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/debug.h"
22 #include "util/session.h"
23 #include "util/symbol.h"
24 #include "util/cpumap.h"
34 static int *fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
36 static unsigned int user_interval
= UINT_MAX
;
37 static long default_interval
= 0;
39 static int nr_cpus
= 0;
40 static unsigned int page_size
;
41 static unsigned int mmap_pages
= 128;
42 static unsigned int user_freq
= UINT_MAX
;
43 static int freq
= 1000;
45 static const char *output_name
= "perf.data";
47 static unsigned int realtime_prio
= 0;
48 static bool raw_samples
= false;
49 static bool system_wide
= false;
50 static int profile_cpu
= -1;
51 static pid_t target_pid
= -1;
52 static pid_t target_tid
= -1;
53 static pid_t
*all_tids
= NULL
;
54 static int thread_num
= 0;
55 static pid_t child_pid
= -1;
56 static bool inherit
= true;
57 static enum write_mode_t write_mode
= WRITE_FORCE
;
58 static bool call_graph
= false;
59 static bool inherit_stat
= false;
60 static bool no_samples
= false;
61 static bool sample_address
= false;
62 static bool multiplex
= false;
63 static int multiplex_fd
= -1;
65 static long samples
= 0;
66 static struct timeval last_read
;
67 static struct timeval this_read
;
69 static u64 bytes_written
= 0;
71 static struct pollfd
*event_array
;
73 static int nr_poll
= 0;
74 static int nr_cpu
= 0;
76 static int file_new
= 1;
77 static off_t post_processing_offset
;
79 static struct perf_session
*session
;
88 static struct mmap_data
*mmap_array
[MAX_NR_CPUS
][MAX_COUNTERS
];
90 static unsigned long mmap_read_head(struct mmap_data
*md
)
92 struct perf_event_mmap_page
*pc
= md
->base
;
101 static void mmap_write_tail(struct mmap_data
*md
, unsigned long tail
)
103 struct perf_event_mmap_page
*pc
= md
->base
;
106 * ensure all reads are done before we write the tail out.
109 pc
->data_tail
= tail
;
112 static void write_output(void *buf
, size_t size
)
115 int ret
= write(output
, buf
, size
);
118 die("failed to write");
123 bytes_written
+= ret
;
127 static int process_synthesized_event(event_t
*event
,
128 struct perf_session
*self __used
)
130 write_output(event
, event
->header
.size
);
134 static void mmap_read(struct mmap_data
*md
)
136 unsigned int head
= mmap_read_head(md
);
137 unsigned int old
= md
->prev
;
138 unsigned char *data
= md
->base
+ page_size
;
143 gettimeofday(&this_read
, NULL
);
146 * If we're further behind than half the buffer, there's a chance
147 * the writer will bite our tail and mess up the samples under us.
149 * If we somehow ended up ahead of the head, we got messed up.
151 * In either case, truncate and restart at head.
158 timersub(&this_read
, &last_read
, &iv
);
159 msecs
= iv
.tv_sec
*1000 + iv
.tv_usec
/1000;
161 fprintf(stderr
, "WARNING: failed to keep up with mmap data."
162 " Last read %lu msecs ago.\n", msecs
);
165 * head points to a known good entry, start there.
