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/symbol.h"
25 static int fd[MAX_NR_CPUS][MAX_COUNTERS];
27 static long default_interval = 0;
29 static int nr_cpus = 0;
30 static unsigned int page_size;
31 static unsigned int mmap_pages = 128;
32 static int freq = 1000;
34 static const char *output_name = "perf.data";
36 static unsigned int realtime_prio = 0;
37 static int raw_samples = 0;
38 static int system_wide = 0;
39 static int profile_cpu = -1;
40 static pid_t target_pid = -1;
41 static pid_t child_pid = -1;
42 static int inherit = 1;
44 static int append_file = 0;
45 static int call_graph = 0;
46 static int inherit_stat = 0;
47 static int no_samples = 0;
48 static int sample_address = 0;
49 static int multiplex = 0;
50 static int multiplex_fd = -1;
52 static long samples = 0;
53 static struct timeval last_read;
54 static struct timeval this_read;
56 static u64 bytes_written = 0;
58 static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
60 static int nr_poll = 0;
61 static int nr_cpu = 0;
63 static int file_new = 1;
65 struct perf_header *header = NULL;
74 static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
76 static unsigned long mmap_read_head(struct mmap_data *md)
78 struct perf_event_mmap_page *pc = md->base;
87 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
89 struct perf_event_mmap_page *pc = md->base;
92 * ensure all reads are done before we write the tail out.
98 static void write_output(void *buf, size_t size)
101 int ret = write(output, buf, size);
104 die("failed to write");
109 bytes_written += ret;
113 static void write_event(event_t *buf, size_t size)
116 * Add it to the list of DSOs, so that when we finish this
117 * record session we can pick the available build-ids.
119 if (buf->header.type == PERF_RECORD_MMAP)
120 dsos__findnew(buf->mmap.filename);
122 write_output(buf, size);
125 static int process_synthesized_event(event_t *event)
127 write_event(event, event->header.size);
131 static void mmap_read(struct mmap_data *md)
133 unsigned int head = mmap_read_head(md);
134 unsigned int old = md->prev;
135 unsigned char *data = md->base + page_size;
140 gettimeofday(&this_read, NULL);
143 * If we're further behind than half the buffer, there's a chance
144 * the writer will bite our tail and mess up the samples under us.
146 * If we somehow ended up ahead of the head, we got messed up.
148 * In either case, truncate and restart at head.
155 timersub(&this_read, &last_read, &iv);
156 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
158 fprintf(stderr, "WARNING: failed to keep up with mmap data."
159 " Last read %lu msecs ago.\n", msecs);
162 * head points to a known good entry, start there.
167 last_read = this_read;
174 if ((old & md->mask) + size != (head & md->mask)) {
175 buf = &data[old & md->mask];
176 size = md->mask + 1 - (old & md->mask);
179 write_event(buf, size);
182 buf = &data[old & md->mask];
186 write_event(buf, size);
189 mmap_write_tail(md, old);
192 static volatile int done = 0;
193 static volatile int signr = -1;
195 static void sig_handler(int sig)
201 static void sig_atexit(void)
204 kill(child_pid, SIGTERM);
209 signal(signr, SIG_DFL);
210 kill(getpid(), signr);
215 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
217 struct perf_header_attr *h_attr;
219 if (nr < header->attrs) {
220 h_attr = header->attr[nr];
222 h_attr = perf_header_attr__new(a);
223 perf_header__add_attr(header, h_attr);
229 static void create_counter(int counter, int cpu, pid_t pid)
231 char *filter = filters[counter];
232 struct perf_event_attr *attr = attrs + counter;
233 struct perf_header_attr *h_attr;
234 int track = !counter; /* only the first counter needs these */
243 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
244 PERF_FORMAT_TOTAL_TIME_RUNNING |
247 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
250 attr->sample_type |= PERF_SAMPLE_PERIOD;
252 attr->sample_freq = freq;
256 attr->sample_freq = 0;
259 attr->inherit_stat = 1;
262 attr->sample_type |= PERF_SAMPLE_ADDR;
265 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
268 attr->sample_type |= PERF_SAMPLE_TIME;
269 attr->sample_type |= PERF_SAMPLE_RAW;
