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/evsel.h"
22 #include "util/debug.h"
23 #include "util/session.h"
24 #include "util/symbol.h"
25 #include "util/cpumap.h"
31 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
38 static u64 user_interval = ULLONG_MAX;
39 static u64 default_interval = 0;
40 static u64 sample_type;
42 static struct cpu_map *cpus;
43 static unsigned int page_size;
44 static unsigned int mmap_pages = 128;
45 static unsigned int user_freq = UINT_MAX;
46 static int freq = 1000;
48 static int pipe_output = 0;
49 static const char *output_name = "perf.data";
51 static int realtime_prio = 0;
52 static bool raw_samples = false;
53 static bool sample_id_all_avail = true;
54 static bool system_wide = false;
55 static pid_t target_pid = -1;
56 static pid_t target_tid = -1;
57 static pid_t *all_tids = NULL;
58 static int thread_num = 0;
59 static pid_t child_pid = -1;
60 static bool no_inherit = false;
61 static enum write_mode_t write_mode = WRITE_FORCE;
62 static bool call_graph = false;
63 static bool inherit_stat = false;
64 static bool no_samples = false;
65 static bool sample_address = false;
66 static bool sample_time = false;
67 static bool no_buildid = false;
68 static bool no_buildid_cache = false;
70 static long samples = 0;
71 static u64 bytes_written = 0;
73 static struct pollfd *event_array;
75 static int nr_poll = 0;
76 static int nr_cpu = 0;
78 static int file_new = 1;
79 static off_t post_processing_offset;
81 static struct perf_session *session;
82 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(struct perf_evsel *evsel, int cpu)
234 char *filter = evsel->filter;
235 struct perf_event_attr *attr = &evsel->attr;
236 struct perf_header_attr *h_attr;
237 int track = !evsel->idx; /* only the first counter needs these */
247 * Check if parse_single_tracepoint_event has already asked for
250 * XXX this is kludgy but short term fix for problems introduced by
251 * eac23d1c that broke 'perf script' by having different sample_types
252 * when using multiple tracepoint events when we use a perf binary
253 * that tries to use sample_id_all on an older kernel.
255 * We need to move counter creation to perf_session, support
256 * different sample_types, etc.
258 bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
260 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
261 PERF_FORMAT_TOTAL_TIME_RUNNING |
264 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
267 attr->sample_type |= PERF_SAMPLE_ID;
270 * We default some events to a 1 default interval. But keep
271 * it a weak assumption overridable by the user.
273 if (!attr->sample_period || (user_freq != UINT_MAX &&
274 user_interval != ULLONG_MAX)) {
276 attr->sample_type |= PERF_SAMPLE_PERIOD;
278 attr->sample_freq = freq;
280 attr->sample_period = default_interval;
285 attr->sample_freq = 0;
288 attr->inherit_stat = 1;
290 if (sample_address) {
291 attr->sample_type |= PERF_SAMPLE_ADDR;
292 attr->mmap_data = track;
296 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
299 attr->sample_type |= PERF_SAMPLE_CPU;
301 if (sample_id_all_avail &&
302 (sample_time || system_wide || !no_inherit || cpu_list))
303 attr->sample_type |= PERF_SAMPLE_TIME;
306 attr->sample_type |= PERF_SAMPLE_TIME;
307 attr->sample_type |= PERF_SAMPLE_RAW;
308 attr->sample_type |= PERF_SAMPLE_CPU;
313 attr->inherit = !no_inherit;
314 if (target_pid == -1 && target_tid == -1 && !system_wide) {
316 attr->enable_on_exec = 1;
319 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
321 for (thread_index = 0; thread_index < thread_num; thread_index++) {
323 FD(evsel, nr_cpu, thread_index) = sys_perf_event_open(attr, all_tids[thread_index], cpu, group_fd, 0);
325 if (FD(evsel, nr_cpu, thread_index) < 0) {
328 if (err == EPERM || err == EACCES)
329 die("Permission error - are you root?\n"
330 "\t Consider tweaking"
331 " /proc/sys/kernel/perf_event_paranoid.\n");
332 else if (err == ENODEV && cpu_list) {
333 die("No such device - did you specify"
334 " an out-of-range profile CPU?\n");
335 } else if (err == EINVAL && sample_id_all_avail) {
337 * Old kernel, no attr->sample_id_type_all field
339 sample_id_all_avail = false;
340 if (!sample_time && !raw_samples && !time_needed)
341 attr->sample_type &= ~PERF_SAMPLE_TIME;
343 goto retry_sample_id;
347 * If it's cycles then fall back to hrtimer
348 * based cpu-clock-tick sw counter, which
349 * is always available even if no PMU support:
351 if (attr->type == PERF_TYPE_HARDWARE
352 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
355 warning(" ... trying to fall back to cpu-clock-ticks\n");
356 attr->type = PERF_TYPE_SOFTWARE;
357 attr->config = PERF_COUNT_SW_CPU_CLOCK;
361 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
362 FD(evsel, nr_cpu, thread_index), strerror(err));
364 #if defined(__i386__) || defined(__x86_64__)
365 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
366 die("No hardware sampling interrupt available."
