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"
34 static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
36 static long default_interval = 0;
38 static int nr_cpus = 0;
39 static unsigned int page_size;
40 static unsigned int mmap_pages = 128;
41 static int freq = 1000;
43 static const char *output_name = "perf.data";
45 static unsigned int realtime_prio = 0;
46 static bool raw_samples = false;
47 static bool system_wide = false;
48 static int profile_cpu = -1;
49 static pid_t target_pid = -1;
50 static pid_t target_tid = -1;
51 static pid_t *all_tids = NULL;
52 static int thread_num = 0;
53 static pid_t child_pid = -1;
54 static bool inherit = true;
55 static enum write_mode_t write_mode = WRITE_FORCE;
56 static bool call_graph = false;
57 static bool inherit_stat = false;
58 static bool no_samples = false;
59 static bool sample_address = false;
60 static bool multiplex = false;
61 static int multiplex_fd = -1;
63 static long samples = 0;
64 static struct timeval last_read;
65 static struct timeval this_read;
67 static u64 bytes_written = 0;
69 static struct pollfd *event_array;
71 static int nr_poll = 0;
72 static int nr_cpu = 0;
74 static int file_new = 1;
75 static off_t post_processing_offset;
77 static struct perf_session *session;
86 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
88 static unsigned long mmap_read_head(struct mmap_data *md)
90 struct perf_event_mmap_page *pc = md->base;
99 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
101 struct perf_event_mmap_page *pc = md->base;
104 * ensure all reads are done before we write the tail out.
107 pc->data_tail = tail;
110 static void write_output(void *buf, size_t size)
113 int ret = write(output, buf, size);
116 die("failed to write");
121 bytes_written += ret;
125 static int process_synthesized_event(event_t *event,
126 struct perf_session *self __used)
128 write_output(event, event->header.size);
132 static void mmap_read(struct mmap_data *md)
134 unsigned int head = mmap_read_head(md);
135 unsigned int old = md->prev;
136 unsigned char *data = md->base + page_size;
141 gettimeofday(&this_read, NULL);
144 * If we're further behind than half the buffer, there's a chance
145 * the writer will bite our tail and mess up the samples under us.
147 * If we somehow ended up ahead of the head, we got messed up.
149 * In either case, truncate and restart at head.
156 timersub(&this_read, &last_read, &iv);
157 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
159 fprintf(stderr, "WARNING: failed to keep up with mmap data."
160 " Last read %lu msecs ago.\n", msecs);
163 * head points to a known good entry, start there.
168 last_read = this_read;
175 if ((old & md->mask) + size != (head & md->mask)) {
176 buf = &data[old & md->mask];
177 size = md->mask + 1 - (old & md->mask);
180 write_output(buf, size);
183 buf = &data[old & md->mask];
187 write_output(buf, size);
190 mmap_write_tail(md, old);
193 static volatile int done = 0;
194 static volatile int signr = -1;
196 static void sig_handler(int sig)
202 static void sig_atexit(void)
205 kill(child_pid, SIGTERM);
210 signal(signr, SIG_DFL);
211 kill(getpid(), signr);
216 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
218 struct perf_header_attr *h_attr;
220 if (nr < session->header.attrs) {
221 h_attr = session->header.attr[nr];
223 h_attr = perf_header_attr__new(a);
225 if (perf_header__add_attr(&session->header, h_attr) < 0) {
226 perf_header_attr__delete(h_attr);
234 static void create_counter(int counter, int cpu)
236 char *filter = filters[counter];
237 struct perf_event_attr *attr = attrs + counter;
238 struct perf_header_attr *h_attr;
239 int track = !counter; /* only the first counter needs these */
249 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
250 PERF_FORMAT_TOTAL_TIME_RUNNING |
253 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
256 attr->sample_type |= PERF_SAMPLE_ID;
259 attr->sample_type |= PERF_SAMPLE_PERIOD;
261 attr->sample_freq = freq;
265 attr->sample_freq = 0;
268 attr->inherit_stat = 1;
271 attr->sample_type |= PERF_SAMPLE_ADDR;
274 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
277 attr->sample_type |= PERF_SAMPLE_TIME;
278 attr->sample_type |= PERF_SAMPLE_RAW;
279 attr->sample_type |= PERF_SAMPLE_CPU;
284 attr->inherit = inherit;
285 if (target_pid == -1 && !system_wide) {
287 attr->enable_on_exec = 1;
290 for (thread_index = 0; thread_index < thread_num; thread_index++) {
292 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
293 all_tids[thread_index], cpu, group_fd, 0);
295 if (fd[nr_cpu][counter][thread_index] < 0) {
298 if (err == EPERM || err == EACCES)
299 die("Permission error - are you root?\n"
300 "\t Consider tweaking"
301 " /proc/sys/kernel/perf_event_paranoid.\n");
302 else if (err == ENODEV && profile_cpu != -1) {
303 die("No such device - did you specify"
304 " an out-of-range profile CPU?\n");
308 * If it's cycles then fall back to hrtimer
309 * based cpu-clock-tick sw counter, which
310 * is always available even if no PMU support:
312 if (attr->type == PERF_TYPE_HARDWARE
313 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
316 warning(" ... trying to fall back to cpu-clock-ticks\n");
317 attr->type = PERF_TYPE_SOFTWARE;
318 attr->config = PERF_COUNT_SW_CPU_CLOCK;
322 error("perfcounter syscall returned with %d (%s)\n",
323 fd[nr_cpu][counter][thread_index], strerror(err));
325 #if defined(__i386__) || defined(__x86_64__)
326 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
327 die("No hardware sampling interrupt available."
