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 *session)
21 struct perf_data_file *file = session->file;
23 if (perf_session__read_header(session) < 0) {
24 pr_err("incompatible file format (rerun with -v to learn more)");
28 if (perf_data_file__is_pipe(file))
31 if (!perf_evlist__valid_sample_type(session->evlist)) {
32 pr_err("non matching sample_type");
36 if (!perf_evlist__valid_sample_id_all(session->evlist)) {
37 pr_err("non matching sample_id_all");
41 if (!perf_evlist__valid_read_format(session->evlist)) {
42 pr_err("non matching read_format");
49 void perf_session__set_id_hdr_size(struct perf_session *session)
51 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
53 machines__set_id_hdr_size(&session->machines, id_hdr_size);
56 int perf_session__create_kernel_maps(struct perf_session *session)
58 int ret = machine__create_kernel_maps(&session->machines.host);
61 ret = machines__create_guest_kernel_maps(&session->machines);
65 static void perf_session__destroy_kernel_maps(struct perf_session *session)
67 machines__destroy_kernel_maps(&session->machines);
70 struct perf_session *perf_session__new(struct perf_data_file *file,
71 bool repipe, struct perf_tool *tool)
73 struct perf_session *session = zalloc(sizeof(*session));
78 session->repipe = repipe;
79 INIT_LIST_HEAD(&session->ordered_samples.samples);
80 INIT_LIST_HEAD(&session->ordered_samples.sample_cache);
81 INIT_LIST_HEAD(&session->ordered_samples.to_free);
82 machines__init(&session->machines);
85 if (perf_data_file__open(file))
90 if (perf_data_file__is_read(file)) {
91 if (perf_session__open(session) < 0)
94 perf_session__set_id_hdr_size(session);
98 if (!file || perf_data_file__is_write(file)) {
100 * In O_RDONLY mode this will be performed when reading the
101 * kernel MMAP event, in perf_event__process_mmap().
103 if (perf_session__create_kernel_maps(session) < 0)
107 if (tool && tool->ordering_requires_timestamps &&
108 tool->ordered_samples && !perf_evlist__sample_id_all(session->evlist)) {
109 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
110 tool->ordered_samples = false;
116 perf_data_file__close(file);
118 perf_session__delete(session);
123 static void perf_session__delete_dead_threads(struct perf_session *session)
125 machine__delete_dead_threads(&session->machines.host);
128 static void perf_session__delete_threads(struct perf_session *session)
130 machine__delete_threads(&session->machines.host);
133 static void perf_session_env__delete(struct perf_session_env *env)
136 free(env->os_release);
143 free(env->sibling_cores);
144 free(env->sibling_threads);
145 free(env->numa_nodes);
146 free(env->pmu_mappings);
149 void perf_session__delete(struct perf_session *session)
151 perf_session__destroy_kernel_maps(session);
152 perf_session__delete_dead_threads(session);
153 perf_session__delete_threads(session);
154 perf_session_env__delete(&session->header.env);
155 machines__exit(&session->machines);
157 perf_data_file__close(session->file);
162 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
164 union perf_event *event
166 struct perf_session *session
169 dump_printf(": unhandled!\n");
173 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
174 union perf_event *event __maybe_unused,
175 struct perf_evlist **pevlist
178 dump_printf(": unhandled!\n");
182 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
183 union perf_event *event __maybe_unused,
184 struct perf_sample *sample __maybe_unused,
185 struct perf_evsel *evsel __maybe_unused,
186 struct machine *machine __maybe_unused)
188 dump_printf(": unhandled!\n");
192 static int process_event_stub(struct perf_tool *tool __maybe_unused,
193 union perf_event *event __maybe_unused,
194 struct perf_sample *sample __maybe_unused,
195 struct machine *machine __maybe_unused)
197 dump_printf(": unhandled!\n");
201 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
202 union perf_event *event __maybe_unused,
203 struct perf_session *perf_session
206 dump_printf(": unhandled!\n");
210 static int process_finished_round(struct perf_tool *tool,
211 union perf_event *event,
212 struct perf_session *session);
214 void perf_tool__fill_defaults(struct perf_tool *tool)
216 if (tool->sample == NULL)
217 tool->sample = process_event_sample_stub;
218 if (tool->mmap == NULL)
219 tool->mmap = process_event_stub;
220 if (tool->mmap2 == NULL)
221 tool->mmap2 = process_event_stub;
222 if (tool->comm == NULL)
223 tool->comm = process_event_stub;
