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#include <linux/kernel.h>
2#include <traceevent/event-parse.h>
3
4#include <byteswap.h>
5#include <unistd.h>
6#include <sys/types.h>
7#include <sys/mman.h>
8
9#include "evlist.h"
10#include "evsel.h"
11#include "session.h"
12#include "tool.h"
13#include "sort.h"
14#include "util.h"
15#include "cpumap.h"
16#include "perf_regs.h"
17#include "vdso.h"
18
19static int perf_session__open(struct perf_session *session)
20{
21 struct perf_data_file *file = session->file;
22
23 if (perf_session__read_header(session) < 0) {
24 pr_err("incompatible file format (rerun with -v to learn more)");
25 return -1;
26 }
27
28 if (perf_data_file__is_pipe(file))
29 return 0;
30
31 if (!perf_evlist__valid_sample_type(session->evlist)) {
32 pr_err("non matching sample_type");
33 return -1;
34 }
35
36 if (!perf_evlist__valid_sample_id_all(session->evlist)) {
37 pr_err("non matching sample_id_all");
38 return -1;
39 }
40
41 if (!perf_evlist__valid_read_format(session->evlist)) {
42 pr_err("non matching read_format");
43 return -1;
44 }
45
46 return 0;
47}
48
49void perf_session__set_id_hdr_size(struct perf_session *session)
50{
51 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
52
53 machines__set_id_hdr_size(&session->machines, id_hdr_size);
54}
55
56int perf_session__create_kernel_maps(struct perf_session *session)
57{
58 int ret = machine__create_kernel_maps(&session->machines.host);
59
60 if (ret >= 0)
61 ret = machines__create_guest_kernel_maps(&session->machines);
62 return ret;
63}
64
65static void perf_session__destroy_kernel_maps(struct perf_session *session)
66{
67 machines__destroy_kernel_maps(&session->machines);
68}
69
70struct perf_session *perf_session__new(struct perf_data_file *file,
71 bool repipe, struct perf_tool *tool)
72{
73 struct perf_session *session = zalloc(sizeof(*session));
74
75 if (!session)
76 goto out;
77
78 session->repipe = repipe;
79 INIT_LIST_HEAD(&session->ordered_samples.samples);
80 INIT_LIST_HEAD(&session->ordered_samples.sample_cache);
81 INIT_LIST_HEAD(&session->ordered_samples.to_free);
82 machines__init(&session->machines);
83
84 if (file) {
85 if (perf_data_file__open(file))
86 goto out_delete;
87
88 session->file = file;
89
90 if (perf_data_file__is_read(file)) {
91 if (perf_session__open(session) < 0)
92 goto out_close;
93
94 perf_session__set_id_hdr_size(session);
95 }
96 }
97
98 if (!file || perf_data_file__is_write(file)) {
99 /*
100 * In O_RDONLY mode this will be performed when reading the
101 * kernel MMAP event, in perf_event__process_mmap().
102 */
103 if (perf_session__create_kernel_maps(session) < 0)
104 goto out_delete;
105 }
106
107 if (tool && tool->ordering_requires_timestamps &&
108 tool->ordered_samples && !perf_evlist__sample_id_all(session->evlist)) {
109 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
110 tool->ordered_samples = false;
111 }
112
113 return session;
114
115 out_close:
116 perf_data_file__close(file);
117 out_delete:
118 perf_session__delete(session);
119 out:
120 return NULL;
121}
122
123static void perf_session__delete_dead_threads(struct perf_session *session)
124{
125 machine__delete_dead_threads(&session->machines.host);
126}
127
128static void perf_session__delete_threads(struct perf_session *session)
129{
130 machine__delete_threads(&session->machines.host);
131}
132
133static void perf_session_env__delete(struct perf_session_env *env)
134{
135 zfree(&env->hostname);
136 zfree(&env->os_release);
137 zfree(&env->version);
138 zfree(&env->arch);
139 zfree(&env->cpu_desc);
140 zfree(&env->cpuid);
141
142 zfree(&env->cmdline);
143 zfree(&env->sibling_cores);
144 zfree(&env->sibling_threads);
145 zfree(&env->numa_nodes);
146 zfree(&env->pmu_mappings);
147}
148
149void perf_session__delete(struct perf_session *session)
150{
151 perf_session__destroy_kernel_maps(session);
152 perf_session__delete_dead_threads(session);
153 perf_session__delete_threads(session);
154 perf_session_env__delete(&session->header.env);
155 machines__exit(&session->machines);
156 if (session->file)
157 perf_data_file__close(session->file);
158 free(session);
159 vdso__exit();
160}
161
162static int process_event_synth_tracing_data_stub(struct perf_tool *tool
163 __maybe_unused,
164 union perf_event *event
165 __maybe_unused,
166 struct perf_session *session
167 __maybe_unused)
168{
169 dump_printf(": unhandled!\n");
170 return 0;
171}
172
173static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
174 union perf_event *event __maybe_unused,
175 struct perf_evlist **pevlist
176 __maybe_unused)
177{
178 dump_printf(": unhandled!\n");
179 return 0;
180}
181
182static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
183 union perf_event *event __maybe_unused,
184 struct perf_sample *sample __maybe_unused,
185 struct perf_evsel *evsel __maybe_unused,
186 struct machine *machine __maybe_unused)
187{
188 dump_printf(": unhandled!\n");
189 return 0;
190}
191
192static int process_event_stub(struct perf_tool *tool __maybe_unused,
193 union perf_event *event __maybe_unused,
194 struct perf_sample *sample __maybe_unused,
195 struct machine *machine __maybe_unused)
196{
197 dump_printf(": unhandled!\n");
198 return 0;
199}
200
201static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
202 union perf_event *event __maybe_unused,
203 struct perf_session *perf_session
204 __maybe_unused)
205{
206 dump_printf(": unhandled!\n");
207 return 0;
208}
209
