1 #define _FILE_OFFSET_BITS 64
3 #include <linux/kernel.h>
18 static int perf_session__open(struct perf_session *self, bool force)
20 struct stat input_stat;
22 if (!strcmp(self->filename, "-")) {
24 self->fd = STDIN_FILENO;
26 if (perf_session__read_header(self, self->fd) < 0)
27 pr_err("incompatible file format (rerun with -v to learn more)");
32 self->fd = open(self->filename, O_RDONLY);
36 pr_err("failed to open %s: %s", self->filename, strerror(err));
37 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
38 pr_err(" (try 'perf record' first)");
43 if (fstat(self->fd, &input_stat) < 0)
46 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
47 pr_err("file %s not owned by current user or root\n",
52 if (!input_stat.st_size) {
53 pr_info("zero-sized file (%s), nothing to do!\n",
58 if (perf_session__read_header(self, self->fd) < 0) {
59 pr_err("incompatible file format (rerun with -v to learn more)");
63 if (!perf_evlist__valid_sample_type(self->evlist)) {
64 pr_err("non matching sample_type");
68 if (!perf_evlist__valid_sample_id_all(self->evlist)) {
69 pr_err("non matching sample_id_all");
73 self->size = input_stat.st_size;
82 void perf_session__update_sample_type(struct perf_session *self)
84 self->sample_type = perf_evlist__sample_type(self->evlist);
85 self->sample_size = __perf_evsel__sample_size(self->sample_type);
86 self->sample_id_all = perf_evlist__sample_id_all(self->evlist);
87 self->id_hdr_size = perf_evlist__id_hdr_size(self->evlist);
88 self->host_machine.id_hdr_size = self->id_hdr_size;
91 int perf_session__create_kernel_maps(struct perf_session *self)
93 int ret = machine__create_kernel_maps(&self->host_machine);
96 ret = machines__create_guest_kernel_maps(&self->machines);
100 static void perf_session__destroy_kernel_maps(struct perf_session *self)
102 machine__destroy_kernel_maps(&self->host_machine);
103 machines__destroy_guest_kernel_maps(&self->machines);
106 struct perf_session *perf_session__new(const char *filename, int mode,
107 bool force, bool repipe,
108 struct perf_tool *tool)
110 struct perf_session *self;
114 if (!filename || !strlen(filename)) {
115 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
118 filename = "perf.data";
121 len = strlen(filename);
122 self = zalloc(sizeof(*self) + len);
127 memcpy(self->filename, filename, len);
129 * On 64bit we can mmap the data file in one go. No need for tiny mmap
130 * slices. On 32bit we use 32MB.
132 #if BITS_PER_LONG == 64
133 self->mmap_window = ULLONG_MAX;
135 self->mmap_window = 32 * 1024 * 1024ULL;
137 self->machines = RB_ROOT;
138 self->repipe = repipe;
139 INIT_LIST_HEAD(&self->ordered_samples.samples);
140 INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
141 INIT_LIST_HEAD(&self->ordered_samples.to_free);
142 machine__init(&self->host_machine, "", HOST_KERNEL_ID);
143 hists__init(&self->hists);
145 if (mode == O_RDONLY) {
146 if (perf_session__open(self, force) < 0)
148 perf_session__update_sample_type(self);
149 } else if (mode == O_WRONLY) {
151 * In O_RDONLY mode this will be performed when reading the
152 * kernel MMAP event, in perf_event__process_mmap().
154 if (perf_session__create_kernel_maps(self) < 0)
158 if (tool && tool->ordering_requires_timestamps &&
159 tool->ordered_samples && !self->sample_id_all) {
160 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
161 tool->ordered_samples = false;
167 perf_session__delete(self);
171 static void machine__delete_dead_threads(struct machine *machine)
173 struct thread *n, *t;
175 list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
181 static void perf_session__delete_dead_threads(struct perf_session *session)
183 machine__delete_dead_threads(&session->host_machine);
186 static void machine__delete_threads(struct machine *self)
188 struct rb_node *nd = rb_first(&self->threads);
191 struct thread *t = rb_entry(nd, struct thread, rb_node);
193 rb_erase(&t->rb_node, &self->threads);
199 static void perf_session__delete_threads(struct perf_session *session)
201 machine__delete_threads(&session->host_machine);
204 void perf_session__delete(struct perf_session *self)
206 perf_session__destroy_kernel_maps(self);
207 perf_session__delete_dead_threads(self);
208 perf_session__delete_threads(self);
209 machine__exit(&self->host_machine);
214 void machine__remove_thread(struct machine *self, struct thread *th)
216 self->last_match = NULL;
217 rb_erase(&th->rb_node, &self->threads);
219 * We may have references to this thread, for instance in some hist_entry
220 * instances, so just move them to a separate list.
