proc: Add proc_mkdir_data()
[linux-2.6.git] / tools / perf / builtin-record.c
blobf1a939ebc19c5d4aad0afbe1b96201aadbdb0511
1 /*
2 * builtin-record.c
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.
7 */
8 #define _FILE_OFFSET_BITS 64
10 #include "builtin.h"
12 #include "perf.h"
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/tool.h"
26 #include "util/symbol.h"
27 #include "util/cpumap.h"
28 #include "util/thread_map.h"
30 #include <unistd.h>
31 #include <sched.h>
32 #include <sys/mman.h>
34 #ifndef HAVE_ON_EXIT
35 #ifndef ATEXIT_MAX
36 #define ATEXIT_MAX 32
37 #endif
38 static int __on_exit_count = 0;
39 typedef void (*on_exit_func_t) (int, void *);
40 static on_exit_func_t __on_exit_funcs[ATEXIT_MAX];
41 static void *__on_exit_args[ATEXIT_MAX];
42 static int __exitcode = 0;
43 static void __handle_on_exit_funcs(void);
44 static int on_exit(on_exit_func_t function, void *arg);
45 #define exit(x) (exit)(__exitcode = (x))
47 static int on_exit(on_exit_func_t function, void *arg)
49 if (__on_exit_count == ATEXIT_MAX)
50 return -ENOMEM;
51 else if (__on_exit_count == 0)
52 atexit(__handle_on_exit_funcs);
53 __on_exit_funcs[__on_exit_count] = function;
54 __on_exit_args[__on_exit_count++] = arg;
55 return 0;
58 static void __handle_on_exit_funcs(void)
60 int i;
61 for (i = 0; i < __on_exit_count; i++)
62 __on_exit_funcs[i] (__exitcode, __on_exit_args[i]);
64 #endif
66 enum write_mode_t {
67 WRITE_FORCE,
68 WRITE_APPEND
71 struct perf_record {
72 struct perf_tool tool;
73 struct perf_record_opts opts;
74 u64 bytes_written;
75 const char *output_name;
76 struct perf_evlist *evlist;
77 struct perf_session *session;
78 const char *progname;
79 int output;
80 unsigned int page_size;
81 int realtime_prio;
82 enum write_mode_t write_mode;
83 bool no_buildid;
84 bool no_buildid_cache;
85 bool force;
86 bool file_new;
87 bool append_file;
88 long samples;
89 off_t post_processing_offset;
92 static void advance_output(struct perf_record *rec, size_t size)
94 rec->bytes_written += size;
97 static int write_output(struct perf_record *rec, void *buf, size_t size)
99 while (size) {
100 int ret = write(rec->output, buf, size);
102 if (ret < 0) {
103 pr_err("failed to write\n");
104 return -1;
107 size -= ret;
108 buf += ret;
110 rec->bytes_written += ret;
113 return 0;
116 static int process_synthesized_event(struct perf_tool *tool,
117 union perf_event *event,
118 struct perf_sample *sample __maybe_unused,
119 struct machine *machine __maybe_unused)
121 struct perf_record *rec = container_of(tool, struct perf_record, tool);
122 if (write_output(rec, event, event->header.size) < 0)
123 return -1;
125 return 0;
128 static int perf_record__mmap_read(struct perf_record *rec,
129 struct perf_mmap *md)
131 unsigned int head = perf_mmap__read_head(md);
132 unsigned int old = md->prev;
133 unsigned char *data = md->base + rec->page_size;
134 unsigned long size;
135 void *buf;
136 int rc = 0;
138 if (old == head)
139 return 0;
141 rec->samples++;
143 size = head - old;
145 if ((old & md->mask) + size != (head & md->mask)) {
146 buf = &data[old & md->mask];
147 size = md->mask + 1 - (old & md->mask);
148 old += size;
150 if (write_output(rec, buf, size) < 0) {
151 rc = -1;
152 goto out;
156 buf = &data[old & md->mask];
157 size = head - old;
158 old += size;
160 if (write_output(rec, buf, size) < 0) {
161 rc = -1;
162 goto out;
165 md->prev = old;
166 perf_mmap__write_tail(md, old);
168 out:
169 return rc;
172 static volatile int done = 0;
173 static volatile int signr = -1;
174 static volatile int child_finished = 0;
176 static void sig_handler(int sig)
178 if (sig == SIGCHLD)
179 child_finished = 1;
181 done = 1;
182 signr = sig;
185 static void perf_record__sig_exit(int exit_status __maybe_unused, void *arg)
187 struct perf_record *rec = arg;
