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 struct timeval last_read;
69 static struct timeval this_read;
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;
90 static struct mmap_data *mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
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 perf_session *self __used)
137 write_output(event, event->header.size);
141 static void mmap_read(struct mmap_data *md)
143 unsigned int head = mmap_read_head(md);
144 unsigned int old = md->prev;
145 unsigned char *data = md->base + page_size;
150 gettimeofday(&this_read, NULL);
153 * If we're further behind than half the buffer, there's a chance
154 * the writer will bite our tail and mess up the samples under us.
156 * If we somehow ended up ahead of the head, we got messed up.
158 * In either case, truncate and restart at head.
165 timersub(&this_read, &last_read, &iv);
166 msecs = iv.tv_sec*1000 + iv.tv_usec/1000;
168 fprintf(stderr, "WARNING: failed to keep up with mmap data."
169 " Last read %lu msecs ago.\n", msecs);
172 * head points to a known good entry, start there.
177 last_read = this_read;
184 if ((old & md->mask) + size != (head & md->mask)) {
185 buf = &data[old & md->mask];
186 size = md->mask + 1 - (old & md->mask);
189 write_output(buf, size);
192 buf = &data[old & md->mask];
196 write_output(buf, size);
199 mmap_write_tail(md, old);
202 static volatile int done = 0;
203 static volatile int signr = -1;
205 static void sig_handler(int sig)
211 static void sig_atexit(void)
214 kill(child_pid, SIGTERM);
219 signal(signr, SIG_DFL);
220 kill(getpid(), signr);
225 static struct perf_header_attr *get_header_attr(struct perf_event_attr *a, int nr)
227 struct perf_header_attr *h_attr;
229 if (nr < session->header.attrs) {
230 h_attr = session->header.attr[nr];
232 h_attr = perf_header_attr__new(a);
234 if (perf_header__add_attr(&session->header, h_attr) < 0) {
235 perf_header_attr__delete(h_attr);
243 static void create_counter(int counter, int cpu)
245 char *filter = filters[counter];
246 struct perf_event_attr *attr = attrs + counter;
247 struct perf_header_attr *h_attr;
248 int track = !counter; /* only the first counter needs these */
258 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
259 PERF_FORMAT_TOTAL_TIME_RUNNING |
262 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
265 attr->sample_type |= PERF_SAMPLE_ID;
268 * We default some events to a 1 default interval. But keep
269 * it a weak assumption overridable by the user.
271 if (!attr->sample_period || (user_freq != UINT_MAX &&
272 user_interval != ULLONG_MAX)) {
274 attr->sample_type |= PERF_SAMPLE_PERIOD;
276 attr->sample_freq = freq;
278 attr->sample_period = default_interval;
283 attr->sample_freq = 0;
286 attr->inherit_stat = 1;
289 attr->sample_type |= PERF_SAMPLE_ADDR;
292 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
295 attr->sample_type |= PERF_SAMPLE_TIME;
296 attr->sample_type |= PERF_SAMPLE_RAW;
297 attr->sample_type |= PERF_SAMPLE_CPU;
302 attr->inherit = !no_inherit;
303 if (target_pid == -1 && target_tid == -1 && !system_wide) {
305 attr->enable_on_exec = 1;
308 for (thread_index = 0; thread_index < thread_num; thread_index++) {
310 fd[nr_cpu][counter][thread_index] = sys_perf_event_open(attr,
311 all_tids[thread_index], cpu, group_fd, 0);
313 if (fd[nr_cpu][counter][thread_index] < 0) {
316 if (err == EPERM || err == EACCES)
317 die("Permission error - are you root?\n"
318 "\t Consider tweaking"
319 " /proc/sys/kernel/perf_event_paranoid.\n");
320 else if (err == ENODEV && profile_cpu != -1) {
321 die("No such device - did you specify"
322 " an out-of-range profile CPU?\n");
326 * If it's cycles then fall back to hrtimer
327 * based cpu-clock-tick sw counter, which
328 * is always available even if no PMU support:
330 if (attr->type == PERF_TYPE_HARDWARE
331 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
334 warning(" ... trying to fall back to cpu-clock-ticks\n");
335 attr->type = PERF_TYPE_SOFTWARE;
336 attr->config = PERF_COUNT_SW_CPU_CLOCK;
340 error("perfcounter syscall returned with %d (%s)\n",
341 fd[nr_cpu][counter][thread_index], strerror(err));
343 #if defined(__i386__) || defined(__x86_64__)
344 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
345 die("No hardware sampling interrupt available."
