1 #include <linux/kernel.h>
2 #include <traceevent/event-parse.h>
16 #include "perf_regs.h"
19 static int perf_session__open(struct perf_session *self, bool force)
21 struct stat input_stat;
23 if (!strcmp(self->filename, "-")) {
25 self->fd = STDIN_FILENO;
27 if (perf_session__read_header(self) < 0)
28 pr_err("incompatible file format (rerun with -v to learn more)");
33 self->fd = open(self->filename, O_RDONLY);
37 pr_err("failed to open %s: %s", self->filename, strerror(err));
38 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
39 pr_err(" (try 'perf record' first)");
44 if (fstat(self->fd, &input_stat) < 0)
47 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
48 pr_err("file %s not owned by current user or root\n",
53 if (!input_stat.st_size) {
54 pr_info("zero-sized file (%s), nothing to do!\n",
59 if (perf_session__read_header(self) < 0) {
60 pr_err("incompatible file format (rerun with -v to learn more)");
64 if (!perf_evlist__valid_sample_type(self->evlist)) {
65 pr_err("non matching sample_type");
69 if (!perf_evlist__valid_sample_id_all(self->evlist)) {
70 pr_err("non matching sample_id_all");
74 if (!perf_evlist__valid_read_format(self->evlist)) {
75 pr_err("non matching read_format");
79 self->size = input_stat.st_size;
88 void perf_session__set_id_hdr_size(struct perf_session *session)
90 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
92 machines__set_id_hdr_size(&session->machines, id_hdr_size);
95 int perf_session__create_kernel_maps(struct perf_session *self)
97 int ret = machine__create_kernel_maps(&self->machines.host);
100 ret = machines__create_guest_kernel_maps(&self->machines);
104 static void perf_session__destroy_kernel_maps(struct perf_session *self)
106 machines__destroy_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->repipe = repipe;
132 INIT_LIST_HEAD(&self->ordered_samples.samples);
133 INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
134 INIT_LIST_HEAD(&self->ordered_samples.to_free);
135 machines__init(&self->machines);
137 if (mode == O_RDONLY) {
138 if (perf_session__open(self, force) < 0)
140 perf_session__set_id_hdr_size(self);
141 } else if (mode == O_WRONLY) {
143 * In O_RDONLY mode this will be performed when reading the
144 * kernel MMAP event, in perf_event__process_mmap().
146 if (perf_session__create_kernel_maps(self) < 0)
150 if (tool && tool->ordering_requires_timestamps &&
151 tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
152 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
153 tool->ordered_samples = false;
159 perf_session__delete(self);
163 static void perf_session__delete_dead_threads(struct perf_session *session)
165 machine__delete_dead_threads(&session->machines.host);
168 static void perf_session__delete_threads(struct perf_session *session)
170 machine__delete_threads(&session->machines.host);
173 static void perf_session_env__delete(struct perf_session_env *env)
176 free(env->os_release);
183 free(env->sibling_cores);
184 free(env->sibling_threads);
185 free(env->numa_nodes);
186 free(env->pmu_mappings);
189 void perf_session__delete(struct perf_session *self)
191 perf_session__destroy_kernel_maps(self);
192 perf_session__delete_dead_threads(self);
193 perf_session__delete_threads(self);
194 perf_session_env__delete(&self->header.env);
195 machines__exit(&self->machines);
201 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
203 union perf_event *event
205 struct perf_session *session
208 dump_printf(": unhandled!\n");
212 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
213 union perf_event *event __maybe_unused,
214 struct perf_evlist **pevlist
217 dump_printf(": unhandled!\n");
221 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
222 union perf_event *event __maybe_unused,
223 struct perf_sample *sample __maybe_unused,
224 struct perf_evsel *evsel __maybe_unused,
225 struct machine *machine __maybe_unused)
227 dump_printf(": unhandled!\n");
231 static int process_event_stub(struct perf_tool *tool __maybe_unused,
232 union perf_event *event __maybe_unused,
233 struct perf_sample *sample __maybe_unused,
234 struct machine *machine __maybe_unused)
