4 * Builtin record command: Record the profile of a workload
5 * (or a CPU, or a PID) into the perf.data output file - for
6 * later analysis via perf report.
12 #include "util/util.h"
13 #include "util/parse-options.h"
14 #include "util/parse-events.h"
15 #include "util/string.h"
20 #define ALIGN(x, a) __ALIGN_MASK(x, (typeof(x))(a)-1)
21 #define __ALIGN_MASK(x, mask) (((x)+(mask))&~(mask))
23 static long default_interval
= 100000;
24 static long event_count
[MAX_COUNTERS
];
26 static int fd
[MAX_NR_CPUS
][MAX_COUNTERS
];
27 static int nr_cpus
= 0;
28 static unsigned int page_size
;
29 static unsigned int mmap_pages
= 128;
32 static const char *output_name
= "perf.data";
34 static unsigned int realtime_prio
= 0;
35 static int system_wide
= 0;
36 static pid_t target_pid
= -1;
37 static int inherit
= 1;
39 static int append_file
= 0;
41 const unsigned int default_count
[] = {
57 static unsigned int mmap_read_head(struct mmap_data
*md
)
59 struct perf_counter_mmap_page
*pc
= md
->base
;
69 static struct timeval last_read
, this_read
;
71 static __u64 bytes_written
;
73 static void mmap_read(struct mmap_data
*md
)
75 unsigned int head
= mmap_read_head(md
);
76 unsigned int old
= md
->prev
;
77 unsigned char *data
= md
->base
+ page_size
;
82 gettimeofday(&this_read
, NULL
);
85 * If we're further behind than half the buffer, there's a chance
86 * the writer will bite our tail and mess up the samples under us.
88 * If we somehow ended up ahead of the head, we got messed up.
90 * In either case, truncate and restart at head.
93 if (diff
> md
->mask
/ 2 || diff
< 0) {
97 timersub(&this_read
, &last_read
, &iv
);
98 msecs
= iv
.tv_sec
*1000 + iv
.tv_usec
/1000;
100 fprintf(stderr
, "WARNING: failed to keep up with mmap data."
101 " Last read %lu msecs ago.\n", msecs
);
104 * head points to a known good entry, start there.
109 last_read
= this_read
;
116 if ((old
& md
->mask
) + size
!= (head
& md
->mask
)) {
117 buf
= &data
[old
& md
->mask
];
118 size
= md
->mask
+ 1 - (old
& md
->mask
);
122 int ret
= write(output
, buf
, size
);
125 die("failed to write");
130 bytes_written
+= ret
;
134 buf
= &data
[old
& md
->mask
];
139 int ret
= write(output
, buf
, size
);
142 die("failed to write");
147 bytes_written
+= ret
;
153 static volatile int done
= 0;
155 static void sig_handler(int sig
)
160 static struct pollfd event_array
[MAX_NR_CPUS
* MAX_COUNTERS
];
161 static struct mmap_data mmap_array
[MAX_NR_CPUS
][MAX_COUNTERS
];
167 struct perf_event_header header
;
173 char filename
[PATH_MAX
];
177 struct perf_event_header header
;
183 static void pid_synthesize_comm_event(pid_t pid
, int full
)
185 struct comm_event comm_ev
;
186 char filename
[PATH_MAX
];
192 struct dirent dirent
, *next
;
194 snprintf(filename
, sizeof(filename
), "/proc/%d/stat", pid
);
196 fd
= open(filename
, O_RDONLY
);
198 fprintf(stderr
, "couldn't open %s\n", filename
);
201 if (read(fd
, bf
, sizeof(bf
)) < 0) {
202 fprintf(stderr
, "couldn't read %s\n", filename
);
207 /* 9027 (cat) R 6747 9027 6747 34816 9027 ... */
208 memset(&comm_ev
, 0, sizeof(comm_ev
));
209 field
= strchr(bf
, '(');
212 sep
= strchr(++field
, ')');
216 memcpy(comm_ev
.comm
, field
, size
++);
219 comm_ev
.header
.type
= PERF_EVENT_COMM
;
220 size
= ALIGN(size
