1 #define _FILE_OFFSET_BITS 64
3 #include <linux/kernel.h>
17 #include "event-parse.h"
18 #include "perf_regs.h"
22 static int perf_session__open(struct perf_session *self, bool force)
24 struct stat input_stat;
26 if (!strcmp(self->filename, "-")) {
28 self->fd = STDIN_FILENO;
30 if (perf_session__read_header(self, self->fd) < 0)
31 pr_err("incompatible file format (rerun with -v to learn more)");
36 self->fd = open(self->filename, O_RDONLY);
40 pr_err("failed to open %s: %s", self->filename, strerror(err));
41 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
42 pr_err(" (try 'perf record' first)");
47 if (fstat(self->fd, &input_stat) < 0)
50 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
51 pr_err("file %s not owned by current user or root\n",
56 if (!input_stat.st_size) {
57 pr_info("zero-sized file (%s), nothing to do!\n",
62 if (perf_session__read_header(self, self->fd) < 0) {
63 pr_err("incompatible file format (rerun with -v to learn more)");
67 if (!perf_evlist__valid_sample_type(self->evlist)) {
68 pr_err("non matching sample_type");
72 if (!perf_evlist__valid_sample_id_all(self->evlist)) {
73 pr_err("non matching sample_id_all");
77 self->size = input_stat.st_size;
86 void perf_session__set_id_hdr_size(struct perf_session *session)
88 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
90 session->host_machine.id_hdr_size = id_hdr_size;
91 machines__set_id_hdr_size(&session->machines, id_hdr_size);
94 int perf_session__create_kernel_maps(struct perf_session *self)
96 int ret = machine__create_kernel_maps(&self->host_machine);
99 ret = machines__create_guest_kernel_maps(&self->machines);
103 static void perf_session__destroy_kernel_maps(struct perf_session *self)
105 machine__destroy_kernel_maps(&self->host_machine);
106 machines__destroy_guest_kernel_maps(&self->machines);
109 struct perf_session *perf_session__new(const char *filename, int mode,
110 bool force, bool repipe,
111 struct perf_tool *tool)
113 struct perf_session *self;
117 if (!filename || !strlen(filename)) {
118 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
121 filename = "perf.data";
124 len = strlen(filename);
125 self = zalloc(sizeof(*self) + len);
130 memcpy(self->filename, filename, len);
131 self->machines = RB_ROOT;
132 self->repipe = repipe;
133 INIT_LIST_HEAD(&self->ordered_samples.samples);
134 INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
135 INIT_LIST_HEAD(&self->ordered_samples.to_free);
136 machine__init(&self->host_machine, "", HOST_KERNEL_ID);
137 hists__init(&self->hists);
139 if (mode == O_RDONLY) {
140 if (perf_session__open(self, force) < 0)
142 perf_session__set_id_hdr_size(self);
143 } else if (mode == O_WRONLY) {
145 * In O_RDONLY mode this will be performed when reading the
146 * kernel MMAP event, in perf_event__process_mmap().
148 if (perf_session__create_kernel_maps(self) < 0)
152 if (tool && tool->ordering_requires_timestamps &&
153 tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
154 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
155 tool->ordered_samples = false;
161 perf_session__delete(self);
165 static void machine__delete_dead_threads(struct machine *machine)
167 struct thread *n, *t;
169 list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
175 static void perf_session__delete_dead_threads(struct perf_session *session)
177 machine__delete_dead_threads(&session->host_machine);
180 static void machine__delete_threads(struct machine *self)
182 struct rb_node *nd = rb_first(&self->threads);
185 struct thread *t = rb_entry(nd, struct thread, rb_node);
187 rb_erase(&t->rb_node, &self->threads);
193 static void perf_session__delete_threads(struct perf_session *session)
195 machine__delete_threads(&session->host_machine);
198 static void perf_session_env__delete(struct perf_session_env *env)
201 free(env->os_release);
208 free(env->sibling_cores);
209 free(env->sibling_threads);
210 free(env->numa_nodes);
211 free(env->pmu_mappings);
214 void perf_session__delete(struct perf_session *self)
216 perf_session__destroy_kernel_maps(self);
217 perf_session__delete_dead_threads(self);
218 perf_session__delete_threads(self);
219 perf_session_env__delete(&self->header.env);
220 machine__exit(&self->host_machine);
226 void machine__remove_thread(struct machine *self, struct thread *th)
228 self->last_match = NULL;
229 rb_erase(&th->rb_node, &self->threads);
231 * We may have references to this thread, for instance in some hist_entry
232 * instances, so just move them to a separate list.
234 list_add_tail(&th->node, &self->dead_threads);
237 static bool symbol__match_parent_regex(struct symbol *sym)
239 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
245 static const u8 cpumodes[] = {
246 PERF_RECORD_MISC_USER,
247 PERF_RECORD_MISC_KERNEL,
248 PERF_RECORD_MISC_GUEST_USER,
249 PERF_RECORD_MISC_GUEST_KERNEL
251 #define NCPUMODES (sizeof(cpumodes)/sizeof(u8))
253 static void ip__resolve_ams(struct machine *self, struct thread *thread,
254 struct addr_map_symbol *ams,
257 struct addr_location al;
261 memset(&al, 0, sizeof(al));
263 for (i = 0; i < NCPUMODES; i++) {
266 * We cannot use the header.misc hint to determine whether a
267 * branch stack address is user, kernel, guest, hypervisor.
