1 #include <linux/kernel.h>
2 #include <traceevent/event-parse.h>
16 #include "perf_regs.h"
18 static int perf_session__open(struct perf_session *session)
20 struct perf_data_file *file = session->file;
22 if (perf_session__read_header(session) < 0) {
23 pr_err("incompatible file format (rerun with -v to learn more)");
27 if (perf_data_file__is_pipe(file))
30 if (!perf_evlist__valid_sample_type(session->evlist)) {
31 pr_err("non matching sample_type");
35 if (!perf_evlist__valid_sample_id_all(session->evlist)) {
36 pr_err("non matching sample_id_all");
40 if (!perf_evlist__valid_read_format(session->evlist)) {
41 pr_err("non matching read_format");
48 void perf_session__set_id_hdr_size(struct perf_session *session)
50 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
52 machines__set_id_hdr_size(&session->machines, id_hdr_size);
55 int perf_session__create_kernel_maps(struct perf_session *session)
57 int ret = machine__create_kernel_maps(&session->machines.host);
60 ret = machines__create_guest_kernel_maps(&session->machines);
64 static void perf_session__destroy_kernel_maps(struct perf_session *session)
66 machines__destroy_kernel_maps(&session->machines);
69 struct perf_session *perf_session__new(struct perf_data_file *file,
70 bool repipe, struct perf_tool *tool)
72 struct perf_session *session = zalloc(sizeof(*session));
77 session->repipe = repipe;
78 INIT_LIST_HEAD(&session->ordered_samples.samples);
79 INIT_LIST_HEAD(&session->ordered_samples.sample_cache);
80 INIT_LIST_HEAD(&session->ordered_samples.to_free);
81 machines__init(&session->machines);
84 if (perf_data_file__open(file))
89 if (perf_data_file__is_read(file)) {
90 if (perf_session__open(session) < 0)
93 perf_session__set_id_hdr_size(session);
97 if (!file || perf_data_file__is_write(file)) {
99 * In O_RDONLY mode this will be performed when reading the
100 * kernel MMAP event, in perf_event__process_mmap().
102 if (perf_session__create_kernel_maps(session) < 0)
106 if (tool && tool->ordering_requires_timestamps &&
107 tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) {
108 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
109 tool->ordered_events = false;
115 perf_data_file__close(file);
117 perf_session__delete(session);
122 static void perf_session__delete_dead_threads(struct perf_session *session)
124 machine__delete_dead_threads(&session->machines.host);
127 static void perf_session__delete_threads(struct perf_session *session)
129 machine__delete_threads(&session->machines.host);
132 static void perf_session_env__delete(struct perf_session_env *env)
134 zfree(&env->hostname);
135 zfree(&env->os_release);
136 zfree(&env->version);
138 zfree(&env->cpu_desc);
141 zfree(&env->cmdline);
142 zfree(&env->sibling_cores);
143 zfree(&env->sibling_threads);
144 zfree(&env->numa_nodes);
145 zfree(&env->pmu_mappings);
148 void perf_session__delete(struct perf_session *session)
150 perf_session__destroy_kernel_maps(session);
151 perf_session__delete_dead_threads(session);
152 perf_session__delete_threads(session);
153 perf_session_env__delete(&session->header.env);
154 machines__exit(&session->machines);
156 perf_data_file__close(session->file);
160 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
162 union perf_event *event
164 struct perf_session *session
167 dump_printf(": unhandled!\n");
171 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
172 union perf_event *event __maybe_unused,
173 struct perf_evlist **pevlist
176 dump_printf(": unhandled!\n");
180 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
181 union perf_event *event __maybe_unused,
182 struct perf_sample *sample __maybe_unused,
183 struct perf_evsel *evsel __maybe_unused,
184 struct machine *machine __maybe_unused)
186 dump_printf(": unhandled!\n");
190 static int process_event_stub(struct perf_tool *tool __maybe_unused,
191 union perf_event *event __maybe_unused,
192 struct perf_sample *sample __maybe_unused,
193 struct machine *machine __maybe_unused)
195 dump_printf(": unhandled!\n");
199 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
200 union perf_event *event __maybe_unused,
201 struct perf_session *perf_session
204 dump_printf(": unhandled!\n");
208 static int process_finished_round(struct perf_tool *tool,
209 union perf_event *event,
210 struct perf_session *session);
212 void perf_tool__fill_defaults(struct perf_tool *tool)
214 if (tool->sample == NULL)
215 tool->sample = process_event_sample_stub;
