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"
18 #include "util/string.h"
20 #include "util/header.h"
21 #include "util/event.h"
22 #include "util/debug.h"
23 #include "util/session.h"
24 #include "util/symbol.h"
25 #include "util/cpumap.h"
30 static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
32 static long default_interval = 0;
34 static int nr_cpus = 0;
35 static unsigned int page_size;
36 static unsigned int mmap_pages = 128;
37 static int freq = 1000;
39 static const char *output_name = "perf.data";
41 static unsigned int realtime_prio = 0;
42 static int raw_samples = 0;
43 static int system_wide = 0;
44 static int profile_cpu = -1;
45 static pid_t target_pid = -1;
46 static pid_t target_tid = -1;
47 static pid_t *all_tids = NULL;
48 static int thread_num = 0;
49 static pid_t child_pid = -1;
50 static int inherit = 1;
52 static int append_file = 0;
53 static int call_graph = 0;
54 static int inherit_stat = 0;
55 static int no_samples = 0;
56 static int sample_address = 0;
57 static int multiplex = 0;
58 static int multiplex_fd = -1;
60 static long samples = 0;
61 static struct timeval last_read;
62 static struct timeval this_read;
64 static u64 bytes_written = 0;
66 static struct pollfd *event_array;
68 static int nr_poll = 0;
69 static int nr_cpu = 0;
71 static int file_new = 1;
72 static off_t post_processing_offset;
74 static struct perf_session *session;
83 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
85 static unsigned long mmap_read_head(struct mmap_data *md)
87 struct perf_event_mmap_page *pc = md->base;
96 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
98 struct perf_event_mmap_page *pc = md->base;
101 * ensure all reads are done before we write the tail out.
104 pc->data_tail = tail;
107 static void write_output(void *buf, size_t size)
110 int ret = write(output, buf, size);
113 die("failed to write");
118 bytes_written += ret;
122 static int process_synthesized_event(event_t *event,
123 struct perf_session *self __used)
125 write_output(event, event->header.size);
129 static void mmap_read(struct mmap_data *md)
131 unsigned int head = mmap_read_head(md);
132 unsigned int old = md->prev;
133 unsigned char *data = md->base + page_size;
138 gettimeofday(&this_read, NULL);
141 * If we're further behind than half the buffer, there's a chance
142 * the writer will bite our tail and mess up the samples under us.
144 * If we somehow ended up ahead of the head, we got messed up.
146 * In either case, truncate and restart at head.
153 timersub(&this_read, &last_read, &iv);
154 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
156 fprintf(stderr, "WARNING: failed to keep up with mmap data."
157 " Last read %lu msecs ago.\n", msecs);
160 * head points to a known good entry, start there.
165 last_read = this_read;
172 if ((old & md->mask) + size != (head & md->mask)) {
173 buf = &data[old & md->mask];
174 size = md->mask + 1 - (old & md->mask);
177 write_output(buf, size);
180 buf = &data[old & md->mask];
184 write_output(buf, size);
187 mmap_write_tail(md, old);
190 static volatile int done = 0;
191 static volatile int signr = -1;
193 static void sig_handler(int sig)
199 static void sig_atexit(void)
202 kill(child_pid, SIGTERM);
207 signal(signr, SIG_DFL);
208 kill(getpid(), signr);
213 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
215 struct perf_header_attr *h_attr;
217 if (nr < session->header.attrs) {
218 h_attr = session->header.attr[nr];
220 h_attr = perf_header_attr__new(a);
222 if (perf_header__add_attr(&session->header, h_attr) < 0) {
223 perf_header_attr__delete(h_attr);
231 static void create_counter(int counter, int cpu)
233 char *filter = filters[counter];
234 struct perf_event_attr *attr = attrs + counter;
235 struct perf_header_attr *h_attr;
236 int track = !counter; /* only the first counter needs these */
246 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
247 PERF_FORMAT_TOTAL_TIME_RUNNING |
250 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
253 attr->sample_type |= PERF_SAMPLE_ID;
256 attr->sample_type |= PERF_SAMPLE_PERIOD;
258 attr->sample_freq = freq;
262 attr->sample_freq = 0;
265 attr->inherit_stat = 1;
268 attr->sample_type |= PERF_SAMPLE_ADDR;
271 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
274 attr->sample_type |= PERF_SAMPLE_TIME;
275 attr->sample_type |= PERF_SAMPLE_RAW;
276 attr->sample_type |= PERF_SAMPLE_CPU;
281 attr->inherit = inherit;
282 if (target_pid == -1 && !system_wide) {
284 attr->enable_on_exec = 1;
287 for (thread_index = 0; thread_index < thread_num; thread_index++) {
289 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
290 all_tids[thread_index], cpu, group_fd, 0);
292 if (fd[nr_cpu][counter][thread_index] < 0) {
295 if (err == EPERM || err == EACCES)
296 die("Permission error - are you root?\n"
297 "\t Consider tweaking"
298 " /proc/sys/kernel/perf_event_paranoid.\n");
299 else if (err == ENODEV && profile_cpu != -1) {
300 die("No such device - did you specify"
301 " an out-of-range profile CPU?\n");
305 * If it's cycles then fall back to hrtimer
306 * based cpu-clock-tick sw counter, which
307 * is always available even if no PMU support:
309 if (attr->type == PERF_TYPE_HARDWARE
310 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
313 warning(" ... trying to fall back to cpu-clock-ticks\n");
314 attr->type = PERF_TYPE_SOFTWARE;
315 attr->config = PERF_COUNT_SW_CPU_CLOCK;
319 error("perfcounter syscall returned with %d (%s)\n",
320 fd[nr_cpu][counter][thread_index], strerror(err));
322 #if defined(__i386__) || defined(__x86_64__)
323 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
324 die("No hardware sampling interrupt available."