224 if (tool->fork == NULL)
225 tool->fork = process_event_stub;
226 if (tool->exit == NULL)
227 tool->exit = process_event_stub;
228 if (tool->lost == NULL)
229 tool->lost = perf_event__process_lost;
230 if (tool->read == NULL)
231 tool->read = process_event_sample_stub;
232 if (tool->throttle == NULL)
233 tool->throttle = process_event_stub;
234 if (tool->unthrottle == NULL)
235 tool->unthrottle = process_event_stub;
236 if (tool->attr == NULL)
237 tool->attr = process_event_synth_attr_stub;
238 if (tool->tracing_data == NULL)
239 tool->tracing_data = process_event_synth_tracing_data_stub;
240 if (tool->build_id == NULL)
241 tool->build_id = process_finished_round_stub;
242 if (tool->finished_round == NULL) {
243 if (tool->ordered_samples)
244 tool->finished_round = process_finished_round;
246 tool->finished_round = process_finished_round_stub;
250 void mem_bswap_32(void *src, int byte_size)
253 while (byte_size > 0) {
255 byte_size -= sizeof(u32);
260 void mem_bswap_64(void *src, int byte_size)
264 while (byte_size > 0) {
266 byte_size -= sizeof(u64);
271 static void swap_sample_id_all(union perf_event *event, void *data)
273 void *end = (void *) event + event->header.size;
274 int size = end - data;
276 BUG_ON(size % sizeof(u64));
277 mem_bswap_64(data, size);
280 static void perf_event__all64_swap(union perf_event *event,
281 bool sample_id_all __maybe_unused)
283 struct perf_event_header *hdr = &event->header;
284 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
287 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
289 event->comm.pid = bswap_32(event->comm.pid);
290 event->comm.tid = bswap_32(event->comm.tid);
293 void *data = &event->comm.comm;
295 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
296 swap_sample_id_all(event, data);
300 static void perf_event__mmap_swap(union perf_event *event,
303 event->mmap.pid = bswap_32(event->mmap.pid);
304 event->mmap.tid = bswap_32(event->mmap.tid);
305 event->mmap.start = bswap_64(event->mmap.start);
306 event->mmap.len = bswap_64(event->mmap.len);
307 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
310 void *data = &event->mmap.filename;
312 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
313 swap_sample_id_all(event, data);
317 static void perf_event__mmap2_swap(union perf_event *event,
320 event->mmap2.pid = bswap_32(event->mmap2.pid);
321 event->mmap2.tid = bswap_32(event->mmap2.tid);
322 event->mmap2.start = bswap_64(event->mmap2.start);
323 event->mmap2.len = bswap_64(event->mmap2.len);
324 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
325 event->mmap2.maj = bswap_32(event->mmap2.maj);
326 event->mmap2.min = bswap_32(event->mmap2.min);
327 event->mmap2.ino = bswap_64(event->mmap2.ino);
330 void *data = &event->mmap2.filename;
332 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
333 swap_sample_id_all(event, data);
336 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
338 event->fork.pid = bswap_32(event->fork.pid);
339 event->fork.tid = bswap_32(event->fork.tid);
340 event->fork.ppid = bswap_32(event->fork.ppid);
341 event->fork.ptid = bswap_32(event->fork.ptid);
342 event->fork.time = bswap_64(event->fork.time);
345 swap_sample_id_all(event, &event->fork + 1);
348 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
350 event->read.pid = bswap_32(event->read.pid);
351 event->read.tid = bswap_32(event->read.tid);
352 event->read.value = bswap_64(event->read.value);
353 event->read.time_enabled = bswap_64(event->read.time_enabled);
354 event->read.time_running = bswap_64(event->read.time_running);
355 event->read.id = bswap_64(event->read.id);
358 swap_sample_id_all(event, &event->read + 1);
361 static void perf_event__throttle_swap(union perf_event *event,
364 event->throttle.time = bswap_64(event->throttle.time);
365 event->throttle.id = bswap_64(event->throttle.id);
366 event->throttle.stream_id = bswap_64(event->throttle.stream_id);
369 swap_sample_id_all(event, &event->throttle + 1);
372 static u8 revbyte(u8 b)
374 int rev = (b >> 4) | ((b & 0xf) << 4);
375 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
376 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
381 * XXX this is hack in attempt to carry flags bitfield
382 * throught endian village. ABI says:
384 * Bit-fields are allocated from right to left (least to most significant)
385 * on little-endian implementations and from left to right (most to least
386 * significant) on big-endian implementations.
