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"
35 static int *fd[MAX_NR_CPUS][MAX_COUNTERS];
37 static u64 user_interval = ULLONG_MAX;
38 static u64 default_interval = 0;
39 static u64 sample_type;
41 static int nr_cpus = 0;
42 static unsigned int page_size;
43 static unsigned int mmap_pages = 128;
44 static unsigned int user_freq = UINT_MAX;
45 static int freq = 1000;
47 static int pipe_output = 0;
48 static const char *output_name = "perf.data";
50 static int realtime_prio = 0;
51 static bool raw_samples = false;
52 static bool sample_id_all_avail = true;
53 static bool system_wide = false;
54 static pid_t target_pid = -1;
55 static pid_t target_tid = -1;
56 static pid_t *all_tids = NULL;
57 static int thread_num = 0;
58 static pid_t child_pid = -1;
59 static bool no_inherit = false;
60 static enum write_mode_t write_mode = WRITE_FORCE;
61 static bool call_graph = false;
62 static bool inherit_stat = false;
63 static bool no_samples = false;
64 static bool sample_address = false;
65 static bool sample_time = false;
66 static bool no_buildid = false;
67 static bool no_buildid_cache = false;
69 static long samples = 0;
70 static u64 bytes_written = 0;
72 static struct pollfd *event_array;
74 static int nr_poll = 0;
75 static int nr_cpu = 0;
77 static int file_new = 1;
78 static off_t post_processing_offset;
80 static struct perf_session *session;
81 static const char *cpu_list;
90 static struct mmap_data mmap_array[MAX_NR_CPUS];
92 static unsigned long mmap_read_head(struct mmap_data *md)
94 struct perf_event_mmap_page *pc = md->base;
103 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
105 struct perf_event_mmap_page *pc = md->base;
108 * ensure all reads are done before we write the tail out.
111 pc->data_tail = tail;
114 static void advance_output(size_t size)
116 bytes_written += size;
119 static void write_output(void *buf, size_t size)
122 int ret = write(output, buf, size);
125 die("failed to write");
130 bytes_written += ret;
134 static int process_synthesized_event(event_t *event,
135 struct sample_data *sample __used,
136 struct perf_session *self __used)
138 write_output(event, event->header.size);
142 static void mmap_read(struct mmap_data *md)
144 unsigned int head = mmap_read_head(md);
145 unsigned int old = md->prev;
146 unsigned char *data = md->base + page_size;
152 * If we're further behind than half the buffer, there's a chance
153 * the writer will bite our tail and mess up the samples under us.
155 * If we somehow ended up ahead of the head, we got messed up.
157 * In either case, truncate and restart at head.
161 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
163 * head points to a known good entry, start there.
173 if ((old & md->mask) + size != (head & md->mask)) {
174 buf = &data[old & md->mask];
175 size = md->mask + 1 - (old & md->mask);
178 write_output(buf, size);
181 buf = &data[old & md->mask];
185 write_output(buf, size);
188 mmap_write_tail(md, old);
191 static volatile int done = 0;
192 static volatile int signr = -1;
194 static void sig_handler(int sig)
200 static void sig_atexit(void)
203 kill(child_pid, SIGTERM);
205 if (signr == -1 || signr == SIGUSR1)
208 signal(signr, SIG_DFL);
209 kill(getpid(), signr);
214 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
216 struct perf_header_attr *h_attr;
218 if (nr < session->header.attrs) {
219 h_attr = session->header.attr[nr];
221 h_attr = perf_header_attr__new(a);
223 if (perf_header__add_attr(&session->header, h_attr) < 0) {
224 perf_header_attr__delete(h_attr);
232 static void create_counter(int counter, int cpu)
234 char *filter = filters[counter];
235 struct perf_event_attr *attr = attrs + counter;
236 struct perf_header_attr *h_attr;
237 int track = !counter; /* only the first counter needs these */
247 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
248 PERF_FORMAT_TOTAL_TIME_RUNNING |
251 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
254 attr->sample_type |= PERF_SAMPLE_ID;
257 * We default some events to a 1 default interval. But keep
258 * it a weak assumption overridable by the user.
