1 #define _FILE_OFFSET_BITS 64
3 #include <linux/kernel.h>
13 static int perf_session__open(struct perf_session *self, bool force)
15 struct stat input_stat;
17 if (!strcmp(self->filename, "-")) {
19 self->fd = STDIN_FILENO;
21 if (perf_header__read(self, self->fd) < 0)
22 pr_err("incompatible file format");
27 self->fd = open(self->filename, O_RDONLY);
29 pr_err("failed to open file: %s", self->filename);
30 if (!strcmp(self->filename, "perf.data"))
31 pr_err(" (try 'perf record' first)");
36 if (fstat(self->fd, &input_stat) < 0)
39 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
40 pr_err("file %s not owned by current user or root\n",
45 if (!input_stat.st_size) {
46 pr_info("zero-sized file (%s), nothing to do!\n",
51 if (perf_header__read(self, self->fd) < 0) {
52 pr_err("incompatible file format");
56 self->size = input_stat.st_size;
65 void perf_session__update_sample_type(struct perf_session *self)
67 self->sample_type = perf_header__sample_type(&self->header);
70 int perf_session__create_kernel_maps(struct perf_session *self)
72 int ret = machine__create_kernel_maps(&self->host_machine);
75 ret = machines__create_guest_kernel_maps(&self->machines);
79 struct perf_session *perf_session__new(const char *filename, int mode, bool force, bool repipe)
81 size_t len = filename ? strlen(filename) + 1 : 0;
82 struct perf_session *self = zalloc(sizeof(*self) + len);
87 if (perf_header__init(&self->header) < 0)
90 memcpy(self->filename, filename, len);
91 self->threads = RB_ROOT;
92 self->hists_tree = RB_ROOT;
93 self->last_match = NULL;
94 self->mmap_window = 32;
97 self->machines = RB_ROOT;
98 self->repipe = repipe;
99 INIT_LIST_HEAD(&self->ordered_samples.samples_head);
100 machine__init(&self->host_machine, "", HOST_KERNEL_ID);
102 if (mode == O_RDONLY) {
103 if (perf_session__open(self, force) < 0)
105 } else if (mode == O_WRONLY) {
107 * In O_RDONLY mode this will be performed when reading the
108 * kernel MMAP event, in event__process_mmap().
110 if (perf_session__create_kernel_maps(self) < 0)
114 perf_session__update_sample_type(self);
121 perf_session__delete(self);
125 void perf_session__delete(struct perf_session *self)
127 perf_header__exit(&self->header);
133 static bool symbol__match_parent_regex(struct symbol *sym)
135 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
141 struct map_symbol *perf_session__resolve_callchain(struct perf_session *self,
142 struct thread *thread,
143 struct ip_callchain *chain,
144 struct symbol **parent)
146 u8 cpumode = PERF_RECORD_MISC_USER;
148 struct map_symbol *syms = calloc(chain->nr, sizeof(*syms));
153 for (i = 0; i < chain->nr; i++) {
154 u64 ip = chain->ips[i];
155 struct addr_location al;
157 if (ip >= PERF_CONTEXT_MAX) {
159 case PERF_CONTEXT_HV:
160 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
161 case PERF_CONTEXT_KERNEL:
162 cpumode = PERF_RECORD_MISC_KERNEL; break;
163 case PERF_CONTEXT_USER:
164 cpumode = PERF_RECORD_MISC_USER; break;
172 thread__find_addr_location(thread, self, cpumode,
173 MAP__FUNCTION, thread->pid, ip, &al, NULL);
174 if (al.sym != NULL) {
175 if (sort__has_parent && !*parent &&
176 symbol__match_parent_regex(al.sym))
178 if (!symbol_conf.use_callchain)
180 syms[i].map = al.map;
181 syms[i].sym = al.sym;
188 static int process_event_stub(event_t *event __used,
189 struct perf_session *session __used)
191 dump_printf(": unhandled!\n");
195 static int process_finished_round_stub(event_t *event __used,
196 struct perf_session *session __used,
197 struct perf_event_ops *ops __used)
199 dump_printf(": unhandled!\n");
203 static int process_finished_round(event_t *event,
204 struct perf_session *session,
205 struct perf_event_ops *ops);
207 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