388 * The above seems to be byte specific, so we need to reverse each
389 * byte of the bitfield. 'Internet' also says this might be implementation
390 * specific and we probably need proper fix and carry perf_event_attr
391 * bitfield flags in separate data file FEAT_ section. Thought this seems
394 static void swap_bitfield(u8 *p, unsigned len)
398 for (i = 0; i < len; i++) {
404 /* exported for swapping attributes in file header */
405 void perf_event__attr_swap(struct perf_event_attr *attr)
407 attr->type = bswap_32(attr->type);
408 attr->size = bswap_32(attr->size);
409 attr->config = bswap_64(attr->config);
410 attr->sample_period = bswap_64(attr->sample_period);
411 attr->sample_type = bswap_64(attr->sample_type);
412 attr->read_format = bswap_64(attr->read_format);
413 attr->wakeup_events = bswap_32(attr->wakeup_events);
414 attr->bp_type = bswap_32(attr->bp_type);
415 attr->bp_addr = bswap_64(attr->bp_addr);
416 attr->bp_len = bswap_64(attr->bp_len);
417 attr->branch_sample_type = bswap_64(attr->branch_sample_type);
418 attr->sample_regs_user = bswap_64(attr->sample_regs_user);
419 attr->sample_stack_user = bswap_32(attr->sample_stack_user);
421 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
424 static void perf_event__hdr_attr_swap(union perf_event *event,
425 bool sample_id_all __maybe_unused)
429 perf_event__attr_swap(&event->attr.attr);
431 size = event->header.size;
432 size -= (void *)&event->attr.id - (void *)event;
433 mem_bswap_64(event->attr.id, size);
436 static void perf_event__event_type_swap(union perf_event *event,
437 bool sample_id_all __maybe_unused)
439 event->event_type.event_type.event_id =
440 bswap_64(event->event_type.event_type.event_id);
443 static void perf_event__tracing_data_swap(union perf_event *event,
444 bool sample_id_all __maybe_unused)
446 event->tracing_data.size = bswap_32(event->tracing_data.size);
449 typedef void (*perf_event__swap_op)(union perf_event *event,
452 static perf_event__swap_op perf_event__swap_ops[] = {
453 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
454 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap,
455 [PERF_RECORD_COMM] = perf_event__comm_swap,
456 [PERF_RECORD_FORK] = perf_event__task_swap,
457 [PERF_RECORD_EXIT] = perf_event__task_swap,
458 [PERF_RECORD_LOST] = perf_event__all64_swap,
459 [PERF_RECORD_READ] = perf_event__read_swap,
460 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap,
461 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap,
462 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
463 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
464 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
465 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
466 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
467 [PERF_RECORD_HEADER_MAX] = NULL,
470 struct sample_queue {
473 union perf_event *event;
474 struct list_head list;
477 static void perf_session_free_sample_buffers(struct perf_session *session)
479 struct ordered_samples *os = &session->ordered_samples;
481 while (!list_empty(&os->to_free)) {
482 struct sample_queue *sq;
484 sq = list_entry(os->to_free.next, struct sample_queue, list);
490 static int perf_session_deliver_event(struct perf_session *session,
491 union perf_event *event,
492 struct perf_sample *sample,
493 struct perf_tool *tool,
496 static int flush_sample_queue(struct perf_session *s,
497 struct perf_tool *tool)
499 struct ordered_samples *os = &s->ordered_samples;
500 struct list_head *head = &os->samples;
501 struct sample_queue *tmp, *iter;
502 struct perf_sample sample;
503 u64 limit = os->next_flush;
504 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
505 bool show_progress = limit == ULLONG_MAX;
506 struct ui_progress prog;
509 if (!tool->ordered_samples || !limit)
513 ui_progress__init(&prog, os->nr_samples, "Processing time ordered events...");
515 list_for_each_entry_safe(iter, tmp, head, list) {
519 if (iter->timestamp > limit)
522 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
524 pr_err("Can't parse sample, err = %d\n", ret);
526 ret = perf_session_deliver_event(s, iter->event, &sample, tool,
532 os->last_flush = iter->timestamp;
533 list_del(&iter->list);
534 list_add(&iter->list, &os->sample_cache);
537 ui_progress__update(&prog, 1);
540 if (list_empty(head)) {
541 os->last_sample = NULL;
542 } else if (last_ts <= limit) {
544 list_entry(head->prev, struct sample_queue, list);
553 * When perf record finishes a pass on every buffers, it records this pseudo
555 * We record the max timestamp t found in the pass n.
556 * Assuming these timestamps are monotonic across cpus, we know that if
557 * a buffer still has events with timestamps below t, they will be all
558 * available and then read in the pass n + 1.
559 * Hence when we start to read the pass n + 2, we can safely flush every
560 * events with timestamps below t.
