1 #define _FILE_OFFSET_BITS 64
3 #include <linux/kernel.h>
14 static int perf_session__open(struct perf_session *self, bool force)
16 struct stat input_stat;
18 if (!strcmp(self->filename, "-")) {
20 self->fd = STDIN_FILENO;
22 if (perf_header__read(self, self->fd) < 0)
23 pr_err("incompatible file format");
28 self->fd = open(self->filename, O_RDONLY);
32 pr_err("failed to open %s: %s", self->filename, strerror(err));
33 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
34 pr_err(" (try 'perf record' first)");
39 if (fstat(self->fd, &input_stat) < 0)
42 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
43 pr_err("file %s not owned by current user or root\n",
48 if (!input_stat.st_size) {
49 pr_info("zero-sized file (%s), nothing to do!\n",
54 if (perf_header__read(self, self->fd) < 0) {
55 pr_err("incompatible file format");
59 self->size = input_stat.st_size;
68 void perf_session__update_sample_type(struct perf_session *self)
70 self->sample_type = perf_header__sample_type(&self->header);
73 int perf_session__create_kernel_maps(struct perf_session *self)
75 int ret = machine__create_kernel_maps(&self->host_machine);
78 ret = machines__create_guest_kernel_maps(&self->machines);
82 struct perf_session *perf_session__new(const char *filename, int mode, bool force, bool repipe)
84 size_t len = filename ? strlen(filename) + 1 : 0;
85 struct perf_session *self = zalloc(sizeof(*self) + len);
90 if (perf_header__init(&self->header) < 0)
93 memcpy(self->filename, filename, len);
94 self->threads = RB_ROOT;
95 INIT_LIST_HEAD(&self->dead_threads);
96 self->hists_tree = RB_ROOT;
97 self->last_match = NULL;
98 self->mmap_window = 32;
101 self->machines = RB_ROOT;
102 self->repipe = repipe;
103 INIT_LIST_HEAD(&self->ordered_samples.samples_head);
104 machine__init(&self->host_machine, "", HOST_KERNEL_ID);
106 if (mode == O_RDONLY) {
107 if (perf_session__open(self, force) < 0)
109 } else if (mode == O_WRONLY) {
111 * In O_RDONLY mode this will be performed when reading the
112 * kernel MMAP event, in event__process_mmap().
114 if (perf_session__create_kernel_maps(self) < 0)
118 perf_session__update_sample_type(self);
125 perf_session__delete(self);
129 void perf_session__delete(struct perf_session *self)
131 perf_header__exit(&self->header);
137 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
139 rb_erase(&th->rb_node, &self->threads);
141 * We may have references to this thread, for instance in some hist_entry
142 * instances, so just move them to a separate list.
144 list_add_tail(&th->node, &self->dead_threads);
147 static bool symbol__match_parent_regex(struct symbol *sym)
149 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
155 struct map_symbol *perf_session__resolve_callchain(struct perf_session *self,
156 struct thread *thread,
157 struct ip_callchain *chain,
158 struct symbol **parent)
160 u8 cpumode = PERF_RECORD_MISC_USER;
162 struct map_symbol *syms = calloc(chain->nr, sizeof(*syms));
167 for (i = 0; i < chain->nr; i++) {
168 u64 ip = chain->ips[i];
169 struct addr_location al;
171 if (ip >= PERF_CONTEXT_MAX) {
173 case PERF_CONTEXT_HV:
174 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
175 case PERF_CONTEXT_KERNEL:
176 cpumode = PERF_RECORD_MISC_KERNEL; break;
177 case PERF_CONTEXT_USER:
178 cpumode = PERF_RECORD_MISC_USER; break;
186 thread__find_addr_location(thread, self, cpumode,
187 MAP__FUNCTION, thread->pid, ip, &al, NULL);
188 if (al.sym != NULL) {
189 if (sort__has_parent && !*parent &&
190 symbol__match_parent_regex(al.sym))
192 if (!symbol_conf.use_callchain)
194 syms[i].map = al.map;
195 syms[i].sym = al.sym;
202 static int process_event_stub(event_t *event __used,
203 struct perf_session *session __used)
205 dump_printf(": unhandled!\n");
209 static int process_finished_round_stub(event_t *event __used,
210 struct perf_session *session __used,
211 struct perf_event_ops *ops __used)
