Loading...
1#define _FILE_OFFSET_BITS 64
2
3#include <linux/kernel.h>
4
5#include <byteswap.h>
6#include <unistd.h>
7#include <sys/types.h>
8#include <sys/mman.h>
9
10#include "evlist.h"
11#include "evsel.h"
12#include "session.h"
13#include "sort.h"
14#include "util.h"
15#include "cpumap.h"
16
17static int perf_session__open(struct perf_session *self, bool force)
18{
19 struct stat input_stat;
20
21 if (!strcmp(self->filename, "-")) {
22 self->fd_pipe = true;
23 self->fd = STDIN_FILENO;
24
25 if (perf_session__read_header(self, self->fd) < 0)
26 pr_err("incompatible file format");
27
28 return 0;
29 }
30
31 self->fd = open(self->filename, O_RDONLY);
32 if (self->fd < 0) {
33 int err = errno;
34
35 pr_err("failed to open %s: %s", self->filename, strerror(err));
36 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
37 pr_err(" (try 'perf record' first)");
38 pr_err("\n");
39 return -errno;
40 }
41
42 if (fstat(self->fd, &input_stat) < 0)
43 goto out_close;
44
45 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
46 pr_err("file %s not owned by current user or root\n",
47 self->filename);
48 goto out_close;
49 }
50
51 if (!input_stat.st_size) {
52 pr_info("zero-sized file (%s), nothing to do!\n",
53 self->filename);
54 goto out_close;
55 }
56
57 if (perf_session__read_header(self, self->fd) < 0) {
58 pr_err("incompatible file format");
59 goto out_close;
60 }
61
62 if (!perf_evlist__valid_sample_type(self->evlist)) {
63 pr_err("non matching sample_type");
64 goto out_close;
65 }
66
67 if (!perf_evlist__valid_sample_id_all(self->evlist)) {
68 pr_err("non matching sample_id_all");
69 goto out_close;
70 }
71
72 self->size = input_stat.st_size;
73 return 0;
74
75out_close:
76 close(self->fd);
77 self->fd = -1;
78 return -1;
79}
80
81static void perf_session__id_header_size(struct perf_session *session)
82{
83 struct perf_sample *data;
84 u64 sample_type = session->sample_type;
85 u16 size = 0;
86
87 if (!session->sample_id_all)
88 goto out;
89
90 if (sample_type & PERF_SAMPLE_TID)
91 size += sizeof(data->tid) * 2;
92
93 if (sample_type & PERF_SAMPLE_TIME)
94 size += sizeof(data->time);
95
96 if (sample_type & PERF_SAMPLE_ID)
97 size += sizeof(data->id);
98
99 if (sample_type & PERF_SAMPLE_STREAM_ID)
100 size += sizeof(data->stream_id);
101
102 if (sample_type & PERF_SAMPLE_CPU)
103 size += sizeof(data->cpu) * 2;
104out:
105 session->id_hdr_size = size;
106}
107
108void perf_session__update_sample_type(struct perf_session *self)
109{
110 self->sample_type = perf_evlist__sample_type(self->evlist);
111 self->sample_size = __perf_evsel__sample_size(self->sample_type);
112 self->sample_id_all = perf_evlist__sample_id_all(self->evlist);
113 perf_session__id_header_size(self);
114}
115
116int perf_session__create_kernel_maps(struct perf_session *self)
117{
118 int ret = machine__create_kernel_maps(&self->host_machine);
119
120 if (ret >= 0)
121 ret = machines__create_guest_kernel_maps(&self->machines);
122 return ret;
123}
124
125static void perf_session__destroy_kernel_maps(struct perf_session *self)
126{
127 machine__destroy_kernel_maps(&self->host_machine);
128 machines__destroy_guest_kernel_maps(&self->machines);
129}
130
131struct perf_session *perf_session__new(const char *filename, int mode,
132 bool force, bool repipe,
133 struct perf_event_ops *ops)
134{
135 size_t len = filename ? strlen(filename) + 1 : 0;
136 struct perf_session *self = zalloc(sizeof(*self) + len);
137
138 if (self == NULL)
139 goto out;
140
141 memcpy(self->filename, filename, len);
142 self->threads = RB_ROOT;
143 INIT_LIST_HEAD(&self->dead_threads);
144 self->last_match = NULL;
145 /*
146 * On 64bit we can mmap the data file in one go. No need for tiny mmap
147 * slices. On 32bit we use 32MB.
148 */
149#if BITS_PER_LONG == 64
150 self->mmap_window = ULLONG_MAX;
151#else
152 self->mmap_window = 32 * 1024 * 1024ULL;
153#endif
154 self->machines = RB_ROOT;
155 self->repipe = repipe;
156 INIT_LIST_HEAD(&self->ordered_samples.samples);
157 INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
158 INIT_LIST_HEAD(&self->ordered_samples.to_free);
159 machine__init(&self->host_machine, "", HOST_KERNEL_ID);
160
161 if (mode == O_RDONLY) {
162 if (perf_session__open(self, force) < 0)
163 goto out_delete;
164 perf_session__update_sample_type(self);
165 } else if (mode == O_WRONLY) {
166 /*
167 * In O_RDONLY mode this will be performed when reading the
168 * kernel MMAP event, in perf_event__process_mmap().
169 */
170 if (perf_session__create_kernel_maps(self) < 0)
171 goto out_delete;
172 }
173
174 if (ops && ops->ordering_requires_timestamps &&
175 ops->ordered_samples && !self->sample_id_all) {
176 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
177 ops->ordered_samples = false;
178 }
179
180out:
181 return self;
182out_delete:
183 perf_session__delete(self);
184 return NULL;
185}
186
187static void perf_session__delete_dead_threads(struct perf_session *self)
188{
189 struct thread *n, *t;
190
191 list_for_each_entry_safe(t, n, &self->dead_threads, node) {
192 list_del(&t->node);
193 thread__delete(t);
194 }
195}
196
197static void perf_session__delete_threads(struct perf_session *self)
198{
199 struct rb_node *nd = rb_first(&self->threads);
200
201 while (nd) {
202 struct thread *t = rb_entry(nd, struct thread, rb_node);
203
204 rb_erase(&t->rb_node, &self->threads);
205 nd = rb_next(nd);
206 thread__delete(t);
207 }
208}
209
210void perf_session__delete(struct perf_session *self)
211{
212 perf_session__destroy_kernel_maps(self);
213 perf_session__delete_dead_threads(self);
214 perf_session__delete_threads(self);
215 machine__exit(&self->host_machine);
216 close(self->fd);
217 free(self);
218}
219
220void perf_session__remove_thread(struct perf_session *self, struct thread *th)
221{
222 self->last_match = NULL;
223 rb_erase(&th->rb_node, &self->threads);
224 /*
225 * We may have references to this thread, for instance in some hist_entry
226 * instances, so just move them to a separate list.
227 */
228 list_add_tail(&th->node, &self->dead_threads);
229}
230
231static bool symbol__match_parent_regex(struct symbol *sym)
232{
233 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
234 return 1;
235
236 return 0;
237}
238
239int perf_session__resolve_callchain(struct perf_session *self,
240 struct thread *thread,
241 struct ip_callchain *chain,
242 struct symbol **parent)
243{
244 u8 cpumode = PERF_RECORD_MISC_USER;
245 unsigned int i;
246 int err;
247
248 callchain_cursor_reset(&self->callchain_cursor);
249
250 for (i = 0; i < chain->nr; i++) {
251 u64 ip;
252 struct addr_location al;
253
254 if (callchain_param.order == ORDER_CALLEE)
255 ip = chain->ips[i];
256 else
257 ip = chain->ips[chain->nr - i - 1];
258
259 if (ip >= PERF_CONTEXT_MAX) {
260 switch (ip) {
261 case PERF_CONTEXT_HV:
262 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
263 case PERF_CONTEXT_KERNEL:
264 cpumode = PERF_RECORD_MISC_KERNEL; break;
265 case PERF_CONTEXT_USER:
266 cpumode = PERF_RECORD_MISC_USER; break;
267 default:
268 break;
269 }
270 continue;
271 }
272
273 al.filtered = false;
274 thread__find_addr_location(thread, self, cpumode,
275 MAP__FUNCTION, thread->pid, ip, &al, NULL);
276 if (al.sym != NULL) {
277 if (sort__has_parent && !*parent &&
278 symbol__match_parent_regex(al.sym))
279 *parent = al.sym;
280 if (!symbol_conf.use_callchain)
281 break;
282 }
283
284 err = callchain_cursor_append(&self->callchain_cursor,
285 ip, al.map, al.sym);
286 if (err)
287 return err;
288 }
289
290 return 0;
291}
292
293static int process_event_synth_stub(union perf_event *event __used,
294 struct perf_session *session __used)
295{
296 dump_printf(": unhandled!\n");
297 return 0;
298}
299
300static int process_event_sample_stub(union perf_event *event __used,
301 struct perf_sample *sample __used,
302 struct perf_evsel *evsel __used,
303 struct perf_session *session __used)
304{
305 dump_printf(": unhandled!\n");
306 return 0;
307}
308
309static int process_event_stub(union perf_event *event __used,
310 struct perf_sample *sample __used,
311 struct perf_session *session __used)
312{
313 dump_printf(": unhandled!\n");
314 return 0;
315}
316
317static int process_finished_round_stub(union perf_event *event __used,
318 struct perf_session *session __used,
319 struct perf_event_ops *ops __used)
320{
321 dump_printf(": unhandled!\n");
322 return 0;
323}
324
325static int process_finished_round(union perf_event *event,
326 struct perf_session *session,
327 struct perf_event_ops *ops);
328
329static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
330{
331 if (handler->sample == NULL)
332 handler->sample = process_event_sample_stub;
333 if (handler->mmap == NULL)
334 handler->mmap = process_event_stub;
335 if (handler->comm == NULL)
336 handler->comm = process_event_stub;
337 if (handler->fork == NULL)
338 handler->fork = process_event_stub;
339 if (handler->exit == NULL)
340 handler->exit = process_event_stub;
341 if (handler->lost == NULL)
342 handler->lost = perf_event__process_lost;
343 if (handler->read == NULL)
344 handler->read = process_event_stub;
345 if (handler->throttle == NULL)
346 handler->throttle = process_event_stub;
347 if (handler->unthrottle == NULL)
348 handler->unthrottle = process_event_stub;
349 if (handler->attr == NULL)
350 handler->attr = process_event_synth_stub;
351 if (handler->event_type == NULL)
352 handler->event_type = process_event_synth_stub;
353 if (handler->tracing_data == NULL)
354 handler->tracing_data = process_event_synth_stub;
355 if (handler->build_id == NULL)
356 handler->build_id = process_event_synth_stub;
357 if (handler->finished_round == NULL) {
358 if (handler->ordered_samples)
359 handler->finished_round = process_finished_round;
360 else
361 handler->finished_round = process_finished_round_stub;
362 }
363}
364
365void mem_bswap_64(void *src, int byte_size)
366{
367 u64 *m = src;
368
369 while (byte_size > 0) {
370 *m = bswap_64(*m);
371 byte_size -= sizeof(u64);
372 ++m;
373 }
374}
375
376static void perf_event__all64_swap(union perf_event *event)
377{
378 struct perf_event_header *hdr = &event->header;
379 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
380}
381
382static void perf_event__comm_swap(union perf_event *event)
383{
384 event->comm.pid = bswap_32(event->comm.pid);
