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);
30 pr_err("failed to open file: %s", self->filename);
31 if (!strcmp(self->filename, "perf.data"))
32 pr_err(" (try 'perf record' first)");
37 if (fstat(self->fd, &input_stat) < 0)
40 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
41 pr_err("file %s not owned by current user or root\n",
46 if (!input_stat.st_size) {
47 pr_info("zero-sized file (%s), nothing to do!\n",
52 if (perf_header__read(self, self->fd) < 0) {
53 pr_err("incompatible file format");
57 self->size = input_stat.st_size;
66 void perf_session__update_sample_type(struct perf_session *self)
68 self->sample_type = perf_header__sample_type(&self->header);
71 int perf_session__create_kernel_maps(struct perf_session *self)
73 int ret = machine__create_kernel_maps(&self->host_machine);
76 ret = machines__create_guest_kernel_maps(&self->machines);
80 struct perf_session *perf_session__new(const char *filename, int mode, bool force, bool repipe)
82 size_t len = filename ? strlen(filename) + 1 : 0;
83 struct perf_session *self = zalloc(sizeof(*self) + len);
88 if (perf_header__init(&self->header) < 0)
91 memcpy(self->filename, filename, len);
92 self->threads = RB_ROOT;
93 INIT_LIST_HEAD(&self->dead_threads);
94 self->hists_tree = RB_ROOT;
95 self->last_match = NULL;
96 self->mmap_window = 32;
99 self->machines = RB_ROOT;
100 self->repipe = repipe;
101 INIT_LIST_HEAD(&self->ordered_samples.samples_head);
102 machine__init(&self->host_machine, "", HOST_KERNEL_ID);
104 if (mode == O_RDONLY) {
105 if (perf_session__open(self, force) < 0)
107 } else if (mode == O_WRONLY) {
109 * In O_RDONLY mode this will be performed when reading the
110 * kernel MMAP event, in event__process_mmap().
112 if (perf_session__create_kernel_maps(self) < 0)
116 perf_session__update_sample_type(self);
123 perf_session__delete(self);
127 void perf_session__delete(struct perf_session *self)
129 perf_header__exit(&self->header);
135 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
137 rb_erase(&th->rb_node, &self->threads);
139 * We may have references to this thread, for instance in some hist_entry
140 * instances, so just move them to a separate list.
142 list_add_tail(&th->node, &self->dead_threads);
145 static bool symbol__match_parent_regex(struct symbol *sym)
147 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
153 struct map_symbol *perf_session__resolve_callchain(struct perf_session *self,
154 struct thread *thread,
155 struct ip_callchain *chain,
156 struct symbol **parent)
158 u8 cpumode = PERF_RECORD_MISC_USER;
160 struct map_symbol *syms = calloc(chain->nr, sizeof(*syms));
165 for (i = 0; i < chain->nr; i++) {
166 u64 ip = chain->ips[i];
167 struct addr_location al;
169 if (ip >= PERF_CONTEXT_MAX) {
171 case PERF_CONTEXT_HV:
172 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
173 case PERF_CONTEXT_KERNEL:
174 cpumode = PERF_RECORD_MISC_KERNEL; break;
175 case PERF_CONTEXT_USER:
176 cpumode = PERF_RECORD_MISC_USER; break;
184 thread__find_addr_location(thread, self, cpumode,
185 MAP__FUNCTION, thread->pid, ip, &al, NULL);
186 if (al.sym != NULL) {
187 if (sort__has_parent && !*parent &&
188 symbol__match_parent_regex(al.sym))
190 if (!symbol_conf.use_callchain)
192 syms[i].map = al.map;
193 syms[i].sym = al.sym;
200 static int process_event_stub(event_t *event __used,
201 struct perf_session *session __used)
203 dump_printf(": unhandled!\n");
207 static int process_finished_round_stub(event_t *event __used,
208 struct perf_session *session __used,
209 struct perf_event_ops *ops __used)
211 dump_printf(": unhandled!\n");
215 static int process_finished_round(event_t *event,
216 struct perf_session *session,
217 struct perf_event_ops *ops);
219 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