270 attr->sample_type |= PERF_SAMPLE_CPU;
275 attr->inherit = (cpu < 0) && inherit;
279 fd[nr_cpu][counter] = sys_perf_event_open(attr, pid, cpu, group_fd, 0);
281 if (fd[nr_cpu][counter] < 0) {
285 die("Permission error - are you root?\n");
286 else if (err == ENODEV && profile_cpu != -1)
287 die("No such device - did you specify an out-of-range profile CPU?\n");
290 * If it's cycles then fall back to hrtimer
291 * based cpu-clock-tick sw counter, which
292 * is always available even if no PMU support:
294 if (attr->type == PERF_TYPE_HARDWARE
295 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
298 warning(" ... trying to fall back to cpu-clock-ticks\n");
299 attr->type = PERF_TYPE_SOFTWARE;
300 attr->config = PERF_COUNT_SW_CPU_CLOCK;
304 error("perfcounter syscall returned with %d (%s)\n",
305 fd[nr_cpu][counter], strerror(err));
306 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
310 h_attr = get_header_attr(attr, counter);
313 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
314 fprintf(stderr, "incompatible append\n");
319 if (read(fd[nr_cpu][counter], &read_data, sizeof(read_data)) == -1) {
320 perror("Unable to read perf file descriptor\n");
324 perf_header_attr__add_id(h_attr, read_data.id);
326 assert(fd[nr_cpu][counter] >= 0);
327 fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
330 * First counter acts as the group leader:
332 if (group && group_fd == -1)
333 group_fd = fd[nr_cpu][counter];
334 if (multiplex && multiplex_fd == -1)
335 multiplex_fd = fd[nr_cpu][counter];
337 if (multiplex && fd[nr_cpu][counter] != multiplex_fd) {
339 ret = ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
342 event_array[nr_poll].fd = fd[nr_cpu][counter];
343 event_array[nr_poll].events = POLLIN;
346 mmap_array[nr_cpu][counter].counter = counter;
347 mmap_array[nr_cpu][counter].prev = 0;
348 mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
349 mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
350 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter], 0);
351 if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
352 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
357 if (filter != NULL) {
358 ret = ioctl(fd[nr_cpu][counter],
359 PERF_EVENT_IOC_SET_FILTER, filter);
361 error("failed to set filter with %d (%s)\n", errno,
367 ioctl(fd[nr_cpu][counter], PERF_EVENT_IOC_ENABLE);
370 static void open_counters(int cpu, pid_t pid)
375 for (counter = 0; counter < nr_counters; counter++)
376 create_counter(counter, cpu, pid);
381 static void atexit_header(void)
383 header->data_size += bytes_written;
385 perf_header__write(header, output, true);
388 static int __cmd_record(int argc, const char **argv)
395 unsigned long waking = 0;
397 page_size = sysconf(_SC_PAGE_SIZE);
398 nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);
399 assert(nr_cpus <= MAX_NR_CPUS);
400 assert(nr_cpus >= 0);
403 signal(SIGCHLD, sig_handler);
404 signal(SIGINT, sig_handler);
406 if (!stat(output_name, &st) && st.st_size) {
407 if (!force && !append_file) {
408 fprintf(stderr, "Error, output file %s exists, use -A to append or -f to overwrite.\n",
416 flags = O_CREAT|O_RDWR;
422 output = open(output_name, flags, S_IRUSR|S_IWUSR);
424 perror("failed to create output file");
429 header = perf_header__read(output);
431 header = perf_header__new();
434 perf_header__set_feat(header, HEADER_TRACE_INFO);
436 for (i = 0; i < nr_counters; i++) {
437 if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
438 perf_header__set_feat(header, HEADER_TRACE_INFO);
444 atexit(atexit_header);
451 open_counters(profile_cpu, pid);
453 if (profile_cpu != -1) {
454 open_counters(profile_cpu, target_pid);
456 for (i = 0; i < nr_cpus; i++)
457 open_counters(i, target_pid);
462 perf_header__write(header, output, false);
465 event__synthesize_thread(pid, process_synthesized_event);
467 event__synthesize_threads(process_synthesized_event);
469 if (target_pid == -1 && argc) {
472 die("failed to fork");
475 if (execvp(argv[0], (char **)argv)) {
481 * Wait a bit for the execv'ed child to appear
482 * and be updated in /proc
483 * FIXME: Do you know a less heuristical solution?
