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"
19 #include "util/header.h"
20 #include "util/event.h"
21 #include "util/debug.h"
22 #include "util/session.h"
23 #include "util/symbol.h"
24 #include "util/cpumap.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 target_tid = -1;
46 static pid_t *all_tids = NULL;
47 static int thread_num = 0;
48 static pid_t child_pid = -1;
49 static int inherit = 1;
51 static int append_file = 0;
52 static int call_graph = 0;
53 static int inherit_stat = 0;
54 static int no_samples = 0;
55 static int sample_address = 0;
56 static int multiplex = 0;
57 static int multiplex_fd = -1;
59 static long samples = 0;
60 static struct timeval last_read;
61 static struct timeval this_read;
63 static u64 bytes_written = 0;
65 static struct pollfd *event_array;
67 static int nr_poll = 0;
68 static int nr_cpu = 0;
70 static int file_new = 1;
71 static off_t post_processing_offset;
73 static struct perf_session *session;
82 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
84 static unsigned long mmap_read_head(struct mmap_data *md)
86 struct perf_event_mmap_page *pc = md->base;
95 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
97 struct perf_event_mmap_page *pc = md->base;
100 * ensure all reads are done before we write the tail out.
103 pc->data_tail = tail;
106 static void write_output(void *buf, size_t size)
109 int ret = write(output, buf, size);
112 die("failed to write");
117 bytes_written += ret;
121 static int process_synthesized_event(event_t *event,
122 struct perf_session *self __used)
124 write_output(event, event->header.size);
128 static void mmap_read(struct mmap_data *md)
130 unsigned int head = mmap_read_head(md);
131 unsigned int old = md->prev;
132 unsigned char *data = md->base + page_size;
137 gettimeofday(&this_read, NULL);
140 * If we're further behind than half the buffer, there's a chance
141 * the writer will bite our tail and mess up the samples under us.
143 * If we somehow ended up ahead of the head, we got messed up.
145 * In either case, truncate and restart at head.
152 timersub(&this_read, &last_read, &iv);
153 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
155 fprintf(stderr, "WARNING: failed to keep up with mmap data."
156 " Last read %lu msecs ago.\n", msecs);
159 * head points to a known good entry, start there.
164 last_read = this_read;
171 if ((old & md->mask) + size != (head & md->mask)) {
172 buf = &data[old & md->mask];
173 size = md->mask + 1 - (old & md->mask);
176 write_output(buf, size);
179 buf = &data[old & md->mask];
183 write_output(buf, size);
186 mmap_write_tail(md, old);
189 static volatile int done = 0;
190 static volatile int signr = -1;
192 static void sig_handler(int sig)
198 static void sig_atexit(void)
201 kill(child_pid, SIGTERM);
206 signal(signr, SIG_DFL);
207 kill(getpid(), signr);
212 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
214 struct perf_header_attr *h_attr;
216 if (nr < session->header.attrs) {
217 h_attr = session->header.attr[nr];
219 h_attr = perf_header_attr__new(a);
221 if (perf_header__add_attr(&session->header, h_attr) < 0) {
222 perf_header_attr__delete(h_attr);
230 static void create_counter(int counter, int cpu)
232 char *filter = filters[counter];
233 struct perf_event_attr *attr = attrs + counter;
234 struct perf_header_attr *h_attr;
235 int track = !counter; /* only the first counter needs these */
245 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
246 PERF_FORMAT_TOTAL_TIME_RUNNING |
249 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
252 attr->sample_type |= PERF_SAMPLE_ID;
255 attr->sample_type |= PERF_SAMPLE_PERIOD;
257 attr->sample_freq = freq;
261 attr->sample_freq = 0;
264 attr->inherit_stat = 1;
267 attr->sample_type |= PERF_SAMPLE_ADDR;
270 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
273 attr->sample_type |= PERF_SAMPLE_TIME;
274 attr->sample_type |= PERF_SAMPLE_RAW;
275 attr->sample_type |= PERF_SAMPLE_CPU;
280 attr->inherit = inherit;
281 if (target_pid == -1 && !system_wide) {
283 attr->enable_on_exec = 1;
286 for (thread_index = 0; thread_index < thread_num; thread_index++) {
288 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
289 all_tids[thread_index], cpu, group_fd, 0);
291 if (fd[nr_cpu][counter][thread_index] < 0) {
294 if (err == EPERM || err == EACCES)
295 die("Permission error - are you root?\n"
296 "\t Consider tweaking"
297 " /proc/sys/kernel/perf_event_paranoid.\n");
298 else if (err == ENODEV && profile_cpu != -1) {
299 die("No such device - did you specify"
300 " an out-of-range profile CPU?\n");
304 * If it's cycles then fall back to hrtimer
305 * based cpu-clock-tick sw counter, which
306 * is always available even if no PMU support:
308 if (attr->type == PERF_TYPE_HARDWARE
309 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
312 warning(" ... trying to fall back to cpu-clock-ticks\n");
313 attr->type = PERF_TYPE_SOFTWARE;
314 attr->config = PERF_COUNT_SW_CPU_CLOCK;
318 error("perfcounter syscall returned with %d (%s)\n",
319 fd[nr_cpu][counter][thread_index], strerror(err));
321 #if defined(__i386__) || defined(__x86_64__)
322 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
323 die("No hardware sampling interrupt available."
