1 #include <linux/kernel.h>
2 #include <traceevent/event-parse.h>
16 #include "perf_regs.h"
19 static int perf_session__open(struct perf_session *session)
21 struct perf_data_file *file = session->file;
23 if (perf_session__read_header(session) < 0) {
24 pr_err("incompatible file format (rerun with -v to learn more)");
28 if (perf_data_file__is_pipe(file))
31 if (!perf_evlist__valid_sample_type(session->evlist)) {
32 pr_err("non matching sample_type");
36 if (!perf_evlist__valid_sample_id_all(session->evlist)) {
37 pr_err("non matching sample_id_all");
41 if (!perf_evlist__valid_read_format(session->evlist)) {
42 pr_err("non matching read_format");
49 void perf_session__set_id_hdr_size(struct perf_session *session)
51 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
53 machines__set_id_hdr_size(&session->machines, id_hdr_size);
56 int perf_session__create_kernel_maps(struct perf_session *session)
58 int ret = machine__create_kernel_maps(&session->machines.host);
61 ret = machines__create_guest_kernel_maps(&session->machines);
65 static void perf_session__destroy_kernel_maps(struct perf_session *session)
67 machines__destroy_kernel_maps(&session->machines);
70 struct perf_session *perf_session__new(struct perf_data_file *file,
71 bool repipe, struct perf_tool *tool)
73 struct perf_session *session = zalloc(sizeof(*session));
78 session->repipe = repipe;
79 INIT_LIST_HEAD(&session->ordered_samples.samples);
80 INIT_LIST_HEAD(&session->ordered_samples.sample_cache);
81 INIT_LIST_HEAD(&session->ordered_samples.to_free);
82 machines__init(&session->machines);
85 if (perf_data_file__open(file))
90 if (perf_data_file__is_read(file)) {
91 if (perf_session__open(session) < 0)
94 perf_session__set_id_hdr_size(session);
98 if (!file || perf_data_file__is_write(file)) {
100 * In O_RDONLY mode this will be performed when reading the
101 * kernel MMAP event, in perf_event__process_mmap().
103 if (perf_session__create_kernel_maps(session) < 0)
107 if (tool && tool->ordering_requires_timestamps &&
108 tool->ordered_samples && !perf_evlist__sample_id_all(session->evlist)) {
109 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
110 tool->ordered_samples = false;
116 perf_data_file__close(file);
118 perf_session__delete(session);
123 static void perf_session__delete_dead_threads(struct perf_session *session)
125 machine__delete_dead_threads(&session->machines.host);
128 static void perf_session__delete_threads(struct perf_session *session)
130 machine__delete_threads(&session->machines.host);
133 static void perf_session_env__delete(struct perf_session_env *env)
135 zfree(&env->hostname);
136 zfree(&env->os_release);
137 zfree(&env->version);
139 zfree(&env->cpu_desc);
142 zfree(&env->cmdline);
143 zfree(&env->sibling_cores);
144 zfree(&env->sibling_threads);
145 zfree(&env->numa_nodes);
146 zfree(&env->pmu_mappings);
149 void perf_session__delete(struct perf_session *session)
151 perf_session__destroy_kernel_maps(session);
152 perf_session__delete_dead_threads(session);
153 perf_session__delete_threads(session);
154 perf_session_env__delete(&session->header.env);
155 machines__exit(&session->machines);
157 perf_data_file__close(session->file);
162 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
164 union perf_event *event
166 struct perf_session *session
169 dump_printf(": unhandled!\n");
173 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
174 union perf_event *event __maybe_unused,
175 struct perf_evlist **pevlist
178 dump_printf(": unhandled!\n");
182 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
183 union perf_event *event __maybe_unused,
184 struct perf_sample *sample __maybe_unused,
185 struct perf_evsel *evsel __maybe_unused,
186 struct machine *machine __maybe_unused)
188 dump_printf(": unhandled!\n");
192 static int process_event_stub(struct perf_tool *tool __maybe_unused,
193 union perf_event *event __maybe_unused,
194 struct perf_sample *sample __maybe_unused,
195 struct machine *machine __maybe_unused)
197 dump_printf(": unhandled!\n");
201 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
202 union perf_event *event __maybe_unused,
203 struct perf_session *perf_session
206 dump_printf(": unhandled!\n");
210 static int process_finished_round(struct perf_tool *tool,
211 union perf_event *event,
212 struct perf_session *session);
214 void perf_tool__fill_defaults(struct perf_tool *tool)
216 if (tool->sample == NULL)
217 tool->sample = process_event_sample_stub;
