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;
40 static int nr_cpus = 0;
41 static unsigned int page_size;
42 static unsigned int mmap_pages = 128;
43 static unsigned int user_freq = UINT_MAX;
44 static int freq = 1000;
46 static int pipe_output = 0;
47 static const char *output_name = "perf.data";
49 static int realtime_prio = 0;
50 static bool raw_samples = false;
51 static bool system_wide = false;
52 static int profile_cpu = -1;
53 static pid_t target_pid = -1;
54 static pid_t target_tid = -1;
55 static pid_t *all_tids = NULL;
56 static int thread_num = 0;
57 static pid_t child_pid = -1;
58 static bool no_inherit = false;
59 static enum write_mode_t write_mode = WRITE_FORCE;
60 static bool call_graph = false;
61 static bool inherit_stat = false;
62 static bool no_samples = false;
63 static bool sample_address = false;
64 static bool multiplex = false;
65 static int multiplex_fd = -1;
67 static long samples = 0;
68 static u64 bytes_written = 0;
70 static struct pollfd *event_array;
72 static int nr_poll = 0;
73 static int nr_cpu = 0;
75 static int file_new = 1;
76 static off_t post_processing_offset;
78 static struct perf_session *session;
87 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
89 static unsigned long mmap_read_head(struct mmap_data *md)
91 struct perf_event_mmap_page *pc = md->base;
100 static void mmap_write_tail(struct mmap_data *md, unsigned long tail)
102 struct perf_event_mmap_page *pc = md->base;
105 * ensure all reads are done before we write the tail out.
108 pc->data_tail = tail;
111 static void advance_output(size_t size)
113 bytes_written += size;
116 static void write_output(void *buf, size_t size)
119 int ret = write(output, buf, size);
122 die("failed to write");
127 bytes_written += ret;
131 static int process_synthesized_event(event_t *event,
132 struct perf_session *self __used)
134 write_output(event, event->header.size);
138 static void mmap_read(struct mmap_data *md)
140 unsigned int head = mmap_read_head(md);
141 unsigned int old = md->prev;
142 unsigned char *data = md->base + page_size;
148 * If we're further behind than half the buffer, there's a chance
149 * the writer will bite our tail and mess up the samples under us.
151 * If we somehow ended up ahead of the head, we got messed up.
153 * In either case, truncate and restart at head.
157 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
159 * head points to a known good entry, start there.
169 if ((old & md->mask) + size != (head & md->mask)) {
170 buf = &data[old & md->mask];
171 size = md->mask + 1 - (old & md->mask);
174 write_output(buf, size);
177 buf = &data[old & md->mask];
181 write_output(buf, size);
184 mmap_write_tail(md, old);
187 static volatile int done = 0;
188 static volatile int signr = -1;
190 static void sig_handler(int sig)
196 static void sig_atexit(void)
199 kill(child_pid, SIGTERM);
204 signal(signr, SIG_DFL);
205 kill(getpid(), signr);
210 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
212 struct perf_header_attr *h_attr;
214 if (nr < session->header.attrs) {
215 h_attr = session->header.attr[nr];
217 h_attr = perf_header_attr__new(a);
219 if (perf_header__add_attr(&session->header, h_attr) < 0) {
220 perf_header_attr__delete(h_attr);
228 static void create_counter(int counter, int cpu)
230 char *filter = filters[counter];
231 struct perf_event_attr *attr = attrs + counter;
232 struct perf_header_attr *h_attr;
233 int track = !counter; /* only the first counter needs these */
243 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
244 PERF_FORMAT_TOTAL_TIME_RUNNING |
247 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
250 attr->sample_type |= PERF_SAMPLE_ID;
253 * We default some events to a 1 default interval. But keep
254 * it a weak assumption overridable by the user.