188 int status;
190 if (rec->evlist->workload.pid > 0) {
191 if (!child_finished)
192 kill(rec->evlist->workload.pid, SIGTERM);
194 wait(&status);
195 if (WIFSIGNALED(status))
196 psignal(WTERMSIG(status), rec->progname);
199 if (signr == -1 || signr == SIGUSR1)
200 return;
202 signal(signr, SIG_DFL);
203 kill(getpid(), signr);
206 static bool perf_evlist__equal(struct perf_evlist *evlist,
207 struct perf_evlist *other)
209 struct perf_evsel *pos, *pair;
211 if (evlist->nr_entries != other->nr_entries)
212 return false;
214 pair = perf_evlist__first(other);
216 list_for_each_entry(pos, &evlist->entries, node) {
217 if (memcmp(&pos->attr, &pair->attr, sizeof(pos->attr) != 0))
218 return false;
219 pair = perf_evsel__next(pair);
222 return true;
225 static int perf_record__open(struct perf_record *rec)
227 char msg[512];
228 struct perf_evsel *pos;
229 struct perf_evlist *evlist = rec->evlist;
230 struct perf_session *session = rec->session;
231 struct perf_record_opts *opts = &rec->opts;
232 int rc = 0;
234 perf_evlist__config(evlist, opts);
236 list_for_each_entry(pos, &evlist->entries, node) {
237 try_again:
238 if (perf_evsel__open(pos, evlist->cpus, evlist->threads) < 0) {
239 if (perf_evsel__fallback(pos, errno, msg, sizeof(msg))) {
240 if (verbose)
241 ui__warning("%s\n", msg);
242 goto try_again;
245 rc = -errno;
246 perf_evsel__open_strerror(pos, &opts->target,
247 errno, msg, sizeof(msg));
248 ui__error("%s\n", msg);
249 goto out;
253 if (perf_evlist__apply_filters(evlist)) {
254 error("failed to set filter with %d (%s)\n", errno,
255 strerror(errno));
256 rc = -1;
257 goto out;
260 if (perf_evlist__mmap(evlist, opts->mmap_pages, false) < 0) {
261 if (errno == EPERM) {
262 pr_err("Permission error mapping pages.\n"
263 "Consider increasing "
264 "/proc/sys/kernel/perf_event_mlock_kb,\n"
265 "or try again with a smaller value of -m/--mmap_pages.\n"
266 "(current value: %d)\n", opts->mmap_pages);
267 rc = -errno;
268 } else if (!is_power_of_2(opts->mmap_pages) &&
269 (opts->mmap_pages != UINT_MAX)) {
270 pr_err("--mmap_pages/-m value must be a power of two.");
271 rc = -EINVAL;
272 } else {
273 pr_err("failed to mmap with %d (%s)\n", errno, strerror(errno));
274 rc = -errno;
276 goto out;
279 if (rec->file_new)
280 session->evlist = evlist;
281 else {
282 if (!perf_evlist__equal(session->evlist, evlist)) {
283 fprintf(stderr, "incompatible append\n");
284 rc = -1;
285 goto out;
289 perf_session__set_id_hdr_size(session);
290 out:
291 return rc;
294 static int process_buildids(struct perf_record *rec)
296 u64 size = lseek(rec->output, 0, SEEK_CUR);
298 if (size == 0)
299 return 0;
301 rec->session->fd = rec->output;
302 return __perf_session__process_events(rec->session, rec->post_processing_offset,
303 size - rec->post_processing_offset,
304 size, &build_id__mark_dso_hit_ops);
307 static void perf_record__exit(int status, void *arg)
309 struct perf_record *rec = arg;
311 if (status != 0)
312 return;
314 if (!rec->opts.pipe_output) {
315 rec->session->header.data_size += rec->bytes_written;
317 if (!rec->no_buildid)
318 process_buildids(rec);
319 perf_session__write_header(rec->session, rec->evlist,
320 rec->output, true);
321 perf_session__delete(rec->session);
322 perf_evlist__delete(rec->evlist);
323 symbol__exit();
327 static void perf_event__synthesize_guest_os(struct machine *machine, void *data)
329 int err;
330 struct perf_tool *tool = data;
332 *As for guest kernel when processing subcommand record&report,
333 *we arrange module mmap prior to guest kernel mmap and trigger
334 *a preload dso because default guest module symbols are loaded
335 *from guest kallsyms instead of /lib/modules/XXX/XXX. This
336 *method is used to avoid symbol missing when the first addr is
337 *in module instead of in guest kernel.
