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);
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;
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 goto retry_sample_id;
334 * If it's cycles then fall back to hrtimer
335 * based cpu-clock-tick sw counter, which
336 * is always available even if no PMU support:
338 if (attr->type == PERF_TYPE_HARDWARE
339 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
342 warning(" ... trying to fall back to cpu-clock-ticks\n");
343 attr->type = PERF_TYPE_SOFTWARE;
344 attr->config = PERF_COUNT_SW_CPU_CLOCK;
348 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
349 fd[nr_cpu][counter][thread_index], strerror(err));
351 #if defined(__i386__) || defined(__x86_64__)
352 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
353 die("No hardware sampling interrupt available."
354 " No APIC? If so then you can boot the kernel"
355 " with the \"lapic\" boot parameter to"
356 " force-enable it.\n");
359 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
363 h_attr = get_header_attr(attr, counter);
368 if (memcmp(&h_attr->attr, attr, sizeof(*attr))) {
369 fprintf(stderr, "incompatible append\n");
374 if (read(fd[nr_cpu][counter][thread_index], &read_data, sizeof(read_data)) == -1) {
375 perror("Unable to read perf file descriptor");
379 if (perf_header_attr__add_id(h_attr, read_data.id) < 0) {
380 pr_warning("Not enough memory to add id\n");
384 assert(fd[nr_cpu][counter][thread_index] >= 0);
385 fcntl(fd[nr_cpu][counter][thread_index], F_SETFL, O_NONBLOCK);
388 * First counter acts as the group leader:
390 if (group && group_fd == -1)
391 group_fd = fd[nr_cpu][counter][thread_index];
393 if (counter || thread_index) {
394 ret = ioctl(fd[nr_cpu][counter][thread_index],
395 PERF_EVENT_IOC_SET_OUTPUT,
398 error("failed to set output: %d (%s)\n", errno,
403 mmap_array[nr_cpu].counter = counter;
404 mmap_array[nr_cpu].prev = 0;
405 mmap_array[nr_cpu].mask = mmap_pages*page_size - 1;
406 mmap_array[nr_cpu].base = mmap(NULL, (mmap_pages+1)*page_size,
407 PROT_READ|PROT_WRITE, MAP_SHARED, fd[nr_cpu][counter][thread_index], 0);
408 if (mmap_array[nr_cpu].base == MAP_FAILED) {
409 error("failed to mmap with %d (%s)\n", errno, strerror(errno));
413 event_array[nr_poll].fd = fd[nr_cpu][counter][thread_index];
414 event_array[nr_poll].events = POLLIN;
418 if (filter != NULL) {
419 ret = ioctl(fd[nr_cpu][counter][thread_index],
420 PERF_EVENT_IOC_SET_FILTER, filter);
422 error("failed to set filter with %d (%s)\n", errno,
430 static void open_counters(int cpu)
435 for (counter = 0; counter < nr_counters; counter++)
436 create_counter(counter, cpu);
441 static int process_buildids(void)
443 u64 size = lseek(output, 0, SEEK_CUR);
448 session->fd = output;
449 return __perf_session__process_events(session, post_processing_offset,
450 size - post_processing_offset,
451 size, &build_id__mark_dso_hit_ops);
454 static void atexit_header(void)
457 session->header.data_size += bytes_written;
461 perf_header__write(&session->header, output, true);
462 perf_session__delete(session);
467 static void event__synthesize_guest_os(struct machine *machine, void *data)
470 struct perf_session *psession = data;
472 if (machine__is_host(machine))
476 *As for guest kernel when processing subcommand record&report,
477 *we arrange module mmap prior to guest kernel mmap and trigger
478 *a preload dso because default guest module symbols are loaded
479 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
480 *method is used to avoid symbol missing when the first addr is
481 *in module instead of in guest kernel.
483 err = event__synthesize_modules(process_synthesized_event,
486 pr_err("Couldn't record guest kernel [%d]'s reference"
487 " relocation symbol.\n", machine->pid);
490 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
491 * have no _text sometimes.