236 dump_printf(": unhandled!\n");
240 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
241 union perf_event *event __maybe_unused,
242 struct perf_session *perf_session
245 dump_printf(": unhandled!\n");
249 static int process_finished_round(struct perf_tool *tool,
250 union perf_event *event,
251 struct perf_session *session);
253 void perf_tool__fill_defaults(struct perf_tool *tool)
255 if (tool->sample == NULL)
256 tool->sample = process_event_sample_stub;
257 if (tool->mmap == NULL)
258 tool->mmap = process_event_stub;
259 if (tool->comm == NULL)
260 tool->comm = process_event_stub;
261 if (tool->fork == NULL)
262 tool->fork = process_event_stub;
263 if (tool->exit == NULL)
264 tool->exit = process_event_stub;
265 if (tool->lost == NULL)
266 tool->lost = perf_event__process_lost;
267 if (tool->read == NULL)
268 tool->read = process_event_sample_stub;
269 if (tool->throttle == NULL)
270 tool->throttle = process_event_stub;
271 if (tool->unthrottle == NULL)
272 tool->unthrottle = process_event_stub;
273 if (tool->attr == NULL)
274 tool->attr = process_event_synth_attr_stub;
275 if (tool->tracing_data == NULL)
276 tool->tracing_data = process_event_synth_tracing_data_stub;
277 if (tool->build_id == NULL)
278 tool->build_id = process_finished_round_stub;
279 if (tool->finished_round == NULL) {
280 if (tool->ordered_samples)
281 tool->finished_round = process_finished_round;
283 tool->finished_round = process_finished_round_stub;
287 void mem_bswap_32(void *src, int byte_size)
290 while (byte_size > 0) {
292 byte_size -= sizeof(u32);
297 void mem_bswap_64(void *src, int byte_size)
301 while (byte_size > 0) {
303 byte_size -= sizeof(u64);
308 static void swap_sample_id_all(union perf_event *event, void *data)
310 void *end = (void *) event + event->header.size;
311 int size = end - data;
313 BUG_ON(size % sizeof(u64));
314 mem_bswap_64(data, size);
317 static void perf_event__all64_swap(union perf_event *event,
318 bool sample_id_all __maybe_unused)
320 struct perf_event_header *hdr = &event->header;
321 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
324 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
326 event->comm.pid = bswap_32(event->comm.pid);
327 event->comm.tid = bswap_32(event->comm.tid);
330 void *data = &event->comm.comm;
332 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
333 swap_sample_id_all(event, data);
337 static void perf_event__mmap_swap(union perf_event *event,
340 event->mmap.pid = bswap_32(event->mmap.pid);
341 event->mmap.tid = bswap_32(event->mmap.tid);
342 event->mmap.start = bswap_64(event->mmap.start);
343 event->mmap.len = bswap_64(event->mmap.len);
344 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
347 void *data = &event->mmap.filename;
349 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
350 swap_sample_id_all(event, data);
354 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
356 event->fork.pid = bswap_32(event->fork.pid);
357 event->fork.tid = bswap_32(event->fork.tid);
358 event->fork.ppid = bswap_32(event->fork.ppid);
359 event->fork.ptid = bswap_32(event->fork.ptid);
360 event->fork.time = bswap_64(event->fork.time);
363 swap_sample_id_all(event, &event->fork + 1);
366 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
368 event->read.pid = bswap_32(event->read.pid);
369 event->read.tid = bswap_32(event->read.tid);
370 event->read.value = bswap_64(event->read.value);
371 event->read.time_enabled = bswap_64(event->read.time_enabled);
372 event->read.time_running = bswap_64(event->read.time_running);
373 event->read.id = bswap_64(event->read.id);
376 swap_sample_id_all(event, &event->read + 1);
379 static u8 revbyte(u8 b)
381 int rev = (b >> 4) | ((b & 0xf) << 4);
382 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
383 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
388 * XXX this is hack in attempt to carry flags bitfield
389 * throught endian village. ABI says:
391 * Bit-fields are allocated from right to left (least to most significant)
392 * on little-endian implementations and from left to right (most to least
393 * significant) on big-endian implementations.