222 list_add_tail(&th->node, &self->dead_threads);
225 static bool symbol__match_parent_regex(struct symbol *sym)
227 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
233 static const u8 cpumodes[] = {
234 PERF_RECORD_MISC_USER,
235 PERF_RECORD_MISC_KERNEL,
236 PERF_RECORD_MISC_GUEST_USER,
237 PERF_RECORD_MISC_GUEST_KERNEL
239 #define NCPUMODES (sizeof(cpumodes)/sizeof(u8))
241 static void ip__resolve_ams(struct machine *self, struct thread *thread,
242 struct addr_map_symbol *ams,
245 struct addr_location al;
249 memset(&al, 0, sizeof(al));
251 for (i = 0; i < NCPUMODES; i++) {
254 * We cannot use the header.misc hint to determine whether a
255 * branch stack address is user, kernel, guest, hypervisor.
256 * Branches may straddle the kernel/user/hypervisor boundaries.
257 * Thus, we have to try consecutively until we find a match
258 * or else, the symbol is unknown
260 thread__find_addr_location(thread, self, m, MAP__FUNCTION,
267 ams->al_addr = al.addr;
272 struct branch_info *machine__resolve_bstack(struct machine *self,
274 struct branch_stack *bs)
276 struct branch_info *bi;
279 bi = calloc(bs->nr, sizeof(struct branch_info));
283 for (i = 0; i < bs->nr; i++) {
284 ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to);
285 ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from);
286 bi[i].flags = bs->entries[i].flags;
291 int machine__resolve_callchain(struct machine *self, struct perf_evsel *evsel,
292 struct thread *thread,
293 struct ip_callchain *chain,
294 struct symbol **parent)
296 u8 cpumode = PERF_RECORD_MISC_USER;
300 callchain_cursor_reset(&evsel->hists.callchain_cursor);
302 for (i = 0; i < chain->nr; i++) {
304 struct addr_location al;
306 if (callchain_param.order == ORDER_CALLEE)
309 ip = chain->ips[chain->nr - i - 1];
311 if (ip >= PERF_CONTEXT_MAX) {
313 case PERF_CONTEXT_HV:
314 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
315 case PERF_CONTEXT_KERNEL:
316 cpumode = PERF_RECORD_MISC_KERNEL; break;
317 case PERF_CONTEXT_USER:
318 cpumode = PERF_RECORD_MISC_USER; break;
326 thread__find_addr_location(thread, self, cpumode,
327 MAP__FUNCTION, ip, &al, NULL);
328 if (al.sym != NULL) {
329 if (sort__has_parent && !*parent &&
330 symbol__match_parent_regex(al.sym))
332 if (!symbol_conf.use_callchain)
336 err = callchain_cursor_append(&evsel->hists.callchain_cursor,
345 static int process_event_synth_tracing_data_stub(union perf_event *event __used,
346 struct perf_session *session __used)
348 dump_printf(": unhandled!\n");
352 static int process_event_synth_attr_stub(union perf_event *event __used,
353 struct perf_evlist **pevlist __used)
355 dump_printf(": unhandled!\n");
359 static int process_event_sample_stub(struct perf_tool *tool __used,
360 union perf_event *event __used,
361 struct perf_sample *sample __used,
362 struct perf_evsel *evsel __used,
363 struct machine *machine __used)
365 dump_printf(": unhandled!\n");
369 static int process_event_stub(struct perf_tool *tool __used,
370 union perf_event *event __used,
371 struct perf_sample *sample __used,
372 struct machine *machine __used)
374 dump_printf(": unhandled!\n");
378 static int process_finished_round_stub(struct perf_tool *tool __used,
379 union perf_event *event __used,
380 struct perf_session *perf_session __used)
382 dump_printf(": unhandled!\n");
386 static int process_event_type_stub(struct perf_tool *tool __used,
387 union perf_event *event __used)
389 dump_printf(": unhandled!\n");
393 static int process_finished_round(struct perf_tool *tool,
394 union perf_event *event,
395 struct perf_session *session);
397 static void perf_tool__fill_defaults(struct perf_tool *tool)
399 if (tool->sample == NULL)
400 tool->sample = process_event_sample_stub;
401 if (tool->mmap == NULL)
402 tool->mmap = process_event_stub;
403 if (tool->comm == NULL)
404 tool->comm = process_event_stub;
405 if (tool->fork == NULL)
406 tool->fork = process_event_stub;
407 if (tool->exit == NULL)
408 tool->exit = process_event_stub;
409 if (tool->lost == NULL)
410 tool->lost = perf_event__process_lost;