339 err = perf_event__synthesize_modules(tool, process_synthesized_event,
340 machine);
341 if (err < 0)
342 pr_err("Couldn't record guest kernel [%d]'s reference"
343 " relocation symbol.\n", machine->pid);
346 * We use _stext for guest kernel because guest kernel's /proc/kallsyms
347 * have no _text sometimes.
349 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
350 machine, "_text");
351 if (err < 0)
352 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
353 machine, "_stext");
354 if (err < 0)
355 pr_err("Couldn't record guest kernel [%d]'s reference"
356 " relocation symbol.\n", machine->pid);
359 static struct perf_event_header finished_round_event = {
360 .size = sizeof(struct perf_event_header),
361 .type = PERF_RECORD_FINISHED_ROUND,
364 static int perf_record__mmap_read_all(struct perf_record *rec)
366 int i;
367 int rc = 0;
369 for (i = 0; i < rec->evlist->nr_mmaps; i++) {
370 if (rec->evlist->mmap[i].base) {
371 if (perf_record__mmap_read(rec, &rec->evlist->mmap[i]) != 0) {
372 rc = -1;
373 goto out;
378 if (perf_header__has_feat(&rec->session->header, HEADER_TRACING_DATA))
379 rc = write_output(rec, &finished_round_event,
380 sizeof(finished_round_event));
382 out:
383 return rc;
386 static int __cmd_record(struct perf_record *rec, int argc, const char **argv)
388 struct stat st;
389 int flags;
390 int err, output, feat;
391 unsigned long waking = 0;
392 const bool forks = argc > 0;
393 struct machine *machine;
394 struct perf_tool *tool = &rec->tool;
395 struct perf_record_opts *opts = &rec->opts;
396 struct perf_evlist *evsel_list = rec->evlist;
397 const char *output_name = rec->output_name;
398 struct perf_session *session;
399 bool disabled = false;
401 rec->progname = argv[0];
403 rec->page_size = sysconf(_SC_PAGE_SIZE);
405 on_exit(perf_record__sig_exit, rec);
406 signal(SIGCHLD, sig_handler);
407 signal(SIGINT, sig_handler);
408 signal(SIGUSR1, sig_handler);
410 if (!output_name) {
411 if (!fstat(STDOUT_FILENO, &st) && S_ISFIFO(st.st_mode))
412 opts->pipe_output = true;
413 else
414 rec->output_name = output_name = "perf.data";
416 if (output_name) {
417 if (!strcmp(output_name, "-"))
418 opts->pipe_output = true;
419 else if (!stat(output_name, &st) && st.st_size) {
420 if (rec->write_mode == WRITE_FORCE) {
421 char oldname[PATH_MAX];
422 snprintf(oldname, sizeof(oldname), "%s.old",
423 output_name);
424 unlink(oldname);
425 rename(output_name, oldname);
427 } else if (rec->write_mode == WRITE_APPEND) {
428 rec->write_mode = WRITE_FORCE;
432 flags = O_CREAT|O_RDWR;
433 if (rec->write_mode == WRITE_APPEND)
434 rec->file_new = 0;
435 else
436 flags |= O_TRUNC;
438 if (opts->pipe_output)
439 output = STDOUT_FILENO;
440 else
441 output = open(output_name, flags, S_IRUSR | S_IWUSR);
442 if (output < 0) {
443 perror("failed to create output file");
444 return -1;
447 rec->output = output;
449 session = perf_session__new(output_name, O_WRONLY,
450 rec->write_mode == WRITE_FORCE, false, NULL);
451 if (session == NULL) {
452 pr_err("Not enough memory for reading perf file header\n");
453 return -1;
456 rec->session = session;
458 for (feat = HEADER_FIRST_FEATURE; feat < HEADER_LAST_FEATURE; feat++)
459 perf_header__set_feat(&session->header, feat);
461 if (rec->no_buildid)
462 perf_header__clear_feat(&session->header, HEADER_BUILD_ID);
464 if (!have_tracepoints(&evsel_list->entries))
465 perf_header__clear_feat(&session->header, HEADER_TRACING_DATA);
467 if (!rec->opts.branch_stack)
468 perf_header__clear_feat(&session->header, HEADER_BRANCH_STACK);
470 if (!rec->file_new) {
471 err = perf_session__read_header(session, output);
472 if (err < 0)
473 goto out_delete_session;