221 if (handler->sample == NULL)
222 handler->sample = process_event_stub;
223 if (handler->mmap == NULL)
224 handler->mmap = process_event_stub;
225 if (handler->comm == NULL)
226 handler->comm = process_event_stub;
227 if (handler->fork == NULL)
228 handler->fork = process_event_stub;
229 if (handler->exit == NULL)
230 handler->exit = process_event_stub;
231 if (handler->lost == NULL)
232 handler->lost = process_event_stub;
233 if (handler->read == NULL)
234 handler->read = process_event_stub;
235 if (handler->throttle == NULL)
236 handler->throttle = process_event_stub;
237 if (handler->unthrottle == NULL)
238 handler->unthrottle = process_event_stub;
239 if (handler->attr == NULL)
240 handler->attr = process_event_stub;
241 if (handler->event_type == NULL)
242 handler->event_type = process_event_stub;
243 if (handler->tracing_data == NULL)
244 handler->tracing_data = process_event_stub;
245 if (handler->build_id == NULL)
246 handler->build_id = process_event_stub;
247 if (handler->finished_round == NULL) {
248 if (handler->ordered_samples)
249 handler->finished_round = process_finished_round;
251 handler->finished_round = process_finished_round_stub;
255 void mem_bswap_64(void *src, int byte_size)
259 while (byte_size > 0) {
261 byte_size -= sizeof(u64);
266 static void event__all64_swap(event_t *self)
268 struct perf_event_header *hdr = &self->header;
269 mem_bswap_64(hdr + 1, self->header.size - sizeof(*hdr));
272 static void event__comm_swap(event_t *self)
274 self->comm.pid = bswap_32(self->comm.pid);
275 self->comm.tid = bswap_32(self->comm.tid);
278 static void event__mmap_swap(event_t *self)
280 self->mmap.pid = bswap_32(self->mmap.pid);
281 self->mmap.tid = bswap_32(self->mmap.tid);
282 self->mmap.start = bswap_64(self->mmap.start);
283 self->mmap.len = bswap_64(self->mmap.len);
284 self->mmap.pgoff = bswap_64(self->mmap.pgoff);
287 static void event__task_swap(event_t *self)
289 self->fork.pid = bswap_32(self->fork.pid);
290 self->fork.tid = bswap_32(self->fork.tid);
291 self->fork.ppid = bswap_32(self->fork.ppid);
292 self->fork.ptid = bswap_32(self->fork.ptid);
293 self->fork.time = bswap_64(self->fork.time);
296 static void event__read_swap(event_t *self)
298 self->read.pid = bswap_32(self->read.pid);
299 self->read.tid = bswap_32(self->read.tid);
300 self->read.value = bswap_64(self->read.value);
301 self->read.time_enabled = bswap_64(self->read.time_enabled);
302 self->read.time_running = bswap_64(self->read.time_running);
303 self->read.id = bswap_64(self->read.id);
306 static void event__attr_swap(event_t *self)
310 self->attr.attr.type = bswap_32(self->attr.attr.type);
311 self->attr.attr.size = bswap_32(self->attr.attr.size);
312 self->attr.attr.config = bswap_64(self->attr.attr.config);
313 self->attr.attr.sample_period = bswap_64(self->attr.attr.sample_period);
314 self->attr.attr.sample_type = bswap_64(self->attr.attr.sample_type);
315 self->attr.attr.read_format = bswap_64(self->attr.attr.read_format);
316 self->attr.attr.wakeup_events = bswap_32(self->attr.attr.wakeup_events);
317 self->attr.attr.bp_type = bswap_32(self->attr.attr.bp_type);
318 self->attr.attr.bp_addr = bswap_64(self->attr.attr.bp_addr);
319 self->attr.attr.bp_len = bswap_64(self->attr.attr.bp_len);
321 size = self->header.size;
322 size -= (void *)&self->attr.id - (void *)self;
323 mem_bswap_64(self->attr.id, size);
326 static void event__event_type_swap(event_t *self)
328 self->event_type.event_type.event_id =
329 bswap_64(self->event_type.event_type.event_id);
332 static void event__tracing_data_swap(event_t *self)
334 self->tracing_data.size = bswap_32(self->tracing_data.size);
337 typedef void (*event__swap_op)(event_t *self);
339 static event__swap_op event__swap_ops[] = {
340 [PERF_RECORD_MMAP] = event__mmap_swap,