411 if (tool->read == NULL)
412 tool->read = process_event_sample_stub;
413 if (tool->throttle == NULL)
414 tool->throttle = process_event_stub;
415 if (tool->unthrottle == NULL)
416 tool->unthrottle = process_event_stub;
417 if (tool->attr == NULL)
418 tool->attr = process_event_synth_attr_stub;
419 if (tool->event_type == NULL)
420 tool->event_type = process_event_type_stub;
421 if (tool->tracing_data == NULL)
422 tool->tracing_data = process_event_synth_tracing_data_stub;
423 if (tool->build_id == NULL)
424 tool->build_id = process_finished_round_stub;
425 if (tool->finished_round == NULL) {
426 if (tool->ordered_samples)
427 tool->finished_round = process_finished_round;
429 tool->finished_round = process_finished_round_stub;
433 void mem_bswap_64(void *src, int byte_size)
437 while (byte_size > 0) {
439 byte_size -= sizeof(u64);
444 static void perf_event__all64_swap(union perf_event *event)
446 struct perf_event_header *hdr = &event->header;
447 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
450 static void perf_event__comm_swap(union perf_event *event)
452 event->comm.pid = bswap_32(event->comm.pid);
453 event->comm.tid = bswap_32(event->comm.tid);
456 static void perf_event__mmap_swap(union perf_event *event)
458 event->mmap.pid = bswap_32(event->mmap.pid);
459 event->mmap.tid = bswap_32(event->mmap.tid);
460 event->mmap.start = bswap_64(event->mmap.start);
461 event->mmap.len = bswap_64(event->mmap.len);
462 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
465 static void perf_event__task_swap(union perf_event *event)
467 event->fork.pid = bswap_32(event->fork.pid);
468 event->fork.tid = bswap_32(event->fork.tid);
469 event->fork.ppid = bswap_32(event->fork.ppid);
470 event->fork.ptid = bswap_32(event->fork.ptid);
471 event->fork.time = bswap_64(event->fork.time);
474 static void perf_event__read_swap(union perf_event *event)
476 event->read.pid = bswap_32(event->read.pid);
477 event->read.tid = bswap_32(event->read.tid);
478 event->read.value = bswap_64(event->read.value);
479 event->read.time_enabled = bswap_64(event->read.time_enabled);
480 event->read.time_running = bswap_64(event->read.time_running);
481 event->read.id = bswap_64(event->read.id);
484 /* exported for swapping attributes in file header */
485 void perf_event__attr_swap(struct perf_event_attr *attr)
487 attr->type = bswap_32(attr->type);
488 attr->size = bswap_32(attr->size);
489 attr->config = bswap_64(attr->config);
490 attr->sample_period = bswap_64(attr->sample_period);
491 attr->sample_type = bswap_64(attr->sample_type);
492 attr->read_format = bswap_64(attr->read_format);
493 attr->wakeup_events = bswap_32(attr->wakeup_events);
494 attr->bp_type = bswap_32(attr->bp_type);
495 attr->bp_addr = bswap_64(attr->bp_addr);
496 attr->bp_len = bswap_64(attr->bp_len);
499 static void perf_event__hdr_attr_swap(union perf_event *event)
503 perf_event__attr_swap(&event->attr.attr);
505 size = event->header.size;
506 size -= (void *)&event->attr.id - (void *)event;
507 mem_bswap_64(event->attr.id, size);
510 static void perf_event__event_type_swap(union perf_event *event)
512 event->event_type.event_type.event_id =
513 bswap_64(event->event_type.event_type.event_id);
516 static void perf_event__tracing_data_swap(union perf_event *event)
518 event->tracing_data.size = bswap_32(event->tracing_data.size);
521 typedef void (*perf_event__swap_op)(union perf_event *event);
523 static perf_event__swap_op perf_event__swap_ops[] = {
524 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
525 [PERF_RECORD_COMM] = perf_event__comm_swap,
526 [PERF_RECORD_FORK] = perf_event__task_swap,
527 [PERF_RECORD_EXIT] = perf_event__task_swap,
528 [PERF_RECORD_LOST] = perf_event__all64_swap,
529 [PERF_RECORD_READ] = perf_event__read_swap,
530 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
531 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
532 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
533 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
534 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
