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 pid_t target_pid = -1;
53 static pid_t target_tid = -1;
54 static pid_t *all_tids = NULL;
55 static int thread_num = 0;
56 static pid_t child_pid = -1;
57 static bool no_inherit = false;
58 static enum write_mode_t write_mode = WRITE_FORCE;
59 static bool call_graph = false;
60 static bool inherit_stat = false;
61 static bool no_samples = false;
62 static bool sample_address = false;
63 static bool no_buildid = false;
65 static long samples = 0;
66 static u64 bytes_written = 0;
68 static struct pollfd *event_array;
70 static int nr_poll = 0;
71 static int nr_cpu = 0;
73 static int file_new = 1;
74 static off_t post_processing_offset;
76 static struct perf_session *session;
77 static const char *cpu_list;
86 static struct mmap_data mmap_array[MAX_NR_CPUS];
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 advance_output(size_t size)
112 bytes_written += size;
115 static void write_output(void *buf, size_t size)
118 int ret = write(output, buf, size);
121 die("failed to write");
126 bytes_written += ret;
130 static int process_synthesized_event(event_t *event,
131 struct perf_session *self __used)
133 write_output(event, event->header.size);
137 static void mmap_read(struct mmap_data *md)
139 unsigned int head = mmap_read_head(md);
140 unsigned int old = md->prev;
141 unsigned char *data = md->base + page_size;
147 * If we're further behind than half the buffer, there's a chance
148 * the writer will bite our tail and mess up the samples under us.
150 * If we somehow ended up ahead of the head, we got messed up.
152 * In either case, truncate and restart at head.
156 fprintf(stderr, "WARNING: failed to keep up with mmap data\n");
158 * head points to a known good entry, start there.
168 if ((old & md->mask) + size != (head & md->mask)) {
169 buf = &data[old & md->mask];
170 size = md->mask + 1 - (old & md->mask);
173 write_output(buf, size);
176 buf = &data[old & md->mask];
180 write_output(buf, size);
183 mmap_write_tail(md, old);
186 static volatile int done = 0;
187 static volatile int signr = -1;
189 static void sig_handler(int sig)
195 static void sig_atexit(void)
198 kill(child_pid, SIGTERM);
200 if (signr == -1 || signr == SIGUSR1)
203 signal(signr, SIG_DFL);
204 kill(getpid(), signr);
209 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
211 struct perf_header_attr *h_attr;
213 if (nr < session->header.attrs) {
214 h_attr = session->header.attr[nr];
216 h_attr = perf_header_attr__new(a);
218 if (perf_header__add_attr(&session->header, h_attr) < 0) {
219 perf_header_attr__delete(h_attr);
227 static void create_counter(int counter, int cpu)
229 char *filter = filters[counter];
230 struct perf_event_attr *attr = attrs + counter;
231 struct perf_header_attr *h_attr;
232 int track = !counter; /* only the first counter needs these */
242 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
243 PERF_FORMAT_TOTAL_TIME_RUNNING |
246 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
249 attr->sample_type |= PERF_SAMPLE_ID;
252 * We default some events to a 1 default interval. But keep
253 * it a weak assumption overridable by the user.
255 if (!attr->sample_period || (user_freq != UINT_MAX &&
256 user_interval != ULLONG_MAX)) {
258 attr->sample_type |= PERF_SAMPLE_PERIOD;
260 attr->sample_freq = freq;
262 attr->sample_period = default_interval;
267 attr->sample_freq = 0;
270 attr->inherit_stat = 1;
272 if (sample_address) {
273 attr->sample_type |= PERF_SAMPLE_ADDR;
274 attr->mmap_data = track;
278 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
281 attr->sample_type |= PERF_SAMPLE_CPU;
284 attr->sample_type |= PERF_SAMPLE_TIME;
285 attr->sample_type |= PERF_SAMPLE_RAW;
286 attr->sample_type |= PERF_SAMPLE_CPU;
291 attr->inherit = !no_inherit;
292 if (target_pid == -1 && target_tid == -1 && !system_wide) {
294 attr->enable_on_exec = 1;
297 for (thread_index = 0; thread_index < thread_num; thread_index++) {
299 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
300 all_tids[thread_index], cpu, group_fd, 0);
302 if (fd[nr_cpu][counter][thread_index] < 0) {
305 if (err == EPERM || err == EACCES)
306 die("Permission error - are you root?\n"
307 "\t Consider tweaking"
308 " /proc/sys/kernel/perf_event_paranoid.\n");
309 else if (err == ENODEV && cpu_list) {
310 die("No such device - did you specify"
311 " an out-of-range profile CPU?\n");
315 * If it's cycles then fall back to hrtimer
316 * based cpu-clock-tick sw counter, which
317 * is always available even if no PMU support:
319 if (attr->type == PERF_TYPE_HARDWARE
320 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
323 warning(" ... trying to fall back to cpu-clock-ticks\n");
324 attr->type = PERF_TYPE_SOFTWARE;
325 attr->config = PERF_COUNT_SW_CPU_CLOCK;
329 error("perfcounter syscall returned with %d (%s)\n",
330 fd[nr_cpu][counter][thread_index], strerror(err));
332 #if defined(__i386__) || defined(__x86_64__)
333 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
334 die("No hardware sampling interrupt available."