395 * The above seems to be byte specific, so we need to reverse each
396 * byte of the bitfield. 'Internet' also says this might be implementation
397 * specific and we probably need proper fix and carry perf_event_attr
398 * bitfield flags in separate data file FEAT_ section. Thought this seems
401 static void swap_bitfield(u8 *p, unsigned len)
405 for (i = 0; i < len; i++) {
411 /* exported for swapping attributes in file header */
412 void perf_event__attr_swap(struct perf_event_attr *attr)
414 attr->type = bswap_32(attr->type);
415 attr->size = bswap_32(attr->size);
416 attr->config = bswap_64(attr->config);
417 attr->sample_period = bswap_64(attr->sample_period);
418 attr->sample_type = bswap_64(attr->sample_type);
419 attr->read_format = bswap_64(attr->read_format);
420 attr->wakeup_events = bswap_32(attr->wakeup_events);
421 attr->bp_type = bswap_32(attr->bp_type);
422 attr->bp_addr = bswap_64(attr->bp_addr);
423 attr->bp_len = bswap_64(attr->bp_len);
425 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
428 static void perf_event__hdr_attr_swap(union perf_event *event,
429 bool sample_id_all __maybe_unused)
433 perf_event__attr_swap(&event->attr.attr);
435 size = event->header.size;
436 size -= (void *)&event->attr.id - (void *)event;
437 mem_bswap_64(event->attr.id, size);
440 static void perf_event__event_type_swap(union perf_event *event,
441 bool sample_id_all __maybe_unused)
443 event->event_type.event_type.event_id =
444 bswap_64(event->event_type.event_type.event_id);
447 static void perf_event__tracing_data_swap(union perf_event *event,
448 bool sample_id_all __maybe_unused)
450 event->tracing_data.size = bswap_32(event->tracing_data.size);
453 typedef void (*perf_event__swap_op)(union perf_event *event,
456 static perf_event__swap_op perf_event__swap_ops[] = {
457 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
458 [PERF_RECORD_COMM] = perf_event__comm_swap,
459 [PERF_RECORD_FORK] = perf_event__task_swap,
460 [PERF_RECORD_EXIT] = perf_event__task_swap,
461 [PERF_RECORD_LOST] = perf_event__all64_swap,
462 [PERF_RECORD_READ] = perf_event__read_swap,
463 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
464 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
465 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
466 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
467 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
468 [PERF_RECORD_HEADER_MAX] = NULL,
471 struct sample_queue {
474 union perf_event *event;
475 struct list_head list;
478 static void perf_session_free_sample_buffers(struct perf_session *session)
480 struct ordered_samples *os = &session->ordered_samples;
482 while (!list_empty(&os->to_free)) {
483 struct sample_queue *sq;
485 sq = list_entry(os->to_free.next, struct sample_queue, list);
491 static int perf_session_deliver_event(struct perf_session *session,
492 union perf_event *event,
493 struct perf_sample *sample,
494 struct perf_tool *tool,
497 static int flush_sample_queue(struct perf_session *s,
498 struct perf_tool *tool)
500 struct ordered_samples *os = &s->ordered_samples;
501 struct list_head *head = &os->samples;
502 struct sample_queue *tmp, *iter;
503 struct perf_sample sample;
504 u64 limit = os->next_flush;
505 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
506 unsigned idx = 0, progress_next = os->nr_samples / 16;
509 if (!tool->ordered_samples || !limit)
512 list_for_each_entry_safe(iter, tmp, head, list) {
513 if (iter->timestamp > limit)
516 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
518 pr_err("Can't parse sample, err = %d\n", ret);
520 ret = perf_session_deliver_event(s, iter->event, &sample, tool,
526 os->last_flush = iter->timestamp;
527 list_del(&iter->list);
528 list_add(&iter->list, &os->sample_cache);
529 if (++idx >= progress_next) {
530 progress_next += os->nr_samples / 16;
531 ui_progress__update(idx, os->nr_samples,
532 "Processing time ordered events...");
536 if (list_empty(head)) {
537 os->last_sample = NULL;
538 } else if (last_ts <= limit) {
540 list_entry(head->prev, struct sample_queue, list);
549 * When perf record finishes a pass on every buffers, it records this pseudo
551 * We record the max timestamp t found in the pass n.
552 * Assuming these timestamps are monotonic across cpus, we know that if
553 * a buffer still has events with timestamps below t, they will be all
554 * available and then read in the pass n + 1.
555 * Hence when we start to read the pass n + 2, we can safely flush every
556 * events with timestamps below t.
