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 static bool perf_session__has_comm_exec(struct perf_session *session)
72 struct perf_evsel *evsel;
74 evlist__for_each(session->evlist, evsel) {
75 if (evsel->attr.comm_exec)
82 static void perf_session__set_comm_exec(struct perf_session *session)
84 bool comm_exec = perf_session__has_comm_exec(session);
86 machines__set_comm_exec(&session->machines, comm_exec);
89 struct perf_session *perf_session__new(struct perf_data_file *file,
90 bool repipe, struct perf_tool *tool)
92 struct perf_session *session = zalloc(sizeof(*session));
97 session->repipe = repipe;
98 ordered_events__init(&session->ordered_events);
99 machines__init(&session->machines);
102 if (perf_data_file__open(file))
105 session->file = file;
107 if (perf_data_file__is_read(file)) {
108 if (perf_session__open(session) < 0)
111 perf_session__set_id_hdr_size(session);
112 perf_session__set_comm_exec(session);
116 if (!file || perf_data_file__is_write(file)) {
118 * In O_RDONLY mode this will be performed when reading the
119 * kernel MMAP event, in perf_event__process_mmap().
121 if (perf_session__create_kernel_maps(session) < 0)
122 pr_warning("Cannot read kernel map\n");
125 if (tool && tool->ordering_requires_timestamps &&
126 tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) {
127 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
128 tool->ordered_events = false;
134 perf_data_file__close(file);
136 perf_session__delete(session);
141 static void perf_session__delete_dead_threads(struct perf_session *session)
143 machine__delete_dead_threads(&session->machines.host);
146 static void perf_session__delete_threads(struct perf_session *session)
148 machine__delete_threads(&session->machines.host);
151 static void perf_session_env__delete(struct perf_session_env *env)
153 zfree(&env->hostname);
154 zfree(&env->os_release);
155 zfree(&env->version);
157 zfree(&env->cpu_desc);
160 zfree(&env->cmdline);
161 zfree(&env->sibling_cores);
162 zfree(&env->sibling_threads);
163 zfree(&env->numa_nodes);
164 zfree(&env->pmu_mappings);
167 void perf_session__delete(struct perf_session *session)
169 perf_session__destroy_kernel_maps(session);
170 perf_session__delete_dead_threads(session);
171 perf_session__delete_threads(session);
172 perf_session_env__delete(&session->header.env);
173 machines__exit(&session->machines);
175 perf_data_file__close(session->file);
179 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
181 union perf_event *event
183 struct perf_session *session
186 dump_printf(": unhandled!\n");
190 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
191 union perf_event *event __maybe_unused,
192 struct perf_evlist **pevlist
195 dump_printf(": unhandled!\n");
199 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
200 union perf_event *event __maybe_unused,
201 struct perf_sample *sample __maybe_unused,
202 struct perf_evsel *evsel __maybe_unused,
203 struct machine *machine __maybe_unused)
205 dump_printf(": unhandled!\n");
209 static int process_event_stub(struct perf_tool *tool __maybe_unused,
210 union perf_event *event __maybe_unused,
211 struct perf_sample *sample __maybe_unused,
212 struct machine *machine __maybe_unused)
214 dump_printf(": unhandled!\n");
218 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
219 union perf_event *event __maybe_unused,
220 struct perf_session *perf_session
223 dump_printf(": unhandled!\n");
227 static int process_finished_round(struct perf_tool *tool,
228 union perf_event *event,
229 struct perf_session *session);
231 void perf_tool__fill_defaults(struct perf_tool *tool)
233 if (tool->sample == NULL)
234 tool->sample = process_event_sample_stub;
235 if (tool->mmap == NULL)
236 tool->mmap = process_event_stub;
237 if (tool->mmap2 == NULL)
238 tool->mmap2 = process_event_stub;
239 if (tool->comm == NULL)
240 tool->comm = process_event_stub;
241 if (tool->fork == NULL)
242 tool->fork = process_event_stub;
243 if (tool->exit == NULL)
244 tool->exit = process_event_stub;
245 if (tool->lost == NULL)