385 event->comm.tid = bswap_32(event->comm.tid);
386}
387
388static void perf_event__mmap_swap(union perf_event *event)
389{
390 event->mmap.pid = bswap_32(event->mmap.pid);
391 event->mmap.tid = bswap_32(event->mmap.tid);
392 event->mmap.start = bswap_64(event->mmap.start);
393 event->mmap.len = bswap_64(event->mmap.len);
394 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
395}
396
397static void perf_event__task_swap(union perf_event *event)
398{
399 event->fork.pid = bswap_32(event->fork.pid);
400 event->fork.tid = bswap_32(event->fork.tid);
401 event->fork.ppid = bswap_32(event->fork.ppid);
402 event->fork.ptid = bswap_32(event->fork.ptid);
403 event->fork.time = bswap_64(event->fork.time);
404}
405
406static void perf_event__read_swap(union perf_event *event)
407{
408 event->read.pid = bswap_32(event->read.pid);
409 event->read.tid = bswap_32(event->read.tid);
410 event->read.value = bswap_64(event->read.value);
411 event->read.time_enabled = bswap_64(event->read.time_enabled);
412 event->read.time_running = bswap_64(event->read.time_running);
413 event->read.id = bswap_64(event->read.id);
414}
415
416/* exported for swapping attributes in file header */
417void perf_event__attr_swap(struct perf_event_attr *attr)
418{
419 attr->type = bswap_32(attr->type);
420 attr->size = bswap_32(attr->size);
421 attr->config = bswap_64(attr->config);
422 attr->sample_period = bswap_64(attr->sample_period);
423 attr->sample_type = bswap_64(attr->sample_type);
424 attr->read_format = bswap_64(attr->read_format);
425 attr->wakeup_events = bswap_32(attr->wakeup_events);
426 attr->bp_type = bswap_32(attr->bp_type);
427 attr->bp_addr = bswap_64(attr->bp_addr);
428 attr->bp_len = bswap_64(attr->bp_len);
429}
430
431static void perf_event__hdr_attr_swap(union perf_event *event)
432{
433 size_t size;
434
435 perf_event__attr_swap(&event->attr.attr);
436
437 size = event->header.size;
438 size -= (void *)&event->attr.id - (void *)event;
439 mem_bswap_64(event->attr.id, size);
440}
441
442static void perf_event__event_type_swap(union perf_event *event)
443{
444 event->event_type.event_type.event_id =
445 bswap_64(event->event_type.event_type.event_id);
446}
447
448static void perf_event__tracing_data_swap(union perf_event *event)
449{
450 event->tracing_data.size = bswap_32(event->tracing_data.size);
451}
452
453typedef void (*perf_event__swap_op)(union perf_event *event);
454
455static perf_event__swap_op perf_event__swap_ops[] = {
456 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
457 [PERF_RECORD_COMM] = perf_event__comm_swap,
458 [PERF_RECORD_FORK] = perf_event__task_swap,
459 [PERF_RECORD_EXIT] = perf_event__task_swap,
460 [PERF_RECORD_LOST] = perf_event__all64_swap,
461 [PERF_RECORD_READ] = perf_event__read_swap,
462 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
463 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
464 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
465 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
466 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
467 [PERF_RECORD_HEADER_MAX] = NULL,
468};
469
470struct sample_queue {
471 u64 timestamp;
472 u64 file_offset;
473 union perf_event *event;
474 struct list_head list;
475};
476
477static void perf_session_free_sample_buffers(struct perf_session *session)
478{
479 struct ordered_samples *os = &session->ordered_samples;
480
481 while (!list_empty(&os->to_free)) {
482 struct sample_queue *sq;
483
484 sq = list_entry(os->to_free.next, struct sample_queue, list);
485 list_del(&sq->list);
486 free(sq);
487 }
488}
489
490static int perf_session_deliver_event(struct perf_session *session,
491 union perf_event *event,
492 struct perf_sample *sample,
493 struct perf_event_ops *ops,
494 u64 file_offset);
495
496static void flush_sample_queue(struct perf_session *s,
497 struct perf_event_ops *ops)
498{
499 struct ordered_samples *os = &s->ordered_samples;
500 struct list_head *head = &os->samples;
501 struct sample_queue *tmp, *iter;
502 struct perf_sample sample;
503 u64 limit = os->next_flush;
504 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
505 int ret;
506
507 if (!ops->ordered_samples || !limit)
508 return;
509
510 list_for_each_entry_safe(iter, tmp, head, list) {
511 if (iter->timestamp > limit)
512 break;
513
514 ret = perf_session__parse_sample(s, iter->event, &sample);
515 if (ret)
516 pr_err("Can't parse sample, err = %d\n", ret);
517 else
518 perf_session_deliver_event(s, iter->event, &sample, ops,
519 iter->file_offset);
520
521 os->last_flush = iter->timestamp;
522 list_del(&iter->list);
523 list_add(&iter->list, &os->sample_cache);
524 }
525
526 if (list_empty(head)) {
527 os->last_sample = NULL;
528 } else if (last_ts <= limit) {
529 os->last_sample =
530 list_entry(head->prev, struct sample_queue, list);
531 }
532}
533
534/*
535 * When perf record finishes a pass on every buffers, it records this pseudo
536 * event.
537 * We record the max timestamp t found in the pass n.
538 * Assuming these timestamps are monotonic across cpus, we know that if
539 * a buffer still has events with timestamps below t, they will be all
540 * available and then read in the pass n + 1.
541 * Hence when we start to read the pass n + 2, we can safely flush every
542 * events with timestamps below t.
543 *
544 * ============ PASS n =================
545 * CPU 0 | CPU 1
546 * |
547 * cnt1 timestamps | cnt2 timestamps
548 * 1 | 2
549 * 2 | 3
550 * - | 4 <--- max recorded
551 *
552 * ============ PASS n + 1 ==============
553 * CPU 0 | CPU 1
554 * |
555 * cnt1 timestamps | cnt2 timestamps
556 * 3 | 5
557 * 4 | 6
558 * 5 | 7 <---- max recorded
559 *
560 * Flush every events below timestamp 4
561 *
562 * ============ PASS n + 2 ==============
563 * CPU 0 | CPU 1
564 * |
565 * cnt1 timestamps | cnt2 timestamps
566 * 6 | 8
567 * 7 | 9
568 * - | 10
569 *
570 * Flush every events below timestamp 7
571 * etc...
572 */
573static int process_finished_round(union perf_event *event __used,
574 struct perf_session *session,
575 struct perf_event_ops *ops)
576{
577 flush_sample_queue(session, ops);
578 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
579
580 return 0;
581}
582
583/* The queue is ordered by time */
584static void __queue_event(struct sample_queue *new, struct perf_session *s)
585{
586 struct ordered_samples *os = &s->ordered_samples;
587 struct sample_queue *sample = os->last_sample;
588 u64 timestamp = new->timestamp;
589 struct list_head *p;
590
591 os->last_sample = new;
592
593 if (!sample) {
594 list_add(&new->list, &os->samples);
595 os->max_timestamp = timestamp;
596 return;
597 }
598
599 /*
600 * last_sample might point to some random place in the list as it's
601 * the last queued event. We expect that the new event is close to
602 * this.
603 */
604 if (sample->timestamp <= timestamp) {
605 while (sample->timestamp <= timestamp) {
606 p = sample->list.next;
607 if (p == &os->samples) {
608 list_add_tail(&new->list, &os->samples);
609 os->max_timestamp = timestamp;
610 return;
611 }
612 sample = list_entry(p, struct sample_queue, list);
613 }
614 list_add_tail(&new->list, &sample->list);
615 } else {
616 while (sample->timestamp > timestamp) {
617 p = sample->list.prev;
618 if (p == &os->samples) {
619 list_add(&new->list, &os->samples);
620 return;
621 }
622 sample = list_entry(p, struct sample_queue, list);
623 }
624 list_add(&new->list, &sample->list);
625 }
626}
627
628#define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
629
630static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
631 struct perf_sample *sample, u64 file_offset)
632{
633 struct ordered_samples *os = &s->ordered_samples;
634 struct list_head *sc = &os->sample_cache;
635 u64 timestamp = sample->time;
636 struct sample_queue *new;
637
638 if (!timestamp || timestamp == ~0ULL)
639 return -ETIME;
640
641 if (timestamp < s->ordered_samples.last_flush) {
642 printf("Warning: Timestamp below last timeslice flush\n");
643 return -EINVAL;
644 }
645
646 if (!list_empty(sc)) {
647 new = list_entry(sc->next, struct sample_queue, list);
648 list_del(&new->list);
649 } else if (os->sample_buffer) {
650 new = os->sample_buffer + os->sample_buffer_idx;
651 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
652 os->sample_buffer = NULL;
653 } else {
654 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
655 if (!os->sample_buffer)
656 return -ENOMEM;
657 list_add(&os->sample_buffer->list, &os->to_free);
658 os->sample_buffer_idx = 2;
659 new = os->sample_buffer + 1;
660 }
661
662 new->timestamp = timestamp;
663 new->file_offset = file_offset;
664 new->event = event;
665
666 __queue_event(new, s);
667
668 return 0;
669}
670
671static void callchain__printf(struct perf_sample *sample)
672{
673 unsigned int i;
674
675 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
676
677 for (i = 0; i < sample->callchain->nr; i++)
678 printf("..... %2d: %016" PRIx64 "\n",
679 i, sample->callchain->ips[i]);
680}
681
682static void perf_session__print_tstamp(struct perf_session *session,
683 union perf_event *event,
684 struct perf_sample *sample)
685{
686 if (event->header.type != PERF_RECORD_SAMPLE &&
687 !session->sample_id_all) {
688 fputs("-1 -1 ", stdout);
689 return;
690 }
691
692 if ((session->sample_type & PERF_SAMPLE_CPU))
693 printf("%u ", sample->cpu);
694
695 if (session->sample_type & PERF_SAMPLE_TIME)
696 printf("%" PRIu64 " ", sample->time);
697}
698
699static void dump_event(struct perf_session *session, union perf_event *event,
700 u64 file_offset, struct perf_sample *sample)
701{
702 if (!dump_trace)
703 return;
704
705 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
706 file_offset, event->header.size, event->header.type);
707
708 trace_event(event);
709
710 if (sample)
711 perf_session__print_tstamp(session, event, sample);
712
713 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
714 event->header.size, perf_event__name(event->header.type));
715}
716
717static void dump_sample(struct perf_session *session, union perf_event *event,
718 struct perf_sample *sample)
719{
720 if (!dump_trace)
721 return;
722
723 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
724 event->header.misc, sample->pid, sample->tid, sample->ip,
725 sample->period, sample->addr);
726
727 if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