210static int process_finished_round(struct perf_tool *tool,
211 union perf_event *event,
212 struct perf_session *session);
213
214void perf_tool__fill_defaults(struct perf_tool *tool)
215{
216 if (tool->sample == NULL)
217 tool->sample = process_event_sample_stub;
218 if (tool->mmap == NULL)
219 tool->mmap = process_event_stub;
220 if (tool->mmap2 == NULL)
221 tool->mmap2 = process_event_stub;
222 if (tool->comm == NULL)
223 tool->comm = process_event_stub;
224 if (tool->fork == NULL)
225 tool->fork = process_event_stub;
226 if (tool->exit == NULL)
227 tool->exit = process_event_stub;
228 if (tool->lost == NULL)
229 tool->lost = perf_event__process_lost;
230 if (tool->read == NULL)
231 tool->read = process_event_sample_stub;
232 if (tool->throttle == NULL)
233 tool->throttle = process_event_stub;
234 if (tool->unthrottle == NULL)
235 tool->unthrottle = process_event_stub;
236 if (tool->attr == NULL)
237 tool->attr = process_event_synth_attr_stub;
238 if (tool->tracing_data == NULL)
239 tool->tracing_data = process_event_synth_tracing_data_stub;
240 if (tool->build_id == NULL)
241 tool->build_id = process_finished_round_stub;
242 if (tool->finished_round == NULL) {
243 if (tool->ordered_samples)
244 tool->finished_round = process_finished_round;
245 else
246 tool->finished_round = process_finished_round_stub;
247 }
248}
249
250static void swap_sample_id_all(union perf_event *event, void *data)
251{
252 void *end = (void *) event + event->header.size;
253 int size = end - data;
254
255 BUG_ON(size % sizeof(u64));
256 mem_bswap_64(data, size);
257}
258
259static void perf_event__all64_swap(union perf_event *event,
260 bool sample_id_all __maybe_unused)
261{
262 struct perf_event_header *hdr = &event->header;
263 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
264}
265
266static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
267{
268 event->comm.pid = bswap_32(event->comm.pid);
269 event->comm.tid = bswap_32(event->comm.tid);
270
271 if (sample_id_all) {
272 void *data = &event->comm.comm;
273
274 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
275 swap_sample_id_all(event, data);
276 }
277}
278
279static void perf_event__mmap_swap(union perf_event *event,
280 bool sample_id_all)
281{
282 event->mmap.pid = bswap_32(event->mmap.pid);
283 event->mmap.tid = bswap_32(event->mmap.tid);
284 event->mmap.start = bswap_64(event->mmap.start);
285 event->mmap.len = bswap_64(event->mmap.len);
286 event->mmap.pgoff = bswap_64(event->mmap.pgoff);
287
288 if (sample_id_all) {
289 void *data = &event->mmap.filename;
290
291 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
292 swap_sample_id_all(event, data);
293 }
294}
295
296static void perf_event__mmap2_swap(union perf_event *event,
297 bool sample_id_all)
298{
299 event->mmap2.pid = bswap_32(event->mmap2.pid);
300 event->mmap2.tid = bswap_32(event->mmap2.tid);
301 event->mmap2.start = bswap_64(event->mmap2.start);
302 event->mmap2.len = bswap_64(event->mmap2.len);
303 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
304 event->mmap2.maj = bswap_32(event->mmap2.maj);
305 event->mmap2.min = bswap_32(event->mmap2.min);
306 event->mmap2.ino = bswap_64(event->mmap2.ino);
307
308 if (sample_id_all) {
309 void *data = &event->mmap2.filename;
310
311 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
312 swap_sample_id_all(event, data);
313 }
314}
315static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
316{
317 event->fork.pid = bswap_32(event->fork.pid);
318 event->fork.tid = bswap_32(event->fork.tid);
319 event->fork.ppid = bswap_32(event->fork.ppid);
320 event->fork.ptid = bswap_32(event->fork.ptid);
321 event->fork.time = bswap_64(event->fork.time);
322
323 if (sample_id_all)
324 swap_sample_id_all(event, &event->fork + 1);
325}
326
327static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
328{
329 event->read.pid = bswap_32(event->read.pid);
330 event->read.tid = bswap_32(event->read.tid);
331 event->read.value = bswap_64(event->read.value);
332 event->read.time_enabled = bswap_64(event->read.time_enabled);
333 event->read.time_running = bswap_64(event->read.time_running);
334 event->read.id = bswap_64(event->read.id);
335
336 if (sample_id_all)
337 swap_sample_id_all(event, &event->read + 1);
338}
339
340static void perf_event__throttle_swap(union perf_event *event,
341 bool sample_id_all)
342{
343 event->throttle.time = bswap_64(event->throttle.time);
344 event->throttle.id = bswap_64(event->throttle.id);
345 event->throttle.stream_id = bswap_64(event->throttle.stream_id);
346
347 if (sample_id_all)
348 swap_sample_id_all(event, &event->throttle + 1);
349}
350
351static u8 revbyte(u8 b)
352{
353 int rev = (b >> 4) | ((b & 0xf) << 4);
354 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
355 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
356 return (u8) rev;
357}
358
359/*
360 * XXX this is hack in attempt to carry flags bitfield
361 * throught endian village. ABI says:
362 *
363 * Bit-fields are allocated from right to left (least to most significant)
364 * on little-endian implementations and from left to right (most to least
365 * significant) on big-endian implementations.
366 *
367 * The above seems to be byte specific, so we need to reverse each
368 * byte of the bitfield. 'Internet' also says this might be implementation
369 * specific and we probably need proper fix and carry perf_event_attr
370 * bitfield flags in separate data file FEAT_ section. Thought this seems
371 * to work for now.
