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;
507 bool show_progress = limit == ULLONG_MAX;
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 (show_progress && (++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 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__combined_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 sample_read__printf(struct perf_sample *sample, u64 read_format)
760 printf("... sample_read:\n");
762 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
763 printf("...... time enabled %016" PRIx64 "\n",
764 sample->read.time_enabled);
766 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
767 printf("...... time running %016" PRIx64 "\n",
768 sample->read.time_running);
770 if (read_format & PERF_FORMAT_GROUP) {
773 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
775 for (i = 0; i < sample->read.group.nr; i++) {
776 struct sample_read_value *value;
778 value = &sample->read.group.values[i];
779 printf("..... id %016" PRIx64
780 ", value %016" PRIx64 "\n",
781 value->id, value->value);
784 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
785 sample->read.one.id, sample->read.one.value);
788 static void dump_event(struct perf_session *session, union perf_event *event,
789 u64 file_offset, struct perf_sample *sample)
794 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
795 file_offset, event->header.size, event->header.type);
800 perf_session__print_tstamp(session, event, sample);
802 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
803 event->header.size, perf_event__name(event->header.type));
806 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
807 struct perf_sample *sample)
814 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
815 event->header.misc, sample->pid, sample->tid, sample->ip,
816 sample->period, sample->addr);
818 sample_type = evsel->attr.sample_type;
820 if (sample_type & PERF_SAMPLE_CALLCHAIN)
821 callchain__printf(sample);
823 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
824 branch_stack__printf(sample);
826 if (sample_type & PERF_SAMPLE_REGS_USER)
827 regs_user__printf(sample, evsel->attr.sample_regs_user);
829 if (sample_type & PERF_SAMPLE_STACK_USER)
830 stack_user__printf(&sample->user_stack);
832 if (sample_type & PERF_SAMPLE_WEIGHT)
833 printf("... weight: %" PRIu64 "\n", sample->weight);
835 if (sample_type & PERF_SAMPLE_DATA_SRC)
836 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
838 if (sample_type & PERF_SAMPLE_READ)
839 sample_read__printf(sample, evsel->attr.read_format);
842 static struct machine *
843 perf_session__find_machine_for_cpumode(struct perf_session *session,
844 union perf_event *event,
845 struct perf_sample *sample)
847 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
850 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
851 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
854 if (event->header.type == PERF_RECORD_MMAP)
855 pid = event->mmap.pid;
859 return perf_session__findnew_machine(session, pid);
862 return &session->machines.host;
865 static int deliver_sample_value(struct perf_session *session,
866 struct perf_tool *tool,
867 union perf_event *event,
868 struct perf_sample *sample,
869 struct sample_read_value *v,
870 struct machine *machine)
872 struct perf_sample_id *sid;
874 sid = perf_evlist__id2sid(session->evlist, v->id);
877 sample->period = v->value - sid->period;
878 sid->period = v->value;
881 if (!sid || sid->evsel == NULL) {
882 ++session->stats.nr_unknown_id;
886 return tool->sample(tool, event, sample, sid->evsel, machine);
889 static int deliver_sample_group(struct perf_session *session,
890 struct perf_tool *tool,
891 union perf_event *event,
892 struct perf_sample *sample,
893 struct machine *machine)
898 for (i = 0; i < sample->read.group.nr; i++) {
899 ret = deliver_sample_value(session, tool, event, sample,
900 &sample->read.group.values[i],
910 perf_session__deliver_sample(struct perf_session *session,
911 struct perf_tool *tool,
912 union perf_event *event,
913 struct perf_sample *sample,
914 struct perf_evsel *evsel,
915 struct machine *machine)
917 /* We know evsel != NULL. */
918 u64 sample_type = evsel->attr.sample_type;
919 u64 read_format = evsel->attr.read_format;
921 /* Standard sample delievery. */
922 if (!(sample_type & PERF_SAMPLE_READ))
923 return tool->sample(tool, event, sample, evsel, machine);
925 /* For PERF_SAMPLE_READ we have either single or group mode. */
926 if (read_format & PERF_FORMAT_GROUP)
927 return deliver_sample_group(session, tool, event, sample,
930 return deliver_sample_value(session, tool, event, sample,
931 &sample->read.one, machine);
934 static int perf_session_deliver_event(struct perf_session *session,
935 union perf_event *event,
936 struct perf_sample *sample,
937 struct perf_tool *tool,
940 struct perf_evsel *evsel;
941 struct machine *machine;
943 dump_event(session, event, file_offset, sample);
945 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
946 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
948 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
949 * because the tools right now may apply filters, discarding
950 * some of the samples. For consistency, in the future we
951 * should have something like nr_filtered_samples and remove
952 * the sample->period from total_sample_period, etc, KISS for
955 * Also testing against NULL allows us to handle files without
956 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
957 * future probably it'll be a good idea to restrict event
958 * processing via perf_session to files with both set.