486 event__synthesize_thread(pid,
487 process_synthesized_event);
494 struct sched_param param;
496 param.sched_priority = realtime_prio;
497 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
498 pr_err("Could not set realtime priority.\n");
506 for (i = 0; i < nr_cpu; i++) {
507 for (counter = 0; counter < nr_counters; counter++) {
508 if (mmap_array[i][counter].base)
509 mmap_read(&mmap_array[i][counter]);
513 if (hits == samples) {
516 ret = poll(event_array, nr_poll, -1);
521 for (i = 0; i < nr_cpu; i++) {
522 for (counter = 0; counter < nr_counters; counter++)
523 ioctl(fd[i][counter], PERF_EVENT_IOC_DISABLE);
528 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
531 * Approximate RIP event size: 24 bytes.
534 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
535 (double)bytes_written / 1024.0 / 1024.0,
542 static const char * const record_usage[] = {
543 "perf record [<options>] [<command>]",
544 "perf record [<options>] -- <command> [<options>]",
548 static const struct option options[] = {
549 OPT_CALLBACK('e', "event", NULL, "event",
550 "event selector. use 'perf list' to list available events",
552 OPT_CALLBACK(0, "filter", NULL, "filter",
553 "event filter", parse_filter),
554 OPT_INTEGER('p', "pid", &target_pid,
555 "record events on existing pid"),
556 OPT_INTEGER('r', "realtime", &realtime_prio,
557 "collect data with this RT SCHED_FIFO priority"),
558 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
559 "collect raw sample records from all opened counters"),
560 OPT_BOOLEAN('a', "all-cpus", &system_wide,
561 "system-wide collection from all CPUs"),
562 OPT_BOOLEAN('A', "append", &append_file,
563 "append to the output file to do incremental profiling"),
564 OPT_INTEGER('C', "profile_cpu", &profile_cpu,
565 "CPU to profile on"),
566 OPT_BOOLEAN('f', "force", &force,
567 "overwrite existing data file"),
568 OPT_LONG('c', "count", &default_interval,
569 "event period to sample"),
570 OPT_STRING('o', "output", &output_name, "file",
572 OPT_BOOLEAN('i', "inherit", &inherit,
573 "child tasks inherit counters"),
574 OPT_INTEGER('F', "freq", &freq,
575 "profile at this frequency"),
576 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
577 "number of mmap data pages"),
578 OPT_BOOLEAN('g', "call-graph", &call_graph,
579 "do call-graph (stack chain/backtrace) recording"),
580 OPT_BOOLEAN('v', "verbose", &verbose,
581 "be more verbose (show counter open errors, etc)"),
582 OPT_BOOLEAN('s', "stat", &inherit_stat,
583 "per thread counts"),
584 OPT_BOOLEAN('d', "data", &sample_address,
586 OPT_BOOLEAN('n', "no-samples", &no_samples,
588 OPT_BOOLEAN('M', "multiplex", &multiplex,
589 "multiplex counter output in a single channel"),
593 int cmd_record(int argc, const char **argv, const char *prefix __used)
599 argc = parse_options(argc, argv, options, record_usage,
600 PARSE_OPT_STOP_AT_NON_OPTION);
601 if (!argc && target_pid == -1 && !system_wide)
602 usage_with_options(record_usage, options);
606 attrs[0].type = PERF_TYPE_HARDWARE;
607 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
611 * User specified count overrides default frequency.
613 if (default_interval)
616 default_interval = freq;
618 fprintf(stderr, "frequency and count are zero, aborting\n");
622 for (counter = 0; counter < nr_counters; counter++) {
623 if (attrs[counter].sample_period)
626 attrs[counter].sample_period = default_interval;
629 return __cmd_record(argc, argv);