372 */
373static void swap_bitfield(u8 *p, unsigned len)
374{
375 unsigned i;
376
377 for (i = 0; i < len; i++) {
378 *p = revbyte(*p);
379 p++;
380 }
381}
382
383/* exported for swapping attributes in file header */
384void perf_event__attr_swap(struct perf_event_attr *attr)
385{
386 attr->type = bswap_32(attr->type);
387 attr->size = bswap_32(attr->size);
388 attr->config = bswap_64(attr->config);
389 attr->sample_period = bswap_64(attr->sample_period);
390 attr->sample_type = bswap_64(attr->sample_type);
391 attr->read_format = bswap_64(attr->read_format);
392 attr->wakeup_events = bswap_32(attr->wakeup_events);
393 attr->bp_type = bswap_32(attr->bp_type);
394 attr->bp_addr = bswap_64(attr->bp_addr);
395 attr->bp_len = bswap_64(attr->bp_len);
396 attr->branch_sample_type = bswap_64(attr->branch_sample_type);
397 attr->sample_regs_user = bswap_64(attr->sample_regs_user);
398 attr->sample_stack_user = bswap_32(attr->sample_stack_user);
399
400 swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
401}
402
403static void perf_event__hdr_attr_swap(union perf_event *event,
404 bool sample_id_all __maybe_unused)
405{
406 size_t size;
407
408 perf_event__attr_swap(&event->attr.attr);
409
410 size = event->header.size;
411 size -= (void *)&event->attr.id - (void *)event;
412 mem_bswap_64(event->attr.id, size);
413}
414
415static void perf_event__event_type_swap(union perf_event *event,
416 bool sample_id_all __maybe_unused)
417{
418 event->event_type.event_type.event_id =
419 bswap_64(event->event_type.event_type.event_id);
420}
421
422static void perf_event__tracing_data_swap(union perf_event *event,
423 bool sample_id_all __maybe_unused)
424{
425 event->tracing_data.size = bswap_32(event->tracing_data.size);
426}
427
428typedef void (*perf_event__swap_op)(union perf_event *event,
429 bool sample_id_all);
430
431static perf_event__swap_op perf_event__swap_ops[] = {
432 [PERF_RECORD_MMAP] = perf_event__mmap_swap,
433 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap,
434 [PERF_RECORD_COMM] = perf_event__comm_swap,
435 [PERF_RECORD_FORK] = perf_event__task_swap,
436 [PERF_RECORD_EXIT] = perf_event__task_swap,
437 [PERF_RECORD_LOST] = perf_event__all64_swap,
438 [PERF_RECORD_READ] = perf_event__read_swap,
439 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap,
440 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap,
441 [PERF_RECORD_SAMPLE] = perf_event__all64_swap,
442 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap,
443 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap,
444 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
445 [PERF_RECORD_HEADER_BUILD_ID] = NULL,
446 [PERF_RECORD_HEADER_MAX] = NULL,
447};
448
449struct sample_queue {
450 u64 timestamp;
451 u64 file_offset;
452 union perf_event *event;
453 struct list_head list;
454};
455
456static void perf_session_free_sample_buffers(struct perf_session *session)
457{
458 struct ordered_samples *os = &session->ordered_samples;
459
460 while (!list_empty(&os->to_free)) {
461 struct sample_queue *sq;
462
463 sq = list_entry(os->to_free.next, struct sample_queue, list);
464 list_del(&sq->list);
465 free(sq);
466 }
467}
468
469static int perf_session_deliver_event(struct perf_session *session,
470 union perf_event *event,
471 struct perf_sample *sample,
472 struct perf_tool *tool,
473 u64 file_offset);
474
475static int flush_sample_queue(struct perf_session *s,
476 struct perf_tool *tool)
477{
478 struct ordered_samples *os = &s->ordered_samples;
479 struct list_head *head = &os->samples;
480 struct sample_queue *tmp, *iter;
481 struct perf_sample sample;
482 u64 limit = os->next_flush;
483 u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
484 bool show_progress = limit == ULLONG_MAX;
485 struct ui_progress prog;
486 int ret;
487
488 if (!tool->ordered_samples || !limit)
489 return 0;
490
491 if (show_progress)
492 ui_progress__init(&prog, os->nr_samples, "Processing time ordered events...");
493
494 list_for_each_entry_safe(iter, tmp, head, list) {
495 if (session_done())
496 return 0;
497
498 if (iter->timestamp > limit)
499 break;
500
501 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
502 if (ret)
503 pr_err("Can't parse sample, err = %d\n", ret);
504 else {
505 ret = perf_session_deliver_event(s, iter->event, &sample, tool,
506 iter->file_offset);
507 if (ret)
508 return ret;
509 }
510
511 os->last_flush = iter->timestamp;
512 list_del(&iter->list);
513 list_add(&iter->list, &os->sample_cache);
514
515 if (show_progress)
516 ui_progress__update(&prog, 1);
517 }
518
519 if (list_empty(head)) {
520 os->last_sample = NULL;
521 } else if (last_ts <= limit) {
522 os->last_sample =
523 list_entry(head->prev, struct sample_queue, list);
524 }
525
526 os->nr_samples = 0;
527
528 return 0;
529}
530
531/*
532 * When perf record finishes a pass on every buffers, it records this pseudo
533 * event.
534 * We record the max timestamp t found in the pass n.
535 * Assuming these timestamps are monotonic across cpus, we know that if
536 * a buffer still has events with timestamps below t, they will be all
537 * available and then read in the pass n + 1.
538 * Hence when we start to read the pass n + 2, we can safely flush every
539 * events with timestamps below t.
540 *
541 * ============ PASS n =================
542 * CPU 0 | CPU 1
543 * |
544 * cnt1 timestamps | cnt2 timestamps
545 * 1 | 2
546 * 2 | 3
547 * - | 4 <--- max recorded
548 *
549 * ============ PASS n + 1 ==============
550 * CPU 0 | CPU 1
551 * |
552 * cnt1 timestamps | cnt2 timestamps
553 * 3 | 5
554 * 4 | 6
555 * 5 | 7 <---- max recorded
556 *
557 * Flush every events below timestamp 4
558 *
559 * ============ PASS n + 2 ==============
560 * CPU 0 | CPU 1
561 * |
562 * cnt1 timestamps | cnt2 timestamps
563 * 6 | 8
564 * 7 | 9
565 * - | 10
566 *
567 * Flush every events below timestamp 7
568 * etc...