268 * Branches may straddle the kernel/user/hypervisor boundaries.
269 * Thus, we have to try consecutively until we find a match
270 * or else, the symbol is unknown
272 thread__find_addr_location(thread, self, m, MAP__FUNCTION,
279 ams->al_addr = al.addr;
284 struct branch_info *machine__resolve_bstack(struct machine *self,
286 struct branch_stack *bs)
288 struct branch_info *bi;
291 bi = calloc(bs->nr, sizeof(struct branch_info));
295 for (i = 0; i < bs->nr; i++) {
296 ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to);
297 ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from);
298 bi[i].flags = bs->entries[i].flags;
303 static int machine__resolve_callchain_sample(struct machine *machine,
304 struct thread *thread,
305 struct ip_callchain *chain,
306 struct symbol **parent)
309 u8 cpumode = PERF_RECORD_MISC_USER;
313 callchain_cursor_reset(&callchain_cursor);
315 if (chain->nr > PERF_MAX_STACK_DEPTH) {
316 pr_warning("corrupted callchain. skipping...\n");
320 for (i = 0; i < chain->nr; i++) {
322 struct addr_location al;
324 if (callchain_param.order == ORDER_CALLEE)
327 ip = chain->ips[chain->nr - i - 1];
329 if (ip >= PERF_CONTEXT_MAX) {
331 case PERF_CONTEXT_HV:
332 cpumode = PERF_RECORD_MISC_HYPERVISOR;
334 case PERF_CONTEXT_KERNEL:
335 cpumode = PERF_RECORD_MISC_KERNEL;
337 case PERF_CONTEXT_USER:
338 cpumode = PERF_RECORD_MISC_USER;
341 pr_debug("invalid callchain context: "
342 "%"PRId64"\n", (s64) ip);
344 * It seems the callchain is corrupted.
347 callchain_cursor_reset(&callchain_cursor);
354 thread__find_addr_location(thread, machine, cpumode,
355 MAP__FUNCTION, ip, &al, NULL);
356 if (al.sym != NULL) {
357 if (sort__has_parent && !*parent &&
358 symbol__match_parent_regex(al.sym))
360 if (!symbol_conf.use_callchain)
364 err = callchain_cursor_append(&callchain_cursor,
373 static int unwind_entry(struct unwind_entry *entry, void *arg)
375 struct callchain_cursor *cursor = arg;
376 return callchain_cursor_append(cursor, entry->ip,
377 entry->map, entry->sym);
380 int machine__resolve_callchain(struct machine *machine,
381 struct perf_evsel *evsel,
382 struct thread *thread,
383 struct perf_sample *sample,
384 struct symbol **parent)
389 callchain_cursor_reset(&callchain_cursor);
391 ret = machine__resolve_callchain_sample(machine, thread,
392 sample->callchain, parent);
396 /* Can we do dwarf post unwind? */
397 if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
398 (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
401 /* Bail out if nothing was captured. */
402 if ((!sample->user_regs.regs) ||
403 (!sample->user_stack.size))
406 return unwind__get_entries(unwind_entry, &callchain_cursor, machine,
407 thread, evsel->attr.sample_regs_user,
412 static int process_event_synth_tracing_data_stub(union perf_event *event
414 struct perf_session *session
417 dump_printf(": unhandled!\n");
421 static int process_event_synth_attr_stub(union perf_event *event __maybe_unused,
422 struct perf_evlist **pevlist
425 dump_printf(": unhandled!\n");
429 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
430 union perf_event *event __maybe_unused,
431 struct perf_sample *sample __maybe_unused,
432 struct perf_evsel *evsel __maybe_unused,
433 struct machine *machine __maybe_unused)
435 dump_printf(": unhandled!\n");
439 static int process_event_stub(struct perf_tool *tool __maybe_unused,
440 union perf_event *event __maybe_unused,
441 struct perf_sample *sample __maybe_unused,
442 struct machine *machine __maybe_unused)
444 dump_printf(": unhandled!\n");
448 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
449 union perf_event *event __maybe_unused,
450 struct perf_session *perf_session
453 dump_printf(": unhandled!\n");
457 static int process_event_type_stub(struct perf_tool *tool __maybe_unused,
458 union perf_event *event __maybe_unused)
460 dump_printf(": unhandled!\n");
464 static int process_finished_round(struct perf_tool *tool,