728 callchain__printf(sample);
729}
730
731static int perf_session_deliver_event(struct perf_session *session,
732 union perf_event *event,
733 struct perf_sample *sample,
734 struct perf_event_ops *ops,
735 u64 file_offset)
736{
737 struct perf_evsel *evsel;
738
739 dump_event(session, event, file_offset, sample);
740
741 switch (event->header.type) {
742 case PERF_RECORD_SAMPLE:
743 dump_sample(session, event, sample);
744 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
745 if (evsel == NULL) {
746 ++session->hists.stats.nr_unknown_id;
747 return -1;
748 }
749 return ops->sample(event, sample, evsel, session);
750 case PERF_RECORD_MMAP:
751 return ops->mmap(event, sample, session);
752 case PERF_RECORD_COMM:
753 return ops->comm(event, sample, session);
754 case PERF_RECORD_FORK:
755 return ops->fork(event, sample, session);
756 case PERF_RECORD_EXIT:
757 return ops->exit(event, sample, session);
758 case PERF_RECORD_LOST:
759 return ops->lost(event, sample, session);
760 case PERF_RECORD_READ:
761 return ops->read(event, sample, session);
762 case PERF_RECORD_THROTTLE:
763 return ops->throttle(event, sample, session);
764 case PERF_RECORD_UNTHROTTLE:
765 return ops->unthrottle(event, sample, session);
766 default:
767 ++session->hists.stats.nr_unknown_events;
768 return -1;
769 }
770}
771
772static int perf_session__preprocess_sample(struct perf_session *session,
773 union perf_event *event, struct perf_sample *sample)
774{
775 if (event->header.type != PERF_RECORD_SAMPLE ||
776 !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
777 return 0;
778
779 if (!ip_callchain__valid(sample->callchain, event)) {
780 pr_debug("call-chain problem with event, skipping it.\n");
781 ++session->hists.stats.nr_invalid_chains;
782 session->hists.stats.total_invalid_chains += sample->period;
783 return -EINVAL;
784 }
785 return 0;
786}
787
788static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
789 struct perf_event_ops *ops, u64 file_offset)
790{
791 dump_event(session, event, file_offset, NULL);
792
793 /* These events are processed right away */
794 switch (event->header.type) {
795 case PERF_RECORD_HEADER_ATTR:
796 return ops->attr(event, session);
797 case PERF_RECORD_HEADER_EVENT_TYPE:
798 return ops->event_type(event, session);
799 case PERF_RECORD_HEADER_TRACING_DATA:
800 /* setup for reading amidst mmap */
801 lseek(session->fd, file_offset, SEEK_SET);
802 return ops->tracing_data(event, session);
803 case PERF_RECORD_HEADER_BUILD_ID:
804 return ops->build_id(event, session);
805 case PERF_RECORD_FINISHED_ROUND:
806 return ops->finished_round(event, session, ops);
807 default:
808 return -EINVAL;
809 }
810}
811
812static int perf_session__process_event(struct perf_session *session,
813 union perf_event *event,
814 struct perf_event_ops *ops,
815 u64 file_offset)
816{
817 struct perf_sample sample;
818 int ret;
819
820 if (session->header.needs_swap &&
821 perf_event__swap_ops[event->header.type])
822 perf_event__swap_ops[event->header.type](event);
823
824 if (event->header.type >= PERF_RECORD_HEADER_MAX)
825 return -EINVAL;
826
827 hists__inc_nr_events(&session->hists, event->header.type);
828
829 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
830 return perf_session__process_user_event(session, event, ops, file_offset);
831
832 /*
833 * For all kernel events we get the sample data
834 */
835 ret = perf_session__parse_sample(session, event, &sample);
836 if (ret)
837 return ret;
838
839 /* Preprocess sample records - precheck callchains */
840 if (perf_session__preprocess_sample(session, event, &sample))
841 return 0;
842
843 if (ops->ordered_samples) {
844 ret = perf_session_queue_event(session, event, &sample,
845 file_offset);
846 if (ret != -ETIME)
847 return ret;
848 }
849
850 return perf_session_deliver_event(session, event, &sample, ops,
851 file_offset);
852}
853
854void perf_event_header__bswap(struct perf_event_header *self)
855{
856 self->type = bswap_32(self->type);
857 self->misc = bswap_16(self->misc);
858 self->size = bswap_16(self->size);
859}
860
861static struct thread *perf_session__register_idle_thread(struct perf_session *self)
862{
863 struct thread *thread = perf_session__findnew(self, 0);
864
865 if (thread == NULL || thread__set_comm(thread, "swapper")) {
866 pr_err("problem inserting idle task.\n");
867 thread = NULL;
868 }
869
870 return thread;
871}
872
873static void perf_session__warn_about_errors(const struct perf_session *session,
874 const struct perf_event_ops *ops)
875{
876 if (ops->lost == perf_event__process_lost &&
877 session->hists.stats.total_lost != 0) {
878 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64
879 "!\n\nCheck IO/CPU overload!\n\n",
880 session->hists.stats.total_period,
881 session->hists.stats.total_lost);
882 }
883
884 if (session->hists.stats.nr_unknown_events != 0) {
885 ui__warning("Found %u unknown events!\n\n"
886 "Is this an older tool processing a perf.data "
887 "file generated by a more recent tool?\n\n"
888 "If that is not the case, consider "
889 "reporting to linux-kernel@vger.kernel.org.\n\n",
890 session->hists.stats.nr_unknown_events);
891 }
892
893 if (session->hists.stats.nr_unknown_id != 0) {
894 ui__warning("%u samples with id not present in the header\n",
895 session->hists.stats.nr_unknown_id);
896 }
897
898 if (session->hists.stats.nr_invalid_chains != 0) {
899 ui__warning("Found invalid callchains!\n\n"
900 "%u out of %u events were discarded for this reason.\n\n"
901 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
902 session->hists.stats.nr_invalid_chains,
903 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
904 }
905}
906
907#define session_done() (*(volatile int *)(&session_done))
908volatile int session_done;
909
910static int __perf_session__process_pipe_events(struct perf_session *self,
911 struct perf_event_ops *ops)
912{
913 union perf_event event;
914 uint32_t size;
915 int skip = 0;
916 u64 head;
917 int err;
918 void *p;
919
920 perf_event_ops__fill_defaults(ops);
921
922 head = 0;
923more:
924 err = readn(self->fd, &event, sizeof(struct perf_event_header));
925 if (err <= 0) {
926 if (err == 0)
927 goto done;
928
929 pr_err("failed to read event header\n");
930 goto out_err;
931 }
932
933 if (self->header.needs_swap)
934 perf_event_header__bswap(&event.header);
935
936 size = event.header.size;
937 if (size == 0)
938 size = 8;
939
940 p = &event;
941 p += sizeof(struct perf_event_header);
942
943 if (size - sizeof(struct perf_event_header)) {
944 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
945 if (err <= 0) {
946 if (err == 0) {
947 pr_err("unexpected end of event stream\n");
948 goto done;
949 }
950
951 pr_err("failed to read event data\n");
952 goto out_err;
953 }
954 }
955
956 if (size == 0 ||
957 (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
958 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
959 head, event.header.size, event.header.type);
960 /*
961 * assume we lost track of the stream, check alignment, and
962 * increment a single u64 in the hope to catch on again 'soon'.
963 */
964 if (unlikely(head & 7))
965 head &= ~7ULL;
966
967 size = 8;
968 }
969
970 head += size;
971
972 if (skip > 0)
973 head += skip;
974
975 if (!session_done())
976 goto more;
977done:
978 err = 0;
979out_err:
980 perf_session__warn_about_errors(self, ops);
981 perf_session_free_sample_buffers(self);
982 return err;
983}
984
985static union perf_event *
986fetch_mmaped_event(struct perf_session *session,
987 u64 head, size_t mmap_size, char *buf)
988{
989 union perf_event *event;
990
991 /*
992 * Ensure we have enough space remaining to read
993 * the size of the event in the headers.
994 */
995 if (head + sizeof(event->header) > mmap_size)
996 return NULL;
997
998 event = (union perf_event *)(buf + head);
999
1000 if (session->header.needs_swap)
1001 perf_event_header__bswap(&event->header);
1002
1003 if (head + event->header.size > mmap_size)
1004 return NULL;
1005
1006 return event;
1007}
1008
1009int __perf_session__process_events(struct perf_session *session,
1010 u64 data_offset, u64 data_size,
1011 u64 file_size, struct perf_event_ops *ops)
1012{
1013 u64 head, page_offset, file_offset, file_pos, progress_next;
1014 int err, mmap_prot, mmap_flags, map_idx = 0;
1015 struct ui_progress *progress;
1016 size_t page_size, mmap_size;
1017 char *buf, *mmaps[8];
1018 union perf_event *event;
1019 uint32_t size;
1020
1021 perf_event_ops__fill_defaults(ops);
1022
1023 page_size = sysconf(_SC_PAGESIZE);
1024
1025 page_offset = page_size * (data_offset / page_size);
1026 file_offset = page_offset;
1027 head = data_offset - page_offset;
1028
1029 if (data_offset + data_size < file_size)
1030 file_size = data_offset + data_size;
1031
1032 progress_next = file_size / 16;
1033 progress = ui_progress__new("Processing events...", file_size);
1034 if (progress == NULL)
1035 return -1;
1036
1037 mmap_size = session->mmap_window;
1038 if (mmap_size > file_size)
1039 mmap_size = file_size;
1040
1041 memset(mmaps, 0, sizeof(mmaps));
1042
1043 mmap_prot = PROT_READ;
1044 mmap_flags = MAP_SHARED;
1045
1046 if (session->header.needs_swap) {
1047 mmap_prot |= PROT_WRITE;
1048 mmap_flags = MAP_PRIVATE;
1049 }
1050remap:
1051 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1052 file_offset);
1053 if (buf == MAP_FAILED) {
1054 pr_err("failed to mmap file\n");
1055 err = -errno;
1056 goto out_err;
1057 }
1058 mmaps[map_idx] = buf;
1059 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1060 file_pos = file_offset + head;
1061
1062more:
1063 event = fetch_mmaped_event(session, head, mmap_size, buf);
1064 if (!event) {
1065 if (mmaps[map_idx]) {
1066 munmap(mmaps[map_idx], mmap_size);
1067 mmaps[map_idx] = NULL;
1068 }
1069
1070 page_offset = page_size * (head / page_size);
1071 file_offset += page_offset;
1072 head -= page_offset;
1073 goto remap;
1074 }
1075
1076 size = event->header.size;
1077
1078 if (size == 0 ||
1079 perf_session__process_event(session, event, ops, file_pos) < 0) {
1080 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1081 file_offset + head, event->header.size,
1082 event->header.type);
1083 /*
1084 * assume we lost track of the stream, check alignment, and
1085 * increment a single u64 in the hope to catch on again 'soon'.