569 */
570static int process_finished_round(struct perf_tool *tool,
571 union perf_event *event __maybe_unused,
572 struct perf_session *session)
573{
574 int ret = flush_sample_queue(session, tool);
575 if (!ret)
576 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
577
578 return ret;
579}
580
581/* The queue is ordered by time */
582static void __queue_event(struct sample_queue *new, struct perf_session *s)
583{
584 struct ordered_samples *os = &s->ordered_samples;
585 struct sample_queue *sample = os->last_sample;
586 u64 timestamp = new->timestamp;
587 struct list_head *p;
588
589 ++os->nr_samples;
590 os->last_sample = new;
591
592 if (!sample) {
593 list_add(&new->list, &os->samples);
594 os->max_timestamp = timestamp;
595 return;
596 }
597
598 /*
599 * last_sample might point to some random place in the list as it's
600 * the last queued event. We expect that the new event is close to
601 * this.
602 */
603 if (sample->timestamp <= timestamp) {
604 while (sample->timestamp <= timestamp) {
605 p = sample->list.next;
606 if (p == &os->samples) {
607 list_add_tail(&new->list, &os->samples);
608 os->max_timestamp = timestamp;
609 return;
610 }
611 sample = list_entry(p, struct sample_queue, list);
612 }
613 list_add_tail(&new->list, &sample->list);
614 } else {
615 while (sample->timestamp > timestamp) {
616 p = sample->list.prev;
617 if (p == &os->samples) {
618 list_add(&new->list, &os->samples);
619 return;
620 }
621 sample = list_entry(p, struct sample_queue, list);
622 }
623 list_add(&new->list, &sample->list);
624 }
625}
626
627#define MAX_SAMPLE_BUFFER (64 * 1024 / sizeof(struct sample_queue))
628
629int perf_session_queue_event(struct perf_session *s, union perf_event *event,
630 struct perf_sample *sample, u64 file_offset)
631{
632 struct ordered_samples *os = &s->ordered_samples;
633 struct list_head *sc = &os->sample_cache;
634 u64 timestamp = sample->time;
635 struct sample_queue *new;
636
637 if (!timestamp || timestamp == ~0ULL)
638 return -ETIME;
639
640 if (timestamp < s->ordered_samples.last_flush) {
641 printf("Warning: Timestamp below last timeslice flush\n");
642 return -EINVAL;
643 }
644
645 if (!list_empty(sc)) {
646 new = list_entry(sc->next, struct sample_queue, list);
647 list_del(&new->list);
648 } else if (os->sample_buffer) {
649 new = os->sample_buffer + os->sample_buffer_idx;
650 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
651 os->sample_buffer = NULL;
652 } else {
653 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
654 if (!os->sample_buffer)
655 return -ENOMEM;
656 list_add(&os->sample_buffer->list, &os->to_free);
657 os->sample_buffer_idx = 2;
658 new = os->sample_buffer + 1;
659 }
660
661 new->timestamp = timestamp;
662 new->file_offset = file_offset;
663 new->event = event;
664
665 __queue_event(new, s);
666
667 return 0;
668}
669
670static void callchain__printf(struct perf_sample *sample)
671{
672 unsigned int i;
673
674 printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
675
676 for (i = 0; i < sample->callchain->nr; i++)
677 printf("..... %2d: %016" PRIx64 "\n",
678 i, sample->callchain->ips[i]);
679}
680
681static void branch_stack__printf(struct perf_sample *sample)
682{
683 uint64_t i;
684
685 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
686
687 for (i = 0; i < sample->branch_stack->nr; i++)
688 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
689 i, sample->branch_stack->entries[i].from,
690 sample->branch_stack->entries[i].to);
691}
692
693static void regs_dump__printf(u64 mask, u64 *regs)
694{
695 unsigned rid, i = 0;
696
697 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
698 u64 val = regs[i++];
699
700 printf(".... %-5s 0x%" PRIx64 "\n",
701 perf_reg_name(rid), val);
702 }
703}
704
705static void regs_user__printf(struct perf_sample *sample)
706{
707 struct regs_dump *user_regs = &sample->user_regs;
708
709 if (user_regs->regs) {
710 u64 mask = user_regs->mask;
711 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
712 regs_dump__printf(mask, user_regs->regs);
713 }
714}
715
716static void stack_user__printf(struct stack_dump *dump)
717{
718 printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
719 dump->size, dump->offset);
720}
721
722static void perf_session__print_tstamp(struct perf_session *session,
723 union perf_event *event,
724 struct perf_sample *sample)
725{
726 u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
727
728 if (event->header.type != PERF_RECORD_SAMPLE &&
729 !perf_evlist__sample_id_all(session->evlist)) {
730 fputs("-1 -1 ", stdout);
731 return;
732 }
733
734 if ((sample_type & PERF_SAMPLE_CPU))
735 printf("%u ", sample->cpu);
736
737 if (sample_type & PERF_SAMPLE_TIME)
738 printf("%" PRIu64 " ", sample->time);
739}
740
741static void sample_read__printf(struct perf_sample *sample, u64 read_format)
742{
743 printf("... sample_read:\n");
744
745 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
746 printf("...... time enabled %016" PRIx64 "\n",
747 sample->read.time_enabled);
748
749 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
750 printf("...... time running %016" PRIx64 "\n",
751 sample->read.time_running);
752
753 if (read_format & PERF_FORMAT_GROUP) {
754 u64 i;
755
756 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
757
758 for (i = 0; i < sample->read.group.nr; i++) {
759 struct sample_read_value *value;
760
761 value = &sample->read.group.values[i];
762 printf("..... id %016" PRIx64
763 ", value %016" PRIx64 "\n",
764 value->id, value->value);
765 }
766 } else
767 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
768 sample->read.one.id, sample->read.one.value);
769}
770
771static void dump_event(struct perf_session *session, union perf_event *event,
772 u64 file_offset, struct perf_sample *sample)
773{
774 if (!dump_trace)
775 return;
776
777 printf("\n%#" PRIx64 " [%#x]: event: %d\n",
778 file_offset, event->header.size, event->header.type);
779
780 trace_event(event);
781
782 if (sample)
783 perf_session__print_tstamp(session, event, sample);
784
785 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
786 event->header.size, perf_event__name(event->header.type));
787}
788
789static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
790 struct perf_sample *sample)
791{
792 u64 sample_type;
793
794 if (!dump_trace)
795 return;
796
797 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
798 event->header.misc, sample->pid, sample->tid, sample->ip,
799 sample->period, sample->addr);
800
801 sample_type = evsel->attr.sample_type;
802