218 if (tool->mmap == NULL)
219 tool->mmap = process_event_stub;
220 if (tool->mmap2 == NULL)
221 tool->mmap2 = process_event_stub;
222 if (tool->comm == NULL)
223 tool->comm = process_event_stub;
224 if (tool->fork == NULL)
225 tool->fork = process_event_stub;
226 if (tool->exit == NULL)
227 tool->exit = process_event_stub;
228 if (tool->lost == NULL)
229 tool->lost = perf_event__process_lost;
230 if (tool->read == NULL)
231 tool->read = process_event_sample_stub;
232 if (tool->throttle == NULL)
233 tool->throttle = process_event_stub;
234 if (tool->unthrottle == NULL)
235 tool->unthrottle = process_event_stub;
236 if (tool->attr == NULL)
237 tool->attr = process_event_synth_attr_stub;
238 if (tool->tracing_data == NULL)
239 tool->tracing_data = process_event_synth_tracing_data_stub;
240 if (tool->build_id == NULL)
241 tool->build_id = process_finished_round_stub;
242 if (tool->finished_round == NULL) {
243 if (tool->ordered_samples)
244 tool->finished_round = process_finished_round;
246 tool->finished_round = process_finished_round_stub;
250 static void swap_sample_id_all(union perf_event *event, void *data)
252 void *end = (void *) event + event->header.size;
253 int size = end - data;
255 BUG_ON(size % sizeof(u64));
256 mem_bswap_64(data, size);
259 static void perf_event__all64_swap(union perf_event *event,
260 bool sample_id_all __maybe_unused)
262 struct perf_event_header *hdr = &event->header;
263 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
266 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
268 event->comm.pid = bswap_32(event->comm.pid);
269 event->comm.tid = bswap_32(event->comm.tid);
272 void *data = &event->comm.comm;
274 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
275 swap_sample_id_all(event, data);
279 static void perf_event__mmap_swap(union perf_event *event,
282 event->mmap.pid = bswap_32(event->mmap.pid);
283 event->mmap.tid = bswap_32(event->mmap.tid);
284 event->mmap.start = bswap_64(event->mmap.start);
285 event->mmap.len = bswap_64(event->mmap.len);
286 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
289 void *data = &event->mmap.filename;
291 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
292 swap_sample_id_all(event, data);
296 static void perf_event__mmap2_swap(union perf_event *event,
299 event->mmap2.pid = bswap_32(event->mmap2.pid);
300 event->mmap2.tid = bswap_32(event->mmap2.tid);
301 event->mmap2.start = bswap_64(event->mmap2.start);
302 event->mmap2.len = bswap_64(event->mmap2.len);
303 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
304 event->mmap2.maj = bswap_32(event->mmap2.maj);
305 event->mmap2.min = bswap_32(event->mmap2.min);
306 event->mmap2.ino = bswap_64(event->mmap2.ino);
309 void *data = &event->mmap2.filename;
311 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
312 swap_sample_id_all(event, data);
315 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
317 event->fork.pid = bswap_32(event->fork.pid);
318 event->fork.tid = bswap_32(event->fork.tid);
319 event->fork.ppid = bswap_32(event->fork.ppid);
320 event->fork.ptid = bswap_32(event->fork.ptid);
321 event->fork.time = bswap_64(event->fork.time);
324 swap_sample_id_all(event, &event->fork + 1);
327 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
329 event->read.pid = bswap_32(event->read.pid);
330 event->read.tid = bswap_32(event->read.tid);
331 event->read.value = bswap_64(event->read.value);
332 event->read.time_enabled = bswap_64(event->read.time_enabled);
333 event->read.time_running = bswap_64(event->read.time_running);
334 event->read.id = bswap_64(event->read.id);
337 swap_sample_id_all(event, &event->read + 1);
340 static void perf_event__throttle_swap(union perf_event *event,
343 event->throttle.time = bswap_64(event->throttle.time);
344 event->throttle.id = bswap_64(event->throttle.id);
345 event->throttle.stream_id = bswap_64(event->throttle.stream_id);
348 swap_sample_id_all(event, &event->throttle + 1);
351 static u8 revbyte(u8 b)
353 int rev = (b >> 4) | ((b & 0xf) << 4);
354 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
355 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
360 * XXX this is hack in attempt to carry flags bitfield
361 * throught endian village. ABI says:
363 * Bit-fields are allocated from right to left (least to most significant)
364 * on little-endian implementations and from left to right (most to least
365 * significant) on big-endian implementations.
