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,
292 struct thread *thread,
293 struct ip_callchain *chain,
294 struct symbol **parent)
296 u8 cpumode = PERF_RECORD_MISC_USER;
300 callchain_cursor_reset(&callchain_cursor);
302 if (chain->nr > PERF_MAX_STACK_DEPTH) {
303 pr_warning("corrupted callchain. skipping...\n");
307 for (i = 0; i < chain->nr; i++) {
309 struct addr_location al;
311 if (callchain_param.order == ORDER_CALLEE)
314 ip = chain->ips[chain->nr - i - 1];
316 if (ip >= PERF_CONTEXT_MAX) {
318 case PERF_CONTEXT_HV:
319 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
320 case PERF_CONTEXT_KERNEL:
321 cpumode = PERF_RECORD_MISC_KERNEL; break;
322 case PERF_CONTEXT_USER:
323 cpumode = PERF_RECORD_MISC_USER; break;
325 pr_debug("invalid callchain context: "
326 "%"PRId64"\n", (s64) ip);
328 * It seems the callchain is corrupted.
331 callchain_cursor_reset(&callchain_cursor);
338 thread__find_addr_location(thread, self, cpumode,
339 MAP__FUNCTION, ip, &al, NULL);
340 if (al.sym != NULL) {
341 if (sort__has_parent && !*parent &&
342 symbol__match_parent_regex(al.sym))
344 if (!symbol_conf.use_callchain)
348 err = callchain_cursor_append(&callchain_cursor,
357 static int process_event_synth_tracing_data_stub(union perf_event *event __used,
358 struct perf_session *session __used)
360 dump_printf(": unhandled!\n");
364 static int process_event_synth_attr_stub(union perf_event *event __used,
365 struct perf_evlist **pevlist __used)
367 dump_printf(": unhandled!\n");
371 static int process_event_sample_stub(struct perf_tool *tool __used,
372 union perf_event *event __used,
373 struct perf_sample *sample __used,
374 struct perf_evsel *evsel __used,
375 struct machine *machine __used)
377 dump_printf(": unhandled!\n");
381 static int process_event_stub(struct perf_tool *tool __used,
382 union perf_event *event __used,
383 struct perf_sample *sample __used,
384 struct machine *machine __used)
386 dump_printf(": unhandled!\n");
390 static int process_finished_round_stub(struct perf_tool *tool __used,
391 union perf_event *event __used,
392 struct perf_session *perf_session __used)
394 dump_printf(": unhandled!\n");
398 static int process_event_type_stub(struct perf_tool *tool __used,
399 union perf_event *event __used)
401 dump_printf(": unhandled!\n");
405 static int process_finished_round(struct perf_tool *tool,
406 union perf_event *event,
407 struct perf_session *session);
409 static void perf_tool__fill_defaults(struct perf_tool *tool)
411 if (tool->sample == NULL)
412 tool->sample = process_event_sample_stub;
413 if (tool->mmap == NULL)
414 tool->mmap = process_event_stub;
415 if (tool->comm == NULL)
416 tool->comm = process_event_stub;
417 if (tool->fork == NULL)
418 tool->fork = process_event_stub;
419 if (tool->exit == NULL)
420 tool->exit = process_event_stub;
421 if (tool->lost == NULL)
422 tool->lost = perf_event__process_lost;
423 if (tool->read == NULL)
424 tool->read = process_event_sample_stub;
425 if (tool->throttle == NULL)
426 tool->throttle = process_event_stub;
427 if (tool->unthrottle == NULL)