535 [PERF_RECORD_HEADER_MAX] = NULL,
538 struct sample_queue {
541 union perf_event *event;
542 struct list_head list;
545 static void perf_session_free_sample_buffers(struct perf_session *session)
547 struct ordered_samples *os = &session->ordered_samples;
549 while (!list_empty(&os->to_free)) {
550 struct sample_queue *sq;
552 sq = list_entry(os->to_free.next, struct sample_queue, list);
558 static int perf_session_deliver_event(struct perf_session *session,
559 union perf_event *event,
560 struct perf_sample *sample,
561 struct perf_tool *tool,
564 static void flush_sample_queue(struct perf_session *s,
565 struct perf_tool *tool)
567 struct ordered_samples *os = &s->ordered_samples;
568 struct list_head *head = &os->samples;
569 struct sample_queue *tmp, *iter;
570 struct perf_sample sample;
571 u64 limit = os->next_flush;
572 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
573 unsigned idx = 0, progress_next = os->nr_samples / 16;
576 if (!tool->ordered_samples || !limit)
579 list_for_each_entry_safe(iter, tmp, head, list) {
580 if (iter->timestamp > limit)
583 ret = perf_session__parse_sample(s, iter->event, &sample);
585 pr_err("Can't parse sample, err = %d\n", ret);
587 perf_session_deliver_event(s, iter->event, &sample, tool,
590 os->last_flush = iter->timestamp;
591 list_del(&iter->list);
592 list_add(&iter->list, &os->sample_cache);
593 if (++idx >= progress_next) {
594 progress_next += os->nr_samples / 16;
595 ui_progress__update(idx, os->nr_samples,
596 "Processing time ordered events...");
600 if (list_empty(head)) {
601 os->last_sample = NULL;
602 } else if (last_ts <= limit) {
604 list_entry(head->prev, struct sample_queue, list);
611 * When perf record finishes a pass on every buffers, it records this pseudo
613 * We record the max timestamp t found in the pass n.
614 * Assuming these timestamps are monotonic across cpus, we know that if
615 * a buffer still has events with timestamps below t, they will be all
616 * available and then read in the pass n + 1.
617 * Hence when we start to read the pass n + 2, we can safely flush every
618 * events with timestamps below t.
620 * ============ PASS n =================
623 * cnt1 timestamps | cnt2 timestamps
626 * - | 4 <--- max recorded
628 * ============ PASS n + 1 ==============
631 * cnt1 timestamps | cnt2 timestamps
634 * 5 | 7 <---- max recorded
636 * Flush every events below timestamp 4
638 * ============ PASS n + 2 ==============
641 * cnt1 timestamps | cnt2 timestamps
646 * Flush every events below timestamp 7
649 static int process_finished_round(struct perf_tool *tool,
650 union perf_event *event __used,
651 struct perf_session *session)
653 flush_sample_queue(session, tool);
654 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
659 /* The queue is ordered by time */
660 static void __queue_event(struct sample_queue *new, struct perf_session *s)
662 struct ordered_samples *os = &s->ordered_samples;
663 struct sample_queue *sample = os->last_sample;
664 u64 timestamp = new->timestamp;
668 os->last_sample = new;
671 list_add(&new->list, &os->samples);
672 os->max_timestamp = timestamp;
677 * last_sample might point to some random place in the list as it's
678 * the last queued event. We expect that the new event is close to
681 if (sample->timestamp <= timestamp) {
682 while (sample->timestamp <= timestamp) {
683 p = sample->list.next;
684 if (p == &os->samples) {
685 list_add_tail(&new->list, &os->samples);
686 os->max_timestamp = timestamp;
689 sample = list_entry(p, struct sample_queue, list);
691 list_add_tail(&new->list, &sample->list);
693 while (sample->timestamp > timestamp) {
694 p = sample->list.prev;
695 if (p == &os->samples) {
696 list_add(&new->list, &os->samples);
699 sample = list_entry(p, struct sample_queue, list);
701 list_add(&new->list, &sample->list);
705 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
707 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
708 struct perf_sample *sample, u64 file_offset)
710 struct ordered_samples *os = &s->ordered_samples;