246 tool->lost = perf_event__process_lost;
247 if (tool->read == NULL)
248 tool->read = process_event_sample_stub;
249 if (tool->throttle == NULL)
250 tool->throttle = process_event_stub;
251 if (tool->unthrottle == NULL)
252 tool->unthrottle = process_event_stub;
253 if (tool->attr == NULL)
254 tool->attr = process_event_synth_attr_stub;
255 if (tool->tracing_data == NULL)
256 tool->tracing_data = process_event_synth_tracing_data_stub;
257 if (tool->build_id == NULL)
258 tool->build_id = process_finished_round_stub;
259 if (tool->finished_round == NULL) {
260 if (tool->ordered_events)
261 tool->finished_round = process_finished_round;
263 tool->finished_round = process_finished_round_stub;
267 static void swap_sample_id_all(union perf_event *event, void *data)
269 void *end = (void *) event + event->header.size;
270 int size = end - data;
272 BUG_ON(size % sizeof(u64));
273 mem_bswap_64(data, size);
276 static void perf_event__all64_swap(union perf_event *event,
277 bool sample_id_all __maybe_unused)
279 struct perf_event_header *hdr = &event->header;
280 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
283 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
285 event->comm.pid = bswap_32(event->comm.pid);
286 event->comm.tid = bswap_32(event->comm.tid);
289 void *data = &event->comm.comm;
291 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
292 swap_sample_id_all(event, data);
296 static void perf_event__mmap_swap(union perf_event *event,
299 event->mmap.pid = bswap_32(event->mmap.pid);
300 event->mmap.tid = bswap_32(event->mmap.tid);
301 event->mmap.start = bswap_64(event->mmap.start);
302 event->mmap.len = bswap_64(event->mmap.len);
303 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
306 void *data = &event->mmap.filename;
308 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
309 swap_sample_id_all(event, data);
313 static void perf_event__mmap2_swap(union perf_event *event,
316 event->mmap2.pid = bswap_32(event->mmap2.pid);
317 event->mmap2.tid = bswap_32(event->mmap2.tid);
318 event->mmap2.start = bswap_64(event->mmap2.start);
319 event->mmap2.len = bswap_64(event->mmap2.len);
320 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
321 event->mmap2.maj = bswap_32(event->mmap2.maj);
322 event->mmap2.min = bswap_32(event->mmap2.min);
323 event->mmap2.ino = bswap_64(event->mmap2.ino);
326 void *data = &event->mmap2.filename;
328 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
329 swap_sample_id_all(event, data);
332 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
334 event->fork.pid = bswap_32(event->fork.pid);
335 event->fork.tid = bswap_32(event->fork.tid);
336 event->fork.ppid = bswap_32(event->fork.ppid);
337 event->fork.ptid = bswap_32(event->fork.ptid);
338 event->fork.time = bswap_64(event->fork.time);
341 swap_sample_id_all(event, &event->fork + 1);
344 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
346 event->read.pid = bswap_32(event->read.pid);
347 event->read.tid = bswap_32(event->read.tid);
348 event->read.value = bswap_64(event->read.value);
349 event->read.time_enabled = bswap_64(event->read.time_enabled);
350 event->read.time_running = bswap_64(event->read.time_running);
351 event->read.id = bswap_64(event->read.id);
354 swap_sample_id_all(event, &event->read + 1);
357 static void perf_event__throttle_swap(union perf_event *event,
360 event->throttle.time = bswap_64(event->throttle.time);
361 event->throttle.id = bswap_64(event->throttle.id);
362 event->throttle.stream_id = bswap_64(event->throttle.stream_id);
365 swap_sample_id_all(event, &event->throttle + 1);
368 static u8 revbyte(u8 b)
370 int rev = (b >> 4) | ((b & 0xf) << 4);
371 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
372 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
377 * XXX this is hack in attempt to carry flags bitfield
378 * throught endian village. ABI says:
380 * Bit-fields are allocated from right to left (least to most significant)
381 * on little-endian implementations and from left to right (most to least
382 * significant) on big-endian implementations.