335 " No APIC? If so then you can boot the kernel"
336 " with the \"lapic\" boot parameter to"
337 " force-enable it.\n");
340 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
344 h_attr = get_header_attr(attr, counter);
349 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
350 fprintf(stderr, "incompatible append\n");
355 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
356 perror("Unable to read perf file descriptor");
360 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
361 pr_warning("Not enough memory to add id\n");
365 assert(fd[nr_cpu][counter][thread_index] >= 0);
366 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
369 * First counter acts as the group leader:
371 if (group && group_fd == -1)
372 group_fd = fd[nr_cpu][counter][thread_index];
374 if (counter || thread_index) {
375 ret = ioctl(fd[nr_cpu][counter][thread_index],
376 PERF_EVENT_IOC_SET_OUTPUT,
379 error("failed to set output: %d (%s)\n", errno,
384 mmap_array[nr_cpu].counter = counter;
385 mmap_array[nr_cpu].prev = 0;
386 mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
387 mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
388 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
389 if (mmap_array[nr_cpu].base == MAP_FAILED) {
390 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
394 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
395 event_array[nr_poll].events = POLLIN;
399 if (filter != NULL) {
400 ret = ioctl(fd[nr_cpu][counter][thread_index],
401 PERF_EVENT_IOC_SET_FILTER, filter);
403 error("failed to set filter with %d (%s)\n", errno,
411 static void open_counters(int cpu)
416 for (counter = 0; counter < nr_counters; counter++)
417 create_counter(counter, cpu);
422 static int process_buildids(void)
424 u64 size = lseek(output, 0, SEEK_CUR);
429 session->fd = output;
430 return __perf_session__process_events(session, post_processing_offset,
431 size - post_processing_offset,
432 size, &build_id__mark_dso_hit_ops);
435 static void atexit_header(void)
438 session->header.data_size += bytes_written;
441 perf_header__write(&session->header, output, true);
442 perf_session__delete(session);
447 static void event__synthesize_guest_os(struct machine *machine, void *data)
450 struct perf_session *psession = data;
452 if (machine__is_host(machine))
456 *As for guest kernel when processing subcommand record&report,
457 *we arrange module mmap prior to guest kernel mmap and trigger
458 *a preload dso because default guest module symbols are loaded
459 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
460 *method is used to avoid symbol missing when the first addr is
461 *in module instead of in guest kernel.
