1 #include <linux/kernel.h>
15 #include "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, self->fd) < 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, self->fd) < 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 self->size = input_stat.st_size;
83 void perf_session__set_id_hdr_size(struct perf_session *session)
85 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
87 machines__set_id_hdr_size(&session->machines, id_hdr_size);
90 int perf_session__create_kernel_maps(struct perf_session *self)
92 int ret = machine__create_kernel_maps(&self->machines.host);
95 ret = machines__create_guest_kernel_maps(&self->machines);
99 static void perf_session__destroy_kernel_maps(struct perf_session *self)
101 machines__destroy_kernel_maps(&self->machines);
104 struct perf_session *perf_session__new(const char *filename, int mode,
105 bool force, bool repipe,
106 struct perf_tool *tool)
108 struct perf_session *self;
112 if (!filename || !strlen(filename)) {
113 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
116 filename = "perf.data";
119 len = strlen(filename);
120 self = zalloc(sizeof(*self) + len);
125 memcpy(self->filename, filename, len);
126 self->repipe = repipe;
127 INIT_LIST_HEAD(&self->ordered_samples.samples);
128 INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
129 INIT_LIST_HEAD(&self->ordered_samples.to_free);
130 machines__init(&self->machines);
132 if (mode == O_RDONLY) {
133 if (perf_session__open(self, force) < 0)
135 perf_session__set_id_hdr_size(self);
136 } else if (mode == O_WRONLY) {
138 * In O_RDONLY mode this will be performed when reading the
139 * kernel MMAP event, in perf_event__process_mmap().
141 if (perf_session__create_kernel_maps(self) < 0)
145 if (tool && tool->ordering_requires_timestamps &&
146 tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
147 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
148 tool->ordered_samples = false;
154 perf_session__delete(self);
158 static void perf_session__delete_dead_threads(struct perf_session *session)
160 machine__delete_dead_threads(&session->machines.host);
163 static void perf_session__delete_threads(struct perf_session *session)
165 machine__delete_threads(&session->machines.host);
168 static void perf_session_env__delete(struct perf_session_env *env)
171 free(env->os_release);
178 free(env->sibling_cores);
179 free(env->sibling_threads);
180 free(env->numa_nodes);
181 free(env->pmu_mappings);
184 void perf_session__delete(struct perf_session *self)
186 perf_session__destroy_kernel_maps(self);
187 perf_session__delete_dead_threads(self);
188 perf_session__delete_threads(self);
189 perf_session_env__delete(&self->header.env);
190 machines__exit(&self->machines);
196 static int process_event_synth_tracing_data_stub(union perf_event *event
198 struct perf_session *session
201 dump_printf(": unhandled!\n");
205 static int process_event_synth_attr_stub(union perf_event *event __maybe_unused,
206 struct perf_evlist **pevlist
209 dump_printf(": unhandled!\n");
213 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
214 union perf_event *event __maybe_unused,
215 struct perf_sample *sample __maybe_unused,
216 struct perf_evsel *evsel __maybe_unused,
217 struct machine *machine __maybe_unused)
219 dump_printf(": unhandled!\n");
223 static int process_event_stub(struct perf_tool *tool __maybe_unused,
224 union perf_event *event __maybe_unused,
225 struct perf_sample *sample __maybe_unused,
226 struct machine *machine __maybe_unused)
228 dump_printf(": unhandled!\n");
232 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
233 union perf_event *event __maybe_unused,
234 struct perf_session *perf_session
237 dump_printf(": unhandled!\n");
241 static int process_event_type_stub(struct perf_tool *tool __maybe_unused,
242 union perf_event *event __maybe_unused)
244 dump_printf(": unhandled!\n");
248 static int process_finished_round(struct perf_tool *tool,