328 " No APIC? If so then you can boot the kernel"
329 " with the \"lapic\" boot parameter to"
330 " force-enable it.\n");
333 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
337 h_attr = get_header_attr(attr, counter);
342 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
343 fprintf(stderr, "incompatible append\n");
348 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
349 perror("Unable to read perf file descriptor\n");
353 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
354 pr_warning("Not enough memory to add id\n");
358 assert(fd[nr_cpu][counter][thread_index] >= 0);
359 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
362 * First counter acts as the group leader:
364 if (group && group_fd == -1)
365 group_fd = fd[nr_cpu][counter][thread_index];
366 if (multiplex && multiplex_fd == -1)
367 multiplex_fd = fd[nr_cpu][counter][thread_index];
369 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
371 ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
374 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
375 event_array[nr_poll].events = POLLIN;
378 mmap_array[nr_cpu][counter][thread_index].counter = counter;
379 mmap_array[nr_cpu][counter][thread_index].prev = 0;
380 mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
381 mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
382 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
383 if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
384 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
389 if (filter != NULL) {
390 ret = ioctl(fd[nr_cpu][counter][thread_index],
391 PERF_EVENT_IOC_SET_FILTER, filter);
393 error("failed to set filter with %d (%s)\n", errno,
401 static void open_counters(int cpu)
406 for (counter = 0; counter < nr_counters; counter++)
407 create_counter(counter, cpu);
412 static int process_buildids(void)
414 u64 size = lseek(output, 0, SEEK_CUR);
419 session->fd = output;
420 return __perf_session__process_events(session, post_processing_offset,
421 size - post_processing_offset,
422 size, &build_id__mark_dso_hit_ops);
425 static void atexit_header(void)
427 session->header.data_size += bytes_written;
430 perf_header__write(&session->header, output, true);
433 static int __cmd_record(int argc, const char **argv)
440 unsigned long waking = 0;
441 int child_ready_pipe[2], go_pipe[2];
442 const bool forks = argc > 0;
445 page_size = sysconf(_SC_PAGE_SIZE);
448 signal(SIGCHLD, sig_handler);
449 signal(SIGINT, sig_handler);
451 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
452 perror("failed to create pipes");
456 if (!stat(output_name, &st) && st.st_size) {
457 if (write_mode == WRITE_FORCE) {
458 char oldname[PATH_MAX];
459 snprintf(oldname, sizeof(oldname), "%s.old",
462 rename(output_name, oldname);
464 } else if (write_mode == WRITE_APPEND) {
465 write_mode = WRITE_FORCE;
468 flags = O_CREAT|O_RDWR;
469 if (write_mode == WRITE_APPEND)
474 output = open(output_name, flags, S_IRUSR|S_IWUSR);
476 perror("failed to create output file");
480 session = perf_session__new(output_name, O_WRONLY,
481 write_mode == WRITE_FORCE);
482 if (session == NULL) {
483 pr_err("Not enough memory for reading perf file header\n");
488 err = perf_header__read(&session->header, output);
494 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
496 for (i = 0; i < nr_counters; i++) {
497 if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
498 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
504 atexit(atexit_header);
509 perror("failed to fork");
514 close(child_ready_pipe[0]);
516 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
519 * Do a dummy execvp to get the PLT entry resolved,
520 * so we avoid the resolver overhead on the real
523 execvp("", (char **)argv);
526 * Tell the parent we're ready to go
528 close(child_ready_pipe[1]);
531 * Wait until the parent tells us to go.