562 * ============ PASS n =================
565 * cnt1 timestamps | cnt2 timestamps
568 * - | 4 <--- max recorded
570 * ============ PASS n + 1 ==============
573 * cnt1 timestamps | cnt2 timestamps
576 * 5 | 7 <---- max recorded
578 * Flush every events below timestamp 4
580 * ============ PASS n + 2 ==============
583 * cnt1 timestamps | cnt2 timestamps
588 * Flush every events below timestamp 7
591 static int process_finished_round(struct perf_tool *tool,
592 union perf_event *event __maybe_unused,
593 struct perf_session *session)
595 int ret = flush_sample_queue(session, tool);
597 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
602 /* The queue is ordered by time */
603 static void __queue_event(struct sample_queue *new, struct perf_session *s)
605 struct ordered_samples *os = &s->ordered_samples;
606 struct sample_queue *sample = os->last_sample;
607 u64 timestamp = new->timestamp;
611 os->last_sample = new;
614 list_add(&new->list, &os->samples);
615 os->max_timestamp = timestamp;
620 * last_sample might point to some random place in the list as it's
621 * the last queued event. We expect that the new event is close to
624 if (sample->timestamp <= timestamp) {
625 while (sample->timestamp <= timestamp) {
626 p = sample->list.next;
627 if (p == &os->samples) {
628 list_add_tail(&new->list, &os->samples);
629 os->max_timestamp = timestamp;
632 sample = list_entry(p, struct sample_queue, list);
634 list_add_tail(&new->list, &sample->list);
636 while (sample->timestamp > timestamp) {
637 p = sample->list.prev;
638 if (p == &os->samples) {
639 list_add(&new->list, &os->samples);
642 sample = list_entry(p, struct sample_queue, list);
644 list_add(&new->list, &sample->list);
648 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
650 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
651 struct perf_sample *sample, u64 file_offset)
653 struct ordered_samples *os = &s->ordered_samples;
654 struct list_head *sc = &os->sample_cache;
655 u64 timestamp = sample->time;
656 struct sample_queue *new;
658 if (!timestamp || timestamp == ~0ULL)
661 if (timestamp < s->ordered_samples.last_flush) {
662 printf("Warning: Timestamp below last timeslice flush\n");
666 if (!list_empty(sc)) {
667 new = list_entry(sc->next, struct sample_queue, list);
668 list_del(&new->list);
669 } else if (os->sample_buffer) {
670 new = os->sample_buffer + os->sample_buffer_idx;
671 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
672 os->sample_buffer = NULL;
674 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
675 if (!os->sample_buffer)
677 list_add(&os->sample_buffer->list, &os->to_free);
678 os->sample_buffer_idx = 2;
679 new = os->sample_buffer + 1;
682 new->timestamp = timestamp;
683 new->file_offset = file_offset;
686 __queue_event(new, s);
691 static void callchain__printf(struct perf_sample *sample)
695 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
697 for (i = 0; i < sample->callchain->nr; i++)
698 printf("..... %2d: %016" PRIx64 "\n",
699 i, sample->callchain->ips[i]);
702 static void branch_stack__printf(struct perf_sample *sample)
706 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
708 for (i = 0; i < sample->branch_stack->nr; i++)
709 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
710 i, sample->branch_stack->entries[i].from,
711 sample->branch_stack->entries[i].to);
714 static void regs_dump__printf(u64 mask, u64 *regs)
718 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
721 printf(".... %-5s 0x%" PRIx64 "\n",
722 perf_reg_name(rid), val);
726 static void regs_user__printf(struct perf_sample *sample, u64 mask)
728 struct regs_dump *user_regs = &sample->user_regs;
730 if (user_regs->regs) {
731 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
732 regs_dump__printf(mask, user_regs->regs);
736 static void stack_user__printf(struct stack_dump *dump)
738 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
739 dump->size, dump->offset);
742 static void perf_session__print_tstamp(struct perf_session *session,
743 union perf_event *event,
744 struct perf_sample *sample)
746 u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
748 if (event->header.type != PERF_RECORD_SAMPLE &&
749 !perf_evlist__sample_id_all(session->evlist)) {
750 fputs("-1 -1 ", stdout);
754 if ((sample_type & PERF_SAMPLE_CPU))
755 printf("%u ", sample->cpu);
757 if (sample_type & PERF_SAMPLE_TIME)
758 printf("%" PRIu64 " ", sample->time);
761 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
763 printf("... sample_read:\n");
765 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
766 printf("...... time enabled %016" PRIx64 "\n",
767 sample->read.time_enabled);
769 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
770 printf("...... time running %016" PRIx64 "\n",
771 sample->read.time_running);
773 if (read_format & PERF_FORMAT_GROUP) {
776 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