249 union perf_event *event,
250 struct perf_session *session);
252 static void perf_tool__fill_defaults(struct perf_tool *tool)
254 if (tool->sample == NULL)
255 tool->sample = process_event_sample_stub;
256 if (tool->mmap == NULL)
257 tool->mmap = process_event_stub;
258 if (tool->comm == NULL)
259 tool->comm = process_event_stub;
260 if (tool->fork == NULL)
261 tool->fork = process_event_stub;
262 if (tool->exit == NULL)
263 tool->exit = process_event_stub;
264 if (tool->lost == NULL)
265 tool->lost = perf_event__process_lost;
266 if (tool->read == NULL)
267 tool->read = process_event_sample_stub;
268 if (tool->throttle == NULL)
269 tool->throttle = process_event_stub;
270 if (tool->unthrottle == NULL)
271 tool->unthrottle = process_event_stub;
272 if (tool->attr == NULL)
273 tool->attr = process_event_synth_attr_stub;
274 if (tool->event_type == NULL)
275 tool->event_type = process_event_type_stub;
276 if (tool->tracing_data == NULL)
277 tool->tracing_data = process_event_synth_tracing_data_stub;
278 if (tool->build_id == NULL)
279 tool->build_id = process_finished_round_stub;
280 if (tool->finished_round == NULL) {
281 if (tool->ordered_samples)
282 tool->finished_round = process_finished_round;
284 tool->finished_round = process_finished_round_stub;
288 void mem_bswap_32(void *src, int byte_size)
291 while (byte_size > 0) {
293 byte_size -= sizeof(u32);
298 void mem_bswap_64(void *src, int byte_size)
302 while (byte_size > 0) {
304 byte_size -= sizeof(u64);
309 static void swap_sample_id_all(union perf_event *event, void *data)
311 void *end = (void *) event + event->header.size;
312 int size = end - data;
314 BUG_ON(size % sizeof(u64));
315 mem_bswap_64(data, size);
318 static void perf_event__all64_swap(union perf_event *event,
319 bool sample_id_all __maybe_unused)
321 struct perf_event_header *hdr = &event->header;
322 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
325 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
327 event->comm.pid = bswap_32(event->comm.pid);
328 event->comm.tid = bswap_32(event->comm.tid);
331 void *data = &event->comm.comm;
333 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
334 swap_sample_id_all(event, data);
338 static void perf_event__mmap_swap(union perf_event *event,
341 event->mmap.pid = bswap_32(event->mmap.pid);
342 event->mmap.tid = bswap_32(event->mmap.tid);
343 event->mmap.start = bswap_64(event->mmap.start);
344 event->mmap.len = bswap_64(event->mmap.len);
345 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
348 void *data = &event->mmap.filename;
350 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
351 swap_sample_id_all(event, data);
355 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
357 event->fork.pid = bswap_32(event->fork.pid);
358 event->fork.tid = bswap_32(event->fork.tid);
359 event->fork.ppid = bswap_32(event->fork.ppid);
360 event->fork.ptid = bswap_32(event->fork.ptid);
361 event->fork.time = bswap_64(event->fork.time);
364 swap_sample_id_all(event, &event->fork + 1);
367 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
369 event->read.pid = bswap_32(event->read.pid);
370 event->read.tid = bswap_32(event->read.tid);
371 event->read.value = bswap_64(event->read.value);
372 event->read.time_enabled = bswap_64(event->read.time_enabled);
373 event->read.time_running = bswap_64(event->read.time_running);
374 event->read.id = bswap_64(event->read.id);
377 swap_sample_id_all(event, &event->read + 1);
380 static u8 revbyte(u8 b)
382 int rev = (b >> 4) | ((b & 0xf) << 4);
383 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
384 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
389 * XXX this is hack in attempt to carry flags bitfield
390 * throught endian village. ABI says:
392 * Bit-fields are allocated from right to left (least to most significant)
393 * on little-endian implementations and from left to right (most to least
394 * significant) on big-endian implementations.