256 if (!attr->sample_period || (user_freq != UINT_MAX &&
257 user_interval != ULLONG_MAX)) {
259 attr->sample_type |= PERF_SAMPLE_PERIOD;
261 attr->sample_freq = freq;
263 attr->sample_period = default_interval;
268 attr->sample_freq = 0;
271 attr->inherit_stat = 1;
274 attr->sample_type |= PERF_SAMPLE_ADDR;
277 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
280 attr->sample_type |= PERF_SAMPLE_TIME;
281 attr->sample_type |= PERF_SAMPLE_RAW;
282 attr->sample_type |= PERF_SAMPLE_CPU;
287 attr->inherit = !no_inherit;
288 if (target_pid == -1 && target_tid == -1 && !system_wide) {
290 attr->enable_on_exec = 1;
293 for (thread_index = 0; thread_index < thread_num; thread_index++) {
295 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
296 all_tids[thread_index], cpu, group_fd, 0);
298 if (fd[nr_cpu][counter][thread_index] < 0) {
301 if (err == EPERM || err == EACCES)
302 die("Permission error - are you root?\n"
303 "\t Consider tweaking"
304 " /proc/sys/kernel/perf_event_paranoid.\n");
305 else if (err == ENODEV && profile_cpu != -1) {
306 die("No such device - did you specify"
307 " an out-of-range profile CPU?\n");
311 * If it's cycles then fall back to hrtimer
312 * based cpu-clock-tick sw counter, which
313 * is always available even if no PMU support:
315 if (attr->type == PERF_TYPE_HARDWARE
316 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
319 warning(" ... trying to fall back to cpu-clock-ticks\n");
320 attr->type = PERF_TYPE_SOFTWARE;
321 attr->config = PERF_COUNT_SW_CPU_CLOCK;
325 error("perfcounter syscall returned with %d (%s)\n",
326 fd[nr_cpu][counter][thread_index], strerror(err));
328 #if defined(__i386__) || defined(__x86_64__)
329 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
330 die("No hardware sampling interrupt available."
331 " No APIC? If so then you can boot the kernel"
332 " with the \"lapic\" boot parameter to"
333 " force-enable it.\n");
336 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
340 h_attr = get_header_attr(attr, counter);
345 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
346 fprintf(stderr, "incompatible append\n");
351 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
352 perror("Unable to read perf file descriptor\n");
356 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
357 pr_warning("Not enough memory to add id\n");
361 assert(fd[nr_cpu][counter][thread_index] >= 0);
362 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
365 * First counter acts as the group leader:
367 if (group && group_fd == -1)
368 group_fd = fd[nr_cpu][counter][thread_index];
369 if (multiplex && multiplex_fd == -1)
370 multiplex_fd = fd[nr_cpu][counter][thread_index];
372 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
374 ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
377 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
378 event_array[nr_poll].events = POLLIN;
381 mmap_array[nr_cpu][counter][thread_index].counter = counter;
382 mmap_array[nr_cpu][counter][thread_index].prev = 0;
383 mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
384 mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
385 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
386 if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
387 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
392 if (filter != NULL) {
393 ret = ioctl(fd[nr_cpu][counter][thread_index],
394 PERF_EVENT_IOC_SET_FILTER, filter);
396 error("failed to set filter with %d (%s)\n", errno,
404 static void open_counters(int cpu)
409 for (counter = 0; counter < nr_counters; counter++)
410 create_counter(counter, cpu);
415 static int process_buildids(void)
417 u64 size = lseek(output, 0, SEEK_CUR);
422 session->fd = output;
423 return __perf_session__process_events(session, post_processing_offset,
424 size - post_processing_offset,
425 size, &build_id__mark_dso_hit_ops);
428 static void atexit_header(void)
431 session->header.data_size += bytes_written;
434 perf_header__write(&session->header, output, true);
438 static void event__synthesize_guest_os(struct machine *machine, void *data)
441 char *guest_kallsyms;
443 struct perf_session *psession = data;
445 if (machine__is_host(machine))
449 *As for guest kernel when processing subcommand record&report,
450 *we arrange module mmap prior to guest kernel mmap and trigger
451 *a preload dso because default guest module symbols are loaded
452 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
453 *method is used to avoid symbol missing when the first addr is
454 *in module instead of in guest kernel.
456 err = event__synthesize_modules(process_synthesized_event,
459 pr_err("Couldn't record guest kernel [%d]'s reference"
460 " relocation symbol.\n", machine->pid);
462 if (machine__is_default_guest(machine))
463 guest_kallsyms = (char *) symbol_conf.default_guest_kallsyms;
465 sprintf(path, "%s/proc/kallsyms", machine->root_dir);
466 guest_kallsyms = path;
470 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
471 * have no _text sometimes.