1086 */
1087 if (unlikely(head & 7))
1088 head &= ~7ULL;
1089
1090 size = 8;
1091 }
1092
1093 head += size;
1094 file_pos += size;
1095
1096 if (file_pos >= progress_next) {
1097 progress_next += file_size / 16;
1098 ui_progress__update(progress, file_pos);
1099 }
1100
1101 if (file_pos < file_size)
1102 goto more;
1103
1104 err = 0;
1105 /* do the final flush for ordered samples */
1106 session->ordered_samples.next_flush = ULLONG_MAX;
1107 flush_sample_queue(session, ops);
1108out_err:
1109 ui_progress__delete(progress);
1110 perf_session__warn_about_errors(session, ops);
1111 perf_session_free_sample_buffers(session);
1112 return err;
1113}
1114
1115int perf_session__process_events(struct perf_session *self,
1116 struct perf_event_ops *ops)
1117{
1118 int err;
1119
1120 if (perf_session__register_idle_thread(self) == NULL)
1121 return -ENOMEM;
1122
1123 if (!self->fd_pipe)
1124 err = __perf_session__process_events(self,
1125 self->header.data_offset,
1126 self->header.data_size,
1127 self->size, ops);
1128 else
1129 err = __perf_session__process_pipe_events(self, ops);
1130
1131 return err;
1132}
1133
1134bool perf_session__has_traces(struct perf_session *self, const char *msg)
1135{
1136 if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1137 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1138 return false;
1139 }
1140
1141 return true;
1142}
1143
1144int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1145 const char *symbol_name,
1146 u64 addr)
1147{
1148 char *bracket;
1149 enum map_type i;
1150 struct ref_reloc_sym *ref;
1151
1152 ref = zalloc(sizeof(struct ref_reloc_sym));
1153 if (ref == NULL)
1154 return -ENOMEM;
1155
1156 ref->name = strdup(symbol_name);
1157 if (ref->name == NULL) {
1158 free(ref);
1159 return -ENOMEM;
1160 }
1161
1162 bracket = strchr(ref->name, ']');
1163 if (bracket)
1164 *bracket = '\0';
1165
1166 ref->addr = addr;
1167
1168 for (i = 0; i < MAP__NR_TYPES; ++i) {
1169 struct kmap *kmap = map__kmap(maps[i]);
1170 kmap->ref_reloc_sym = ref;
1171 }
1172
1173 return 0;
1174}
1175
1176size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1177{
1178 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1179 __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1180 machines__fprintf_dsos(&self->machines, fp);
1181}
1182
1183size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1184 bool with_hits)
1185{
1186 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1187 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1188}
1189
1190size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1191{
1192 struct perf_evsel *pos;
1193 size_t ret = fprintf(fp, "Aggregated stats:\n");
1194
1195 ret += hists__fprintf_nr_events(&session->hists, fp);
1196
1197 list_for_each_entry(pos, &session->evlist->entries, node) {
1198 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1199 ret += hists__fprintf_nr_events(&pos->hists, fp);
1200 }
1201
1202 return ret;
1203}
1204
1205struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1206 unsigned int type)
1207{
1208 struct perf_evsel *pos;
1209
1210 list_for_each_entry(pos, &session->evlist->entries, node) {
1211 if (pos->attr.type == type)
1212 return pos;
1213 }
1214 return NULL;
1215}
1216
1217void perf_session__print_ip(union perf_event *event,
1218 struct perf_sample *sample,
1219 struct perf_session *session,
1220 int print_sym, int print_dso)
1221{
1222 struct addr_location al;
1223 const char *symname, *dsoname;
1224 struct callchain_cursor *cursor = &session->callchain_cursor;
1225 struct callchain_cursor_node *node;
1226
1227 if (perf_event__preprocess_sample(event, session, &al, sample,
1228 NULL) < 0) {
1229 error("problem processing %d event, skipping it.\n",
1230 event->header.type);
1231 return;
1232 }
1233
1234 if (symbol_conf.use_callchain && sample->callchain) {
1235
1236 if (perf_session__resolve_callchain(session, al.thread,
1237 sample->callchain, NULL) != 0) {
1238 if (verbose)
1239 error("Failed to resolve callchain. Skipping\n");
1240 return;
1241 }
1242 callchain_cursor_commit(cursor);
1243
1244 while (1) {
1245 node = callchain_cursor_current(cursor);
1246 if (!node)
1247 break;
1248
1249 printf("\t%16" PRIx64, node->ip);
1250 if (print_sym) {
1251 if (node->sym && node->sym->name)
1252 symname = node->sym->name;
1253 else
1254 symname = "";
1255
1256 printf(" %s", symname);
1257 }
1258 if (print_dso) {
1259 if (node->map && node->map->dso && node->map->dso->name)
1260 dsoname = node->map->dso->name;
1261 else
1262 dsoname = "";
1263
1264 printf(" (%s)", dsoname);
1265 }
1266 printf("\n");
1267
1268 callchain_cursor_advance(cursor);
1269 }
1270
1271 } else {
1272 printf("%16" PRIx64, sample->ip);
1273 if (print_sym) {
1274 if (al.sym && al.sym->name)
1275 symname = al.sym->name;
1276 else
1277 symname = "";
1278
1279 printf(" %s", symname);
1280 }
1281
1282 if (print_dso) {
1283 if (al.map && al.map->dso && al.map->dso->name)
1284 dsoname = al.map->dso->name;
1285 else
1286 dsoname = "";
1287
1288 printf(" (%s)", dsoname);
1289 }
1290 }
1291}
1292
1293int perf_session__cpu_bitmap(struct perf_session *session,
1294 const char *cpu_list, unsigned long *cpu_bitmap)
1295{
1296 int i;
1297 struct cpu_map *map;
1298
1299 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1300 struct perf_evsel *evsel;
1301
1302 evsel = perf_session__find_first_evtype(session, i);
1303 if (!evsel)
1304 continue;
1305
1306 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1307 pr_err("File does not contain CPU events. "
1308 "Remove -c option to proceed.\n");
1309 return -1;
1310 }
1311 }
1312
1313 map = cpu_map__new(cpu_list);
1314
1315 for (i = 0; i < map->nr; i++) {
1316 int cpu = map->map[i];
1317
1318 if (cpu >= MAX_NR_CPUS) {
1319 pr_err("Requested CPU %d too large. "
1320 "Consider raising MAX_NR_CPUS\n", cpu);
1321 return -1;
1322 }
1323
1324 set_bit(cpu, cpu_bitmap);
1325 }
1326
1327 return 0;
1328}
1#define _FILE_OFFSET_BITS 64
2
3#include <linux/kernel.h>
4
5#include <byteswap.h>
6#include <unistd.h>
7#include <sys/types.h>
8#include <sys/mman.h>
9
10#include "evlist.h"
11#include "evsel.h"
12#include "session.h"
13#include "tool.h"
14#include "sort.h"
15#include "util.h"
16#include "cpumap.h"
17
18static int perf_session__open(struct perf_session *self, bool force)
19{
20 struct stat input_stat;
21
22 if (!strcmp(self->filename, "-")) {
23 self->fd_pipe = true;
24 self->fd = STDIN_FILENO;
25
26 if (perf_session__read_header(self, self->fd) < 0)
27 pr_err("incompatible file format (rerun with -v to learn more)");
28
29 return 0;
30 }
31
32 self->fd = open(self->filename, O_RDONLY);
33 if (self->fd < 0) {
34 int err = errno;
35
36 pr_err("failed to open %s: %s", self->filename, strerror(err));
37 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
38 pr_err(" (try 'perf record' first)");
39 pr_err("\n");
40 return -errno;
41 }
42
43 if (fstat(self->fd, &input_stat) < 0)
44 goto out_close;
45
46 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
47 pr_err("file %s not owned by current user or root\n",
48 self->filename);
49 goto out_close;
50 }
51
52 if (!input_stat.st_size) {
53 pr_info("zero-sized file (%s), nothing to do!\n",
54 self->filename);
55 goto out_close;
56 }
57
58 if (perf_session__read_header(self, self->fd) < 0) {
59 pr_err("incompatible file format (rerun with -v to learn more)");
60 goto out_close;
61 }
62
63 if (!perf_evlist__valid_sample_type(self->evlist)) {
64 pr_err("non matching sample_type");
65 goto out_close;
66 }
67
68 if (!perf_evlist__valid_sample_id_all(self->evlist)) {
69 pr_err("non matching sample_id_all");
70 goto out_close;
71 }
72
73 self->size = input_stat.st_size;
74 return 0;
75
76out_close:
77 close(self->fd);
78 self->fd = -1;
79 return -1;
80}
81
82void perf_session__update_sample_type(struct perf_session *self)
83{
84 self->sample_type = perf_evlist__sample_type(self->evlist);
85 self->sample_size = __perf_evsel__sample_size(self->sample_type);
86 self->sample_id_all = perf_evlist__sample_id_all(self->evlist);
87 self->id_hdr_size = perf_evlist__id_hdr_size(self->evlist);
88 self->host_machine.id_hdr_size = self->id_hdr_size;
89}
90
91int perf_session__create_kernel_maps(struct perf_session *self)
92{
93 int ret = machine__create_kernel_maps(&self->host_machine);
94
95 if (ret >= 0)
96 ret = machines__create_guest_kernel_maps(&self->machines);
97 return ret;
98}
99
100static void perf_session__destroy_kernel_maps(struct perf_session *self)
101{
102 machine__destroy_kernel_maps(&self->host_machine);
103 machines__destroy_guest_kernel_maps(&self->machines);
104}
105
106struct perf_session *perf_session__new(const char *filename, int mode,
107 bool force, bool repipe,
108 struct perf_tool *tool)
109{
110 struct perf_session *self;
111 struct stat st;
112 size_t len;
113
114 if (!filename || !strlen(filename)) {
115 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
116 filename = "-";
117 else
118 filename = "perf.data";
119 }
120
121 len = strlen(filename);
122 self = zalloc(sizeof(*self) + len);
123
124 if (self == NULL)
125 goto out;
126
127 memcpy(self->filename, filename, len);
128 /*
129 * On 64bit we can mmap the data file in one go. No need for tiny mmap
130 * slices. On 32bit we use 32MB.
131 */
132#if BITS_PER_LONG == 64
133 self->mmap_window = ULLONG_MAX;
134#else
135 self->mmap_window = 32 * 1024 * 1024ULL;
136#endif
137 self->machines = RB_ROOT;
138 self->repipe = repipe;
139 INIT_LIST_HEAD(&self->ordered_samples.samples);
140 INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
141 INIT_LIST_HEAD(&self->ordered_samples.to_free);
142 machine__init(&self->host_machine, "", HOST_KERNEL_ID);
143 hists__init(&self->hists);
144
145 if (mode == O_RDONLY) {
146 if (perf_session__open(self, force) < 0)
147 goto out_delete;
148 perf_session__update_sample_type(self);
149 } else if (mode == O_WRONLY) {
150 /*
151 * In O_RDONLY mode this will be performed when reading the
152 * kernel MMAP event, in perf_event__process_mmap().