558 * ============ PASS n =================
561 * cnt1 timestamps | cnt2 timestamps
564 * - | 4 <--- max recorded
566 * ============ PASS n + 1 ==============
569 * cnt1 timestamps | cnt2 timestamps
572 * 5 | 7 <---- max recorded
574 * Flush every events below timestamp 4
576 * ============ PASS n + 2 ==============
579 * cnt1 timestamps | cnt2 timestamps
584 * Flush every events below timestamp 7
587 static int process_finished_round(struct perf_tool *tool,
588 union perf_event *event __maybe_unused,
589 struct perf_session *session)
591 int ret = flush_sample_queue(session, tool);
593 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
598 /* The queue is ordered by time */
599 static void __queue_event(struct sample_queue *new, struct perf_session *s)
601 struct ordered_samples *os = &s->ordered_samples;
602 struct sample_queue *sample = os->last_sample;
603 u64 timestamp = new->timestamp;
607 os->last_sample = new;
610 list_add(&new->list, &os->samples);
611 os->max_timestamp = timestamp;
616 * last_sample might point to some random place in the list as it's
617 * the last queued event. We expect that the new event is close to
620 if (sample->timestamp <= timestamp) {
621 while (sample->timestamp <= timestamp) {
622 p = sample->list.next;
623 if (p == &os->samples) {
624 list_add_tail(&new->list, &os->samples);
625 os->max_timestamp = timestamp;
628 sample = list_entry(p, struct sample_queue, list);
630 list_add_tail(&new->list, &sample->list);
632 while (sample->timestamp > timestamp) {
633 p = sample->list.prev;
634 if (p == &os->samples) {
635 list_add(&new->list, &os->samples);
638 sample = list_entry(p, struct sample_queue, list);
640 list_add(&new->list, &sample->list);
644 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
646 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
647 struct perf_sample *sample, u64 file_offset)
649 struct ordered_samples *os = &s->ordered_samples;
650 struct list_head *sc = &os->sample_cache;
651 u64 timestamp = sample->time;
652 struct sample_queue *new;
654 if (!timestamp || timestamp == ~0ULL)
657 if (timestamp < s->ordered_samples.last_flush) {
658 printf("Warning: Timestamp below last timeslice flush\n");
662 if (!list_empty(sc)) {
663 new = list_entry(sc->next, struct sample_queue, list);
664 list_del(&new->list);
665 } else if (os->sample_buffer) {
666 new = os->sample_buffer + os->sample_buffer_idx;
667 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
668 os->sample_buffer = NULL;
670 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
671 if (!os->sample_buffer)
673 list_add(&os->sample_buffer->list, &os->to_free);
674 os->sample_buffer_idx = 2;
675 new = os->sample_buffer + 1;
678 new->timestamp = timestamp;
679 new->file_offset = file_offset;
682 __queue_event(new, s);
687 static void callchain__printf(struct perf_sample *sample)
691 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
693 for (i = 0; i < sample->callchain->nr; i++)
694 printf("..... %2d: %016" PRIx64 "\n",
695 i, sample->callchain->ips[i]);
698 static void branch_stack__printf(struct perf_sample *sample)
702 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
704 for (i = 0; i < sample->branch_stack->nr; i++)
705 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
706 i, sample->branch_stack->entries[i].from,
707 sample->branch_stack->entries[i].to);
710 static void regs_dump__printf(u64 mask, u64 *regs)
714 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
717 printf(".... %-5s 0x%" PRIx64 "\n",
718 perf_reg_name(rid), val);
722 static void regs_user__printf(struct perf_sample *sample, u64 mask)
724 struct regs_dump *user_regs = &sample->user_regs;
726 if (user_regs->regs) {
727 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
728 regs_dump__printf(mask, user_regs->regs);
732 static void stack_user__printf(struct stack_dump *dump)
734 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
735 dump->size, dump->offset);
738 static void perf_session__print_tstamp(struct perf_session *session,
739 union perf_event *event,
740 struct perf_sample *sample)
742 u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
744 if (event->header.type != PERF_RECORD_SAMPLE &&
745 !perf_evlist__sample_id_all(session->evlist)) {
746 fputs("-1 -1 ", stdout);
750 if ((sample_type & PERF_SAMPLE_CPU))
751 printf("%u ", sample->cpu);
753 if (sample_type & PERF_SAMPLE_TIME)
754 printf("%" PRIu64 " ", sample->time);
757 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
759 printf("... sample_read:\n");
761 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
762 printf("...... time enabled %016" PRIx64 "\n",
763 sample->read.time_enabled);
765 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
766 printf("...... time running %016" PRIx64 "\n",
767 sample->read.time_running);
769 if (read_format & PERF_FORMAT_GROUP) {
