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"
18 #include "util/string.h"
20 #include "util/header.h"
21 #include "util/event.h"
22 #include "util/debug.h"
23 #include "util/session.h"
24 #include "util/symbol.h"
29 static int fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
31 static long default_interval
= 0;
33 static int nr_cpus
= 0;
34 static unsigned int page_size
;
35 static unsigned int mmap_pages
= 128;
36 static int freq
= 1000;
38 static const char *output_name
= "perf.data";
40 static unsigned int realtime_prio
= 0;
41 static int raw_samples
= 0;
42 static int system_wide
= 0;
43 static int profile_cpu
= -1;
44 static pid_t target_pid
= -1;
45 static pid_t child_pid
= -1;
46 static int inherit
= 1;
48 static int append_file
= 0;
49 static int call_graph
= 0;
50 static int inherit_stat
= 0;
51 static int no_samples
= 0;
52 static int sample_address
= 0;
53 static int multiplex
= 0;
54 static int multiplex_fd
= -1;
56 static long samples
= 0;
57 static struct timeval last_read
;
58 static struct timeval this_read
;
60 static u64 bytes_written
= 0;
62 static struct pollfd event_array
[MAX_NR_CPUS
* MAX_COUNTERS
];
64 static int nr_poll
= 0;
65 static int nr_cpu
= 0;
67 static int file_new
= 1;
68 static off_t post_processing_offset
;
70 static struct perf_session
*session
;
79 static struct mmap_data mmap_array
[MAX_NR_CPUS
][MAX_COUNTERS
];
81 static unsigned long mmap_read_head(struct mmap_data
*md
)
83 struct perf_event_mmap_page
*pc
= md
->base
;
92 static void mmap_write_tail(struct mmap_data
*md
, unsigned long tail
)
94 struct perf_event_mmap_page
*pc
= md
->base
;
97 * ensure all reads are done before we write the tail out.
100 pc
->data_tail
= tail
;
103 static void write_output(void *buf
, size_t size
)
106 int ret
= write(output
, buf
, size
);
109 die("failed to write");
114 bytes_written
+= ret
;
118 static int process_synthesized_event(event_t
*event
,
119 struct perf_session
*self __used
)
121 write_output(event
, event
->header
.size
);
125 static void mmap_read(struct mmap_data
*md
)
127 unsigned int head
= mmap_read_head(md
);
128 unsigned int old
= md
->prev
;
129 unsigned char *data
= md
->base
+ page_size
;
134 gettimeofday(&this_read
, NULL
);
137 * If we're further behind than half the buffer, there's a chance
138 * the writer will bite our tail and mess up the samples under us.
140 * If we somehow ended up ahead of the head, we got messed up.
142 * In either case, truncate and restart at head.
149 timersub(&this_read
, &last_read
, &iv
);
150 msecs
= iv
.tv_sec
*1000 + iv
.tv_usec
/1000;
152 fprintf(stderr
, "WARNING: failed to keep up with mmap data."
153 " Last read %lu msecs ago.\n", msecs
);
156 * head points to a known good entry, start there.