803 if (sample_type & PERF_SAMPLE_CALLCHAIN)
804 callchain__printf(sample);
805
806 if (sample_type & PERF_SAMPLE_BRANCH_STACK)
807 branch_stack__printf(sample);
808
809 if (sample_type & PERF_SAMPLE_REGS_USER)
810 regs_user__printf(sample);
811
812 if (sample_type & PERF_SAMPLE_STACK_USER)
813 stack_user__printf(&sample->user_stack);
814
815 if (sample_type & PERF_SAMPLE_WEIGHT)
816 printf("... weight: %" PRIu64 "\n", sample->weight);
817
818 if (sample_type & PERF_SAMPLE_DATA_SRC)
819 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
820
821 if (sample_type & PERF_SAMPLE_TRANSACTION)
822 printf("... transaction: %" PRIx64 "\n", sample->transaction);
823
824 if (sample_type & PERF_SAMPLE_READ)
825 sample_read__printf(sample, evsel->attr.read_format);
826}
827
828static struct machine *
829 perf_session__find_machine_for_cpumode(struct perf_session *session,
830 union perf_event *event,
831 struct perf_sample *sample)
832{
833 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
834 struct machine *machine;
835
836 if (perf_guest &&
837 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
838 (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
839 u32 pid;
840
841 if (event->header.type == PERF_RECORD_MMAP
842 || event->header.type == PERF_RECORD_MMAP2)
843 pid = event->mmap.pid;
844 else
845 pid = sample->pid;
846
847 machine = perf_session__find_machine(session, pid);
848 if (!machine)
849 machine = perf_session__findnew_machine(session,
850 DEFAULT_GUEST_KERNEL_ID);
851 return machine;
852 }
853
854 return &session->machines.host;
855}
856
857static int deliver_sample_value(struct perf_session *session,
858 struct perf_tool *tool,
859 union perf_event *event,
860 struct perf_sample *sample,
861 struct sample_read_value *v,
862 struct machine *machine)
863{
864 struct perf_sample_id *sid;
865
866 sid = perf_evlist__id2sid(session->evlist, v->id);
867 if (sid) {
868 sample->id = v->id;
869 sample->period = v->value - sid->period;
870 sid->period = v->value;
871 }
872
873 if (!sid || sid->evsel == NULL) {
874 ++session->stats.nr_unknown_id;
875 return 0;
876 }
877
878 return tool->sample(tool, event, sample, sid->evsel, machine);
879}
880
881static int deliver_sample_group(struct perf_session *session,
882 struct perf_tool *tool,
883 union perf_event *event,
884 struct perf_sample *sample,
885 struct machine *machine)
886{
887 int ret = -EINVAL;
888 u64 i;
889
890 for (i = 0; i < sample->read.group.nr; i++) {
891 ret = deliver_sample_value(session, tool, event, sample,
892 &sample->read.group.values[i],
893 machine);
894 if (ret)
895 break;
896 }
897
898 return ret;
899}
900
901static int
902perf_session__deliver_sample(struct perf_session *session,
903 struct perf_tool *tool,
904 union perf_event *event,
905 struct perf_sample *sample,
906 struct perf_evsel *evsel,
907 struct machine *machine)
908{
909 /* We know evsel != NULL. */
910 u64 sample_type = evsel->attr.sample_type;
911 u64 read_format = evsel->attr.read_format;
912
913 /* Standard sample delievery. */
914 if (!(sample_type & PERF_SAMPLE_READ))
915 return tool->sample(tool, event, sample, evsel, machine);
916
917 /* For PERF_SAMPLE_READ we have either single or group mode. */
918 if (read_format & PERF_FORMAT_GROUP)
919 return deliver_sample_group(session, tool, event, sample,
920 machine);
921 else
922 return deliver_sample_value(session, tool, event, sample,
923 &sample->read.one, machine);
924}
925
926static int perf_session_deliver_event(struct perf_session *session,
927 union perf_event *event,
928 struct perf_sample *sample,
929 struct perf_tool *tool,
930 u64 file_offset)
931{
932 struct perf_evsel *evsel;
933 struct machine *machine;
934
935 dump_event(session, event, file_offset, sample);
936
937 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
938 if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
939 /*
940 * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
941 * because the tools right now may apply filters, discarding
942 * some of the samples. For consistency, in the future we
943 * should have something like nr_filtered_samples and remove
944 * the sample->period from total_sample_period, etc, KISS for
945 * now tho.
946 *
947 * Also testing against NULL allows us to handle files without
948 * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
949 * future probably it'll be a good idea to restrict event
950 * processing via perf_session to files with both set.
951 */
952 hists__inc_nr_events(&evsel->hists, event->header.type);
953 }
954
955 machine = perf_session__find_machine_for_cpumode(session, event,
956 sample);
957
958 switch (event->header.type) {
959 case PERF_RECORD_SAMPLE:
960 dump_sample(evsel, event, sample);
961 if (evsel == NULL) {
962 ++session->stats.nr_unknown_id;
963 return 0;
964 }
965 if (machine == NULL) {
966 ++session->stats.nr_unprocessable_samples;
967 return 0;
968 }
969 return perf_session__deliver_sample(session, tool, event,
970 sample, evsel, machine);
971 case PERF_RECORD_MMAP:
972 return tool->mmap(tool, event, sample, machine);
973 case PERF_RECORD_MMAP2:
974 return tool->mmap2(tool, event, sample, machine);
975 case PERF_RECORD_COMM:
976 return tool->comm(tool, event, sample, machine);
977 case PERF_RECORD_FORK:
978 return tool->fork(tool, event, sample, machine);
979 case PERF_RECORD_EXIT:
980 return tool->exit(tool, event, sample, machine);
981 case PERF_RECORD_LOST:
982 if (tool->lost == perf_event__process_lost)
983 session->stats.total_lost += event->lost.lost;
984 return tool->lost(tool, event, sample, machine);
985 case PERF_RECORD_READ:
986 return tool->read(tool, event, sample, evsel, machine);
987 case PERF_RECORD_THROTTLE:
988 return tool->throttle(tool, event, sample, machine);
989 case PERF_RECORD_UNTHROTTLE:
990 return tool->unthrottle(tool, event, sample, machine);
991 default:
992 ++session->stats.nr_unknown_events;
993 return -1;
994 }
995}
996
997static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
998 struct perf_tool *tool, u64 file_offset)
999{
1000 int fd = perf_data_file__fd(session->file);
1001 int err;
1002
1003 dump_event(session, event, file_offset, NULL);
1004
1005 /* These events are processed right away */
1006 switch (event->header.type) {
1007 case PERF_RECORD_HEADER_ATTR:
1008 err = tool->attr(tool, event, &session->evlist);
1009 if (err == 0)
1010 perf_session__set_id_hdr_size(session);
1011 return err;
1012 case PERF_RECORD_HEADER_EVENT_TYPE:
1013 /*
1014 * Depreceated, but we need to handle it for sake
1015 * of old data files create in pipe mode.