960 hists__inc_nr_events(&evsel->hists, event->header.type);
963 machine = perf_session__find_machine_for_cpumode(session, event,
966 switch (event->header.type) {
967 case PERF_RECORD_SAMPLE:
968 dump_sample(evsel, event, sample);
970 ++session->stats.nr_unknown_id;
973 if (machine == NULL) {
974 ++session->stats.nr_unprocessable_samples;
977 return perf_session__deliver_sample(session, tool, event,
978 sample, evsel, machine);
979 case PERF_RECORD_MMAP:
980 return tool->mmap(tool, event, sample, machine);
981 case PERF_RECORD_COMM:
982 return tool->comm(tool, event, sample, machine);
983 case PERF_RECORD_FORK:
984 return tool->fork(tool, event, sample, machine);
985 case PERF_RECORD_EXIT:
986 return tool->exit(tool, event, sample, machine);
987 case PERF_RECORD_LOST:
988 if (tool->lost == perf_event__process_lost)
989 session->stats.total_lost += event->lost.lost;
990 return tool->lost(tool, event, sample, machine);
991 case PERF_RECORD_READ:
992 return tool->read(tool, event, sample, evsel, machine);
993 case PERF_RECORD_THROTTLE:
994 return tool->throttle(tool, event, sample, machine);
995 case PERF_RECORD_UNTHROTTLE:
996 return tool->unthrottle(tool, event, sample, machine);
998 ++session->stats.nr_unknown_events;
1003 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1004 struct perf_tool *tool, u64 file_offset)
1008 dump_event(session, event, file_offset, NULL);
1010 /* These events are processed right away */
1011 switch (event->header.type) {
1012 case PERF_RECORD_HEADER_ATTR:
1013 err = tool->attr(tool, event, &session->evlist);
1015 perf_session__set_id_hdr_size(session);
1017 case PERF_RECORD_HEADER_TRACING_DATA:
1018 /* setup for reading amidst mmap */
1019 lseek(session->fd, file_offset, SEEK_SET);
1020 return tool->tracing_data(tool, event, session);
1021 case PERF_RECORD_HEADER_BUILD_ID:
1022 return tool->build_id(tool, event, session);
1023 case PERF_RECORD_FINISHED_ROUND:
1024 return tool->finished_round(tool, event, session);
1030 static void event_swap(union perf_event *event, bool sample_id_all)
1032 perf_event__swap_op swap;
1034 swap = perf_event__swap_ops[event->header.type];
1036 swap(event, sample_id_all);
1039 static int perf_session__process_event(struct perf_session *session,
1040 union perf_event *event,
1041 struct perf_tool *tool,
1044 struct perf_sample sample;
1047 if (session->header.needs_swap)
1048 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1050 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1053 events_stats__inc(&session->stats, event->header.type);
1055 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1056 return perf_session__process_user_event(session, event, tool, file_offset);
1059 * For all kernel events we get the sample data
1061 ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1065 if (tool->ordered_samples) {
1066 ret = perf_session_queue_event(session, event, &sample,
1072 return perf_session_deliver_event(session, event, &sample, tool,
1076 void perf_event_header__bswap(struct perf_event_header *self)
1078 self->type = bswap_32(self->type);
1079 self->misc = bswap_16(self->misc);
1080 self->size = bswap_16(self->size);
1083 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1085 return machine__findnew_thread(&session->machines.host, 0, pid);
1088 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1090 struct thread *thread = perf_session__findnew(self, 0);
1092 if (thread == NULL || thread__set_comm(thread, "swapper")) {
1093 pr_err("problem inserting idle task.\n");