367 * The above seems to be byte specific, so we need to reverse each
368 * byte of the bitfield. 'Internet' also says this might be implementation
369 * specific and we probably need proper fix and carry perf_event_attr
370 * bitfield flags in separate data file FEAT_ section. Thought this seems
373 static void swap_bitfield(u8 *p, unsigned len)
377 for (i = 0; i < len; i++) {
383 /* exported for swapping attributes in file header */
384 void perf_event__attr_swap(struct perf_event_attr *attr)
386 attr->type = bswap_32(attr->type);
387 attr->size = bswap_32(attr->size);
388 attr->config = bswap_64(attr->config);
389 attr->sample_period = bswap_64(attr->sample_period);
390 attr->sample_type = bswap_64(attr->sample_type);
391 attr->read_format = bswap_64(attr->read_format);
392 attr->wakeup_events = bswap_32(attr->wakeup_events);
393 attr->bp_type = bswap_32(attr->bp_type);
394 attr->bp_addr = bswap_64(attr->bp_addr);
395 attr->bp_len = bswap_64(attr->bp_len);
396 attr->branch_sample_type = bswap_64(attr->branch_sample_type);
397 attr->sample_regs_user = bswap_64(attr->sample_regs_user);
398 attr->sample_stack_user = bswap_32(attr->sample_stack_user);
400 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
403 static void perf_event__hdr_attr_swap(union perf_event *event,
404 bool sample_id_all __maybe_unused)
408 perf_event__attr_swap(&event->attr.attr);
410 size = event->header.size;
411 size -= (void *)&event->attr.id - (void *)event;
412 mem_bswap_64(event->attr.id, size);
415 static void perf_event__event_type_swap(union perf_event *event,
416 bool sample_id_all __maybe_unused)
418 event->event_type.event_type.event_id =
419 bswap_64(event->event_type.event_type.event_id);
422 static void perf_event__tracing_data_swap(union perf_event *event,
423 bool sample_id_all __maybe_unused)
425 event->tracing_data.size = bswap_32(event->tracing_data.size);
428 typedef void (*perf_event__swap_op)(union perf_event *event,
431 static perf_event__swap_op perf_event__swap_ops[] = {
432 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
433 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap,
434 [PERF_RECORD_COMM] = perf_event__comm_swap,
435 [PERF_RECORD_FORK] = perf_event__task_swap,
436 [PERF_RECORD_EXIT] = perf_event__task_swap,
437 [PERF_RECORD_LOST] = perf_event__all64_swap,
438 [PERF_RECORD_READ] = perf_event__read_swap,
439 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap,
440 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap,
441 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
442 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
443 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
444 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
445 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
446 [PERF_RECORD_HEADER_MAX] = NULL,
449 struct sample_queue {
452 union perf_event *event;
453 struct list_head list;
456 static void perf_session_free_sample_buffers(struct perf_session *session)
458 struct ordered_samples *os = &session->ordered_samples;
460 while (!list_empty(&os->to_free)) {
461 struct sample_queue *sq;
463 sq = list_entry(os->to_free.next, struct sample_queue, list);
469 static int perf_session_deliver_event(struct perf_session *session,
470 union perf_event *event,
471 struct perf_sample *sample,
472 struct perf_tool *tool,
475 static int flush_sample_queue(struct perf_session *s,
476 struct perf_tool *tool)
478 struct ordered_samples *os = &s->ordered_samples;
479 struct list_head *head = &os->samples;
480 struct sample_queue *tmp, *iter;
481 struct perf_sample sample;
482 u64 limit = os->next_flush;
483 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
484 bool show_progress = limit == ULLONG_MAX;
485 struct ui_progress prog;
488 if (!tool->ordered_samples || !limit)
492 ui_progress__init(&prog, os->nr_samples, "Processing time ordered events...");
494 list_for_each_entry_safe(iter, tmp, head, list) {
498 if (iter->timestamp > limit)
501 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
503 pr_err("Can't parse sample, err = %d\n", ret);
505 ret = perf_session_deliver_event(s, iter->event, &sample, tool,
511 os->last_flush = iter->timestamp;
512 list_del(&iter->list);
513 list_add(&iter->list, &os->sample_cache);
516 ui_progress__update(&prog, 1);
519 if (list_empty(head)) {
520 os->last_sample = NULL;
521 } else if (last_ts <= limit) {
523 list_entry(head->prev, struct sample_queue, list);
532 * When perf record finishes a pass on every buffers, it records this pseudo
534 * We record the max timestamp t found in the pass n.
535 * Assuming these timestamps are monotonic across cpus, we know that if
536 * a buffer still has events with timestamps below t, they will be all
537 * available and then read in the pass n + 1.
538 * Hence when we start to read the pass n + 2, we can safely flush every
539 * events with timestamps below t.