1016 */
1017 return 0;
1018 case PERF_RECORD_HEADER_TRACING_DATA:
1019 /* setup for reading amidst mmap */
1020 lseek(fd, file_offset, SEEK_SET);
1021 return tool->tracing_data(tool, event, session);
1022 case PERF_RECORD_HEADER_BUILD_ID:
1023 return tool->build_id(tool, event, session);
1024 case PERF_RECORD_FINISHED_ROUND:
1025 return tool->finished_round(tool, event, session);
1026 default:
1027 return -EINVAL;
1028 }
1029}
1030
1031static void event_swap(union perf_event *event, bool sample_id_all)
1032{
1033 perf_event__swap_op swap;
1034
1035 swap = perf_event__swap_ops[event->header.type];
1036 if (swap)
1037 swap(event, sample_id_all);
1038}
1039
1040static int perf_session__process_event(struct perf_session *session,
1041 union perf_event *event,
1042 struct perf_tool *tool,
1043 u64 file_offset)
1044{
1045 struct perf_sample sample;
1046 int ret;
1047
1048 if (session->header.needs_swap)
1049 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1050
1051 if (event->header.type >= PERF_RECORD_HEADER_MAX)
1052 return -EINVAL;
1053
1054 events_stats__inc(&session->stats, event->header.type);
1055
1056 if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1057 return perf_session__process_user_event(session, event, tool, file_offset);
1058
1059 /*
1060 * For all kernel events we get the sample data
1061 */
1062 ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1063 if (ret)
1064 return ret;
1065
1066 if (tool->ordered_samples) {
1067 ret = perf_session_queue_event(session, event, &sample,
1068 file_offset);
1069 if (ret != -ETIME)
1070 return ret;
1071 }
1072
1073 return perf_session_deliver_event(session, event, &sample, tool,
1074 file_offset);
1075}
1076
1077void perf_event_header__bswap(struct perf_event_header *hdr)
1078{
1079 hdr->type = bswap_32(hdr->type);
1080 hdr->misc = bswap_16(hdr->misc);
1081 hdr->size = bswap_16(hdr->size);
1082}
1083
1084struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1085{
1086 return machine__findnew_thread(&session->machines.host, 0, pid);
1087}
1088
1089static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1090{
1091 struct thread *thread = perf_session__findnew(session, 0);
1092
1093 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1094 pr_err("problem inserting idle task.\n");
1095 thread = NULL;
1096 }
1097
1098 return thread;
1099}
1100
1101static void perf_session__warn_about_errors(const struct perf_session *session,
1102 const struct perf_tool *tool)
1103{
1104 if (tool->lost == perf_event__process_lost &&
1105 session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1106 ui__warning("Processed %d events and lost %d chunks!\n\n"
1107 "Check IO/CPU overload!\n\n",
1108 session->stats.nr_events[0],
1109 session->stats.nr_events[PERF_RECORD_LOST]);
1110 }
1111
1112 if (session->stats.nr_unknown_events != 0) {
1113 ui__warning("Found %u unknown events!\n\n"
1114 "Is this an older tool processing a perf.data "
1115 "file generated by a more recent tool?\n\n"
1116 "If that is not the case, consider "
1117 "reporting to linux-kernel@vger.kernel.org.\n\n",
1118 session->stats.nr_unknown_events);
1119 }
1120
1121 if (session->stats.nr_unknown_id != 0) {
1122 ui__warning("%u samples with id not present in the header\n",
1123 session->stats.nr_unknown_id);
1124 }
1125
1126 if (session->stats.nr_invalid_chains != 0) {
1127 ui__warning("Found invalid callchains!\n\n"
1128 "%u out of %u events were discarded for this reason.\n\n"
1129 "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1130 session->stats.nr_invalid_chains,
1131 session->stats.nr_events[PERF_RECORD_SAMPLE]);
1132 }
1133
1134 if (session->stats.nr_unprocessable_samples != 0) {
1135 ui__warning("%u unprocessable samples recorded.\n"
1136 "Do you have a KVM guest running and not using 'perf kvm'?\n",
1137 session->stats.nr_unprocessable_samples);
1138 }
1139}
1140
1141volatile int session_done;
1142
1143static int __perf_session__process_pipe_events(struct perf_session *session,
1144 struct perf_tool *tool)
1145{
1146 int fd = perf_data_file__fd(session->file);
1147 union perf_event *event;
1148 uint32_t size, cur_size = 0;
1149 void *buf = NULL;
1150 int skip = 0;
1151 u64 head;
1152 ssize_t err;
1153 void *p;
1154
1155 perf_tool__fill_defaults(tool);
1156
1157 head = 0;
1158 cur_size = sizeof(union perf_event);
1159
1160 buf = malloc(cur_size);
1161 if (!buf)
1162 return -errno;
1163more:
1164 event = buf;
1165 err = readn(fd, event, sizeof(struct perf_event_header));
1166 if (err <= 0) {
1167 if (err == 0)
1168 goto done;
1169
1170 pr_err("failed to read event header\n");
1171 goto out_err;
1172 }
1173
1174 if (session->header.needs_swap)
1175 perf_event_header__bswap(&event->header);
1176
1177 size = event->header.size;
1178 if (size < sizeof(struct perf_event_header)) {
1179 pr_err("bad event header size\n");
1180 goto out_err;
1181 }
1182
1183 if (size > cur_size) {
1184 void *new = realloc(buf, size);
1185 if (!new) {
1186 pr_err("failed to allocate memory to read event\n");
1187 goto out_err;
1188 }
1189 buf = new;
1190 cur_size = size;
1191 event = buf;
1192 }
1193 p = event;
1194 p += sizeof(struct perf_event_header);
1195
1196 if (size - sizeof(struct perf_event_header)) {
1197 err = readn(fd, p, size - sizeof(struct perf_event_header));
1198 if (err <= 0) {
1199 if (err == 0) {
1200 pr_err("unexpected end of event stream\n");
1201 goto done;
1202 }
1203
1204 pr_err("failed to read event data\n");