476 if (forks) {
477 err = perf_evlist__prepare_workload(evsel_list, opts, argv);
478 if (err < 0) {
479 pr_err("Couldn't run the workload!\n");
480 goto out_delete_session;
484 if (perf_record__open(rec) != 0) {
485 err = -1;
486 goto out_delete_session;
489 if (!evsel_list->nr_groups)
490 perf_header__clear_feat(&session->header, HEADER_GROUP_DESC);
493 * perf_session__delete(session) will be called at perf_record__exit()
495 on_exit(perf_record__exit, rec);
497 if (opts->pipe_output) {
498 err = perf_header__write_pipe(output);
499 if (err < 0)
500 goto out_delete_session;
501 } else if (rec->file_new) {
502 err = perf_session__write_header(session, evsel_list,
503 output, false);
504 if (err < 0)
505 goto out_delete_session;
508 if (!rec->no_buildid
509 && !perf_header__has_feat(&session->header, HEADER_BUILD_ID)) {
510 pr_err("Couldn't generate buildids. "
511 "Use --no-buildid to profile anyway.\n");
512 err = -1;
513 goto out_delete_session;
516 rec->post_processing_offset = lseek(output, 0, SEEK_CUR);
518 machine = &session->machines.host;
520 if (opts->pipe_output) {
521 err = perf_event__synthesize_attrs(tool, session,
522 process_synthesized_event);
523 if (err < 0) {
524 pr_err("Couldn't synthesize attrs.\n");
525 goto out_delete_session;
528 err = perf_event__synthesize_event_types(tool, process_synthesized_event,
529 machine);
530 if (err < 0) {
531 pr_err("Couldn't synthesize event_types.\n");
532 goto out_delete_session;
535 if (have_tracepoints(&evsel_list->entries)) {
537 * FIXME err <= 0 here actually means that
538 * there were no tracepoints so its not really
539 * an error, just that we don't need to
540 * synthesize anything. We really have to
541 * return this more properly and also
542 * propagate errors that now are calling die()
544 err = perf_event__synthesize_tracing_data(tool, output, evsel_list,
545 process_synthesized_event);
546 if (err <= 0) {
547 pr_err("Couldn't record tracing data.\n");
548 goto out_delete_session;
550 advance_output(rec, err);
554 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
555 machine, "_text");
556 if (err < 0)
557 err = perf_event__synthesize_kernel_mmap(tool, process_synthesized_event,
558 machine, "_stext");
559 if (err < 0)
560 pr_err("Couldn't record kernel reference relocation symbol\n"
561 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
562 "Check /proc/kallsyms permission or run as root.\n");
564 err = perf_event__synthesize_modules(tool, process_synthesized_event,
565 machine);
566 if (err < 0)
567 pr_err("Couldn't record kernel module information.\n"
568 "Symbol resolution may be skewed if relocation was used (e.g. kexec).\n"
569 "Check /proc/modules permission or run as root.\n");
571 if (perf_guest) {
572 machines__process_guests(&session->machines,
573 perf_event__synthesize_guest_os, tool);
576 if (perf_target__has_task(&opts->target))
577 err = perf_event__synthesize_thread_map(tool, evsel_list->threads,
578 process_synthesized_event,
579 machine);
580 else if (perf_target__has_cpu(&opts->target))
581 err = perf_event__synthesize_threads(tool, process_synthesized_event,
582 machine);
583 else /* command specified */
584 err = 0;
586 if (err != 0)
587 goto out_delete_session;
589 if (rec->realtime_prio) {
590 struct sched_param param;
592 param.sched_priority = rec->realtime_prio;
593 if (sched_setscheduler(0, SCHED_FIFO, &param)) {
594 pr_err("Could not set realtime priority.\n");
595 err = -1;
596 goto out_delete_session;
601 * When perf is starting the traced process, all the events
602 * (apart from group members) have enable_on_exec=1 set,
603 * so don't spoil it by prematurely enabling them.