428 tool->unthrottle = process_event_stub;
429 if (tool->attr == NULL)
430 tool->attr = process_event_synth_attr_stub;
431 if (tool->event_type == NULL)
432 tool->event_type = process_event_type_stub;
433 if (tool->tracing_data == NULL)
434 tool->tracing_data = process_event_synth_tracing_data_stub;
435 if (tool->build_id == NULL)
436 tool->build_id = process_finished_round_stub;
437 if (tool->finished_round == NULL) {
438 if (tool->ordered_samples)
439 tool->finished_round = process_finished_round;
441 tool->finished_round = process_finished_round_stub;
445 void mem_bswap_64(void *src, int byte_size)
449 while (byte_size > 0) {
451 byte_size -= sizeof(u64);
456 static void swap_sample_id_all(union perf_event *event, void *data)
458 void *end = (void *) event + event->header.size;
459 int size = end - data;
461 BUG_ON(size % sizeof(u64));
462 mem_bswap_64(data, size);
465 static void perf_event__all64_swap(union perf_event *event,
466 bool sample_id_all __used)
468 struct perf_event_header *hdr = &event->header;
469 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
472 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
474 event->comm.pid = bswap_32(event->comm.pid);
475 event->comm.tid = bswap_32(event->comm.tid);
478 void *data = &event->comm.comm;
480 data += ALIGN(strlen(data) + 1, sizeof(u64));
481 swap_sample_id_all(event, data);
485 static void perf_event__mmap_swap(union perf_event *event,
488 event->mmap.pid = bswap_32(event->mmap.pid);
489 event->mmap.tid = bswap_32(event->mmap.tid);
490 event->mmap.start = bswap_64(event->mmap.start);
491 event->mmap.len = bswap_64(event->mmap.len);
492 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
495 void *data = &event->mmap.filename;
497 data += ALIGN(strlen(data) + 1, sizeof(u64));
498 swap_sample_id_all(event, data);
502 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
504 event->fork.pid = bswap_32(event->fork.pid);
505 event->fork.tid = bswap_32(event->fork.tid);
506 event->fork.ppid = bswap_32(event->fork.ppid);
507 event->fork.ptid = bswap_32(event->fork.ptid);
508 event->fork.time = bswap_64(event->fork.time);
511 swap_sample_id_all(event, &event->fork + 1);
514 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
516 event->read.pid = bswap_32(event->read.pid);
517 event->read.tid = bswap_32(event->read.tid);
518 event->read.value = bswap_64(event->read.value);
519 event->read.time_enabled = bswap_64(event->read.time_enabled);
520 event->read.time_running = bswap_64(event->read.time_running);
521 event->read.id = bswap_64(event->read.id);
524 swap_sample_id_all(event, &event->read + 1);
527 static u8 revbyte(u8 b)
529 int rev = (b >> 4) | ((b & 0xf) << 4);
530 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
531 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
536 * XXX this is hack in attempt to carry flags bitfield
537 * throught endian village. ABI says:
539 * Bit-fields are allocated from right to left (least to most significant)
540 * on little-endian implementations and from left to right (most to least
541 * significant) on big-endian implementations.