493 err = event__synthesize_kernel_mmap(process_synthesized_event,
494 psession, machine, "_text");
496 err = event__synthesize_kernel_mmap(process_synthesized_event,
497 psession, machine, "_stext");
499 pr_err("Couldn't record guest kernel [%d]'s reference"
500 " relocation symbol.\n", machine->pid);
503 static struct perf_event_header finished_round_event = {
504 .size = sizeof(struct perf_event_header),
505 .type = PERF_RECORD_FINISHED_ROUND,
508 static void mmap_read_all(void)
512 for (i = 0; i < nr_cpu; i++) {
513 if (mmap_array[i].base)
514 mmap_read(&mmap_array[i]);
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)
527 unsigned long waking = 0;
528 int child_ready_pipe[2], go_pipe[2];
529 const bool forks = argc > 0;
531 struct machine *machine;
533 page_size = sysconf(_SC_PAGE_SIZE);
536 signal(SIGCHLD, sig_handler);
537 signal(SIGINT, sig_handler);
539 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
540 perror("failed to create pipes");
544 if (!strcmp(output_name, "-"))
546 else if (!stat(output_name, &st) && st.st_size) {
547 if (write_mode == WRITE_FORCE) {
548 char oldname[PATH_MAX];
549 snprintf(oldname, sizeof(oldname), "%s.old",
552 rename(output_name, oldname);
554 } else if (write_mode == WRITE_APPEND) {
555 write_mode = WRITE_FORCE;
558 flags = O_CREAT|O_RDWR;
559 if (write_mode == WRITE_APPEND)
565 output = STDOUT_FILENO;
567 output = open(output_name, flags, S_IRUSR | S_IWUSR);
569 perror("failed to create output file");
573 session = perf_session__new(output_name, O_WRONLY,
574 write_mode == WRITE_FORCE, false);
575 if (session == NULL) {
576 pr_err("Not enough memory for reading perf file header\n");
581 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
584 err = perf_header__read(session, output);
586 goto out_delete_session;
589 if (have_tracepoints(attrs, nr_counters))
590 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
593 * perf_session__delete(session) will be called at atexit_header()
595 atexit(atexit_header);
600 perror("failed to fork");
607 close(child_ready_pipe[0]);
609 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
612 * Do a dummy execvp to get the PLT entry resolved,
613 * so we avoid the resolver overhead on the real
616 execvp("", (char **)argv);
619 * Tell the parent we're ready to go
621 close(child_ready_pipe[1]);
624 * Wait until the parent tells us to go.
626 if (read(go_pipe[0], &buf, 1) == -1)
627 perror("unable to read pipe");
629 execvp(argv[0], (char **)argv);
635 if (!system_wide && target_tid == -1 && target_pid == -1)
636 all_tids[0] = child_pid;
638 close(child_ready_pipe[1]);
641 * wait for child to settle
643 if (read(child_ready_pipe[0], &buf, 1) == -1) {
644 perror("unable to read pipe");
647 close(child_ready_pipe[0]);
650 nr_cpus = read_cpu_map(cpu_list);
652 perror("failed to collect number of CPUs");
656 if (!system_wide && no_inherit && !cpu_list) {
659 for (i = 0; i < nr_cpus; i++)
660 open_counters(cpumap[i]);
663 perf_session__set_sample_type(session, sample_type);
666 err = perf_header__write_pipe(output);
669 } else if (file_new) {
670 err = perf_header__write(&session->header, output, false);
675 post_processing_offset = lseek(output, 0, SEEK_CUR);
677 perf_session__set_sample_id_all(session, sample_id_all_avail);
680 err = event__synthesize_attrs(&session->header,
681 process_synthesized_event,
684 pr_err("Couldn't synthesize attrs.\n");
688 err = event__synthesize_event_types(process_synthesized_event,
691 pr_err("Couldn't synthesize event_types.\n");
695 if (have_tracepoints(attrs, nr_counters)) {
697 * FIXME err <= 0 here actually means that
698 * there were no tracepoints so its not really
699 * an error, just that we don't need to
700 * synthesize anything. We really have to
701 * return this more properly and also
702 * propagate errors that now are calling die()
704 err = event__synthesize_tracing_data(output, attrs,
706 process_synthesized_event,
709 pr_err("Couldn't record tracing data.\n");
716 machine = perf_session__find_host_machine(session);
718 pr_err("Couldn't find native kernel information.\n");
722 err = event__synthesize_kernel_mmap(process_synthesized_event,
723 session, machine, "_text");
725 err = event__synthesize_kernel_mmap(process_synthesized_event,
726 session, machine, "_stext");
728 pr_err("Couldn't record kernel reference relocation symbol\n"
729 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
730 "Check /proc/kallsyms permission or run as root.\n");
732 err = event__synthesize_modules(process_synthesized_event,
735 pr_err("Couldn't record kernel module information.\n"
736 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
737 "Check /proc/modules permission or run as root.\n");
740 perf_session__process_machines(session, event__synthesize_guest_os);
743 event__synthesize_thread(target_tid, process_synthesized_event,
746 event__synthesize_threads(process_synthesized_event, session);
749 struct sched_param param;
751 param.sched_priority = realtime_prio;
752 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
753 pr_err("Could not set realtime priority.\n");
770 if (hits == samples) {
773 err = poll(event_array, nr_poll, -1);
778 for (i = 0; i < nr_cpu; i++) {
780 counter < nr_counters;
785 ioctl(fd[i][counter][thread],
786 PERF_EVENT_IOC_DISABLE);
795 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
798 * Approximate RIP event size: 24 bytes.