711 struct list_head *sc = &os->sample_cache;
712 u64 timestamp = sample->time;
713 struct sample_queue *new;
715 if (!timestamp || timestamp == ~0ULL)
718 if (timestamp < s->ordered_samples.last_flush) {
719 printf("Warning: Timestamp below last timeslice flush\n");
723 if (!list_empty(sc)) {
724 new = list_entry(sc->next, struct sample_queue, list);
725 list_del(&new->list);
726 } else if (os->sample_buffer) {
727 new = os->sample_buffer + os->sample_buffer_idx;
728 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
729 os->sample_buffer = NULL;
731 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
732 if (!os->sample_buffer)
734 list_add(&os->sample_buffer->list, &os->to_free);
735 os->sample_buffer_idx = 2;
736 new = os->sample_buffer + 1;
739 new->timestamp = timestamp;
740 new->file_offset = file_offset;
743 __queue_event(new, s);
748 static void callchain__printf(struct perf_sample *sample)
752 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
754 for (i = 0; i < sample->callchain->nr; i++)
755 printf("..... %2d: %016" PRIx64 "\n",
756 i, sample->callchain->ips[i]);
759 static void branch_stack__printf(struct perf_sample *sample)
763 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
765 for (i = 0; i < sample->branch_stack->nr; i++)
766 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
767 i, sample->branch_stack->entries[i].from,
768 sample->branch_stack->entries[i].to);
771 static void perf_session__print_tstamp(struct perf_session *session,
772 union perf_event *event,
773 struct perf_sample *sample)
775 if (event->header.type != PERF_RECORD_SAMPLE &&
776 !session->sample_id_all) {
777 fputs("-1 -1 ", stdout);
781 if ((session->sample_type & PERF_SAMPLE_CPU))
782 printf("%u ", sample->cpu);
784 if (session->sample_type & PERF_SAMPLE_TIME)
785 printf("%" PRIu64 " ", sample->time);
788 static void dump_event(struct perf_session *session, union perf_event *event,
789 u64 file_offset, struct perf_sample *sample)
794 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
795 file_offset, event->header.size, event->header.type);
800 perf_session__print_tstamp(session, event, sample);
802 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
803 event->header.size, perf_event__name(event->header.type));
806 static void dump_sample(struct perf_session *session, union perf_event *event,
807 struct perf_sample *sample)
812 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
813 event->header.misc, sample->pid, sample->tid, sample->ip,
814 sample->period, sample->addr);
816 if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
817 callchain__printf(sample);
819 if (session->sample_type & PERF_SAMPLE_BRANCH_STACK)
820 branch_stack__printf(sample);
823 static struct machine *
824 perf_session__find_machine_for_cpumode(struct perf_session *session,
825 union perf_event *event)
827 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
829 if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest)
830 return perf_session__find_machine(session, event->ip.pid);
832 return perf_session__find_host_machine(session);
835 static int perf_session_deliver_event(struct perf_session *session,
836 union perf_event *event,
837 struct perf_sample *sample,
838 struct perf_tool *tool,
841 struct perf_evsel *evsel;
842 struct machine *machine;
844 dump_event(session, event, file_offset, sample);
846 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
847 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
849 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
850 * because the tools right now may apply filters, discarding
851 * some of the samples. For consistency, in the future we
852 * should have something like nr_filtered_samples and remove
853 * the sample->period from total_sample_period, etc, KISS for
856 * Also testing against NULL allows us to handle files without
857 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
858 * future probably it'll be a good idea to restrict event
859 * processing via perf_session to files with both set.