778 for (i = 0; i < sample->read.group.nr; i++) {
779 struct sample_read_value *value;
781 value = &sample->read.group.values[i];
782 printf("..... id %016" PRIx64
783 ", value %016" PRIx64 "\n",
784 value->id, value->value);
787 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
788 sample->read.one.id, sample->read.one.value);
791 static void dump_event(struct perf_session *session, union perf_event *event,
792 u64 file_offset, struct perf_sample *sample)
797 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
798 file_offset, event->header.size, event->header.type);
803 perf_session__print_tstamp(session, event, sample);
805 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
806 event->header.size, perf_event__name(event->header.type));
809 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
810 struct perf_sample *sample)
817 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
818 event->header.misc, sample->pid, sample->tid, sample->ip,
819 sample->period, sample->addr);
821 sample_type = evsel->attr.sample_type;
823 if (sample_type & PERF_SAMPLE_CALLCHAIN)
824 callchain__printf(sample);
826 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
827 branch_stack__printf(sample);
829 if (sample_type & PERF_SAMPLE_REGS_USER)
830 regs_user__printf(sample, evsel->attr.sample_regs_user);
832 if (sample_type & PERF_SAMPLE_STACK_USER)
833 stack_user__printf(&sample->user_stack);
835 if (sample_type & PERF_SAMPLE_WEIGHT)
836 printf("... weight: %" PRIu64 "\n", sample->weight);
838 if (sample_type & PERF_SAMPLE_DATA_SRC)
839 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
841 if (sample_type & PERF_SAMPLE_TRANSACTION)
842 printf("... transaction: %" PRIx64 "\n", sample->transaction);
844 if (sample_type & PERF_SAMPLE_READ)
845 sample_read__printf(sample, evsel->attr.read_format);
848 static struct machine *
849 perf_session__find_machine_for_cpumode(struct perf_session *session,
850 union perf_event *event,
851 struct perf_sample *sample)
853 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
856 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
857 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
860 if (event->header.type == PERF_RECORD_MMAP
861 || event->header.type == PERF_RECORD_MMAP2)
862 pid = event->mmap.pid;
866 return perf_session__findnew_machine(session, pid);
869 return &session->machines.host;
872 static int deliver_sample_value(struct perf_session *session,
873 struct perf_tool *tool,
874 union perf_event *event,
875 struct perf_sample *sample,
876 struct sample_read_value *v,
877 struct machine *machine)
879 struct perf_sample_id *sid;
881 sid = perf_evlist__id2sid(session->evlist, v->id);
884 sample->period = v->value - sid->period;
885 sid->period = v->value;
888 if (!sid || sid->evsel == NULL) {
889 ++session->stats.nr_unknown_id;
893 return tool->sample(tool, event, sample, sid->evsel, machine);
896 static int deliver_sample_group(struct perf_session *session,
897 struct perf_tool *tool,
898 union perf_event *event,
899 struct perf_sample *sample,
900 struct machine *machine)
905 for (i = 0; i < sample->read.group.nr; i++) {
906 ret = deliver_sample_value(session, tool, event, sample,
907 &sample->read.group.values[i],
917 perf_session__deliver_sample(struct perf_session *session,
918 struct perf_tool *tool,
919 union perf_event *event,
920 struct perf_sample *sample,
921 struct perf_evsel *evsel,
922 struct machine *machine)
924 /* We know evsel != NULL. */
925 u64 sample_type = evsel->attr.sample_type;
926 u64 read_format = evsel->attr.read_format;
928 /* Standard sample delievery. */
929 if (!(sample_type & PERF_SAMPLE_READ))
930 return tool->sample(tool, event, sample, evsel, machine);
932 /* For PERF_SAMPLE_READ we have either single or group mode. */
933 if (read_format & PERF_FORMAT_GROUP)
934 return deliver_sample_group(session, tool, event, sample,
937 return deliver_sample_value(session, tool, event, sample,
938 &sample->read.one, machine);
941 static int perf_session_deliver_event(struct perf_session *session,
942 union perf_event *event,
943 struct perf_sample *sample,
944 struct perf_tool *tool,
947 struct perf_evsel *evsel;
948 struct machine *machine;
950 dump_event(session, event, file_offset, sample);
952 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
953 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
955 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
956 * because the tools right now may apply filters, discarding
957 * some of the samples. For consistency, in the future we
958 * should have something like nr_filtered_samples and remove
959 * the sample->period from total_sample_period, etc, KISS for
962 * Also testing against NULL allows us to handle files without
963 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
964 * future probably it'll be a good idea to restrict event
965 * processing via perf_session to files with both set.