216 if (tool->mmap == NULL)
217 tool->mmap = process_event_stub;
218 if (tool->mmap2 == NULL)
219 tool->mmap2 = process_event_stub;
220 if (tool->comm == NULL)
221 tool->comm = process_event_stub;
222 if (tool->fork == NULL)
223 tool->fork = process_event_stub;
224 if (tool->exit == NULL)
225 tool->exit = process_event_stub;
226 if (tool->lost == NULL)
227 tool->lost = perf_event__process_lost;
228 if (tool->read == NULL)
229 tool->read = process_event_sample_stub;
230 if (tool->throttle == NULL)
231 tool->throttle = process_event_stub;
232 if (tool->unthrottle == NULL)
233 tool->unthrottle = process_event_stub;
234 if (tool->attr == NULL)
235 tool->attr = process_event_synth_attr_stub;
236 if (tool->tracing_data == NULL)
237 tool->tracing_data = process_event_synth_tracing_data_stub;
238 if (tool->build_id == NULL)
239 tool->build_id = process_finished_round_stub;
240 if (tool->finished_round == NULL) {
241 if (tool->ordered_events)
242 tool->finished_round = process_finished_round;
244 tool->finished_round = process_finished_round_stub;
248 static void swap_sample_id_all(union perf_event *event, void *data)
250 void *end = (void *) event + event->header.size;
251 int size = end - data;
253 BUG_ON(size % sizeof(u64));
254 mem_bswap_64(data, size);
257 static void perf_event__all64_swap(union perf_event *event,
258 bool sample_id_all __maybe_unused)
260 struct perf_event_header *hdr = &event->header;
261 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
264 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
266 event->comm.pid = bswap_32(event->comm.pid);
267 event->comm.tid = bswap_32(event->comm.tid);
270 void *data = &event->comm.comm;
272 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
273 swap_sample_id_all(event, data);
277 static void perf_event__mmap_swap(union perf_event *event,
280 event->mmap.pid = bswap_32(event->mmap.pid);
281 event->mmap.tid = bswap_32(event->mmap.tid);
282 event->mmap.start = bswap_64(event->mmap.start);
283 event->mmap.len = bswap_64(event->mmap.len);
284 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
287 void *data = &event->mmap.filename;
289 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
290 swap_sample_id_all(event, data);
294 static void perf_event__mmap2_swap(union perf_event *event,
297 event->mmap2.pid = bswap_32(event->mmap2.pid);
298 event->mmap2.tid = bswap_32(event->mmap2.tid);
299 event->mmap2.start = bswap_64(event->mmap2.start);
300 event->mmap2.len = bswap_64(event->mmap2.len);
301 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
302 event->mmap2.maj = bswap_32(event->mmap2.maj);
303 event->mmap2.min = bswap_32(event->mmap2.min);
304 event->mmap2.ino = bswap_64(event->mmap2.ino);
307 void *data = &event->mmap2.filename;
309 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
310 swap_sample_id_all(event, data);
313 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
315 event->fork.pid = bswap_32(event->fork.pid);
316 event->fork.tid = bswap_32(event->fork.tid);
317 event->fork.ppid = bswap_32(event->fork.ppid);
318 event->fork.ptid = bswap_32(event->fork.ptid);
319 event->fork.time = bswap_64(event->fork.time);
322 swap_sample_id_all(event, &event->fork + 1);
325 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
327 event->read.pid = bswap_32(event->read.pid);
328 event->read.tid = bswap_32(event->read.tid);
329 event->read.value = bswap_64(event->read.value);
330 event->read.time_enabled = bswap_64(event->read.time_enabled);
331 event->read.time_running = bswap_64(event->read.time_running);
332 event->read.id = bswap_64(event->read.id);
335 swap_sample_id_all(event, &event->read + 1);
338 static void perf_event__throttle_swap(union perf_event *event,
341 event->throttle.time = bswap_64(event->throttle.time);
342 event->throttle.id = bswap_64(event->throttle.id);
343 event->throttle.stream_id = bswap_64(event->throttle.stream_id);
346 swap_sample_id_all(event, &event->throttle + 1);
349 static u8 revbyte(u8 b)
351 int rev = (b >> 4) | ((b & 0xf) << 4);
352 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
353 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
358 * XXX this is hack in attempt to carry flags bitfield
359 * throught endian village. ABI says:
361 * Bit-fields are allocated from right to left (least to most significant)
362 * on little-endian implementations and from left to right (most to least
363 * significant) on big-endian implementations.