, sizeof(uint64_t));
221 comm_ev
.header
.size
= sizeof(comm_ev
) - (sizeof(comm_ev
.comm
) - size
);
226 ret
= write(output
, &comm_ev
, comm_ev
.header
.size
);
228 perror("failed to write");
234 snprintf(filename
, sizeof(filename
), "/proc/%d/task", pid
);
236 tasks
= opendir(filename
);
237 while (!readdir_r(tasks
, &dirent
, &next
) && next
) {
239 pid
= strtol(dirent
.d_name
, &end
, 10);
245 ret
= write(output
, &comm_ev
, comm_ev
.header
.size
);
247 perror("failed to write");
255 fprintf(stderr
, "couldn't get COMM and pgid, malformed %s\n",
260 static void pid_synthesize_mmap_samples(pid_t pid
)
262 char filename
[PATH_MAX
];
265 snprintf(filename
, sizeof(filename
), "/proc/%d/maps", pid
);
267 fp
= fopen(filename
, "r");
269 fprintf(stderr
, "couldn't open %s\n", filename
);
273 char bf
[BUFSIZ
], *pbf
= bf
;
274 struct mmap_event mmap_ev
= {
275 .header
.type
= PERF_EVENT_MMAP
,
279 if (fgets(bf
, sizeof(bf
), fp
) == NULL
)
282 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
283 n
= hex2u64(pbf
, &mmap_ev
.start
);
287 n
= hex2u64(pbf
, &mmap_ev
.len
);
291 if (*pbf
== 'x') { /* vm_exec */
292 char *execname
= strrchr(bf
, ' ');
294 if (execname
== NULL
|| execname
[1] != '/')
298 size
= strlen(execname
);
299 execname
[size
- 1] = '\0'; /* Remove \n */
300 memcpy(mmap_ev
.filename
, execname
, size
);
301 size
= ALIGN(size
, sizeof(uint64_t));
302 mmap_ev
.len
-= mmap_ev
.start
;
303 mmap_ev
.header
.size
= (sizeof(mmap_ev
) -
304 (sizeof(mmap_ev
.filename
) - size
));
308 if (write(output
, &mmap_ev
, mmap_ev
.header
.size
) < 0) {
309 perror("failed to write");
318 static void synthesize_samples(void)
321 struct dirent dirent
, *next
;
323 proc
= opendir("/proc");
325 while (!readdir_r(proc
, &dirent
, &next
) && next
) {
329 pid
= strtol(dirent
.d_name
, &end
, 10);
330 if (*end
) /* only interested in proper numerical dirents */
333 pid_synthesize_comm_event(pid
, 1);
334 pid_synthesize_mmap_samples(pid
);
342 static void create_counter(int counter
, int cpu
, pid_t pid
)
344 struct perf_counter_attr attr
;
347 memset(&attr
, 0, sizeof(attr
));
348 attr
.config
= event_id
[counter
];
349 attr
.sample_period
= event_count
[counter
];
350 attr
.sample_type
= PERF_SAMPLE_IP
| PERF_SAMPLE_TID
| PERF_SAMPLE_PERIOD
;
353 attr
.sample_freq
= freq
;
357 attr
.inherit
= (cpu
< 0) && inherit
;
359 track
= 0; /* only the first counter needs these */
361 fd
[nr_cpu
][counter
] = sys_perf_counter_open(&attr
, pid
, cpu
, group_fd
, 0);
363 if (fd
[nr_cpu
][counter
] < 0) {
366 error("syscall returned with %d (%s)\n",
367 fd
[nr_cpu
][counter
], strerror(err
));
369 printf("Are you root?\n");
372 assert(fd
[nr_cpu
][counter
] >= 0);
373 fcntl(fd
[nr_cpu
][counter
], F_SETFL
, O_NONBLOCK
);
376 * First counter acts as the group leader:
378 if (group
&& group_fd
== -1)
379 group_fd
= fd
[nr_cpu
][counter
];
381 event_array
[nr_poll
].fd
= fd
[nr_cpu
][counter
];
382 event_array
[nr_poll
].events
= POLLIN
;
385 mmap_array
[nr_cpu
][counter
].counter
= counter
;
386 mmap_array
[nr_cpu
][counter
].prev
= 0;
387 mmap_array
[nr_cpu
][counter
].mask
= mmap_pages
*page_size
- 1;
388 mmap_array
[nr_cpu
][counter
].base
= mmap(NULL
, (mmap_pages
+1)*page_size
,
389 PROT_READ
, MAP_SHARED