772 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
774 for (i = 0; i < sample->read.group.nr; i++) {
775 struct sample_read_value *value;
777 value = &sample->read.group.values[i];
778 printf("..... id %016" PRIx64
779 ", value %016" PRIx64 "\n",
780 value->id, value->value);
783 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
784 sample->read.one.id, sample->read.one.value);
787 static void dump_event(struct perf_session *session, union perf_event *event,
788 u64 file_offset, struct perf_sample *sample)
793 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
794 file_offset, event->header.size, event->header.type);
799 perf_session__print_tstamp(session, event, sample);
801 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
802 event->header.size, perf_event__name(event->header.type));
805 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
806 struct perf_sample *sample)
813 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
814 event->header.misc, sample->pid, sample->tid, sample->ip,
815 sample->period, sample->addr);
817 sample_type = evsel->attr.sample_type;
819 if (sample_type & PERF_SAMPLE_CALLCHAIN)
820 callchain__printf(sample);
822 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
823 branch_stack__printf(sample);
825 if (sample_type & PERF_SAMPLE_REGS_USER)
826 regs_user__printf(sample, evsel->attr.sample_regs_user);
828 if (sample_type & PERF_SAMPLE_STACK_USER)
829 stack_user__printf(&sample->user_stack);
831 if (sample_type & PERF_SAMPLE_WEIGHT)
832 printf("... weight: %" PRIu64 "\n", sample->weight);
834 if (sample_type & PERF_SAMPLE_DATA_SRC)
835 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
837 if (sample_type & PERF_SAMPLE_READ)
838 sample_read__printf(sample, evsel->attr.read_format);
841 static struct machine *
842 perf_session__find_machine_for_cpumode(struct perf_session *session,
843 union perf_event *event,
844 struct perf_sample *sample)
846 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
849 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
850 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
853 if (event->header.type == PERF_RECORD_MMAP)
854 pid = event->mmap.pid;
858 return perf_session__findnew_machine(session, pid);
861 return &session->machines.host;
864 static int deliver_sample_value(struct perf_session *session,
865 struct perf_tool *tool,
866 union perf_event *event,
867 struct perf_sample *sample,
868 struct sample_read_value *v,
869 struct machine *machine)
871 struct perf_sample_id *sid;
873 sid = perf_evlist__id2sid(session->evlist, v->id);
876 sample->period = v->value - sid->period;
877 sid->period = v->value;
880 if (!sid || sid->evsel == NULL) {
881 ++session->stats.nr_unknown_id;
885 return tool->sample(tool, event, sample, sid->evsel, machine);
888 static int deliver_sample_group(struct perf_session *session,
889 struct perf_tool *tool,
890 union perf_event *event,
891 struct perf_sample *sample,
892 struct machine *machine)
897 for (i = 0; i < sample->read.group.nr; i++) {
898 ret = deliver_sample_value(session, tool, event, sample,
899 &sample->read.group.values[i],
909 perf_session__deliver_sample(struct perf_session *session,
910 struct perf_tool *tool,
911 union perf_event *event,
912 struct perf_sample *sample,
913 struct perf_evsel *evsel,
914 struct machine *machine)
916 /* We know evsel != NULL. */
917 u64 sample_type = evsel->attr.sample_type;
918 u64 read_format = evsel->attr.read_format;
920 /* Standard sample delievery. */
921 if (!(sample_type & PERF_SAMPLE_READ))
922 return tool->sample(tool, event, sample, evsel, machine);
924 /* For PERF_SAMPLE_READ we have either single or group mode. */
925 if (read_format & PERF_FORMAT_GROUP)
926 return deliver_sample_group(session, tool, event, sample,
929 return deliver_sample_value(session, tool, event, sample,
930 &sample->read.one, machine);
933 static int perf_session_deliver_event(struct perf_session *session,
934 union perf_event *event,
935 struct perf_sample *sample,
936 struct perf_tool *tool,
939 struct perf_evsel *evsel;
940 struct machine *machine;
942 dump_event(session, event, file_offset, sample);
944 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
945 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
947 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
948 * because the tools right now may apply filters, discarding
949 * some of the samples. For consistency, in the future we
950 * should have something like nr_filtered_samples and remove
951 * the sample->period from total_sample_period, etc, KISS for
954 * Also testing against NULL allows us to handle files without
955 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
956 * future probably it'll be a good idea to restrict event
957 * processing via perf_session to files with both set.