365 * The above seems to be byte specific, so we need to reverse each
366 * byte of the bitfield. 'Internet' also says this might be implementation
367 * specific and we probably need proper fix and carry perf_event_attr
368 * bitfield flags in separate data file FEAT_ section. Thought this seems
371 static void swap_bitfield(u8 *p, unsigned len)
375 for (i = 0; i < len; i++) {
381 /* exported for swapping attributes in file header */
382 void perf_event__attr_swap(struct perf_event_attr *attr)
384 attr->type = bswap_32(attr->type);
385 attr->size = bswap_32(attr->size);
386 attr->config = bswap_64(attr->config);
387 attr->sample_period = bswap_64(attr->sample_period);
388 attr->sample_type = bswap_64(attr->sample_type);
389 attr->read_format = bswap_64(attr->read_format);
390 attr->wakeup_events = bswap_32(attr->wakeup_events);
391 attr->bp_type = bswap_32(attr->bp_type);
392 attr->bp_addr = bswap_64(attr->bp_addr);
393 attr->bp_len = bswap_64(attr->bp_len);
394 attr->branch_sample_type = bswap_64(attr->branch_sample_type);
395 attr->sample_regs_user = bswap_64(attr->sample_regs_user);
396 attr->sample_stack_user = bswap_32(attr->sample_stack_user);
398 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
401 static void perf_event__hdr_attr_swap(union perf_event *event,
402 bool sample_id_all __maybe_unused)
406 perf_event__attr_swap(&event->attr.attr);
408 size = event->header.size;
409 size -= (void *)&event->attr.id - (void *)event;
410 mem_bswap_64(event->attr.id, size);
413 static void perf_event__event_type_swap(union perf_event *event,
414 bool sample_id_all __maybe_unused)
416 event->event_type.event_type.event_id =
417 bswap_64(event->event_type.event_type.event_id);
420 static void perf_event__tracing_data_swap(union perf_event *event,
421 bool sample_id_all __maybe_unused)
423 event->tracing_data.size = bswap_32(event->tracing_data.size);
426 typedef void (*perf_event__swap_op)(union perf_event *event,
429 static perf_event__swap_op perf_event__swap_ops[] = {
430 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
431 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap,
432 [PERF_RECORD_COMM] = perf_event__comm_swap,
433 [PERF_RECORD_FORK] = perf_event__task_swap,
434 [PERF_RECORD_EXIT] = perf_event__task_swap,
435 [PERF_RECORD_LOST] = perf_event__all64_swap,
436 [PERF_RECORD_READ] = perf_event__read_swap,
437 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap,
438 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap,
439 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
440 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
441 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
442 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
443 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
444 [PERF_RECORD_HEADER_MAX] = NULL,
447 struct sample_queue {
450 union perf_event *event;
451 struct list_head list;
454 static void perf_session_free_sample_buffers(struct perf_session *session)
456 struct ordered_samples *os = &session->ordered_samples;
458 while (!list_empty(&os->to_free)) {
459 struct sample_queue *sq;
461 sq = list_entry(os->to_free.next, struct sample_queue, list);
467 static int perf_session_deliver_event(struct perf_session *session,
468 union perf_event *event,
469 struct perf_sample *sample,
470 struct perf_tool *tool,
473 static int flush_sample_queue(struct perf_session *s,
474 struct perf_tool *tool)
476 struct ordered_samples *os = &s->ordered_samples;
477 struct list_head *head = &os->samples;
478 struct sample_queue *tmp, *iter;
479 struct perf_sample sample;
480 u64 limit = os->next_flush;
481 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
482 bool show_progress = limit == ULLONG_MAX;
483 struct ui_progress prog;
486 if (!tool->ordered_events || !limit)
490 ui_progress__init(&prog, os->nr_samples, "Processing time ordered events...");
492 list_for_each_entry_safe(iter, tmp, head, list) {
496 if (iter->timestamp > limit)
499 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
501 pr_err("Can't parse sample, err = %d\n", ret);
503 ret = perf_session_deliver_event(s, iter->event, &sample, tool,
509 os->last_flush = iter->timestamp;
510 list_del(&iter->list);
511 list_add(&iter->list, &os->sample_cache);
515 ui_progress__update(&prog, 1);
518 if (list_empty(head)) {
519 os->last_sample = NULL;
520 } else if (last_ts <= limit) {
522 list_entry(head->prev, struct sample_queue, list);
529 * When perf record finishes a pass on every buffers, it records this pseudo
531 * We record the max timestamp t found in the pass n.