1205 goto out_err;
1206 }
1207 }
1208
1209 if ((skip = perf_session__process_event(session, event, tool, head)) < 0) {
1210 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1211 head, event->header.size, event->header.type);
1212 err = -EINVAL;
1213 goto out_err;
1214 }
1215
1216 head += size;
1217
1218 if (skip > 0)
1219 head += skip;
1220
1221 if (!session_done())
1222 goto more;
1223done:
1224 /* do the final flush for ordered samples */
1225 session->ordered_samples.next_flush = ULLONG_MAX;
1226 err = flush_sample_queue(session, tool);
1227out_err:
1228 free(buf);
1229 perf_session__warn_about_errors(session, tool);
1230 perf_session_free_sample_buffers(session);
1231 return err;
1232}
1233
1234static union perf_event *
1235fetch_mmaped_event(struct perf_session *session,
1236 u64 head, size_t mmap_size, char *buf)
1237{
1238 union perf_event *event;
1239
1240 /*
1241 * Ensure we have enough space remaining to read
1242 * the size of the event in the headers.
1243 */
1244 if (head + sizeof(event->header) > mmap_size)
1245 return NULL;
1246
1247 event = (union perf_event *)(buf + head);
1248
1249 if (session->header.needs_swap)
1250 perf_event_header__bswap(&event->header);
1251
1252 if (head + event->header.size > mmap_size) {
1253 /* We're not fetching the event so swap back again */
1254 if (session->header.needs_swap)
1255 perf_event_header__bswap(&event->header);
1256 return NULL;
1257 }
1258
1259 return event;
1260}
1261
1262/*
1263 * On 64bit we can mmap the data file in one go. No need for tiny mmap
1264 * slices. On 32bit we use 32MB.
1265 */
1266#if BITS_PER_LONG == 64
1267#define MMAP_SIZE ULLONG_MAX
1268#define NUM_MMAPS 1
1269#else
1270#define MMAP_SIZE (32 * 1024 * 1024ULL)
1271#define NUM_MMAPS 128
1272#endif
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 int fd = perf_data_file__fd(session->file);
1279 u64 head, page_offset, file_offset, file_pos;
1280 int err, mmap_prot, mmap_flags, map_idx = 0;
1281 size_t mmap_size;
1282 char *buf, *mmaps[NUM_MMAPS];
1283 union perf_event *event;
1284 uint32_t size;
1285 struct ui_progress prog;
1286
1287 perf_tool__fill_defaults(tool);
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_size && (data_offset + data_size < file_size))
1294 file_size = data_offset + data_size;
1295
1296 ui_progress__init(&prog, file_size, "Processing events...");
1297
1298 mmap_size = MMAP_SIZE;
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, 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 < sizeof(struct perf_event_header) ||
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 ui_progress__update(&prog, size);
1352
1353 if (session_done())
1354 goto out;
1355
1356 if (file_pos < file_size)
1357 goto more;
1358
1359out:
1360 /* do the final flush for ordered samples */
1361 session->ordered_samples.next_flush = ULLONG_MAX;
1362 err = flush_sample_queue(session, tool);
1363out_err:
1364 ui_progress__finish();
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 *session,
1371 struct perf_tool *tool)
1372{
1373 u64 size = perf_data_file__size(session->file);
1374 int err;
1375
1376 if (perf_session__register_idle_thread(session) == NULL)
1377 return -ENOMEM;
1378
1379 if (!perf_data_file__is_pipe(session->file))
1380 err = __perf_session__process_events(session,
1381 session->header.data_offset,
1382 session->header.data_size,
1383 size, tool);
1384 else
1385 err = __perf_session__process_pipe_events(session, tool);
1386
1387 return err;
1388}
1389
1390bool perf_session__has_traces(struct perf_session *session, const char *msg)
1391{
1392 struct perf_evsel *evsel;
1393
1394 evlist__for_each(session->evlist, evsel) {
1395 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1396 return true;
1397 }
1398
1399 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1400 return false;
1401}
1402
1403int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1404 const char *symbol_name, u64 addr)
1405{
1406 char *bracket;
1407 enum map_type i;
1408 struct ref_reloc_sym *ref;
1409
1410 ref = zalloc(sizeof(struct ref_reloc_sym));
1411 if (ref == NULL)
1412 return -ENOMEM;
1413
1414 ref->name = strdup(symbol_name);
1415 if (ref->name == NULL) {
1416 free(ref);
1417 return -ENOMEM;
1418 }
1419
1420 bracket = strchr(ref->name, ']');
1421 if (bracket)
1422 *bracket = '\0';
1423
1424 ref->addr = addr;
1425
1426 for (i = 0; i < MAP__NR_TYPES; ++i) {
1427 struct kmap *kmap = map__kmap(maps[i]);
1428 kmap->ref_reloc_sym = ref;
1429 }
1430
1431 return 0;
1432}
1433
1434size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1435{
1436 return machines__fprintf_dsos(&session->machines, fp);
1437}
1438
1439size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1440 bool (skip)(struct dso *dso, int parm), int parm)
1441{
1442 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1443}
1444
1445size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1446{
1447 struct perf_evsel *pos;
1448 size_t ret = fprintf(fp, "Aggregated stats:\n");
1449
1450 ret += events_stats__fprintf(&session->stats, fp);
1451
1452 evlist__for_each(session->evlist, pos) {
1453 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1454 ret += events_stats__fprintf(&pos->hists.stats, fp);
1455 }
1456
1457 return ret;
1458}
1459
1460size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1461{
1462 /*
1463 * FIXME: Here we have to actually print all the machines in this
1464 * session, not just the host...
