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"
24 #define ALIGN(x, a) __ALIGN_MASK(x, (typeof(x))(a)-1)
25 #define __ALIGN_MASK(x, mask) (((x)+(mask))&~(mask))
27 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
29 static long default_interval = 100000;
31 static int nr_cpus = 0;
32 static unsigned int page_size;
33 static unsigned int mmap_pages = 128;
36 static const char *output_name = "perf.data";
38 static unsigned int realtime_prio = 0;
39 static int raw_samples = 0;
40 static int system_wide = 0;
41 static int profile_cpu = -1;
42 static pid_t target_pid = -1;
43 static int inherit = 1;
45 static int append_file = 0;
46 static int call_graph = 0;
47 static int inherit_stat = 0;
48 static int no_samples = 0;
49 static int sample_address = 0;
52 static struct timeval last_read;
53 static struct timeval this_read;
55 static u64 bytes_written;
57 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
62 static int file_new = 1;
64 struct perf_header *header;
73 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
75 static unsigned long mmap_read_head(struct mmap_data *md)
77 struct perf_counter_mmap_page *pc = md->base;
86 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
88 struct perf_counter_mmap_page *pc = md->base;
91 * ensure all reads are done before we write the tail out.
97 static void write_output(void *buf, size_t size)
100 int ret = write(output, buf, size);
103 die("failed to write");
108 bytes_written += ret;
112 static void mmap_read(struct mmap_data *md)
114 unsigned int head = mmap_read_head(md);
115 unsigned int old = md->prev;
116 unsigned char *data = md->base + page_size;
121 gettimeofday(&this_read, NULL);
124 * If we're further behind than half the buffer, there's a chance
125 * the writer will bite our tail and mess up the samples under us.
127 * If we somehow ended up ahead of the head, we got messed up.
129 * In either case, truncate and restart at head.
136 timersub(&this_read, &last_read, &iv);
137 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
139 fprintf(stderr, "WARNING: failed to keep up with mmap data."
140 " Last read %lu msecs ago.\n", msecs);
143 * head points to a known good entry, start there.
148 last_read = this_read;
155 if ((old & md->mask) + size != (head & md->mask)) {
156 buf = &data[old & md->mask];
157 size = md->mask + 1 - (old & md->mask);
160 write_output(buf, size);
163 buf = &data[old & md->mask];
167 write_output(buf, size);
170 mmap_write_tail(md, old);
173 static volatile int done = 0;
174 static volatile int signr = -1;
176 static void sig_handler(int sig)
182 static void sig_atexit(void)
187 signal(signr, SIG_DFL);
188 kill(getpid(), signr);
191 static pid_t pid_synthesize_comm_event(pid_t pid, int full)
193 struct comm_event comm_ev;
194 char filename[PATH_MAX];
199 struct dirent dirent, *next;
202 snprintf(filename, sizeof(filename), "/proc/%d/status", pid);
204 fp = fopen(filename, "r");
207 * We raced with a task exiting - just return:
210 fprintf(stderr, "couldn't open %s\n", filename);
214 memset(&comm_ev, 0, sizeof(comm_ev));
215 while (!comm_ev.comm[0] || !comm_ev.pid) {
216 if (fgets(bf, sizeof(bf), fp) == NULL)
219 if (memcmp(bf, "Name:", 5) == 0) {
221 while (*name && isspace(*name))
223 size = strlen(name) - 1;
224 memcpy(comm_ev.comm, name, size++);
225 } else if (memcmp(bf, "Tgid:", 5) == 0) {
226 char *tgids = bf + 5;
227 while (*tgids && isspace(*tgids))
229 tgid = comm_ev.pid = atoi(tgids);
233 comm_ev.header.type = PERF_EVENT_COMM;
234 size = ALIGN(size, sizeof(u64));
235 comm_ev.header.size = sizeof(comm_ev) - (sizeof(comm_ev.comm) - size);
240 write_output(&comm_ev, comm_ev.header.size);
244 snprintf(filename, sizeof(filename), "/proc/%d/task", pid);
246 tasks = opendir(filename);
247 while (!readdir_r(tasks, &dirent, &next) && next) {
249 pid = strtol(dirent.d_name, &end, 10);
255 write_output(&comm_ev, comm_ev.header.size);
264 fprintf(stderr, "couldn't get COMM and pgid, malformed %s\n",
269 static void pid_synthesize_mmap_samples(pid_t pid, pid_t tgid)
271 char filename[PATH_MAX];
274 snprintf(filename, sizeof(filename), "/proc/%d/maps", pid);