213 dump_printf(": unhandled!\n");
217 static int process_finished_round(event_t *event,
218 struct perf_session *session,
219 struct perf_event_ops *ops);
221 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
223 if (handler->sample == NULL)
224 handler->sample = process_event_stub;
225 if (handler->mmap == NULL)
226 handler->mmap = process_event_stub;
227 if (handler->comm == NULL)
228 handler->comm = process_event_stub;
229 if (handler->fork == NULL)
230 handler->fork = process_event_stub;
231 if (handler->exit == NULL)
232 handler->exit = process_event_stub;
233 if (handler->lost == NULL)
234 handler->lost = process_event_stub;
235 if (handler->read == NULL)
236 handler->read = process_event_stub;
237 if (handler->throttle == NULL)
238 handler->throttle = process_event_stub;
239 if (handler->unthrottle == NULL)
240 handler->unthrottle = process_event_stub;
241 if (handler->attr == NULL)
242 handler->attr = process_event_stub;
243 if (handler->event_type == NULL)
244 handler->event_type = process_event_stub;
245 if (handler->tracing_data == NULL)
246 handler->tracing_data = process_event_stub;
247 if (handler->build_id == NULL)
248 handler->build_id = process_event_stub;
249 if (handler->finished_round == NULL) {
250 if (handler->ordered_samples)
251 handler->finished_round = process_finished_round;
253 handler->finished_round = process_finished_round_stub;
257 void mem_bswap_64(void *src, int byte_size)
261 while (byte_size > 0) {
263 byte_size -= sizeof(u64);
268 static void event__all64_swap(event_t *self)
270 struct perf_event_header *hdr = &self->header;
271 mem_bswap_64(hdr + 1, self->header.size - sizeof(*hdr));
274 static void event__comm_swap(event_t *self)
276 self->comm.pid = bswap_32(self->comm.pid);
277 self->comm.tid = bswap_32(self->comm.tid);
280 static void event__mmap_swap(event_t *self)
282 self->mmap.pid = bswap_32(self->mmap.pid);
283 self->mmap.tid = bswap_32(self->mmap.tid);
284 self->mmap.start = bswap_64(self->mmap.start);
285 self->mmap.len = bswap_64(self->mmap.len);
286 self->mmap.pgoff = bswap_64(self->mmap.pgoff);
289 static void event__task_swap(event_t *self)
291 self->fork.pid = bswap_32(self->fork.pid);
292 self->fork.tid = bswap_32(self->fork.tid);
293 self->fork.ppid = bswap_32(self->fork.ppid);
294 self->fork.ptid = bswap_32(self->fork.ptid);
295 self->fork.time = bswap_64(self->fork.time);
298 static void event__read_swap(event_t *self)
300 self->read.pid = bswap_32(self->read.pid);
301 self->read.tid = bswap_32(self->read.tid);
302 self->read.value = bswap_64(self->read.value);
303 self->read.time_enabled = bswap_64(self->read.time_enabled);
304 self->read.time_running = bswap_64(self->read.time_running);
305 self->read.id = bswap_64(self->read.id);
308 static void event__attr_swap(event_t *self)
312 self->attr.attr.type = bswap_32(self->attr.attr.type);
313 self->attr.attr.size = bswap_32(self->attr.attr.size);
314 self->attr.attr.config = bswap_64(self->attr.attr.config);
315 self->attr.attr.sample_period = bswap_64(self->attr.attr.sample_period);
316 self->attr.attr.sample_type = bswap_64(self->attr.attr.sample_type);
317 self->attr.attr.read_format = bswap_64(self->attr.attr.read_format);
318 self->attr.attr.wakeup_events = bswap_32(self->attr.attr.wakeup_events);
319 self->attr.attr.bp_type = bswap_32(self->attr.attr.bp_type);
320 self->attr.attr.bp_addr = bswap_64(self->attr.attr.bp_addr);
321 self->attr.attr.bp_len = bswap_64(self->attr.attr.bp_len);
323 size = self->header.size;
324 size -= (void *)&self->attr.id - (void *)self;
325 mem_bswap_64(self->attr.id, size);
328 static void event__event_type_swap(event_t *self)
330 self->event_type.event_type.event_id =
331 bswap_64(self->event_type.event_type.event_id);
334 static void event__tracing_data_swap(event_t *self)
336 self->tracing_data.size = bswap_32(self->tracing_data.size);
339 typedef void (*event__swap_op)(event_t *self);