153 */
154 if (perf_session__create_kernel_maps(self) < 0)
155 goto out_delete;
156 }
157
158 if (tool && tool->ordering_requires_timestamps &&
159 tool->ordered_samples && !self->sample_id_all) {
160 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
161 tool->ordered_samples = false;
162 }
163
164out:
165 return self;
166out_delete:
167 perf_session__delete(self);
168 return NULL;
169}
170
171static void machine__delete_dead_threads(struct machine *machine)
172{
173 struct thread *n, *t;
174
175 list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
176 list_del(&t->node);
177 thread__delete(t);
178 }
179}
180
181static void perf_session__delete_dead_threads(struct perf_session *session)
182{
183 machine__delete_dead_threads(&session->host_machine);
184}
185
186static void machine__delete_threads(struct machine *self)
187{
188 struct rb_node *nd = rb_first(&self->threads);
189
190 while (nd) {
191 struct thread *t = rb_entry(nd, struct thread, rb_node);
192
193 rb_erase(&t->rb_node, &self->threads);
194 nd = rb_next(nd);
195 thread__delete(t);
196 }
197}
198
199static void perf_session__delete_threads(struct perf_session *session)
200{
201 machine__delete_threads(&session->host_machine);
202}
203
204void perf_session__delete(struct perf_session *self)
205{
206 perf_session__destroy_kernel_maps(self);
207 perf_session__delete_dead_threads(self);
208 perf_session__delete_threads(self);
209 machine__exit(&self->host_machine);
210 close(self->fd);
211 free(self);
212}
213
214void machine__remove_thread(struct machine *self, struct thread *th)
215{
216 self->last_match = NULL;
217 rb_erase(&th->rb_node, &self->threads);
218 /*
219 * We may have references to this thread, for instance in some hist_entry
220 * instances, so just move them to a separate list.
221 */
222 list_add_tail(&th->node, &self->dead_threads);
223}
224
225static bool symbol__match_parent_regex(struct symbol *sym)
226{
227 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
228 return 1;
229
230 return 0;
231}
232
233static const u8 cpumodes[] = {
234 PERF_RECORD_MISC_USER,
235 PERF_RECORD_MISC_KERNEL,
236 PERF_RECORD_MISC_GUEST_USER,
237 PERF_RECORD_MISC_GUEST_KERNEL
238};
239#define NCPUMODES (sizeof(cpumodes)/sizeof(u8))
240
241static void ip__resolve_ams(struct machine *self, struct thread *thread,
242 struct addr_map_symbol *ams,
243 u64 ip)
244{
245 struct addr_location al;
246 size_t i;
247 u8 m;
248
249 memset(&al, 0, sizeof(al));
250
251 for (i = 0; i < NCPUMODES; i++) {
252 m = cpumodes[i];
253 /*
254 * We cannot use the header.misc hint to determine whether a
255 * branch stack address is user, kernel, guest, hypervisor.
256 * Branches may straddle the kernel/user/hypervisor boundaries.
257 * Thus, we have to try consecutively until we find a match
258 * or else, the symbol is unknown
259 */
260 thread__find_addr_location(thread, self, m, MAP__FUNCTION,
261 ip, &al, NULL);
262 if (al.sym)
263 goto found;
264 }
265found:
266 ams->addr = ip;
267 ams->al_addr = al.addr;
268 ams->sym = al.sym;
269 ams->map = al.map;
270}
271
272struct branch_info *machine__resolve_bstack(struct machine *self,
273 struct thread *thr,
274 struct branch_stack *bs)
275{
276 struct branch_info *bi;
277 unsigned int i;
278
279 bi = calloc(bs->nr, sizeof(struct branch_info));
280 if (!bi)
281 return NULL;
282
283 for (i = 0; i < bs->nr; i++) {
284 ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to);
285 ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from);
286 bi[i].flags = bs->entries[i].flags;
287 }
288 return bi;
289}
290
291int machine__resolve_callchain(struct machine *self,
292 struct perf_evsel *evsel __used,
293 struct thread *thread,
294 struct ip_callchain *chain,
295 struct symbol **parent)
296{
297 u8 cpumode = PERF_RECORD_MISC_USER;
298 unsigned int i;
299 int err;
300
301 callchain_cursor_reset(&callchain_cursor);
302
303 if (chain->nr > PERF_MAX_STACK_DEPTH) {
304 pr_warning("corrupted callchain. skipping...\n");
305 return 0;
306 }
307
308 for (i = 0; i < chain->nr; i++) {
309 u64 ip;
310 struct addr_location al;
311
312 if (callchain_param.order == ORDER_CALLEE)
313 ip = chain->ips[i];
314 else
315 ip = chain->ips[chain->nr - i - 1];
316
317 if (ip >= PERF_CONTEXT_MAX) {
318 switch (ip) {
319 case PERF_CONTEXT_HV:
320 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
321 case PERF_CONTEXT_KERNEL:
322 cpumode = PERF_RECORD_MISC_KERNEL; break;
323 case PERF_CONTEXT_USER:
324 cpumode = PERF_RECORD_MISC_USER; break;
325 default:
326 pr_debug("invalid callchain context: "
327 "%"PRId64"\n", (s64) ip);
328 /*
329 * It seems the callchain is corrupted.
330 * Discard all.
331 */
332 callchain_cursor_reset(&callchain_cursor);
333 return 0;
334 }
335 continue;
336 }
337
338 al.filtered = false;
339 thread__find_addr_location(thread, self, cpumode,
340 MAP__FUNCTION, ip, &al, NULL);
341 if (al.sym != NULL) {
342 if (sort__has_parent && !*parent &&
343 symbol__match_parent_regex(al.sym))
344 *parent = al.sym;
345 if (!symbol_conf.use_callchain)
346 break;
347 }
348
349 err = callchain_cursor_append(&callchain_cursor,
350 ip, al.map, al.sym);
351 if (err)
352 return err;
353 }
354
355 return 0;
356}
357
358static int process_event_synth_tracing_data_stub(union perf_event *event __used,
359 struct perf_session *session __used)
360{
361 dump_printf(": unhandled!\n");
362 return 0;
363}
364
365static int process_event_synth_attr_stub(union perf_event *event __used,
366 struct perf_evlist **pevlist __used)
367{
368 dump_printf(": unhandled!\n");
369 return 0;
370}
371
372static int process_event_sample_stub(struct perf_tool *tool __used,
373 union perf_event *event __used,
374 struct perf_sample *sample __used,
375 struct perf_evsel *evsel __used,
376 struct machine *machine __used)
377{
378 dump_printf(": unhandled!\n");
379 return 0;
380}
381
382static int process_event_stub(struct perf_tool *tool __used,
383 union perf_event *event __used,
384 struct perf_sample *sample __used,
385 struct machine *machine __used)
386{
387 dump_printf(": unhandled!\n");
388 return 0;
389}
390
391static int process_finished_round_stub(struct perf_tool *tool __used,
392 union perf_event *event __used,
393 struct perf_session *perf_session __used)
394{
395 dump_printf(": unhandled!\n");
396 return 0;
397}
398
399static int process_event_type_stub(struct perf_tool *tool __used,
400 union perf_event *event __used)
401{
402 dump_printf(": unhandled!\n");
403 return 0;
404}
405
406static int process_finished_round(struct perf_tool *tool,
407 union perf_event *event,
408 struct perf_session *session);
409
410static void perf_tool__fill_defaults(struct perf_tool *tool)
411{
412 if (tool->sample == NULL)
413 tool->sample = process_event_sample_stub;
414 if (tool->mmap == NULL)
415 tool->mmap = process_event_stub;
416 if (tool->comm == NULL)
417 tool->comm = process_event_stub;
418 if (tool->fork == NULL)
419 tool->fork = process_event_stub;
420 if (tool->exit == NULL)
421 tool->exit = process_event_stub;
422 if (tool->lost == NULL)
423 tool->lost = perf_event__process_lost;
424 if (tool->read == NULL)
425 tool->read = process_event_sample_stub;
426 if (tool->throttle == NULL)
427 tool->throttle = process_event_stub;
428 if (tool->unthrottle == NULL)
429 tool->unthrottle = process_event_stub;
430 if (tool->attr == NULL)
431 tool->attr = process_event_synth_attr_stub;
432 if (tool->event_type == NULL)
433 tool->event_type = process_event_type_stub;
434 if (tool->tracing_data == NULL)
435 tool->tracing_data = process_event_synth_tracing_data_stub;
436 if (tool->build_id == NULL)
437 tool->build_id = process_finished_round_stub;
438 if (tool->finished_round == NULL) {
439 if (tool->ordered_samples)
440 tool->finished_round = process_finished_round;
441 else
442 tool->finished_round = process_finished_round_stub;
443 }
444}
445
446void mem_bswap_32(void *src, int byte_size)
447{
448 u32 *m = src;
449 while (byte_size > 0) {
450 *m = bswap_32(*m);
451 byte_size -= sizeof(u32);
452 ++m;
453 }
454}
455
456void mem_bswap_64(void *src, int byte_size)
457{
458 u64 *m = src;
459
460 while (byte_size > 0) {
461 *m = bswap_64(*m);
462 byte_size -= sizeof(u64);
463 ++m;
464 }
465}
466
467static void swap_sample_id_all(union perf_event *event, void *data)
468{
469 void *end = (void *) event + event->header.size;
470 int size = end - data;
471
472 BUG_ON(size % sizeof(u64));
473 mem_bswap_64(data, size);
474}
475
476static void perf_event__all64_swap(union perf_event *event,
477 bool sample_id_all __used)
478{
479 struct perf_event_header *hdr = &event->header;
480 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
481}
482
483static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
484{
485 event->comm.pid = bswap_32(event->comm.pid);
486 event->comm.tid = bswap_32(event->comm.tid);
487
488 if (sample_id_all) {
489 void *data = &event->comm.comm;
490
491 data += ALIGN(strlen(data) + 1, sizeof(u64));
492 swap_sample_id_all(event, data);
493 }
494}
495
496static void perf_event__mmap_swap(union perf_event *event,
497 bool sample_id_all)
498{
499 event->mmap.pid = bswap_32(event->mmap.pid);
500 event->mmap.tid = bswap_32(event->mmap.tid);
501 event->mmap.start = bswap_64(event->mmap.start);
502 event->mmap.len = bswap_64(event->mmap.len);
503 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
504
505 if (sample_id_all) {
506 void *data = &event->mmap.filename;
507
508 data += ALIGN(strlen(data) + 1, sizeof(u64));
509 swap_sample_id_all(event, data);
510 }
511}
512
513static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
514{
515 event->fork.pid = bswap_32(event->fork.pid);
516 event->fork.tid = bswap_32(event->fork.tid);
517 event->fork.ppid = bswap_32(event->fork.ppid);
518 event->fork.ptid = bswap_32(event->fork.ptid);
519 event->fork.time = bswap_64(event->fork.time);
520
521 if (sample_id_all)
522 swap_sample_id_all(event, &event->fork + 1);
523}
524
525static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
526{
527 event->read.pid = bswap_32(event->read.pid);
528 event->read.tid = bswap_32(event->read.tid);
529 event->read.value = bswap_64(event->read.value);
530 event->read.time_enabled = bswap_64(event->read.time_enabled);
531 event->read.time_running = bswap_64(event->read.time_running);
532 event->read.id = bswap_64(event->read.id);
533
534 if (sample_id_all)
535 swap_sample_id_all(event, &event->read + 1);
536}
537
538static u8 revbyte(u8 b)
539{
540 int rev = (b >> 4) | ((b & 0xf) << 4);
541 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
542 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
543 return (u8) rev;
544}
545
546/*
547 * XXX this is hack in attempt to carry flags bitfield
548 * throught endian village. ABI says:
549 *
550 * Bit-fields are allocated from right to left (least to most significant)
551 * on little-endian implementations and from left to right (most to least
552 * significant) on big-endian implementations.