209 if (handler->sample == NULL)
210 handler->sample = process_event_stub;
211 if (handler->mmap == NULL)
212 handler->mmap = process_event_stub;
213 if (handler->comm == NULL)
214 handler->comm = process_event_stub;
215 if (handler->fork == NULL)
216 handler->fork = process_event_stub;
217 if (handler->exit == NULL)
218 handler->exit = process_event_stub;
219 if (handler->lost == NULL)
220 handler->lost = process_event_stub;
221 if (handler->read == NULL)
222 handler->read = process_event_stub;
223 if (handler->throttle == NULL)
224 handler->throttle = process_event_stub;
225 if (handler->unthrottle == NULL)
226 handler->unthrottle = process_event_stub;
227 if (handler->attr == NULL)
228 handler->attr = process_event_stub;
229 if (handler->event_type == NULL)
230 handler->event_type = process_event_stub;
231 if (handler->tracing_data == NULL)
232 handler->tracing_data = process_event_stub;
233 if (handler->build_id == NULL)
234 handler->build_id = process_event_stub;
235 if (handler->finished_round == NULL) {
236 if (handler->ordered_samples)
237 handler->finished_round = process_finished_round;
239 handler->finished_round = process_finished_round_stub;
243 void mem_bswap_64(void *src, int byte_size)
247 while (byte_size > 0) {
249 byte_size -= sizeof(u64);
254 static void event__all64_swap(event_t *self)
256 struct perf_event_header *hdr = &self->header;
257 mem_bswap_64(hdr + 1, self->header.size - sizeof(*hdr));
260 static void event__comm_swap(event_t *self)
262 self->comm.pid = bswap_32(self->comm.pid);
263 self->comm.tid = bswap_32(self->comm.tid);
266 static void event__mmap_swap(event_t *self)
268 self->mmap.pid = bswap_32(self->mmap.pid);
269 self->mmap.tid = bswap_32(self->mmap.tid);
270 self->mmap.start = bswap_64(self->mmap.start);
271 self->mmap.len = bswap_64(self->mmap.len);
272 self->mmap.pgoff = bswap_64(self->mmap.pgoff);
275 static void event__task_swap(event_t *self)
277 self->fork.pid = bswap_32(self->fork.pid);
278 self->fork.tid = bswap_32(self->fork.tid);
279 self->fork.ppid = bswap_32(self->fork.ppid);
280 self->fork.ptid = bswap_32(self->fork.ptid);
281 self->fork.time = bswap_64(self->fork.time);
284 static void event__read_swap(event_t *self)
286 self->read.pid = bswap_32(self->read.pid);
287 self->read.tid = bswap_32(self->read.tid);
288 self->read.value = bswap_64(self->read.value);
289 self->read.time_enabled = bswap_64(self->read.time_enabled);
290 self->read.time_running = bswap_64(self->read.time_running);
291 self->read.id = bswap_64(self->read.id);
294 static void event__attr_swap(event_t *self)
298 self->attr.attr.type = bswap_32(self->attr.attr.type);
299 self->attr.attr.size = bswap_32(self->attr.attr.size);
300 self->attr.attr.config = bswap_64(self->attr.attr.config);
301 self->attr.attr.sample_period = bswap_64(self->attr.attr.sample_period);
302 self->attr.attr.sample_type = bswap_64(self->attr.attr.sample_type);
303 self->attr.attr.read_format = bswap_64(self->attr.attr.read_format);
304 self->attr.attr.wakeup_events = bswap_32(self->attr.attr.wakeup_events);
305 self->attr.attr.bp_type = bswap_32(self->attr.attr.bp_type);
306 self->attr.attr.bp_addr = bswap_64(self->attr.attr.bp_addr);
307 self->attr.attr.bp_len = bswap_64(self->attr.attr.bp_len);
309 size = self->header.size;
310 size -= (void *)&self->attr.id - (void *)self;
311 mem_bswap_64(self->attr.id, size);
314 static void event__event_type_swap(event_t *self)
316 self->event_type.event_type.event_id =
317 bswap_64(self->event_type.event_type.event_id);
320 static void event__tracing_data_swap(event_t *self)
322 self->tracing_data.size = bswap_32(self->tracing_data.size);
325 typedef void (*event__swap_op)(event_t *self);
327 static event__swap_op event__swap_ops[] = {