967 hists__inc_nr_events(&evsel->hists, event->header.type);
970 machine = perf_session__find_machine_for_cpumode(session, event,
973 switch (event->header.type) {
974 case PERF_RECORD_SAMPLE:
975 dump_sample(evsel, event, sample);
977 ++session->stats.nr_unknown_id;
980 if (machine == NULL) {
981 ++session->stats.nr_unprocessable_samples;
984 return perf_session__deliver_sample(session, tool, event,
985 sample, evsel, machine);
986 case PERF_RECORD_MMAP:
987 return tool->mmap(tool, event, sample, machine);
988 case PERF_RECORD_MMAP2:
989 return tool->mmap2(tool, event, sample, machine);
990 case PERF_RECORD_COMM:
991 return tool->comm(tool, event, sample, machine);
992 case PERF_RECORD_FORK:
993 return tool->fork(tool, event, sample, machine);
994 case PERF_RECORD_EXIT:
995 return tool->exit(tool, event, sample, machine);
996 case PERF_RECORD_LOST:
997 if (tool->lost == perf_event__process_lost)
998 session->stats.total_lost += event->lost.lost;
999 return tool->lost(tool, event, sample, machine);
1000 case PERF_RECORD_READ:
1001 return tool->read(tool, event, sample, evsel, machine);
1002 case PERF_RECORD_THROTTLE:
1003 return tool->throttle(tool, event, sample, machine);
1004 case PERF_RECORD_UNTHROTTLE:
1005 return tool->unthrottle(tool, event, sample, machine);
1007 ++session->stats.nr_unknown_events;
1012 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1013 struct perf_tool *tool, u64 file_offset)
1015 int fd = perf_data_file__fd(session->file);
1018 dump_event(session, event, file_offset, NULL);
1020 /* These events are processed right away */
1021 switch (event->header.type) {
1022 case PERF_RECORD_HEADER_ATTR:
1023 err = tool->attr(tool, event, &session->evlist);
1025 perf_session__set_id_hdr_size(session);
1027 case PERF_RECORD_HEADER_TRACING_DATA:
1028 /* setup for reading amidst mmap */
1029 lseek(fd, file_offset, SEEK_SET);
1030 return tool->tracing_data(tool, event, session);
1031 case PERF_RECORD_HEADER_BUILD_ID:
1032 return tool->build_id(tool, event, session);
1033 case PERF_RECORD_FINISHED_ROUND:
1034 return tool->finished_round(tool, event, session);
1040 static void event_swap(union perf_event *event, bool sample_id_all)
1042 perf_event__swap_op swap;
1044 swap = perf_event__swap_ops[event->header.type];
1046 swap(event, sample_id_all);
1049 static int perf_session__process_event(struct perf_session *session,
1050 union perf_event *event,
1051 struct perf_tool *tool,
1054 struct perf_sample sample;
1057 if (session->header.needs_swap)
1058 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1060 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1063 events_stats__inc(&session->stats, event->header.type);
1065 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1066 return perf_session__process_user_event(session, event, tool, file_offset);
1069 * For all kernel events we get the sample data
1071 ret = perf_evlist__parse_sample(session->evlist, 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 *hdr)
1088 hdr->type = bswap_32(hdr->type);
1089 hdr->misc = bswap_16(hdr->misc);
1090 hdr->size = bswap_16(hdr->size);
1093 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1095 return machine__findnew_thread(&session->machines.host, 0, pid);
1098 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1100 struct thread *thread = perf_session__findnew(session, 0);
1102 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
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->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->stats.nr_events[0],
1118 session->stats.nr_events[PERF_RECORD_LOST]);
1121 if (session->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->stats.nr_unknown_events);
1130 if (session->stats.nr_unknown_id != 0) {
1131 ui__warning("%u samples with id not present in the header\n",
1132 session->stats.nr_unknown_id);
1135 if (session->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->stats.nr_invalid_chains,
1140 session->stats.nr_events[PERF_RECORD_SAMPLE]);
1143 if (session->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->stats.nr_unprocessable_samples);
1150 volatile int session_done;
1152 static int __perf_session__process_pipe_events(struct perf_session *session,
1153 struct perf_tool *tool)
1155 int fd = perf_data_file__fd(session->file);
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(fd, event, sizeof(struct perf_event_header));
1179 pr_err("failed to read event header\n");
1183 if (session->header.needs_swap)
1184 perf_event_header__bswap(&event->header);
1186 size = event->header.size;
1187 if (size < sizeof(struct perf_event_header)) {
1188 pr_err("bad event header size\n");
1192 if (size > cur_size) {
1193 void *new = realloc(buf, size);
1195 pr_err("failed to allocate memory to read event\n");
1203 p += sizeof(struct perf_event_header);
1205 if (size - sizeof(struct perf_event_header)) {
1206 err = readn(fd, p, size - sizeof(struct perf_event_header));
1209 pr_err("unexpected end of event stream\n");
1213 pr_err("failed to read event data\n");
1218 if ((skip = perf_session__process_event(session, event, tool, head)) < 0) {
1219 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1220 head, event->header.size, event->header.type);
1230 if (!session_done())
1233 /* do the final flush for ordered samples */
1234 session->ordered_samples.next_flush = ULLONG_MAX;
1235 err = flush_sample_queue(session, tool);
1238 perf_session__warn_about_errors(session, tool);
1239 perf_session_free_sample_buffers(session);
1243 static union perf_event *
1244 fetch_mmaped_event(struct perf_session *session,
1245 u64 head, size_t mmap_size, char *buf)
1247 union perf_event *event;
1250 * Ensure we have enough space remaining to read
1251 * the size of the event in the headers.