801 "[ perf record: Captured and wrote %.3f MB %s (~%lld samples) ]\n",
802 (double)bytes_written / 1024.0 / 1024.0,
809 perf_session__delete(session);
813 static const char * const record_usage[] = {
814 "perf record [<options>] [<command>]",
815 "perf record [<options>] -- <command> [<options>]",
819 static bool force, append_file;
821 const struct option record_options[] = {
822 OPT_CALLBACK('e', "event", NULL, "event",
823 "event selector. use 'perf list' to list available events",
825 OPT_CALLBACK(0, "filter", NULL, "filter",
826 "event filter", parse_filter),
827 OPT_INTEGER('p', "pid", &target_pid,
828 "record events on existing process id"),
829 OPT_INTEGER('t', "tid", &target_tid,
830 "record events on existing thread id"),
831 OPT_INTEGER('r', "realtime", &realtime_prio,
832 "collect data with this RT SCHED_FIFO priority"),
833 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
834 "collect raw sample records from all opened counters"),
835 OPT_BOOLEAN('a', "all-cpus", &system_wide,
836 "system-wide collection from all CPUs"),
837 OPT_BOOLEAN('A', "append", &append_file,
838 "append to the output file to do incremental profiling"),
839 OPT_STRING('C', "cpu", &cpu_list, "cpu",
840 "list of cpus to monitor"),
841 OPT_BOOLEAN('f', "force", &force,
842 "overwrite existing data file (deprecated)"),
843 OPT_U64('c', "count", &user_interval, "event period to sample"),
844 OPT_STRING('o', "output", &output_name, "file",
846 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
847 "child tasks do not inherit counters"),
848 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
849 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
850 OPT_BOOLEAN('g', "call-graph", &call_graph,
851 "do call-graph (stack chain/backtrace) recording"),
852 OPT_INCR('v', "verbose", &verbose,
853 "be more verbose (show counter open errors, etc)"),
854 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
855 OPT_BOOLEAN('s', "stat", &inherit_stat,
856 "per thread counts"),
857 OPT_BOOLEAN('d', "data", &sample_address,
859 OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
860 OPT_BOOLEAN('n', "no-samples", &no_samples,
862 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
863 "do not update the buildid cache"),
864 OPT_BOOLEAN('B', "no-buildid", &no_buildid,
865 "do not collect buildids in perf.data"),
869 int cmd_record(int argc, const char **argv, const char *prefix __used)
871 int i, j, err = -ENOMEM;
873 argc = parse_options(argc, argv, record_options, record_usage,
874 PARSE_OPT_STOP_AT_NON_OPTION);
875 if (!argc && target_pid == -1 && target_tid == -1 &&
876 !system_wide && !cpu_list)
877 usage_with_options(record_usage, record_options);
879 if (force && append_file) {
880 fprintf(stderr, "Can't overwrite and append at the same time."
881 " You need to choose between -f and -A");
882 usage_with_options(record_usage, record_options);
883 } else if (append_file) {
884 write_mode = WRITE_APPEND;
886 write_mode = WRITE_FORCE;
891 if (no_buildid_cache || no_buildid)
892 disable_buildid_cache();
896 attrs[0].type = PERF_TYPE_HARDWARE;
897 attrs[0].config = PERF_COUNT_HW_CPU_CYCLES;
900 if (target_pid != -1) {
901 target_tid = target_pid;
902 thread_num = find_all_tid(target_pid, &all_tids);
903 if (thread_num <= 0) {
904 fprintf(stderr, "Can't find all threads of pid %d\n",
906 usage_with_options(record_usage, record_options);
909 all_tids=malloc(sizeof(pid_t));
911 goto out_symbol_exit;
913 all_tids[0] = target_tid;
917 for (i = 0; i < MAX_NR_CPUS; i++) {
918 for (j = 0; j < MAX_COUNTERS; j++) {
919 fd[i][j] = malloc(sizeof(int)*thread_num);
924 event_array = malloc(
925 sizeof(struct pollfd)*MAX_NR_CPUS*MAX_COUNTERS*thread_num);
929 if (user_interval != ULLONG_MAX)
930 default_interval = user_interval;
931 if (user_freq != UINT_MAX)
935 * User specified count overrides default frequency.
937 if (default_interval)
940 default_interval = freq;
942 fprintf(stderr, "frequency and count are zero, aborting\n");
944 goto out_free_event_array;
947 err = __cmd_record(argc, argv);
949 out_free_event_array:
952 for (i = 0; i < MAX_NR_CPUS; i++) {
953 for (j = 0; j < MAX_COUNTERS; j++)