532 * Assuming these timestamps are monotonic across cpus, we know that if
533 * a buffer still has events with timestamps below t, they will be all
534 * available and then read in the pass n + 1.
535 * Hence when we start to read the pass n + 2, we can safely flush every
536 * events with timestamps below t.
538 * ============ PASS n =================
541 * cnt1 timestamps | cnt2 timestamps
544 * - | 4 <--- max recorded
546 * ============ PASS n + 1 ==============
549 * cnt1 timestamps | cnt2 timestamps
552 * 5 | 7 <---- max recorded
554 * Flush every events below timestamp 4
556 * ============ PASS n + 2 ==============
559 * cnt1 timestamps | cnt2 timestamps
564 * Flush every events below timestamp 7
567 static int process_finished_round(struct perf_tool *tool,
568 union perf_event *event __maybe_unused,
569 struct perf_session *session)
571 int ret = flush_sample_queue(session, tool);
573 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
578 /* The queue is ordered by time */
579 static void __queue_event(struct sample_queue *new, struct perf_session *s)
581 struct ordered_samples *os = &s->ordered_samples;
582 struct sample_queue *sample = os->last_sample;
583 u64 timestamp = new->timestamp;
587 os->last_sample = new;
590 list_add(&new->list, &os->samples);
591 os->max_timestamp = timestamp;
596 * last_sample might point to some random place in the list as it's
597 * the last queued event. We expect that the new event is close to
600 if (sample->timestamp <= timestamp) {
601 while (sample->timestamp <= timestamp) {
602 p = sample->list.next;
603 if (p == &os->samples) {
604 list_add_tail(&new->list, &os->samples);
605 os->max_timestamp = timestamp;
608 sample = list_entry(p, struct sample_queue, list);
610 list_add_tail(&new->list, &sample->list);
612 while (sample->timestamp > timestamp) {
613 p = sample->list.prev;
614 if (p == &os->samples) {
615 list_add(&new->list, &os->samples);
618 sample = list_entry(p, struct sample_queue, list);
620 list_add(&new->list, &sample->list);
624 #define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
626 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
627 struct perf_sample *sample, u64 file_offset)
629 struct ordered_samples *os = &s->ordered_samples;
630 struct list_head *sc = &os->sample_cache;
631 u64 timestamp = sample->time;
632 struct sample_queue *new;
634 if (!timestamp || timestamp == ~0ULL)
637 if (timestamp < s->ordered_samples.last_flush) {
638 printf("Warning: Timestamp below last timeslice flush\n");
642 if (!list_empty(sc)) {
643 new = list_entry(sc->next, struct sample_queue, list);
644 list_del(&new->list);
645 } else if (os->sample_buffer) {
646 new = os->sample_buffer + os->sample_buffer_idx;
647 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
648 os->sample_buffer = NULL;
650 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
651 if (!os->sample_buffer)
653 list_add(&os->sample_buffer->list, &os->to_free);
654 os->sample_buffer_idx = 2;
655 new = os->sample_buffer + 1;
658 new->timestamp = timestamp;
659 new->file_offset = file_offset;
662 __queue_event(new, s);
667 static void callchain__printf(struct perf_sample *sample)
671 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
673 for (i = 0; i < sample->callchain->nr; i++)
674 printf("..... %2d: %016" PRIx64 "\n",
675 i, sample->callchain->ips[i]);
678 static void branch_stack__printf(struct perf_sample *sample)
682 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
684 for (i = 0; i < sample->branch_stack->nr; i++)
685 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
686 i, sample->branch_stack->entries[i].from,
687 sample->branch_stack->entries[i].to);
690 static void regs_dump__printf(u64 mask, u64 *regs)
694 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
697 printf(".... %-5s 0x%" PRIx64 "\n",
698 perf_reg_name(rid), val);
702 static void regs_user__printf(struct perf_sample *sample)
704 struct regs_dump *user_regs = &sample->user_regs;
706 if (user_regs->regs) {
707 u64 mask = user_regs->mask;