328 [PERF_RECORD_MMAP] = event__mmap_swap,
329 [PERF_RECORD_COMM] = event__comm_swap,
330 [PERF_RECORD_FORK] = event__task_swap,
331 [PERF_RECORD_EXIT] = event__task_swap,
332 [PERF_RECORD_LOST] = event__all64_swap,
333 [PERF_RECORD_READ] = event__read_swap,
334 [PERF_RECORD_SAMPLE] = event__all64_swap,
335 [PERF_RECORD_HEADER_ATTR] = event__attr_swap,
336 [PERF_RECORD_HEADER_EVENT_TYPE] = event__event_type_swap,
337 [PERF_RECORD_HEADER_TRACING_DATA] = event__tracing_data_swap,
338 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
339 [PERF_RECORD_HEADER_MAX] = NULL,
342 struct sample_queue {
344 struct sample_event *event;
345 struct list_head list;
348 static void flush_sample_queue(struct perf_session *s,
349 struct perf_event_ops *ops)
351 struct list_head *head = &s->ordered_samples.samples_head;
352 u64 limit = s->ordered_samples.next_flush;
353 struct sample_queue *tmp, *iter;
355 if (!ops->ordered_samples || !limit)
358 list_for_each_entry_safe(iter, tmp, head, list) {
359 if (iter->timestamp > limit)
362 if (iter == s->ordered_samples.last_inserted)
363 s->ordered_samples.last_inserted = NULL;
365 ops->sample((event_t *)iter->event, s);
367 s->ordered_samples.last_flush = iter->timestamp;
368 list_del(&iter->list);
375 * When perf record finishes a pass on every buffers, it records this pseudo
377 * We record the max timestamp t found in the pass n.
378 * Assuming these timestamps are monotonic across cpus, we know that if
379 * a buffer still has events with timestamps below t, they will be all
380 * available and then read in the pass n + 1.
381 * Hence when we start to read the pass n + 2, we can safely flush every
382 * events with timestamps below t.
384 * ============ PASS n =================
387 * cnt1 timestamps | cnt2 timestamps
390 * - | 4 <--- max recorded
392 * ============ PASS n + 1 ==============
395 * cnt1 timestamps | cnt2 timestamps
398 * 5 | 7 <---- max recorded
400 * Flush every events below timestamp 4
402 * ============ PASS n + 2 ==============
405 * cnt1 timestamps | cnt2 timestamps
410 * Flush every events below timestamp 7
413 static int process_finished_round(event_t *event __used,
414 struct perf_session *session,
415 struct perf_event_ops *ops)
417 flush_sample_queue(session, ops);
418 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
423 static void __queue_sample_end(struct sample_queue *new, struct list_head *head)
425 struct sample_queue *iter;
427 list_for_each_entry_reverse(iter, head, list) {
428 if (iter->timestamp < new->timestamp) {
429 list_add(&new->list, &iter->list);
434 list_add(&new->list, head);
437 static void __queue_sample_before(struct sample_queue *new,
438 struct sample_queue *iter,
439 struct list_head *head)
441 list_for_each_entry_continue_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_after(struct sample_queue *new,
452 struct sample_queue *iter,
453 struct list_head *head)
455 list_for_each_entry_continue(iter, head, list) {
456 if (iter->timestamp > new->timestamp) {
457 list_add_tail(&new->list, &iter->list);
461 list_add_tail(&new->list, head);
464 /* The queue is ordered by time */
465 static void __queue_sample_event(struct sample_queue *new,
466 struct perf_session *s)
468 struct sample_queue *last_inserted = s->ordered_samples.last_inserted;
469 struct list_head *head = &s->ordered_samples.samples_head;
472 if (!last_inserted) {
473 __queue_sample_end(new, head);
478 * Most of the time the current event has a timestamp
479 * very close to the last event inserted, unless we just switched
480 * to another event buffer. Having a sorting based on a list and
481 * on the last inserted event that is close to the current one is
482 * probably more efficient than an rbtree based sorting.