1465 */
1466 return machine__fprintf(&session->machines.host, fp);
1467}
1468
1469struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1470 unsigned int type)
1471{
1472 struct perf_evsel *pos;
1473
1474 evlist__for_each(session->evlist, pos) {
1475 if (pos->attr.type == type)
1476 return pos;
1477 }
1478 return NULL;
1479}
1480
1481void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1482 struct addr_location *al,
1483 unsigned int print_opts, unsigned int stack_depth)
1484{
1485 struct callchain_cursor_node *node;
1486 int print_ip = print_opts & PRINT_IP_OPT_IP;
1487 int print_sym = print_opts & PRINT_IP_OPT_SYM;
1488 int print_dso = print_opts & PRINT_IP_OPT_DSO;
1489 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1490 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1491 int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1492 char s = print_oneline ? ' ' : '\t';
1493
1494 if (symbol_conf.use_callchain && sample->callchain) {
1495 struct addr_location node_al;
1496
1497 if (machine__resolve_callchain(al->machine, evsel, al->thread,
1498 sample, NULL, NULL,
1499 PERF_MAX_STACK_DEPTH) != 0) {
1500 if (verbose)
1501 error("Failed to resolve callchain. Skipping\n");
1502 return;
1503 }
1504 callchain_cursor_commit(&callchain_cursor);
1505
1506 if (print_symoffset)
1507 node_al = *al;
1508
1509 while (stack_depth) {
1510 u64 addr = 0;
1511
1512 node = callchain_cursor_current(&callchain_cursor);
1513 if (!node)
1514 break;
1515
1516 if (node->sym && node->sym->ignore)
1517 goto next;
1518
1519 if (print_ip)
1520 printf("%c%16" PRIx64, s, node->ip);
1521
1522 if (node->map)
1523 addr = node->map->map_ip(node->map, node->ip);
1524
1525 if (print_sym) {
1526 printf(" ");
1527 if (print_symoffset) {
1528 node_al.addr = addr;
1529 node_al.map = node->map;
1530 symbol__fprintf_symname_offs(node->sym, &node_al, stdout);
1531 } else
1532 symbol__fprintf_symname(node->sym, stdout);
1533 }
1534
1535 if (print_dso) {
1536 printf(" (");
1537 map__fprintf_dsoname(node->map, stdout);
1538 printf(")");
1539 }
1540
1541 if (print_srcline)
1542 map__fprintf_srcline(node->map, addr, "\n ",
1543 stdout);
1544
1545 if (!print_oneline)
1546 printf("\n");
1547
1548 stack_depth--;
1549next:
1550 callchain_cursor_advance(&callchain_cursor);
1551 }
1552
1553 } else {
1554 if (al->sym && al->sym->ignore)
1555 return;
1556
1557 if (print_ip)
1558 printf("%16" PRIx64, sample->ip);
1559
1560 if (print_sym) {
1561 printf(" ");
1562 if (print_symoffset)
1563 symbol__fprintf_symname_offs(al->sym, al,
1564 stdout);
1565 else
1566 symbol__fprintf_symname(al->sym, stdout);
1567 }
1568
1569 if (print_dso) {
1570 printf(" (");
1571 map__fprintf_dsoname(al->map, stdout);
1572 printf(")");
1573 }
1574
1575 if (print_srcline)
1576 map__fprintf_srcline(al->map, al->addr, "\n ", stdout);
1577 }
1578}
1579
1580int perf_session__cpu_bitmap(struct perf_session *session,
1581 const char *cpu_list, unsigned long *cpu_bitmap)
1582{
1583 int i, err = -1;
1584 struct cpu_map *map;
1585
1586 for (i = 0; i < PERF_TYPE_MAX; ++i) {
1587 struct perf_evsel *evsel;
1588
1589 evsel = perf_session__find_first_evtype(session, i);
1590 if (!evsel)
1591 continue;
1592
1593 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1594 pr_err("File does not contain CPU events. "
1595 "Remove -c option to proceed.\n");
1596 return -1;
1597 }
1598 }
1599
1600 map = cpu_map__new(cpu_list);
1601 if (map == NULL) {
1602 pr_err("Invalid cpu_list\n");
1603 return -1;
1604 }
1605
1606 for (i = 0; i < map->nr; i++) {
1607 int cpu = map->map[i];
1608
1609 if (cpu >= MAX_NR_CPUS) {
1610 pr_err("Requested CPU %d too large. "
1611 "Consider raising MAX_NR_CPUS\n", cpu);
1612 goto out_delete_map;
1613 }
1614
1615 set_bit(cpu, cpu_bitmap);
1616 }
1617
1618 err = 0;
1619
1620out_delete_map:
1621 cpu_map__delete(map);
1622 return err;
1623}
1624
1625void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1626 bool full)
1627{
1628 int fd = perf_data_file__fd(session->file);
1629 struct stat st;
1630 int ret;
1631
1632 if (session == NULL || fp == NULL)
1633 return;
1634
1635 ret = fstat(fd, &st);
1636 if (ret == -1)
1637 return;
1638
1639 fprintf(fp, "# ========\n");
1640 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1641 perf_header__fprintf_info(session, fp, full);
1642 fprintf(fp, "# ========\n#\n");
1643}
1644
1645
1646int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1647 const struct perf_evsel_str_handler *assocs,
1648 size_t nr_assocs)
1649{
1650 struct perf_evsel *evsel;
1651 size_t i;
1652 int err;
1653
1654 for (i = 0; i < nr_assocs; i++) {
1655 /*
1656 * Adding a handler for an event not in the session,
1657 * just ignore it.
1658 */
1659 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1660 if (evsel == NULL)
1661 continue;
1662
1663 err = -EEXIST;
1664 if (evsel->handler != NULL)
1665 goto out;
1666 evsel->handler = assocs[i].handler;
1667 }
1668
1669 err = 0;
1670out:
1671 return err;
1672}