708 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
709 regs_dump__printf(mask, user_regs->regs);
713 static void stack_user__printf(struct stack_dump *dump)
715 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
716 dump->size, dump->offset);
719 static void perf_session__print_tstamp(struct perf_session *session,
720 union perf_event *event,
721 struct perf_sample *sample)
723 u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
725 if (event->header.type != PERF_RECORD_SAMPLE &&
726 !perf_evlist__sample_id_all(session->evlist)) {
727 fputs("-1 -1 ", stdout);
731 if ((sample_type & PERF_SAMPLE_CPU))
732 printf("%u ", sample->cpu);
734 if (sample_type & PERF_SAMPLE_TIME)
735 printf("%" PRIu64 " ", sample->time);
738 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
740 printf("... sample_read:\n");
742 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
743 printf("...... time enabled %016" PRIx64 "\n",
744 sample->read.time_enabled);
746 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
747 printf("...... time running %016" PRIx64 "\n",
748 sample->read.time_running);
750 if (read_format & PERF_FORMAT_GROUP) {
753 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
755 for (i = 0; i < sample->read.group.nr; i++) {
756 struct sample_read_value *value;
758 value = &sample->read.group.values[i];
759 printf("..... id %016" PRIx64
760 ", value %016" PRIx64 "\n",
761 value->id, value->value);
764 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
765 sample->read.one.id, sample->read.one.value);
768 static void dump_event(struct perf_session *session, union perf_event *event,
769 u64 file_offset, struct perf_sample *sample)
774 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
775 file_offset, event->header.size, event->header.type);
780 perf_session__print_tstamp(session, event, sample);
782 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
783 event->header.size, perf_event__name(event->header.type));
786 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
787 struct perf_sample *sample)
794 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
795 event->header.misc, sample->pid, sample->tid, sample->ip,
796 sample->period, sample->addr);
798 sample_type = evsel->attr.sample_type;
800 if (sample_type & PERF_SAMPLE_CALLCHAIN)
801 callchain__printf(sample);
803 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
804 branch_stack__printf(sample);
806 if (sample_type & PERF_SAMPLE_REGS_USER)
807 regs_user__printf(sample);
809 if (sample_type & PERF_SAMPLE_STACK_USER)
810 stack_user__printf(&sample->user_stack);
812 if (sample_type & PERF_SAMPLE_WEIGHT)
813 printf("... weight: %" PRIu64 "\n", sample->weight);
815 if (sample_type & PERF_SAMPLE_DATA_SRC)
816 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
818 if (sample_type & PERF_SAMPLE_TRANSACTION)
819 printf("... transaction: %" PRIx64 "\n", sample->transaction);
821 if (sample_type & PERF_SAMPLE_READ)
822 sample_read__printf(sample, evsel->attr.read_format);
825 static struct machine *
826 perf_session__find_machine_for_cpumode(struct perf_session *session,
827 union perf_event *event,
828 struct perf_sample *sample)
830 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
831 struct machine *machine;
834 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
835 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
838 if (event->header.type == PERF_RECORD_MMAP
839 || event->header.type == PERF_RECORD_MMAP2)
840 pid = event->mmap.pid;
844 machine = perf_session__find_machine(session, pid);
846 machine = perf_session__findnew_machine(session,
847 DEFAULT_GUEST_KERNEL_ID);
851 return &session->machines.host;
854 static int deliver_sample_value(struct perf_session *session,
855 struct perf_tool *tool,
856 union perf_event *event,
857 struct perf_sample *sample,
858 struct sample_read_value *v,
859 struct machine *machine)
861 struct perf_sample_id *sid;
863 sid = perf_evlist__id2sid(session->evlist, v->id);
866 sample->period = v->value - sid->period;
867 sid->period = v->value;
870 if (!sid || sid->evsel == NULL) {