465 union perf_event *event,
466 struct perf_session *session);
468 static void perf_tool__fill_defaults(struct perf_tool *tool)
470 if (tool->sample == NULL)
471 tool->sample = process_event_sample_stub;
472 if (tool->mmap == NULL)
473 tool->mmap = process_event_stub;
474 if (tool->comm == NULL)
475 tool->comm = process_event_stub;
476 if (tool->fork == NULL)
477 tool->fork = process_event_stub;
478 if (tool->exit == NULL)
479 tool->exit = process_event_stub;
480 if (tool->lost == NULL)
481 tool->lost = perf_event__process_lost;
482 if (tool->read == NULL)
483 tool->read = process_event_sample_stub;
484 if (tool->throttle == NULL)
485 tool->throttle = process_event_stub;
486 if (tool->unthrottle == NULL)
487 tool->unthrottle = process_event_stub;
488 if (tool->attr == NULL)
489 tool->attr = process_event_synth_attr_stub;
490 if (tool->event_type == NULL)
491 tool->event_type = process_event_type_stub;
492 if (tool->tracing_data == NULL)
493 tool->tracing_data = process_event_synth_tracing_data_stub;
494 if (tool->build_id == NULL)
495 tool->build_id = process_finished_round_stub;
496 if (tool->finished_round == NULL) {
497 if (tool->ordered_samples)
498 tool->finished_round = process_finished_round;
500 tool->finished_round = process_finished_round_stub;
504 void mem_bswap_32(void *src, int byte_size)
507 while (byte_size > 0) {
509 byte_size -= sizeof(u32);
514 void mem_bswap_64(void *src, int byte_size)
518 while (byte_size > 0) {
520 byte_size -= sizeof(u64);
525 static void swap_sample_id_all(union perf_event *event, void *data)
527 void *end = (void *) event + event->header.size;
528 int size = end - data;
530 BUG_ON(size % sizeof(u64));
531 mem_bswap_64(data, size);
534 static void perf_event__all64_swap(union perf_event *event,
535 bool sample_id_all __maybe_unused)
537 struct perf_event_header *hdr = &event->header;
538 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
541 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
543 event->comm.pid = bswap_32(event->comm.pid);
544 event->comm.tid = bswap_32(event->comm.tid);
547 void *data = &event->comm.comm;
549 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
550 swap_sample_id_all(event, data);
554 static void perf_event__mmap_swap(union perf_event *event,
557 event->mmap.pid = bswap_32(event->mmap.pid);
558 event->mmap.tid = bswap_32(event->mmap.tid);
559 event->mmap.start = bswap_64(event->mmap.start);
560 event->mmap.len = bswap_64(event->mmap.len);
561 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
564 void *data = &event->mmap.filename;
566 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
567 swap_sample_id_all(event, data);
571 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
573 event->fork.pid = bswap_32(event->fork.pid);
574 event->fork.tid = bswap_32(event->fork.tid);
575 event->fork.ppid = bswap_32(event->fork.ppid);
576 event->fork.ptid = bswap_32(event->fork.ptid);
577 event->fork.time = bswap_64(event->fork.time);
580 swap_sample_id_all(event, &event->fork + 1);
583 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
585 event->read.pid = bswap_32(event->read.pid);
586 event->read.tid = bswap_32(event->read.tid);
587 event->read.value = bswap_64(event->read.value);
588 event->read.time_enabled = bswap_64(event->read.time_enabled);
589 event->read.time_running = bswap_64(event->read.time_running);
590 event->read.id = bswap_64(event->read.id);
593 swap_sample_id_all(event, &event->read + 1);
596 static u8 revbyte(u8 b)
598 int rev = (b >> 4) | ((b & 0xf) << 4);
599 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
600 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
605 * XXX this is hack in attempt to carry flags bitfield
606 * throught endian village. ABI says:
608 * Bit-fields are allocated from right to left (least to most significant)
609 * on little-endian implementations and from left to right (most to least
610 * significant) on big-endian implementations.