384 * The above seems to be byte specific, so we need to reverse each
385 * byte of the bitfield. 'Internet' also says this might be implementation
386 * specific and we probably need proper fix and carry perf_event_attr
387 * bitfield flags in separate data file FEAT_ section. Thought this seems
390 static void swap_bitfield(u8 *p, unsigned len)
394 for (i = 0; i < len; i++) {
400 /* exported for swapping attributes in file header */
401 void perf_event__attr_swap(struct perf_event_attr *attr)
403 attr->type = bswap_32(attr->type);
404 attr->size = bswap_32(attr->size);
405 attr->config = bswap_64(attr->config);
406 attr->sample_period = bswap_64(attr->sample_period);
407 attr->sample_type = bswap_64(attr->sample_type);
408 attr->read_format = bswap_64(attr->read_format);
409 attr->wakeup_events = bswap_32(attr->wakeup_events);
410 attr->bp_type = bswap_32(attr->bp_type);
411 attr->bp_addr = bswap_64(attr->bp_addr);
412 attr->bp_len = bswap_64(attr->bp_len);
413 attr->branch_sample_type = bswap_64(attr->branch_sample_type);
414 attr->sample_regs_user = bswap_64(attr->sample_regs_user);
415 attr->sample_stack_user = bswap_32(attr->sample_stack_user);
417 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
420 static void perf_event__hdr_attr_swap(union perf_event *event,
421 bool sample_id_all __maybe_unused)
425 perf_event__attr_swap(&event->attr.attr);
427 size = event->header.size;
428 size -= (void *)&event->attr.id - (void *)event;
429 mem_bswap_64(event->attr.id, size);
432 static void perf_event__event_type_swap(union perf_event *event,
433 bool sample_id_all __maybe_unused)
435 event->event_type.event_type.event_id =
436 bswap_64(event->event_type.event_type.event_id);
439 static void perf_event__tracing_data_swap(union perf_event *event,
440 bool sample_id_all __maybe_unused)
442 event->tracing_data.size = bswap_32(event->tracing_data.size);
445 typedef void (*perf_event__swap_op)(union perf_event *event,
448 static perf_event__swap_op perf_event__swap_ops[] = {
449 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
450 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap,
451 [PERF_RECORD_COMM] = perf_event__comm_swap,
452 [PERF_RECORD_FORK] = perf_event__task_swap,
453 [PERF_RECORD_EXIT] = perf_event__task_swap,
454 [PERF_RECORD_LOST] = perf_event__all64_swap,
455 [PERF_RECORD_READ] = perf_event__read_swap,
456 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap,
457 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap,
458 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
459 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
460 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
461 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
462 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
463 [PERF_RECORD_HEADER_MAX] = NULL,
467 * When perf record finishes a pass on every buffers, it records this pseudo
469 * We record the max timestamp t found in the pass n.
470 * Assuming these timestamps are monotonic across cpus, we know that if
471 * a buffer still has events with timestamps below t, they will be all
472 * available and then read in the pass n + 1.
473 * Hence when we start to read the pass n + 2, we can safely flush every
474 * events with timestamps below t.
476 * ============ PASS n =================
479 * cnt1 timestamps | cnt2 timestamps
482 * - | 4 <--- max recorded
484 * ============ PASS n + 1 ==============
487 * cnt1 timestamps | cnt2 timestamps
490 * 5 | 7 <---- max recorded
492 * Flush every events below timestamp 4
494 * ============ PASS n + 2 ==============
497 * cnt1 timestamps | cnt2 timestamps
502 * Flush every events below timestamp 7
505 static int process_finished_round(struct perf_tool *tool,
506 union perf_event *event __maybe_unused,
507 struct perf_session *session)
509 return ordered_events__flush(session, tool, OE_FLUSH__ROUND);
512 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
513 struct perf_tool *tool, struct perf_sample *sample,
516 struct ordered_events *oe = &s->ordered_events;
517 u64 timestamp = sample->time;
518 struct ordered_event *new;
520 if (!timestamp || timestamp == ~0ULL)
523 if (timestamp < oe->last_flush) {
524 WARN_ONCE(1, "Timestamp below last timeslice flush\n");
526 pr_oe_time(timestamp, "out of order event");
527 pr_oe_time(oe->last_flush, "last flush, last_flush_type %d\n",