346 " No APIC? If so then you can boot the kernel"
347 " with the \"lapic\" boot parameter to"
348 " force-enable it.\n");
351 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
355 h_attr = get_header_attr(attr, counter);
360 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
361 fprintf(stderr, "incompatible append\n");
366 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
367 perror("Unable to read perf file descriptor\n");
371 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
372 pr_warning("Not enough memory to add id\n");
376 assert(fd[nr_cpu][counter][thread_index] >= 0);
377 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
380 * First counter acts as the group leader:
382 if (group && group_fd == -1)
383 group_fd = fd[nr_cpu][counter][thread_index];
384 if (multiplex && multiplex_fd == -1)
385 multiplex_fd = fd[nr_cpu][counter][thread_index];
387 if (multiplex && fd[nr_cpu][counter][thread_index] != multiplex_fd) {
389 ret = ioctl(fd[nr_cpu][counter][thread_index], PERF_EVENT_IOC_SET_OUTPUT, multiplex_fd);
392 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
393 event_array[nr_poll].events = POLLIN;
396 mmap_array[nr_cpu][counter][thread_index].counter = counter;
397 mmap_array[nr_cpu][counter][thread_index].prev = 0;
398 mmap_array[nr_cpu][counter][thread_index].mask = mmap_pages*page_size - 1;
399 mmap_array[nr_cpu][counter][thread_index].base = mmap(NULL, (mmap_pages+1)*page_size,
400 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
401 if (mmap_array[nr_cpu][counter][thread_index].base == MAP_FAILED) {
402 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
407 if (filter != NULL) {
408 ret = ioctl(fd[nr_cpu][counter][thread_index],
409 PERF_EVENT_IOC_SET_FILTER, filter);
411 error("failed to set filter with %d (%s)\n", errno,
419 static void open_counters(int cpu)
424 for (counter = 0; counter < nr_counters; counter++)
425 create_counter(counter, cpu);
430 static int process_buildids(void)
432 u64 size = lseek(output, 0, SEEK_CUR);
437 session->fd = output;
438 return __perf_session__process_events(session, post_processing_offset,
439 size - post_processing_offset,
440 size, &build_id__mark_dso_hit_ops);
443 static void atexit_header(void)
446 session->header.data_size += bytes_written;
449 perf_header__write(&session->header, output, true);
453 static void event__synthesize_guest_os(struct machine *machine, void *data)
456 char *guest_kallsyms;
458 struct perf_session *psession = data;
460 if (machine__is_host(machine))
464 *As for guest kernel when processing subcommand record&report,
465 *we arrange module mmap prior to guest kernel mmap and trigger
466 *a preload dso because default guest module symbols are loaded
467 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
468 *method is used to avoid symbol missing when the first addr is
469 *in module instead of in guest kernel.
471 err = event__synthesize_modules(process_synthesized_event,
474 pr_err("Couldn't record guest kernel [%d]'s reference"
475 " relocation symbol.\n", machine->pid);
477 if (machine__is_default_guest(machine))
478 guest_kallsyms = (char *) symbol_conf.default_guest_kallsyms;
480 sprintf(path, "%s/proc/kallsyms", machine->root_dir);
481 guest_kallsyms = path;
485 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
486 * have no _text sometimes.