612 * The above seems to be byte specific, so we need to reverse each
613 * byte of the bitfield. 'Internet' also says this might be implementation
614 * specific and we probably need proper fix and carry perf_event_attr
615 * bitfield flags in separate data file FEAT_ section. Thought this seems
618 static void swap_bitfield(u8 *p, unsigned len)
622 for (i = 0; i < len; i++) {
628 /* exported for swapping attributes in file header */
629 void perf_event__attr_swap(struct perf_event_attr *attr)
631 attr->type = bswap_32(attr->type);
632 attr->size = bswap_32(attr->size);
633 attr->config = bswap_64(attr->config);
634 attr->sample_period = bswap_64(attr->sample_period);
635 attr->sample_type = bswap_64(attr->sample_type);
636 attr->read_format = bswap_64(attr->read_format);
637 attr->wakeup_events = bswap_32(attr->wakeup_events);
638 attr->bp_type = bswap_32(attr->bp_type);
639 attr->bp_addr = bswap_64(attr->bp_addr);
640 attr->bp_len = bswap_64(attr->bp_len);
642 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
645 static void perf_event__hdr_attr_swap(union perf_event *event,
646 bool sample_id_all __maybe_unused)
650 perf_event__attr_swap(&event->attr.attr);
652 size = event->header.size;
653 size -= (void *)&event->attr.id - (void *)event;
654 mem_bswap_64(event->attr.id, size);
657 static void perf_event__event_type_swap(union perf_event *event,
658 bool sample_id_all __maybe_unused)
660 event->event_type.event_type.event_id =
661 bswap_64(event->event_type.event_type.event_id);
664 static void perf_event__tracing_data_swap(union perf_event *event,
665 bool sample_id_all __maybe_unused)
667 event->tracing_data.size = bswap_32(event->tracing_data.size);
670 typedef void (*perf_event__swap_op)(union perf_event *event,
673 static perf_event__swap_op perf_event__swap_ops[] = {
674 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
675 [PERF_RECORD_COMM] = perf_event__comm_swap,
676 [PERF_RECORD_FORK] = perf_event__task_swap,
677 [PERF_RECORD_EXIT] = perf_event__task_swap,
678 [PERF_RECORD_LOST] = perf_event__all64_swap,
679 [PERF_RECORD_READ] = perf_event__read_swap,
680 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
681 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
682 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
683 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
684 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
685 [PERF_RECORD_HEADER_MAX] = NULL,
688 struct sample_queue {
691 union perf_event *event;
692 struct list_head list;
695 static void perf_session_free_sample_buffers(struct perf_session *session)
697 struct ordered_samples *os = &session->ordered_samples;
699 while (!list_empty(&os->to_free)) {
700 struct sample_queue *sq;
702 sq = list_entry(os->to_free.next, struct sample_queue, list);
708 static int perf_session_deliver_event(struct perf_session *session,
709 union perf_event *event,
710 struct perf_sample *sample,
711 struct perf_tool *tool,
714 static int flush_sample_queue(struct perf_session *s,
715 struct perf_tool *tool)
717 struct ordered_samples *os = &s->ordered_samples;
718 struct list_head *head = &os->samples;
719 struct sample_queue *tmp, *iter;
720 struct perf_sample sample;
721 u64 limit = os->next_flush;
722 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
723 unsigned idx = 0, progress_next = os->nr_samples / 16;
726 if (!tool->ordered_samples || !limit)
729 list_for_each_entry_safe(iter, tmp, head, list) {
730 if (iter->timestamp > limit)
733 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
735 pr_err("Can't parse sample, err = %d\n", ret);
737 ret = perf_session_deliver_event(s, iter->event, &sample, tool,
743 os->last_flush = iter->timestamp;
744 list_del(&iter->list);
745 list_add(&iter->list, &os->sample_cache);
746 if (++idx >= progress_next) {
747 progress_next += os->nr_samples / 16;
748 ui_progress__update(idx, os->nr_samples,
749 "Processing time ordered events...");
753 if (list_empty(head)) {
754 os->last_sample = NULL;
755 } else if (last_ts <= limit) {
757 list_entry(head->prev, struct sample_queue, list);
766 * When perf record finishes a pass on every buffers, it records this pseudo
768 * We record the max timestamp t found in the pass n.
769 * Assuming these timestamps are monotonic across cpus, we know that if
770 * a buffer still has events with timestamps below t, they will be all
771 * available and then read in the pass n + 1.
772 * Hence when we start to read the pass n + 2, we can safely flush every
773 * events with timestamps below t.
775 * ============ PASS n =================
778 * cnt1 timestamps | cnt2 timestamps
781 * - | 4 <--- max recorded
783 * ============ PASS n + 1 ==============
786 * cnt1 timestamps | cnt2 timestamps
789 * 5 | 7 <---- max recorded
791 * Flush every events below timestamp 4
793 * ============ PASS n + 2 ==============
796 * cnt1 timestamps | cnt2 timestamps
801 * Flush every events below timestamp 7
804 static int process_finished_round(struct perf_tool *tool,
805 union perf_event *event __maybe_unused,
806 struct perf_session *session)
808 int ret = flush_sample_queue(session, tool);
810 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
815 /* The queue is ordered by time */
816 static void __queue_event(struct sample_queue *new, struct perf_session *s)
818 struct ordered_samples *os = &s->ordered_samples;
819 struct sample_queue *sample = os->last_sample;
820 u64 timestamp = new->timestamp;
824 os->last_sample = new;
827 list_add(&new->list, &os->samples);
828 os->max_timestamp = timestamp;
833 * last_sample might point to some random place in the list as it's
834 * the last queued event. We expect that the new event is close to
837 if (sample->timestamp <= timestamp) {
838 while (sample->timestamp <= timestamp) {