528 oe->last_flush_type);
530 /* We could get out of order messages after forced flush. */
531 if (oe->last_flush_type != OE_FLUSH__HALF)
535 new = ordered_events__new(oe, timestamp, event);
537 ordered_events__flush(s, tool, OE_FLUSH__HALF);
538 new = ordered_events__new(oe, timestamp, event);
544 new->file_offset = file_offset;
548 static void callchain__printf(struct perf_sample *sample)
552 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
554 for (i = 0; i < sample->callchain->nr; i++)
555 printf("..... %2d: %016" PRIx64 "\n",
556 i, sample->callchain->ips[i]);
559 static void branch_stack__printf(struct perf_sample *sample)
563 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
565 for (i = 0; i < sample->branch_stack->nr; i++)
566 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
567 i, sample->branch_stack->entries[i].from,
568 sample->branch_stack->entries[i].to);
571 static void regs_dump__printf(u64 mask, u64 *regs)
575 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
578 printf(".... %-5s 0x%" PRIx64 "\n",
579 perf_reg_name(rid), val);
583 static void regs_user__printf(struct perf_sample *sample)
585 struct regs_dump *user_regs = &sample->user_regs;
587 if (user_regs->regs) {
588 u64 mask = user_regs->mask;
589 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
590 regs_dump__printf(mask, user_regs->regs);
594 static void stack_user__printf(struct stack_dump *dump)
596 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
597 dump->size, dump->offset);
600 static void perf_session__print_tstamp(struct perf_session *session,
601 union perf_event *event,
602 struct perf_sample *sample)
604 u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
606 if (event->header.type != PERF_RECORD_SAMPLE &&
607 !perf_evlist__sample_id_all(session->evlist)) {
608 fputs("-1 -1 ", stdout);
612 if ((sample_type & PERF_SAMPLE_CPU))
613 printf("%u ", sample->cpu);
615 if (sample_type & PERF_SAMPLE_TIME)
616 printf("%" PRIu64 " ", sample->time);
619 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
621 printf("... sample_read:\n");
623 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
624 printf("...... time enabled %016" PRIx64 "\n",
625 sample->read.time_enabled);
627 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
628 printf("...... time running %016" PRIx64 "\n",
629 sample->read.time_running);
631 if (read_format & PERF_FORMAT_GROUP) {
634 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
636 for (i = 0; i < sample->read.group.nr; i++) {
637 struct sample_read_value *value;
639 value = &sample->read.group.values[i];
640 printf("..... id %016" PRIx64
641 ", value %016" PRIx64 "\n",
642 value->id, value->value);
645 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
646 sample->read.one.id, sample->read.one.value);
649 static void dump_event(struct perf_session *session, union perf_event *event,
650 u64 file_offset, struct perf_sample *sample)
655 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
656 file_offset, event->header.size, event->header.type);
661 perf_session__print_tstamp(session, event, sample);
663 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
664 event->header.size, perf_event__name(event->header.type));
667 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
668 struct perf_sample *sample)
675 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
676 event->header.misc, sample->pid, sample->tid, sample->ip,
677 sample->period, sample->addr);
679 sample_type = evsel->attr.sample_type;
681 if (sample_type & PERF_SAMPLE_CALLCHAIN)
682 callchain__printf(sample);
684 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
685 branch_stack__printf(sample);
687 if (sample_type & PERF_SAMPLE_REGS_USER)
688 regs_user__printf(sample);
690 if (sample_type & PERF_SAMPLE_STACK_USER)
691 stack_user__printf(&sample->user_stack);
693 if (sample_type & PERF_SAMPLE_WEIGHT)
694 printf("... weight: %" PRIu64 "\n", sample->weight);
696 if (sample_type & PERF_SAMPLE_DATA_SRC)
697 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
699 if (sample_type & PERF_SAMPLE_TRANSACTION)
700 printf("... transaction: %" PRIx64 "\n", sample->transaction);
702 if (sample_type & PERF_SAMPLE_READ)
703 sample_read__printf(sample, evsel->attr.read_format);
706 static struct machine *
707 perf_session__find_machine_for_cpumode(struct perf_session *session,
708 union perf_event *event,