324 " No APIC? If so then you can boot the kernel"
325 " with the \"lapic\" boot parameter to"
326 " force-enable it.\n");
329 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
333 h_attr = get_header_attr(attr, counter);
338 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
339 fprintf(stderr, "incompatible append\n");
344 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
345 perror("Unable to read perf file descriptor\n");
349 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
350 pr_warning("Not enough memory to add id\n");
354 assert(fd[nr_cpu][counter][thread_index] >= 0);
355 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
358 * First counter acts as the group leader:
360 if (group && group_fd == -1)
361 group_fd = fd[nr_cpu][counter][thread_index];
362 if (multiplex && multiplex_fd == -1)
363 multiplex_fd = fd[nr_cpu][counter][thread_index];
365 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
367 ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
370 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
371 event_array[nr_poll].events = POLLIN;
374 mmap_array[nr_cpu][counter][thread_index].counter = counter;
375 mmap_array[nr_cpu][counter][thread_index].prev = 0;
376 mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
377 mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
378 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
379 if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
380 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
385 if (filter != NULL) {
386 ret = ioctl(fd[nr_cpu][counter][thread_index],
387 PERF_EVENT_IOC_SET_FILTER, filter);
389 error("failed to set filter with %d (%s)\n", errno,
397 static void open_counters(int cpu)
402 for (counter = 0; counter < nr_counters; counter++)
403 create_counter(counter, cpu);
408 static int process_buildids(void)
410 u64 size = lseek(output, 0, SEEK_CUR);
415 session->fd = output;
416 return __perf_session__process_events(session, post_processing_offset,
417 size - post_processing_offset,
418 size, &build_id__mark_dso_hit_ops);
421 static void atexit_header(void)
423 session->header.data_size += bytes_written;
426 perf_header__write(&session->header, output, true);
429 static int __cmd_record(int argc, const char **argv)
436 unsigned long waking = 0;
437 int child_ready_pipe[2], go_pipe[2];
438 const bool forks = argc > 0;
441 page_size = sysconf(_SC_PAGE_SIZE);
444 signal(SIGCHLD, sig_handler);
445 signal(SIGINT, sig_handler);
447 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
448 perror("failed to create pipes");
452 if (!stat(output_name, &st) && st.st_size) {
455 pr_err("Error, output file %s exists, use -A "
456 "to append or -f to overwrite.\n",
461 char oldname[PATH_MAX];
462 snprintf(oldname, sizeof(oldname), "%s.old",
465 rename(output_name, oldname);
471 flags = O_CREAT|O_RDWR;
477 output = open(output_name, flags, S_IRUSR|S_IWUSR);
479 perror("failed to create output file");
483 session = perf_session__new(output_name, O_WRONLY, force);
484 if (session == NULL) {
485 pr_err("Not enough memory for reading perf file header\n");
490 err = perf_header__read(&session->header, output);
496 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
498 for (i = 0; i < nr_counters; i++) {
499 if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
500 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
506 atexit(atexit_header);
511 perror("failed to fork");
516 close(child_ready_pipe[0]);
518 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
521 * Do a dummy execvp to get the PLT entry resolved,
522 * so we avoid the resolver overhead on the real
525 execvp("", (char **)argv);
528 * Tell the parent we're ready to go
530 close(child_ready_pipe[1]);
533 * Wait until the parent tells us to go.
535 if (read(go_pipe[0], &buf, 1) == -1)
536 perror("unable to read pipe");
538 execvp(argv[0], (char **)argv);
544 if (!system_wide && target_tid == -1 && target_pid == -1)
545 all_tids[0] = child_pid;
547 close(child_ready_pipe[1]);
550 * wait for child to settle
552 if (read(child_ready_pipe[0], &buf, 1) == -1) {
553 perror("unable to read pipe");
556 close(child_ready_pipe[0]);
559 if ((!system_wide && !inherit) || profile_cpu != -1) {
560 open_counters(profile_cpu);
562 nr_cpus = read_cpu_map();
563 for (i = 0; i < nr_cpus; i++)
564 open_counters(cpumap[i]);
568 err = perf_header__write(&session->header, output, false);
573 post_processing_offset = lseek(output, 0, SEEK_CUR);
575 err = event__synthesize_kernel_mmap(process_synthesized_event,
578 err = event__synthesize_kernel_mmap(process_synthesized_event,
581 pr_err("Couldn't record kernel reference relocation symbol.\n");
585 err = event__synthesize_modules(process_synthesized_event, session);
587 pr_err("Couldn't record kernel reference relocation symbol.\n");
591 if (!system_wide && profile_cpu == -1)
592 event__synthesize_thread(target_tid, process_synthesized_event,
595 event__synthesize_threads(process_synthesized_event, session);
598 struct sched_param param;
600 param.sched_priority = realtime_prio;
601 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
602 pr_err("Could not set realtime priority.\n");
617 for (i = 0; i < nr_cpu; i++) {
618 for (counter = 0; counter < nr_counters; counter++) {
620 thread < thread_num; thread++) {
621 if (mmap_array[i][counter][thread].base)
622 mmap_read(&mmap_array[i][counter][thread]);
628 if (hits == samples) {
631 err = poll(event_array, nr_poll, -1);
636 for (i = 0; i < nr_cpu; i++) {
638 counter < nr_counters;
643 ioctl(fd[i][counter][thread],
644 PERF_EVENT_IOC_DISABLE);
650 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
653 * Approximate RIP event size: 24 bytes.
