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"
17 #include "util/header.h"
18 #include "util/event.h"
19 #include "util/debug.h"
20 #include "util/trace-event.h"
25 #define ALIGN(x, a) __ALIGN_MASK(x, (typeof(x))(a)-1)
26 #define __ALIGN_MASK(x, mask) (((x)+(mask))&~(mask))
28 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
30 static long default_interval = 100000;
32 static int nr_cpus = 0;
33 static unsigned int page_size;
34 static unsigned int mmap_pages = 128;
37 static const char *output_name = "perf.data";
39 static unsigned int realtime_prio = 0;
40 static int raw_samples = 0;
41 static int system_wide = 0;
42 static int profile_cpu = -1;
43 static pid_t target_pid = -1;
44 static int inherit = 1;
46 static int append_file = 0;
47 static int call_graph = 0;
48 static int inherit_stat = 0;
49 static int no_samples = 0;
50 static int sample_address = 0;
51 static int multiplex = 0;
54 static struct timeval last_read;
55 static struct timeval this_read;
57 static u64 bytes_written;
59 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
64 static int file_new = 1;
66 struct perf_header *header;
75 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
77 static unsigned long mmap_read_head(struct mmap_data *md)
79 struct perf_counter_mmap_page *pc = md->base;
88 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
90 struct perf_counter_mmap_page *pc = md->base;
93 * ensure all reads are done before we write the tail out.
99 static void write_output(void *buf, size_t size)
102 int ret = write(output, buf, size);
105 die("failed to write");
110 bytes_written += ret;
114 static void mmap_read(struct mmap_data *md)
116 unsigned int head = mmap_read_head(md);
117 unsigned int old = md->prev;
118 unsigned char *data = md->base + page_size;
123 gettimeofday(&this_read, NULL);
126 * If we're further behind than half the buffer, there's a chance
127 * the writer will bite our tail and mess up the samples under us.
129 * If we somehow ended up ahead of the head, we got messed up.
131 * In either case, truncate and restart at head.
138 timersub(&this_read, &last_read, &iv);
139 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
141 fprintf(stderr, "WARNING: failed to keep up with mmap data."
142 " Last read %lu msecs ago.\n", msecs);
145 * head points to a known good entry, start there.
150 last_read = this_read;
157 if ((old & md->mask) + size != (head & md->mask)) {
158 buf = &data[old & md->mask];
159 size = md->mask + 1 - (old & md->mask);
162 write_output(buf, size);
165 buf = &data[old & md->mask];
169 write_output(buf, size);
172 mmap_write_tail(md, old);
175 static volatile int done = 0;
176 static volatile int signr = -1;
178 static void sig_handler(int sig)
184 static void sig_atexit(void)
189 signal(signr, SIG_DFL);
190 kill(getpid(), signr);
193 static pid_t pid_synthesize_comm_event(pid_t pid, int full)
195 struct comm_event comm_ev;
196 char filename[PATH_MAX];
201 struct dirent dirent, *next;
204 snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
206 fp = fopen(filename, "r");
209 * We raced with a task exiting - just return:
212 fprintf(stderr, "couldn't open %s\n", filename);
216 memset(&comm_ev, 0, sizeof(comm_ev));
217 while (!comm_ev.comm[0] || !comm_ev.pid) {
218 if (fgets(bf, sizeof(bf), fp) == NULL)
221 if (memcmp(bf, "Name:", 5) == 0) {
223 while (*name && isspace(*name))
225 size = strlen(name) - 1;
226 memcpy(comm_ev.comm, name, size++);
227 } else if (memcmp(bf, "Tgid:", 5) == 0) {
228 char *tgids = bf + 5;
229 while (*tgids && isspace(*tgids))
231 tgid = comm_ev.pid = atoi(tgids);
235 comm_ev.header.type = PERF_EVENT_COMM;
236 size = ALIGN(size, sizeof(u64));
237 comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
242 write_output(&comm_ev, comm_ev.header.size);
246 snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
248 tasks = opendir(filename);
249 while (!readdir_r(tasks, &dirent, &next) && next) {
251 pid = strtol(dirent.d_name, &end, 10);
257 write_output(&comm_ev, comm_ev.header.size);
266 fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
271 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
273 char filename[PATH_MAX];
276 snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
278 fp = fopen(filename, "r");
281 * We raced with a task exiting - just return:
284 fprintf(stderr, "couldn't open %s\n", filename);
288 char bf[BUFSIZ], *pbf = bf;
289 struct mmap_event mmap_ev = {
290 .header = { .type = PERF_EVENT_MMAP },
294 if (fgets(bf, sizeof(bf), fp) == NULL)
297 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
298 n = hex2u64(pbf, &mmap_ev.start);
302 n = hex2u64(pbf, &mmap_ev.len);
306 if (*pbf == 'x') { /* vm_exec */
307 char *execname = strchr(bf, '/');
310 if (execname == NULL)
311 execname = strstr(bf, "[vdso]");
313 if (execname == NULL)
316 size = strlen(execname);
317 execname[size - 1] = '\0'; /* Remove \n */
318 memcpy(mmap_ev.filename, execname, size);
319 size = ALIGN(size, sizeof(u64));
320 mmap_ev.len -= mmap_ev.start;
321 mmap_ev.header.size = (sizeof(mmap_ev) -
322 (sizeof(mmap_ev.filename) - size));
326 write_output(&mmap_ev, mmap_ev.header.size);
333 static void synthesize_all(void)
336 struct dirent dirent, *next;
338 proc = opendir("/proc");
340 while (!readdir_r(proc, &dirent, &next) && next) {
344 pid = strtol(dirent.d_name, &end, 10);
345 if (*end) /* only interested in proper numerical dirents */
348 tgid = pid_synthesize_comm_event(pid, 1);
349 pid_synthesize_mmap_samples(pid, tgid);
357 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
359 struct perf_header_attr *h_attr;
361 if (nr < header->attrs) {
362 h_attr = header->attr[nr];
364 h_attr = perf_header_attr__new(a);
365 perf_header__add_attr(header, h_attr);
371 static void create_counter(int counter, int cpu, pid_t pid)
373 struct perf_counter_attr *attr = attrs + counter;
374 struct perf_header_attr *h_attr;
375 int track = !counter; /* only the first counter needs these */
383 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
384 PERF_FORMAT_TOTAL_TIME_RUNNING |
387 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
390 attr->sample_type |= PERF_SAMPLE_PERIOD;
392 attr->sample_freq = freq;
396 attr->sample_freq = 0;
399 attr->inherit_stat = 1;
402 attr->sample_type |= PERF_SAMPLE_ADDR;
405 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
408 attr->sample_type |= PERF_SAMPLE_TIME;
409 attr->sample_type |= PERF_SAMPLE_RAW;
410 attr->sample_type |= PERF_SAMPLE_CPU;
415 attr->inherit = (cpu < 0) && inherit;
419 fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
421 if (fd[nr_cpu][counter] < 0) {
425 die("Permission error - are you root?\n");
426 else if (err == ENODEV && profile_cpu != -1)
427 die("No such device - did you specify an out-of-range profile CPU?\n");
430 * If it's cycles then fall back to hrtimer
431 * based cpu-clock-tick sw counter, which
432 * is always available even if no PMU support:
434 if (attr->type == PERF_TYPE_HARDWARE
435 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
438 warning(" ... trying to fall back to cpu-clock-ticks\n");
439 attr->type = PERF_TYPE_SOFTWARE;
440 attr->config = PERF_COUNT_SW_CPU_CLOCK;
444 error("perfcounter syscall returned with %d (%s)\n",
445 fd[nr_cpu][counter], strerror(err));
446 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
450 h_attr = get_header_attr(attr, counter);
453 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
454 fprintf(stderr, "incompatible append\n");
459 if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
460 perror("Unable to read perf file descriptor\n");
464 perf_header_attr__add_id(h_attr, read_data.id);
466 assert(fd[nr_cpu][counter] >= 0);
467 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
470 * First counter acts as the group leader:
472 if (group && group_fd == -1)
473 group_fd = fd[nr_cpu][counter];
475 event_array[nr_poll].fd = fd[nr_cpu][counter];
476 event_array[nr_poll].events = POLLIN;
479 mmap_array[nr_cpu][counter].counter = counter;
480 mmap_array[nr_cpu][counter].prev = 0;
481 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
482 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
483 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
484 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
485 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
489 if (multiplex && fd[nr_cpu][counter] != group_fd)
490 ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_SET_OUTPUT, group_fd);
492 ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
495 static void open_counters(int cpu, pid_t pid)
500 for (counter = 0; counter < nr_counters; counter++)