325 " No APIC? If so then you can boot the kernel"
326 " with the \"lapic\" boot parameter to"
327 " force-enable it.\n");
330 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
334 h_attr = get_header_attr(attr, counter);
339 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
340 fprintf(stderr, "incompatible append\n");
345 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
346 perror("Unable to read perf file descriptor\n");
350 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
351 pr_warning("Not enough memory to add id\n");
355 assert(fd[nr_cpu][counter][thread_index] >= 0);
356 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
359 * First counter acts as the group leader:
361 if (group && group_fd == -1)
362 group_fd = fd[nr_cpu][counter][thread_index];
363 if (multiplex && multiplex_fd == -1)
364 multiplex_fd = fd[nr_cpu][counter][thread_index];
366 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
368 ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
371 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
372 event_array[nr_poll].events = POLLIN;
375 mmap_array[nr_cpu][counter][thread_index].counter = counter;
376 mmap_array[nr_cpu][counter][thread_index].prev = 0;
377 mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
378 mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
379 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
380 if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
381 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
386 if (filter != NULL) {
387 ret = ioctl(fd[nr_cpu][counter][thread_index],
388 PERF_EVENT_IOC_SET_FILTER, filter);
390 error("failed to set filter with %d (%s)\n", errno,
398 static void open_counters(int cpu)
403 for (counter = 0; counter < nr_counters; counter++)
404 create_counter(counter, cpu);
409 static int process_buildids(void)
411 u64 size = lseek(output, 0, SEEK_CUR);
416 session->fd = output;
417 return __perf_session__process_events(session, post_processing_offset,
418 size - post_processing_offset,
419 size, &build_id__mark_dso_hit_ops);
422 static void atexit_header(void)
424 session->header.data_size += bytes_written;
427 perf_header__write(&session->header, output, true);
430 static int __cmd_record(int argc, const char **argv)
437 unsigned long waking = 0;
438 int child_ready_pipe[2], go_pipe[2];
439 const bool forks = argc > 0;
442 page_size = sysconf(_SC_PAGE_SIZE);
445 signal(SIGCHLD, sig_handler);
446 signal(SIGINT, sig_handler);
448 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
449 perror("failed to create pipes");
453 if (!stat(output_name, &st) && st.st_size) {
456 pr_err("Error, output file %s exists, use -A "
457 "to append or -f to overwrite.\n",
462 char oldname[PATH_MAX];
463 snprintf(oldname, sizeof(oldname), "%s.old",
466 rename(output_name, oldname);
472 flags = O_CREAT|O_RDWR;
478 output = open(output_name, flags, S_IRUSR|S_IWUSR);
480 perror("failed to create output file");
484 session = perf_session__new(output_name, O_WRONLY, force);
485 if (session == NULL) {
486 pr_err("Not enough memory for reading perf file header\n");
491 err = perf_header__read(&session->header, output);
497 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
499 for (i = 0; i < nr_counters; i++) {
500 if (attrs[i].sample_type & PERF_SAMPLE_RAW) {
501 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
507 atexit(atexit_header);
512 perror("failed to fork");
517 close(child_ready_pipe[0]);
519 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
522 * Do a dummy execvp to get the PLT entry resolved,
523 * so we avoid the resolver overhead on the real
526 execvp("", (char **)argv);
529 * Tell the parent we're ready to go
531 close(child_ready_pipe[1]);
534 * Wait until the parent tells us to go.