871 ++session->stats.nr_unknown_id;
875 return tool->sample(tool, event, sample, sid->evsel, machine);
878 static int deliver_sample_group(struct perf_session *session,
879 struct perf_tool *tool,
880 union perf_event *event,
881 struct perf_sample *sample,
882 struct machine *machine)
887 for (i = 0; i < sample->read.group.nr; i++) {
888 ret = deliver_sample_value(session, tool, event, sample,
889 &sample->read.group.values[i],
899 perf_session__deliver_sample(struct perf_session *session,
900 struct perf_tool *tool,
901 union perf_event *event,
902 struct perf_sample *sample,
903 struct perf_evsel *evsel,
904 struct machine *machine)
906 /* We know evsel != NULL. */
907 u64 sample_type = evsel->attr.sample_type;
908 u64 read_format = evsel->attr.read_format;
910 /* Standard sample delievery. */
911 if (!(sample_type & PERF_SAMPLE_READ))
912 return tool->sample(tool, event, sample, evsel, machine);
914 /* For PERF_SAMPLE_READ we have either single or group mode. */
915 if (read_format & PERF_FORMAT_GROUP)
916 return deliver_sample_group(session, tool, event, sample,
919 return deliver_sample_value(session, tool, event, sample,
920 &sample->read.one, machine);
923 static int perf_session_deliver_event(struct perf_session *session,
924 union perf_event *event,
925 struct perf_sample *sample,
926 struct perf_tool *tool,
929 struct perf_evsel *evsel;
930 struct machine *machine;
932 dump_event(session, event, file_offset, sample);
934 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
935 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
937 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
938 * because the tools right now may apply filters, discarding
939 * some of the samples. For consistency, in the future we
940 * should have something like nr_filtered_samples and remove
941 * the sample->period from total_sample_period, etc, KISS for
944 * Also testing against NULL allows us to handle files without
945 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
946 * future probably it'll be a good idea to restrict event
947 * processing via perf_session to files with both set.
949 hists__inc_nr_events(&evsel->hists, event->header.type);
952 machine = perf_session__find_machine_for_cpumode(session, event,
955 switch (event->header.type) {
956 case PERF_RECORD_SAMPLE:
957 dump_sample(evsel, event, sample);
959 ++session->stats.nr_unknown_id;
962 if (machine == NULL) {
963 ++session->stats.nr_unprocessable_samples;
966 return perf_session__deliver_sample(session, tool, event,
967 sample, evsel, machine);
968 case PERF_RECORD_MMAP:
969 return tool->mmap(tool, event, sample, machine);
970 case PERF_RECORD_MMAP2:
971 return tool->mmap2(tool, event, sample, machine);
972 case PERF_RECORD_COMM:
973 return tool->comm(tool, event, sample, machine);
974 case PERF_RECORD_FORK:
975 return tool->fork(tool, event, sample, machine);
976 case PERF_RECORD_EXIT:
977 return tool->exit(tool, event, sample, machine);
978 case PERF_RECORD_LOST:
979 if (tool->lost == perf_event__process_lost)
980 session->stats.total_lost += event->lost.lost;
981 return tool->lost(tool, event, sample, machine);
982 case PERF_RECORD_READ:
983 return tool->read(tool, event, sample, evsel, machine);
984 case PERF_RECORD_THROTTLE:
985 return tool->throttle(tool, event, sample, machine);
986 case PERF_RECORD_UNTHROTTLE:
987 return tool->unthrottle(tool, event, sample, machine);
989 ++session->stats.nr_unknown_events;
994 static s64 perf_session__process_user_event(struct perf_session *session,
995 union perf_event *event,
996 struct perf_tool *tool,
999 int fd = perf_data_file__fd(session->file);
1002 dump_event(session, event, file_offset, NULL);
1004 /* These events are processed right away */
1005 switch (event->header.type) {
1006 case PERF_RECORD_HEADER_ATTR:
1007 err = tool->attr(tool, event, &session->evlist);
1009 perf_session__set_id_hdr_size(session);
1011 case PERF_RECORD_HEADER_EVENT_TYPE:
1013 * Depreceated, but we need to handle it for sake
1014 * of old data files create in pipe mode.