1100 static void perf_session__warn_about_errors(const struct perf_session *session,
1101 const struct perf_tool *tool)
1103 if (tool->lost == perf_event__process_lost &&
1104 session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1105 ui__warning("Processed %d events and lost %d chunks!\n\n"
1106 "Check IO/CPU overload!\n\n",
1107 session->stats.nr_events[0],
1108 session->stats.nr_events[PERF_RECORD_LOST]);
1111 if (session->stats.nr_unknown_events != 0) {
1112 ui__warning("Found %u unknown events!\n\n"
1113 "Is this an older tool processing a perf.data "
1114 "file generated by a more recent tool?\n\n"
1115 "If that is not the case, consider "
1116 "reporting to linux-kernel@vger.kernel.org.\n\n",
1117 session->stats.nr_unknown_events);
1120 if (session->stats.nr_unknown_id != 0) {
1121 ui__warning("%u samples with id not present in the header\n",
1122 session->stats.nr_unknown_id);
1125 if (session->stats.nr_invalid_chains != 0) {
1126 ui__warning("Found invalid callchains!\n\n"
1127 "%u out of %u events were discarded for this reason.\n\n"
1128 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1129 session->stats.nr_invalid_chains,
1130 session->stats.nr_events[PERF_RECORD_SAMPLE]);
1133 if (session->stats.nr_unprocessable_samples != 0) {
1134 ui__warning("%u unprocessable samples recorded.\n"
1135 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1136 session->stats.nr_unprocessable_samples);
1140 #define session_done() (*(volatile int *)(&session_done))
1141 volatile int session_done;
1143 static int __perf_session__process_pipe_events(struct perf_session *self,
1144 struct perf_tool *tool)
1146 union perf_event *event;
1147 uint32_t size, cur_size = 0;
1154 perf_tool__fill_defaults(tool);
1157 cur_size = sizeof(union perf_event);
1159 buf = malloc(cur_size);
1164 err = readn(self->fd, event, sizeof(struct perf_event_header));
1169 pr_err("failed to read event header\n");
1173 if (self->header.needs_swap)
1174 perf_event_header__bswap(&event->header);
1176 size = event->header.size;
1177 if (size < sizeof(struct perf_event_header)) {
1178 pr_err("bad event header size\n");
1182 if (size > cur_size) {
1183 void *new = realloc(buf, size);
1185 pr_err("failed to allocate memory to read event\n");
1193 p += sizeof(struct perf_event_header);
1195 if (size - sizeof(struct perf_event_header)) {
1196 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1199 pr_err("unexpected end of event stream\n");
1203 pr_err("failed to read event data\n");
1208 if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1209 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1210 head, event->header.size, event->header.type);
1220 if (!session_done())
1226 perf_session__warn_about_errors(self, tool);
1227 perf_session_free_sample_buffers(self);
1231 static union perf_event *
1232 fetch_mmaped_event(struct perf_session *session,
1233 u64 head, size_t mmap_size, char *buf)
1235 union perf_event *event;
1238 * Ensure we have enough space remaining to read
1239 * the size of the event in the headers.
1241 if (head + sizeof(event->header) > mmap_size)
1244 event = (union perf_event *)(buf + head);
1246 if (session->header.needs_swap)
1247 perf_event_header__bswap(&event->header);
1249 if (head + event->header.size > mmap_size) {
1250 /* We're not fetching the event so swap back again */
1251 if (session->header.needs_swap)
1252 perf_event_header__bswap(&event->header);
1260 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1261 * slices. On 32bit we use 32MB.
