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/evlist.h"
22 #include "util/evsel.h"
23 #include "util/debug.h"
24 #include "util/session.h"
25 #include "util/symbol.h"
26 #include "util/cpumap.h"
27 #include "util/thread_map.h"
33 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
40 static u64 user_interval = ULLONG_MAX;
41 static u64 default_interval = 0;
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 = NULL;
51 static int realtime_prio = 0;
52 static bool nodelay = false;
53 static bool raw_samples = false;
54 static bool sample_id_all_avail = true;
55 static bool system_wide = false;
56 static pid_t target_pid = -1;
57 static pid_t target_tid = -1;
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;
68 static struct perf_evlist *evsel_list;
70 static long samples = 0;
71 static u64 bytes_written = 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;
79 static void advance_output(size_t size)
81 bytes_written += size;
84 static void write_output(void *buf, size_t size)
87 int ret = write(output, buf, size);
90 die("failed to write");
99 static int process_synthesized_event(union perf_event *event,
100 struct perf_sample *sample __used,
101 struct perf_session *self __used)
103 write_output(event, event->header.size);
107 static void mmap_read(struct perf_mmap *md)
109 unsigned int head = perf_mmap__read_head(md);
110 unsigned int old = md->prev;
111 unsigned char *data = md->base + page_size;
122 if ((old & md->mask) + size != (head & md->mask)) {
123 buf = &data[old & md->mask];
124 size = md->mask + 1 - (old & md->mask);
127 write_output(buf, size);
130 buf = &data[old & md->mask];
134 write_output(buf, size);
137 perf_mmap__write_tail(md, old);
140 static volatile int done = 0;
141 static volatile int signr = -1;
143 static void sig_handler(int sig)
149 static void sig_atexit(void)
152 kill(child_pid, SIGTERM);
154 if (signr == -1 || signr == SIGUSR1)
157 signal(signr, SIG_DFL);
158 kill(getpid(), signr);
161 static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
163 struct perf_event_attr *attr = &evsel->attr;
164 int track = !evsel->idx; /* only the first counter needs these */
166 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
167 PERF_FORMAT_TOTAL_TIME_RUNNING |
170 attr->sample_type |= PERF_SAMPLE_IP | PERF_SAMPLE_TID;
172 if (evlist->nr_entries > 1)
173 attr->sample_type |= PERF_SAMPLE_ID;
176 * We default some events to a 1 default interval. But keep
177 * it a weak assumption overridable by the user.
179 if (!attr->sample_period || (user_freq != UINT_MAX &&
180 user_interval != ULLONG_MAX)) {
182 attr->sample_type |= PERF_SAMPLE_PERIOD;
184 attr->sample_freq = freq;
186 attr->sample_period = default_interval;
191 attr->sample_freq = 0;
194 attr->inherit_stat = 1;
196 if (sample_address) {
197 attr->sample_type |= PERF_SAMPLE_ADDR;
198 attr->mmap_data = track;
202 attr->sample_type |= PERF_SAMPLE_CALLCHAIN;
205 attr->sample_type |= PERF_SAMPLE_CPU;
207 if (sample_id_all_avail &&
208 (sample_time || system_wide || !no_inherit || cpu_list))
209 attr->sample_type |= PERF_SAMPLE_TIME;
212 attr->sample_type |= PERF_SAMPLE_TIME;
213 attr->sample_type |= PERF_SAMPLE_RAW;
214 attr->sample_type |= PERF_SAMPLE_CPU;
219 attr->wakeup_events = 1;
225 if (target_pid == -1 && target_tid == -1 && !system_wide) {
227 attr->enable_on_exec = 1;
231 static bool perf_evlist__equal(struct perf_evlist *evlist,
232 struct perf_evlist *other)
234 struct perf_evsel *pos, *pair;
236 if (evlist->nr_entries != other->nr_entries)
239 pair = list_entry(other->entries.next, struct perf_evsel, node);
241 list_for_each_entry(pos, &evlist->entries, node) {
242 if (memcmp(&pos->attr, &pair->attr, sizeof(pos->attr) != 0))
244 pair = list_entry(pair->node.next, struct perf_evsel, node);
250 static void open_counters(struct perf_evlist *evlist)
252 struct perf_evsel *pos;
254 list_for_each_entry(pos, &evlist->entries, node) {
255 struct perf_event_attr *attr = &pos->attr;
257 * Check if parse_single_tracepoint_event has already asked for
260 * XXX this is kludgy but short term fix for problems introduced by
261 * eac23d1c that broke 'perf script' by having different sample_types
262 * when using multiple tracepoint events when we use a perf binary
263 * that tries to use sample_id_all on an older kernel.