553 *
554 * The above seems to be byte specific, so we need to reverse each
555 * byte of the bitfield. 'Internet' also says this might be implementation
556 * specific and we probably need proper fix and carry perf_event_attr
557 * bitfield flags in separate data file FEAT_ section. Thought this seems
558 * to work for now.
559 */
560static void swap_bitfield(u8 *p, unsigned len)
561{
562 unsigned i;
563
564 for (i = 0; i < len; i++) {
565 *p = revbyte(*p);
566 p++;
567 }
568}
569
570/* exported for swapping attributes in file header */
571void perf_event__attr_swap(struct perf_event_attr *attr)
572{
573 attr->type = bswap_32(attr->type);
574 attr->size = bswap_32(attr->size);
575 attr->config = bswap_64(attr->config);
576 attr->sample_period = bswap_64(attr->sample_period);
577 attr->sample_type = bswap_64(attr->sample_type);
578 attr->read_format = bswap_64(attr->read_format);
579 attr->wakeup_events = bswap_32(attr->wakeup_events);
580 attr->bp_type = bswap_32(attr->bp_type);
581 attr->bp_addr = bswap_64(attr->bp_addr);
582 attr->bp_len = bswap_64(attr->bp_len);
583
584 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
585}
586
587static void perf_event__hdr_attr_swap(union perf_event *event,
588 bool sample_id_all __used)
589{
590 size_t size;
591
592 perf_event__attr_swap(&event->attr.attr);
593
594 size = event->header.size;
595 size -= (void *)&event->attr.id - (void *)event;
596 mem_bswap_64(event->attr.id, size);
597}
598
599static void perf_event__event_type_swap(union perf_event *event,
600 bool sample_id_all __used)
601{
602 event->event_type.event_type.event_id =
603 bswap_64(event->event_type.event_type.event_id);
604}
605
606static void perf_event__tracing_data_swap(union perf_event *event,
607 bool sample_id_all __used)
608{
609 event->tracing_data.size = bswap_32(event->tracing_data.size);
610}
611
612typedef void (*perf_event__swap_op)(union perf_event *event,
613 bool sample_id_all);
614
615static perf_event__swap_op perf_event__swap_ops[] = {
616 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
617 [PERF_RECORD_COMM] = perf_event__comm_swap,
618 [PERF_RECORD_FORK] = perf_event__task_swap,
619 [PERF_RECORD_EXIT] = perf_event__task_swap,
620 [PERF_RECORD_LOST] = perf_event__all64_swap,
621 [PERF_RECORD_READ] = perf_event__read_swap,
622 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
623 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
624 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
625 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
626 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
627 [PERF_RECORD_HEADER_MAX] = NULL,
628};
629
630struct sample_queue {
631 u64 timestamp;
632 u64 file_offset;
633 union perf_event *event;
634 struct list_head list;
635};
636
637static void perf_session_free_sample_buffers(struct perf_session *session)
638{
639 struct ordered_samples *os = &session->ordered_samples;
640
641 while (!list_empty(&os->to_free)) {
642 struct sample_queue *sq;
643
644 sq = list_entry(os->to_free.next, struct sample_queue, list);
645 list_del(&sq->list);
646 free(sq);
647 }
648}
649
650static int perf_session_deliver_event(struct perf_session *session,
651 union perf_event *event,
652 struct perf_sample *sample,
653 struct perf_tool *tool,
654 u64 file_offset);
655
656static void flush_sample_queue(struct perf_session *s,
657 struct perf_tool *tool)
658{
659 struct ordered_samples *os = &s->ordered_samples;
660 struct list_head *head = &os->samples;
661 struct sample_queue *tmp, *iter;
662 struct perf_sample sample;
663 u64 limit = os->next_flush;
664 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
665 unsigned idx = 0, progress_next = os->nr_samples / 16;
666 int ret;
667
668 if (!tool->ordered_samples || !limit)
669 return;
670
671 list_for_each_entry_safe(iter, tmp, head, list) {
672 if (iter->timestamp > limit)
673 break;
674
675 ret = perf_session__parse_sample(s, iter->event, &sample);
676 if (ret)
677 pr_err("Can't parse sample, err = %d\n", ret);
678 else
679 perf_session_deliver_event(s, iter->event, &sample, tool,
680 iter->file_offset);
681
682 os->last_flush = iter->timestamp;
683 list_del(&iter->list);
684 list_add(&iter->list, &os->sample_cache);
685 if (++idx >= progress_next) {
686 progress_next += os->nr_samples / 16;
687 ui_progress__update(idx, os->nr_samples,
688 "Processing time ordered events...");
689 }
690 }
691
692 if (list_empty(head)) {
693 os->last_sample = NULL;
694 } else if (last_ts <= limit) {
695 os->last_sample =
696 list_entry(head->prev, struct sample_queue, list);
697 }
698
699 os->nr_samples = 0;
700}
701
702/*
703 * When perf record finishes a pass on every buffers, it records this pseudo
704 * event.
705 * We record the max timestamp t found in the pass n.
706 * Assuming these timestamps are monotonic across cpus, we know that if
707 * a buffer still has events with timestamps below t, they will be all
708 * available and then read in the pass n + 1.
709 * Hence when we start to read the pass n + 2, we can safely flush every
710 * events with timestamps below t.
711 *
712 * ============ PASS n =================
713 * CPU 0 | CPU 1
714 * |
715 * cnt1 timestamps | cnt2 timestamps
716 * 1 | 2
717 * 2 | 3
718 * - | 4 <--- max recorded
719 *
720 * ============ PASS n + 1 ==============
721 * CPU 0 | CPU 1
722 * |
723 * cnt1 timestamps | cnt2 timestamps
724 * 3 | 5
725 * 4 | 6
726 * 5 | 7 <---- max recorded
727 *
728 * Flush every events below timestamp 4
729 *
730 * ============ PASS n + 2 ==============
731 * CPU 0 | CPU 1
732 * |
733 * cnt1 timestamps | cnt2 timestamps
734 * 6 | 8
735 * 7 | 9
736 * - | 10
737 *
738 * Flush every events below timestamp 7
739 * etc...
740 */
741static int process_finished_round(struct perf_tool *tool,
742 union perf_event *event __used,
743 struct perf_session *session)
744{
745 flush_sample_queue(session, tool);
746 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
747
748 return 0;
749}
750
751/* The queue is ordered by time */
752static void __queue_event(struct sample_queue *new, struct perf_session *s)
753{
754 struct ordered_samples *os = &s->ordered_samples;
755 struct sample_queue *sample = os->last_sample;
756 u64 timestamp = new->timestamp;
757 struct list_head *p;
758
759 ++os->nr_samples;
760 os->last_sample = new;
761
762 if (!sample) {
763 list_add(&new->list, &os->samples);
764 os->max_timestamp = timestamp;
765 return;
766 }
767
768 /*
769 * last_sample might point to some random place in the list as it's
770 * the last queued event. We expect that the new event is close to
771 * this.
772 */
773 if (sample->timestamp <= timestamp) {
774 while (sample->timestamp <= timestamp) {
775 p = sample->list.next;
776 if (p == &os->samples) {
777 list_add_tail(&new->list, &os->samples);
778 os->max_timestamp = timestamp;
779 return;
780 }
781 sample = list_entry(p, struct sample_queue, list);
782 }
783 list_add_tail(&new->list, &sample->list);
784 } else {
785 while (sample->timestamp > timestamp) {
786 p = sample->list.prev;
787 if (p == &os->samples) {
788 list_add(&new->list, &os->samples);
789 return;
790 }
791 sample = list_entry(p, struct sample_queue, list);
792 }
793 list_add(&new->list, &sample->list);
794 }
795}
796
797#define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
798
799static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
800 struct perf_sample *sample, u64 file_offset)
801{
802 struct ordered_samples *os = &s->ordered_samples;
803 struct list_head *sc = &os->sample_cache;
804 u64 timestamp = sample->time;
805 struct sample_queue *new;
806
807 if (!timestamp || timestamp == ~0ULL)
808 return -ETIME;
809
810 if (timestamp < s->ordered_samples.last_flush) {
811 printf("Warning: Timestamp below last timeslice flush\n");
812 return -EINVAL;
813 }
814
815 if (!list_empty(sc)) {
816 new = list_entry(sc->next, struct sample_queue, list);
817 list_del(&new->list);
818 } else if (os->sample_buffer) {
819 new = os->sample_buffer + os->sample_buffer_idx;
820 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
821 os->sample_buffer = NULL;
822 } else {
823 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
824 if (!os->sample_buffer)
825 return -ENOMEM;
826 list_add(&os->sample_buffer->list, &os->to_free);
827 os->sample_buffer_idx = 2;
828 new = os->sample_buffer + 1;
829 }
830
831 new->timestamp = timestamp;
832 new->file_offset = file_offset;
833 new->event = event;
834
835 __queue_event(new, s);
836
837 return 0;
838}
839
840static void callchain__printf(struct perf_sample *sample)
841{
842 unsigned int i;
843
844 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
845
846 for (i = 0; i < sample->callchain->nr; i++)
847 printf("..... %2d: %016" PRIx64 "\n",
848 i, sample->callchain->ips[i]);
849}
850
851static void branch_stack__printf(struct perf_sample *sample)
852{
853 uint64_t i;
854
855 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
856
857 for (i = 0; i < sample->branch_stack->nr; i++)
858 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
859 i, sample->branch_stack->entries[i].from,
860 sample->branch_stack->entries[i].to);
861}
862
863static void perf_session__print_tstamp(struct perf_session *session,
864 union perf_event *event,
865 struct perf_sample *sample)
866{
867 if (event->header.type != PERF_RECORD_SAMPLE &&
868 !session->sample_id_all) {
869 fputs("-1 -1 ", stdout);
870 return;
871 }
872
873 if ((session->sample_type & PERF_SAMPLE_CPU))
874 printf("%u ", sample->cpu);
875
876 if (session->sample_type & PERF_SAMPLE_TIME)
877 printf("%" PRIu64 " ", sample->time);
878}
879
880static void dump_event(struct perf_session *session, union perf_event *event,
881 u64 file_offset, struct perf_sample *sample)
882{
883 if (!dump_trace)
884 return;
885
886 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
887 file_offset, event->header.size, event->header.type);
888
889 trace_event(event);
890
891 if (sample)
892 perf_session__print_tstamp(session, event, sample);
893
894 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
895 event->header.size, perf_event__name(event->header.type));
896}
897
898static void dump_sample(struct perf_session *session, union perf_event *event,
899 struct perf_sample *sample)
900{
901 if (!dump_trace)
902 return;
903
904 printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
905 event->header.misc, sample->pid, sample->tid, sample->ip,
906 sample->period, sample->addr);
907
908 if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
909 callchain__printf(sample);
910
911 if (session->sample_type & PERF_SAMPLE_BRANCH_STACK)
912 branch_stack__printf(sample);
913}
914
915static struct machine *
916 perf_session__find_machine_for_cpumode(struct perf_session *session,
917 union perf_event *event)
918{
919 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
920
921 if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) {
922 u32 pid;
923
924 if (event->header.type == PERF_RECORD_MMAP)
925 pid = event->mmap.pid;
926 else
927 pid = event->ip.pid;
928
929 return perf_session__findnew_machine(session, pid);
930 }
931
932 return perf_session__find_host_machine(session);
933}
934
935static int perf_session_deliver_event(struct perf_session *session,
936 union perf_event *event,
937 struct perf_sample *sample,
938 struct perf_tool *tool,
939 u64 file_offset)
940{
941 struct perf_evsel *evsel;
942 struct machine *machine;
943
944 dump_event(session, event, file_offset, sample);
945
946 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
947 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
948 /*
949 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
950 * because the tools right now may apply filters, discarding
951 * some of the samples. For consistency, in the future we
952 * should have something like nr_filtered_samples and remove
953 * the sample->period from total_sample_period, etc, KISS for
954 * now tho.