488 err = event__synthesize_kernel_mmap(process_synthesized_event,
489 psession, machine, "_text");
491 err = event__synthesize_kernel_mmap(process_synthesized_event,
492 psession, machine, "_stext");
494 pr_err("Couldn't record guest kernel [%d]'s reference"
495 " relocation symbol.\n", machine->pid);
498 static struct perf_event_header finished_round_event = {
499 .size = sizeof(struct perf_event_header),
500 .type = PERF_RECORD_FINISHED_ROUND,
503 static void mmap_read_all(void)
505 int i, counter, thread;
507 for (i = 0; i < nr_cpu; i++) {
508 for (counter = 0; counter < nr_counters; counter++) {
509 for (thread = 0; thread < thread_num; thread++) {
510 if (mmap_array[i][counter][thread].base)
511 mmap_read(&mmap_array[i][counter][thread]);
517 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
518 write_output(&finished_round_event, sizeof(finished_round_event));
521 static int __cmd_record(int argc, const char **argv)
528 unsigned long waking = 0;
529 int child_ready_pipe[2], go_pipe[2];
530 const bool forks = argc > 0;
532 struct machine *machine;
534 page_size = sysconf(_SC_PAGE_SIZE);
537 signal(SIGCHLD, sig_handler);
538 signal(SIGINT, sig_handler);
540 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
541 perror("failed to create pipes");
545 if (!strcmp(output_name, "-"))
547 else if (!stat(output_name, &st) && st.st_size) {
548 if (write_mode == WRITE_FORCE) {
549 char oldname[PATH_MAX];
550 snprintf(oldname, sizeof(oldname), "%s.old",
553 rename(output_name, oldname);
555 } else if (write_mode == WRITE_APPEND) {
556 write_mode = WRITE_FORCE;
559 flags = O_CREAT|O_RDWR;
560 if (write_mode == WRITE_APPEND)
566 output = STDOUT_FILENO;
568 output = open(output_name, flags, S_IRUSR | S_IWUSR);
570 perror("failed to create output file");
574 session = perf_session__new(output_name, O_WRONLY,
575 write_mode == WRITE_FORCE, false);
576 if (session == NULL) {
577 pr_err("Not enough memory for reading perf file header\n");
582 err = perf_header__read(session, output);
587 if (have_tracepoints(attrs, nr_counters))
588 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
590 atexit(atexit_header);
595 perror("failed to fork");
602 close(child_ready_pipe[0]);
604 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
607 * Do a dummy execvp to get the PLT entry resolved,
608 * so we avoid the resolver overhead on the real
611 execvp("", (char **)argv);
614 * Tell the parent we're ready to go
616 close(child_ready_pipe[1]);
619 * Wait until the parent tells us to go.
621 if (read(go_pipe[0], &buf, 1) == -1)
622 perror("unable to read pipe");
624 execvp(argv[0], (char **)argv);
630 if (!system_wide && target_tid == -1 && target_pid == -1)
631 all_tids[0] = child_pid;
633 close(child_ready_pipe[1]);
636 * wait for child to settle
638 if (read(child_ready_pipe[0], &buf, 1) == -1) {
639 perror("unable to read pipe");
642 close(child_ready_pipe[0]);
645 if ((!system_wide && no_inherit) || profile_cpu != -1) {
646 open_counters(profile_cpu);
648 nr_cpus = read_cpu_map();
649 for (i = 0; i < nr_cpus; i++)
650 open_counters(cpumap[i]);
654 err = perf_header__write_pipe(output);
657 } else if (file_new) {
658 err = perf_header__write(&session->header, output, false);
663 post_processing_offset = lseek(output, 0, SEEK_CUR);
666 err = event__synthesize_attrs(&session->header,
667 process_synthesized_event,
670 pr_err("Couldn't synthesize attrs.\n");
674 err = event__synthesize_event_types(process_synthesized_event,
677 pr_err("Couldn't synthesize event_types.\n");
681 if (have_tracepoints(attrs, nr_counters)) {
683 * FIXME err <= 0 here actually means that
684 * there were no tracepoints so its not really
685 * an error, just that we don't need to
686 * synthesize anything. We really have to
687 * return this more properly and also
688 * propagate errors that now are calling die()
690 err = event__synthesize_tracing_data(output, attrs,
692 process_synthesized_event,
695 pr_err("Couldn't record tracing data.\n");
702 machine = perf_session__find_host_machine(session);
704 pr_err("Couldn't find native kernel information.\n");
708 err = event__synthesize_kernel_mmap(process_synthesized_event,
709 session, machine, "_text");
711 err = event__synthesize_kernel_mmap(process_synthesized_event,
712 session, machine, "_stext");
714 pr_err("Couldn't record kernel reference relocation symbol.\n");
718 err = event__synthesize_modules(process_synthesized_event,
721 pr_err("Couldn't record kernel reference relocation symbol.\n");
725 perf_session__process_machines(session, event__synthesize_guest_os);
727 if (!system_wide && profile_cpu == -1)
728 event__synthesize_thread(target_tid, process_synthesized_event,
731 event__synthesize_threads(process_synthesized_event, session);
734 struct sched_param param;
736 param.sched_priority = realtime_prio;
737 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
738 pr_err("Could not set realtime priority.\n");
755 if (hits == samples) {
758 err = poll(event_array, nr_poll, -1);
763 for (i = 0; i < nr_cpu; i++) {
765 counter < nr_counters;
770 ioctl(fd[i][counter][thread],
771 PERF_EVENT_IOC_DISABLE);
777 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
780 * Approximate RIP event size: 24 bytes.