543 * The above seems to be byte specific, so we need to reverse each
544 * byte of the bitfield. 'Internet' also says this might be implementation
545 * specific and we probably need proper fix and carry perf_event_attr
546 * bitfield flags in separate data file FEAT_ section. Thought this seems
549 static void swap_bitfield(u8 *p, unsigned len)
553 for (i = 0; i < len; i++) {
559 /* exported for swapping attributes in file header */
560 void perf_event__attr_swap(struct perf_event_attr *attr)
562 attr->type = bswap_32(attr->type);
563 attr->size = bswap_32(attr->size);
564 attr->config = bswap_64(attr->config);
565 attr->sample_period = bswap_64(attr->sample_period);
566 attr->sample_type = bswap_64(attr->sample_type);
567 attr->read_format = bswap_64(attr->read_format);
568 attr->wakeup_events = bswap_32(attr->wakeup_events);
569 attr->bp_type = bswap_32(attr->bp_type);
570 attr->bp_addr = bswap_64(attr->bp_addr);
571 attr->bp_len = bswap_64(attr->bp_len);
573 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
576 static void perf_event__hdr_attr_swap(union perf_event *event,
577 bool sample_id_all __used)
581 perf_event__attr_swap(&event->attr.attr);
583 size = event->header.size;
584 size -= (void *)&event->attr.id - (void *)event;
585 mem_bswap_64(event->attr.id, size);
588 static void perf_event__event_type_swap(union perf_event *event,
589 bool sample_id_all __used)
591 event->event_type.event_type.event_id =
592 bswap_64(event->event_type.event_type.event_id);
595 static void perf_event__tracing_data_swap(union perf_event *event,
596 bool sample_id_all __used)
598 event->tracing_data.size = bswap_32(event->tracing_data.size);
601 typedef void (*perf_event__swap_op)(union perf_event *event,
604 static perf_event__swap_op perf_event__swap_ops[] = {
605 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
606 [PERF_RECORD_COMM] = perf_event__comm_swap,
607 [PERF_RECORD_FORK] = perf_event__task_swap,
608 [PERF_RECORD_EXIT] = perf_event__task_swap,
609 [PERF_RECORD_LOST] = perf_event__all64_swap,
610 [PERF_RECORD_READ] = perf_event__read_swap,
611 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
612 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
613 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
614 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
615 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
616 [PERF_RECORD_HEADER_MAX] = NULL,
619 struct sample_queue {
622 union perf_event *event;
623 struct list_head list;
626 static void perf_session_free_sample_buffers(struct perf_session *session)
628 struct ordered_samples *os = &session->ordered_samples;
630 while (!list_empty(&os->to_free)) {
631 struct sample_queue *sq;
633 sq = list_entry(os->to_free.next, struct sample_queue, list);
639 static int perf_session_deliver_event(struct perf_session *session,
640 union perf_event *event,
641 struct perf_sample *sample,
642 struct perf_tool *tool,
645 static void flush_sample_queue(struct perf_session *s,
646 struct perf_tool *tool)
648 struct ordered_samples *os = &s->ordered_samples;
649 struct list_head *head = &os->samples;
650 struct sample_queue *tmp, *iter;
651 struct perf_sample sample;
652 u64 limit = os->next_flush;
653 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
654 unsigned idx = 0, progress_next = os->nr_samples / 16;
657 if (!tool->ordered_samples || !limit)
660 list_for_each_entry_safe(iter, tmp, head, list) {
661 if (iter->timestamp > limit)
664 ret = perf_session__parse_sample(s, iter->event, &sample);
666 pr_err("Can't parse sample, err = %d\n", ret);
668 perf_session_deliver_event(s, iter->event, &sample, tool,
671 os->last_flush = iter->timestamp;
672 list_del(&iter->list);
673 list_add(&iter->list, &os->sample_cache);
674 if (++idx >= progress_next) {
675 progress_next += os->nr_samples / 16;
676 ui_progress__update(idx, os->nr_samples,
677 "Processing time ordered events...");
681 if (list_empty(head)) {
682 os->last_sample = NULL;
683 } else if (last_ts <= limit) {
685 list_entry(head->prev, struct sample_queue, list);
692 * When perf record finishes a pass on every buffers, it records this pseudo
694 * We record the max timestamp t found in the pass n.
695 * Assuming these timestamps are monotonic across cpus, we know that if
696 * a buffer still has events with timestamps below t, they will be all
697 * available and then read in the pass n + 1.
698 * Hence when we start to read the pass n + 2, we can safely flush every
699 * events with timestamps below t.