341 static event__swap_op event__swap_ops[] = {
342 [PERF_RECORD_MMAP] = event__mmap_swap,
343 [PERF_RECORD_COMM] = event__comm_swap,
344 [PERF_RECORD_FORK] = event__task_swap,
345 [PERF_RECORD_EXIT] = event__task_swap,
346 [PERF_RECORD_LOST] = event__all64_swap,
347 [PERF_RECORD_READ] = event__read_swap,
348 [PERF_RECORD_SAMPLE] = event__all64_swap,
349 [PERF_RECORD_HEADER_ATTR] = event__attr_swap,
350 [PERF_RECORD_HEADER_EVENT_TYPE] = event__event_type_swap,
351 [PERF_RECORD_HEADER_TRACING_DATA] = event__tracing_data_swap,
352 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
353 [PERF_RECORD_HEADER_MAX] = NULL,
356 struct sample_queue {
358 struct sample_event *event;
359 struct list_head list;
362 static void flush_sample_queue(struct perf_session *s,
363 struct perf_event_ops *ops)
365 struct list_head *head = &s->ordered_samples.samples_head;
366 u64 limit = s->ordered_samples.next_flush;
367 struct sample_queue *tmp, *iter;
369 if (!ops->ordered_samples || !limit)
372 list_for_each_entry_safe(iter, tmp, head, list) {
373 if (iter->timestamp > limit)
376 if (iter == s->ordered_samples.last_inserted)
377 s->ordered_samples.last_inserted = NULL;
379 ops->sample((event_t *)iter->event, s);
381 s->ordered_samples.last_flush = iter->timestamp;
382 list_del(&iter->list);
389 * When perf record finishes a pass on every buffers, it records this pseudo
391 * We record the max timestamp t found in the pass n.
392 * Assuming these timestamps are monotonic across cpus, we know that if
393 * a buffer still has events with timestamps below t, they will be all
394 * available and then read in the pass n + 1.
395 * Hence when we start to read the pass n + 2, we can safely flush every
396 * events with timestamps below t.
398 * ============ PASS n =================
401 * cnt1 timestamps | cnt2 timestamps
404 * - | 4 <--- max recorded
406 * ============ PASS n + 1 ==============
409 * cnt1 timestamps | cnt2 timestamps
412 * 5 | 7 <---- max recorded
414 * Flush every events below timestamp 4
416 * ============ PASS n + 2 ==============
419 * cnt1 timestamps | cnt2 timestamps
424 * Flush every events below timestamp 7
427 static int process_finished_round(event_t *event __used,
428 struct perf_session *session,
429 struct perf_event_ops *ops)
431 flush_sample_queue(session, ops);
432 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
437 static void __queue_sample_end(struct sample_queue *new, struct list_head *head)
439 struct sample_queue *iter;
441 list_for_each_entry_reverse(iter, head, list) {
442 if (iter->timestamp < new->timestamp) {
443 list_add(&new->list, &iter->list);
448 list_add(&new->list, head);
451 static void __queue_sample_before(struct sample_queue *new,
452 struct sample_queue *iter,
453 struct list_head *head)
455 list_for_each_entry_continue_reverse(iter, head, list) {
456 if (iter->timestamp < new->timestamp) {
457 list_add(&new->list, &iter->list);
462 list_add(&new->list, head);
465 static void __queue_sample_after(struct sample_queue *new,
466 struct sample_queue *iter,
467 struct list_head *head)
469 list_for_each_entry_continue(iter, head, list) {
470 if (iter->timestamp > new->timestamp) {
471 list_add_tail(&new->list, &iter->list);
475 list_add_tail(&new->list, head);
478 /* The queue is ordered by time */
479 static void __queue_sample_event(struct sample_queue *new,
480 struct perf_session *s)
482 struct sample_queue *last_inserted = s->ordered_samples.last_inserted;
483 struct list_head *head = &s->ordered_samples.samples_head;
486 if (!last_inserted) {
487 __queue_sample_end(new, head);
492 * Most of the time the current event has a timestamp
493 * very close to the last event inserted, unless we just switched
494 * to another event buffer. Having a sorting based on a list and
495 * on the last inserted event that is close to the current one is
496 * probably more efficient than an rbtree based sorting.