861 hists__inc_nr_events(&evsel->hists, event->header.type);
864 machine = perf_session__find_machine_for_cpumode(session, event);
866 switch (event->header.type) {
867 case PERF_RECORD_SAMPLE:
868 dump_sample(session, event, sample);
870 ++session->hists.stats.nr_unknown_id;
873 if (machine == NULL) {
874 ++session->hists.stats.nr_unprocessable_samples;
877 return tool->sample(tool, event, sample, evsel, machine);
878 case PERF_RECORD_MMAP:
879 return tool->mmap(tool, event, sample, machine);
880 case PERF_RECORD_COMM:
881 return tool->comm(tool, event, sample, machine);
882 case PERF_RECORD_FORK:
883 return tool->fork(tool, event, sample, machine);
884 case PERF_RECORD_EXIT:
885 return tool->exit(tool, event, sample, machine);
886 case PERF_RECORD_LOST:
887 if (tool->lost == perf_event__process_lost)
888 session->hists.stats.total_lost += event->lost.lost;
889 return tool->lost(tool, event, sample, machine);
890 case PERF_RECORD_READ:
891 return tool->read(tool, event, sample, evsel, machine);
892 case PERF_RECORD_THROTTLE:
893 return tool->throttle(tool, event, sample, machine);
894 case PERF_RECORD_UNTHROTTLE:
895 return tool->unthrottle(tool, event, sample, machine);
897 ++session->hists.stats.nr_unknown_events;
902 static int perf_session__preprocess_sample(struct perf_session *session,
903 union perf_event *event, struct perf_sample *sample)
905 if (event->header.type != PERF_RECORD_SAMPLE ||
906 !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
909 if (!ip_callchain__valid(sample->callchain, event)) {
910 pr_debug("call-chain problem with event, skipping it.\n");
911 ++session->hists.stats.nr_invalid_chains;
912 session->hists.stats.total_invalid_chains += sample->period;
918 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
919 struct perf_tool *tool, u64 file_offset)
923 dump_event(session, event, file_offset, NULL);
925 /* These events are processed right away */
926 switch (event->header.type) {
927 case PERF_RECORD_HEADER_ATTR:
928 err = tool->attr(event, &session->evlist);
930 perf_session__update_sample_type(session);
932 case PERF_RECORD_HEADER_EVENT_TYPE:
933 return tool->event_type(tool, event);
934 case PERF_RECORD_HEADER_TRACING_DATA:
935 /* setup for reading amidst mmap */
936 lseek(session->fd, file_offset, SEEK_SET);
937 return tool->tracing_data(event, session);
938 case PERF_RECORD_HEADER_BUILD_ID:
939 return tool->build_id(tool, event, session);
940 case PERF_RECORD_FINISHED_ROUND:
941 return tool->finished_round(tool, event, session);
947 static int perf_session__process_event(struct perf_session *session,
948 union perf_event *event,
949 struct perf_tool *tool,
952 struct perf_sample sample;
955 if (session->header.needs_swap &&
956 perf_event__swap_ops[event->header.type])
957 perf_event__swap_ops[event->header.type](event);
959 if (event->header.type >= PERF_RECORD_HEADER_MAX)
962 hists__inc_nr_events(&session->hists, event->header.type);
964 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
965 return perf_session__process_user_event(session, event, tool, file_offset);
968 * For all kernel events we get the sample data
970 ret = perf_session__parse_sample(session, event, &sample);
974 /* Preprocess sample records - precheck callchains */
975 if (perf_session__preprocess_sample(session, event, &sample))
978 if (tool->ordered_samples) {
979 ret = perf_session_queue_event(session, event, &sample,
985 return perf_session_deliver_event(session, event, &sample, tool,
989 void perf_event_header__bswap(struct perf_event_header *self)
991 self->type = bswap_32(self->type);
992 self->misc = bswap_16(self->misc);
993 self->size = bswap_16(self->size);
996 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
998 return machine__findnew_thread(&session->host_machine, pid);
1001 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1003 struct thread *thread = perf_session__findnew(self, 0);
1005 if (thread == NULL || thread__set_comm(thread, "swapper")) {
1006 pr_err("problem inserting idle task.\n");
1013 static void perf_session__warn_about_errors(const struct perf_session *session,
1014 const struct perf_tool *tool)
1016 if (tool->lost == perf_event__process_lost &&
1017 session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1018 ui__warning("Processed %d events and lost %d chunks!\n\n"
1019 "Check IO/CPU overload!\n\n",
1020 session->hists.stats.nr_events[0],
1021 session->hists.stats.nr_events[PERF_RECORD_LOST]);
1024 if (session->hists.stats.nr_unknown_events != 0) {
1025 ui__warning("Found %u unknown events!\n\n"
1026 "Is this an older tool processing a perf.data "
1027 "file generated by a more recent tool?\n\n"
1028 "If that is not the case, consider "
1029 "reporting to linux-kernel@vger.kernel.org.\n\n",
1030 session->hists.stats.nr_unknown_events);
1033 if (session->hists.stats.nr_unknown_id != 0) {
1034 ui__warning("%u samples with id not present in the header\n",
1035 session->hists.stats.nr_unknown_id);
1038 if (session->hists.stats.nr_invalid_chains != 0) {
1039 ui__warning("Found invalid callchains!\n\n"
1040 "%u out of %u events were discarded for this reason.\n\n"
1041 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1042 session->hists.stats.nr_invalid_chains,
1043 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1046 if (session->hists.stats.nr_unprocessable_samples != 0) {
1047 ui__warning("%u unprocessable samples recorded.\n"
1048 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1049 session->hists.stats.nr_unprocessable_samples);
1053 #define session_done() (*(volatile int *)(&session_done))
1054 volatile int session_done;
1056 static int __perf_session__process_pipe_events(struct perf_session *self,
1057 struct perf_tool *tool)
1059 union perf_event event;
1066 perf_tool__fill_defaults(tool);
1070 err = readn(self->fd, &event, sizeof(struct perf_event_header));
1075 pr_err("failed to read event header\n");
1079 if (self->header.needs_swap)
1080 perf_event_header__bswap(&event.header);
1082 size = event.header.size;
1087 p += sizeof(struct perf_event_header);
1089 if (size - sizeof(struct perf_event_header)) {
1090 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1093 pr_err("unexpected end of event stream\n");
1097 pr_err("failed to read event data\n");
1102 if ((skip = perf_session__process_event(self, &event, tool, head)) < 0) {
1103 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1104 head, event.header.size, event.header.type);
1106 * assume we lost track of the stream, check alignment, and
1107 * increment a single u64 in the hope to catch on again 'soon'.