701 * ============ PASS n =================
704 * cnt1 timestamps | cnt2 timestamps
707 * - | 4 <--- max recorded
709 * ============ PASS n + 1 ==============
712 * cnt1 timestamps | cnt2 timestamps
715 * 5 | 7 <---- max recorded
717 * Flush every events below timestamp 4
719 * ============ PASS n + 2 ==============
722 * cnt1 timestamps | cnt2 timestamps
727 * Flush every events below timestamp 7
730 static int process_finished_round(struct perf_tool *tool,
731 union perf_event *event __used,
732 struct perf_session *session)
734 flush_sample_queue(session, tool);
735 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
740 /* The queue is ordered by time */
741 static void __queue_event(struct sample_queue *new, struct perf_session *s)
743 struct ordered_samples *os = &s->ordered_samples;
744 struct sample_queue *sample = os->last_sample;
745 u64 timestamp = new->timestamp;
749 os->last_sample = new;
752 list_add(&new->list, &os->samples);
753 os->max_timestamp = timestamp;
758 * last_sample might point to some random place in the list as it's
759 * the last queued event. We expect that the new event is close to
762 if (sample->timestamp <= timestamp) {
763 while (sample->timestamp <= timestamp) {
764 p = sample->list.next;
765 if (p == &os->samples) {
766 list_add_tail(&new->list, &os->samples);
767 os->max_timestamp = timestamp;
770 sample = list_entry(p, struct sample_queue, list);
772 list_add_tail(&new->list, &sample->list);
774 while (sample->timestamp > timestamp) {
775 p = sample->list.prev;
776 if (p == &os->samples) {
777 list_add(&new->list, &os->samples);
780 sample = list_entry(p, struct sample_queue, list);
782 list_add(&new->list, &sample->list);
786 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
788 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
789 struct perf_sample *sample, u64 file_offset)
791 struct ordered_samples *os = &s->ordered_samples;
792 struct list_head *sc = &os->sample_cache;
793 u64 timestamp = sample->time;
794 struct sample_queue *new;
796 if (!timestamp || timestamp == ~0ULL)
799 if (timestamp < s->ordered_samples.last_flush) {
800 printf("Warning: Timestamp below last timeslice flush\n");
804 if (!list_empty(sc)) {
805 new = list_entry(sc->next, struct sample_queue, list);
806 list_del(&new->list);
807 } else if (os->sample_buffer) {
808 new = os->sample_buffer + os->sample_buffer_idx;
809 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
810 os->sample_buffer = NULL;
812 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
813 if (!os->sample_buffer)
815 list_add(&os->sample_buffer->list, &os->to_free);
816 os->sample_buffer_idx = 2;
817 new = os->sample_buffer + 1;
820 new->timestamp = timestamp;
821 new->file_offset = file_offset;
824 __queue_event(new, s);
829 static void callchain__printf(struct perf_sample *sample)
833 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
835 for (i = 0; i < sample->callchain->nr; i++)
836 printf("..... %2d: %016" PRIx64 "\n",
837 i, sample->callchain->ips[i]);
840 static void branch_stack__printf(struct perf_sample *sample)
844 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
846 for (i = 0; i < sample->branch_stack->nr; i++)
847 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
848 i, sample->branch_stack->entries[i].from,
849 sample->branch_stack->entries[i].to);
852 static void perf_session__print_tstamp(struct perf_session *session,
853 union perf_event *event,
854 struct perf_sample *sample)
856 if (event->header.type != PERF_RECORD_SAMPLE &&
857 !session->sample_id_all) {
858 fputs("-1 -1 ", stdout);
862 if ((session->sample_type & PERF_SAMPLE_CPU))
863 printf("%u ", sample->cpu);
865 if (session->sample_type & PERF_SAMPLE_TIME)
866 printf("%" PRIu64 " ", sample->time);
869 static void dump_event(struct perf_session *session, union perf_event *event,
870 u64 file_offset, struct perf_sample *sample)
875 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
876 file_offset, event->header.size, event->header.type);
881 perf_session__print_tstamp(session, event, sample);
883 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
884 event->header.size, perf_event__name(event->header.type));
887 static void dump_sample(struct perf_session *session, union perf_event *event,
888 struct perf_sample *sample)
893 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
894 event->header.misc, sample->pid, sample->tid, sample->ip,
895 sample->period, sample->addr);
897 if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
898 callchain__printf(sample);
900 if (session->sample_type & PERF_SAMPLE_BRANCH_STACK)
901 branch_stack__printf(sample);
904 static struct machine *
905 perf_session__find_machine_for_cpumode(struct perf_session *session,
906 union perf_event *event)
908 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
910 if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) {
913 if (event->header.type == PERF_RECORD_MMAP)
914 pid = event->mmap.pid;
918 return perf_session__find_machine(session, pid);
921 return perf_session__find_host_machine(session);
924 static int perf_session_deliver_event(struct perf_session *session,
925 union perf_event *event,
926 struct perf_sample *sample,
927 struct perf_tool *tool,
930 struct perf_evsel *evsel;
931 struct machine *machine;
933 dump_event(session, event, file_offset, sample);
935 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
936 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
938 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
939 * because the tools right now may apply filters, discarding
940 * some of the samples. For consistency, in the future we
941 * should have something like nr_filtered_samples and remove
942 * the sample->period from total_sample_period, etc, KISS for
945 * Also testing against NULL allows us to handle files without
946 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
947 * future probably it'll be a good idea to restrict event
948 * processing via perf_session to files with both set.