839 p = sample->list.next;
840 if (p == &os->samples) {
841 list_add_tail(&new->list, &os->samples);
842 os->max_timestamp = timestamp;
845 sample = list_entry(p, struct sample_queue, list);
847 list_add_tail(&new->list, &sample->list);
849 while (sample->timestamp > timestamp) {
850 p = sample->list.prev;
851 if (p == &os->samples) {
852 list_add(&new->list, &os->samples);
855 sample = list_entry(p, struct sample_queue, list);
857 list_add(&new->list, &sample->list);
861 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
863 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
864 struct perf_sample *sample, u64 file_offset)
866 struct ordered_samples *os = &s->ordered_samples;
867 struct list_head *sc = &os->sample_cache;
868 u64 timestamp = sample->time;
869 struct sample_queue *new;
871 if (!timestamp || timestamp == ~0ULL)
874 if (timestamp < s->ordered_samples.last_flush) {
875 printf("Warning: Timestamp below last timeslice flush\n");
879 if (!list_empty(sc)) {
880 new = list_entry(sc->next, struct sample_queue, list);
881 list_del(&new->list);
882 } else if (os->sample_buffer) {
883 new = os->sample_buffer + os->sample_buffer_idx;
884 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
885 os->sample_buffer = NULL;
887 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
888 if (!os->sample_buffer)
890 list_add(&os->sample_buffer->list, &os->to_free);
891 os->sample_buffer_idx = 2;
892 new = os->sample_buffer + 1;
895 new->timestamp = timestamp;
896 new->file_offset = file_offset;
899 __queue_event(new, s);
904 static void callchain__printf(struct perf_sample *sample)
908 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
910 for (i = 0; i < sample->callchain->nr; i++)
911 printf("..... %2d: %016" PRIx64 "\n",
912 i, sample->callchain->ips[i]);
915 static void branch_stack__printf(struct perf_sample *sample)
919 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
921 for (i = 0; i < sample->branch_stack->nr; i++)
922 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
923 i, sample->branch_stack->entries[i].from,
924 sample->branch_stack->entries[i].to);
927 static void regs_dump__printf(u64 mask, u64 *regs)
931 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
934 printf(".... %-5s 0x%" PRIx64 "\n",
935 perf_reg_name(rid), val);
939 static void regs_user__printf(struct perf_sample *sample, u64 mask)
941 struct regs_dump *user_regs = &sample->user_regs;
943 if (user_regs->regs) {
944 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
945 regs_dump__printf(mask, user_regs->regs);
949 static void stack_user__printf(struct stack_dump *dump)
951 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
952 dump->size, dump->offset);
955 static void perf_session__print_tstamp(struct perf_session *session,
956 union perf_event *event,
957 struct perf_sample *sample)
959 u64 sample_type = perf_evlist__sample_type(session->evlist);
961 if (event->header.type != PERF_RECORD_SAMPLE &&
962 !perf_evlist__sample_id_all(session->evlist)) {
963 fputs("-1 -1 ", stdout);
967 if ((sample_type & PERF_SAMPLE_CPU))
968 printf("%u ", sample->cpu);
970 if (sample_type & PERF_SAMPLE_TIME)
971 printf("%" PRIu64 " ", sample->time);
974 static void dump_event(struct perf_session *session, union perf_event *event,
975 u64 file_offset, struct perf_sample *sample)
980 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
981 file_offset, event->header.size, event->header.type);
986 perf_session__print_tstamp(session, event, sample);
988 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
989 event->header.size, perf_event__name(event->header.type));
992 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
993 struct perf_sample *sample)
1000 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
1001 event->header.misc, sample->pid, sample->tid, sample->ip,
1002 sample->period, sample->addr);
1004 sample_type = evsel->attr.sample_type;
1006 if (sample_type & PERF_SAMPLE_CALLCHAIN)
1007 callchain__printf(sample);
1009 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
1010 branch_stack__printf(sample);
1012 if (sample_type & PERF_SAMPLE_REGS_USER)
1013 regs_user__printf(sample, evsel->attr.sample_regs_user);
1015 if (sample_type & PERF_SAMPLE_STACK_USER)
1016 stack_user__printf(&sample->user_stack);
1019 static struct machine *
1020 perf_session__find_machine_for_cpumode(struct perf_session *session,
1021 union perf_event *event)
1023 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1026 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
1027 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
1030 if (event->header.type == PERF_RECORD_MMAP)
1031 pid = event->mmap.pid;
1033 pid = event->ip.pid;
1035 return perf_session__findnew_machine(session, pid);
1038 return perf_session__find_host_machine(session);
1041 static int perf_session_deliver_event(struct perf_session *session,
1042 union perf_event *event,
1043 struct perf_sample *sample,
1044 struct perf_tool *tool,
1047 struct perf_evsel *evsel;
1048 struct machine *machine;
1050 dump_event(session, event, file_offset, sample);
1052 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
1053 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
1055 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
1056 * because the tools right now may apply filters, discarding
1057 * some of the samples. For consistency, in the future we
1058 * should have something like nr_filtered_samples and remove
1059 * the sample->period from total_sample_period, etc, KISS for
1062 * Also testing against NULL allows us to handle files without
1063 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
1064 * future probably it'll be a good idea to restrict event
1065 * processing via perf_session to files with both set.