341 [PERF_RECORD_COMM] = event__comm_swap,
342 [PERF_RECORD_FORK] = event__task_swap,
343 [PERF_RECORD_EXIT] = event__task_swap,
344 [PERF_RECORD_LOST] = event__all64_swap,
345 [PERF_RECORD_READ] = event__read_swap,
346 [PERF_RECORD_SAMPLE] = event__all64_swap,
347 [PERF_RECORD_HEADER_ATTR] = event__attr_swap,
348 [PERF_RECORD_HEADER_EVENT_TYPE] = event__event_type_swap,
349 [PERF_RECORD_HEADER_TRACING_DATA] = event__tracing_data_swap,
350 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
351 [PERF_RECORD_HEADER_MAX] = NULL,
354 struct sample_queue {
356 struct sample_event *event;
357 struct list_head list;
360 static void flush_sample_queue(struct perf_session *s,
361 struct perf_event_ops *ops)
363 struct list_head *head = &s->ordered_samples.samples_head;
364 u64 limit = s->ordered_samples.next_flush;
365 struct sample_queue *tmp, *iter;
367 if (!ops->ordered_samples || !limit)
370 list_for_each_entry_safe(iter, tmp, head, list) {
371 if (iter->timestamp > limit)
374 if (iter == s->ordered_samples.last_inserted)
375 s->ordered_samples.last_inserted = NULL;
377 ops->sample((event_t *)iter->event, s);
379 s->ordered_samples.last_flush = iter->timestamp;
380 list_del(&iter->list);
387 * When perf record finishes a pass on every buffers, it records this pseudo
389 * We record the max timestamp t found in the pass n.
390 * Assuming these timestamps are monotonic across cpus, we know that if
391 * a buffer still has events with timestamps below t, they will be all
392 * available and then read in the pass n + 1.
393 * Hence when we start to read the pass n + 2, we can safely flush every
394 * events with timestamps below t.
396 * ============ PASS n =================
399 * cnt1 timestamps | cnt2 timestamps
402 * - | 4 <--- max recorded
404 * ============ PASS n + 1 ==============
407 * cnt1 timestamps | cnt2 timestamps
410 * 5 | 7 <---- max recorded
412 * Flush every events below timestamp 4
414 * ============ PASS n + 2 ==============
417 * cnt1 timestamps | cnt2 timestamps
422 * Flush every events below timestamp 7
425 static int process_finished_round(event_t *event __used,
426 struct perf_session *session,
427 struct perf_event_ops *ops)
429 flush_sample_queue(session, ops);
430 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
435 static void __queue_sample_end(struct sample_queue *new, struct list_head *head)
437 struct sample_queue *iter;
439 list_for_each_entry_reverse(iter, head, list) {
440 if (iter->timestamp < new->timestamp) {
441 list_add(&new->list, &iter->list);
446 list_add(&new->list, head);
449 static void __queue_sample_before(struct sample_queue *new,
450 struct sample_queue *iter,
451 struct list_head *head)
453 list_for_each_entry_continue_reverse(iter, head, list) {
454 if (iter->timestamp < new->timestamp) {
455 list_add(&new->list, &iter->list);
460 list_add(&new->list, head);
463 static void __queue_sample_after(struct sample_queue *new,
464 struct sample_queue *iter,
465 struct list_head *head)
467 list_for_each_entry_continue(iter, head, list) {
468 if (iter->timestamp > new->timestamp) {
469 list_add_tail(&new->list, &iter->list);
473 list_add_tail(&new->list, head);
476 /* The queue is ordered by time */
477 static void __queue_sample_event(struct sample_queue *new,
478 struct perf_session *s)
480 struct sample_queue *last_inserted = s->ordered_samples.last_inserted;
481 struct list_head *head = &s->ordered_samples.samples_head;
484 if (!last_inserted) {
485 __queue_sample_end(new, head);
490 * Most of the time the current event has a timestamp
491 * very close to the last event inserted, unless we just switched
492 * to another event buffer. Having a sorting based on a list and
493 * on the last inserted event that is close to the current one is
494 * probably more efficient than an rbtree based sorting.