541 * ============ PASS n =================
544 * cnt1 timestamps | cnt2 timestamps
547 * - | 4 <--- max recorded
549 * ============ PASS n + 1 ==============
552 * cnt1 timestamps | cnt2 timestamps
555 * 5 | 7 <---- max recorded
557 * Flush every events below timestamp 4
559 * ============ PASS n + 2 ==============
562 * cnt1 timestamps | cnt2 timestamps
567 * Flush every events below timestamp 7
570 static int process_finished_round(struct perf_tool *tool,
571 union perf_event *event __maybe_unused,
572 struct perf_session *session)
574 int ret = flush_sample_queue(session, tool);
576 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
581 /* The queue is ordered by time */
582 static void __queue_event(struct sample_queue *new, struct perf_session *s)
584 struct ordered_samples *os = &s->ordered_samples;
585 struct sample_queue *sample = os->last_sample;
586 u64 timestamp = new->timestamp;
590 os->last_sample = new;
593 list_add(&new->list, &os->samples);
594 os->max_timestamp = timestamp;
599 * last_sample might point to some random place in the list as it's
600 * the last queued event. We expect that the new event is close to
603 if (sample->timestamp <= timestamp) {
604 while (sample->timestamp <= timestamp) {
605 p = sample->list.next;
606 if (p == &os->samples) {
607 list_add_tail(&new->list, &os->samples);
608 os->max_timestamp = timestamp;
611 sample = list_entry(p, struct sample_queue, list);
613 list_add_tail(&new->list, &sample->list);
615 while (sample->timestamp > timestamp) {
616 p = sample->list.prev;
617 if (p == &os->samples) {
618 list_add(&new->list, &os->samples);
621 sample = list_entry(p, struct sample_queue, list);
623 list_add(&new->list, &sample->list);
627 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
629 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
630 struct perf_sample *sample, u64 file_offset)
632 struct ordered_samples *os = &s->ordered_samples;
633 struct list_head *sc = &os->sample_cache;
634 u64 timestamp = sample->time;
635 struct sample_queue *new;
637 if (!timestamp || timestamp == ~0ULL)
640 if (timestamp < s->ordered_samples.last_flush) {
641 printf("Warning: Timestamp below last timeslice flush\n");
645 if (!list_empty(sc)) {
646 new = list_entry(sc->next, struct sample_queue, list);
647 list_del(&new->list);
648 } else if (os->sample_buffer) {
649 new = os->sample_buffer + os->sample_buffer_idx;
650 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
651 os->sample_buffer = NULL;
653 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
654 if (!os->sample_buffer)
656 list_add(&os->sample_buffer->list, &os->to_free);
657 os->sample_buffer_idx = 2;
658 new = os->sample_buffer + 1;
661 new->timestamp = timestamp;
662 new->file_offset = file_offset;
665 __queue_event(new, s);
670 static void callchain__printf(struct perf_sample *sample)
674 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
676 for (i = 0; i < sample->callchain->nr; i++)
677 printf("..... %2d: %016" PRIx64 "\n",
678 i, sample->callchain->ips[i]);
681 static void branch_stack__printf(struct perf_sample *sample)
685 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
687 for (i = 0; i < sample->branch_stack->nr; i++)
688 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
689 i, sample->branch_stack->entries[i].from,
690 sample->branch_stack->entries[i].to);
693 static void regs_dump__printf(u64 mask, u64 *regs)
697 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
700 printf(".... %-5s 0x%" PRIx64 "\n",
701 perf_reg_name(rid), val);
705 static void regs_user__printf(struct perf_sample *sample)
707 struct regs_dump *user_regs = &sample->user_regs;
709 if (user_regs->regs) {
710 u64 mask = user_regs->mask;
711 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
712 regs_dump__printf(mask, user_regs->regs);
716 static void stack_user__printf(struct stack_dump *dump)
718 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
719 dump->size, dump->offset);
722 static void perf_session__print_tstamp(struct perf_session *session,
723 union perf_event *event,
724 struct perf_sample *sample)
726 u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
728 if (event->header.type != PERF_RECORD_SAMPLE &&
729 !perf_evlist__sample_id_all(session->evlist)) {
730 fputs("-1 -1 ", stdout);
734 if ((sample_type & PERF_SAMPLE_CPU))
735 printf("%u ", sample->cpu);
737 if (sample_type & PERF_SAMPLE_TIME)
738 printf("%" PRIu64 " ", sample->time);
741 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
743 printf("... sample_read:\n");
745 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
746 printf("...... time enabled %016" PRIx64 "\n",
747 sample->read.time_enabled);
749 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
750 printf("...... time running %016" PRIx64 "\n",
751 sample->read.time_running);
753 if (read_format & PERF_FORMAT_GROUP) {
756 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
758 for (i = 0; i < sample->read.group.nr; i++) {
759 struct sample_read_value *value;
761 value = &sample->read.group.values[i];
762 printf("..... id %016" PRIx64
763 ", value %016" PRIx64 "\n",
764 value->id, value->value);
767 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
768 sample->read.one.id, sample->read.one.value);
771 static void dump_event(struct perf_session *session, union perf_event *event,
772 u64 file_offset, struct perf_sample *sample)
777 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
778 file_offset, event->header.size, event->header.type);
783 perf_session__print_tstamp(session, event, sample);
785 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
786 event->header.size, perf_event__name(event->header.type));
789 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
790 struct perf_sample *sample)
797 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
798 event->header.misc, sample->pid, sample->tid, sample->ip,