396 * The above seems to be byte specific, so we need to reverse each
397 * byte of the bitfield. 'Internet' also says this might be implementation
398 * specific and we probably need proper fix and carry perf_event_attr
399 * bitfield flags in separate data file FEAT_ section. Thought this seems
402 static void swap_bitfield(u8 *p, unsigned len)
406 for (i = 0; i < len; i++) {
412 /* exported for swapping attributes in file header */
413 void perf_event__attr_swap(struct perf_event_attr *attr)
415 attr->type = bswap_32(attr->type);
416 attr->size = bswap_32(attr->size);
417 attr->config = bswap_64(attr->config);
418 attr->sample_period = bswap_64(attr->sample_period);
419 attr->sample_type = bswap_64(attr->sample_type);
420 attr->read_format = bswap_64(attr->read_format);
421 attr->wakeup_events = bswap_32(attr->wakeup_events);
422 attr->bp_type = bswap_32(attr->bp_type);
423 attr->bp_addr = bswap_64(attr->bp_addr);
424 attr->bp_len = bswap_64(attr->bp_len);
426 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
429 static void perf_event__hdr_attr_swap(union perf_event *event,
430 bool sample_id_all __maybe_unused)
434 perf_event__attr_swap(&event->attr.attr);
436 size = event->header.size;
437 size -= (void *)&event->attr.id - (void *)event;
438 mem_bswap_64(event->attr.id, size);
441 static void perf_event__event_type_swap(union perf_event *event,
442 bool sample_id_all __maybe_unused)
444 event->event_type.event_type.event_id =
445 bswap_64(event->event_type.event_type.event_id);
448 static void perf_event__tracing_data_swap(union perf_event *event,
449 bool sample_id_all __maybe_unused)
451 event->tracing_data.size = bswap_32(event->tracing_data.size);
454 typedef void (*perf_event__swap_op)(union perf_event *event,
457 static perf_event__swap_op perf_event__swap_ops[] = {
458 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
459 [PERF_RECORD_COMM] = perf_event__comm_swap,
460 [PERF_RECORD_FORK] = perf_event__task_swap,
461 [PERF_RECORD_EXIT] = perf_event__task_swap,
462 [PERF_RECORD_LOST] = perf_event__all64_swap,
463 [PERF_RECORD_READ] = perf_event__read_swap,
464 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
465 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
466 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
467 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
468 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
469 [PERF_RECORD_HEADER_MAX] = NULL,
472 struct sample_queue {
475 union perf_event *event;
476 struct list_head list;
479 static void perf_session_free_sample_buffers(struct perf_session *session)
481 struct ordered_samples *os = &session->ordered_samples;
483 while (!list_empty(&os->to_free)) {
484 struct sample_queue *sq;
486 sq = list_entry(os->to_free.next, struct sample_queue, list);
492 static int perf_session_deliver_event(struct perf_session *session,
493 union perf_event *event,
494 struct perf_sample *sample,
495 struct perf_tool *tool,
498 static int flush_sample_queue(struct perf_session *s,
499 struct perf_tool *tool)
501 struct ordered_samples *os = &s->ordered_samples;
502 struct list_head *head = &os->samples;
503 struct sample_queue *tmp, *iter;
504 struct perf_sample sample;
505 u64 limit = os->next_flush;
506 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
507 unsigned idx = 0, progress_next = os->nr_samples / 16;
510 if (!tool->ordered_samples || !limit)
513 list_for_each_entry_safe(iter, tmp, head, list) {
514 if (iter->timestamp > limit)
517 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
519 pr_err("Can't parse sample, err = %d\n", ret);
521 ret = perf_session_deliver_event(s, iter->event, &sample, tool,
527 os->last_flush = iter->timestamp;
528 list_del(&iter->list);
529 list_add(&iter->list, &os->sample_cache);
530 if (++idx >= progress_next) {
531 progress_next += os->nr_samples / 16;
532 ui_progress__update(idx, os->nr_samples,
533 "Processing time ordered events...");
537 if (list_empty(head)) {
538 os->last_sample = NULL;
539 } else if (last_ts <= limit) {
541 list_entry(head->prev, struct sample_queue, list);
550 * When perf record finishes a pass on every buffers, it records this pseudo
552 * We record the max timestamp t found in the pass n.
553 * Assuming these timestamps are monotonic across cpus, we know that if
554 * a buffer still has events with timestamps below t, they will be all
555 * available and then read in the pass n + 1.
556 * Hence when we start to read the pass n + 2, we can safely flush every
557 * events with timestamps below t.