496 if (last_inserted->timestamp >= new->timestamp)
497 __queue_sample_before(new, last_inserted, head);
499 __queue_sample_after(new, last_inserted, head);
502 static int queue_sample_event(event_t *event, struct sample_data *data,
503 struct perf_session *s)
505 u64 timestamp = data->time;
506 struct sample_queue *new;
509 if (timestamp < s->ordered_samples.last_flush) {
510 printf("Warning: Timestamp below last timeslice flush\n");
514 new = malloc(sizeof(*new));
518 new->timestamp = timestamp;
520 new->event = malloc(event->header.size);
526 memcpy(new->event, event, event->header.size);
528 __queue_sample_event(new, s);
529 s->ordered_samples.last_inserted = new;
531 if (new->timestamp > s->ordered_samples.max_timestamp)
532 s->ordered_samples.max_timestamp = new->timestamp;
537 static int perf_session__process_sample(event_t *event, struct perf_session *s,
538 struct perf_event_ops *ops)
540 struct sample_data data;
542 if (!ops->ordered_samples)
543 return ops->sample(event, s);
545 bzero(&data, sizeof(struct sample_data));
546 event__parse_sample(event, s->sample_type, &data);
548 queue_sample_event(event, &data, s);
553 static int perf_session__process_event(struct perf_session *self,
555 struct perf_event_ops *ops,
556 u64 offset, u64 head)
560 if (event->header.type < PERF_RECORD_HEADER_MAX) {
561 dump_printf("%#Lx [%#x]: PERF_RECORD_%s",
562 offset + head, event->header.size,
563 event__name[event->header.type]);
564 hists__inc_nr_events(&self->hists, event->header.type);
567 if (self->header.needs_swap && event__swap_ops[event->header.type])
568 event__swap_ops[event->header.type](event);
570 switch (event->header.type) {
571 case PERF_RECORD_SAMPLE:
572 return perf_session__process_sample(event, self, ops);
573 case PERF_RECORD_MMAP:
574 return ops->mmap(event, self);
575 case PERF_RECORD_COMM:
576 return ops->comm(event, self);
577 case PERF_RECORD_FORK:
578 return ops->fork(event, self);
579 case PERF_RECORD_EXIT:
580 return ops->exit(event, self);
581 case PERF_RECORD_LOST:
582 return ops->lost(event, self);
583 case PERF_RECORD_READ:
584 return ops->read(event, self);
585 case PERF_RECORD_THROTTLE:
586 return ops->throttle(event, self);
587 case PERF_RECORD_UNTHROTTLE:
588 return ops->unthrottle(event, self);
589 case PERF_RECORD_HEADER_ATTR:
590 return ops->attr(event, self);
591 case PERF_RECORD_HEADER_EVENT_TYPE:
592 return ops->event_type(event, self);
593 case PERF_RECORD_HEADER_TRACING_DATA:
594 /* setup for reading amidst mmap */
595 lseek(self->fd, offset + head, SEEK_SET);
596 return ops->tracing_data(event, self);
597 case PERF_RECORD_HEADER_BUILD_ID:
598 return ops->build_id(event, self);
599 case PERF_RECORD_FINISHED_ROUND:
600 return ops->finished_round(event, self, ops);
602 ++self->hists.stats.nr_unknown_events;
607 void perf_event_header__bswap(struct perf_event_header *self)
609 self->type = bswap_32(self->type);
610 self->misc = bswap_16(self->misc);
611 self->size = bswap_16(self->size);
614 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
616 struct thread *thread = perf_session__findnew(self, 0);
618 if (thread == NULL || thread__set_comm(thread, "swapper")) {
619 pr_err("problem inserting idle task.\n");
626 int do_read(int fd, void *buf, size_t size)
628 void *buf_start = buf;
631 int ret = read(fd, buf, size);
640 return buf - buf_start;
643 #define session_done() (*(volatile int *)(&session_done))
644 volatile int session_done;
646 static int __perf_session__process_pipe_events(struct perf_session *self,
647 struct perf_event_ops *ops)
656 perf_event_ops__fill_defaults(ops);
660 err = do_read(self->fd, &event, sizeof(struct perf_event_header));
665 pr_err("failed to read event header\n");
669 if (self->header.needs_swap)
670 perf_event_header__bswap(&event.header);
672 size = event.header.size;
677 p += sizeof(struct perf_event_header);
679 if (size - sizeof(struct perf_event_header)) {
680 err = do_read(self->fd, p,
681 size - sizeof(struct perf_event_header));
684 pr_err("unexpected end of event stream\n");
688 pr_err("failed to read event data\n");
694 (skip = perf_session__process_event(self, &event, ops,
696 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
697 head, event.header.size, event.header.type);
699 * assume we lost track of the stream, check alignment, and
700 * increment a single u64 in the hope to catch on again 'soon'.