367 " No APIC? If so then you can boot the kernel"
368 " with the \"lapic\" boot parameter to"
369 " force-enable it.\n");
372 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
376 h_attr = get_header_attr(attr, evsel->idx);
381 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
382 fprintf(stderr, "incompatible append\n");
387 if (read(FD(evsel, nr_cpu, thread_index), &read_data, sizeof(read_data)) == -1) {
388 perror("Unable to read perf file descriptor");
392 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
393 pr_warning("Not enough memory to add id\n");
397 assert(FD(evsel, nr_cpu, thread_index) >= 0);
398 fcntl(FD(evsel, nr_cpu, thread_index), F_SETFL, O_NONBLOCK);
401 * First counter acts as the group leader:
403 if (group && group_fd == -1)
404 group_fd = FD(evsel, nr_cpu, thread_index);
406 if (evsel->idx || thread_index) {
407 struct perf_evsel *first;
408 first = list_entry(evsel_list.next, struct perf_evsel, node);
409 ret = ioctl(FD(evsel, nr_cpu, thread_index),
410 PERF_EVENT_IOC_SET_OUTPUT,
411 FD(first, nr_cpu, 0));
413 error("failed to set output: %d (%s)\n", errno,
418 mmap_array[nr_cpu].prev = 0;
419 mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
420 mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
421 PROT_READ | PROT_WRITE, MAP_SHARED, FD(evsel, nr_cpu, thread_index), 0);
422 if (mmap_array[nr_cpu].base == MAP_FAILED) {
423 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
427 event_array[nr_poll].fd = FD(evsel, nr_cpu, thread_index);
428 event_array[nr_poll].events = POLLIN;
432 if (filter != NULL) {
433 ret = ioctl(FD(evsel, nr_cpu, thread_index),
434 PERF_EVENT_IOC_SET_FILTER, filter);
436 error("failed to set filter with %d (%s)\n", errno,
444 sample_type = attr->sample_type;
447 static void open_counters(int cpu)
449 struct perf_evsel *pos;
453 list_for_each_entry(pos, &evsel_list, node)
454 create_counter(pos, cpu);
459 static int process_buildids(void)
461 u64 size = lseek(output, 0, SEEK_CUR);
466 session->fd = output;
467 return __perf_session__process_events(session, post_processing_offset,
468 size - post_processing_offset,
469 size, &build_id__mark_dso_hit_ops);
472 static void atexit_header(void)
475 session->header.data_size += bytes_written;
479 perf_header__write(&session->header, output, true);
480 perf_session__delete(session);
485 static void event__synthesize_guest_os(struct machine *machine, void *data)
488 struct perf_session *psession = data;
490 if (machine__is_host(machine))
494 *As for guest kernel when processing subcommand record&report,
495 *we arrange module mmap prior to guest kernel mmap and trigger
496 *a preload dso because default guest module symbols are loaded
497 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
498 *method is used to avoid symbol missing when the first addr is
499 *in module instead of in guest kernel.
501 err = event__synthesize_modules(process_synthesized_event,
504 pr_err("Couldn't record guest kernel [%d]'s reference"
505 " relocation symbol.\n", machine->pid);
508 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
509 * have no _text sometimes.