533 if (read(go_pipe[0], &buf, 1) == -1)
534 perror("unable to read pipe");
536 execvp(argv[0], (char **)argv);
542 if (!system_wide && target_tid == -1 && target_pid == -1)
543 all_tids[0] = child_pid;
545 close(child_ready_pipe[1]);
548 * wait for child to settle
550 if (read(child_ready_pipe[0], &buf, 1) == -1) {
551 perror("unable to read pipe");
554 close(child_ready_pipe[0]);
557 if ((!system_wide && !inherit) || profile_cpu != -1) {
558 open_counters(profile_cpu);
560 nr_cpus = read_cpu_map();
561 for (i = 0; i < nr_cpus; i++)
562 open_counters(cpumap[i]);
566 err = perf_header__write(&session->header, output, false);
571 post_processing_offset = lseek(output, 0, SEEK_CUR);
573 err = event__synthesize_kernel_mmap(process_synthesized_event,
576 err = event__synthesize_kernel_mmap(process_synthesized_event,
579 pr_err("Couldn't record kernel reference relocation symbol.\n");
583 err = event__synthesize_modules(process_synthesized_event, session);
585 pr_err("Couldn't record kernel reference relocation symbol.\n");
589 if (!system_wide && profile_cpu == -1)
590 event__synthesize_thread(target_tid, process_synthesized_event,
593 event__synthesize_threads(process_synthesized_event, session);
596 struct sched_param param;
598 param.sched_priority = realtime_prio;
599 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
600 pr_err("Could not set realtime priority.\n");
615 for (i = 0; i < nr_cpu; i++) {
616 for (counter = 0; counter < nr_counters; counter++) {
618 thread < thread_num; thread++) {
619 if (mmap_array[i][counter][thread].base)
620 mmap_read(&mmap_array[i][counter][thread]);
626 if (hits == samples) {
629 err = poll(event_array, nr_poll, -1);
634 for (i = 0; i < nr_cpu; i++) {
636 counter < nr_counters;
641 ioctl(fd[i][counter][thread],
642 PERF_EVENT_IOC_DISABLE);
648 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
651 * Approximate RIP event size: 24 bytes.
654 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
655 (double)bytes_written / 1024.0 / 1024.0,
662 static const char * const record_usage[] = {
663 "perf record [<options>] [<command>]",
664 "perf record [<options>] -- <command> [<options>]",
668 static bool force, append_file;
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 (deprecated)"),
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_INCR('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);
728 if (force && append_file) {
729 fprintf(stderr, "Can't overwrite and append at the same time."
730 " You need to choose between -f and -A");
731 usage_with_options(record_usage, options);
732 } else if (append_file) {
733 write_mode = WRITE_APPEND;
735 write_mode = WRITE_FORCE;
742 attrs[0].type = PERF_TYPE_HARDWARE;
743 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
746 if (target_pid != -1) {
747 target_tid = target_pid;
748 thread_num = find_all_tid(target_pid, &all_tids);
749 if (thread_num <= 0) {
750 fprintf(stderr, "Can't find all threads of pid %d\n",
752 usage_with_options(record_usage, options);
755 all_tids=malloc(sizeof(pid_t));
759 all_tids[0] = target_tid;
763 for (i = 0; i < MAX_NR_CPUS; i++) {
764 for (j = 0; j < MAX_COUNTERS; j++) {
765 fd[i][j] = malloc(sizeof(int)*thread_num);
766 mmap_array[i][j] = zalloc(
767 sizeof(struct mmap_data)*thread_num);
768 if (!fd[i][j] || !mmap_array[i][j])
772 event_array = malloc(
773 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
778 * User specified count overrides default frequency.
780 if (default_interval)
783 default_interval = freq;
785 fprintf(stderr, "frequency and count are zero, aborting\n");
789 for (counter = 0; counter < nr_counters; counter++) {
790 if (attrs[counter].sample_period)
793 attrs[counter].sample_period = default_interval;
796 return __cmd_record(argc, argv);