959 hists__inc_nr_events(&evsel->hists, event->header.type);
962 machine = perf_session__find_machine_for_cpumode(session, event,
965 switch (event->header.type) {
966 case PERF_RECORD_SAMPLE:
967 dump_sample(evsel, event, sample);
969 ++session->stats.nr_unknown_id;
972 if (machine == NULL) {
973 ++session->stats.nr_unprocessable_samples;
976 return perf_session__deliver_sample(session, tool, event,
977 sample, evsel, machine);
978 case PERF_RECORD_MMAP:
979 return tool->mmap(tool, event, sample, machine);
980 case PERF_RECORD_COMM:
981 return tool->comm(tool, event, sample, machine);
982 case PERF_RECORD_FORK:
983 return tool->fork(tool, event, sample, machine);
984 case PERF_RECORD_EXIT:
985 return tool->exit(tool, event, sample, machine);
986 case PERF_RECORD_LOST:
987 if (tool->lost == perf_event__process_lost)
988 session->stats.total_lost += event->lost.lost;
989 return tool->lost(tool, event, sample, machine);
990 case PERF_RECORD_READ:
991 return tool->read(tool, event, sample, evsel, machine);
992 case PERF_RECORD_THROTTLE:
993 return tool->throttle(tool, event, sample, machine);
994 case PERF_RECORD_UNTHROTTLE:
995 return tool->unthrottle(tool, event, sample, machine);
997 ++session->stats.nr_unknown_events;
1002 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1003 struct perf_tool *tool, u64 file_offset)
1007 dump_event(session, event, file_offset, NULL);
1009 /* These events are processed right away */
1010 switch (event->header.type) {
1011 case PERF_RECORD_HEADER_ATTR:
1012 err = tool->attr(tool, event, &session->evlist);
1014 perf_session__set_id_hdr_size(session);
1016 case PERF_RECORD_HEADER_TRACING_DATA:
1017 /* setup for reading amidst mmap */
1018 lseek(session->fd, file_offset, SEEK_SET);
1019 return tool->tracing_data(tool, event, session);
1020 case PERF_RECORD_HEADER_BUILD_ID:
1021 return tool->build_id(tool, event, session);
1022 case PERF_RECORD_FINISHED_ROUND:
1023 return tool->finished_round(tool, event, session);
1029 static void event_swap(union perf_event *event, bool sample_id_all)
1031 perf_event__swap_op swap;
1033 swap = perf_event__swap_ops[event->header.type];
1035 swap(event, sample_id_all);
1038 static int perf_session__process_event(struct perf_session *session,
1039 union perf_event *event,
1040 struct perf_tool *tool,
1043 struct perf_sample sample;
1046 if (session->header.needs_swap)
1047 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1049 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1052 events_stats__inc(&session->stats, event->header.type);
1054 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1055 return perf_session__process_user_event(session, event, tool, file_offset);
1058 * For all kernel events we get the sample data
1060 ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1064 if (tool->ordered_samples) {
1065 ret = perf_session_queue_event(session, event, &sample,
1071 return perf_session_deliver_event(session, event, &sample, tool,
1075 void perf_event_header__bswap(struct perf_event_header *self)
1077 self->type = bswap_32(self->type);
1078 self->misc = bswap_16(self->misc);
1079 self->size = bswap_16(self->size);
1082 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1084 return machine__findnew_thread(&session->machines.host, 0, pid);
1087 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1089 struct thread *thread = perf_session__findnew(self, 0);
1091 if (thread == NULL || thread__set_comm(thread, "swapper")) {
1092 pr_err("problem inserting idle task.\n");
1099 static void perf_session__warn_about_errors(const struct perf_session *session,
1100 const struct perf_tool *tool)
1102 if (tool->lost == perf_event__process_lost &&
1103 session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1104 ui__warning("Processed %d events and lost %d chunks!\n\n"
1105 "Check IO/CPU overload!\n\n",
1106 session->stats.nr_events[0],
1107 session->stats.nr_events[PERF_RECORD_LOST]);
1110 if (session->stats.nr_unknown_events != 0) {
1111 ui__warning("Found %u unknown events!\n\n"
1112 "Is this an older tool processing a perf.data "
1113 "file generated by a more recent tool?\n\n"
1114 "If that is not the case, consider "
1115 "reporting to linux-kernel@vger.kernel.org.\n\n",
1116 session->stats.nr_unknown_events);
1119 if (session->stats.nr_unknown_id != 0) {
1120 ui__warning("%u samples with id not present in the header\n",
1121 session->stats.nr_unknown_id);
1124 if (session->stats.nr_invalid_chains != 0) {
1125 ui__warning("Found invalid callchains!\n\n"
1126 "%u out of %u events were discarded for this reason.\n\n"
1127 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1128 session->stats.nr_invalid_chains,
1129 session->stats.nr_events[PERF_RECORD_SAMPLE]);
1132 if (session->stats.nr_unprocessable_samples != 0) {
1133 ui__warning("%u unprocessable samples recorded.\n"
1134 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1135 session->stats.nr_unprocessable_samples);
1139 #define session_done() (*(volatile int *)(&session_done))
1140 volatile int session_done;
1142 static int __perf_session__process_pipe_events(struct perf_session *self,
1143 struct perf_tool *tool)
1145 union perf_event *event;
1146 uint32_t size, cur_size = 0;
1153 perf_tool__fill_defaults(tool);
1156 cur_size = sizeof(union perf_event);
1158 buf = malloc(cur_size);
1163 err = readn(self->fd, event, sizeof(struct perf_event_header));
1168 pr_err("failed to read event header\n");
1172 if (self->header.needs_swap)
1173 perf_event_header__bswap(&event->header);
1175 size = event->header.size;
1176 if (size < sizeof(struct perf_event_header)) {
1177 pr_err("bad event header size\n");
1181 if (size > cur_size) {
1182 void *new = realloc(buf, size);
1184 pr_err("failed to allocate memory to read event\n");
1192 p += sizeof(struct perf_event_header);
1194 if (size - sizeof(struct perf_event_header)) {
1195 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1198 pr_err("unexpected end of event stream\n");
1202 pr_err("failed to read event data\n");
1207 if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1208 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1209 head, event->header.size, event->header.type);
1219 if (!session_done())
1225 perf_session__warn_about_errors(self, tool);
1226 perf_session_free_sample_buffers(self);
1230 static union perf_event *
1231 fetch_mmaped_event(struct perf_session *session,
1232 u64 head, size_t mmap_size, char *buf)
1234 union perf_event *event;
1237 * Ensure we have enough space remaining to read
1238 * the size of the event in the headers.