656 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
657 (double)bytes_written / 1024.0 / 1024.0,
664 static const char * const record_usage[] = {
665 "perf record [<options>] [<command>]",
666 "perf record [<options>] -- <command> [<options>]",
670 static const struct option options[] = {
671 OPT_CALLBACK('e', "event", NULL, "event",
672 "event selector. use 'perf list' to list available events",
674 OPT_CALLBACK(0, "filter", NULL, "filter",
675 "event filter", parse_filter),
676 OPT_INTEGER('p', "pid", &target_pid,
677 "record events on existing process id"),
678 OPT_INTEGER('t', "tid", &target_tid,
679 "record events on existing thread id"),
680 OPT_INTEGER('r', "realtime", &realtime_prio,
681 "collect data with this RT SCHED_FIFO priority"),
682 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
683 "collect raw sample records from all opened counters"),
684 OPT_BOOLEAN('a', "all-cpus", &system_wide,
685 "system-wide collection from all CPUs"),
686 OPT_BOOLEAN('A', "append", &append_file,
687 "append to the output file to do incremental profiling"),
688 OPT_INTEGER('C', "profile_cpu", &profile_cpu,
689 "CPU to profile on"),
690 OPT_BOOLEAN('f', "force", &force,
691 "overwrite existing data file"),
692 OPT_LONG('c', "count", &default_interval,
693 "event period to sample"),
694 OPT_STRING('o', "output", &output_name, "file",
696 OPT_BOOLEAN('i', "inherit", &inherit,
697 "child tasks inherit counters"),
698 OPT_INTEGER('F', "freq", &freq,
699 "profile at this frequency"),
700 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
701 "number of mmap data pages"),
702 OPT_BOOLEAN('g', "call-graph", &call_graph,
703 "do call-graph (stack chain/backtrace) recording"),
704 OPT_BOOLEAN('v', "verbose", &verbose,
705 "be more verbose (show counter open errors, etc)"),
706 OPT_BOOLEAN('s', "stat", &inherit_stat,
707 "per thread counts"),
708 OPT_BOOLEAN('d', "data", &sample_address,
710 OPT_BOOLEAN('n', "no-samples", &no_samples,
712 OPT_BOOLEAN('M', "multiplex", &multiplex,
713 "multiplex counter output in a single channel"),
717 int cmd_record(int argc, const char **argv, const char *prefix __used)
722 argc = parse_options(argc, argv, options, record_usage,
723 PARSE_OPT_STOP_AT_NON_OPTION);
724 if (!argc && target_pid == -1 && target_tid == -1 &&
725 !system_wide && profile_cpu == -1)
726 usage_with_options(record_usage, options);
732 attrs[0].type = PERF_TYPE_HARDWARE;
733 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
736 if (target_pid != -1) {
737 target_tid = target_pid;
738 thread_num = find_all_tid(target_pid, &all_tids);
739 if (thread_num <= 0) {
740 fprintf(stderr, "Can't find all threads of pid %d\n",
742 usage_with_options(record_usage, options);
745 all_tids=malloc(sizeof(pid_t));
749 all_tids[0] = target_tid;
753 for (i = 0; i < MAX_NR_CPUS; i++) {
754 for (j = 0; j < MAX_COUNTERS; j++) {
755 fd[i][j] = malloc(sizeof(int)*thread_num);
756 mmap_array[i][j] = zalloc(
757 sizeof(struct mmap_data)*thread_num);
758 if (!fd[i][j] || !mmap_array[i][j])
762 event_array = malloc(
763 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
768 * User specified count overrides default frequency.
770 if (default_interval)
773 default_interval = freq;
775 fprintf(stderr, "frequency and count are zero, aborting\n");
779 for (counter = 0; counter < nr_counters; counter++) {
780 if (attrs[counter].sample_period)
783 attrs[counter].sample_period = default_interval;
786 return __cmd_record(argc, argv);