799 sample->period, sample->addr);
801 sample_type = evsel->attr.sample_type;
803 if (sample_type & PERF_SAMPLE_CALLCHAIN)
804 callchain__printf(sample);
806 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
807 branch_stack__printf(sample);
809 if (sample_type & PERF_SAMPLE_REGS_USER)
810 regs_user__printf(sample);
812 if (sample_type & PERF_SAMPLE_STACK_USER)
813 stack_user__printf(&sample->user_stack);
815 if (sample_type & PERF_SAMPLE_WEIGHT)
816 printf("... weight: %" PRIu64 "\n", sample->weight);
818 if (sample_type & PERF_SAMPLE_DATA_SRC)
819 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
821 if (sample_type & PERF_SAMPLE_TRANSACTION)
822 printf("... transaction: %" PRIx64 "\n", sample->transaction);
824 if (sample_type & PERF_SAMPLE_READ)
825 sample_read__printf(sample, evsel->attr.read_format);
828 static struct machine *
829 perf_session__find_machine_for_cpumode(struct perf_session *session,
830 union perf_event *event,
831 struct perf_sample *sample)
833 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
834 struct machine *machine;
837 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
838 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
841 if (event->header.type == PERF_RECORD_MMAP
842 || event->header.type == PERF_RECORD_MMAP2)
843 pid = event->mmap.pid;
847 machine = perf_session__find_machine(session, pid);
849 machine = perf_session__findnew_machine(session,
850 DEFAULT_GUEST_KERNEL_ID);
854 return &session->machines.host;
857 static int deliver_sample_value(struct perf_session *session,
858 struct perf_tool *tool,
859 union perf_event *event,
860 struct perf_sample *sample,
861 struct sample_read_value *v,
862 struct machine *machine)
864 struct perf_sample_id *sid;
866 sid = perf_evlist__id2sid(session->evlist, v->id);
869 sample->period = v->value - sid->period;
870 sid->period = v->value;
873 if (!sid || sid->evsel == NULL) {
874 ++session->stats.nr_unknown_id;
878 return tool->sample(tool, event, sample, sid->evsel, machine);
881 static int deliver_sample_group(struct perf_session *session,
882 struct perf_tool *tool,
883 union perf_event *event,
884 struct perf_sample *sample,
885 struct machine *machine)
890 for (i = 0; i < sample->read.group.nr; i++) {
891 ret = deliver_sample_value(session, tool, event, sample,
892 &sample->read.group.values[i],
902 perf_session__deliver_sample(struct perf_session *session,
903 struct perf_tool *tool,
904 union perf_event *event,
905 struct perf_sample *sample,
906 struct perf_evsel *evsel,
907 struct machine *machine)
909 /* We know evsel != NULL. */
910 u64 sample_type = evsel->attr.sample_type;
911 u64 read_format = evsel->attr.read_format;
913 /* Standard sample delievery. */
914 if (!(sample_type & PERF_SAMPLE_READ))
915 return tool->sample(tool, event, sample, evsel, machine);
917 /* For PERF_SAMPLE_READ we have either single or group mode. */
918 if (read_format & PERF_FORMAT_GROUP)
919 return deliver_sample_group(session, tool, event, sample,
922 return deliver_sample_value(session, tool, event, sample,
923 &sample->read.one, machine);
926 static int perf_session_deliver_event(struct perf_session *session,
927 union perf_event *event,
928 struct perf_sample *sample,
929 struct perf_tool *tool,
932 struct perf_evsel *evsel;
933 struct machine *machine;
935 dump_event(session, event, file_offset, sample);
937 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
938 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
940 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
941 * because the tools right now may apply filters, discarding
942 * some of the samples. For consistency, in the future we
943 * should have something like nr_filtered_samples and remove
944 * the sample->period from total_sample_period, etc, KISS for
947 * Also testing against NULL allows us to handle files without
948 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
949 * future probably it'll be a good idea to restrict event
950 * processing via perf_session to files with both set.
952 hists__inc_nr_events(&evsel->hists, event->header.type);
955 machine = perf_session__find_machine_for_cpumode(session, event,
958 switch (event->header.type) {
959 case PERF_RECORD_SAMPLE:
960 dump_sample(evsel, event, sample);
962 ++session->stats.nr_unknown_id;
965 if (machine == NULL) {
966 ++session->stats.nr_unprocessable_samples;
969 return perf_session__deliver_sample(session, tool, event,
970 sample, evsel, machine);
971 case PERF_RECORD_MMAP:
972 return tool->mmap(tool, event, sample, machine);
973 case PERF_RECORD_MMAP2:
974 return tool->mmap2(tool, event, sample, machine);
975 case PERF_RECORD_COMM:
976 return tool->comm(tool, event, sample, machine);
977 case PERF_RECORD_FORK:
978 return tool->fork(tool, event, sample, machine);
979 case PERF_RECORD_EXIT:
980 return tool->exit(tool, event, sample, machine);
981 case PERF_RECORD_LOST:
982 if (tool->lost == perf_event__process_lost)
983 session->stats.total_lost += event->lost.lost;
984 return tool->lost(tool, event, sample, machine);
985 case PERF_RECORD_READ:
986 return tool->read(tool, event, sample, evsel, machine);
987 case PERF_RECORD_THROTTLE:
988 return tool->throttle(tool, event, sample, machine);
989 case PERF_RECORD_UNTHROTTLE:
990 return tool->unthrottle(tool, event, sample, machine);
992 ++session->stats.nr_unknown_events;
997 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
998 struct perf_tool *tool, u64 file_offset)
1000 int fd = perf_data_file__fd(session->file);
1003 dump_event(session, event, file_offset, NULL);
1005 /* These events are processed right away */
1006 switch (event->header.type) {
1007 case PERF_RECORD_HEADER_ATTR:
1008 err = tool->attr(tool, event, &session->evlist);
1010 perf_session__set_id_hdr_size(session);
1012 case PERF_RECORD_HEADER_EVENT_TYPE:
1014 * Depreceated, but we need to handle it for sake
1015 * of old data files create in pipe mode.