950 hists__inc_nr_events(&evsel->hists, event->header.type);
953 machine = perf_session__find_machine_for_cpumode(session, event);
955 switch (event->header.type) {
956 case PERF_RECORD_SAMPLE:
957 dump_sample(session, event, sample);
959 ++session->hists.stats.nr_unknown_id;
962 if (machine == NULL) {
963 ++session->hists.stats.nr_unprocessable_samples;
966 return tool->sample(tool, event, sample, evsel, machine);
967 case PERF_RECORD_MMAP:
968 return tool->mmap(tool, event, sample, machine);
969 case PERF_RECORD_COMM:
970 return tool->comm(tool, event, sample, machine);
971 case PERF_RECORD_FORK:
972 return tool->fork(tool, event, sample, machine);
973 case PERF_RECORD_EXIT:
974 return tool->exit(tool, event, sample, machine);
975 case PERF_RECORD_LOST:
976 if (tool->lost == perf_event__process_lost)
977 session->hists.stats.total_lost += event->lost.lost;
978 return tool->lost(tool, event, sample, machine);
979 case PERF_RECORD_READ:
980 return tool->read(tool, event, sample, evsel, machine);
981 case PERF_RECORD_THROTTLE:
982 return tool->throttle(tool, event, sample, machine);
983 case PERF_RECORD_UNTHROTTLE:
984 return tool->unthrottle(tool, event, sample, machine);
986 ++session->hists.stats.nr_unknown_events;
991 static int perf_session__preprocess_sample(struct perf_session *session,
992 union perf_event *event, struct perf_sample *sample)
994 if (event->header.type != PERF_RECORD_SAMPLE ||
995 !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
998 if (!ip_callchain__valid(sample->callchain, event)) {
999 pr_debug("call-chain problem with event, skipping it.\n");
1000 ++session->hists.stats.nr_invalid_chains;
1001 session->hists.stats.total_invalid_chains += sample->period;
1007 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1008 struct perf_tool *tool, u64 file_offset)
1012 dump_event(session, event, file_offset, NULL);
1014 /* These events are processed right away */
1015 switch (event->header.type) {
1016 case PERF_RECORD_HEADER_ATTR:
1017 err = tool->attr(event, &session->evlist);
1019 perf_session__update_sample_type(session);
1021 case PERF_RECORD_HEADER_EVENT_TYPE:
1022 return tool->event_type(tool, event);
1023 case PERF_RECORD_HEADER_TRACING_DATA:
1024 /* setup for reading amidst mmap */
1025 lseek(session->fd, file_offset, SEEK_SET);
1026 return tool->tracing_data(event, session);
1027 case PERF_RECORD_HEADER_BUILD_ID:
1028 return tool->build_id(tool, event, session);
1029 case PERF_RECORD_FINISHED_ROUND:
1030 return tool->finished_round(tool, event, session);
1036 static void event_swap(union perf_event *event, bool sample_id_all)
1038 perf_event__swap_op swap;
1040 swap = perf_event__swap_ops[event->header.type];
1042 swap(event, sample_id_all);
1045 static int perf_session__process_event(struct perf_session *session,
1046 union perf_event *event,
1047 struct perf_tool *tool,
1050 struct perf_sample sample;
1053 if (session->header.needs_swap)
1054 event_swap(event, session->sample_id_all);
1056 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1059 hists__inc_nr_events(&session->hists, event->header.type);
1061 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1062 return perf_session__process_user_event(session, event, tool, file_offset);
1065 * For all kernel events we get the sample data
1067 ret = perf_session__parse_sample(session, event, &sample);
1071 /* Preprocess sample records - precheck callchains */
1072 if (perf_session__preprocess_sample(session, event, &sample))
1075 if (tool->ordered_samples) {
1076 ret = perf_session_queue_event(session, event, &sample,