955 *
956 * Also testing against NULL allows us to handle files without
957 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
958 * future probably it'll be a good idea to restrict event
959 * processing via perf_session to files with both set.
960 */
961 hists__inc_nr_events(&evsel->hists, event->header.type);
962 }
963
964 machine = perf_session__find_machine_for_cpumode(session, event);
965
966 switch (event->header.type) {
967 case PERF_RECORD_SAMPLE:
968 dump_sample(session, event, sample);
969 if (evsel == NULL) {
970 ++session->hists.stats.nr_unknown_id;
971 return 0;
972 }
973 if (machine == NULL) {
974 ++session->hists.stats.nr_unprocessable_samples;
975 return 0;
976 }
977 return tool->sample(tool, event, sample, evsel, machine);
978 case PERF_RECORD_MMAP:
979 return tool->mmap(tool, event, sample, machine);
980 case PERF_RECORD_COMM:
981 return tool->comm(tool, event, sample, machine);
982 case PERF_RECORD_FORK:
983 return tool->fork(tool, event, sample, machine);
984 case PERF_RECORD_EXIT:
985 return tool->exit(tool, event, sample, machine);
986 case PERF_RECORD_LOST:
987 if (tool->lost == perf_event__process_lost)
988 session->hists.stats.total_lost += event->lost.lost;
989 return tool->lost(tool, event, sample, machine);
990 case PERF_RECORD_READ:
991 return tool->read(tool, event, sample, evsel, machine);
992 case PERF_RECORD_THROTTLE:
993 return tool->throttle(tool, event, sample, machine);
994 case PERF_RECORD_UNTHROTTLE:
995 return tool->unthrottle(tool, event, sample, machine);
996 default:
997 ++session->hists.stats.nr_unknown_events;
998 return -1;
999 }
1000}
1001
1002static int perf_session__preprocess_sample(struct perf_session *session,
1003 union perf_event *event, struct perf_sample *sample)
1004{
1005 if (event->header.type != PERF_RECORD_SAMPLE ||
1006 !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
1007 return 0;
1008
1009 if (!ip_callchain__valid(sample->callchain, event)) {
1010 pr_debug("call-chain problem with event, skipping it.\n");
1011 ++session->hists.stats.nr_invalid_chains;
1012 session->hists.stats.total_invalid_chains += sample->period;
1013 return -EINVAL;
1014 }
1015 return 0;
1016}
1017
1018static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1019 struct perf_tool *tool, u64 file_offset)
1020{
1021 int err;
1022
1023 dump_event(session, event, file_offset, NULL);
1024
1025 /* These events are processed right away */
1026 switch (event->header.type) {
1027 case PERF_RECORD_HEADER_ATTR:
1028 err = tool->attr(event, &session->evlist);
1029 if (err == 0)
1030 perf_session__update_sample_type(session);
1031 return err;
1032 case PERF_RECORD_HEADER_EVENT_TYPE:
1033 return tool->event_type(tool, event);
1034 case PERF_RECORD_HEADER_TRACING_DATA:
1035 /* setup for reading amidst mmap */
1036 lseek(session->fd, file_offset, SEEK_SET);
1037 return tool->tracing_data(event, session);
1038 case PERF_RECORD_HEADER_BUILD_ID:
1039 return tool->build_id(tool, event, session);
1040 case PERF_RECORD_FINISHED_ROUND:
1041 return tool->finished_round(tool, event, session);
1042 default:
1043 return -EINVAL;
1044 }
1045}
1046
1047static void event_swap(union perf_event *event, bool sample_id_all)
1048{
1049 perf_event__swap_op swap;
1050
1051 swap = perf_event__swap_ops[event->header.type];
1052 if (swap)
1053 swap(event, sample_id_all);
1054}
1055
1056static int perf_session__process_event(struct perf_session *session,
1057 union perf_event *event,
1058 struct perf_tool *tool,
1059 u64 file_offset)
1060{
1061 struct perf_sample sample;
1062 int ret;
1063
1064 if (session->header.needs_swap)
1065 event_swap(event, session->sample_id_all);
1066
1067 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1068 return -EINVAL;
1069
1070 hists__inc_nr_events(&session->hists, event->header.type);
1071
1072 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1073 return perf_session__process_user_event(session, event, tool, file_offset);
1074
1075 /*
1076 * For all kernel events we get the sample data
1077 */
1078 ret = perf_session__parse_sample(session, event, &sample);
1079 if (ret)
1080 return ret;
1081
1082 /* Preprocess sample records - precheck callchains */
1083 if (perf_session__preprocess_sample(session, event, &sample))
1084 return 0;
1085
1086 if (tool->ordered_samples) {
1087 ret = perf_session_queue_event(session, event, &sample,
1088 file_offset);
1089 if (ret != -ETIME)
1090 return ret;
1091 }
1092
1093 return perf_session_deliver_event(session, event, &sample, tool,
1094 file_offset);
1095}
1096
1097void perf_event_header__bswap(struct perf_event_header *self)
1098{
1099 self->type = bswap_32(self->type);
1100 self->misc = bswap_16(self->misc);
1101 self->size = bswap_16(self->size);
1102}
1103
1104struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1105{
1106 return machine__findnew_thread(&session->host_machine, pid);
1107}
1108
1109static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1110{
1111 struct thread *thread = perf_session__findnew(self, 0);
1112
1113 if (thread == NULL || thread__set_comm(thread, "swapper")) {
1114 pr_err("problem inserting idle task.\n");
1115 thread = NULL;
1116 }
1117
1118 return thread;
1119}
1120
1121static void perf_session__warn_about_errors(const struct perf_session *session,
1122 const struct perf_tool *tool)
1123{
1124 if (tool->lost == perf_event__process_lost &&
1125 session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1126 ui__warning("Processed %d events and lost %d chunks!\n\n"
1127 "Check IO/CPU overload!\n\n",
1128 session->hists.stats.nr_events[0],
1129 session->hists.stats.nr_events[PERF_RECORD_LOST]);
1130 }
1131
1132 if (session->hists.stats.nr_unknown_events != 0) {
1133 ui__warning("Found %u unknown events!\n\n"
1134 "Is this an older tool processing a perf.data "
1135 "file generated by a more recent tool?\n\n"
1136 "If that is not the case, consider "
1137 "reporting to linux-kernel@vger.kernel.org.\n\n",
1138 session->hists.stats.nr_unknown_events);
1139 }
1140
1141 if (session->hists.stats.nr_unknown_id != 0) {
1142 ui__warning("%u samples with id not present in the header\n",
1143 session->hists.stats.nr_unknown_id);
1144 }
1145
1146 if (session->hists.stats.nr_invalid_chains != 0) {
1147 ui__warning("Found invalid callchains!\n\n"
1148 "%u out of %u events were discarded for this reason.\n\n"
1149 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1150 session->hists.stats.nr_invalid_chains,
1151 session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1152 }
1153
1154 if (session->hists.stats.nr_unprocessable_samples != 0) {
1155 ui__warning("%u unprocessable samples recorded.\n"
1156 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1157 session->hists.stats.nr_unprocessable_samples);
1158 }
1159}
1160
1161#define session_done() (*(volatile int *)(&session_done))
1162volatile int session_done;
1163
1164static int __perf_session__process_pipe_events(struct perf_session *self,
1165 struct perf_tool *tool)
1166{
1167 union perf_event *event;
1168 uint32_t size, cur_size = 0;
1169 void *buf = NULL;
1170 int skip = 0;
1171 u64 head;
1172 int err;
1173 void *p;
1174
1175 perf_tool__fill_defaults(tool);
1176
1177 head = 0;
1178 cur_size = sizeof(union perf_event);
1179
1180 buf = malloc(cur_size);
1181 if (!buf)
1182 return -errno;
1183more:
1184 event = buf;
1185 err = readn(self->fd, event, sizeof(struct perf_event_header));
1186 if (err <= 0) {
1187 if (err == 0)
1188 goto done;
1189
1190 pr_err("failed to read event header\n");
1191 goto out_err;
1192 }
1193
1194 if (self->header.needs_swap)
1195 perf_event_header__bswap(&event->header);
1196
1197 size = event->header.size;
1198 if (size == 0)
1199 size = 8;
1200
1201 if (size > cur_size) {
1202 void *new = realloc(buf, size);
1203 if (!new) {
1204 pr_err("failed to allocate memory to read event\n");
1205 goto out_err;
1206 }
1207 buf = new;
1208 cur_size = size;
1209 event = buf;
1210 }
1211 p = event;
1212 p += sizeof(struct perf_event_header);
1213
1214 if (size - sizeof(struct perf_event_header)) {
1215 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1216 if (err <= 0) {
1217 if (err == 0) {
1218 pr_err("unexpected end of event stream\n");
1219 goto done;
1220 }
1221
1222 pr_err("failed to read event data\n");
1223 goto out_err;
1224 }
1225 }
1226
1227 if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1228 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1229 head, event->header.size, event->header.type);
1230 err = -EINVAL;
1231 goto out_err;
1232 }
1233
1234 head += size;
1235
1236 if (skip > 0)
1237 head += skip;
1238
1239 if (!session_done())
1240 goto more;
1241done:
1242 err = 0;
1243out_err:
1244 free(buf);
1245 perf_session__warn_about_errors(self, tool);
1246 perf_session_free_sample_buffers(self);
1247 return err;
1248}
1249
1250static union perf_event *
1251fetch_mmaped_event(struct perf_session *session,
1252 u64 head, size_t mmap_size, char *buf)
1253{
1254 union perf_event *event;
1255
1256 /*
1257 * Ensure we have enough space remaining to read
1258 * the size of the event in the headers.