170 last_read
= this_read
;
177 if ((old
& md
->mask
) + size
!= (head
& md
->mask
)) {
178 buf
= &data
[old
& md
->mask
];
179 size
= md
->mask
+ 1 - (old
& md
->mask
);
182 write_output(buf
, size
);
185 buf
= &data
[old
& md
->mask
];
189 write_output(buf
, size
);
192 mmap_write_tail(md
, old
);
195 static volatile int done
= 0;
196 static volatile int signr
= -1;
198 static void sig_handler(int sig
)
204 static void sig_atexit(void)
207 kill(child_pid
, SIGTERM
);
212 signal(signr
, SIG_DFL
);
213 kill(getpid(), signr
);
218 static struct perf_header_attr
*get_header_attr(struct perf_event_attr
*a
, int nr
)
220 struct perf_header_attr
*h_attr
;
222 if (nr
< session
->header
.attrs
) {
223 h_attr
= session
->header
.attr
[nr
];
225 h_attr
= perf_header_attr__new(a
);
227 if (perf_header__add_attr(&session
->header
, h_attr
) < 0) {
228 perf_header_attr__delete(h_attr
);
236 static void create_counter(int counter
, int cpu
)
238 char *filter
= filters
[counter
];
239 struct perf_event_attr
*attr
= attrs
+ counter
;
240 struct perf_header_attr
*h_attr
;
241 int track
= !counter
; /* only the first counter needs these */
251 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
252 PERF_FORMAT_TOTAL_TIME_RUNNING
|
255 attr
->sample_type
|= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
258 attr
->sample_type
|= PERF_SAMPLE_ID
;
261 * We default some events to a 1 default interval. But keep
262 * it a weak assumption overridable by the user.
264 if (!attr
->sample_period
|| (user_freq
!= UINT_MAX
&&
265 user_interval
!= UINT_MAX
)) {
267 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
269 attr
->sample_freq
= freq
;
271 attr
->sample_period
= default_interval
;
276 attr
->sample_freq
= 0;
279 attr
->inherit_stat
= 1;
282 attr
->sample_type
|= PERF_SAMPLE_ADDR
;
285 attr
->sample_type
|= PERF_SAMPLE_CALLCHAIN
;
288 attr
->sample_type
|= PERF_SAMPLE_TIME
;
289 attr
->sample_type
|= PERF_SAMPLE_RAW
;
290 attr
->sample_type
|= PERF_SAMPLE_CPU
;
295 attr
->inherit
= inherit
;
296 if (target_pid
== -1 && !system_wide
) {
298 attr
->enable_on_exec
= 1;
301 for (thread_index
= 0; thread_index
< thread_num
; thread_index
++) {
303 fd
[nr_cpu
][counter
][thread_index
] = sys_perf_event_open(attr
,
304 all_tids
[thread_index
], cpu
, group_fd
, 0);
306 if (fd
[nr_cpu
][counter
][thread_index
] < 0) {
309 if (err
== EPERM
|| err
== EACCES
)
310 die("Permission error - are you root?\n"
311 "\t Consider tweaking"
312 " /proc/sys/kernel/perf_event_paranoid.\n");
313 else if (err
== ENODEV
&& profile_cpu
!= -1) {
314 die("No such device - did you specify"
315 " an out-of-range profile CPU?\n");
319 * If it's cycles then fall back to hrtimer
320 * based cpu-clock-tick sw counter, which
321 * is always available even if no PMU support:
323 if (attr
->type
== PERF_TYPE_HARDWARE
324 && attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
327 warning(" ... trying to fall back to cpu-clock-ticks\n");
328 attr
->type
= PERF_TYPE_SOFTWARE
;
329 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
333 error("perfcounter syscall returned with %d (%s)\n",
334 fd
[nr_cpu
][counter
][thread_index
], strerror(err
));
336 #if defined(__i386__) || defined(__x86_64__)
337 if (attr
->type
== PERF_TYPE_HARDWARE
&& err
== EOPNOTSUPP
)
338 die("No hardware sampling interrupt available."