1018 case PERF_RECORD_HEADER_TRACING_DATA:
1019 /* setup for reading amidst mmap */
1020 lseek(fd, file_offset, SEEK_SET);
1021 return tool->tracing_data(tool, event, session);
1022 case PERF_RECORD_HEADER_BUILD_ID:
1023 return tool->build_id(tool, event, session);
1024 case PERF_RECORD_FINISHED_ROUND:
1025 return tool->finished_round(tool, event, session);
1031 static void event_swap(union perf_event *event, bool sample_id_all)
1033 perf_event__swap_op swap;
1035 swap = perf_event__swap_ops[event->header.type];
1037 swap(event, sample_id_all);
1040 static int perf_session__process_event(struct perf_session *session,
1041 union perf_event *event,
1042 struct perf_tool *tool,
1045 struct perf_sample sample;
1048 if (session->header.needs_swap)
1049 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1051 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1054 events_stats__inc(&session->stats, event->header.type);
1056 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1057 return perf_session__process_user_event(session, event, tool, file_offset);
1060 * For all kernel events we get the sample data
1062 ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1066 if (tool->ordered_samples) {
1067 ret = perf_session_queue_event(session, event, &sample,
1073 return perf_session_deliver_event(session, event, &sample, tool,
1077 void perf_event_header__bswap(struct perf_event_header *hdr)
1079 hdr->type = bswap_32(hdr->type);
1080 hdr->misc = bswap_16(hdr->misc);
1081 hdr->size = bswap_16(hdr->size);
1084 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1086 return machine__findnew_thread(&session->machines.host, 0, pid);
1089 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1091 struct thread *thread = perf_session__findnew(session, 0);
1093 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1094 pr_err("problem inserting idle task.\n");
1101 static void perf_session__warn_about_errors(const struct perf_session *session,
1102 const struct perf_tool *tool)
1104 if (tool->lost == perf_event__process_lost &&
1105 session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1106 ui__warning("Processed %d events and lost %d chunks!\n\n"
1107 "Check IO/CPU overload!\n\n",
1108 session->stats.nr_events[0],
1109 session->stats.nr_events[PERF_RECORD_LOST]);
1112 if (session->stats.nr_unknown_events != 0) {
1113 ui__warning("Found %u unknown events!\n\n"
1114 "Is this an older tool processing a perf.data "
1115 "file generated by a more recent tool?\n\n"
1116 "If that is not the case, consider "
1117 "reporting to linux-kernel@vger.kernel.org.\n\n",
1118 session->stats.nr_unknown_events);
1121 if (session->stats.nr_unknown_id != 0) {
1122 ui__warning("%u samples with id not present in the header\n",
1123 session->stats.nr_unknown_id);
1126 if (session->stats.nr_invalid_chains != 0) {
1127 ui__warning("Found invalid callchains!\n\n"
1128 "%u out of %u events were discarded for this reason.\n\n"
1129 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1130 session->stats.nr_invalid_chains,
1131 session->stats.nr_events[PERF_RECORD_SAMPLE]);
1134 if (session->stats.nr_unprocessable_samples != 0) {
1135 ui__warning("%u unprocessable samples recorded.\n"
1136 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1137 session->stats.nr_unprocessable_samples);
1141 volatile int session_done;
1143 static int __perf_session__process_pipe_events(struct perf_session *session,
1144 struct perf_tool *tool)
1146 int fd = perf_data_file__fd(session->file);
1147 union perf_event *event;
1148 uint32_t size, cur_size = 0;
1155 perf_tool__fill_defaults(tool);
1158 cur_size = sizeof(union perf_event);
1160 buf = malloc(cur_size);
1165 err = readn(fd, event, sizeof(struct perf_event_header));
1170 pr_err("failed to read event header\n");
1174 if (session->header.needs_swap)
1175 perf_event_header__bswap(&event->header);
1177 size = event->header.size;
1178 if (size < sizeof(struct perf_event_header)) {
1179 pr_err("bad event header size\n");
1183 if (size > cur_size) {
1184 void *new = realloc(buf, size);
1186 pr_err("failed to allocate memory to read event\n");
1194 p += sizeof(struct perf_event_header);
1196 if (size - sizeof(struct perf_event_header)) {
1197 err = readn(fd, p, size - sizeof(struct perf_event_header));
1200 pr_err("unexpected end of event stream\n");
1204 pr_err("failed to read event data\n");
1209 if ((skip = perf_session__process_event(session, event, tool, head)) < 0) {
1210 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1211 head, event->header.size, event->header.type);
1221 if (!session_done())
1224 /* do the final flush for ordered samples */
1225 session->ordered_samples.next_flush = ULLONG_MAX;
1226 err = flush_sample_queue(session, tool);
1229 perf_session__warn_about_errors(session, tool);
1230 perf_session_free_sample_buffers(session);
1234 static union perf_event *
1235 fetch_mmaped_event(struct perf_session *session,
1236 u64 head, size_t mmap_size, char *buf)
1238 union perf_event *event;
1241 * Ensure we have enough space remaining to read
1242 * the size of the event in the headers.