463 err = event__synthesize_modules(process_synthesized_event,
466 pr_err("Couldn't record guest kernel [%d]'s reference"
467 " relocation symbol.\n", machine->pid);
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)
492 for (i = 0; i < nr_cpu; i++) {
493 if (mmap_array[i].base)
494 mmap_read(&mmap_array[i]);
497 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
498 write_output(&finished_round_event, sizeof(finished_round_event));
501 static int __cmd_record(int argc, const char **argv)
507 unsigned long waking = 0;
508 int child_ready_pipe[2], go_pipe[2];
509 const bool forks = argc > 0;
511 struct machine *machine;
513 page_size = sysconf(_SC_PAGE_SIZE);
516 signal(SIGCHLD, sig_handler);
517 signal(SIGINT, sig_handler);
518 signal(SIGUSR1, sig_handler);
520 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
521 perror("failed to create pipes");
525 if (!strcmp(output_name, "-"))
527 else if (!stat(output_name, &st) && st.st_size) {
528 if (write_mode == WRITE_FORCE) {
529 char oldname[PATH_MAX];
530 snprintf(oldname, sizeof(oldname), "%s.old",
533 rename(output_name, oldname);
535 } else if (write_mode == WRITE_APPEND) {
536 write_mode = WRITE_FORCE;
539 flags = O_CREAT|O_RDWR;
540 if (write_mode == WRITE_APPEND)
546 output = STDOUT_FILENO;
548 output = open(output_name, flags, S_IRUSR | S_IWUSR);
550 perror("failed to create output file");
554 session = perf_session__new(output_name, O_WRONLY,
555 write_mode == WRITE_FORCE, false);
556 if (session == NULL) {
557 pr_err("Not enough memory for reading perf file header\n");
562 err = perf_header__read(session, output);
564 goto out_delete_session;
567 if (have_tracepoints(attrs, nr_counters))
568 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
571 * perf_session__delete(session) will be called at atexit_header()
573 atexit(atexit_header);
578 perror("failed to fork");
585 close(child_ready_pipe[0]);
587 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
590 * Do a dummy execvp to get the PLT entry resolved,
591 * so we avoid the resolver overhead on the real
594 execvp("", (char **)argv);
597 * Tell the parent we're ready to go
599 close(child_ready_pipe[1]);
602 * Wait until the parent tells us to go.
604 if (read(go_pipe[0], &buf, 1) == -1)
605 perror("unable to read pipe");
607 execvp(argv[0], (char **)argv);
610 kill(getppid(), SIGUSR1);
614 if (!system_wide && target_tid == -1 && target_pid == -1)
615 all_tids[0] = child_pid;
617 close(child_ready_pipe[1]);
620 * wait for child to settle
622 if (read(child_ready_pipe[0], &buf, 1) == -1) {
623 perror("unable to read pipe");
626 close(child_ready_pipe[0]);
629 nr_cpus = read_cpu_map(cpu_list);
631 perror("failed to collect number of CPUs");
635 if (!system_wide && no_inherit && !cpu_list) {
638 for (i = 0; i < nr_cpus; i++)
639 open_counters(cpumap[i]);
643 err = perf_header__write_pipe(output);
646 } else if (file_new) {
647 err = perf_header__write(&session->header, output, false);
652 post_processing_offset = lseek(output, 0, SEEK_CUR);
655 err = event__synthesize_attrs(&session->header,
656 process_synthesized_event,
659 pr_err("Couldn't synthesize attrs.\n");
663 err = event__synthesize_event_types(process_synthesized_event,
666 pr_err("Couldn't synthesize event_types.\n");
670 if (have_tracepoints(attrs, nr_counters)) {
672 * FIXME err <= 0 here actually means that
673 * there were no tracepoints so its not really
674 * an error, just that we don't need to
675 * synthesize anything. We really have to
676 * return this more properly and also
677 * propagate errors that now are calling die()
679 err = event__synthesize_tracing_data(output, attrs,
681 process_synthesized_event,
684 pr_err("Couldn't record tracing data.\n");
691 machine = perf_session__find_host_machine(session);
693 pr_err("Couldn't find native kernel information.\n");
697 err = event__synthesize_kernel_mmap(process_synthesized_event,
698 session, machine, "_text");
700 err = event__synthesize_kernel_mmap(process_synthesized_event,
701 session, machine, "_stext");
703 pr_err("Couldn't record kernel reference relocation symbol\n"
704 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
705 "Check /proc/kallsyms permission or run as root.\n");
707 err = event__synthesize_modules(process_synthesized_event,
710 pr_err("Couldn't record kernel module information.\n"
711 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
712 "Check /proc/modules permission or run as root.\n");
715 perf_session__process_machines(session, event__synthesize_guest_os);
718 event__synthesize_thread(target_tid, process_synthesized_event,
721 event__synthesize_threads(process_synthesized_event, session);
724 struct sched_param param;
726 param.sched_priority = realtime_prio;
727 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
728 pr_err("Could not set realtime priority.\n");
745 if (hits == samples) {
748 err = poll(event_array, nr_poll, -1);
753 for (i = 0; i < nr_cpu; i++) {
755 counter < nr_counters;
760 ioctl(fd[i][counter][thread],
761 PERF_EVENT_IOC_DISABLE);
767 if (quiet || signr == SIGUSR1)
770 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
773 * Approximate RIP event size: 24 bytes.