1017 case PERF_RECORD_HEADER_TRACING_DATA:
1018 /* setup for reading amidst mmap */
1019 lseek(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 s64 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_events) {
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 *hdr)
1078 hdr->type = bswap_32(hdr->type);
1079 hdr->misc = bswap_16(hdr->misc);
1080 hdr->size = bswap_16(hdr->size);
1083 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1085 return machine__findnew_thread(&session->machines.host, -1, pid);
1088 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1090 struct thread *thread;
1092 thread = machine__findnew_thread(&session->machines.host, 0, 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, size;
1280 int err, mmap_prot, mmap_flags, map_idx = 0;
1282 char *buf, *mmaps[NUM_MMAPS];
1283 union perf_event *event;
1284 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;
1301 session->one_mmap = true;
1304 memset(mmaps, 0, sizeof(mmaps));
1306 mmap_prot = PROT_READ;
1307 mmap_flags = MAP_SHARED;
1309 if (session->header.needs_swap) {
1310 mmap_prot |= PROT_WRITE;
1311 mmap_flags = MAP_PRIVATE;
1314 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1316 if (buf == MAP_FAILED) {
1317 pr_err("failed to mmap file\n");
1321 mmaps[map_idx] = buf;
1322 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1323 file_pos = file_offset + head;
1324 if (session->one_mmap) {
1325 session->one_mmap_addr = buf;
1326 session->one_mmap_offset = file_offset;
1330 event = fetch_mmaped_event(session, head, mmap_size, buf);
1332 if (mmaps[map_idx]) {
1333 munmap(mmaps[map_idx], mmap_size);
1334 mmaps[map_idx] = NULL;
1337 page_offset = page_size * (head / page_size);
1338 file_offset += page_offset;
1339 head -= page_offset;
1343 size = event->header.size;
1345 if (size < sizeof(struct perf_event_header) ||
1346 (skip = perf_session__process_event(session, event, tool, file_pos))
1348 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1349 file_offset + head, event->header.size,
1350 event->header.type);
1361 ui_progress__update(&prog, size);
1366 if (file_pos < file_size)
1370 /* do the final flush for ordered samples */
1371 session->ordered_samples.next_flush = ULLONG_MAX;
1372 err = flush_sample_queue(session, tool);
1374 ui_progress__finish();
1375 perf_session__warn_about_errors(session, tool);
1376 perf_session_free_sample_buffers(session);
1377 session->one_mmap = false;
1381 int perf_session__process_events(struct perf_session *session,
1382 struct perf_tool *tool)
1384 u64 size = perf_data_file__size(session->file);
1387 if (perf_session__register_idle_thread(session) == NULL)
1390 if (!perf_data_file__is_pipe(session->file))
1391 err = __perf_session__process_events(session,
1392 session->header.data_offset,
1393 session->header.data_size,
1396 err = __perf_session__process_pipe_events(session, tool);
1401 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1403 struct perf_evsel *evsel;
1405 evlist__for_each(session->evlist, evsel) {
1406 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1410 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1414 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1415 const char *symbol_name, u64 addr)
1419 struct ref_reloc_sym *ref;
1421 ref = zalloc(sizeof(struct ref_reloc_sym));
1425 ref->name = strdup(symbol_name);
1426 if (ref->name == NULL) {
1431 bracket = strchr(ref->name, ']');
1437 for (i = 0; i < MAP__NR_TYPES; ++i) {
1438 struct kmap *kmap = map__kmap(maps[i]);
1439 kmap->ref_reloc_sym = ref;
1445 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1447 return machines__fprintf_dsos(&session->machines, fp);
1450 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1451 bool (skip)(struct dso *dso, int parm), int parm)
1453 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1456 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1458 struct perf_evsel *pos;
1459 size_t ret = fprintf(fp, "Aggregated stats:\n");
1461 ret += events_stats__fprintf(&session->stats, fp);
1463 evlist__for_each(session->evlist, pos) {
1464 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1465 ret += events_stats__fprintf(&pos->hists.stats, fp);
1471 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1474 * FIXME: Here we have to actually print all the machines in this
1475 * session, not just the host...