276 fp = fopen(filename, "r");
279 * We raced with a task exiting - just return:
282 fprintf(stderr, "couldn't open %s\n", filename);
286 char bf[BUFSIZ], *pbf = bf;
287 struct mmap_event mmap_ev = {
288 .header = { .type = PERF_EVENT_MMAP },
292 if (fgets(bf, sizeof(bf), fp) == NULL)
295 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */
296 n = hex2u64(pbf, &mmap_ev.start);
300 n = hex2u64(pbf, &mmap_ev.len);
304 if (*pbf == 'x') { /* vm_exec */
305 char *execname = strchr(bf, '/');
308 if (execname == NULL)
309 execname = strstr(bf, "[vdso]");
311 if (execname == NULL)
314 size = strlen(execname);
315 execname[size - 1] = '\0'; /* Remove \n */
316 memcpy(mmap_ev.filename, execname, size);
317 size = ALIGN(size, sizeof(u64));
318 mmap_ev.len -= mmap_ev.start;
319 mmap_ev.header.size = (sizeof(mmap_ev) -
320 (sizeof(mmap_ev.filename) - size));
324 write_output(&mmap_ev, mmap_ev.header.size);
331 static void synthesize_all(void)
334 struct dirent dirent, *next;
336 proc = opendir("/proc");
338 while (!readdir_r(proc, &dirent, &next) && next) {
342 pid = strtol(dirent.d_name, &end, 10);
343 if (*end) /* only interested in proper numerical dirents */
346 tgid = pid_synthesize_comm_event(pid, 1);
347 pid_synthesize_mmap_samples(pid, tgid);
355 static struct perf_header_attr *get_header_attr(struct perf_counter_attr *a, int nr)
357 struct perf_header_attr *h_attr;
359 if (nr < header->attrs) {
360 h_attr = header->attr[nr];
362 h_attr = perf_header_attr__new(a);
363 perf_header__add_attr(header, h_attr);
369 static void create_counter(int counter, int cpu, pid_t pid)
371 struct perf_counter_attr *attr = attrs + counter;
372 struct perf_header_attr *h_attr;
373 int track = !counter; /* only the first counter needs these */
381 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
382 PERF_FORMAT_TOTAL_TIME_RUNNING |
385 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
388 attr->sample_type |= PERF_SAMPLE_PERIOD;
390 attr->sample_freq = freq;
394 attr->sample_freq = 0;
397 attr->inherit_stat = 1;
400 attr->sample_type |= PERF_SAMPLE_ADDR;
403 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
406 attr->sample_type |= PERF_SAMPLE_RAW;
410 attr->inherit = (cpu < 0) && inherit;
414 fd[nr_cpu][counter] = sys_perf_counter_open(attr, pid, cpu, group_fd, 0);
416 if (fd[nr_cpu][counter] < 0) {
420 die("Permission error - are you root?\n");
421 else if (err == ENODEV && profile_cpu != -1)
422 die("No such device - did you specify an out-of-range profile CPU?\n");
425 * If it's cycles then fall back to hrtimer
426 * based cpu-clock-tick sw counter, which
427 * is always available even if no PMU support:
429 if (attr->type == PERF_TYPE_HARDWARE
430 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
433 warning(" ... trying to fall back to cpu-clock-ticks\n");
434 attr->type = PERF_TYPE_SOFTWARE;
435 attr->config = PERF_COUNT_SW_CPU_CLOCK;
439 error("perfcounter syscall returned with %d (%s)\n",
440 fd[nr_cpu][counter], strerror(err));
441 die("No CONFIG_PERF_COUNTERS=y kernel support configured?\n");
445 h_attr = get_header_attr(attr, counter);
448 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
449 fprintf(stderr, "incompatible append\n");
454 if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
455 perror("Unable to read perf file descriptor\n");
459 perf_header_attr__add_id(h_attr, read_data.id);
461 assert(fd[nr_cpu][counter] >= 0);
462 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
465 * First counter acts as the group leader:
467 if (group && group_fd == -1)
468 group_fd = fd[nr_cpu][counter];
470 event_array[nr_poll].fd = fd[nr_cpu][counter];
471 event_array[nr_poll].events = POLLIN;
474 mmap_array[nr_cpu][counter].counter = counter;
475 mmap_array[nr_cpu][counter].prev = 0;
476 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
477 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
478 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
479 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
480 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
484 ioctl(fd[nr_cpu][counter], PERF_COUNTER_IOC_ENABLE);