1253 if (head + sizeof(event->header) > mmap_size)
1256 event = (union perf_event *)(buf + head);
1258 if (session->header.needs_swap)
1259 perf_event_header__bswap(&event->header);
1261 if (head + event->header.size > mmap_size) {
1262 /* We're not fetching the event so swap back again */
1263 if (session->header.needs_swap)
1264 perf_event_header__bswap(&event->header);
1272 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1273 * slices. On 32bit we use 32MB.
1275 #if BITS_PER_LONG == 64
1276 #define MMAP_SIZE ULLONG_MAX
1279 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1280 #define NUM_MMAPS 128
1283 int __perf_session__process_events(struct perf_session *session,
1284 u64 data_offset, u64 data_size,
1285 u64 file_size, struct perf_tool *tool)
1287 int fd = perf_data_file__fd(session->file);
1288 u64 head, page_offset, file_offset, file_pos;
1289 int err, mmap_prot, mmap_flags, map_idx = 0;
1291 char *buf, *mmaps[NUM_MMAPS];
1292 union perf_event *event;
1294 struct ui_progress prog;
1296 perf_tool__fill_defaults(tool);
1298 page_offset = page_size * (data_offset / page_size);
1299 file_offset = page_offset;
1300 head = data_offset - page_offset;
1302 if (data_size && (data_offset + data_size < file_size))
1303 file_size = data_offset + data_size;
1305 ui_progress__init(&prog, file_size, "Processing events...");
1307 mmap_size = MMAP_SIZE;
1308 if (mmap_size > file_size)
1309 mmap_size = file_size;
1311 memset(mmaps, 0, sizeof(mmaps));
1313 mmap_prot = PROT_READ;
1314 mmap_flags = MAP_SHARED;
1316 if (session->header.needs_swap) {
1317 mmap_prot |= PROT_WRITE;
1318 mmap_flags = MAP_PRIVATE;
1321 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1323 if (buf == MAP_FAILED) {
1324 pr_err("failed to mmap file\n");
1328 mmaps[map_idx] = buf;
1329 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1330 file_pos = file_offset + head;
1333 event = fetch_mmaped_event(session, head, mmap_size, buf);
1335 if (mmaps[map_idx]) {
1336 munmap(mmaps[map_idx], mmap_size);
1337 mmaps[map_idx] = NULL;
1340 page_offset = page_size * (head / page_size);
1341 file_offset += page_offset;
1342 head -= page_offset;
1346 size = event->header.size;
1348 if (size < sizeof(struct perf_event_header) ||
1349 perf_session__process_event(session, event, tool, file_pos) < 0) {
1350 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1351 file_offset + head, event->header.size,
1352 event->header.type);
1360 ui_progress__update(&prog, size);
1365 if (file_pos < file_size)
1369 /* do the final flush for ordered samples */
1370 session->ordered_samples.next_flush = ULLONG_MAX;
1371 err = flush_sample_queue(session, tool);
1373 ui_progress__finish();
1374 perf_session__warn_about_errors(session, tool);
1375 perf_session_free_sample_buffers(session);
1379 int perf_session__process_events(struct perf_session *session,
1380 struct perf_tool *tool)
1382 u64 size = perf_data_file__size(session->file);
1385 if (perf_session__register_idle_thread(session) == NULL)
1388 if (!perf_data_file__is_pipe(session->file))
1389 err = __perf_session__process_events(session,
1390 session->header.data_offset,
1391 session->header.data_size,
1394 err = __perf_session__process_pipe_events(session, tool);
1399 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1401 struct perf_evsel *evsel;
1403 list_for_each_entry(evsel, &session->evlist->entries, node) {
1404 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1408 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1412 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1413 const char *symbol_name, u64 addr)
1417 struct ref_reloc_sym *ref;
1419 ref = zalloc(sizeof(struct ref_reloc_sym));
1423 ref->name = strdup(symbol_name);
1424 if (ref->name == NULL) {
1429 bracket = strchr(ref->name, ']');
1435 for (i = 0; i < MAP__NR_TYPES; ++i) {
1436 struct kmap *kmap = map__kmap(maps[i]);
1437 kmap->ref_reloc_sym = ref;
1443 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1445 return machines__fprintf_dsos(&session->machines, fp);
1448 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1449 bool (skip)(struct dso *dso, int parm), int parm)
1451 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1454 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1456 struct perf_evsel *pos;
1457 size_t ret = fprintf(fp, "Aggregated stats:\n");
1459 ret += events_stats__fprintf(&session->stats, fp);
1461 list_for_each_entry(pos, &session->evlist->entries, node) {
1462 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1463 ret += events_stats__fprintf(&pos->hists.stats, fp);
1469 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1472 * FIXME: Here we have to actually print all the machines in this
1473 * session, not just the host...