702 if (unlikely(head & 7))
710 dump_printf("\n%#Lx [%#x]: event: %d\n",
711 head, event.header.size, event.header.type);
724 int __perf_session__process_events(struct perf_session *self,
725 u64 data_offset, u64 data_size,
726 u64 file_size, struct perf_event_ops *ops)
728 int err, mmap_prot, mmap_flags;
735 struct ui_progress *progress = ui_progress__new("Processing events...",
737 if (progress == NULL)
740 perf_event_ops__fill_defaults(ops);
742 page_size = sysconf(_SC_PAGESIZE);
745 shift = page_size * (head / page_size);
749 mmap_prot = PROT_READ;
750 mmap_flags = MAP_SHARED;
752 if (self->header.needs_swap) {
753 mmap_prot |= PROT_WRITE;
754 mmap_flags = MAP_PRIVATE;
757 buf = mmap(NULL, page_size * self->mmap_window, mmap_prot,
758 mmap_flags, self->fd, offset);
759 if (buf == MAP_FAILED) {
760 pr_err("failed to mmap file\n");
766 event = (event_t *)(buf + head);
767 ui_progress__update(progress, offset);
769 if (self->header.needs_swap)
770 perf_event_header__bswap(&event->header);
771 size = event->header.size;
775 if (head + event->header.size >= page_size * self->mmap_window) {
778 shift = page_size * (head / page_size);
780 munmap_ret = munmap(buf, page_size * self->mmap_window);
781 assert(munmap_ret == 0);
788 size = event->header.size;
790 dump_printf("\n%#Lx [%#x]: event: %d\n",
791 offset + head, event->header.size, event->header.type);
794 perf_session__process_event(self, event, ops, offset, head) < 0) {
795 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
796 offset + head, event->header.size,
799 * assume we lost track of the stream, check alignment, and
800 * increment a single u64 in the hope to catch on again 'soon'.
802 if (unlikely(head & 7))
810 if (offset + head >= data_offset + data_size)
813 if (offset + head < file_size)
817 /* do the final flush for ordered samples */
818 self->ordered_samples.next_flush = ULLONG_MAX;
819 flush_sample_queue(self, ops);
821 ui_progress__delete(progress);
825 int perf_session__process_events(struct perf_session *self,
826 struct perf_event_ops *ops)
830 if (perf_session__register_idle_thread(self) == NULL)
833 if (!symbol_conf.full_paths) {
836 if (getcwd(bf, sizeof(bf)) == NULL) {
839 pr_err("failed to get the current directory\n");
842 self->cwd = strdup(bf);
843 if (self->cwd == NULL) {
847 self->cwdlen = strlen(self->cwd);
851 err = __perf_session__process_events(self,
852 self->header.data_offset,
853 self->header.data_size,
856 err = __perf_session__process_pipe_events(self, ops);
861 bool perf_session__has_traces(struct perf_session *self, const char *msg)
863 if (!(self->sample_type & PERF_SAMPLE_RAW)) {
864 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
871 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
872 const char *symbol_name,
877 struct ref_reloc_sym *ref;
879 ref = zalloc(sizeof(struct ref_reloc_sym));
883 ref->name = strdup(symbol_name);
884 if (ref->name == NULL) {
889 bracket = strchr(ref->name, ']');
895 for (i = 0; i < MAP__NR_TYPES; ++i) {
896 struct kmap *kmap = map__kmap(maps[i]);
897 kmap->ref_reloc_sym = ref;
903 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
905 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
906 __dsos__fprintf(&self->host_machine.user_dsos, fp) +
907 machines__fprintf_dsos(&self->machines, fp);
910 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
913 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
914 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);