559 * ============ PASS n =================
562 * cnt1 timestamps | cnt2 timestamps
565 * - | 4 <--- max recorded
567 * ============ PASS n + 1 ==============
570 * cnt1 timestamps | cnt2 timestamps
573 * 5 | 7 <---- max recorded
575 * Flush every events below timestamp 4
577 * ============ PASS n + 2 ==============
580 * cnt1 timestamps | cnt2 timestamps
585 * Flush every events below timestamp 7
588 static int process_finished_round(struct perf_tool *tool,
589 union perf_event *event __maybe_unused,
590 struct perf_session *session)
592 int ret = flush_sample_queue(session, tool);
594 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
599 /* The queue is ordered by time */
600 static void __queue_event(struct sample_queue *new, struct perf_session *s)
602 struct ordered_samples *os = &s->ordered_samples;
603 struct sample_queue *sample = os->last_sample;
604 u64 timestamp = new->timestamp;
608 os->last_sample = new;
611 list_add(&new->list, &os->samples);
612 os->max_timestamp = timestamp;
617 * last_sample might point to some random place in the list as it's
618 * the last queued event. We expect that the new event is close to
621 if (sample->timestamp <= timestamp) {
622 while (sample->timestamp <= timestamp) {
623 p = sample->list.next;
624 if (p == &os->samples) {
625 list_add_tail(&new->list, &os->samples);
626 os->max_timestamp = timestamp;
629 sample = list_entry(p, struct sample_queue, list);
631 list_add_tail(&new->list, &sample->list);
633 while (sample->timestamp > timestamp) {
634 p = sample->list.prev;
635 if (p == &os->samples) {
636 list_add(&new->list, &os->samples);
639 sample = list_entry(p, struct sample_queue, list);
641 list_add(&new->list, &sample->list);
645 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
647 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
648 struct perf_sample *sample, u64 file_offset)
650 struct ordered_samples *os = &s->ordered_samples;
651 struct list_head *sc = &os->sample_cache;
652 u64 timestamp = sample->time;
653 struct sample_queue *new;
655 if (!timestamp || timestamp == ~0ULL)
658 if (timestamp < s->ordered_samples.last_flush) {
659 printf("Warning: Timestamp below last timeslice flush\n");
663 if (!list_empty(sc)) {
664 new = list_entry(sc->next, struct sample_queue, list);
665 list_del(&new->list);
666 } else if (os->sample_buffer) {
667 new = os->sample_buffer + os->sample_buffer_idx;
668 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
669 os->sample_buffer = NULL;
671 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
672 if (!os->sample_buffer)
674 list_add(&os->sample_buffer->list, &os->to_free);
675 os->sample_buffer_idx = 2;
676 new = os->sample_buffer + 1;
679 new->timestamp = timestamp;
680 new->file_offset = file_offset;
683 __queue_event(new, s);
688 static void callchain__printf(struct perf_sample *sample)
692 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
694 for (i = 0; i < sample->callchain->nr; i++)
695 printf("..... %2d: %016" PRIx64 "\n",
696 i, sample->callchain->ips[i]);
699 static void branch_stack__printf(struct perf_sample *sample)
703 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
705 for (i = 0; i < sample->branch_stack->nr; i++)
706 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
707 i, sample->branch_stack->entries[i].from,
708 sample->branch_stack->entries[i].to);
711 static void regs_dump__printf(u64 mask, u64 *regs)
715 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
718 printf(".... %-5s 0x%" PRIx64 "\n",
719 perf_reg_name(rid), val);
723 static void regs_user__printf(struct perf_sample *sample, u64 mask)
725 struct regs_dump *user_regs = &sample->user_regs;
727 if (user_regs->regs) {
728 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
729 regs_dump__printf(mask, user_regs->regs);
733 static void stack_user__printf(struct stack_dump *dump)
735 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
736 dump->size, dump->offset);
739 static void perf_session__print_tstamp(struct perf_session *session,
740 union perf_event *event,
741 struct perf_sample *sample)
743 u64 sample_type = perf_evlist__sample_type(session->evlist);
745 if (event->header.type != PERF_RECORD_SAMPLE &&
746 !perf_evlist__sample_id_all(session->evlist)) {
747 fputs("-1 -1 ", stdout);
751 if ((sample_type & PERF_SAMPLE_CPU))
752 printf("%u ", sample->cpu);
754 if (sample_type & PERF_SAMPLE_TIME)
755 printf("%" PRIu64 " ", sample->time);
758 static void dump_event(struct perf_session *session, union perf_event *event,
759 u64 file_offset, struct perf_sample *sample)
764 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
765 file_offset, event->header.size, event->header.type);
770 perf_session__print_tstamp(session, event, sample);
772 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