1244 if (head + sizeof(event->header) > mmap_size)
1247 event = (union perf_event *)(buf + head);
1249 if (session->header.needs_swap)
1250 perf_event_header__bswap(&event->header);
1252 if (head + event->header.size > mmap_size) {
1253 /* We're not fetching the event so swap back again */
1254 if (session->header.needs_swap)
1255 perf_event_header__bswap(&event->header);
1263 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1264 * slices. On 32bit we use 32MB.
1266 #if BITS_PER_LONG == 64
1267 #define MMAP_SIZE ULLONG_MAX
1270 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1271 #define NUM_MMAPS 128
1274 int __perf_session__process_events(struct perf_session *session,
1275 u64 data_offset, u64 data_size,
1276 u64 file_size, struct perf_tool *tool)
1278 int fd = perf_data_file__fd(session->file);
1279 u64 head, page_offset, file_offset, file_pos;
1280 int err, mmap_prot, mmap_flags, map_idx = 0;
1282 char *buf, *mmaps[NUM_MMAPS];
1283 union perf_event *event;
1285 struct ui_progress prog;
1287 perf_tool__fill_defaults(tool);
1289 page_offset = page_size * (data_offset / page_size);
1290 file_offset = page_offset;
1291 head = data_offset - page_offset;
1293 if (data_size && (data_offset + data_size < file_size))
1294 file_size = data_offset + data_size;
1296 ui_progress__init(&prog, file_size, "Processing events...");
1298 mmap_size = MMAP_SIZE;
1299 if (mmap_size > file_size)
1300 mmap_size = file_size;
1302 memset(mmaps, 0, sizeof(mmaps));
1304 mmap_prot = PROT_READ;
1305 mmap_flags = MAP_SHARED;
1307 if (session->header.needs_swap) {
1308 mmap_prot |= PROT_WRITE;
1309 mmap_flags = MAP_PRIVATE;
1312 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1314 if (buf == MAP_FAILED) {
1315 pr_err("failed to mmap file\n");
1319 mmaps[map_idx] = buf;
1320 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1321 file_pos = file_offset + head;
1324 event = fetch_mmaped_event(session, head, mmap_size, buf);
1326 if (mmaps[map_idx]) {
1327 munmap(mmaps[map_idx], mmap_size);
1328 mmaps[map_idx] = NULL;
1331 page_offset = page_size * (head / page_size);
1332 file_offset += page_offset;
1333 head -= page_offset;
1337 size = event->header.size;
1339 if (size < sizeof(struct perf_event_header) ||
1340 perf_session__process_event(session, event, tool, file_pos) < 0) {
1341 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1342 file_offset + head, event->header.size,
1343 event->header.type);
1351 ui_progress__update(&prog, size);
1356 if (file_pos < file_size)
1360 /* do the final flush for ordered samples */
1361 session->ordered_samples.next_flush = ULLONG_MAX;
1362 err = flush_sample_queue(session, tool);
1364 ui_progress__finish();
1365 perf_session__warn_about_errors(session, tool);
1366 perf_session_free_sample_buffers(session);
1370 int perf_session__process_events(struct perf_session *session,
1371 struct perf_tool *tool)
1373 u64 size = perf_data_file__size(session->file);
1376 if (perf_session__register_idle_thread(session) == NULL)
1379 if (!perf_data_file__is_pipe(session->file))
1380 err = __perf_session__process_events(session,
1381 session->header.data_offset,
1382 session->header.data_size,
1385 err = __perf_session__process_pipe_events(session, tool);
1390 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1392 struct perf_evsel *evsel;
1394 evlist__for_each(session->evlist, evsel) {
1395 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1399 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1403 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1404 const char *symbol_name, u64 addr)
1408 struct ref_reloc_sym *ref;
1410 ref = zalloc(sizeof(struct ref_reloc_sym));
1414 ref->name = strdup(symbol_name);
1415 if (ref->name == NULL) {
1420 bracket = strchr(ref->name, ']');
1426 for (i = 0; i < MAP__NR_TYPES; ++i) {
1427 struct kmap *kmap = map__kmap(maps[i]);
1428 kmap->ref_reloc_sym = ref;
1434 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1436 return machines__fprintf_dsos(&session->machines, fp);
1439 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1440 bool (skip)(struct dso *dso, int parm), int parm)
1442 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1445 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1447 struct perf_evsel *pos;
1448 size_t ret = fprintf(fp, "Aggregated stats:\n");
1450 ret += events_stats__fprintf(&session->stats, fp);
1452 evlist__for_each(session->evlist, pos) {
1453 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1454 ret += events_stats__fprintf(&pos->hists.stats, fp);
1460 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1463 * FIXME: Here we have to actually print all the machines in this
1464 * session, not just the host...