783 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
784 (double)bytes_written / 1024.0 / 1024.0,
791 static const char * const record_usage[] = {
792 "perf record [<options>] [<command>]",
793 "perf record [<options>] -- <command> [<options>]",
797 static bool force, append_file;
799 static const struct option options[] = {
800 OPT_CALLBACK('e', "event", NULL, "event",
801 "event selector. use 'perf list' to list available events",
803 OPT_CALLBACK(0, "filter", NULL, "filter",
804 "event filter", parse_filter),
805 OPT_INTEGER('p', "pid", &target_pid,
806 "record events on existing process id"),
807 OPT_INTEGER('t', "tid", &target_tid,
808 "record events on existing thread id"),
809 OPT_INTEGER('r', "realtime", &realtime_prio,
810 "collect data with this RT SCHED_FIFO priority"),
811 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
812 "collect raw sample records from all opened counters"),
813 OPT_BOOLEAN('a', "all-cpus", &system_wide,
814 "system-wide collection from all CPUs"),
815 OPT_BOOLEAN('A', "append", &append_file,
816 "append to the output file to do incremental profiling"),
817 OPT_INTEGER('C', "profile_cpu", &profile_cpu,
818 "CPU to profile on"),
819 OPT_BOOLEAN('f', "force", &force,
820 "overwrite existing data file (deprecated)"),
821 OPT_U64('c', "count", &user_interval, "event period to sample"),
822 OPT_STRING('o', "output", &output_name, "file",
824 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
825 "child tasks do not inherit counters"),
826 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
827 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
828 OPT_BOOLEAN('g', "call-graph", &call_graph,
829 "do call-graph (stack chain/backtrace) recording"),
830 OPT_INCR('v', "verbose", &verbose,
831 "be more verbose (show counter open errors, etc)"),
832 OPT_BOOLEAN('s', "stat", &inherit_stat,
833 "per thread counts"),
834 OPT_BOOLEAN('d', "data", &sample_address,
836 OPT_BOOLEAN('n', "no-samples", &no_samples,
838 OPT_BOOLEAN('M', "multiplex", &multiplex,
839 "multiplex counter output in a single channel"),
843 int cmd_record(int argc, const char **argv, const char *prefix __used)
847 argc = parse_options(argc, argv, options, record_usage,
848 PARSE_OPT_STOP_AT_NON_OPTION);
849 if (!argc && target_pid == -1 && target_tid == -1 &&
850 !system_wide && profile_cpu == -1)
851 usage_with_options(record_usage, options);
853 if (force && append_file) {
854 fprintf(stderr, "Can't overwrite and append at the same time."
855 " You need to choose between -f and -A");
856 usage_with_options(record_usage, options);
857 } else if (append_file) {
858 write_mode = WRITE_APPEND;
860 write_mode = WRITE_FORCE;
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, options);
880 all_tids=malloc(sizeof(pid_t));
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);
891 mmap_array[i][j] = zalloc(
892 sizeof(struct mmap_data)*thread_num);
893 if (!fd[i][j] || !mmap_array[i][j])
897 event_array = malloc(
898 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
902 if (user_interval != ULLONG_MAX)
903 default_interval = user_interval;
904 if (user_freq != UINT_MAX)
908 * User specified count overrides default frequency.
910 if (default_interval)
913 default_interval = freq;
915 fprintf(stderr, "frequency and count are zero, aborting\n");
919 return __cmd_record(argc, argv);