339 " No APIC? If so then you can boot the kernel"
340 " with the \"lapic\" boot parameter to"
341 " force-enable it.\n");
344 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
348 h_attr
= get_header_attr(attr
, counter
);
353 if (memcmp(&h_attr
->attr
, attr
, sizeof(*attr
))) {
354 fprintf(stderr
, "incompatible append\n");
359 if (read(fd
[nr_cpu
][counter
][thread_index
], &read_data
, sizeof(read_data
)) == -1) {
360 perror("Unable to read perf file descriptor\n");
364 if (perf_header_attr__add_id(h_attr
, read_data
.id
) < 0) {
365 pr_warning("Not enough memory to add id\n");
369 assert(fd
[nr_cpu
][counter
][thread_index
] >= 0);
370 fcntl(fd
[nr_cpu
][counter
][thread_index
], F_SETFL
, O_NONBLOCK
);
373 * First counter acts as the group leader:
375 if (group
&& group_fd
== -1)
376 group_fd
= fd
[nr_cpu
][counter
][thread_index
];
377 if (multiplex
&& multiplex_fd
== -1)
378 multiplex_fd
= fd
[nr_cpu
][counter
][thread_index
];
380 if (multiplex
&& fd
[nr_cpu
][counter
][thread_index
] != multiplex_fd
) {
382 ret
= ioctl(fd
[nr_cpu
][counter
][thread_index
], PERF_EVENT_IOC_SET_OUTPUT
, multiplex_fd
);
385 event_array
[nr_poll
].fd
= fd
[nr_cpu
][counter
][thread_index
];
386 event_array
[nr_poll
].events
= POLLIN
;
389 mmap_array
[nr_cpu
][counter
][thread_index
].counter
= counter
;
390 mmap_array
[nr_cpu
][counter
][thread_index
].prev
= 0;
391 mmap_array
[nr_cpu
][counter
][thread_index
].mask
= mmap_pages
*page_size
- 1;
392 mmap_array
[nr_cpu
][counter
][thread_index
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
393 PROT_READ
|PROT_WRITE
, MAP_SHARED
, fd
[nr_cpu
][counter
][thread_index
], 0);
394 if (mmap_array
[nr_cpu
][counter
][thread_index
].base
== MAP_FAILED
) {
395 error("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
400 if (filter
!= NULL
) {
401 ret
= ioctl(fd
[nr_cpu
][counter
][thread_index
],
402 PERF_EVENT_IOC_SET_FILTER
, filter
);
404 error("failed to set filter with %d (%s)\n", errno
,
412 static void open_counters(int cpu
)
417 for (counter
= 0; counter
< nr_counters
; counter
++)
418 create_counter(counter
, cpu
);
423 static int process_buildids(void)
425 u64 size
= lseek(output
, 0, SEEK_CUR
);
430 session
->fd
= output
;
431 return __perf_session__process_events(session
, post_processing_offset
,
432 size
- post_processing_offset
,
433 size
, &build_id__mark_dso_hit_ops
);
436 static void atexit_header(void)
438 session
->header
.data_size
+= bytes_written
;
441 perf_header__write(&session
->header
, output
, true);
444 static int __cmd_record(int argc
, const char **argv
)
451 unsigned long waking
= 0;
452 int child_ready_pipe
[2], go_pipe
[2];
453 const bool forks
= argc
> 0;
456 page_size
= sysconf(_SC_PAGE_SIZE
);
459 signal(SIGCHLD
, sig_handler
);
460 signal(SIGINT
, sig_handler
);
462 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
463 perror("failed to create pipes");
467 if (!stat(output_name
, &st
) && st
.st_size
) {
468 if (write_mode
== WRITE_FORCE
) {
469 char oldname
[PATH_MAX
];
470 snprintf(oldname
, sizeof(oldname
), "%s.old",
473 rename(output_name
, oldname
);
475 } else if (write_mode
== WRITE_APPEND
) {
476 write_mode
= WRITE_FORCE
;
479 flags
= O_CREAT
|O_RDWR
;
480 if (write_mode
== WRITE_APPEND
)
485 output
= open(output_name
, flags
, S_IRUSR
|S_IWUSR
);
487 perror("failed to create output file");
491 session
= perf_session__new(output_name
, O_WRONLY
,
492 write_mode
== WRITE_FORCE
);
493 if (session
== NULL
) {
494 pr_err("Not enough memory for reading perf file header\n");
499 err
= perf_header__read(&session
->header
, output
);
505 perf_header__set_feat(&session
->header
, HEADER_TRACE_INFO
);
507 for (i
= 0; i
< nr_counters
; i
++) {
508 if (attrs
[i
].sample_type
& PERF_SAMPLE_RAW
) {
509 perf_header__set_feat(&session
->header
, HEADER_TRACE_INFO
);
515 atexit(atexit_header
);
520 perror("failed to fork");
525 close(child_ready_pipe
[0]);
527 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
530 * Do a dummy execvp to get the PLT entry resolved,
531 * so we avoid the resolver overhead on the real
534 execvp("", (char **)argv
);
537 * Tell the parent we're ready to go
539 close(child_ready_pipe
[1]);
542 * Wait until the parent tells us to go.