776 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
777 (double)bytes_written / 1024.0 / 1024.0,
784 perf_session__delete(session);
788 static const char * const record_usage[] = {
789 "perf record [<options>] [<command>]",
790 "perf record [<options>] -- <command> [<options>]",
794 static bool force, append_file;
796 const struct option record_options[] = {
797 OPT_CALLBACK('e', "event", NULL, "event",
798 "event selector. use 'perf list' to list available events",
800 OPT_CALLBACK(0, "filter", NULL, "filter",
801 "event filter", parse_filter),
802 OPT_INTEGER('p', "pid", &target_pid,
803 "record events on existing process id"),
804 OPT_INTEGER('t', "tid", &target_tid,
805 "record events on existing thread id"),
806 OPT_INTEGER('r', "realtime", &realtime_prio,
807 "collect data with this RT SCHED_FIFO priority"),
808 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
809 "collect raw sample records from all opened counters"),
810 OPT_BOOLEAN('a', "all-cpus", &system_wide,
811 "system-wide collection from all CPUs"),
812 OPT_BOOLEAN('A', "append", &append_file,
813 "append to the output file to do incremental profiling"),
814 OPT_STRING('C', "cpu", &cpu_list, "cpu",
815 "list of cpus to monitor"),
816 OPT_BOOLEAN('f', "force", &force,
817 "overwrite existing data file (deprecated)"),
818 OPT_U64('c', "count", &user_interval, "event period to sample"),
819 OPT_STRING('o', "output", &output_name, "file",
821 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
822 "child tasks do not inherit counters"),
823 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
824 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
825 OPT_BOOLEAN('g', "call-graph", &call_graph,
826 "do call-graph (stack chain/backtrace) recording"),
827 OPT_INCR('v', "verbose", &verbose,
828 "be more verbose (show counter open errors, etc)"),
829 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
830 OPT_BOOLEAN('s', "stat", &inherit_stat,
831 "per thread counts"),
832 OPT_BOOLEAN('d', "data", &sample_address,
834 OPT_BOOLEAN('n', "no-samples", &no_samples,
836 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid,
837 "do not update the buildid cache"),
841 int cmd_record(int argc, const char **argv, const char *prefix __used)
843 int i, j, err = -ENOMEM;
845 argc = parse_options(argc, argv, record_options, record_usage,
846 PARSE_OPT_STOP_AT_NON_OPTION);
847 if (!argc && target_pid == -1 && target_tid == -1 &&
848 !system_wide && !cpu_list)
849 usage_with_options(record_usage, record_options);
851 if (force && append_file) {
852 fprintf(stderr, "Can't overwrite and append at the same time."
853 " You need to choose between -f and -A");
854 usage_with_options(record_usage, record_options);
855 } else if (append_file) {
856 write_mode = WRITE_APPEND;
858 write_mode = WRITE_FORCE;
863 disable_buildid_cache();
867 attrs[0].type = PERF_TYPE_HARDWARE;
868 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
871 if (target_pid != -1) {
872 target_tid = target_pid;
873 thread_num = find_all_tid(target_pid, &all_tids);
874 if (thread_num <= 0) {
875 fprintf(stderr, "Can't find all threads of pid %d\n",
877 usage_with_options(record_usage, record_options);
880 all_tids=malloc(sizeof(pid_t));
882 goto out_symbol_exit;
884 all_tids[0] = target_tid;
888 for (i = 0; i < MAX_NR_CPUS; i++) {
889 for (j = 0; j < MAX_COUNTERS; j++) {
890 fd[i][j] = malloc(sizeof(int)*thread_num);
895 event_array = malloc(
896 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
900 if (user_interval != ULLONG_MAX)
901 default_interval = user_interval;
902 if (user_freq != UINT_MAX)
906 * User specified count overrides default frequency.
908 if (default_interval)
911 default_interval = freq;
913 fprintf(stderr, "frequency and count are zero, aborting\n");
915 goto out_free_event_array;
918 err = __cmd_record(argc, argv);
920 out_free_event_array:
923 for (i = 0; i < MAX_NR_CPUS; i++) {
924 for (j = 0; j < MAX_COUNTERS; j++)