484 if (last_inserted->timestamp >= new->timestamp)
485 __queue_sample_before(new, last_inserted, head);
487 __queue_sample_after(new, last_inserted, head);
490 static int queue_sample_event(event_t *event, struct sample_data *data,
491 struct perf_session *s)
493 u64 timestamp = data->time;
494 struct sample_queue *new;
497 if (timestamp < s->ordered_samples.last_flush) {
498 printf("Warning: Timestamp below last timeslice flush\n");
502 new = malloc(sizeof(*new));
506 new->timestamp = timestamp;
508 new->event = malloc(event->header.size);
514 memcpy(new->event, event, event->header.size);
516 __queue_sample_event(new, s);
517 s->ordered_samples.last_inserted = new;
519 if (new->timestamp > s->ordered_samples.max_timestamp)
520 s->ordered_samples.max_timestamp = new->timestamp;
525 static int perf_session__process_sample(event_t *event, struct perf_session *s,
526 struct perf_event_ops *ops)
528 struct sample_data data;
530 if (!ops->ordered_samples)
531 return ops->sample(event, s);
533 bzero(&data, sizeof(struct sample_data));
534 event__parse_sample(event, s->sample_type, &data);
536 queue_sample_event(event, &data, s);
541 static int perf_session__process_event(struct perf_session *self,
543 struct perf_event_ops *ops,
544 u64 offset, u64 head)
548 if (event->header.type < PERF_RECORD_HEADER_MAX) {
549 dump_printf("%#Lx [%#x]: PERF_RECORD_%s",
550 offset + head, event->header.size,
551 event__name[event->header.type]);
552 hists__inc_nr_events(&self->hists, event->header.type);
555 if (self->header.needs_swap && event__swap_ops[event->header.type])
556 event__swap_ops[event->header.type](event);
558 switch (event->header.type) {
559 case PERF_RECORD_SAMPLE:
560 return perf_session__process_sample(event, self, ops);
561 case PERF_RECORD_MMAP:
562 return ops->mmap(event, self);
563 case PERF_RECORD_COMM:
564 return ops->comm(event, self);
565 case PERF_RECORD_FORK:
566 return ops->fork(event, self);
567 case PERF_RECORD_EXIT:
568 return ops->exit(event, self);
569 case PERF_RECORD_LOST:
570 return ops->lost(event, self);
571 case PERF_RECORD_READ:
572 return ops->read(event, self);
573 case PERF_RECORD_THROTTLE:
574 return ops->throttle(event, self);
575 case PERF_RECORD_UNTHROTTLE:
576 return ops->unthrottle(event, self);
577 case PERF_RECORD_HEADER_ATTR:
578 return ops->attr(event, self);
579 case PERF_RECORD_HEADER_EVENT_TYPE:
580 return ops->event_type(event, self);
581 case PERF_RECORD_HEADER_TRACING_DATA:
582 /* setup for reading amidst mmap */
583 lseek(self->fd, offset + head, SEEK_SET);
584 return ops->tracing_data(event, self);
585 case PERF_RECORD_HEADER_BUILD_ID:
586 return ops->build_id(event, self);
587 case PERF_RECORD_FINISHED_ROUND:
588 return ops->finished_round(event, self, ops);
590 ++self->hists.stats.nr_unknown_events;
595 void perf_event_header__bswap(struct perf_event_header *self)
597 self->type = bswap_32(self->type);
598 self->misc = bswap_16(self->misc);
599 self->size = bswap_16(self->size);
602 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
604 struct thread *thread = perf_session__findnew(self, 0);
606 if (thread == NULL || thread__set_comm(thread, "swapper")) {
607 pr_err("problem inserting idle task.\n");
614 int do_read(int fd, void *buf, size_t size)
616 void *buf_start = buf;
619 int ret = read(fd, buf, size);
628 return buf - buf_start;
631 #define session_done() (*(volatile int *)(&session_done))
632 volatile int session_done;
634 static int __perf_session__process_pipe_events(struct perf_session *self,
635 struct perf_event_ops *ops)
644 perf_event_ops__fill_defaults(ops);
648 err = do_read(self->fd, &event, sizeof(struct perf_event_header));
653 pr_err("failed to read event header\n");
657 if (self->header.needs_swap)
658 perf_event_header__bswap(&event.header);
660 size = event.header.size;
665 p += sizeof(struct perf_event_header);
667 if (size - sizeof(struct perf_event_header)) {
668 err = do_read(self->fd, p,
669 size - sizeof(struct perf_event_header));
672 pr_err("unexpected end of event stream\n");
676 pr_err("failed to read event data\n");
682 (skip = perf_session__process_event(self, &event, ops,
684 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
685 head, event.header.size, event.header.type);
687 * assume we lost track of the stream, check alignment, and
688 * increment a single u64 in the hope to catch on again 'soon'.