1067 hists__inc_nr_events(&evsel->hists, event->header.type);
1070 machine = perf_session__find_machine_for_cpumode(session, event);
1072 switch (event->header.type) {
1073 case PERF_RECORD_SAMPLE:
1074 dump_sample(evsel, event, sample);
1075 if (evsel == NULL) {
1076 ++session->hists.stats.nr_unknown_id;
1079 if (machine == NULL) {
1080 ++session->hists.stats.nr_unprocessable_samples;
1083 return tool->sample(tool, event, sample, evsel, machine);
1084 case PERF_RECORD_MMAP:
1085 return tool->mmap(tool, event, sample, machine);
1086 case PERF_RECORD_COMM:
1087 return tool->comm(tool, event, sample, machine);
1088 case PERF_RECORD_FORK:
1089 return tool->fork(tool, event, sample, machine);
1090 case PERF_RECORD_EXIT:
1091 return tool->exit(tool, event, sample, machine);
1092 case PERF_RECORD_LOST:
1093 if (tool->lost == perf_event__process_lost)
1094 session->hists.stats.total_lost += event->lost.lost;
1095 return tool->lost(tool, event, sample, machine);
1096 case PERF_RECORD_READ:
1097 return tool->read(tool, event, sample, evsel, machine);
1098 case PERF_RECORD_THROTTLE:
1099 return tool->throttle(tool, event, sample, machine);
1100 case PERF_RECORD_UNTHROTTLE:
1101 return tool->unthrottle(tool, event, sample, machine);
1103 ++session->hists.stats.nr_unknown_events;
1108 static int perf_session__preprocess_sample(struct perf_session *session,
1109 union perf_event *event, struct perf_sample *sample)
1111 if (event->header.type != PERF_RECORD_SAMPLE ||
1112 !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN))
1115 if (!ip_callchain__valid(sample->callchain, event)) {
1116 pr_debug("call-chain problem with event, skipping it.\n");
1117 ++session->hists.stats.nr_invalid_chains;
1118 session->hists.stats.total_invalid_chains += sample->period;
1124 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1125 struct perf_tool *tool, u64 file_offset)
1129 dump_event(session, event, file_offset, NULL);
1131 /* These events are processed right away */
1132 switch (event->header.type) {
1133 case PERF_RECORD_HEADER_ATTR:
1134 err = tool->attr(event, &session->evlist);
1136 perf_session__set_id_hdr_size(session);
1138 case PERF_RECORD_HEADER_EVENT_TYPE:
1139 return tool->event_type(tool, event);
1140 case PERF_RECORD_HEADER_TRACING_DATA:
1141 /* setup for reading amidst mmap */
1142 lseek(session->fd, file_offset, SEEK_SET);
1143 return tool->tracing_data(event, session);
1144 case PERF_RECORD_HEADER_BUILD_ID:
1145 return tool->build_id(tool, event, session);
1146 case PERF_RECORD_FINISHED_ROUND:
1147 return tool->finished_round(tool, event, session);
1153 static void event_swap(union perf_event *event, bool sample_id_all)
1155 perf_event__swap_op swap;
1157 swap = perf_event__swap_ops[event->header.type];
1159 swap(event, sample_id_all);
1162 static int perf_session__process_event(struct perf_session *session,
1163 union perf_event *event,
1164 struct perf_tool *tool,
1167 struct perf_sample sample;
1170 if (session->header.needs_swap)
1171 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1173 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1176 hists__inc_nr_events(&session->hists, event->header.type);
1178 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1179 return perf_session__process_user_event(session, event, tool, file_offset);
1182 * For all kernel events we get the sample data
1184 ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1188 /* Preprocess sample records - precheck callchains */
1189 if (perf_session__preprocess_sample(session, event, &sample))
1192 if (tool->ordered_samples) {
1193 ret = perf_session_queue_event(session, event, &sample,
1199 return perf_session_deliver_event(session, event, &sample, tool,
1203 void perf_event_header__bswap(struct perf_event_header *self)
1205 self->type = bswap_32(self->type);
1206 self->misc = bswap_16(self->misc);
1207 self->size = bswap_16(self->size);
1210 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1212 return machine__findnew_thread(&session->host_machine, pid);
1215 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1217 struct thread *thread = perf_session__findnew(self, 0);
1219 if (thread == NULL || thread__set_comm(thread, "swapper")) {
1220 pr_err("problem inserting idle task.\n");
1227 static void perf_session__warn_about_errors(const struct perf_session *session,
1228 const struct perf_tool *tool)
1230 if (tool->lost == perf_event__process_lost &&
1231 session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1232 ui__warning("Processed %d events and lost %d chunks!\n\n"
1233 "Check IO/CPU overload!\n\n",
1234 session->hists.stats.nr_events[0],
1235 session->hists.stats.nr_events[PERF_RECORD_LOST]);
1238 if (session->hists.stats.nr_unknown_events != 0) {
1239 ui__warning("Found %u unknown events!\n\n"
1240 "Is this an older tool processing a perf.data "
1241 "file generated by a more recent tool?\n\n"
1242 "If that is not the case, consider "
1243 "reporting to linux-kernel@vger.kernel.org.\n\n",
1244 session->hists.stats.nr_unknown_events);
1247 if (session->hists.stats.nr_unknown_id != 0) {
1248 ui__warning("%u samples with id not present in the header\n",
1249 session->hists.stats.nr_unknown_id);
1252 if (session->hists.stats.nr_invalid_chains != 0) {
1253 ui__warning("Found invalid callchains!\n\n"
1254 "%u out of %u events were discarded for this reason.\n\n"
1255 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1256 session->hists.stats.nr_invalid_chains,
1257 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1260 if (session->hists.stats.nr_unprocessable_samples != 0) {
1261 ui__warning("%u unprocessable samples recorded.\n"
1262 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1263 session->hists.stats.nr_unprocessable_samples);
1267 #define session_done() (*(volatile int *)(&session_done))
1268 volatile int session_done;
1270 static int __perf_session__process_pipe_events(struct perf_session *self,