161 last_read
= this_read
;
168 if ((old
& md
->mask
) + size
!= (head
& md
->mask
)) {
169 buf
= &data
[old
& md
->mask
];
170 size
= md
->mask
+ 1 - (old
& md
->mask
);
173 write_output(buf
, size
);
176 buf
= &data
[old
& md
->mask
];
180 write_output(buf
, size
);
183 mmap_write_tail(md
, old
);
186 static volatile int done
= 0;
187 static volatile int signr
= -1;
189 static void sig_handler(int sig
)
195 static void sig_atexit(void)
198 kill(child_pid
, SIGTERM
);
203 signal(signr
, SIG_DFL
);
204 kill(getpid(), signr
);
209 static struct perf_header_attr
*get_header_attr(struct perf_event_attr
*a
, int nr
)
211 struct perf_header_attr
*h_attr
;
213 if (nr
< session
->header
.attrs
) {
214 h_attr
= session
->header
.attr
[nr
];
216 h_attr
= perf_header_attr__new(a
);
218 if (perf_header__add_attr(&session
->header
, h_attr
) < 0) {
219 perf_header_attr__delete(h_attr
);
227 static void create_counter(int counter
, int cpu
, pid_t pid
)
229 char *filter
= filters
[counter
];
230 struct perf_event_attr
*attr
= attrs
+ counter
;
231 struct perf_header_attr
*h_attr
;
232 int track
= !counter
; /* only the first counter needs these */
241 attr
->read_format
= PERF_FORMAT_TOTAL_TIME_ENABLED
|
242 PERF_FORMAT_TOTAL_TIME_RUNNING
|
245 attr
->sample_type
|= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
;
248 attr
->sample_type
|= PERF_SAMPLE_PERIOD
;
250 attr
->sample_freq
= freq
;
254 attr
->sample_freq
= 0;
257 attr
->inherit_stat
= 1;
260 attr
->sample_type
|= PERF_SAMPLE_ADDR
;
263 attr
->sample_type
|= PERF_SAMPLE_CALLCHAIN
;
266 attr
->sample_type
|= PERF_SAMPLE_TIME
;
267 attr
->sample_type
|= PERF_SAMPLE_RAW
;
268 attr
->sample_type
|= PERF_SAMPLE_CPU
;
273 attr
->inherit
= inherit
;
277 fd
[nr_cpu
][counter
] = sys_perf_event_open(attr
, pid
, cpu
, group_fd
, 0);
279 if (fd
[nr_cpu
][counter
] < 0) {
282 if (err
== EPERM
|| err
== EACCES
)
283 die("Permission error - are you root?\n");
284 else if (err
== ENODEV
&& profile_cpu
!= -1)
285 die("No such device - did you specify an out-of-range profile CPU?\n");
288 * If it's cycles then fall back to hrtimer
289 * based cpu-clock-tick sw counter, which
290 * is always available even if no PMU support:
292 if (attr
->type
== PERF_TYPE_HARDWARE
293 && attr
->config
== PERF_COUNT_HW_CPU_CYCLES
) {
296 warning(" ... trying to fall back to cpu-clock-ticks\n");
297 attr
->type
= PERF_TYPE_SOFTWARE
;
298 attr
->config
= PERF_COUNT_SW_CPU_CLOCK
;
302 error("perfcounter syscall returned with %d (%s)\n",
303 fd
[nr_cpu
][counter
], strerror(err
));
305 #if defined(__i386__) || defined(__x86_64__)
306 if (attr
->type
== PERF_TYPE_HARDWARE
&& err
== EOPNOTSUPP
)
307 die("No hardware sampling interrupt available. No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.\n");
310 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
314 h_attr
= get_header_attr(attr
, counter
);
319 if (memcmp(&h_attr
->attr
, attr
, sizeof(*attr
))) {
320 fprintf(stderr
, "incompatible append\n");
325 if (read(fd
[nr_cpu
][counter
], &read_data
, sizeof(read_data
)) == -1) {
326 perror("Unable to read perf file descriptor\n");
330 if (perf_header_attr__add_id(h_attr
, read_data
.id
) < 0) {
331 pr_warning("Not enough memory to add id\n");
335 assert(fd
[nr_cpu
][counter
] >= 0);
336 fcntl(fd
[nr_cpu
][counter
], F_SETFL
, O_NONBLOCK
);
339 * First counter acts as the group leader:
341 if (group
&& group_fd
== -1)
342 group_fd
= fd
[nr_cpu
][counter
];
343 if (multiplex
&& multiplex_fd
== -1)
344 multiplex_fd
= fd
[nr_cpu
][counter
];
346 if (multiplex
&& fd
[nr_cpu
][counter
] != multiplex_fd
) {
348 ret
= ioctl(fd
[nr_cpu
][counter
], PERF_EVENT_IOC_SET_OUTPUT
, multiplex_fd
);
351 event_array
[nr_poll
].fd
= fd
[nr_cpu
][counter
];
352 event_array
[nr_poll
].events
= POLLIN
;
355 mmap_array
[nr_cpu
][counter
].counter
= counter
;
356 mmap_array
[nr_cpu
][counter
].prev
= 0;
357 mmap_array
[nr_cpu
][counter
].mask
= mmap_pages
*page_size
- 1;
358 mmap_array
[nr_cpu
][counter
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
359 PROT_READ
|PROT_WRITE
, MAP_SHARED
, fd
[nr_cpu
][counter
], 0);
360 if (mmap_array
[nr_cpu
][counter
].base
== MAP_FAILED
) {
361 error("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
366 if (filter
!= NULL
) {
367 ret
= ioctl(fd