690 if (unlikely(head & 7))
698 dump_printf("\n%#Lx [%#x]: event: %d\n",
699 head, event.header.size, event.header.type);
712 int __perf_session__process_events(struct perf_session *self,
713 u64 data_offset, u64 data_size,
714 u64 file_size, struct perf_event_ops *ops)
716 int err, mmap_prot, mmap_flags;
723 struct ui_progress *progress = ui_progress__new("Processing events...",
725 if (progress == NULL)
728 perf_event_ops__fill_defaults(ops);
730 page_size = sysconf(_SC_PAGESIZE);
733 shift = page_size * (head / page_size);
737 mmap_prot = PROT_READ;
738 mmap_flags = MAP_SHARED;
740 if (self->header.needs_swap) {
741 mmap_prot |= PROT_WRITE;
742 mmap_flags = MAP_PRIVATE;
745 buf = mmap(NULL, page_size * self->mmap_window, mmap_prot,
746 mmap_flags, self->fd, offset);
747 if (buf == MAP_FAILED) {
748 pr_err("failed to mmap file\n");
754 event = (event_t *)(buf + head);
755 ui_progress__update(progress, offset);
757 if (self->header.needs_swap)
758 perf_event_header__bswap(&event->header);
759 size = event->header.size;
763 if (head + event->header.size >= page_size * self->mmap_window) {
766 shift = page_size * (head / page_size);
768 munmap_ret = munmap(buf, page_size * self->mmap_window);
769 assert(munmap_ret == 0);
776 size = event->header.size;
778 dump_printf("\n%#Lx [%#x]: event: %d\n",
779 offset + head, event->header.size, event->header.type);
782 perf_session__process_event(self, event, ops, offset, head) < 0) {
783 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
784 offset + head, event->header.size,
787 * assume we lost track of the stream, check alignment, and
788 * increment a single u64 in the hope to catch on again 'soon'.
790 if (unlikely(head & 7))
798 if (offset + head >= data_offset + data_size)
801 if (offset + head < file_size)
805 /* do the final flush for ordered samples */
806 self->ordered_samples.next_flush = ULLONG_MAX;
807 flush_sample_queue(self, ops);
809 ui_progress__delete(progress);
813 int perf_session__process_events(struct perf_session *self,
814 struct perf_event_ops *ops)
818 if (perf_session__register_idle_thread(self) == NULL)
821 if (!symbol_conf.full_paths) {
824 if (getcwd(bf, sizeof(bf)) == NULL) {
827 pr_err("failed to get the current directory\n");
830 self->cwd = strdup(bf);
831 if (self->cwd == NULL) {
835 self->cwdlen = strlen(self->cwd);
839 err = __perf_session__process_events(self,
840 self->header.data_offset,
841 self->header.data_size,
844 err = __perf_session__process_pipe_events(self, ops);
849 bool perf_session__has_traces(struct perf_session *self, const char *msg)
851 if (!(self->sample_type & PERF_SAMPLE_RAW)) {
852 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
859 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
860 const char *symbol_name,
865 struct ref_reloc_sym *ref;
867 ref = zalloc(sizeof(struct ref_reloc_sym));
871 ref->name = strdup(symbol_name);
872 if (ref->name == NULL) {
877 bracket = strchr(ref->name, ']');
883 for (i = 0; i < MAP__NR_TYPES; ++i) {
884 struct kmap *kmap = map__kmap(maps[i]);
885 kmap->ref_reloc_sym = ref;
891 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
893 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
894 __dsos__fprintf(&self->host_machine.user_dsos, fp) +
895 machines__fprintf_dsos(&self->machines, fp);