536 if (read(go_pipe[0], &buf, 1) == -1)
537 perror("unable to read pipe");
539 execvp(argv[0], (char **)argv);
545 if (!system_wide && target_tid == -1 && target_pid == -1)
546 all_tids[0] = child_pid;
548 close(child_ready_pipe[1]);
551 * wait for child to settle
553 if (read(child_ready_pipe[0], &buf, 1) == -1) {
554 perror("unable to read pipe");
557 close(child_ready_pipe[0]);
560 if ((!system_wide && !inherit) || profile_cpu != -1) {
561 open_counters(profile_cpu);
563 nr_cpus = read_cpu_map();
564 for (i = 0; i < nr_cpus; i++)
565 open_counters(cpumap[i]);
569 err = perf_header__write(&session->header, output, false);
574 post_processing_offset = lseek(output, 0, SEEK_CUR);
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 const struct option options[] = {
669 OPT_CALLBACK('e', "event", NULL, "event",
670 "event selector. use 'perf list' to list available events",
672 OPT_CALLBACK(0, "filter", NULL, "filter",
673 "event filter", parse_filter),
674 OPT_INTEGER('p', "pid", &target_pid,
675 "record events on existing process id"),
676 OPT_INTEGER('t', "tid", &target_tid,
677 "record events on existing thread id"),
678 OPT_INTEGER('r', "realtime", &realtime_prio,
679 "collect data with this RT SCHED_FIFO priority"),
680 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
681 "collect raw sample records from all opened counters"),
682 OPT_BOOLEAN('a', "all-cpus", &system_wide,
683 "system-wide collection from all CPUs"),
684 OPT_BOOLEAN('A', "append", &append_file,
685 "append to the output file to do incremental profiling"),
686 OPT_INTEGER('C', "profile_cpu", &profile_cpu,
687 "CPU to profile on"),
688 OPT_BOOLEAN('f', "force", &force,
689 "overwrite existing data file"),
690 OPT_LONG('c', "count", &default_interval,
691 "event period to sample"),
692 OPT_STRING('o', "output", &output_name, "file",
694 OPT_BOOLEAN('i', "inherit", &inherit,
695 "child tasks inherit counters"),
696 OPT_INTEGER('F', "freq", &freq,
697 "profile at this frequency"),
698 OPT_INTEGER('m', "mmap-pages", &mmap_pages,
699 "number of mmap data pages"),
700 OPT_BOOLEAN('g', "call-graph", &call_graph,
701 "do call-graph (stack chain/backtrace) recording"),
702 OPT_BOOLEAN('v', "verbose", &verbose,
703 "be more verbose (show counter open errors, etc)"),
704 OPT_BOOLEAN('s', "stat", &inherit_stat,
705 "per thread counts"),
706 OPT_BOOLEAN('d', "data", &sample_address,
708 OPT_BOOLEAN('n', "no-samples", &no_samples,
710 OPT_BOOLEAN('M', "multiplex", &multiplex,
711 "multiplex counter output in a single channel"),
715 int cmd_record(int argc, const char **argv, const char *prefix __used)
720 argc = parse_options(argc, argv, options, record_usage,
721 PARSE_OPT_STOP_AT_NON_OPTION);
722 if (!argc && target_pid == -1 && target_tid == -1 &&
723 !system_wide && profile_cpu == -1)
724 usage_with_options(record_usage, options);
730 attrs[0].type = PERF_TYPE_HARDWARE;
731 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
734 if (target_pid != -1) {
735 target_tid = target_pid;
736 thread_num = find_all_tid(target_pid, &all_tids);
737 if (thread_num <= 0) {
738 fprintf(stderr, "Can't find all threads of pid %d\n",
740 usage_with_options(record_usage, options);
743 all_tids=malloc(sizeof(pid_t));
747 all_tids[0] = target_tid;
751 for (i = 0; i < MAX_NR_CPUS; i++) {
752 for (j = 0; j < MAX_COUNTERS; j++) {
753 fd[i][j] = malloc(sizeof(int)*thread_num);
754 mmap_array[i][j] = malloc(
755 sizeof(struct mmap_data)*thread_num);
756 if (!fd[i][j] || !mmap_array[i][j])
760 event_array = malloc(
761 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
766 * User specified count overrides default frequency.
768 if (default_interval)
771 default_interval = freq;
773 fprintf(stderr, "frequency and count are zero, aborting\n");
777 for (counter = 0; counter < nr_counters; counter++) {
778 if (attrs[counter].sample_period)
781 attrs[counter].sample_period = default_interval;
784 return __cmd_record(argc, argv);