773 event->header.size, perf_event__name(event->header.type));
776 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
777 struct perf_sample *sample)
784 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
785 event->header.misc, sample->pid, sample->tid, sample->ip,
786 sample->period, sample->addr);
788 sample_type = evsel->attr.sample_type;
790 if (sample_type & PERF_SAMPLE_CALLCHAIN)
791 callchain__printf(sample);
793 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
794 branch_stack__printf(sample);
796 if (sample_type & PERF_SAMPLE_REGS_USER)
797 regs_user__printf(sample, evsel->attr.sample_regs_user);
799 if (sample_type & PERF_SAMPLE_STACK_USER)
800 stack_user__printf(&sample->user_stack);
802 if (sample_type & PERF_SAMPLE_WEIGHT)
803 printf("... weight: %" PRIu64 "\n", sample->weight);
805 if (sample_type & PERF_SAMPLE_DATA_SRC)
806 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
809 static struct machine *
810 perf_session__find_machine_for_cpumode(struct perf_session *session,
811 union perf_event *event)
813 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
814 struct machine *machine;
817 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
818 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
821 if (event->header.type == PERF_RECORD_MMAP)
822 pid = event->mmap.pid;
826 machine = perf_session__find_machine(session, pid);
828 machine = perf_session__findnew_machine(session,
829 DEFAULT_GUEST_KERNEL_ID);
833 return &session->machines.host;
836 static int perf_session_deliver_event(struct perf_session *session,
837 union perf_event *event,
838 struct perf_sample *sample,
839 struct perf_tool *tool,
842 struct perf_evsel *evsel;
843 struct machine *machine;
845 dump_event(session, event, file_offset, sample);
847 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
848 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
850 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
851 * because the tools right now may apply filters, discarding
852 * some of the samples. For consistency, in the future we
853 * should have something like nr_filtered_samples and remove
854 * the sample->period from total_sample_period, etc, KISS for
857 * Also testing against NULL allows us to handle files without
858 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
859 * future probably it'll be a good idea to restrict event
860 * processing via perf_session to files with both set.
862 hists__inc_nr_events(&evsel->hists, event->header.type);
865 machine = perf_session__find_machine_for_cpumode(session, event);
867 switch (event->header.type) {
868 case PERF_RECORD_SAMPLE:
869 dump_sample(evsel, event, sample);
871 ++session->stats.nr_unknown_id;
874 if (machine == NULL) {
875 ++session->stats.nr_unprocessable_samples;
878 return tool->sample(tool, event, sample, evsel, machine);
879 case PERF_RECORD_MMAP:
880 return tool->mmap(tool, event, sample, machine);
881 case PERF_RECORD_COMM:
882 return tool->comm(tool, event, sample, machine);
883 case PERF_RECORD_FORK:
884 return tool->fork(tool, event, sample, machine);
885 case PERF_RECORD_EXIT:
886 return tool->exit(tool, event, sample, machine);
887 case PERF_RECORD_LOST:
888 if (tool->lost == perf_event__process_lost)
889 session->stats.total_lost += event->lost.lost;
890 return tool->lost(tool, event, sample, machine);
891 case PERF_RECORD_READ:
892 return tool->read(tool, event, sample, evsel, machine);
893 case PERF_RECORD_THROTTLE:
894 return tool->throttle(tool, event, sample, machine);
895 case PERF_RECORD_UNTHROTTLE:
896 return tool->unthrottle(tool, event, sample, machine);
898 ++session->stats.nr_unknown_events;
903 static int perf_session__preprocess_sample(struct perf_session *session,
904 union perf_event *event, struct perf_sample *sample)
906 if (event->header.type != PERF_RECORD_SAMPLE ||
907 !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN))
910 if (!ip_callchain__valid(sample->callchain, event)) {
911 pr_debug("call-chain problem with event, skipping it.\n");
912 ++session->stats.nr_invalid_chains;
913 session->stats.total_invalid_chains += sample->period;
919 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
920 struct perf_tool *tool, u64 file_offset)
924 dump_event(session, event, file_offset, NULL);
926 /* These events are processed right away */
927 switch (event->header.type) {
928 case PERF_RECORD_HEADER_ATTR:
929 err = tool->attr(event, &session->evlist);
931 perf_session__set_id_hdr_size(session);
933 case PERF_RECORD_HEADER_EVENT_TYPE:
934 return tool->event_type(tool, event);
935 case PERF_RECORD_HEADER_TRACING_DATA:
936 /* setup for reading amidst mmap */
937 lseek(session->fd, file_offset, SEEK_SET);
938 return tool->tracing_data(event, session);