1475 return machine__fprintf(&session->machines.host, fp);
1478 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1481 struct perf_evsel *pos;
1483 list_for_each_entry(pos, &session->evlist->entries, node) {
1484 if (pos->attr.type == type)
1490 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1491 struct perf_sample *sample, struct machine *machine,
1492 unsigned int print_opts, unsigned int stack_depth)
1494 struct addr_location al;
1495 struct callchain_cursor_node *node;
1496 int print_ip = print_opts & PRINT_IP_OPT_IP;
1497 int print_sym = print_opts & PRINT_IP_OPT_SYM;
1498 int print_dso = print_opts & PRINT_IP_OPT_DSO;
1499 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1500 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1501 char s = print_oneline ? ' ' : '\t';
1503 if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
1504 error("problem processing %d event, skipping it.\n",
1505 event->header.type);
1509 if (symbol_conf.use_callchain && sample->callchain) {
1511 if (machine__resolve_callchain(machine, evsel, al.thread,
1513 PERF_MAX_STACK_DEPTH) != 0) {
1515 error("Failed to resolve callchain. Skipping\n");
1518 callchain_cursor_commit(&callchain_cursor);
1520 while (stack_depth) {
1521 node = callchain_cursor_current(&callchain_cursor);
1526 printf("%c%16" PRIx64, s, node->ip);
1530 if (print_symoffset) {
1533 symbol__fprintf_symname_offs(node->sym, &al, stdout);
1535 symbol__fprintf_symname(node->sym, stdout);
1540 map__fprintf_dsoname(node->map, stdout);
1547 callchain_cursor_advance(&callchain_cursor);
1554 printf("%16" PRIx64, sample->ip);
1558 if (print_symoffset)
1559 symbol__fprintf_symname_offs(al.sym, &al,
1562 symbol__fprintf_symname(al.sym, stdout);
1567 map__fprintf_dsoname(al.map, stdout);
1573 int perf_session__cpu_bitmap(struct perf_session *session,
1574 const char *cpu_list, unsigned long *cpu_bitmap)
1577 struct cpu_map *map;
1579 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1580 struct perf_evsel *evsel;
1582 evsel = perf_session__find_first_evtype(session, i);
1586 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1587 pr_err("File does not contain CPU events. "
1588 "Remove -c option to proceed.\n");
1593 map = cpu_map__new(cpu_list);
1595 pr_err("Invalid cpu_list\n");
1599 for (i = 0; i < map->nr; i++) {
1600 int cpu = map->map[i];
1602 if (cpu >= MAX_NR_CPUS) {
1603 pr_err("Requested CPU %d too large. "
1604 "Consider raising MAX_NR_CPUS\n", cpu);
1608 set_bit(cpu, cpu_bitmap);
1614 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1617 int fd = perf_data_file__fd(session->file);
1621 if (session == NULL || fp == NULL)
1624 ret = fstat(fd, &st);
1628 fprintf(fp, "# ========\n");
1629 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1630 perf_header__fprintf_info(session, fp, full);
1631 fprintf(fp, "# ========\n#\n");
1635 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1636 const struct perf_evsel_str_handler *assocs,
1639 struct perf_evsel *evsel;
1643 for (i = 0; i < nr_assocs; i++) {
1645 * Adding a handler for an event not in the session,
1648 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1653 if (evsel->handler != NULL)
1655 evsel->handler = assocs[i].handler;