1477 return machine__fprintf(&session->machines.host, fp);
1480 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1483 struct perf_evsel *pos;
1485 evlist__for_each(session->evlist, pos) {
1486 if (pos->attr.type == type)
1492 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1493 struct addr_location *al,
1494 unsigned int print_opts, unsigned int stack_depth)
1496 struct callchain_cursor_node *node;
1497 int print_ip = print_opts & PRINT_IP_OPT_IP;
1498 int print_sym = print_opts & PRINT_IP_OPT_SYM;
1499 int print_dso = print_opts & PRINT_IP_OPT_DSO;
1500 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1501 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1502 int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1503 char s = print_oneline ? ' ' : '\t';
1505 if (symbol_conf.use_callchain && sample->callchain) {
1506 struct addr_location node_al;
1508 if (machine__resolve_callchain(al->machine, evsel, al->thread,
1510 PERF_MAX_STACK_DEPTH) != 0) {
1512 error("Failed to resolve callchain. Skipping\n");
1515 callchain_cursor_commit(&callchain_cursor);
1517 if (print_symoffset)
1520 while (stack_depth) {
1523 node = callchain_cursor_current(&callchain_cursor);
1527 if (node->sym && node->sym->ignore)
1531 printf("%c%16" PRIx64, s, node->ip);
1534 addr = node->map->map_ip(node->map, node->ip);
1538 if (print_symoffset) {
1539 node_al.addr = addr;
1540 node_al.map = node->map;
1541 symbol__fprintf_symname_offs(node->sym, &node_al, stdout);
1543 symbol__fprintf_symname(node->sym, stdout);
1548 map__fprintf_dsoname(node->map, stdout);
1553 map__fprintf_srcline(node->map, addr, "\n ",
1561 callchain_cursor_advance(&callchain_cursor);
1565 if (al->sym && al->sym->ignore)
1569 printf("%16" PRIx64, sample->ip);
1573 if (print_symoffset)
1574 symbol__fprintf_symname_offs(al->sym, al,
1577 symbol__fprintf_symname(al->sym, stdout);
1582 map__fprintf_dsoname(al->map, stdout);
1587 map__fprintf_srcline(al->map, al->addr, "\n ", stdout);
1591 int perf_session__cpu_bitmap(struct perf_session *session,
1592 const char *cpu_list, unsigned long *cpu_bitmap)
1595 struct cpu_map *map;
1597 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1598 struct perf_evsel *evsel;
1600 evsel = perf_session__find_first_evtype(session, i);
1604 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1605 pr_err("File does not contain CPU events. "
1606 "Remove -c option to proceed.\n");
1611 map = cpu_map__new(cpu_list);
1613 pr_err("Invalid cpu_list\n");
1617 for (i = 0; i < map->nr; i++) {
1618 int cpu = map->map[i];
1620 if (cpu >= MAX_NR_CPUS) {
1621 pr_err("Requested CPU %d too large. "
1622 "Consider raising MAX_NR_CPUS\n", cpu);
1623 goto out_delete_map;
1626 set_bit(cpu, cpu_bitmap);
1632 cpu_map__delete(map);
1636 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1642 if (session == NULL || fp == NULL)
1645 fd = perf_data_file__fd(session->file);
1647 ret = fstat(fd, &st);
1651 fprintf(fp, "# ========\n");
1652 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1653 perf_header__fprintf_info(session, fp, full);
1654 fprintf(fp, "# ========\n#\n");
1658 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1659 const struct perf_evsel_str_handler *assocs,
1662 struct perf_evsel *evsel;
1666 for (i = 0; i < nr_assocs; i++) {
1668 * Adding a handler for an event not in the session,
1671 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1676 if (evsel->handler != NULL)
1678 evsel->handler = assocs[i].handler;