265 * We need to move counter creation to perf_session, support
266 * different sample_types, etc.
268 bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
270 config_attr(pos, evlist);
272 attr->sample_id_all = sample_id_all_avail ? 1 : 0;
274 if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group,
278 if (err == EPERM || err == EACCES) {
279 ui__warning_paranoid();
281 } else if (err == ENODEV && cpu_list) {
282 die("No such device - did you specify"
283 " an out-of-range profile CPU?\n");
284 } else if (err == EINVAL && sample_id_all_avail) {
286 * Old kernel, no attr->sample_id_type_all field
288 sample_id_all_avail = false;
289 if (!sample_time && !raw_samples && !time_needed)
290 attr->sample_type &= ~PERF_SAMPLE_TIME;
292 goto retry_sample_id;
296 * If it's cycles then fall back to hrtimer
297 * based cpu-clock-tick sw counter, which
298 * is always available even if no PMU support:
300 if (attr->type == PERF_TYPE_HARDWARE
301 && attr->config == PERF_COUNT_HW_CPU_CYCLES) {
304 ui__warning("The cycles event is not supported, "
305 "trying to fall back to cpu-clock-ticks\n");
306 attr->type = PERF_TYPE_SOFTWARE;
307 attr->config = PERF_COUNT_SW_CPU_CLOCK;
312 ui__warning("The %s event is not supported.\n",
318 error("sys_perf_event_open() syscall returned with %d (%s). /bin/dmesg may provide additional information.\n",
321 #if defined(__i386__) || defined(__x86_64__)
322 if (attr->type == PERF_TYPE_HARDWARE && err == EOPNOTSUPP)
323 die("No hardware sampling interrupt available."
324 " No APIC? If so then you can boot the kernel"
325 " with the \"lapic\" boot parameter to"
326 " force-enable it.\n");
329 die("No CONFIG_PERF_EVENTS=y kernel support configured?\n");
333 if (perf_evlist__set_filters(evlist)) {
334 error("failed to set filter with %d (%s)\n", errno,
339 if (perf_evlist__mmap(evlist, mmap_pages, false) < 0)
340 die("failed to mmap with %d (%s)\n", errno, strerror(errno));
343 session->evlist = evlist;
345 if (!perf_evlist__equal(session->evlist, evlist)) {
346 fprintf(stderr, "incompatible append\n");
351 perf_session__update_sample_type(session);
354 static int process_buildids(void)
356 u64 size = lseek(output, 0, SEEK_CUR);
361 session->fd = output;
362 return __perf_session__process_events(session, post_processing_offset,
363 size - post_processing_offset,
364 size, &build_id__mark_dso_hit_ops);
367 static void atexit_header(void)
370 session->header.data_size += bytes_written;
374 perf_session__write_header(session, evsel_list, output, true);
375 perf_session__delete(session);
376 perf_evlist__delete(evsel_list);
381 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
384 struct perf_session *psession = data;
386 if (machine__is_host(machine))
390 *As for guest kernel when processing subcommand record&report,
391 *we arrange module mmap prior to guest kernel mmap and trigger
392 *a preload dso because default guest module symbols are loaded
393 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
394 *method is used to avoid symbol missing when the first addr is
395 *in module instead of in guest kernel.