1263 #if BITS_PER_LONG == 64
1264 #define MMAP_SIZE ULLONG_MAX
1267 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1268 #define NUM_MMAPS 128
1271 int __perf_session__process_events(struct perf_session *session,
1272 u64 data_offset, u64 data_size,
1273 u64 file_size, struct perf_tool *tool)
1275 u64 head, page_offset, file_offset, file_pos, progress_next;
1276 int err, mmap_prot, mmap_flags, map_idx = 0;
1278 char *buf, *mmaps[NUM_MMAPS];
1279 union perf_event *event;
1282 perf_tool__fill_defaults(tool);
1284 page_offset = page_size * (data_offset / page_size);
1285 file_offset = page_offset;
1286 head = data_offset - page_offset;
1288 if (data_offset + data_size < file_size)
1289 file_size = data_offset + data_size;
1291 progress_next = file_size / 16;
1293 mmap_size = MMAP_SIZE;
1294 if (mmap_size > file_size)
1295 mmap_size = file_size;
1297 memset(mmaps, 0, sizeof(mmaps));
1299 mmap_prot = PROT_READ;
1300 mmap_flags = MAP_SHARED;
1302 if (session->header.needs_swap) {
1303 mmap_prot |= PROT_WRITE;
1304 mmap_flags = MAP_PRIVATE;
1307 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1309 if (buf == MAP_FAILED) {
1310 pr_err("failed to mmap file\n");
1314 mmaps[map_idx] = buf;
1315 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1316 file_pos = file_offset + head;
1319 event = fetch_mmaped_event(session, head, mmap_size, buf);
1321 if (mmaps[map_idx]) {
1322 munmap(mmaps[map_idx], mmap_size);
1323 mmaps[map_idx] = NULL;
1326 page_offset = page_size * (head / page_size);
1327 file_offset += page_offset;
1328 head -= page_offset;
1332 size = event->header.size;
1334 if (size < sizeof(struct perf_event_header) ||
1335 perf_session__process_event(session, event, tool, file_pos) < 0) {
1336 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1337 file_offset + head, event->header.size,
1338 event->header.type);
1346 if (file_pos >= progress_next) {
1347 progress_next += file_size / 16;
1348 ui_progress__update(file_pos, file_size,
1349 "Processing events...");
1352 if (file_pos < file_size)
1356 /* do the final flush for ordered samples */
1357 session->ordered_samples.next_flush = ULLONG_MAX;
1358 err = flush_sample_queue(session, tool);
1360 ui_progress__finish();
1361 perf_session__warn_about_errors(session, tool);
1362 perf_session_free_sample_buffers(session);
1366 int perf_session__process_events(struct perf_session *self,
1367 struct perf_tool *tool)
1371 if (perf_session__register_idle_thread(self) == NULL)
1375 err = __perf_session__process_events(self,
1376 self->header.data_offset,
1377 self->header.data_size,
1380 err = __perf_session__process_pipe_events(self, tool);
1385 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1387 struct perf_evsel *evsel;
1389 list_for_each_entry(evsel, &session->evlist->entries, node) {
1390 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1394 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1398 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1399 const char *symbol_name, u64 addr)
1403 struct ref_reloc_sym *ref;
1405 ref = zalloc(sizeof(struct ref_reloc_sym));
1409 ref->name = strdup(symbol_name);
1410 if (ref->name == NULL) {
1415 bracket = strchr(ref->name, ']');
1421 for (i = 0; i < MAP__NR_TYPES; ++i) {
1422 struct kmap *kmap = map__kmap(maps[i]);
1423 kmap->ref_reloc_sym = ref;
1429 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1431 return machines__fprintf_dsos(&self->machines, fp);
1434 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1435 bool (skip)(struct dso *dso, int parm), int parm)
1437 return machines__fprintf_dsos_buildid(&self->machines, fp, skip, parm);
1440 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1442 struct perf_evsel *pos;
1443 size_t ret = fprintf(fp, "Aggregated stats:\n");
1445 ret += events_stats__fprintf(&session->stats, fp);
1447 list_for_each_entry(pos, &session->evlist->entries, node) {
1448 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1449 ret += events_stats__fprintf(&pos->hists.stats, fp);
1455 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1458 * FIXME: Here we have to actually print all the machines in this
1459 * session, not just the host...