260 if (!attr->sample_period || (user_freq != UINT_MAX &&
261 user_interval != ULLONG_MAX)) {
263 attr->sample_type |= PERF_SAMPLE_PERIOD;
265 attr->sample_freq = freq;
267 attr->sample_period = default_interval;
272 attr->sample_freq = 0;
275 attr->inherit_stat = 1;
277 if (sample_address) {
278 attr->sample_type |= PERF_SAMPLE_ADDR;
279 attr->mmap_data = track;
283 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
286 attr->sample_type |= PERF_SAMPLE_CPU;
288 if (sample_time || system_wide || !no_inherit || cpu_list)
289 attr->sample_type |= PERF_SAMPLE_TIME;
292 attr->sample_type |= PERF_SAMPLE_TIME;
293 attr->sample_type |= PERF_SAMPLE_RAW;
294 attr->sample_type |= PERF_SAMPLE_CPU;
298 sample_type = attr->sample_type;
302 attr->inherit = !no_inherit;
303 if (target_pid == -1 && target_tid == -1 && !system_wide) {
305 attr->enable_on_exec = 1;
308 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
310 for (thread_index = 0; thread_index < thread_num; thread_index++) {
312 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
313 all_tids[thread_index], cpu, group_fd, 0);
315 if (fd[nr_cpu][counter][thread_index] < 0) {
318 if (err == EPERM || err == EACCES)
319 die("Permission error - are you root?\n"
320 "\t Consider tweaking"
321 " /proc/sys/kernel/perf_event_paranoid.\n");
322 else if (err == ENODEV && cpu_list) {
323 die("No such device - did you specify"
324 " an out-of-range profile CPU?\n");
325 } else if (err == EINVAL && sample_id_all_avail) {
327 * Old kernel, no attr->sample_id_type_all field
329 sample_id_all_avail = false;
330 if (!sample_time && !raw_samples)
331 attr->sample_type &= ~PERF_SAMPLE_TIME;
333 goto retry_sample_id;
337 * If it's cycles then fall back to hrtimer
338 * based cpu-clock-tick sw counter, which
339 * is always available even if no PMU support:
341 if (attr->type == PERF_TYPE_HARDWARE
342 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
345 warning(" ... trying to fall back to cpu-clock-ticks\n");
346 attr->type = PERF_TYPE_SOFTWARE;
347 attr->config = PERF_COUNT_SW_CPU_CLOCK;
351 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
352 fd[nr_cpu][counter][thread_index], strerror(err));
354 #if defined(__i386__) || defined(__x86_64__)
355 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
356 die("No hardware sampling interrupt available."