939 case PERF_RECORD_HEADER_BUILD_ID:
940 return tool->build_id(tool, event, session);
941 case PERF_RECORD_FINISHED_ROUND:
942 return tool->finished_round(tool, event, session);
948 static void event_swap(union perf_event *event, bool sample_id_all)
950 perf_event__swap_op swap;
952 swap = perf_event__swap_ops[event->header.type];
954 swap(event, sample_id_all);
957 static int perf_session__process_event(struct perf_session *session,
958 union perf_event *event,
959 struct perf_tool *tool,
962 struct perf_sample sample;
965 if (session->header.needs_swap)
966 event_swap(event, perf_evlist__sample_id_all(session->evlist));
968 if (event->header.type >= PERF_RECORD_HEADER_MAX)
971 events_stats__inc(&session->stats, event->header.type);
973 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
974 return perf_session__process_user_event(session, event, tool, file_offset);
977 * For all kernel events we get the sample data
979 ret = perf_evlist__parse_sample(session->evlist, event, &sample);
983 /* Preprocess sample records - precheck callchains */
984 if (perf_session__preprocess_sample(session, event, &sample))
987 if (tool->ordered_samples) {
988 ret = perf_session_queue_event(session, event, &sample,
994 return perf_session_deliver_event(session, event, &sample, tool,
998 void perf_event_header__bswap(struct perf_event_header *self)
1000 self->type = bswap_32(self->type);
1001 self->misc = bswap_16(self->misc);
1002 self->size = bswap_16(self->size);
1005 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1007 return machine__findnew_thread(&session->machines.host, pid);
1010 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1012 struct thread *thread = perf_session__findnew(self, 0);
1014 if (thread == NULL || thread__set_comm(thread, "swapper")) {
1015 pr_err("problem inserting idle task.\n");
1022 static void perf_session__warn_about_errors(const struct perf_session *session,
1023 const struct perf_tool *tool)
1025 if (tool->lost == perf_event__process_lost &&
1026 session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1027 ui__warning("Processed %d events and lost %d chunks!\n\n"
1028 "Check IO/CPU overload!\n\n",
1029 session->stats.nr_events[0],
1030 session->stats.nr_events[PERF_RECORD_LOST]);
1033 if (session->stats.nr_unknown_events != 0) {
1034 ui__warning("Found %u unknown events!\n\n"
1035 "Is this an older tool processing a perf.data "
1036 "file generated by a more recent tool?\n\n"
1037 "If that is not the case, consider "
1038 "reporting to linux-kernel@vger.kernel.org.\n\n",
1039 session->stats.nr_unknown_events);
1042 if (session->stats.nr_unknown_id != 0) {
1043 ui__warning("%u samples with id not present in the header\n",
1044 session->stats.nr_unknown_id);
1047 if (session->stats.nr_invalid_chains != 0) {
1048 ui__warning("Found invalid callchains!\n\n"
1049 "%u out of %u events were discarded for this reason.\n\n"
1050 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1051 session->stats.nr_invalid_chains,
1052 session->stats.nr_events[PERF_RECORD_SAMPLE]);
1055 if (session->stats.nr_unprocessable_samples != 0) {
1056 ui__warning("%u unprocessable samples recorded.\n"
1057 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1058 session->stats.nr_unprocessable_samples);
1062 #define session_done() (*(volatile int *)(&session_done))
1063 volatile int session_done;
1065 static int __perf_session__process_pipe_events(struct perf_session *self,
1066 struct perf_tool *tool)
1068 union perf_event *event;
1069 uint32_t size, cur_size = 0;
1076 perf_tool__fill_defaults(tool);
1079 cur_size = sizeof(union perf_event);
1081 buf = malloc(cur_size);
1086 err = readn(self->fd, event, sizeof(struct perf_event_header));
1091 pr_err("failed to read event header\n");
1095 if (self->header.needs_swap)
1096 perf_event_header__bswap(&event->header);
1098 size = event->header.size;
1102 if (size > cur_size) {
1103 void *new = realloc(buf, size);
1105 pr_err("failed to allocate memory to read event\n");
1113 p += sizeof(struct perf_event_header);
1115 if (size - sizeof(struct perf_event_header)) {
1116 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1119 pr_err("unexpected end of event stream\n");
1123 pr_err("failed to read event data\n");
1128 if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1129 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1130 head, event->header.size, event->header.type);
1140 if (!session_done())
1146 perf_session__warn_about_errors(self, tool);
1147 perf_session_free_sample_buffers(self);
1151 static union perf_event *
1152 fetch_mmaped_event(struct perf_session *session,
1153 u64 head, size_t mmap_size, char *buf)
1155 union perf_event *event;
1158 * Ensure we have enough space remaining to read
1159 * the size of the event in the headers.