1082 return perf_session_deliver_event(session, event, &sample, tool,
1086 void perf_event_header__bswap(struct perf_event_header *self)
1088 self->type = bswap_32(self->type);
1089 self->misc = bswap_16(self->misc);
1090 self->size = bswap_16(self->size);
1093 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1095 return machine__findnew_thread(&session->host_machine, pid);
1098 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1100 struct thread *thread = perf_session__findnew(self, 0);
1102 if (thread == NULL || thread__set_comm(thread, "swapper")) {
1103 pr_err("problem inserting idle task.\n");
1110 static void perf_session__warn_about_errors(const struct perf_session *session,
1111 const struct perf_tool *tool)
1113 if (tool->lost == perf_event__process_lost &&
1114 session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1115 ui__warning("Processed %d events and lost %d chunks!\n\n"
1116 "Check IO/CPU overload!\n\n",
1117 session->hists.stats.nr_events[0],
1118 session->hists.stats.nr_events[PERF_RECORD_LOST]);
1121 if (session->hists.stats.nr_unknown_events != 0) {
1122 ui__warning("Found %u unknown events!\n\n"
1123 "Is this an older tool processing a perf.data "
1124 "file generated by a more recent tool?\n\n"
1125 "If that is not the case, consider "
1126 "reporting to linux-kernel@vger.kernel.org.\n\n",
1127 session->hists.stats.nr_unknown_events);
1130 if (session->hists.stats.nr_unknown_id != 0) {
1131 ui__warning("%u samples with id not present in the header\n",
1132 session->hists.stats.nr_unknown_id);
1135 if (session->hists.stats.nr_invalid_chains != 0) {
1136 ui__warning("Found invalid callchains!\n\n"
1137 "%u out of %u events were discarded for this reason.\n\n"
1138 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1139 session->hists.stats.nr_invalid_chains,
1140 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1143 if (session->hists.stats.nr_unprocessable_samples != 0) {
1144 ui__warning("%u unprocessable samples recorded.\n"
1145 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1146 session->hists.stats.nr_unprocessable_samples);
1150 #define session_done() (*(volatile int *)(&session_done))
1151 volatile int session_done;
1153 static int __perf_session__process_pipe_events(struct perf_session *self,
1154 struct perf_tool *tool)
1156 union perf_event *event;
1157 uint32_t size, cur_size = 0;
1164 perf_tool__fill_defaults(tool);
1167 cur_size = sizeof(union perf_event);
1169 buf = malloc(cur_size);
1174 err = readn(self->fd, event, sizeof(struct perf_event_header));
1179 pr_err("failed to read event header\n");
1183 if (self->header.needs_swap)
1184 perf_event_header__bswap(&event->header);
1186 size = event->header.size;
1190 if (size > cur_size) {
1191 void *new = realloc(buf, size);
1193 pr_err("failed to allocate memory to read event\n");
1201 p += sizeof(struct perf_event_header);
1203 if (size - sizeof(struct perf_event_header)) {
1204 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1207 pr_err("unexpected end of event stream\n");
1211 pr_err("failed to read event data\n");
1216 if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1217 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1218 head, event->header.size, event->header.type);
1228 if (!session_done())
1234 perf_session__warn_about_errors(self, tool);
1235 perf_session_free_sample_buffers(self);
1239 static union perf_event *
1240 fetch_mmaped_event(struct perf_session *session,
1241 u64 head, size_t mmap_size, char *buf)
1243 union perf_event *event;
1246 * Ensure we have enough space remaining to read
1247 * the size of the event in the headers.