357 " No APIC? If so then you can boot the kernel"
358 " with the \"lapic\" boot parameter to"
359 " force-enable it.\n");
362 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
366 h_attr = get_header_attr(attr, counter);
371 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
372 fprintf(stderr, "incompatible append\n");
377 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
378 perror("Unable to read perf file descriptor");
382 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
383 pr_warning("Not enough memory to add id\n");
387 assert(fd[nr_cpu][counter][thread_index] >= 0);
388 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
391 * First counter acts as the group leader:
393 if (group && group_fd == -1)
394 group_fd = fd[nr_cpu][counter][thread_index];
396 if (counter || thread_index) {
397 ret = ioctl(fd[nr_cpu][counter][thread_index],
398 PERF_EVENT_IOC_SET_OUTPUT,
401 error("failed to set output: %d (%s)\n", errno,
406 mmap_array[nr_cpu].counter = counter;
407 mmap_array[nr_cpu].prev = 0;
408 mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
409 mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
410 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
411 if (mmap_array[nr_cpu].base == MAP_FAILED) {
412 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
416 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
417 event_array[nr_poll].events = POLLIN;
421 if (filter != NULL) {
422 ret = ioctl(fd[nr_cpu][counter][thread_index],
423 PERF_EVENT_IOC_SET_FILTER, filter);
425 error("failed to set filter with %d (%s)\n", errno,
433 static void open_counters(int cpu)
438 for (counter = 0; counter < nr_counters; counter++)
439 create_counter(counter, cpu);
444 static int process_buildids(void)
446 u64 size = lseek(output, 0, SEEK_CUR);
451 session->fd = output;
452 return __perf_session__process_events(session, post_processing_offset,
453 size - post_processing_offset,
454 size, &build_id__mark_dso_hit_ops);
457 static void atexit_header(void)
460 session->header.data_size += bytes_written;
464 perf_header__write(&session->header, output, true);
465 perf_session__delete(session);
470 static void event__synthesize_guest_os(struct machine *machine, void *data)
473 struct perf_session *psession = data;
475 if (machine__is_host(machine))
479 *As for guest kernel when processing subcommand record&report,
480 *we arrange module mmap prior to guest kernel mmap and trigger
481 *a preload dso because default guest module symbols are loaded
482 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
483 *method is used to avoid symbol missing when the first addr is
484 *in module instead of in guest kernel.
486 err = event__synthesize_modules(process_synthesized_event,
489 pr_err("Couldn't record guest kernel [%d]'s reference"
490 " relocation symbol.\n", machine->pid);
493 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
494 * have no _text sometimes.
496 err = event__synthesize_kernel_mmap(process_synthesized_event,
497 psession, machine, "_text");
499 err = event__synthesize_kernel_mmap(process_synthesized_event,
500 psession, machine, "_stext");
502 pr_err("Couldn't record guest kernel [%d]'s reference"
503 " relocation symbol.\n", machine->pid);
506 static struct perf_event_header finished_round_event = {
507 .size = sizeof(struct perf_event_header),
508 .type = PERF_RECORD_FINISHED_ROUND,
511 static void mmap_read_all(void)
515 for (i = 0; i < nr_cpu; i++) {
516 if (mmap_array[i].base)
517 mmap_read(&mmap_array[i]);
520 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
521 write_output(&finished_round_event, sizeof(finished_round_event));
524 static int __cmd_record(int argc, const char **argv)
530 unsigned long waking = 0;
531 int child_ready_pipe[2], go_pipe[2];
532 const bool forks = argc > 0;
534 struct machine *machine;
536 page_size = sysconf(_SC_PAGE_SIZE);
539 signal(SIGCHLD, sig_handler);
540 signal(SIGINT, sig_handler);
541 signal(SIGUSR1, sig_handler);
543 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
544 perror("failed to create pipes");
548 if (!strcmp(output_name, "-"))
550 else if (!stat(output_name, &st) && st.st_size) {
551 if (write_mode == WRITE_FORCE) {
552 char oldname[PATH_MAX];
553 snprintf(oldname, sizeof(oldname), "%s.old",
556 rename(output_name, oldname);
558 } else if (write_mode == WRITE_APPEND) {
559 write_mode = WRITE_FORCE;
562 flags = O_CREAT|O_RDWR;
563 if (write_mode == WRITE_APPEND)
569 output = STDOUT_FILENO;
571 output = open(output_name, flags, S_IRUSR | S_IWUSR);
573 perror("failed to create output file");
577 session = perf_session__new(output_name, O_WRONLY,
578 write_mode == WRITE_FORCE, false, NULL);
579 if (session == NULL) {
580 pr_err("Not enough memory for reading perf file header\n");
585 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
588 err = perf_header__read(session, output);
590 goto out_delete_session;
593 if (have_tracepoints(attrs, nr_counters))
594 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
597 * perf_session__delete(session) will be called at atexit_header()
599 atexit(atexit_header);
604 perror("failed to fork");
611 close(child_ready_pipe[0]);
613 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
616 * Do a dummy execvp to get the PLT entry resolved,
617 * so we avoid the resolver overhead on the real
620 execvp("", (char **)argv);
623 * Tell the parent we're ready to go
625 close(child_ready_pipe[1]);
628 * Wait until the parent tells us to go.