1271 struct perf_tool *tool)
1273 union perf_event *event;
1274 uint32_t size, cur_size = 0;
1281 perf_tool__fill_defaults(tool);
1284 cur_size = sizeof(union perf_event);
1286 buf = malloc(cur_size);
1291 err = readn(self->fd, event, sizeof(struct perf_event_header));
1296 pr_err("failed to read event header\n");
1300 if (self->header.needs_swap)
1301 perf_event_header__bswap(&event->header);
1303 size = event->header.size;
1307 if (size > cur_size) {
1308 void *new = realloc(buf, size);
1310 pr_err("failed to allocate memory to read event\n");
1318 p += sizeof(struct perf_event_header);
1320 if (size - sizeof(struct perf_event_header)) {
1321 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1324 pr_err("unexpected end of event stream\n");
1328 pr_err("failed to read event data\n");
1333 if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1334 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1335 head, event->header.size, event->header.type);
1345 if (!session_done())
1351 perf_session__warn_about_errors(self, tool);
1352 perf_session_free_sample_buffers(self);
1356 static union perf_event *
1357 fetch_mmaped_event(struct perf_session *session,
1358 u64 head, size_t mmap_size, char *buf)
1360 union perf_event *event;
1363 * Ensure we have enough space remaining to read
1364 * the size of the event in the headers.
1366 if (head + sizeof(event->header) > mmap_size)
1369 event = (union perf_event *)(buf + head);
1371 if (session->header.needs_swap)
1372 perf_event_header__bswap(&event->header);
1374 if (head + event->header.size > mmap_size)
1381 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1382 * slices. On 32bit we use 32MB.
1384 #if BITS_PER_LONG == 64
1385 #define MMAP_SIZE ULLONG_MAX
1388 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1389 #define NUM_MMAPS 128
1392 int __perf_session__process_events(struct perf_session *session,
1393 u64 data_offset, u64 data_size,
1394 u64 file_size, struct perf_tool *tool)
1396 u64 head, page_offset, file_offset, file_pos, progress_next;
1397 int err, mmap_prot, mmap_flags, map_idx = 0;
1399 char *buf, *mmaps[NUM_MMAPS];
1400 union perf_event *event;
1403 perf_tool__fill_defaults(tool);
1405 page_offset = page_size * (data_offset / page_size);
1406 file_offset = page_offset;
1407 head = data_offset - page_offset;
1409 if (data_offset + data_size < file_size)
1410 file_size = data_offset + data_size;
1412 progress_next = file_size / 16;
1414 mmap_size = MMAP_SIZE;
1415 if (mmap_size > file_size)
1416 mmap_size = file_size;
1418 memset(mmaps, 0, sizeof(mmaps));
1420 mmap_prot = PROT_READ;
1421 mmap_flags = MAP_SHARED;
1423 if (session->header.needs_swap) {
1424 mmap_prot |= PROT_WRITE;
1425 mmap_flags = MAP_PRIVATE;
1428 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1430 if (buf == MAP_FAILED) {
1431 pr_err("failed to mmap file\n");
1435 mmaps[map_idx] = buf;
1436 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1437 file_pos = file_offset + head;
1440 event = fetch_mmaped_event(session, head, mmap_size, buf);
1442 if (mmaps[map_idx]) {
1443 munmap(mmaps[map_idx], mmap_size);
1444 mmaps[map_idx] = NULL;
1447 page_offset = page_size * (head / page_size);
1448 file_offset += page_offset;
1449 head -= page_offset;
1453 size = event->header.size;
1456 perf_session__process_event(session, event, tool, file_pos) < 0) {
1457 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1458 file_offset + head, event->header.size,
1459 event->header.type);
1467 if (file_pos >= progress_next) {
1468 progress_next += file_size / 16;
1469 ui_progress__update(file_pos, file_size,
1470 "Processing events...");
1473 if (file_pos < file_size)
1477 /* do the final flush for ordered samples */
1478 session->ordered_samples.next_flush = ULLONG_MAX;
1479 err = flush_sample_queue(session, tool);
1481 ui_progress__finish();
1482 perf_session__warn_about_errors(session, tool);
1483 perf_session_free_sample_buffers(session);
1487 int perf_session__process_events(struct perf_session *self,
1488 struct perf_tool *tool)
1492 if (perf_session__register_idle_thread(self) == NULL)
1496 err = __perf_session__process_events(self,
1497 self->header.data_offset,
1498 self->header.data_size,
1501 err = __perf_session__process_pipe_events(self, tool);
1506 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1508 if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) {
1509 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1516 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1517 const char *symbol_name, u64 addr)
1521 struct ref_reloc_sym *ref;
1523 ref = zalloc(sizeof(struct ref_reloc_sym));
1527 ref->name = strdup(symbol_name);
1528 if (ref->name == NULL) {
1533 bracket = strchr(ref->name, ']');
1539 for (i = 0; i < MAP__NR_TYPES; ++i) {
1540 struct kmap *kmap = map__kmap(maps[i]);
1541 kmap->ref_reloc_sym = ref;
1547 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1549 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1550 __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1551 machines__fprintf_dsos(&self->machines, fp);
1554 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1555 bool (skip)(struct dso *dso, int parm), int parm)
1557 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, skip, parm);
1558 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, skip, parm);
1561 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1563 struct perf_evsel *pos;
1564 size_t ret = fprintf(fp, "Aggregated stats:\n");
1566 ret += hists__fprintf_nr_events(&session->hists, fp);
1568 list_for_each_entry(pos, &session->evlist->entries, node) {
1569 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1570 ret += hists__fprintf_nr_events(&pos->hists, fp);
1576 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1579 * FIXME: Here we have to actually print all the machines in this
1580 * session, not just the host...