1466 return machine__fprintf(&session->machines.host, fp);
1469 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1472 struct perf_evsel *pos;
1474 evlist__for_each(session->evlist, pos) {
1475 if (pos->attr.type == type)
1481 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1482 struct addr_location *al,
1483 unsigned int print_opts, unsigned int stack_depth)
1485 struct callchain_cursor_node *node;
1486 int print_ip = print_opts & PRINT_IP_OPT_IP;
1487 int print_sym = print_opts & PRINT_IP_OPT_SYM;
1488 int print_dso = print_opts & PRINT_IP_OPT_DSO;
1489 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1490 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1491 int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1492 char s = print_oneline ? ' ' : '\t';
1494 if (symbol_conf.use_callchain && sample->callchain) {
1495 struct addr_location node_al;
1497 if (machine__resolve_callchain(al->machine, evsel, al->thread,
1499 PERF_MAX_STACK_DEPTH) != 0) {
1501 error("Failed to resolve callchain. Skipping\n");
1504 callchain_cursor_commit(&callchain_cursor);
1506 if (print_symoffset)
1509 while (stack_depth) {
1512 node = callchain_cursor_current(&callchain_cursor);
1516 if (node->sym && node->sym->ignore)
1520 printf("%c%16" PRIx64, s, node->ip);
1523 addr = node->map->map_ip(node->map, node->ip);
1527 if (print_symoffset) {
1528 node_al.addr = addr;
1529 node_al.map = node->map;
1530 symbol__fprintf_symname_offs(node->sym, &node_al, stdout);
1532 symbol__fprintf_symname(node->sym, stdout);
1537 map__fprintf_dsoname(node->map, stdout);
1542 map__fprintf_srcline(node->map, addr, "\n ",
1550 callchain_cursor_advance(&callchain_cursor);
1554 if (al->sym && al->sym->ignore)
1558 printf("%16" PRIx64, sample->ip);
1562 if (print_symoffset)
1563 symbol__fprintf_symname_offs(al->sym, al,
1566 symbol__fprintf_symname(al->sym, stdout);
1571 map__fprintf_dsoname(al->map, stdout);
1576 map__fprintf_srcline(al->map, al->addr, "\n ", stdout);
1580 int perf_session__cpu_bitmap(struct perf_session *session,
1581 const char *cpu_list, unsigned long *cpu_bitmap)
1584 struct cpu_map *map;
1586 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1587 struct perf_evsel *evsel;
1589 evsel = perf_session__find_first_evtype(session, i);
1593 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1594 pr_err("File does not contain CPU events. "
1595 "Remove -c option to proceed.\n");
1600 map = cpu_map__new(cpu_list);
1602 pr_err("Invalid cpu_list\n");
1606 for (i = 0; i < map->nr; i++) {
1607 int cpu = map->map[i];
1609 if (cpu >= MAX_NR_CPUS) {
1610 pr_err("Requested CPU %d too large. "
1611 "Consider raising MAX_NR_CPUS\n", cpu);
1612 goto out_delete_map;
1615 set_bit(cpu, cpu_bitmap);
1621 cpu_map__delete(map);
1625 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1631 if (session == NULL || fp == NULL)
1634 fd = perf_data_file__fd(session->file);
1636 ret = fstat(fd, &st);
1640 fprintf(fp, "# ========\n");
1641 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1642 perf_header__fprintf_info(session, fp, full);
1643 fprintf(fp, "# ========\n#\n");
1647 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1648 const struct perf_evsel_str_handler *assocs,
1651 struct perf_evsel *evsel;
1655 for (i = 0; i < nr_assocs; i++) {
1657 * Adding a handler for an event not in the session,
1660 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1665 if (evsel->handler != NULL)
1667 evsel->handler = assocs[i].handler;