473 err = event__synthesize_kernel_mmap(process_synthesized_event,
474 psession, machine, "_text");
476 err = event__synthesize_kernel_mmap(process_synthesized_event,
477 psession, machine, "_stext");
479 pr_err("Couldn't record guest kernel [%d]'s reference"
480 " relocation symbol.\n", machine->pid);
483 static struct perf_event_header finished_round_event = {
484 .size = sizeof(struct perf_event_header),
485 .type = PERF_RECORD_FINISHED_ROUND,
488 static void mmap_read_all(void)
490 int i, counter, thread;
492 for (i = 0; i < nr_cpu; i++) {
493 for (counter = 0; counter < nr_counters; counter++) {
494 for (thread = 0; thread < thread_num; thread++) {
495 if (mmap_array[i][counter][thread].base)
496 mmap_read(&mmap_array[i][counter][thread]);
502 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
503 write_output(&finished_round_event, sizeof(finished_round_event));
506 static int __cmd_record(int argc, const char **argv)
513 unsigned long waking = 0;
514 int child_ready_pipe[2], go_pipe[2];
515 const bool forks = argc > 0;
517 struct machine *machine;
519 page_size = sysconf(_SC_PAGE_SIZE);
522 signal(SIGCHLD, sig_handler);
523 signal(SIGINT, sig_handler);
525 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
526 perror("failed to create pipes");
530 if (!strcmp(output_name, "-"))
532 else if (!stat(output_name, &st) && st.st_size) {
533 if (write_mode == WRITE_FORCE) {
534 char oldname[PATH_MAX];
535 snprintf(oldname, sizeof(oldname), "%s.old",
538 rename(output_name, oldname);
540 } else if (write_mode == WRITE_APPEND) {
541 write_mode = WRITE_FORCE;
544 flags = O_CREAT|O_RDWR;
545 if (write_mode == WRITE_APPEND)
551 output = STDOUT_FILENO;
553 output = open(output_name, flags, S_IRUSR | S_IWUSR);
555 perror("failed to create output file");
559 session = perf_session__new(output_name, O_WRONLY,
560 write_mode == WRITE_FORCE, false);
561 if (session == NULL) {
562 pr_err("Not enough memory for reading perf file header\n");
567 err = perf_header__read(session, output);
572 if (have_tracepoints(attrs, nr_counters))
573 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
575 atexit(atexit_header);
580 perror("failed to fork");
587 close(child_ready_pipe[0]);
589 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
592 * Do a dummy execvp to get the PLT entry resolved,
593 * so we avoid the resolver overhead on the real
596 execvp("", (char **)argv);
599 * Tell the parent we're ready to go
601 close(child_ready_pipe[1]);
604 * Wait until the parent tells us to go.
606 if (read(go_pipe[0], &buf, 1) == -1)
607 perror("unable to read pipe");
609 execvp(argv[0], (char **)argv);
615 if (!system_wide && target_tid == -1 && target_pid == -1)
616 all_tids[0] = child_pid;
618 close(child_ready_pipe[1]);
621 * wait for child to settle
623 if (read(child_ready_pipe[0], &buf, 1) == -1) {
624 perror("unable to read pipe");
627 close(child_ready_pipe[0]);
630 if ((!system_wide && no_inherit) || profile_cpu != -1) {
631 open_counters(profile_cpu);
633 nr_cpus = read_cpu_map();
634 for (i = 0; i < nr_cpus; i++)
635 open_counters(cpumap[i]);
639 err = perf_header__write_pipe(output);
642 } else if (file_new) {
643 err = perf_header__write(&session->header, output, false);
648 post_processing_offset = lseek(output, 0, SEEK_CUR);
651 err = event__synthesize_attrs(&session->header,
652 process_synthesized_event,
655 pr_err("Couldn't synthesize attrs.\n");
659 err = event__synthesize_event_types(process_synthesized_event,
662 pr_err("Couldn't synthesize event_types.\n");
666 if (have_tracepoints(attrs, nr_counters)) {
668 * FIXME err <= 0 here actually means that
669 * there were no tracepoints so its not really
670 * an error, just that we don't need to
671 * synthesize anything. We really have to
672 * return this more properly and also
673 * propagate errors that now are calling die()
675 err = event__synthesize_tracing_data(output, attrs,
677 process_synthesized_event,
680 pr_err("Couldn't record tracing data.\n");
687 machine = perf_session__find_host_machine(session);
689 pr_err("Couldn't find native kernel information.\n");
693 err = event__synthesize_kernel_mmap(process_synthesized_event,
694 session, machine, "_text");
696 err = event__synthesize_kernel_mmap(process_synthesized_event,
697 session, machine, "_stext");
699 pr_err("Couldn't record kernel reference relocation symbol.\n");
703 err = event__synthesize_modules(process_synthesized_event,
706 pr_err("Couldn't record kernel reference relocation symbol.\n");
710 perf_session__process_machines(session, event__synthesize_guest_os);
712 if (!system_wide && profile_cpu == -1)
713 event__synthesize_thread(target_tid, process_synthesized_event,
716 event__synthesize_threads(process_synthesized_event, session);
719 struct sched_param param;
721 param.sched_priority = realtime_prio;
722 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
723 pr_err("Could not set realtime priority.\n");
740 if (hits == samples) {
743 err = poll(event_array, nr_poll, -1);
748 for (i = 0; i < nr_cpu; i++) {
750 counter < nr_counters;
755 ioctl(fd[i][counter][thread],
756 PERF_EVENT_IOC_DISABLE);
762 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
765 * Approximate RIP event size: 24 bytes.