1161 if (head + sizeof(event->header) > mmap_size)
1164 event = (union perf_event *)(buf + head);
1166 if (session->header.needs_swap)
1167 perf_event_header__bswap(&event->header);
1169 if (head + event->header.size > mmap_size)
1176 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1177 * slices. On 32bit we use 32MB.
1179 #if BITS_PER_LONG == 64
1180 #define MMAP_SIZE ULLONG_MAX
1183 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1184 #define NUM_MMAPS 128
1187 int __perf_session__process_events(struct perf_session *session,
1188 u64 data_offset, u64 data_size,
1189 u64 file_size, struct perf_tool *tool)
1191 u64 head, page_offset, file_offset, file_pos, progress_next;
1192 int err, mmap_prot, mmap_flags, map_idx = 0;
1194 char *buf, *mmaps[NUM_MMAPS];
1195 union perf_event *event;
1198 perf_tool__fill_defaults(tool);
1200 page_offset = page_size * (data_offset / page_size);
1201 file_offset = page_offset;
1202 head = data_offset - page_offset;
1204 if (data_offset + data_size < file_size)
1205 file_size = data_offset + data_size;
1207 progress_next = file_size / 16;
1209 mmap_size = MMAP_SIZE;
1210 if (mmap_size > file_size)
1211 mmap_size = file_size;
1213 memset(mmaps, 0, sizeof(mmaps));
1215 mmap_prot = PROT_READ;
1216 mmap_flags = MAP_SHARED;
1218 if (session->header.needs_swap) {
1219 mmap_prot |= PROT_WRITE;
1220 mmap_flags = MAP_PRIVATE;
1223 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1225 if (buf == MAP_FAILED) {
1226 pr_err("failed to mmap file\n");
1230 mmaps[map_idx] = buf;
1231 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1232 file_pos = file_offset + head;
1235 event = fetch_mmaped_event(session, head, mmap_size, buf);
1237 if (mmaps[map_idx]) {
1238 munmap(mmaps[map_idx], mmap_size);
1239 mmaps[map_idx] = NULL;
1242 page_offset = page_size * (head / page_size);
1243 file_offset += page_offset;
1244 head -= page_offset;
1248 size = event->header.size;
1251 perf_session__process_event(session, event, tool, file_pos) < 0) {
1252 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1253 file_offset + head, event->header.size,
1254 event->header.type);
1262 if (file_pos >= progress_next) {
1263 progress_next += file_size / 16;
1264 ui_progress__update(file_pos, file_size,
1265 "Processing events...");
1268 if (file_pos < file_size)
1272 /* do the final flush for ordered samples */
1273 session->ordered_samples.next_flush = ULLONG_MAX;
1274 err = flush_sample_queue(session, tool);
1276 ui_progress__finish();
1277 perf_session__warn_about_errors(session, tool);
1278 perf_session_free_sample_buffers(session);
1282 int perf_session__process_events(struct perf_session *self,
1283 struct perf_tool *tool)
1287 if (perf_session__register_idle_thread(self) == NULL)
1291 err = __perf_session__process_events(self,
1292 self->header.data_offset,
1293 self->header.data_size,
1296 err = __perf_session__process_pipe_events(self, tool);
1301 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1303 if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) {
1304 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1311 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1312 const char *symbol_name, u64 addr)
1316 struct ref_reloc_sym *ref;
1318 ref = zalloc(sizeof(struct ref_reloc_sym));
1322 ref->name = strdup(symbol_name);
1323 if (ref->name == NULL) {
1328 bracket = strchr(ref->name, ']');
1334 for (i = 0; i < MAP__NR_TYPES; ++i) {
1335 struct kmap *kmap = map__kmap(maps[i]);
1336 kmap->ref_reloc_sym = ref;
1342 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1344 return machines__fprintf_dsos(&self->machines, fp);
1347 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1348 bool (skip)(struct dso *dso, int parm), int parm)
1350 return machines__fprintf_dsos_buildid(&self->machines, fp, skip, parm);
1353 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1355 struct perf_evsel *pos;
1356 size_t ret = fprintf(fp, "Aggregated stats:\n");
1358 ret += events_stats__fprintf(&session->stats, fp);
1360 list_for_each_entry(pos, &session->evlist->entries, node) {
1361 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1362 ret += events_stats__fprintf(&pos->hists.stats, fp);
1368 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1371 * FIXME: Here we have to actually print all the machines in this
1372 * session, not just the host...