511 err = event__synthesize_kernel_mmap(process_synthesized_event,
512 psession, machine, "_text");
514 err = event__synthesize_kernel_mmap(process_synthesized_event,
515 psession, machine, "_stext");
517 pr_err("Couldn't record guest kernel [%d]'s reference"
518 " relocation symbol.\n", machine->pid);
521 static struct perf_event_header finished_round_event = {
522 .size = sizeof(struct perf_event_header),
523 .type = PERF_RECORD_FINISHED_ROUND,
526 static void mmap_read_all(void)
530 for (i = 0; i < nr_cpu; i++) {
531 if (mmap_array[i].base)
532 mmap_read(&mmap_array[i]);
535 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
536 write_output(&finished_round_event, sizeof(finished_round_event));
539 static int __cmd_record(int argc, const char **argv)
545 unsigned long waking = 0;
546 int child_ready_pipe[2], go_pipe[2];
547 const bool forks = argc > 0;
549 struct machine *machine;
551 page_size = sysconf(_SC_PAGE_SIZE);
554 signal(SIGCHLD, sig_handler);
555 signal(SIGINT, sig_handler);
556 signal(SIGUSR1, sig_handler);
558 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
559 perror("failed to create pipes");
563 if (!strcmp(output_name, "-"))
565 else if (!stat(output_name, &st) && st.st_size) {
566 if (write_mode == WRITE_FORCE) {
567 char oldname[PATH_MAX];
568 snprintf(oldname, sizeof(oldname), "%s.old",
571 rename(output_name, oldname);
573 } else if (write_mode == WRITE_APPEND) {
574 write_mode = WRITE_FORCE;
577 flags = O_CREAT|O_RDWR;
578 if (write_mode == WRITE_APPEND)
584 output = STDOUT_FILENO;
586 output = open(output_name, flags, S_IRUSR | S_IWUSR);
588 perror("failed to create output file");
592 session = perf_session__new(output_name, O_WRONLY,
593 write_mode == WRITE_FORCE, false, NULL);
594 if (session == NULL) {
595 pr_err("Not enough memory for reading perf file header\n");
600 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
603 err = perf_header__read(session, output);
605 goto out_delete_session;
608 if (have_tracepoints(&evsel_list))
609 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
612 * perf_session__delete(session) will be called at atexit_header()
614 atexit(atexit_header);
619 perror("failed to fork");
626 close(child_ready_pipe[0]);
628 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
631 * Do a dummy execvp to get the PLT entry resolved,
632 * so we avoid the resolver overhead on the real
635 execvp("", (char **)argv);
638 * Tell the parent we're ready to go
640 close(child_ready_pipe[1]);
643 * Wait until the parent tells us to go.
645 if (read(go_pipe[0], &buf, 1) == -1)
646 perror("unable to read pipe");
648 execvp(argv[0], (char **)argv);
651 kill(getppid(), SIGUSR1);
655 if (!system_wide && target_tid == -1 && target_pid == -1)
656 all_tids[0] = child_pid;
658 close(child_ready_pipe[1]);
661 * wait for child to settle
663 if (read(child_ready_pipe[0], &buf, 1) == -1) {
664 perror("unable to read pipe");
667 close(child_ready_pipe[0]);
670 if (!system_wide && no_inherit && !cpu_list) {
673 for (i = 0; i < cpus->nr; i++)
674 open_counters(cpus->map[i]);
677 perf_session__set_sample_type(session, sample_type);
680 err = perf_header__write_pipe(output);
683 } else if (file_new) {
684 err = perf_header__write(&session->header, output, false);
689 post_processing_offset = lseek(output, 0, SEEK_CUR);
691 perf_session__set_sample_id_all(session, sample_id_all_avail);
694 err = event__synthesize_attrs(&session->header,
695 process_synthesized_event,
698 pr_err("Couldn't synthesize attrs.\n");
702 err = event__synthesize_event_types(process_synthesized_event,
705 pr_err("Couldn't synthesize event_types.\n");
709 if (have_tracepoints(&evsel_list)) {
711 * FIXME err <= 0 here actually means that
712 * there were no tracepoints so its not really
713 * an error, just that we don't need to
714 * synthesize anything. We really have to
715 * return this more properly and also
716 * propagate errors that now are calling die()
718 err = event__synthesize_tracing_data(output, &evsel_list,
719 process_synthesized_event,
722 pr_err("Couldn't record tracing data.\n");
729 machine = perf_session__find_host_machine(session);
731 pr_err("Couldn't find native kernel information.\n");
735 err = event__synthesize_kernel_mmap(process_synthesized_event,
736 session, machine, "_text");
738 err = event__synthesize_kernel_mmap(process_synthesized_event,
739 session, machine, "_stext");
741 pr_err("Couldn't record kernel reference relocation symbol\n"
742 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
743 "Check /proc/kallsyms permission or run as root.\n");
745 err = event__synthesize_modules(process_synthesized_event,
748 pr_err("Couldn't record kernel module information.\n"
749 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
750 "Check /proc/modules permission or run as root.\n");
753 perf_session__process_machines(session, event__synthesize_guest_os);
756 event__synthesize_thread(target_tid, process_synthesized_event,
759 event__synthesize_threads(process_synthesized_event, session);
762 struct sched_param param;
764 param.sched_priority = realtime_prio;
765 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
766 pr_err("Could not set realtime priority.\n");
783 if (hits == samples) {
786 err = poll(event_array, nr_poll, -1);
791 for (i = 0; i < nr_cpu; i++) {
792 struct perf_evsel *pos;
794 list_for_each_entry(pos, &evsel_list, node) {
798 ioctl(FD(pos, i, thread),
799 PERF_EVENT_IOC_DISABLE);
805 if (quiet || signr == SIGUSR1)
808 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
811 * Approximate RIP event size: 24 bytes.