1240 if (head + sizeof(event->header) > mmap_size)
1243 event = (union perf_event *)(buf + head);
1245 if (session->header.needs_swap)
1246 perf_event_header__bswap(&event->header);
1248 if (head + event->header.size > mmap_size) {
1249 /* We're not fetching the event so swap back again */
1250 if (session->header.needs_swap)
1251 perf_event_header__bswap(&event->header);
1259 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1260 * slices. On 32bit we use 32MB.
1262 #if BITS_PER_LONG == 64
1263 #define MMAP_SIZE ULLONG_MAX
1266 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1267 #define NUM_MMAPS 128
1270 int __perf_session__process_events(struct perf_session *session,
1271 u64 data_offset, u64 data_size,
1272 u64 file_size, struct perf_tool *tool)
1274 u64 head, page_offset, file_offset, file_pos, progress_next;
1275 int err, mmap_prot, mmap_flags, map_idx = 0;
1277 char *buf, *mmaps[NUM_MMAPS];
1278 union perf_event *event;
1281 perf_tool__fill_defaults(tool);
1283 page_offset = page_size * (data_offset / page_size);
1284 file_offset = page_offset;
1285 head = data_offset - page_offset;
1287 if (data_offset + data_size < file_size)
1288 file_size = data_offset + data_size;
1290 progress_next = file_size / 16;
1292 mmap_size = MMAP_SIZE;
1293 if (mmap_size > file_size)
1294 mmap_size = file_size;
1296 memset(mmaps, 0, sizeof(mmaps));
1298 mmap_prot = PROT_READ;
1299 mmap_flags = MAP_SHARED;
1301 if (session->header.needs_swap) {
1302 mmap_prot |= PROT_WRITE;
1303 mmap_flags = MAP_PRIVATE;
1306 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1308 if (buf == MAP_FAILED) {
1309 pr_err("failed to mmap file\n");
1313 mmaps[map_idx] = buf;
1314 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1315 file_pos = file_offset + head;
1318 event = fetch_mmaped_event(session, head, mmap_size, buf);
1320 if (mmaps[map_idx]) {
1321 munmap(mmaps[map_idx], mmap_size);
1322 mmaps[map_idx] = NULL;
1325 page_offset = page_size * (head / page_size);
1326 file_offset += page_offset;
1327 head -= page_offset;
1331 size = event->header.size;
1333 if (size < sizeof(struct perf_event_header) ||
1334 perf_session__process_event(session, event, tool, file_pos) < 0) {
1335 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1336 file_offset + head, event->header.size,
1337 event->header.type);
1345 if (file_pos >= progress_next) {
1346 progress_next += file_size / 16;
1347 ui_progress__update(file_pos, file_size,
1348 "Processing events...");
1351 if (file_pos < file_size)
1355 /* do the final flush for ordered samples */
1356 session->ordered_samples.next_flush = ULLONG_MAX;
1357 err = flush_sample_queue(session, tool);
1359 ui_progress__finish();
1360 perf_session__warn_about_errors(session, tool);
1361 perf_session_free_sample_buffers(session);
1365 int perf_session__process_events(struct perf_session *self,
1366 struct perf_tool *tool)
1370 if (perf_session__register_idle_thread(self) == NULL)
1374 err = __perf_session__process_events(self,
1375 self->header.data_offset,
1376 self->header.data_size,
1379 err = __perf_session__process_pipe_events(self, tool);
1384 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1386 struct perf_evsel *evsel;
1388 list_for_each_entry(evsel, &session->evlist->entries, node) {
1389 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1393 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1397 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1398 const char *symbol_name, u64 addr)
1402 struct ref_reloc_sym *ref;
1404 ref = zalloc(sizeof(struct ref_reloc_sym));
1408 ref->name = strdup(symbol_name);
1409 if (ref->name == NULL) {
1414 bracket = strchr(ref->name, ']');
1420 for (i = 0; i < MAP__NR_TYPES; ++i) {
1421 struct kmap *kmap = map__kmap(maps[i]);
1422 kmap->ref_reloc_sym = ref;
1428 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1430 return machines__fprintf_dsos(&self->machines, fp);
1433 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1434 bool (skip)(struct dso *dso, int parm), int parm)
1436 return machines__fprintf_dsos_buildid(&self->machines, fp, skip, parm);
1439 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1441 struct perf_evsel *pos;
1442 size_t ret = fprintf(fp, "Aggregated stats:\n");
1444 ret += events_stats__fprintf(&session->stats, fp);
1446 list_for_each_entry(pos, &session->evlist->entries, node) {
1447 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1448 ret += events_stats__fprintf(&pos->hists.stats, fp);
1454 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1457 * FIXME: Here we have to actually print all the machines in this
1458 * session, not just the host...