709 struct perf_sample *sample)
711 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
712 struct machine *machine;
715 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
716 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
719 if (event->header.type == PERF_RECORD_MMAP
720 || event->header.type == PERF_RECORD_MMAP2)
721 pid = event->mmap.pid;
725 machine = perf_session__find_machine(session, pid);
727 machine = perf_session__findnew_machine(session,
728 DEFAULT_GUEST_KERNEL_ID);
732 return &session->machines.host;
735 static int deliver_sample_value(struct perf_session *session,
736 struct perf_tool *tool,
737 union perf_event *event,
738 struct perf_sample *sample,
739 struct sample_read_value *v,
740 struct machine *machine)
742 struct perf_sample_id *sid;
744 sid = perf_evlist__id2sid(session->evlist, v->id);
747 sample->period = v->value - sid->period;
748 sid->period = v->value;
751 if (!sid || sid->evsel == NULL) {
752 ++session->stats.nr_unknown_id;
756 return tool->sample(tool, event, sample, sid->evsel, machine);
759 static int deliver_sample_group(struct perf_session *session,
760 struct perf_tool *tool,
761 union perf_event *event,
762 struct perf_sample *sample,
763 struct machine *machine)
768 for (i = 0; i < sample->read.group.nr; i++) {
769 ret = deliver_sample_value(session, tool, event, sample,
770 &sample->read.group.values[i],
780 perf_session__deliver_sample(struct perf_session *session,
781 struct perf_tool *tool,
782 union perf_event *event,
783 struct perf_sample *sample,
784 struct perf_evsel *evsel,
785 struct machine *machine)
787 /* We know evsel != NULL. */
788 u64 sample_type = evsel->attr.sample_type;
789 u64 read_format = evsel->attr.read_format;
791 /* Standard sample delievery. */
792 if (!(sample_type & PERF_SAMPLE_READ))
793 return tool->sample(tool, event, sample, evsel, machine);
795 /* For PERF_SAMPLE_READ we have either single or group mode. */
796 if (read_format & PERF_FORMAT_GROUP)
797 return deliver_sample_group(session, tool, event, sample,
800 return deliver_sample_value(session, tool, event, sample,
801 &sample->read.one, machine);
804 int perf_session__deliver_event(struct perf_session *session,
805 union perf_event *event,
806 struct perf_sample *sample,
807 struct perf_tool *tool, u64 file_offset)
809 struct perf_evsel *evsel;
810 struct machine *machine;
812 dump_event(session, event, file_offset, sample);
814 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
816 machine = perf_session__find_machine_for_cpumode(session, event,
819 switch (event->header.type) {
820 case PERF_RECORD_SAMPLE:
821 dump_sample(evsel, event, sample);
823 ++session->stats.nr_unknown_id;
826 if (machine == NULL) {
827 ++session->stats.nr_unprocessable_samples;
830 return perf_session__deliver_sample(session, tool, event,
831 sample, evsel, machine);
832 case PERF_RECORD_MMAP:
833 return tool->mmap(tool, event, sample, machine);
834 case PERF_RECORD_MMAP2:
835 return tool->mmap2(tool, event, sample, machine);
836 case PERF_RECORD_COMM:
837 return tool->comm(tool, event, sample, machine);
838 case PERF_RECORD_FORK:
839 return tool->fork(tool, event, sample, machine);
840 case PERF_RECORD_EXIT:
841 return tool->exit(tool, event, sample, machine);
842 case PERF_RECORD_LOST:
843 if (tool->lost == perf_event__process_lost)
844 session->stats.total_lost += event->lost.lost;
845 return tool->lost(tool, event, sample, machine);
846 case PERF_RECORD_READ:
847 return tool->read(tool, event, sample, evsel, machine);
848 case PERF_RECORD_THROTTLE:
849 return tool->throttle(tool, event, sample, machine);
850 case PERF_RECORD_UNTHROTTLE:
851 return tool->unthrottle(tool, event, sample, machine);
853 ++session->stats.nr_unknown_events;
858 static s64 perf_session__process_user_event(struct perf_session *session,
859 union perf_event *event,
860 struct perf_tool *tool,
863 int fd = perf_data_file__fd(session->file);
866 dump_event(session, event, file_offset, NULL);
868 /* These events are processed right away */
869 switch (event->header.type) {
870 case PERF_RECORD_HEADER_ATTR:
871 err = tool->attr(tool, event, &session->evlist);
873 perf_session__set_id_hdr_size(session);
874 perf_session__set_comm_exec(session);
877 case PERF_RECORD_HEADER_EVENT_TYPE:
879 * Depreceated, but we need to handle it for sake
880 * of old data files create in pipe mode.