501 create_counter(counter, cpu, pid);
506 static void atexit_header(void)
508 header->data_size += bytes_written;
510 perf_header__write(header, output);
513 static int __cmd_record(int argc, const char **argv)
521 page_size = sysconf(_SC_PAGE_SIZE);
522 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
523 assert(nr_cpus <= MAX_NR_CPUS);
524 assert(nr_cpus >= 0);
527 signal(SIGCHLD, sig_handler);
528 signal(SIGINT, sig_handler);
530 if (!stat(output_name, &st) && st.st_size) {
531 if (!force && !append_file) {
532 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
540 flags = O_CREAT|O_RDWR;
546 output = open(output_name, flags, S_IRUSR|S_IWUSR);
548 perror("failed to create output file");
553 header = perf_header__read(output);
555 header = perf_header__new();
559 read_tracing_data(attrs, nr_counters);
561 for (i = 0; i < nr_counters; i++) {
562 if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
563 read_tracing_data(attrs, nr_counters);
568 atexit(atexit_header);
575 open_counters(profile_cpu, pid);
577 if (profile_cpu != -1) {
578 open_counters(profile_cpu, target_pid);
580 for (i = 0; i < nr_cpus; i++)
581 open_counters(i, target_pid);
586 perf_header__write(header, output);
589 pid_t tgid = pid_synthesize_comm_event(pid, 0);
590 pid_synthesize_mmap_samples(pid, tgid);
594 if (target_pid == -1 && argc) {
597 perror("failed to fork");
600 if (execvp(argv[0], (char **)argv)) {
608 struct sched_param param;
610 param.sched_priority = realtime_prio;
611 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
612 printf("Could not set realtime priority.\n");
620 for (i = 0; i < nr_cpu; i++) {
621 for (counter = 0; counter < nr_counters; counter++)
622 mmap_read(&mmap_array[i][counter]);
625 if (hits == samples) {
628 ret = poll(event_array, nr_poll, 100);
633 * Approximate RIP event size: 24 bytes.
636 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
637 (double)bytes_written / 1024.0 / 1024.0,
644 static const char * const record_usage[] = {
645 "perf record [<options>] [<command>]",
646 "perf record [<options>] -- <command> [<options>]",
650 static const struct option options[] = {
651 OPT_CALLBACK('e', "event", NULL, "event",
652 "event selector. use 'perf list' to list available events",
654 OPT_INTEGER('p', "pid", &target_pid,
655 "record events on existing pid"),
656 OPT_INTEGER('r', "realtime", &realtime_prio,
657 "collect data with this RT SCHED_FIFO priority"),
658 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
659 "collect raw sample records from all opened counters"),
660 OPT_BOOLEAN('a', "all-cpus", &system_wide,
661 "system-wide collection from all CPUs"),
662 OPT_BOOLEAN('A', "append", &append_file,
663 "append to the output file to do incremental profiling"),
664 OPT_INTEGER('C', "profile_cpu", &profile_cpu,
665 "CPU to profile on"),
666 OPT_BOOLEAN('f', "force", &force,
667 "overwrite existing data file"),
668 OPT_LONG('c', "count", &default_interval,
669 "event period to sample"),
670 OPT_STRING('o', "output", &output_name, "file",
672 OPT_BOOLEAN('i', "inherit", &inherit,
673 "child tasks inherit counters"),
674 OPT_INTEGER('F', "freq", &freq,
675 "profile at this frequency"),
676 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
677 "number of mmap data pages"),
678 OPT_BOOLEAN('g', "call-graph", &call_graph,
679 "do call-graph (stack chain/backtrace) recording"),
680 OPT_BOOLEAN('v', "verbose", &verbose,
681 "be more verbose (show counter open errors, etc)"),
682 OPT_BOOLEAN('s', "stat", &inherit_stat,
683 "per thread counts"),
684 OPT_BOOLEAN('d', "data", &sample_address,
686 OPT_BOOLEAN('n', "no-samples", &no_samples,
688 OPT_BOOLEAN('M', "multiplex", &multiplex,
689 "multiplex counter output in a single channel"),
693 int cmd_record(int argc, const char **argv, const char *prefix __used)
697 argc = parse_options(argc, argv, options, record_usage,
698 PARSE_OPT_STOP_AT_NON_OPTION);
699 if (!argc && target_pid == -1 && !system_wide)
700 usage_with_options(record_usage, options);
704 attrs[0].type = PERF_TYPE_HARDWARE;
705 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
708 for (counter = 0; counter < nr_counters; counter++) {
709 if (attrs[counter].sample_period)
712 attrs[counter].sample_period = default_interval;
715 return __cmd_record(argc, argv);