487 static void open_counters(int cpu, pid_t pid)
492 for (counter = 0; counter < nr_counters; counter++)
493 create_counter(counter, cpu, pid);
498 static void atexit_header(void)
500 header->data_size += bytes_written;
502 perf_header__write(header, output);
505 static int __cmd_record(int argc, const char **argv)
513 page_size = sysconf(_SC_PAGE_SIZE);
514 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
515 assert(nr_cpus <= MAX_NR_CPUS);
516 assert(nr_cpus >= 0);
519 signal(SIGCHLD, sig_handler);
520 signal(SIGINT, sig_handler);
522 if (!stat(output_name, &st) && st.st_size) {
523 if (!force && !append_file) {
524 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
532 flags = O_CREAT|O_RDWR;
538 output = open(output_name, flags, S_IRUSR|S_IWUSR);
540 perror("failed to create output file");
545 header = perf_header__read(output);
547 header = perf_header__new();
549 atexit(atexit_header);
556 open_counters(profile_cpu, pid);
558 if (profile_cpu != -1) {
559 open_counters(profile_cpu, target_pid);
561 for (i = 0; i < nr_cpus; i++)
562 open_counters(i, target_pid);
567 perf_header__write(header, output);
570 pid_t tgid = pid_synthesize_comm_event(pid, 0);
571 pid_synthesize_mmap_samples(pid, tgid);
575 if (target_pid == -1 && argc) {
578 perror("failed to fork");
581 if (execvp(argv[0], (char **)argv)) {
589 struct sched_param param;
591 param.sched_priority = realtime_prio;
592 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
593 printf("Could not set realtime priority.\n");
601 for (i = 0; i < nr_cpu; i++) {
602 for (counter = 0; counter < nr_counters; counter++)
603 mmap_read(&mmap_array[i][counter]);
606 if (hits == samples) {
609 ret = poll(event_array, nr_poll, 100);
614 * Approximate RIP event size: 24 bytes.
617 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
618 (double)bytes_written / 1024.0 / 1024.0,
625 static const char * const record_usage[] = {
626 "perf record [<options>] [<command>]",
627 "perf record [<options>] -- <command> [<options>]",
631 static const struct option options[] = {
632 OPT_CALLBACK('e', "event", NULL, "event",
633 "event selector. use 'perf list' to list available events",
635 OPT_INTEGER('p', "pid", &target_pid,
636 "record events on existing pid"),
637 OPT_INTEGER('r', "realtime", &realtime_prio,
638 "collect data with this RT SCHED_FIFO priority"),
639 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
640 "collect raw sample records from all opened counters"),
641 OPT_BOOLEAN('a', "all-cpus", &system_wide,
642 "system-wide collection from all CPUs"),
643 OPT_BOOLEAN('A', "append", &append_file,
644 "append to the output file to do incremental profiling"),
645 OPT_INTEGER('C', "profile_cpu", &profile_cpu,
646 "CPU to profile on"),
647 OPT_BOOLEAN('f', "force", &force,
648 "overwrite existing data file"),
649 OPT_LONG('c', "count", &default_interval,
650 "event period to sample"),
651 OPT_STRING('o', "output", &output_name, "file",
653 OPT_BOOLEAN('i', "inherit", &inherit,
654 "child tasks inherit counters"),
655 OPT_INTEGER('F', "freq", &freq,
656 "profile at this frequency"),
657 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
658 "number of mmap data pages"),
659 OPT_BOOLEAN('g', "call-graph", &call_graph,
660 "do call-graph (stack chain/backtrace) recording"),
661 OPT_BOOLEAN('v', "verbose", &verbose,
662 "be more verbose (show counter open errors, etc)"),
663 OPT_BOOLEAN('s', "stat", &inherit_stat,
664 "per thread counts"),
665 OPT_BOOLEAN('d', "data", &sample_address,
667 OPT_BOOLEAN('n', "no-samples", &no_samples,
672 int cmd_record(int argc, const char **argv, const char *prefix __used)
676 argc = parse_options(argc, argv, options, record_usage,
677 PARSE_OPT_STOP_AT_NON_OPTION);
678 if (!argc && target_pid == -1 && !system_wide)
679 usage_with_options(record_usage, options);
683 attrs[0].type = PERF_TYPE_HARDWARE;
684 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
687 for (counter = 0; counter < nr_counters; counter++) {
688 if (attrs[counter].sample_period)
691 attrs[counter].sample_period = default_interval;
694 return __cmd_record(argc, argv);