1249 if (head + sizeof(event->header) > mmap_size)
1252 event = (union perf_event *)(buf + head);
1254 if (session->header.needs_swap)
1255 perf_event_header__bswap(&event->header);
1257 if (head + event->header.size > mmap_size)
1263 int __perf_session__process_events(struct perf_session *session,
1264 u64 data_offset, u64 data_size,
1265 u64 file_size, struct perf_tool *tool)
1267 u64 head, page_offset, file_offset, file_pos, progress_next;
1268 int err, mmap_prot, mmap_flags, map_idx = 0;
1269 size_t page_size, mmap_size;
1270 char *buf, *mmaps[8];
1271 union perf_event *event;
1274 perf_tool__fill_defaults(tool);
1276 page_size = sysconf(_SC_PAGESIZE);
1278 page_offset = page_size * (data_offset / page_size);
1279 file_offset = page_offset;
1280 head = data_offset - page_offset;
1282 if (data_offset + data_size < file_size)
1283 file_size = data_offset + data_size;
1285 progress_next = file_size / 16;
1287 mmap_size = session->mmap_window;
1288 if (mmap_size > file_size)
1289 mmap_size = file_size;
1291 memset(mmaps, 0, sizeof(mmaps));
1293 mmap_prot = PROT_READ;
1294 mmap_flags = MAP_SHARED;
1296 if (session->header.needs_swap) {
1297 mmap_prot |= PROT_WRITE;
1298 mmap_flags = MAP_PRIVATE;
1301 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1303 if (buf == MAP_FAILED) {
1304 pr_err("failed to mmap file\n");
1308 mmaps[map_idx] = buf;
1309 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1310 file_pos = file_offset + head;
1313 event = fetch_mmaped_event(session, head, mmap_size, buf);
1315 if (mmaps[map_idx]) {
1316 munmap(mmaps[map_idx], mmap_size);
1317 mmaps[map_idx] = NULL;
1320 page_offset = page_size * (head / page_size);
1321 file_offset += page_offset;
1322 head -= page_offset;
1326 size = event->header.size;
1329 perf_session__process_event(session, event, tool, file_pos) < 0) {
1330 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1331 file_offset + head, event->header.size,
1332 event->header.type);
1340 if (file_pos >= progress_next) {
1341 progress_next += file_size / 16;
1342 ui_progress__update(file_pos, file_size,
1343 "Processing events...");
1346 if (file_pos < file_size)
1350 /* do the final flush for ordered samples */
1351 session->ordered_samples.next_flush = ULLONG_MAX;
1352 flush_sample_queue(session, tool);
1354 perf_session__warn_about_errors(session, tool);
1355 perf_session_free_sample_buffers(session);
1359 int perf_session__process_events(struct perf_session *self,
1360 struct perf_tool *tool)
1364 if (perf_session__register_idle_thread(self) == NULL)
1368 err = __perf_session__process_events(self,
1369 self->header.data_offset,
1370 self->header.data_size,
1373 err = __perf_session__process_pipe_events(self, tool);
1378 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1380 if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1381 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1388 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1389 const char *symbol_name, u64 addr)
1393 struct ref_reloc_sym *ref;
1395 ref = zalloc(sizeof(struct ref_reloc_sym));
1399 ref->name = strdup(symbol_name);
1400 if (ref->name == NULL) {
1405 bracket = strchr(ref->name, ']');
1411 for (i = 0; i < MAP__NR_TYPES; ++i) {
1412 struct kmap *kmap = map__kmap(maps[i]);
1413 kmap->ref_reloc_sym = ref;
1419 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1421 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1422 __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1423 machines__fprintf_dsos(&self->machines, fp);
1426 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1429 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1430 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1433 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1435 struct perf_evsel *pos;
1436 size_t ret = fprintf(fp, "Aggregated stats:\n");
1438 ret += hists__fprintf_nr_events(&session->hists, fp);
1440 list_for_each_entry(pos, &session->evlist->entries, node) {
1441 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1442 ret += hists__fprintf_nr_events(&pos->hists, fp);
1448 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1451 * FIXME: Here we have to actually print all the machines in this
1452 * session, not just the host...