1461 return machine__fprintf(&session->machines.host, fp);
1464 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1467 struct perf_evsel *pos;
1469 list_for_each_entry(pos, &session->evlist->entries, node) {
1470 if (pos->attr.type == type)
1476 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1477 struct perf_sample *sample, struct machine *machine,
1478 unsigned int print_opts, unsigned int stack_depth)
1480 struct addr_location al;
1481 struct callchain_cursor_node *node;
1482 int print_ip = print_opts & PRINT_IP_OPT_IP;
1483 int print_sym = print_opts & PRINT_IP_OPT_SYM;
1484 int print_dso = print_opts & PRINT_IP_OPT_DSO;
1485 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1486 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1487 char s = print_oneline ? ' ' : '\t';
1489 if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
1490 error("problem processing %d event, skipping it.\n",
1491 event->header.type);
1495 if (symbol_conf.use_callchain && sample->callchain) {
1497 if (machine__resolve_callchain(machine, evsel, al.thread,
1498 sample, NULL, NULL) != 0) {
1500 error("Failed to resolve callchain. Skipping\n");
1503 callchain_cursor_commit(&callchain_cursor);
1505 while (stack_depth) {
1506 node = callchain_cursor_current(&callchain_cursor);
1511 printf("%c%16" PRIx64, s, node->ip);
1515 if (print_symoffset) {
1518 symbol__fprintf_symname_offs(node->sym, &al, stdout);
1520 symbol__fprintf_symname(node->sym, stdout);
1525 map__fprintf_dsoname(node->map, stdout);
1532 callchain_cursor_advance(&callchain_cursor);
1539 printf("%16" PRIx64, sample->ip);
1543 if (print_symoffset)
1544 symbol__fprintf_symname_offs(al.sym, &al,
1547 symbol__fprintf_symname(al.sym, stdout);
1552 map__fprintf_dsoname(al.map, stdout);
1558 int perf_session__cpu_bitmap(struct perf_session *session,
1559 const char *cpu_list, unsigned long *cpu_bitmap)
1562 struct cpu_map *map;
1564 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1565 struct perf_evsel *evsel;
1567 evsel = perf_session__find_first_evtype(session, i);
1571 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1572 pr_err("File does not contain CPU events. "
1573 "Remove -c option to proceed.\n");
1578 map = cpu_map__new(cpu_list);
1580 pr_err("Invalid cpu_list\n");
1584 for (i = 0; i < map->nr; i++) {
1585 int cpu = map->map[i];
1587 if (cpu >= MAX_NR_CPUS) {
1588 pr_err("Requested CPU %d too large. "
1589 "Consider raising MAX_NR_CPUS\n", cpu);
1593 set_bit(cpu, cpu_bitmap);
1599 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1605 if (session == NULL || fp == NULL)
1608 ret = fstat(session->fd, &st);
1612 fprintf(fp, "# ========\n");
1613 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1614 perf_header__fprintf_info(session, fp, full);
1615 fprintf(fp, "# ========\n#\n");
1619 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1620 const struct perf_evsel_str_handler *assocs,
1623 struct perf_evlist *evlist = session->evlist;
1624 struct event_format *format;
1625 struct perf_evsel *evsel;
1626 char *tracepoint, *name;
1630 for (i = 0; i < nr_assocs; i++) {
1632 tracepoint = strdup(assocs[i].name);
1633 if (tracepoint == NULL)
1637 name = strchr(tracepoint, ':');
1642 format = pevent_find_event_by_name(session->pevent,
1644 if (format == NULL) {
1646 * Adding a handler for an event not in the session,
1652 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1657 if (evsel->handler.func != NULL)
1659 evsel->handler.func = assocs[i].handler;