605 if (!perf_target__none(&opts->target))
606 perf_evlist__enable(evsel_list);
609 * Let the child rip
611 if (forks)
612 perf_evlist__start_workload(evsel_list);
614 for (;;) {
615 int hits = rec->samples;
617 if (perf_record__mmap_read_all(rec) < 0) {
618 err = -1;
619 goto out_delete_session;
622 if (hits == rec->samples) {
623 if (done)
624 break;
625 err = poll(evsel_list->pollfd, evsel_list->nr_fds, -1);
626 waking++;
630 * When perf is starting the traced process, at the end events
631 * die with the process and we wait for that. Thus no need to
632 * disable events in this case.
634 if (done && !disabled && !perf_target__none(&opts->target)) {
635 perf_evlist__disable(evsel_list);
636 disabled = true;
640 if (quiet || signr == SIGUSR1)
641 return 0;
643 fprintf(stderr, "[ perf record: Woken up %ld times to write data ]\n", waking);
646 * Approximate RIP event size: 24 bytes.
648 fprintf(stderr,
649 "[ perf record: Captured and wrote %.3f MB %s (~%" PRIu64 " samples) ]\n",
650 (double)rec->bytes_written / 1024.0 / 1024.0,
651 output_name,
652 rec->bytes_written / 24);
654 return 0;
656 out_delete_session:
657 perf_session__delete(session);
658 return err;
661 #define BRANCH_OPT(n, m) \
662 { .name = n, .mode = (m) }
664 #define BRANCH_END { .name = NULL }
666 struct branch_mode {
667 const char *name;
668 int mode;
671 static const struct branch_mode branch_modes[] = {
672 BRANCH_OPT("u", PERF_SAMPLE_BRANCH_USER),
673 BRANCH_OPT("k", PERF_SAMPLE_BRANCH_KERNEL),
674 BRANCH_OPT("hv", PERF_SAMPLE_BRANCH_HV),
675 BRANCH_OPT("any", PERF_SAMPLE_BRANCH_ANY),
676 BRANCH_OPT("any_call", PERF_SAMPLE_BRANCH_ANY_CALL),
677 BRANCH_OPT("any_ret", PERF_SAMPLE_BRANCH_ANY_RETURN),
678 BRANCH_OPT("ind_call", PERF_SAMPLE_BRANCH_IND_CALL),
679 BRANCH_END
682 static int
683 parse_branch_stack(const struct option *opt, const char *str, int unset)
685 #define ONLY_PLM \
686 (PERF_SAMPLE_BRANCH_USER |\
687 PERF_SAMPLE_BRANCH_KERNEL |\
688 PERF_SAMPLE_BRANCH_HV)
690 uint64_t *mode = (uint64_t *)opt->value;
691 const struct branch_mode *br;
692 char *s, *os = NULL, *p;
693 int ret = -1;
695 if (unset)
696 return 0;
699 * cannot set it twice, -b + --branch-filter for instance
701 if (*mode)
702 return -1;
704 /* str may be NULL in case no arg is passed to -b */
705 if (str) {
706 /* because str is read-only */
707 s = os = strdup(str);
708 if (!s)
709 return -1;
711 for (;;) {
712 p = strchr(s, ',');
713 if (p)
714 *p = '\0';
716 for (br = branch_modes; br->name; br++) {
717 if (!strcasecmp(s, br->name))
718 break;
720 if (!br->name) {
721 ui__warning("unknown branch filter %s,"
722 " check man page\n", s);
723 goto error;
726 *mode |= br->mode;
728 if (!p)
729 break;
731 s = p + 1;
734 ret = 0;
736 /* default to any branch */
737 if ((*mode & ~ONLY_PLM) == 0) {