544 if (read(go_pipe
[0], &buf
, 1) == -1)
545 perror("unable to read pipe");
547 execvp(argv
[0], (char **)argv
);
553 if (!system_wide
&& target_tid
== -1 && target_pid
== -1)
554 all_tids
[0] = child_pid
;
556 close(child_ready_pipe
[1]);
559 * wait for child to settle
561 if (read(child_ready_pipe
[0], &buf
, 1) == -1) {
562 perror("unable to read pipe");
565 close(child_ready_pipe
[0]);
568 if ((!system_wide
&& !inherit
) || profile_cpu
!= -1) {
569 open_counters(profile_cpu
);
571 nr_cpus
= read_cpu_map();
572 for (i
= 0; i
< nr_cpus
; i
++)
573 open_counters(cpumap
[i
]);
577 err
= perf_header__write(&session
->header
, output
, false);
582 post_processing_offset
= lseek(output
, 0, SEEK_CUR
);
584 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
587 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
590 pr_err("Couldn't record kernel reference relocation symbol.\n");
594 err
= event__synthesize_modules(process_synthesized_event
, session
);
596 pr_err("Couldn't record kernel reference relocation symbol.\n");
600 if (!system_wide
&& profile_cpu
== -1)
601 event__synthesize_thread(target_tid
, process_synthesized_event
,
604 event__synthesize_threads(process_synthesized_event
, session
);
607 struct sched_param param
;
609 param
.sched_priority
= realtime_prio
;
610 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
611 pr_err("Could not set realtime priority.\n");
626 for (i
= 0; i
< nr_cpu
; i
++) {
627 for (counter
= 0; counter
< nr_counters
; counter
++) {
629 thread
< thread_num
; thread
++) {
630 if (mmap_array
[i
][counter
][thread
].base
)
631 mmap_read(&mmap_array
[i
][counter
][thread
]);
637 if (hits
== samples
) {
640 err
= poll(event_array
, nr_poll
, -1);
645 for (i
= 0; i
< nr_cpu
; i
++) {
647 counter
< nr_counters
;
652 ioctl(fd
[i
][counter
][thread
],
653 PERF_EVENT_IOC_DISABLE
);
659 fprintf(stderr
, "[ perf record: Woken up %ld times to write data ]\n", waking
);
662 * Approximate RIP event size: 24 bytes.