1460 return machine__fprintf(&session->machines.host, fp);
1463 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1466 struct perf_evsel *pos;
1468 list_for_each_entry(pos, &session->evlist->entries, node) {
1469 if (pos->attr.type == type)
1475 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1476 struct perf_sample *sample, struct machine *machine,
1477 unsigned int print_opts, unsigned int stack_depth)
1479 struct addr_location al;
1480 struct callchain_cursor_node *node;
1481 int print_ip = print_opts & PRINT_IP_OPT_IP;
1482 int print_sym = print_opts & PRINT_IP_OPT_SYM;
1483 int print_dso = print_opts & PRINT_IP_OPT_DSO;
1484 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1485 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1486 char s = print_oneline ? ' ' : '\t';
1488 if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
1489 error("problem processing %d event, skipping it.\n",
1490 event->header.type);
1494 if (symbol_conf.use_callchain && sample->callchain) {
1496 if (machine__resolve_callchain(machine, evsel, al.thread,
1497 sample, NULL, NULL) != 0) {
1499 error("Failed to resolve callchain. Skipping\n");
1502 callchain_cursor_commit(&callchain_cursor);
1504 while (stack_depth) {
1505 node = callchain_cursor_current(&callchain_cursor);
1510 printf("%c%16" PRIx64, s, node->ip);
1514 if (print_symoffset) {
1516 symbol__fprintf_symname_offs(node->sym, &al, stdout);
1518 symbol__fprintf_symname(node->sym, stdout);
1523 map__fprintf_dsoname(node->map, stdout);
1530 callchain_cursor_advance(&callchain_cursor);
1537 printf("%16" PRIx64, sample->ip);
1541 if (print_symoffset)
1542 symbol__fprintf_symname_offs(al.sym, &al,
1545 symbol__fprintf_symname(al.sym, stdout);
1550 map__fprintf_dsoname(al.map, stdout);
1556 int perf_session__cpu_bitmap(struct perf_session *session,
1557 const char *cpu_list, unsigned long *cpu_bitmap)
1560 struct cpu_map *map;
1562 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1563 struct perf_evsel *evsel;
1565 evsel = perf_session__find_first_evtype(session, i);
1569 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1570 pr_err("File does not contain CPU events. "
1571 "Remove -c option to proceed.\n");
1576 map = cpu_map__new(cpu_list);
1578 pr_err("Invalid cpu_list\n");
1582 for (i = 0; i < map->nr; i++) {
1583 int cpu = map->map[i];
1585 if (cpu >= MAX_NR_CPUS) {
1586 pr_err("Requested CPU %d too large. "
1587 "Consider raising MAX_NR_CPUS\n", cpu);
1591 set_bit(cpu, cpu_bitmap);
1597 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1603 if (session == NULL || fp == NULL)
1606 ret = fstat(session->fd, &st);
1610 fprintf(fp, "# ========\n");
1611 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1612 perf_header__fprintf_info(session, fp, full);
1613 fprintf(fp, "# ========\n#\n");
1617 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1618 const struct perf_evsel_str_handler *assocs,
1621 struct perf_evlist *evlist = session->evlist;
1622 struct event_format *format;
1623 struct perf_evsel *evsel;
1624 char *tracepoint, *name;
1628 for (i = 0; i < nr_assocs; i++) {
1630 tracepoint = strdup(assocs[i].name);
1631 if (tracepoint == NULL)
1635 name = strchr(tracepoint, ':');
1640 format = pevent_find_event_by_name(session->pevent,
1642 if (format == NULL) {
1644 * Adding a handler for an event not in the session,
1650 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1655 if (evsel->handler.func != NULL)
1657 evsel->handler.func = assocs[i].handler;