1454 return machine__fprintf(&session->host_machine, fp);
1457 void perf_session__remove_thread(struct perf_session *session,
1461 * FIXME: This one makes no sense, we need to remove the thread from
1462 * the machine it belongs to, perf_session can have many machines, so
1463 * doing it always on ->host_machine is wrong. Fix when auditing all
1464 * the 'perf kvm' code.
1466 machine__remove_thread(&session->host_machine, th);
1469 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1472 struct perf_evsel *pos;
1474 list_for_each_entry(pos, &session->evlist->entries, node) {
1475 if (pos->attr.type == type)
1481 void perf_event__print_ip(union perf_event *event, struct perf_sample *sample,
1482 struct machine *machine, int print_sym,
1483 int print_dso, int print_symoffset)
1485 struct addr_location al;
1486 struct callchain_cursor_node *node;
1488 if (perf_event__preprocess_sample(event, machine, &al, sample,
1490 error("problem processing %d event, skipping it.\n",
1491 event->header.type);
1495 if (symbol_conf.use_callchain && sample->callchain) {
1497 if (machine__resolve_callchain(machine, al.thread,
1498 sample->callchain, NULL) != 0) {
1500 error("Failed to resolve callchain. Skipping\n");
1503 callchain_cursor_commit(&callchain_cursor);
1506 node = callchain_cursor_current(&callchain_cursor);
1510 printf("\t%16" PRIx64, node->ip);
1513 symbol__fprintf_symname(node->sym, stdout);
1517 map__fprintf_dsoname(node->map, stdout);
1522 callchain_cursor_advance(&callchain_cursor);
1526 printf("%16" PRIx64, sample->ip);
1529 if (print_symoffset)
1530 symbol__fprintf_symname_offs(al.sym, &al,
1533 symbol__fprintf_symname(al.sym, stdout);
1538 map__fprintf_dsoname(al.map, stdout);
1544 int perf_session__cpu_bitmap(struct perf_session *session,
1545 const char *cpu_list, unsigned long *cpu_bitmap)
1548 struct cpu_map *map;
1550 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1551 struct perf_evsel *evsel;
1553 evsel = perf_session__find_first_evtype(session, i);
1557 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1558 pr_err("File does not contain CPU events. "
1559 "Remove -c option to proceed.\n");
1564 map = cpu_map__new(cpu_list);
1566 pr_err("Invalid cpu_list\n");
1570 for (i = 0; i < map->nr; i++) {
1571 int cpu = map->map[i];
1573 if (cpu >= MAX_NR_CPUS) {
1574 pr_err("Requested CPU %d too large. "
1575 "Consider raising MAX_NR_CPUS\n", cpu);
1579 set_bit(cpu, cpu_bitmap);
1585 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1591 if (session == NULL || fp == NULL)
1594 ret = fstat(session->fd, &st);
1598 fprintf(fp, "# ========\n");
1599 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1600 perf_header__fprintf_info(session, fp, full);
1601 fprintf(fp, "# ========\n#\n");