397 err = perf_event__synthesize_modules(process_synthesized_event,
400 pr_err("Couldn't record guest kernel [%d]'s reference"
401 " relocation symbol.\n", machine->pid);
404 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
405 * have no _text sometimes.
407 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
408 psession, machine, "_text");
410 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
414 pr_err("Couldn't record guest kernel [%d]'s reference"
415 " relocation symbol.\n", machine->pid);
418 static struct perf_event_header finished_round_event = {
419 .size = sizeof(struct perf_event_header),
420 .type = PERF_RECORD_FINISHED_ROUND,
423 static void mmap_read_all(void)
427 for (i = 0; i < evsel_list->cpus->nr; i++) {
428 if (evsel_list->mmap[i].base)
429 mmap_read(&evsel_list->mmap[i]);
432 if (perf_header__has_feat(&session->header, HEADER_TRACE_INFO))
433 write_output(&finished_round_event, sizeof(finished_round_event));
436 static int __cmd_record(int argc, const char **argv)
442 unsigned long waking = 0;
443 int child_ready_pipe[2], go_pipe[2];
444 const bool forks = argc > 0;
446 struct machine *machine;
448 page_size = sysconf(_SC_PAGE_SIZE);
451 signal(SIGCHLD, sig_handler);
452 signal(SIGINT, sig_handler);
453 signal(SIGUSR1, sig_handler);
455 if (forks && (pipe(child_ready_pipe) < 0 || pipe(go_pipe) < 0)) {
456 perror("failed to create pipes");
461 if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
464 output_name = "perf.data";
467 if (!strcmp(output_name, "-"))
469 else if (!stat(output_name, &st) && st.st_size) {
470 if (write_mode == WRITE_FORCE) {
471 char oldname[PATH_MAX];
472 snprintf(oldname, sizeof(oldname), "%s.old",
475 rename(output_name, oldname);
477 } else if (write_mode == WRITE_APPEND) {
478 write_mode = WRITE_FORCE;
482 flags = O_CREAT|O_RDWR;
483 if (write_mode == WRITE_APPEND)
489 output = STDOUT_FILENO;
491 output = open(output_name, flags, S_IRUSR | S_IWUSR);
493 perror("failed to create output file");
497 session = perf_session__new(output_name, O_WRONLY,
498 write_mode == WRITE_FORCE, false, NULL);
499 if (session == NULL) {
500 pr_err("Not enough memory for reading perf file header\n");
505 perf_header__set_feat(&session->header, HEADER_BUILD_ID);
508 err = perf_session__read_header(session, output);
510 goto out_delete_session;
513 if (have_tracepoints(&evsel_list->entries))
514 perf_header__set_feat(&session->header, HEADER_TRACE_INFO);
519 perror("failed to fork");
526 close(child_ready_pipe[0]);
528 fcntl(go_pipe[0], F_SETFD, FD_CLOEXEC);
531 * Do a dummy execvp to get the PLT entry resolved,
532 * so we avoid the resolver overhead on the real
535 execvp("", (char **)argv);
538 * Tell the parent we're ready to go
540 close(child_ready_pipe[1]);
543 * Wait until the parent tells us to go.