1582 return machine__fprintf(&session->host_machine, fp);
1585 void perf_session__remove_thread(struct perf_session *session,
1589 * FIXME: This one makes no sense, we need to remove the thread from
1590 * the machine it belongs to, perf_session can have many machines, so
1591 * doing it always on ->host_machine is wrong. Fix when auditing all
1592 * the 'perf kvm' code.
1594 machine__remove_thread(&session->host_machine, th);
1597 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1600 struct perf_evsel *pos;
1602 list_for_each_entry(pos, &session->evlist->entries, node) {
1603 if (pos->attr.type == type)
1609 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1610 struct perf_sample *sample, struct machine *machine,
1611 int print_sym, int print_dso, int print_symoffset)
1613 struct addr_location al;
1614 struct callchain_cursor_node *node;
1616 if (perf_event__preprocess_sample(event, machine, &al, sample,
1618 error("problem processing %d event, skipping it.\n",
1619 event->header.type);
1623 if (symbol_conf.use_callchain && sample->callchain) {
1626 if (machine__resolve_callchain(machine, evsel, al.thread,
1627 sample, NULL) != 0) {
1629 error("Failed to resolve callchain. Skipping\n");
1632 callchain_cursor_commit(&callchain_cursor);
1635 node = callchain_cursor_current(&callchain_cursor);
1639 printf("\t%16" PRIx64, node->ip);
1642 symbol__fprintf_symname(node->sym, stdout);
1646 map__fprintf_dsoname(node->map, stdout);
1651 callchain_cursor_advance(&callchain_cursor);
1655 printf("%16" PRIx64, sample->ip);
1658 if (print_symoffset)
1659 symbol__fprintf_symname_offs(al.sym, &al,
1662 symbol__fprintf_symname(al.sym, stdout);
1667 map__fprintf_dsoname(al.map, stdout);
1673 int perf_session__cpu_bitmap(struct perf_session *session,
1674 const char *cpu_list, unsigned long *cpu_bitmap)
1677 struct cpu_map *map;
1679 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1680 struct perf_evsel *evsel;
1682 evsel = perf_session__find_first_evtype(session, i);
1686 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1687 pr_err("File does not contain CPU events. "
1688 "Remove -c option to proceed.\n");
1693 map = cpu_map__new(cpu_list);
1695 pr_err("Invalid cpu_list\n");
1699 for (i = 0; i < map->nr; i++) {
1700 int cpu = map->map[i];
1702 if (cpu >= MAX_NR_CPUS) {
1703 pr_err("Requested CPU %d too large. "
1704 "Consider raising MAX_NR_CPUS\n", cpu);
1708 set_bit(cpu, cpu_bitmap);
1714 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1720 if (session == NULL || fp == NULL)
1723 ret = fstat(session->fd, &st);
1727 fprintf(fp, "# ========\n");
1728 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1729 perf_header__fprintf_info(session, fp, full);
1730 fprintf(fp, "# ========\n#\n");
1734 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1735 const struct perf_evsel_str_handler *assocs,
1738 struct perf_evlist *evlist = session->evlist;
1739 struct event_format *format;
1740 struct perf_evsel *evsel;
1741 char *tracepoint, *name;
1745 for (i = 0; i < nr_assocs; i++) {
1747 tracepoint = strdup(assocs[i].name);
1748 if (tracepoint == NULL)
1752 name = strchr(tracepoint, ':');
1757 format = pevent_find_event_by_name(session->pevent,
1759 if (format == NULL) {
1761 * Adding a handler for an event not in the session,
1767 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1772 if (evsel->handler.func != NULL)
1774 evsel->handler.func = assocs[i].handler;