768 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
769 (double)bytes_written / 1024.0 / 1024.0,
776 static const char * const record_usage[] = {
777 "perf record [<options>] [<command>]",
778 "perf record [<options>] -- <command> [<options>]",
782 static bool force, append_file;
784 static const struct option options[] = {
785 OPT_CALLBACK('e', "event", NULL, "event",
786 "event selector. use 'perf list' to list available events",
788 OPT_CALLBACK(0, "filter", NULL, "filter",
789 "event filter", parse_filter),
790 OPT_INTEGER('p', "pid", &target_pid,
791 "record events on existing process id"),
792 OPT_INTEGER('t', "tid", &target_tid,
793 "record events on existing thread id"),
794 OPT_INTEGER('r', "realtime", &realtime_prio,
795 "collect data with this RT SCHED_FIFO priority"),
796 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
797 "collect raw sample records from all opened counters"),
798 OPT_BOOLEAN('a', "all-cpus", &system_wide,
799 "system-wide collection from all CPUs"),
800 OPT_BOOLEAN('A', "append", &append_file,
801 "append to the output file to do incremental profiling"),
802 OPT_INTEGER('C', "profile_cpu", &profile_cpu,
803 "CPU to profile on"),
804 OPT_BOOLEAN('f', "force", &force,
805 "overwrite existing data file (deprecated)"),
806 OPT_U64('c', "count", &user_interval, "event period to sample"),
807 OPT_STRING('o', "output", &output_name, "file",
809 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
810 "child tasks do not inherit counters"),
811 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
812 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
813 OPT_BOOLEAN('g', "call-graph", &call_graph,
814 "do call-graph (stack chain/backtrace) recording"),
815 OPT_INCR('v', "verbose", &verbose,
816 "be more verbose (show counter open errors, etc)"),
817 OPT_BOOLEAN('s', "stat", &inherit_stat,
818 "per thread counts"),
819 OPT_BOOLEAN('d', "data", &sample_address,
821 OPT_BOOLEAN('n', "no-samples", &no_samples,
823 OPT_BOOLEAN('M', "multiplex", &multiplex,
824 "multiplex counter output in a single channel"),
828 int cmd_record(int argc, const char **argv, const char *prefix __used)
832 argc = parse_options(argc, argv, options, record_usage,
833 PARSE_OPT_STOP_AT_NON_OPTION);
834 if (!argc && target_pid == -1 && target_tid == -1 &&
835 !system_wide && profile_cpu == -1)
836 usage_with_options(record_usage, options);
838 if (force && append_file) {
839 fprintf(stderr, "Can't overwrite and append at the same time."
840 " You need to choose between -f and -A");
841 usage_with_options(record_usage, options);
842 } else if (append_file) {
843 write_mode = WRITE_APPEND;
845 write_mode = WRITE_FORCE;
852 attrs[0].type = PERF_TYPE_HARDWARE;
853 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
856 if (target_pid != -1) {
857 target_tid = target_pid;
858 thread_num = find_all_tid(target_pid, &all_tids);
859 if (thread_num <= 0) {
860 fprintf(stderr, "Can't find all threads of pid %d\n",
862 usage_with_options(record_usage, options);
865 all_tids=malloc(sizeof(pid_t));
869 all_tids[0] = target_tid;
873 for (i = 0; i < MAX_NR_CPUS; i++) {
874 for (j = 0; j < MAX_COUNTERS; j++) {
875 fd[i][j] = malloc(sizeof(int)*thread_num);
876 mmap_array[i][j] = zalloc(
877 sizeof(struct mmap_data)*thread_num);
878 if (!fd[i][j] || !mmap_array[i][j])
882 event_array = malloc(
883 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
887 if (user_interval != ULLONG_MAX)
888 default_interval = user_interval;
889 if (user_freq != UINT_MAX)
893 * User specified count overrides default frequency.
895 if (default_interval)
898 default_interval = freq;
900 fprintf(stderr, "frequency and count are zero, aborting\n");
904 return __cmd_record(argc, argv);