are zero, aborting\n");
894 err
= __cmd_record(argc
, argv
);
896 perf_evlist__delete_maps(evsel_list
);