883 case PERF_RECORD_HEADER_TRACING_DATA:
884 /* setup for reading amidst mmap */
885 lseek(fd, file_offset, SEEK_SET);
886 return tool->tracing_data(tool, event, session);
887 case PERF_RECORD_HEADER_BUILD_ID:
888 return tool->build_id(tool, event, session);
889 case PERF_RECORD_FINISHED_ROUND:
890 return tool->finished_round(tool, event, session);
896 static void event_swap(union perf_event *event, bool sample_id_all)
898 perf_event__swap_op swap;
900 swap = perf_event__swap_ops[event->header.type];
902 swap(event, sample_id_all);
905 int perf_session__peek_event(struct perf_session *session, off_t file_offset,
906 void *buf, size_t buf_sz,
907 union perf_event **event_ptr,
908 struct perf_sample *sample)
910 union perf_event *event;
914 if (session->one_mmap && !session->header.needs_swap) {
915 event = file_offset - session->one_mmap_offset +
916 session->one_mmap_addr;
917 goto out_parse_sample;
920 if (perf_data_file__is_pipe(session->file))
923 fd = perf_data_file__fd(session->file);
924 hdr_sz = sizeof(struct perf_event_header);
929 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 ||
930 readn(fd, &buf, hdr_sz) != (ssize_t)hdr_sz)
933 event = (union perf_event *)buf;
935 if (session->header.needs_swap)
936 perf_event_header__bswap(&event->header);
938 if (event->header.size < hdr_sz)
941 rest = event->header.size - hdr_sz;
943 if (readn(fd, &buf, rest) != (ssize_t)rest)
946 if (session->header.needs_swap)
947 event_swap(event, perf_evlist__sample_id_all(session->evlist));
951 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START &&
952 perf_evlist__parse_sample(session->evlist, event, sample))
960 static s64 perf_session__process_event(struct perf_session *session,
961 union perf_event *event,
962 struct perf_tool *tool,
965 struct perf_sample sample;
968 if (session->header.needs_swap)
969 event_swap(event, perf_evlist__sample_id_all(session->evlist));
971 if (event->header.type >= PERF_RECORD_HEADER_MAX)
974 events_stats__inc(&session->stats, event->header.type);
976 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
977 return perf_session__process_user_event(session, event, tool, file_offset);
980 * For all kernel events we get the sample data
982 ret = perf_evlist__parse_sample(session->evlist, event, &sample);
986 if (tool->ordered_events) {
987 ret = perf_session_queue_event(session, event, tool, &sample,
993 return perf_session__deliver_event(session, event, &sample, tool,
997 void perf_event_header__bswap(struct perf_event_header *hdr)
999 hdr->type = bswap_32(hdr->type);
1000 hdr->misc = bswap_16(hdr->misc);
1001 hdr->size = bswap_16(hdr->size);
1004 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1006 return machine__findnew_thread(&session->machines.host, -1, pid);
1009 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1011 struct thread *thread;
1013 thread = machine__findnew_thread(&session->machines.host, 0, 0);
1014 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
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 volatile int session_done;
1064 static int __perf_session__process_pipe_events(struct perf_session *session,
1065 struct perf_tool *tool)
1067 int fd = perf_data_file__fd(session->file);
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(fd, event, sizeof(struct perf_event_header));
1091 pr_err("failed to read event header\n");
1095 if (session->header.needs_swap)
1096 perf_event_header__bswap(&event->header);
1098 size = event->header.size;
1099 if (size < sizeof(struct perf_event_header)) {
1100 pr_err("bad event header size\n");
1104 if (size > cur_size) {
1105 void *new = realloc(buf, size);
1107 pr_err("failed to allocate memory to read event\n");
1115 p += sizeof(struct perf_event_header);
1117 if (size - sizeof(struct perf_event_header)) {
1118 err = readn(fd, p, size - sizeof(struct perf_event_header));
1121 pr_err("unexpected end of event stream\n");
1125 pr_err("failed to read event data\n");
1130 if ((skip = perf_session__process_event(session, event, tool, head)) < 0) {
1131 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1132 head, event->header.size, event->header.type);
1142 if (!session_done())
1145 /* do the final flush for ordered samples */
1146 err = ordered_events__flush(session, tool, OE_FLUSH__FINAL);
1149 perf_session__warn_about_errors(session, tool);
1150 ordered_events__free(&session->ordered_events);
1154 static union perf_event *
1155 fetch_mmaped_event(struct perf_session *session,
1156 u64 head, size_t mmap_size, char *buf)
1158 union perf_event *event;
1161 * Ensure we have enough space remaining to read
1162 * the size of the event in the headers.