[nr_cpu
][counter
],
368 PERF_EVENT_IOC_SET_FILTER
, filter
);
370 error("failed to set filter with %d (%s)\n", errno
,
376 ioctl(fd
[nr_cpu
][counter
], PERF_EVENT_IOC_ENABLE
);
379 static void open_counters(int cpu
, pid_t pid
)
384 for (counter
= 0; counter
< nr_counters
; counter
++)
385 create_counter(counter
, cpu
, pid
);
390 static int process_buildids(void)
392 u64 size
= lseek(output
, 0, SEEK_CUR
);
394 session
->fd
= output
;
395 return __perf_session__process_events(session
, post_processing_offset
,
396 size
- post_processing_offset
,
397 size
, &build_id__mark_dso_hit_ops
);
400 static void atexit_header(void)
402 session
->header
.data_size
+= bytes_written
;
405 perf_header__write(&session
->header
, output
, true);
408 static int __cmd_record(int argc
, const char **argv
)
415 unsigned long waking
= 0;
416 int child_ready_pipe
[2], go_pipe
[2];
417 const bool forks
= target_pid
== -1 && argc
> 0;
420 page_size
= sysconf(_SC_PAGE_SIZE
);
421 nr_cpus
= sysconf(_SC_NPROCESSORS_ONLN
);
422 assert(nr_cpus
<= MAX_NR_CPUS
);
423 assert(nr_cpus
>= 0);
426 signal(SIGCHLD
, sig_handler
);
427 signal(SIGINT
, sig_handler
);
429 if (forks
&& (pipe(child_ready_pipe
) < 0 || pipe(go_pipe
) < 0)) {
430 perror("failed to create pipes");
434 if (!stat(output_name
, &st
) && st
.st_size
) {
437 pr_err("Error, output file %s exists, use -A "
438 "to append or -f to overwrite.\n",
443 char oldname
[PATH_MAX
];
444 snprintf(oldname
, sizeof(oldname
), "%s.old",
447 rename(output_name
, oldname
);
453 flags
= O_CREAT
|O_RDWR
;
459 output
= open(output_name
, flags
, S_IRUSR
|S_IWUSR
);
461 perror("failed to create output file");
465 session
= perf_session__new(output_name
, O_WRONLY
, force
);
466 if (session
== NULL
) {
467 pr_err("Not enough memory for reading perf file header\n");
472 err
= perf_header__read(&session
->header
, output
);
478 perf_header__set_feat(&session
->header
, HEADER_TRACE_INFO
);
480 for (i
= 0; i
< nr_counters
; i
++) {
481 if (attrs
[i
].sample_type
& PERF_SAMPLE_RAW
) {
482 perf_header__set_feat(&session
->header
, HEADER_TRACE_INFO
);
488 atexit(atexit_header
);
493 perror("failed to fork");
498 close(child_ready_pipe
[0]);
500 fcntl(go_pipe
[0], F_SETFD
, FD_CLOEXEC
);
503 * Do a dummy execvp to get the PLT entry resolved,
504 * so we avoid the resolver overhead on the real
507 execvp("", (char **)argv
);
510 * Tell the parent we're ready to go
512 close(child_ready_pipe
[1]);
515 * Wait until the parent tells us to go.
517 if (read(go_pipe
[0], &buf
, 1) == -1)
518 perror("unable to read pipe");
520 execvp(argv
[0], (char **)argv
);
531 close(child_ready_pipe
[1]);
534 * wait for child to settle
536 if (read(child_ready_pipe
[0], &buf
, 1) == -1) {
537 perror("unable to read pipe");
540 close(child_ready_pipe
[0]);
544 if ((!system_wide
&& !inherit
) || profile_cpu
!= -1) {
545 open_counters(profile_cpu
, target_pid
);
547 for (i
= 0; i
< nr_cpus
; i
++)
548 open_counters(i
, target_pid
);
552 err
= perf_header__write(&session
->header
, output
, false);
557 post_processing_offset
= lseek(output
, 0, SEEK_CUR
);
559 err
= event__synthesize_kernel_mmap(process_synthesized_event
,
562 pr_err("Couldn't record kernel reference relocation symbol.\n");
566 err
= event__synthesize_modules(process_synthesized_event
, session
);
568 pr_err("Couldn't record kernel reference relocation symbol.\n");
572 if (!system_wide
&& profile_cpu
== -1)
573 event__synthesize_thread(target_pid
, process_synthesized_event
,
576 event__synthesize_threads(process_synthesized_event
, session
);
579 struct sched_param param
;
581 param
.sched_priority
= realtime_prio
;
582 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
583 pr_err("Could not set realtime priority.\n");
597 for (i
= 0; i
< nr_cpu
; i
++) {
598 for (counter
= 0; counter
< nr_counters
; counter
++) {
599 if (mmap_array
[i
][counter
].base
)
600 mmap_read(&mmap_array
[i
][counter
]);
604 if (hits
== samples
) {
607 err
= poll(event_array
, nr_poll
, -1);
612 for (i
= 0; i
< nr_cpu
; i
++) {
613 for (counter
= 0; counter
< nr_counters
; counter
++)
614 ioctl(fd
[i
][counter
], PERF_EVENT_IOC_DISABLE
);
619 fprintf(stderr
, "[ perf record: Woken up %ld times to write data ]\n", waking
);
622 * Approximate RIP event size: 24 bytes.