738 *mode = PERF_SAMPLE_BRANCH_ANY;
740 error:
741 free(os);
742 return ret;
745 #ifdef LIBUNWIND_SUPPORT
746 static int get_stack_size(char *str, unsigned long *_size)
748 char *endptr;
749 unsigned long size;
750 unsigned long max_size = round_down(USHRT_MAX, sizeof(u64));
752 size = strtoul(str, &endptr, 0);
754 do {
755 if (*endptr)
756 break;
758 size = round_up(size, sizeof(u64));
759 if (!size || size > max_size)
760 break;
762 *_size = size;
763 return 0;
765 } while (0);
767 pr_err("callchain: Incorrect stack dump size (max %ld): %s\n",
768 max_size, str);
769 return -1;
771 #endif /* LIBUNWIND_SUPPORT */
773 int record_parse_callchain_opt(const struct option *opt,
774 const char *arg, int unset)
776 struct perf_record_opts *opts = opt->value;
777 char *tok, *name, *saveptr = NULL;
778 char *buf;
779 int ret = -1;
781 /* --no-call-graph */
782 if (unset)
783 return 0;
785 /* We specified default option if none is provided. */
786 BUG_ON(!arg);
788 /* We need buffer that we know we can write to. */
789 buf = malloc(strlen(arg) + 1);
790 if (!buf)
791 return -ENOMEM;
793 strcpy(buf, arg);
795 tok = strtok_r((char *)buf, ",", &saveptr);
796 name = tok ? : (char *)buf;
798 do {
799 /* Framepointer style */
800 if (!strncmp(name, "fp", sizeof("fp"))) {
801 if (!strtok_r(NULL, ",", &saveptr)) {
802 opts->call_graph = CALLCHAIN_FP;
803 ret = 0;
804 } else
805 pr_err("callchain: No more arguments "
806 "needed for -g fp\n");
807 break;
809 #ifdef LIBUNWIND_SUPPORT
810 /* Dwarf style */
811 } else if (!strncmp(name, "dwarf", sizeof("dwarf"))) {
812 const unsigned long default_stack_dump_size = 8192;
814 ret = 0;
815 opts->call_graph = CALLCHAIN_DWARF;
816 opts->stack_dump_size = default_stack_dump_size;
818 tok = strtok_r(NULL, ",", &saveptr);
819 if (tok) {
820 unsigned long size = 0;
822 ret = get_stack_size(tok, &size);
823 opts->stack_dump_size = size;
826 if (!ret)
827 pr_debug("callchain: stack dump size %d\n",
828 opts->stack_dump_size);
829 #endif /* LIBUNWIND_SUPPORT */
830 } else {
831 pr_err("callchain: Unknown -g option "
832 "value: %s\n", arg);
833 break;
836 } while (0);
838 free(buf);
840 if (!ret)
841 pr_debug("callchain: type %d\n", opts->call_graph);
843 return ret;
846 static const char * const record_usage[] = {
847 "perf record [<options>] [<command>]",
848 "perf record [<options>] -- <command> [<options>]",
849 NULL
853 * XXX Ideally would be local to cmd_record() and passed to a perf_record__new
854 * because we need to have access to it in perf_record__exit, that is called
855 * after cmd_record() exits, but since record_options need to be accessible to
856 * builtin-script, leave it here.
858 * At least we don't ouch it in all the other functions here directly.
860 * Just say no to tons of global variables, sigh.