1164 if (head + sizeof(event->header) > mmap_size)
1167 event = (union perf_event *)(buf + head);
1169 if (session->header.needs_swap)
1170 perf_event_header__bswap(&event->header);
1172 if (head + event->header.size > mmap_size) {
1173 /* We're not fetching the event so swap back again */
1174 if (session->header.needs_swap)
1175 perf_event_header__bswap(&event->header);
1183 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1184 * slices. On 32bit we use 32MB.
1186 #if BITS_PER_LONG == 64
1187 #define MMAP_SIZE ULLONG_MAX
1190 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1191 #define NUM_MMAPS 128
1194 int __perf_session__process_events(struct perf_session *session,
1195 u64 data_offset, u64 data_size,
1196 u64 file_size, struct perf_tool *tool)
1198 int fd = perf_data_file__fd(session->file);
1199 u64 head, page_offset, file_offset, file_pos, size;
1200 int err, mmap_prot, mmap_flags, map_idx = 0;
1202 char *buf, *mmaps[NUM_MMAPS];
1203 union perf_event *event;
1204 struct ui_progress prog;
1207 perf_tool__fill_defaults(tool);
1209 page_offset = page_size * (data_offset / page_size);
1210 file_offset = page_offset;
1211 head = data_offset - page_offset;
1213 if (data_size && (data_offset + data_size < file_size))
1214 file_size = data_offset + data_size;
1216 ui_progress__init(&prog, file_size, "Processing events...");
1218 mmap_size = MMAP_SIZE;
1219 if (mmap_size > file_size) {
1220 mmap_size = file_size;
1221 session->one_mmap = true;
1224 memset(mmaps, 0, sizeof(mmaps));
1226 mmap_prot = PROT_READ;
1227 mmap_flags = MAP_SHARED;
1229 if (session->header.needs_swap) {
1230 mmap_prot |= PROT_WRITE;
1231 mmap_flags = MAP_PRIVATE;
1234 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1236 if (buf == MAP_FAILED) {
1237 pr_err("failed to mmap file\n");
1241 mmaps[map_idx] = buf;
1242 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1243 file_pos = file_offset + head;
1244 if (session->one_mmap) {
1245 session->one_mmap_addr = buf;
1246 session->one_mmap_offset = file_offset;
1250 event = fetch_mmaped_event(session, head, mmap_size, buf);
1252 if (mmaps[map_idx]) {
1253 munmap(mmaps[map_idx], mmap_size);
1254 mmaps[map_idx] = NULL;
1257 page_offset = page_size * (head / page_size);
1258 file_offset += page_offset;
1259 head -= page_offset;
1263 size = event->header.size;
1265 if (size < sizeof(struct perf_event_header) ||
1266 (skip = perf_session__process_event(session, event, tool, file_pos))
1268 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1269 file_offset + head, event->header.size,
1270 event->header.type);
1281 ui_progress__update(&prog, size);
1286 if (file_pos < file_size)
1290 /* do the final flush for ordered samples */
1291 err = ordered_events__flush(session, tool, OE_FLUSH__FINAL);
1293 ui_progress__finish();
1294 perf_session__warn_about_errors(session, tool);
1295 ordered_events__free(&session->ordered_events);
1296 session->one_mmap = false;
1300 int perf_session__process_events(struct perf_session *session,
1301 struct perf_tool *tool)
1303 u64 size = perf_data_file__size(session->file);
1306 if (perf_session__register_idle_thread(session) == NULL)
1309 if (!perf_data_file__is_pipe(session->file))
1310 err = __perf_session__process_events(session,
1311 session->header.data_offset,
1312 session->header.data_size,
1315 err = __perf_session__process_pipe_events(session, tool);
1320 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1322 struct perf_evsel *evsel;
1324 evlist__for_each(session->evlist, evsel) {
1325 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1329 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1333 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1334 const char *symbol_name, u64 addr)
1338 struct ref_reloc_sym *ref;
1340 ref = zalloc(sizeof(struct ref_reloc_sym));
1344 ref->name = strdup(symbol_name);
1345 if (ref->name == NULL) {
1350 bracket = strchr(ref->name, ']');
1356 for (i = 0; i < MAP__NR_TYPES; ++i) {
1357 struct kmap *kmap = map__kmap(maps[i]);
1358 kmap->ref_reloc_sym = ref;
1364 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1366 return machines__fprintf_dsos(&session->machines, fp);
1369 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1370 bool (skip)(struct dso *dso, int parm), int parm)
1372 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1375 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1377 size_t ret = fprintf(fp, "Aggregated stats:\n");
1379 ret += events_stats__fprintf(&session->stats, fp);
1383 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1386 * FIXME: Here we have to actually print all the machines in this
1387 * session, not just the host...