545 if (read(go_pipe[0], &buf, 1) == -1)
546 perror("unable to read pipe");
548 execvp(argv[0], (char **)argv);
551 kill(getppid(), SIGUSR1);
555 if (!system_wide && target_tid == -1 && target_pid == -1)
556 evsel_list->threads->map[0] = child_pid;
558 close(child_ready_pipe[1]);
561 * wait for child to settle
563 if (read(child_ready_pipe[0], &buf, 1) == -1) {
564 perror("unable to read pipe");
567 close(child_ready_pipe[0]);
570 open_counters(evsel_list);
573 * perf_session__delete(session) will be called at atexit_header()
575 atexit(atexit_header);
578 err = perf_header__write_pipe(output);
581 } else if (file_new) {
582 err = perf_session__write_header(session, evsel_list,
588 post_processing_offset = lseek(output, 0, SEEK_CUR);
591 err = perf_session__synthesize_attrs(session,
592 process_synthesized_event);
594 pr_err("Couldn't synthesize attrs.\n");
598 err = perf_event__synthesize_event_types(process_synthesized_event,
601 pr_err("Couldn't synthesize event_types.\n");
605 if (have_tracepoints(&evsel_list->entries)) {
607 * FIXME err <= 0 here actually means that
608 * there were no tracepoints so its not really
609 * an error, just that we don't need to
610 * synthesize anything. We really have to
611 * return this more properly and also
612 * propagate errors that now are calling die()
614 err = perf_event__synthesize_tracing_data(output, evsel_list,
615 process_synthesized_event,
618 pr_err("Couldn't record tracing data.\n");
625 machine = perf_session__find_host_machine(session);
627 pr_err("Couldn't find native kernel information.\n");
631 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
632 session, machine, "_text");
634 err = perf_event__synthesize_kernel_mmap(process_synthesized_event,
635 session, machine, "_stext");
637 pr_err("Couldn't record kernel reference relocation symbol\n"
638 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
639 "Check /proc/kallsyms permission or run as root.\n");
641 err = perf_event__synthesize_modules(process_synthesized_event,
644 pr_err("Couldn't record kernel module information.\n"
645 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
646 "Check /proc/modules permission or run as root.\n");
649 perf_session__process_machines(session,
650 perf_event__synthesize_guest_os);
653 perf_event__synthesize_thread_map(evsel_list->threads,
654 process_synthesized_event,
657 perf_event__synthesize_threads(process_synthesized_event,
661 struct sched_param param;
663 param.sched_priority = realtime_prio;
664 if (sched_setscheduler(0, SCHED_FIFO, ¶m)) {
665 pr_err("Could not set realtime priority.\n");
682 if (hits == samples) {
685 err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
690 for (i = 0; i < evsel_list->cpus->nr; i++) {
691 struct perf_evsel *pos;
693 list_for_each_entry(pos, &evsel_list->entries, node) {
695 thread < evsel_list->threads->nr;
697 ioctl(FD(pos, i, thread),
698 PERF_EVENT_IOC_DISABLE);
704 if (quiet || signr == SIGUSR1)
707 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
710 * Approximate RIP event size: 24 bytes.
713 "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
714 (double)bytes_written / 1024.0 / 1024.0,
721 perf_session__delete(session);
725 static const char * const record_usage[] = {
726 "perf record [<options>] [<command>]",
727 "perf record [<options>] -- <command> [<options>]",
731 static bool force, append_file;
733 const struct option record_options[] = {
734 OPT_CALLBACK('e', "event", &evsel_list, "event",
735 "event selector. use 'perf list' to list available events",
737 OPT_CALLBACK(0, "filter", &evsel_list, "filter",
738 "event filter", parse_filter),
739 OPT_INTEGER('p', "pid", &target_pid,
740 "record events on existing process id"),
741 OPT_INTEGER('t', "tid", &target_tid,