665 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
666 (double)bytes_written
/ 1024.0 / 1024.0,
673 static const char * const record_usage
[] = {
674 "perf record [<options>] [<command>]",
675 "perf record [<options>] -- <command> [<options>]",
679 static bool force
, append_file
;
681 static const struct option options
[] = {
682 OPT_CALLBACK('e', "event", NULL
, "event",
683 "event selector. use 'perf list' to list available events",
685 OPT_CALLBACK(0, "filter", NULL
, "filter",
686 "event filter", parse_filter
),
687 OPT_INTEGER('p', "pid", &target_pid
,
688 "record events on existing process id"),
689 OPT_INTEGER('t', "tid", &target_tid
,
690 "record events on existing thread id"),
691 OPT_INTEGER('r', "realtime", &realtime_prio
,
692 "collect data with this RT SCHED_FIFO priority"),
693 OPT_BOOLEAN('R', "raw-samples", &raw_samples
,
694 "collect raw sample records from all opened counters"),
695 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
696 "system-wide collection from all CPUs"),
697 OPT_BOOLEAN('A', "append", &append_file
,
698 "append to the output file to do incremental profiling"),
699 OPT_INTEGER('C', "profile_cpu", &profile_cpu
,
700 "CPU to profile on"),
701 OPT_BOOLEAN('f', "force", &force
,
702 "overwrite existing data file (deprecated)"),
703 OPT_LONG('c', "count", &user_interval
,
704 "event period to sample"),
705 OPT_STRING('o', "output", &output_name
, "file",
707 OPT_BOOLEAN('i', "inherit", &inherit
,
708 "child tasks inherit counters"),
709 OPT_INTEGER('F', "freq", &user_freq
,
710 "profile at this frequency"),
711 OPT_INTEGER('m', "mmap-pages", &mmap_pages
,
712 "number of mmap data pages"),
713 OPT_BOOLEAN('g', "call-graph", &call_graph
,
714 "do call-graph (stack chain/backtrace) recording"),
715 OPT_INCR('v', "verbose", &verbose
,
716 "be more verbose (show counter open errors, etc)"),
717 OPT_BOOLEAN('s', "stat", &inherit_stat
,
718 "per thread counts"),
719 OPT_BOOLEAN('d', "data", &sample_address
,
721 OPT_BOOLEAN('n', "no-samples", &no_samples
,
723 OPT_BOOLEAN('M', "multiplex", &multiplex
,
724 "multiplex counter output in a single channel"),
728 int cmd_record(int argc
, const char **argv
, const char *prefix __used
)
732 argc
= parse_options(argc
, argv
, options
, record_usage
,
733 PARSE_OPT_STOP_AT_NON_OPTION
);
734 if (!argc
&& target_pid
== -1 && target_tid
== -1 &&
735 !system_wide
&& profile_cpu
== -1)
736 usage_with_options(record_usage
, options
);
738 if (force
&& append_file
) {
739 fprintf(stderr
, "Can't overwrite and append at the same time."
740 " You need to choose between -f and -A");
741 usage_with_options(record_usage
, options
);
742 } else if (append_file
) {
743 write_mode
= WRITE_APPEND
;
745 write_mode
= WRITE_FORCE
;
752 attrs
[0].type
= PERF_TYPE_HARDWARE
;
753 attrs
[0].config
= PERF_COUNT_HW_CPU_CYCLES
;
756 if (target_pid
!= -1) {
757 target_tid
= target_pid
;
758 thread_num
= find_all_tid(target_pid
, &all_tids
);
759 if (thread_num
<= 0) {
760 fprintf(stderr
, "Can't find all threads of pid %d\n",
762 usage_with_options(record_usage
, options
);
765 all_tids
=malloc(sizeof(pid_t
));
769 all_tids
[0] = target_tid
;
773 for (i
= 0; i
< MAX_NR_CPUS
; i
++) {
774 for (j
= 0; j
< MAX_COUNTERS
; j
++) {
775 fd
[i
][j
] = malloc(sizeof(int)*thread_num
);
776 mmap_array
[i
][j
] = zalloc(
777 sizeof(struct mmap_data
)*thread_num
);
778 if (!fd
[i
][j
] || !mmap_array
[i
][j
])
782 event_array
= malloc(
783 sizeof(struct pollfd
)*MAX_NR_CPUS
*MAX_COUNTERS
*thread_num
);
787 if (user_interval
!= UINT_MAX
)
788 default_interval
= user_interval
;
789 if (user_freq
!= UINT_MAX
)
793 * User specified count overrides default frequency.
795 if (default_interval
)
798 default_interval
= freq
;
800 fprintf(stderr
, "frequency and count are zero, aborting\n");
804 return __cmd_record(argc
, argv
);