630 if (read(go_pipe[0], &buf, 1) == -1)
631 perror("unable to read pipe");
633 execvp(argv[0], (char **)argv);
636 kill(getppid(), SIGUSR1);
640 if (!system_wide && target_tid == -1 && target_pid == -1)
641 all_tids[0] = child_pid;
643 close(child_ready_pipe[1]);
646 * wait for child to settle
648 if (read(child_ready_pipe[0], &buf, 1) == -1) {
649 perror("unable to read pipe");
652 close(child_ready_pipe[0]);
655 nr_cpus = read_cpu_map(cpu_list);
657 perror("failed to collect number of CPUs");
661 if (!system_wide && no_inherit && !cpu_list) {
664 for (i = 0; i < nr_cpus; i++)
665 open_counters(cpumap[i]);
668 perf_session__set_sample_type(session, sample_type);
671 err = perf_header__write_pipe(output);
674 } else if (file_new) {
675 err = perf_header__write(&session->header, output, false);
680 post_processing_offset = lseek(output, 0, SEEK_CUR);
682 perf_session__set_sample_id_all(session, sample_id_all_avail);
685 err = event__synthesize_attrs(&session->header,
686 process_synthesized_event,
689 pr_err("Couldn't synthesize attrs.\n");
693 err = event__synthesize_event_types(process_synthesized_event,
696 pr_err("Couldn't synthesize event_types.\n");
700 if (have_tracepoints(attrs, nr_counters)) {
702 * FIXME err <= 0 here actually means that
703 * there were no tracepoints so its not really
704 * an error, just that we don't need to
705 * synthesize anything. We really have to
706 * return this more properly and also
707 * propagate errors that now are calling die()
709 err = event__synthesize_tracing_data(output, attrs,
711 process_synthesized_event,
714 pr_err("Couldn't record tracing data.\n");
721 machine = perf_session__find_host_machine(session);
723 pr_err("Couldn't find native kernel information.\n");
727 err = event__synthesize_kernel_mmap(process_synthesized_event,
728 session, machine, "_text");
730 err = event__synthesize_kernel_mmap(process_synthesized_event,
731 session, machine, "_stext");
733 pr_err("Couldn't record kernel reference relocation symbol\n"
734 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
735 "Check /proc/kallsyms permission or run as root.\n");
737 err = event__synthesize_modules(process_synthesized_event,
740 pr_err("Couldn't record kernel module information.\n"
741 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
742 "Check /proc/modules permission or run as root.\n");
745 perf_session__process_machines(session, event__synthesize_guest_os);
748 event__synthesize_thread(target_tid, process_synthesized_event,
751 event__synthesize_threads(process_synthesized_event, session);
754 struct sched_param param;
756 param.sched_priority = realtime_prio;
757 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
758 pr_err("Could not set realtime priority.\n");
775 if (hits == samples) {
778 err = poll(event_array, nr_poll, -1);
783 for (i = 0; i < nr_cpu; i++) {
785 counter < nr_counters;
790 ioctl(fd[i][counter][thread],
791 PERF_EVENT_IOC_DISABLE);
797 if (quiet || signr == SIGUSR1)
800 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
803 * Approximate RIP event size: 24 bytes.