, fd
[nr_cpu
][counter
], 0);
390 if (mmap_array
[nr_cpu
][counter
].base
== MAP_FAILED
) {
391 error("failed to mmap with %d (%s)\n", errno
, strerror(errno
));
396 static void open_counters(int cpu
, pid_t pid
)
401 pid_synthesize_comm_event(pid
, 0);
402 pid_synthesize_mmap_samples(pid
);
406 for (counter
= 0; counter
< nr_counters
; counter
++)
407 create_counter(counter
, cpu
, pid
);
412 static int __cmd_record(int argc
, const char **argv
)
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);
425 if (!stat(output_name
, &st
) && !force
&& !append_file
) {
426 fprintf(stderr
, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
431 flags
= O_CREAT
|O_RDWR
;
437 output
= open(output_name
, flags
, S_IRUSR
|S_IWUSR
);
439 perror("failed to create output file");
444 open_counters(-1, target_pid
!= -1 ? target_pid
: getpid());
445 } else for (i
= 0; i
< nr_cpus
; i
++)
446 open_counters(i
, target_pid
);
448 signal(SIGCHLD
, sig_handler
);
449 signal(SIGINT
, sig_handler
);
451 if (target_pid
== -1 && argc
) {
454 perror("failed to fork");
457 if (execvp(argv
[0], (char **)argv
)) {
465 struct sched_param param
;
467 param
.sched_priority
= realtime_prio
;
468 if (sched_setscheduler(0, SCHED_FIFO
, ¶m
)) {
469 printf("Could not set realtime priority.\n");
475 synthesize_samples();
480 for (i
= 0; i
< nr_cpu
; i
++) {
481 for (counter
= 0; counter
< nr_counters
; counter
++)
482 mmap_read(&mmap_array
[i
][counter
]);
486 ret
= poll(event_array
, nr_poll
, 100);
490 * Approximate RIP event size: 24 bytes.
493 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
494 (double)bytes_written
/ 1024.0 / 1024.0,
501 static const char * const record_usage
[] = {
502 "perf record [<options>] [<command>]",
503 "perf record [<options>] -- <command> [<options>]",
507 static char events_help_msg
[EVENTS_HELP_MAX
];
509 static const struct option options
[] = {
510 OPT_CALLBACK('e', "event", NULL
, "event",
511 events_help_msg
, parse_events
),
512 OPT_INTEGER('p', "pid", &target_pid
,
513 "record events on existing pid"),
514 OPT_INTEGER('r', "realtime", &realtime_prio
,
515 "collect data with this RT SCHED_FIFO priority"),
516 OPT_BOOLEAN('a', "all-cpus", &system_wide
,
517 "system-wide collection from all CPUs"),
518 OPT_BOOLEAN('A', "append", &append_file
,
519 "append to the output file to do incremental profiling"),
520 OPT_BOOLEAN('f', "force", &force
,
521 "overwrite existing data file"),
522 OPT_LONG('c', "count", &default_interval
,
523 "event period to sample"),
524 OPT_STRING('o', "output", &output_name
, "file",
526 OPT_BOOLEAN('i', "inherit", &inherit
,
527 "child tasks inherit counters"),
528 OPT_INTEGER('F', "freq", &freq
,
529 "profile at this frequency"),
530 OPT_INTEGER('m', "mmap-pages", &mmap_pages
,
531 "number of mmap data pages"),
535 int cmd_record(int argc
, const char **argv
, const char *prefix
)
539 create_events_help(events_help_msg
);
541 argc
= parse_options(argc
, argv
, options
, record_usage
, 0);
542 if (!argc
&& target_pid
== -1 && !system_wide
)
543 usage_with_options(record_usage
, options
);
550 for (counter
= 0; counter
< nr_counters
; counter
++) {
551 if (event_count
[counter
])
554 event_count
[counter
] = default_interval
;
557 return __cmd_record(argc
, argv
);