862 static struct perf_record record = {
863 .opts = {
864 .mmap_pages = UINT_MAX,
865 .user_freq = UINT_MAX,
866 .user_interval = ULLONG_MAX,
867 .freq = 4000,
868 .target = {
869 .uses_mmap = true,
872 .write_mode = WRITE_FORCE,
873 .file_new = true,
876 #define CALLCHAIN_HELP "do call-graph (stack chain/backtrace) recording: "
878 #ifdef LIBUNWIND_SUPPORT
879 const char record_callchain_help[] = CALLCHAIN_HELP "[fp] dwarf";
880 #else
881 const char record_callchain_help[] = CALLCHAIN_HELP "[fp]";
882 #endif
885 * XXX Will stay a global variable till we fix builtin-script.c to stop messing
886 * with it and switch to use the library functions in perf_evlist that came
887 * from builtin-record.c, i.e. use perf_record_opts,
888 * perf_evlist__prepare_workload, etc instead of fork+exec'in 'perf record',
889 * using pipes, etc.
891 const struct option record_options[] = {
892 OPT_CALLBACK('e', "event", &record.evlist, "event",
893 "event selector. use 'perf list' to list available events",
894 parse_events_option),
895 OPT_CALLBACK(0, "filter", &record.evlist, "filter",
896 "event filter", parse_filter),
897 OPT_STRING('p', "pid", &record.opts.target.pid, "pid",
898 "record events on existing process id"),
899 OPT_STRING('t', "tid", &record.opts.target.tid, "tid",
900 "record events on existing thread id"),
901 OPT_INTEGER('r', "realtime", &record.realtime_prio,
902 "collect data with this RT SCHED_FIFO priority"),
903 OPT_BOOLEAN('D', "no-delay", &record.opts.no_delay,
904 "collect data without buffering"),
905 OPT_BOOLEAN('R', "raw-samples", &record.opts.raw_samples,
906 "collect raw sample records from all opened counters"),
907 OPT_BOOLEAN('a', "all-cpus", &record.opts.target.system_wide,
908 "system-wide collection from all CPUs"),
909 OPT_BOOLEAN('A', "append", &record.append_file,
910 "append to the output file to do incremental profiling"),
911 OPT_STRING('C', "cpu", &record.opts.target.cpu_list, "cpu",
912 "list of cpus to monitor"),
913 OPT_BOOLEAN('f', "force", &record.force,
914 "overwrite existing data file (deprecated)"),
915 OPT_U64('c', "count", &record.opts.user_interval, "event period to sample"),
916 OPT_STRING('o', "output", &record.output_name, "file",
917 "output file name"),
918 OPT_BOOLEAN('i', "no-inherit", &record.opts.no_inherit,
919 "child tasks do not inherit counters"),
920 OPT_UINTEGER('F', "freq", &record.opts.user_freq, "profile at this frequency"),
921 OPT_UINTEGER('m', "mmap-pages", &record.opts.mmap_pages,
922 "number of mmap data pages"),
923 OPT_BOOLEAN(0, "group", &record.opts.group,
924 "put the counters into a counter group"),
925 OPT_CALLBACK_DEFAULT('g', "call-graph", &record.opts,
926 "mode[,dump_size]", record_callchain_help,
927 &record_parse_callchain_opt, "fp"),
928 OPT_INCR('v', "verbose", &verbose,
929 "be more verbose (show counter open errors, etc)"),
930 OPT_BOOLEAN('q', "quiet", &quiet, "don't print any message"),
931 OPT_BOOLEAN('s', "stat", &record.opts.inherit_stat,
932 "per thread counts"),
933 OPT_BOOLEAN('d', "data", &record.opts.sample_address,
934 "Sample addresses"),
935 OPT_BOOLEAN('T', "timestamp", &record.opts.sample_time, "Sample timestamps"),
936 OPT_BOOLEAN('P', "period", &record.opts.period, "Sample period"),
937 OPT_BOOLEAN('n', "no-samples", &record.opts.no_samples,
938 "don't sample"),
939 OPT_BOOLEAN('N', "no-buildid-cache", &record.no_buildid_cache,
940 "do not update the buildid cache"),
941 OPT_BOOLEAN('B', "no-buildid", &record.no_buildid,
942 "do not collect buildids in perf.data"),
943 OPT_CALLBACK('G', "cgroup", &record.evlist, "name",
944 "monitor event in cgroup name only",
945 parse_cgroups),
946 OPT_STRING('u', "uid", &record.opts.target.uid_str, "user",
947 "user to profile"),
949 OPT_CALLBACK_NOOPT('b', "branch-any", &record.opts.branch_stack,
950 "branch any", "sample any taken branches",
951 parse_branch_stack),
953 OPT_CALLBACK('j', "branch-filter", &record.opts.branch_stack,
954 "branch filter mask", "branch stack filter modes",
955 parse_branch_stack),
956 OPT_END()
959 int cmd_record(int argc, const char **argv, const char *prefix __maybe_unused)
961 int err = -ENOMEM;
962 struct perf_evsel *pos;
963 struct perf_evlist *evsel_list;
964 struct perf_record *rec = &record;
965 char errbuf[BUFSIZ];
967 evsel_list = perf_evlist__new(NULL, NULL);
968 if (evsel_list == NULL)
969 return -ENOMEM;
971 rec->evlist = evsel_list;
973 argc = parse_options(argc, argv, record_options, record_usage,
974 PARSE_OPT_STOP_AT_NON_OPTION);
975 if (!argc && perf_target__none(&rec->opts.target))
976 usage_with_options(record_usage, record_options);
978 if (rec->force && rec->append_file) {
979 ui__error("Can't overwrite and append at the same time."