1259 */
1260 if (head + sizeof(event->header) > mmap_size)
1261 return NULL;
1262
1263 event = (union perf_event *)(buf + head);
1264
1265 if (session->header.needs_swap)
1266 perf_event_header__bswap(&event->header);
1267
1268 if (head + event->header.size > mmap_size)
1269 return NULL;
1270
1271 return event;
1272}
1273
1274int __perf_session__process_events(struct perf_session *session,
1275 u64 data_offset, u64 data_size,
1276 u64 file_size, struct perf_tool *tool)
1277{
1278 u64 head, page_offset, file_offset, file_pos, progress_next;
1279 int err, mmap_prot, mmap_flags, map_idx = 0;
1280 size_t page_size, mmap_size;
1281 char *buf, *mmaps[8];
1282 union perf_event *event;
1283 uint32_t size;
1284
1285 perf_tool__fill_defaults(tool);
1286
1287 page_size = sysconf(_SC_PAGESIZE);
1288
1289 page_offset = page_size * (data_offset / page_size);
1290 file_offset = page_offset;
1291 head = data_offset - page_offset;
1292
1293 if (data_offset + data_size < file_size)
1294 file_size = data_offset + data_size;
1295
1296 progress_next = file_size / 16;
1297
1298 mmap_size = session->mmap_window;
1299 if (mmap_size > file_size)
1300 mmap_size = file_size;
1301
1302 memset(mmaps, 0, sizeof(mmaps));
1303
1304 mmap_prot = PROT_READ;
1305 mmap_flags = MAP_SHARED;
1306
1307 if (session->header.needs_swap) {
1308 mmap_prot |= PROT_WRITE;
1309 mmap_flags = MAP_PRIVATE;
1310 }
1311remap:
1312 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1313 file_offset);
1314 if (buf == MAP_FAILED) {
1315 pr_err("failed to mmap file\n");
1316 err = -errno;
1317 goto out_err;
1318 }
1319 mmaps[map_idx] = buf;
1320 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1321 file_pos = file_offset + head;
1322
1323more:
1324 event = fetch_mmaped_event(session, head, mmap_size, buf);
1325 if (!event) {
1326 if (mmaps[map_idx]) {
1327 munmap(mmaps[map_idx], mmap_size);
1328 mmaps[map_idx] = NULL;
1329 }
1330
1331 page_offset = page_size * (head / page_size);
1332 file_offset += page_offset;
1333 head -= page_offset;
1334 goto remap;
1335 }
1336
1337 size = event->header.size;
1338
1339 if (size == 0 ||
1340 perf_session__process_event(session, event, tool, file_pos) < 0) {
1341 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1342 file_offset + head, event->header.size,
1343 event->header.type);
1344 err = -EINVAL;
1345 goto out_err;
1346 }
1347
1348 head += size;
1349 file_pos += size;
1350
1351 if (file_pos >= progress_next) {
1352 progress_next += file_size / 16;
1353 ui_progress__update(file_pos, file_size,
1354 "Processing events...");
1355 }
1356
1357 if (file_pos < file_size)
1358 goto more;
1359
1360 err = 0;
1361 /* do the final flush for ordered samples */
1362 session->ordered_samples.next_flush = ULLONG_MAX;
1363 flush_sample_queue(session, tool);
1364out_err:
1365 perf_session__warn_about_errors(session, tool);
1366 perf_session_free_sample_buffers(session);
1367 return err;
1368}
1369
1370int perf_session__process_events(struct perf_session *self,
1371 struct perf_tool *tool)
1372{
1373 int err;
1374
1375 if (perf_session__register_idle_thread(self) == NULL)
1376 return -ENOMEM;
1377
1378 if (!self->fd_pipe)
1379 err = __perf_session__process_events(self,
1380 self->header.data_offset,
1381 self->header.data_size,
1382 self->size, tool);
1383 else
1384 err = __perf_session__process_pipe_events(self, tool);
1385
1386 return err;
1387}
1388
1389bool perf_session__has_traces(struct perf_session *self, const char *msg)
1390{
1391 if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1392 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1393 return false;
1394 }
1395
1396 return true;
1397}
1398
1399int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1400 const char *symbol_name, u64 addr)
1401{
1402 char *bracket;
1403 enum map_type i;
1404 struct ref_reloc_sym *ref;
1405
1406 ref = zalloc(sizeof(struct ref_reloc_sym));
1407 if (ref == NULL)
1408 return -ENOMEM;
1409
1410 ref->name = strdup(symbol_name);
1411 if (ref->name == NULL) {
1412 free(ref);
1413 return -ENOMEM;
1414 }
1415
1416 bracket = strchr(ref->name, ']');
1417 if (bracket)
1418 *bracket = '\0';
1419
1420 ref->addr = addr;
1421
1422 for (i = 0; i < MAP__NR_TYPES; ++i) {
1423 struct kmap *kmap = map__kmap(maps[i]);
1424 kmap->ref_reloc_sym = ref;
1425 }
1426
1427 return 0;
1428}
1429
1430size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1431{
1432 return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1433 __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1434 machines__fprintf_dsos(&self->machines, fp);
1435}
1436
1437size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1438 bool with_hits)
1439{
1440 size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1441 return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1442}
1443
1444size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1445{
1446 struct perf_evsel *pos;
1447 size_t ret = fprintf(fp, "Aggregated stats:\n");
1448
1449 ret += hists__fprintf_nr_events(&session->hists, fp);
1450
1451 list_for_each_entry(pos, &session->evlist->entries, node) {
1452 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1453 ret += hists__fprintf_nr_events(&pos->hists, fp);
1454 }
1455
1456 return ret;
1457}
1458
1459size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1460{
1461 /*
1462 * FIXME: Here we have to actually print all the machines in this
1463 * session, not just the host...
1464 */
1465 return machine__fprintf(&session->host_machine, fp);
1466}
1467
1468void perf_session__remove_thread(struct perf_session *session,
1469 struct thread *th)
1470{
1471 /*
1472 * FIXME: This one makes no sense, we need to remove the thread from
1473 * the machine it belongs to, perf_session can have many machines, so
1474 * doing it always on ->host_machine is wrong. Fix when auditing all
1475 * the 'perf kvm' code.
1476 */
1477 machine__remove_thread(&session->host_machine, th);
1478}
1479
1480struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1481 unsigned int type)
1482{
1483 struct perf_evsel *pos;
1484
1485 list_for_each_entry(pos, &session->evlist->entries, node) {
1486 if (pos->attr.type == type)
1487 return pos;
1488 }
1489 return NULL;
1490}
1491
1492void perf_event__print_ip(union perf_event *event, struct perf_sample *sample,
1493 struct machine *machine, struct perf_evsel *evsel,
1494 int print_sym, int print_dso, int print_symoffset)
1495{
1496 struct addr_location al;
1497 struct callchain_cursor_node *node;
1498
1499 if (perf_event__preprocess_sample(event, machine, &al, sample,
1500 NULL) < 0) {
1501 error("problem processing %d event, skipping it.\n",
1502 event->header.type);
1503 return;
1504 }
1505
1506 if (symbol_conf.use_callchain && sample->callchain) {
1507
1508 if (machine__resolve_callchain(machine, evsel, al.thread,
1509 sample->callchain, NULL) != 0) {
1510 if (verbose)
1511 error("Failed to resolve callchain. Skipping\n");
1512 return;
1513 }
1514 callchain_cursor_commit(&callchain_cursor);
1515
1516 while (1) {
1517 node = callchain_cursor_current(&callchain_cursor);
1518 if (!node)
1519 break;
1520
1521 printf("\t%16" PRIx64, node->ip);
1522 if (print_sym) {
1523 printf(" ");
1524 symbol__fprintf_symname(node->sym, stdout);
1525 }
1526 if (print_dso) {
1527 printf(" (");
1528 map__fprintf_dsoname(node->map, stdout);
1529 printf(")");
1530 }
1531 printf("\n");
1532
1533 callchain_cursor_advance(&callchain_cursor);
1534 }
1535
1536 } else {
1537 printf("%16" PRIx64, sample->ip);
1538 if (print_sym) {
1539 printf(" ");
1540 if (print_symoffset)
1541 symbol__fprintf_symname_offs(al.sym, &al,
1542 stdout);
1543 else
1544 symbol__fprintf_symname(al.sym, stdout);
1545 }
1546
1547 if (print_dso) {
1548 printf(" (");
1549 map__fprintf_dsoname(al.map, stdout);
1550 printf(")");
1551 }
1552 }
1553}
1554
1555int perf_session__cpu_bitmap(struct perf_session *session,
1556 const char *cpu_list, unsigned long *cpu_bitmap)
1557{
1558 int i;
1559 struct cpu_map *map;
1560
1561 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1562 struct perf_evsel *evsel;
1563
1564 evsel = perf_session__find_first_evtype(session, i);
1565 if (!evsel)
1566 continue;
1567
1568 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1569 pr_err("File does not contain CPU events. "
1570 "Remove -c option to proceed.\n");
1571 return -1;
1572 }
1573 }
1574
1575 map = cpu_map__new(cpu_list);
1576 if (map == NULL) {
1577 pr_err("Invalid cpu_list\n");
1578 return -1;
1579 }
1580
1581 for (i = 0; i < map->nr; i++) {
1582 int cpu = map->map[i];
1583
1584 if (cpu >= MAX_NR_CPUS) {
1585 pr_err("Requested CPU %d too large. "
1586 "Consider raising MAX_NR_CPUS\n", cpu);
1587 return -1;
1588 }
1589
1590 set_bit(cpu, cpu_bitmap);
1591 }
1592
1593 return 0;
1594}
1595
1596void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1597 bool full)
1598{
1599 struct stat st;
1600 int ret;
1601
1602 if (session == NULL || fp == NULL)
1603 return;
1604
1605 ret = fstat(session->fd, &st);
1606 if (ret == -1)
1607 return;
1608
1609 fprintf(fp, "# ========\n");
1610 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1611 perf_header__fprintf_info(session, fp, full);
1612 fprintf(fp, "# ========\n#\n");
1613}