1389 return machine__fprintf(&session->machines.host, fp);
1392 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1395 struct perf_evsel *pos;
1397 evlist__for_each(session->evlist, pos) {
1398 if (pos->attr.type == type)
1404 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1405 struct addr_location *al,
1406 unsigned int print_opts, unsigned int stack_depth)
1408 struct callchain_cursor_node *node;
1409 int print_ip = print_opts & PRINT_IP_OPT_IP;
1410 int print_sym = print_opts & PRINT_IP_OPT_SYM;
1411 int print_dso = print_opts & PRINT_IP_OPT_DSO;
1412 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1413 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1414 int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1415 char s = print_oneline ? ' ' : '\t';
1417 if (symbol_conf.use_callchain && sample->callchain) {
1418 struct addr_location node_al;
1420 if (machine__resolve_callchain(al->machine, evsel, al->thread,
1422 PERF_MAX_STACK_DEPTH) != 0) {
1424 error("Failed to resolve callchain. Skipping\n");
1427 callchain_cursor_commit(&callchain_cursor);
1429 if (print_symoffset)
1432 while (stack_depth) {
1435 node = callchain_cursor_current(&callchain_cursor);
1439 if (node->sym && node->sym->ignore)
1443 printf("%c%16" PRIx64, s, node->ip);
1446 addr = node->map->map_ip(node->map, node->ip);
1450 if (print_symoffset) {
1451 node_al.addr = addr;
1452 node_al.map = node->map;
1453 symbol__fprintf_symname_offs(node->sym, &node_al, stdout);
1455 symbol__fprintf_symname(node->sym, stdout);
1460 map__fprintf_dsoname(node->map, stdout);
1465 map__fprintf_srcline(node->map, addr, "\n ",
1473 callchain_cursor_advance(&callchain_cursor);
1477 if (al->sym && al->sym->ignore)
1481 printf("%16" PRIx64, sample->ip);
1485 if (print_symoffset)
1486 symbol__fprintf_symname_offs(al->sym, al,
1489 symbol__fprintf_symname(al->sym, stdout);
1494 map__fprintf_dsoname(al->map, stdout);
1499 map__fprintf_srcline(al->map, al->addr, "\n ", stdout);
1503 int perf_session__cpu_bitmap(struct perf_session *session,
1504 const char *cpu_list, unsigned long *cpu_bitmap)
1507 struct cpu_map *map;
1509 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1510 struct perf_evsel *evsel;
1512 evsel = perf_session__find_first_evtype(session, i);
1516 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1517 pr_err("File does not contain CPU events. "
1518 "Remove -c option to proceed.\n");
1523 map = cpu_map__new(cpu_list);
1525 pr_err("Invalid cpu_list\n");
1529 for (i = 0; i < map->nr; i++) {
1530 int cpu = map->map[i];
1532 if (cpu >= MAX_NR_CPUS) {
1533 pr_err("Requested CPU %d too large. "
1534 "Consider raising MAX_NR_CPUS\n", cpu);
1535 goto out_delete_map;
1538 set_bit(cpu, cpu_bitmap);
1544 cpu_map__delete(map);
1548 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1554 if (session == NULL || fp == NULL)
1557 fd = perf_data_file__fd(session->file);
1559 ret = fstat(fd, &st);
1563 fprintf(fp, "# ========\n");
1564 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1565 perf_header__fprintf_info(session, fp, full);
1566 fprintf(fp, "# ========\n#\n");
1570 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1571 const struct perf_evsel_str_handler *assocs,
1574 struct perf_evsel *evsel;
1578 for (i = 0; i < nr_assocs; i++) {
1580 * Adding a handler for an event not in the session,
1583 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1588 if (evsel->handler != NULL)
1590 evsel->handler = assocs[i].handler;