1374 return machine__fprintf(&session->machines.host, fp);
1377 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1380 struct perf_evsel *pos;
1382 list_for_each_entry(pos, &session->evlist->entries, node) {
1383 if (pos->attr.type == type)
1389 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1390 struct perf_sample *sample, struct machine *machine,
1391 int print_sym, int print_dso, int print_symoffset)
1393 struct addr_location al;
1394 struct callchain_cursor_node *node;
1396 if (perf_event__preprocess_sample(event, machine, &al, sample,
1398 error("problem processing %d event, skipping it.\n",
1399 event->header.type);
1403 if (symbol_conf.use_callchain && sample->callchain) {
1406 if (machine__resolve_callchain(machine, evsel, al.thread,
1407 sample, NULL) != 0) {
1409 error("Failed to resolve callchain. Skipping\n");
1412 callchain_cursor_commit(&callchain_cursor);
1415 node = callchain_cursor_current(&callchain_cursor);
1419 printf("\t%16" PRIx64, node->ip);
1422 symbol__fprintf_symname(node->sym, stdout);
1426 map__fprintf_dsoname(node->map, stdout);
1431 callchain_cursor_advance(&callchain_cursor);
1435 printf("%16" PRIx64, sample->ip);
1438 if (print_symoffset)
1439 symbol__fprintf_symname_offs(al.sym, &al,
1442 symbol__fprintf_symname(al.sym, stdout);
1447 map__fprintf_dsoname(al.map, stdout);
1453 int perf_session__cpu_bitmap(struct perf_session *session,
1454 const char *cpu_list, unsigned long *cpu_bitmap)
1457 struct cpu_map *map;
1459 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1460 struct perf_evsel *evsel;
1462 evsel = perf_session__find_first_evtype(session, i);
1466 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1467 pr_err("File does not contain CPU events. "
1468 "Remove -c option to proceed.\n");
1473 map = cpu_map__new(cpu_list);
1475 pr_err("Invalid cpu_list\n");
1479 for (i = 0; i < map->nr; i++) {
1480 int cpu = map->map[i];
1482 if (cpu >= MAX_NR_CPUS) {
1483 pr_err("Requested CPU %d too large. "
1484 "Consider raising MAX_NR_CPUS\n", cpu);
1488 set_bit(cpu, cpu_bitmap);
1494 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1500 if (session == NULL || fp == NULL)
1503 ret = fstat(session->fd, &st);
1507 fprintf(fp, "# ========\n");
1508 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1509 perf_header__fprintf_info(session, fp, full);
1510 fprintf(fp, "# ========\n#\n");
1514 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1515 const struct perf_evsel_str_handler *assocs,
1518 struct perf_evlist *evlist = session->evlist;
1519 struct event_format *format;
1520 struct perf_evsel *evsel;
1521 char *tracepoint, *name;
1525 for (i = 0; i < nr_assocs; i++) {
1527 tracepoint = strdup(assocs[i].name);
1528 if (tracepoint == NULL)
1532 name = strchr(tracepoint, ':');
1537 format = pevent_find_event_by_name(session->pevent,
1539 if (format == NULL) {
1541 * Adding a handler for an event not in the session,
1547 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1552 if (evsel->handler.func != NULL)
1554 evsel->handler.func = assocs[i].handler;