498 if (last_inserted->timestamp >= new->timestamp)
499 __queue_sample_before(new, last_inserted, head);
501 __queue_sample_after(new, last_inserted, head);
504 static int queue_sample_event(event_t *event, struct sample_data *data,
505 struct perf_session *s)
507 u64 timestamp = data->time;
508 struct sample_queue *new;
511 if (timestamp < s->ordered_samples.last_flush) {
512 printf("Warning: Timestamp below last timeslice flush\n");
516 new = malloc(sizeof(*new));
520 new->timestamp = timestamp;
522 new->event = malloc(event->header.size);
528 memcpy(new->event, event, event->header.size);
530 __queue_sample_event(new, s);
531 s->ordered_samples.last_inserted = new;
533 if (new->timestamp > s->ordered_samples.max_timestamp)
534 s->ordered_samples.max_timestamp = new->timestamp;
539 static int perf_session__process_sample(event_t *event, struct perf_session *s,
540 struct perf_event_ops *ops)
542 struct sample_data data;
544 if (!ops->ordered_samples)
545 return ops->sample(event, s);
547 bzero(&data, sizeof(struct sample_data));
548 event__parse_sample(event, s->sample_type, &data);
550 queue_sample_event(event, &data, s);
555 static int perf_session__process_event(struct perf_session *self,
557 struct perf_event_ops *ops,
558 u64 offset, u64 head)
562 if (event->header.type < PERF_RECORD_HEADER_MAX) {
563 dump_printf("%#Lx [%#x]: PERF_RECORD_%s",
564 offset + head, event->header.size,
565 event__name[event->header.type]);
566 hists__inc_nr_events(&self->hists, event->header.type);
569 if (self->header.needs_swap && event__swap_ops[event->header.type])
570 event__swap_ops[event->header.type](event);
572 switch (event->header.type) {
573 case PERF_RECORD_SAMPLE:
574 return perf_session__process_sample(event, self, ops);
575 case PERF_RECORD_MMAP:
576 return ops->mmap(event, self);
577 case PERF_RECORD_COMM:
578 return ops->comm(event, self);
579 case PERF_RECORD_FORK:
580 return ops->fork(event, self);
581 case PERF_RECORD_EXIT:
582 return ops->exit(event, self);
583 case PERF_RECORD_LOST:
584 return ops->lost(event, self);
585 case PERF_RECORD_READ:
586 return ops->read(event, self);
587 case PERF_RECORD_THROTTLE:
588 return ops->throttle(event, self);
589 case PERF_RECORD_UNTHROTTLE:
590 return ops->unthrottle(event, self);
591 case PERF_RECORD_HEADER_ATTR:
592 return ops->attr(event, self);
593 case PERF_RECORD_HEADER_EVENT_TYPE:
594 return ops->event_type(event, self);
595 case PERF_RECORD_HEADER_TRACING_DATA:
596 /* setup for reading amidst mmap */
597 lseek(self->fd, offset + head, SEEK_SET);
598 return ops->tracing_data(event, self);
599 case PERF_RECORD_HEADER_BUILD_ID:
600 return ops->build_id(event, self);
601 case PERF_RECORD_FINISHED_ROUND:
602 return ops->finished_round(event, self, ops);
604 ++self->hists.stats.nr_unknown_events;
609 void perf_event_header__bswap(struct perf_event_header *self)
611 self->type = bswap_32(self->type);
612 self->misc = bswap_16(self->misc);
613 self->size = bswap_16(self->size);
616 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
618 struct thread *thread = perf_session__findnew(self, 0);
620 if (thread == NULL || thread__set_comm(thread, "swapper")) {
621 pr_err("problem inserting idle task.\n");
628 int do_read(int fd, void *buf, size_t size)
630 void *buf_start = buf;
633 int ret = read(fd, buf, size);
642 return buf - buf_start;
645 #define session_done() (*(volatile int *)(&session_done))
646 volatile int session_done;
648 static int __perf_session__process_pipe_events(struct perf_session *self,
649 struct perf_event_ops *ops)
658 perf_event_ops__fill_defaults(ops);
662 err = do_read(self->fd, &event, sizeof(struct perf_event_header));
667 pr_err("failed to read event header\n");
671 if (self->header.needs_swap)
672 perf_event_header__bswap(&event.header);
674 size = event.header.size;
679 p += sizeof(struct perf_event_header);
681 if (size - sizeof(struct perf_event_header)) {
682 err = do_read(self->fd, p,
683 size - sizeof(struct perf_event_header));
686 pr_err("unexpected end of event stream\n");
690 pr_err("failed to read event data\n");
696 (skip = perf_session__process_event(self, &event, ops,
698 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
699 head, event.header.size, event.header.type);
701 * assume we lost track of the stream, check alignment, and
702 * increment a single u64 in the hope to catch on again 'soon'.