1109 if (unlikely(head & 7))
1120 if (!session_done())
1125 perf_session__warn_about_errors(self, tool);
1126 perf_session_free_sample_buffers(self);
1130 static union perf_event *
1131 fetch_mmaped_event(struct perf_session *session,
1132 u64 head, size_t mmap_size, char *buf)
1134 union perf_event *event;
1137 * Ensure we have enough space remaining to read
1138 * the size of the event in the headers.
1140 if (head + sizeof(event->header) > mmap_size)
1143 event = (union perf_event *)(buf + head);
1145 if (session->header.needs_swap)
1146 perf_event_header__bswap(&event->header);
1148 if (head + event->header.size > mmap_size)
1154 int __perf_session__process_events(struct perf_session *session,
1155 u64 data_offset, u64 data_size,
1156 u64 file_size, struct perf_tool *tool)
1158 u64 head, page_offset, file_offset, file_pos, progress_next;
1159 int err, mmap_prot, mmap_flags, map_idx = 0;
1160 size_t page_size, mmap_size;
1161 char *buf, *mmaps[8];
1162 union perf_event *event;
1165 perf_tool__fill_defaults(tool);
1167 page_size = sysconf(_SC_PAGESIZE);
1169 page_offset = page_size * (data_offset / page_size);
1170 file_offset = page_offset;
1171 head = data_offset - page_offset;
1173 if (data_offset + data_size < file_size)
1174 file_size = data_offset + data_size;
1176 progress_next = file_size / 16;
1178 mmap_size = session->mmap_window;
1179 if (mmap_size > file_size)
1180 mmap_size = file_size;
1182 memset(mmaps, 0, sizeof(mmaps));
1184 mmap_prot = PROT_READ;
1185 mmap_flags = MAP_SHARED;
1187 if (session->header.needs_swap) {
1188 mmap_prot |= PROT_WRITE;
1189 mmap_flags = MAP_PRIVATE;
1192 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1194 if (buf == MAP_FAILED) {
1195 pr_err("failed to mmap file\n");
1199 mmaps[map_idx] = buf;
1200 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1201 file_pos = file_offset + head;
1204 event = fetch_mmaped_event(session, head, mmap_size, buf);
1206 if (mmaps[map_idx]) {
1207 munmap(mmaps[map_idx], mmap_size);
1208 mmaps[map_idx] = NULL;
1211 page_offset = page_size * (head / page_size);
1212 file_offset += page_offset;
1213 head -= page_offset;
1217 size = event->header.size;
1220 perf_session__process_event(session, event, tool, file_pos) < 0) {
1221 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1222 file_offset + head, event->header.size,
1223 event->header.type);
1225 * assume we lost track of the stream, check alignment, and
1226 * increment a single u64 in the hope to catch on again 'soon'.