625 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
626 (double)bytes_written
/ 1024.0 / 1024.0,
633 static const char * const record_usage
[] = {
634 "perf record [<options>] [<command>]",
635 "perf record [<options>] -- <command> [<options>]",
639 static const struct option options
[] = {
640 OPT_CALLBACK('e', "event", NULL
, "event",
641 "event selector. use 'perf list' to list available events",
643 OPT_CALLBACK(0, "filter", NULL
, "filter",
644 "event filter", parse_filter
),
645 OPT_INTEGER('p', "pid", &target_pid
,
646 "record events on existing pid"),
647 OPT_INTEGER('r', "realtime", &realtime_prio
,
648 "collect data with this RT SCHED_FIFO priority"),
649 OPT_BOOLEAN('R', "raw-samples", &raw_samples
,
650 "collect raw sample records from all opened counters"),
651 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
652 "system-wide collection from all CPUs"),
653 OPT_BOOLEAN('A', "append", &append_file
,
654 "append to the output file to do incremental profiling"),
655 OPT_INTEGER('C', "profile_cpu", &profile_cpu
,
656 "CPU to profile on"),
657 OPT_BOOLEAN('f', "force", &force
,
658 "overwrite existing data file"),
659 OPT_LONG('c', "count", &default_interval
,
660 "event period to sample"),
661 OPT_STRING('o', "output", &output_name
, "file",
663 OPT_BOOLEAN('i', "inherit", &inherit
,
664 "child tasks inherit counters"),
665 OPT_INTEGER('F', "freq", &freq
,
666 "profile at this frequency"),
667 OPT_INTEGER('m', "mmap-pages", &mmap_pages
,
668 "number of mmap data pages"),
669 OPT_BOOLEAN('g', "call-graph", &call_graph
,
670 "do call-graph (stack chain/backtrace) recording"),
671 OPT_BOOLEAN('v', "verbose", &verbose
,
672 "be more verbose (show counter open errors, etc)"),
673 OPT_BOOLEAN('s', "stat", &inherit_stat
,
674 "per thread counts"),
675 OPT_BOOLEAN('d', "data", &sample_address
,
677 OPT_BOOLEAN('n', "no-samples", &no_samples
,
679 OPT_BOOLEAN('M', "multiplex", &multiplex
,
680 "multiplex counter output in a single channel"),
684 int cmd_record(int argc
, const char **argv
, const char *prefix __used
)
688 argc
= parse_options(argc
, argv
, options
, record_usage
,
689 PARSE_OPT_STOP_AT_NON_OPTION
);
690 if (!argc
&& target_pid
== -1 && !system_wide
&& profile_cpu
== -1)
691 usage_with_options(record_usage
, options
);
697 attrs
[0].type
= PERF_TYPE_HARDWARE
;
698 attrs
[0].config
= PERF_COUNT_HW_CPU_CYCLES
;
702 * User specified count overrides default frequency.
704 if (default_interval
)
707 default_interval
= freq
;
709 fprintf(stderr
, "frequency and count are zero, aborting\n");
713 for (counter
= 0; counter
< nr_counters
; counter
++) {
714 if (attrs
[counter
].sample_period
)
717 attrs
[counter
].sample_period
= default_interval
;
720 return __cmd_record(argc
, argv
);