704 if (unlikely(head & 7))
712 dump_printf("\n%#Lx [%#x]: event: %d\n",
713 head, event.header.size, event.header.type);
726 int __perf_session__process_events(struct perf_session *self,
727 u64 data_offset, u64 data_size,
728 u64 file_size, struct perf_event_ops *ops)
730 int err, mmap_prot, mmap_flags;
737 struct ui_progress *progress = ui_progress__new("Processing events...",
739 if (progress == NULL)
742 perf_event_ops__fill_defaults(ops);
744 page_size = sysconf(_SC_PAGESIZE);
747 shift = page_size * (head / page_size);
751 mmap_prot = PROT_READ;
752 mmap_flags = MAP_SHARED;
754 if (self->header.needs_swap) {
755 mmap_prot |= PROT_WRITE;
756 mmap_flags = MAP_PRIVATE;
759 buf = mmap(NULL, page_size * self->mmap_window, mmap_prot,
760 mmap_flags, self->fd, offset);
761 if (buf == MAP_FAILED) {
762 pr_err("failed to mmap file\n");
768 event = (event_t *)(buf + head);
769 ui_progress__update(progress, offset);
771 if (self->header.needs_swap)
772 perf_event_header__bswap(&event->header);
773 size = event->header.size;
777 if (head + event->header.size >= page_size * self->mmap_window) {
780 shift = page_size * (head / page_size);
782 munmap_ret = munmap(buf, page_size * self->mmap_window);
783 assert(munmap_ret == 0);
790 size = event->header.size;
792 dump_printf("\n%#Lx [%#x]: event: %d\n",
793 offset + head, event->header.size, event->header.type);
796 perf_session__process_event(self, event, ops, offset, head) < 0) {
797 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
798 offset + head, event->header.size,
801 * assume we lost track of the stream, check alignment, and
802 * increment a single u64 in the hope to catch on again 'soon'.
804 if (unlikely(head & 7))
812 if (offset + head >= data_offset + data_size)
815 if (offset + head < file_size)
819 /* do the final flush for ordered samples */
820 self->ordered_samples.next_flush = ULLONG_MAX;
821 flush_sample_queue(self, ops);
823 ui_progress__delete(progress);
827 int perf_session__process_events(struct perf_session *self,
828 struct perf_event_ops *ops)
832 if (perf_session__register_idle_thread(self) == NULL)
835 if (!symbol_conf.full_paths) {
838 if (getcwd(bf, sizeof(bf)) == NULL) {
841 pr_err("failed to get the current directory\n");
844 self->cwd = strdup(bf);
845 if (self->cwd == NULL) {
849 self->cwdlen = strlen(self->cwd);
853 err = __perf_session__process_events(self,
854 self->header.data_offset,
855 self->header.data_size,
858 err = __perf_session__process_pipe_events(self, ops);
863 bool perf_session__has_traces(struct perf_session *self, const char *msg)
865 if (!(self->sample_type & PERF_SAMPLE_RAW)) {
866 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
873 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
874 const char *symbol_name,
879 struct ref_reloc_sym *ref;
881 ref = zalloc(sizeof(struct ref_reloc_sym));
885 ref->name = strdup(symbol_name);
886 if (ref->name == NULL) {
891 bracket = strchr(ref->name, ']');
897 for (i = 0; i < MAP__NR_TYPES; ++i) {
898 struct kmap *kmap = map__kmap(maps[i]);
899 kmap->ref_reloc_sym = ref;
905 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
907 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
908 __dsos__fprintf(&self->host_machine.user_dsos, fp) +
909 machines__fprintf_dsos(&self->machines, fp);
912 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
915 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
916 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);