1228 if (unlikely(head & 7))
1237 if (file_pos >= progress_next) {
1238 progress_next += file_size / 16;
1239 ui_progress__update(file_pos, file_size,
1240 "Processing events...");
1243 if (file_pos < file_size)
1247 /* do the final flush for ordered samples */
1248 session->ordered_samples.next_flush = ULLONG_MAX;
1249 flush_sample_queue(session, tool);
1251 perf_session__warn_about_errors(session, tool);
1252 perf_session_free_sample_buffers(session);
1256 int perf_session__process_events(struct perf_session *self,
1257 struct perf_tool *tool)
1261 if (perf_session__register_idle_thread(self) == NULL)
1265 err = __perf_session__process_events(self,
1266 self->header.data_offset,
1267 self->header.data_size,
1270 err = __perf_session__process_pipe_events(self, tool);
1275 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1277 if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1278 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1285 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1286 const char *symbol_name, u64 addr)
1290 struct ref_reloc_sym *ref;
1292 ref = zalloc(sizeof(struct ref_reloc_sym));
1296 ref->name = strdup(symbol_name);
1297 if (ref->name == NULL) {
1302 bracket = strchr(ref->name, ']');
1308 for (i = 0; i < MAP__NR_TYPES; ++i) {
1309 struct kmap *kmap = map__kmap(maps[i]);
1310 kmap->ref_reloc_sym = ref;
1316 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1318 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1319 __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1320 machines__fprintf_dsos(&self->machines, fp);
1323 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1326 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1327 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1330 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1332 struct perf_evsel *pos;
1333 size_t ret = fprintf(fp, "Aggregated stats:\n");
1335 ret += hists__fprintf_nr_events(&session->hists, fp);
1337 list_for_each_entry(pos, &session->evlist->entries, node) {
1338 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1339 ret += hists__fprintf_nr_events(&pos->hists, fp);
1345 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1348 * FIXME: Here we have to actually print all the machines in this
1349 * session, not just the host...
1351 return machine__fprintf(&session->host_machine, fp);
1354 void perf_session__remove_thread(struct perf_session *session,
1358 * FIXME: This one makes no sense, we need to remove the thread from
1359 * the machine it belongs to, perf_session can have many machines, so
1360 * doing it always on ->host_machine is wrong. Fix when auditing all
1361 * the 'perf kvm' code.
1363 machine__remove_thread(&session->host_machine, th);
1366 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1369 struct perf_evsel *pos;
1371 list_for_each_entry(pos, &session->evlist->entries, node) {
1372 if (pos->attr.type == type)
1378 void perf_event__print_ip(union perf_event *event, struct perf_sample *sample,
1379 struct machine *machine, struct perf_evsel *evsel,
1380 int print_sym, int print_dso, int print_symoffset)
1382 struct addr_location al;
1383 struct callchain_cursor *cursor = &evsel->hists.callchain_cursor;
1384 struct callchain_cursor_node *node;
1386 if (perf_event__preprocess_sample(event, machine, &al, sample,
1388 error("problem processing %d event, skipping it.\n",
1389 event->header.type);
1393 if (symbol_conf.use_callchain && sample->callchain) {
1395 if (machine__resolve_callchain(machine, evsel, al.thread,
1396 sample->callchain, NULL) != 0) {
1398 error("Failed to resolve callchain. Skipping\n");
1401 callchain_cursor_commit(cursor);
1404 node = callchain_cursor_current(cursor);
1408 printf("\t%16" PRIx64, node->ip);
1411 symbol__fprintf_symname(node->sym, stdout);
1415 map__fprintf_dsoname(al.map, stdout);
1420 callchain_cursor_advance(cursor);
1424 printf("%16" PRIx64, sample->ip);
1427 if (print_symoffset)
1428 symbol__fprintf_symname_offs(al.sym, &al,
1431 symbol__fprintf_symname(al.sym, stdout);
1436 map__fprintf_dsoname(al.map, stdout);
1442 int perf_session__cpu_bitmap(struct perf_session *session,
1443 const char *cpu_list, unsigned long *cpu_bitmap)
1446 struct cpu_map *map;
1448 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1449 struct perf_evsel *evsel;
1451 evsel = perf_session__find_first_evtype(session, i);
1455 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1456 pr_err("File does not contain CPU events. "
1457 "Remove -c option to proceed.\n");
1462 map = cpu_map__new(cpu_list);
1464 pr_err("Invalid cpu_list\n");
1468 for (i = 0; i < map->nr; i++) {
1469 int cpu = map->map[i];
1471 if (cpu >= MAX_NR_CPUS) {
1472 pr_err("Requested CPU %d too large. "
1473 "Consider raising MAX_NR_CPUS\n", cpu);
1477 set_bit(cpu, cpu_bitmap);
1483 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1489 if (session == NULL || fp == NULL)
1492 ret = fstat(session->fd, &st);
1496 fprintf(fp, "# ========\n");
1497 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1498 perf_header__fprintf_info(session, fp, full);
1499 fprintf(fp, "# ========\n#\n");