806 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
807 (double)bytes_written / 1024.0 / 1024.0,
814 perf_session__delete(session);
818 static const char * const record_usage[] = {
819 "perf record [<options>] [<command>]",
820 "perf record [<options>] -- <command> [<options>]",
824 static bool force, append_file;
826 const struct option record_options[] = {
827 OPT_CALLBACK('e', "event", NULL, "event",
828 "event selector. use 'perf list' to list available events",
830 OPT_CALLBACK(0, "filter", NULL, "filter",
831 "event filter", parse_filter),
832 OPT_INTEGER('p', "pid", &target_pid,
833 "record events on existing process id"),
834 OPT_INTEGER('t', "tid", &target_tid,
835 "record events on existing thread id"),
836 OPT_INTEGER('r', "realtime", &realtime_prio,
837 "collect data with this RT SCHED_FIFO priority"),
838 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
839 "collect raw sample records from all opened counters"),
840 OPT_BOOLEAN('a', "all-cpus", &system_wide,
841 "system-wide collection from all CPUs"),
842 OPT_BOOLEAN('A', "append", &append_file,
843 "append to the output file to do incremental profiling"),
844 OPT_STRING('C', "cpu", &cpu_list, "cpu",
845 "list of cpus to monitor"),
846 OPT_BOOLEAN('f', "force", &force,
847 "overwrite existing data file (deprecated)"),
848 OPT_U64('c', "count", &user_interval, "event period to sample"),
849 OPT_STRING('o', "output", &output_name, "file",
851 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
852 "child tasks do not inherit counters"),
853 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
854 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
855 OPT_BOOLEAN('g', "call-graph", &call_graph,
856 "do call-graph (stack chain/backtrace) recording"),
857 OPT_INCR('v', "verbose", &verbose,
858 "be more verbose (show counter open errors, etc)"),
859 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
860 OPT_BOOLEAN('s', "stat", &inherit_stat,
861 "per thread counts"),
862 OPT_BOOLEAN('d', "data", &sample_address,
864 OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
865 OPT_BOOLEAN('n', "no-samples", &no_samples,
867 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
868 "do not update the buildid cache"),
869 OPT_BOOLEAN('B', "no-buildid", &no_buildid,
870 "do not collect buildids in perf.data"),
874 int cmd_record(int argc, const char **argv, const char *prefix __used)
876 int i, j, err = -ENOMEM;
878 argc = parse_options(argc, argv, record_options, record_usage,
879 PARSE_OPT_STOP_AT_NON_OPTION);
880 if (!argc && target_pid == -1 && target_tid == -1 &&
881 !system_wide && !cpu_list)
882 usage_with_options(record_usage, record_options);
884 if (force && append_file) {
885 fprintf(stderr, "Can't overwrite and append at the same time."
886 " You need to choose between -f and -A");
887 usage_with_options(record_usage, record_options);
888 } else if (append_file) {
889 write_mode = WRITE_APPEND;
891 write_mode = WRITE_FORCE;
896 if (no_buildid_cache || no_buildid)
897 disable_buildid_cache();
901 attrs[0].type = PERF_TYPE_HARDWARE;
902 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
905 if (target_pid != -1) {
906 target_tid = target_pid;
907 thread_num = find_all_tid(target_pid, &all_tids);
908 if (thread_num <= 0) {
909 fprintf(stderr, "Can't find all threads of pid %d\n",
911 usage_with_options(record_usage, record_options);
914 all_tids=malloc(sizeof(pid_t));
916 goto out_symbol_exit;
918 all_tids[0] = target_tid;
922 for (i = 0; i < MAX_NR_CPUS; i++) {
923 for (j = 0; j < MAX_COUNTERS; j++) {
924 fd[i][j] = malloc(sizeof(int)*thread_num);
929 event_array = malloc(
930 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
934 if (user_interval != ULLONG_MAX)
935 default_interval = user_interval;
936 if (user_freq != UINT_MAX)
940 * User specified count overrides default frequency.
942 if (default_interval)
945 default_interval = freq;
947 fprintf(stderr, "frequency and count are zero, aborting\n");
949 goto out_free_event_array;
952 err = __cmd_record(argc, argv);
954 out_free_event_array:
957 for (i = 0; i < MAX_NR_CPUS; i++) {
958 for (j = 0; j < MAX_COUNTERS; j++)