742 "record events on existing thread id"),
743 OPT_INTEGER('r', "realtime", &realtime_prio,
744 "collect data with this RT SCHED_FIFO priority"),
745 OPT_BOOLEAN('D', "no-delay", &nodelay,
746 "collect data without buffering"),
747 OPT_BOOLEAN('R', "raw-samples", &raw_samples,
748 "collect raw sample records from all opened counters"),
749 OPT_BOOLEAN('a', "all-cpus", &system_wide,
750 "system-wide collection from all CPUs"),
751 OPT_BOOLEAN('A', "append", &append_file,
752 "append to the output file to do incremental profiling"),
753 OPT_STRING('C', "cpu", &cpu_list, "cpu",
754 "list of cpus to monitor"),
755 OPT_BOOLEAN('f', "force", &force,
756 "overwrite existing data file (deprecated)"),
757 OPT_U64('c', "count", &user_interval, "event period to sample"),
758 OPT_STRING('o', "output", &output_name, "file",
760 OPT_BOOLEAN('i', "no-inherit", &no_inherit,
761 "child tasks do not inherit counters"),
762 OPT_UINTEGER('F', "freq", &user_freq, "profile at this frequency"),
763 OPT_UINTEGER('m', "mmap-pages", &mmap_pages, "number of mmap data pages"),
764 OPT_BOOLEAN('g', "call-graph", &call_graph,
765 "do call-graph (stack chain/backtrace) recording"),
766 OPT_INCR('v', "verbose", &verbose,
767 "be more verbose (show counter open errors, etc)"),
768 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
769 OPT_BOOLEAN('s', "stat", &inherit_stat,
770 "per thread counts"),
771 OPT_BOOLEAN('d', "data", &sample_address,
773 OPT_BOOLEAN('T', "timestamp", &sample_time, "Sample timestamps"),
774 OPT_BOOLEAN('n', "no-samples", &no_samples,
776 OPT_BOOLEAN('N', "no-buildid-cache", &no_buildid_cache,
777 "do not update the buildid cache"),
778 OPT_BOOLEAN('B', "no-buildid", &no_buildid,
779 "do not collect buildids in perf.data"),
780 OPT_CALLBACK('G', "cgroup", &evsel_list, "name",
781 "monitor event in cgroup name only",
786 int cmd_record(int argc, const char **argv, const char *prefix __used)
789 struct perf_evsel *pos;
791 evsel_list = perf_evlist__new(NULL, NULL);
792 if (evsel_list == NULL)
795 argc = parse_options(argc, argv, record_options, record_usage,
796 PARSE_OPT_STOP_AT_NON_OPTION);
797 if (!argc && target_pid == -1 && target_tid == -1 &&
798 !system_wide && !cpu_list)
799 usage_with_options(record_usage, record_options);
801 if (force && append_file) {
802 fprintf(stderr, "Can't overwrite and append at the same time."
803 " You need to choose between -f and -A");
804 usage_with_options(record_usage, record_options);
805 } else if (append_file) {
806 write_mode = WRITE_APPEND;
808 write_mode = WRITE_FORCE;
811 if (nr_cgroups && !system_wide) {
812 fprintf(stderr, "cgroup monitoring only available in"
813 " system-wide mode\n");
814 usage_with_options(record_usage, record_options);
819 if (no_buildid_cache || no_buildid)
820 disable_buildid_cache();
822 if (evsel_list->nr_entries == 0 &&
823 perf_evlist__add_default(evsel_list) < 0) {
824 pr_err("Not enough memory for event selector list\n");
825 goto out_symbol_exit;
828 if (target_pid != -1)
829 target_tid = target_pid;
831 if (perf_evlist__create_maps(evsel_list, target_pid,
832 target_tid, cpu_list) < 0)
833 usage_with_options(record_usage, record_options);
835 list_for_each_entry(pos, &evsel_list->entries, node) {
836 if (perf_evsel__alloc_fd(pos, evsel_list->cpus->nr,
837 evsel_list->threads->nr) < 0)
839 if (perf_header__push_event(pos->attr.config, event_name(pos)))
843 if (perf_evlist__alloc_pollfd(evsel_list) < 0)
846 if (user_interval != ULLONG_MAX)
847 default_interval = user_interval;
848 if (user_freq != UINT_MAX)
852 * User specified count overrides default frequency.
854 if (default_interval)
857 default_interval = freq;
859 fprintf(stderr, "frequency and count are zero, aborting\n");
864 err = __cmd_record(argc, argv);
866 perf_evlist__delete_maps(evsel_list);