980 " You need to choose between -f and -A");
981 usage_with_options(record_usage, record_options);
982 } else if (rec->append_file) {
983 rec->write_mode = WRITE_APPEND;
984 } else {
985 rec->write_mode = WRITE_FORCE;
988 if (nr_cgroups && !rec->opts.target.system_wide) {
989 ui__error("cgroup monitoring only available in"
990 " system-wide mode\n");
991 usage_with_options(record_usage, record_options);
994 symbol__init();
996 if (symbol_conf.kptr_restrict)
997 pr_warning(
998 "WARNING: Kernel address maps (/proc/{kallsyms,modules}) are restricted,\n"
999 "check /proc/sys/kernel/kptr_restrict.\n\n"
1000 "Samples in kernel functions may not be resolved if a suitable vmlinux\n"
1001 "file is not found in the buildid cache or in the vmlinux path.\n\n"
1002 "Samples in kernel modules won't be resolved at all.\n\n"
1003 "If some relocation was applied (e.g. kexec) symbols may be misresolved\n"
1004 "even with a suitable vmlinux or kallsyms file.\n\n");
1006 if (rec->no_buildid_cache || rec->no_buildid)
1007 disable_buildid_cache();
1009 if (evsel_list->nr_entries == 0 &&
1010 perf_evlist__add_default(evsel_list) < 0) {
1011 pr_err("Not enough memory for event selector list\n");
1012 goto out_symbol_exit;
1015 err = perf_target__validate(&rec->opts.target);
1016 if (err) {
1017 perf_target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1018 ui__warning("%s", errbuf);
1021 err = perf_target__parse_uid(&rec->opts.target);
1022 if (err) {
1023 int saved_errno = errno;
1025 perf_target__strerror(&rec->opts.target, err, errbuf, BUFSIZ);
1026 ui__error("%s", errbuf);
1028 err = -saved_errno;
1029 goto out_free_fd;
1032 err = -ENOMEM;
1033 if (perf_evlist__create_maps(evsel_list, &rec->opts.target) < 0)
1034 usage_with_options(record_usage, record_options);
1036 list_for_each_entry(pos, &evsel_list->entries, node) {
1037 if (perf_header__push_event(pos->attr.config, perf_evsel__name(pos)))
1038 goto out_free_fd;
1041 if (rec->opts.user_interval != ULLONG_MAX)
1042 rec->opts.default_interval = rec->opts.user_interval;
1043 if (rec->opts.user_freq != UINT_MAX)
1044 rec->opts.freq = rec->opts.user_freq;
1047 * User specified count overrides default frequency.
1049 if (rec->opts.default_interval)
1050 rec->opts.freq = 0;
1051 else if (rec->opts.freq) {
1052 rec->opts.default_interval = rec->opts.freq;
1053 } else {
1054 ui__error("frequency and count are zero, aborting\n");
1055 err = -EINVAL;
1056 goto out_free_fd;
1059 err = __cmd_record(&record, argc, argv);
1060 out_free_fd:
1061 perf_evlist__delete_maps(evsel_list);
1062 out_symbol_exit:
1063 symbol__exit();
1064 return err;