814 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
815 (double)bytes_written / 1024.0 / 1024.0,
822 perf_session__delete(session);
826 static const char * const record_usage[] = {
827 "perf record [<options>] [<command>]",
828 "perf record [<options>] -- <command> [<options>]",
832 static bool force, append_file;
834 const struct option record_options[] = {
835 OPT_CALLBACK('e', "event", NULL, "event",
836 "event selector. use 'perf list' to list available events",
838 OPT_CALLBACK(0, "filter", NULL, "filter",
839 "event filter", parse_filter),
840 OPT_INTEGER('p', "pid", &target_pid,
841 "record events on existing process id"),
842 OPT_INTEGER('t', "tid", &target_tid,
843 "record events on existing thread id"),
844 OPT_INTEGER('r', "realtime", &realtime_prio,
845 "collect data with this RT SCHED_FIFO priority"),
846 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
847 "collect raw sample records from all opened counters"),
848 OPT_BOOLEAN('a', "all-cpus", &system_wide,
849 "system-wide collection from all CPUs"),
850 OPT_BOOLEAN('A', "append", &append_file,
851 "append to the output file to do incremental profiling"),
852 OPT_STRING('C', "cpu", &cpu_list, "cpu",
853 "list of cpus to monitor"),
854 OPT_BOOLEAN('f', "force", &force,
855 "overwrite existing data file (deprecated)"),
856 OPT_U64('c', "count", &user_interval, "event period to sample"),
857 OPT_STRING('o', "output", &output_name, "file",
859 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
860 "child tasks do not inherit counters"),
861 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
862 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
863 OPT_BOOLEAN('g', "call-graph", &call_graph,
864 "do call-graph (stack chain/backtrace) recording"),
865 OPT_INCR('v', "verbose", &verbose,
866 "be more verbose (show counter open errors, etc)"),
867 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
868 OPT_BOOLEAN('s', "stat", &inherit_stat,
869 "per thread counts"),
870 OPT_BOOLEAN('d', "data", &sample_address,
872 OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
873 OPT_BOOLEAN('n', "no-samples", &no_samples,
875 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
876 "do not update the buildid cache"),
877 OPT_BOOLEAN('B', "no-buildid", &no_buildid,
878 "do not collect buildids in perf.data"),
882 int cmd_record(int argc, const char **argv, const char *prefix __used)
885 struct perf_evsel *pos;
887 argc = parse_options(argc, argv, record_options, record_usage,
888 PARSE_OPT_STOP_AT_NON_OPTION);
889 if (!argc && target_pid == -1 && target_tid == -1 &&
890 !system_wide && !cpu_list)
891 usage_with_options(record_usage, record_options);
893 if (force && append_file) {
894 fprintf(stderr, "Can't overwrite and append at the same time."
895 " You need to choose between -f and -A");
896 usage_with_options(record_usage, record_options);
897 } else if (append_file) {
898 write_mode = WRITE_APPEND;
900 write_mode = WRITE_FORCE;
905 if (no_buildid_cache || no_buildid)
906 disable_buildid_cache();
908 if (list_empty(&evsel_list) && perf_evsel_list__create_default() < 0) {
909 pr_err("Not enough memory for event selector list\n");
910 goto out_symbol_exit;
913 if (target_pid != -1) {
914 target_tid = target_pid;
915 thread_num = find_all_tid(target_pid, &all_tids);
916 if (thread_num <= 0) {
917 fprintf(stderr, "Can't find all threads of pid %d\n",
919 usage_with_options(record_usage, record_options);
922 all_tids=malloc(sizeof(pid_t));
924 goto out_symbol_exit;
926 all_tids[0] = target_tid;
930 cpus = cpu_map__new(cpu_list);
932 perror("failed to parse CPUs map");
936 list_for_each_entry(pos, &evsel_list, node) {
937 if (perf_evsel__alloc_fd(pos, cpus->nr, thread_num) < 0)
940 event_array = malloc(
941 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
945 if (user_interval != ULLONG_MAX)
946 default_interval = user_interval;
947 if (user_freq != UINT_MAX)
951 * User specified count overrides default frequency.
953 if (default_interval)
956 default_interval = freq;
958 fprintf(stderr, "frequency and count are zero, aborting\n");
960 goto out_free_event_array;
963 err = __cmd_record(argc, argv);
965 out_free_event_array: