Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[~shefty/rdma-dev.git] / tools / perf / util / session.c
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
17 static 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
75 out_close:
76         close(self->fd);
77         self->fd = -1;
78         return -1;
79 }
80
81 static 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;
104 out:
105        session->id_hdr_size = size;
106 }
107
108 void 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
116 int 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
125 static 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
131 struct 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
180 out:
181         return self;
182 out_delete:
183         perf_session__delete(self);
184         return NULL;
185 }
186
187 static 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
197 static 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
210 void 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
220 void 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
231 static 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
239 int 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
293 static 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
300 static 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
309 static 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
317 static 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
325 static int process_finished_round(union perf_event *event,
326                                   struct perf_session *session,
327                                   struct perf_event_ops *ops);
328
329 static 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
365 void 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
376 static 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
382 static 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
388 static 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
397 static 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
406 static 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 */
417 void 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
431 static 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
442 static 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
448 static void perf_event__tracing_data_swap(union perf_event *event)
449 {
450         event->tracing_data.size = bswap_32(event->tracing_data.size);
451 }
452
453 typedef void (*perf_event__swap_op)(union perf_event *event);
454
455 static 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
470 struct sample_queue {
471         u64                     timestamp;
472         u64                     file_offset;
473         union perf_event        *event;
474         struct list_head        list;
475 };
476
477 static 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
490 static 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
496 static 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  */
573 static 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 */
584 static 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
630 static 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
671 static 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
682 static 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
699 static 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
717 static 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
731 static 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
772 static 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
788 static 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
812 static 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
854 void 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
861 static 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
873 static 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))
908 volatile int session_done;
909
910 static 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;
923 more:
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;
977 done:
978         err = 0;
979 out_err:
980         perf_session__warn_about_errors(self, ops);
981         perf_session_free_sample_buffers(self);
982         return err;
983 }
984
985 static union perf_event *
986 fetch_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
1009 int __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         }
1050 remap:
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
1062 more:
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);
1108 out_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
1115 int 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
1134 bool 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
1144 int 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
1176 size_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
1183 size_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
1190 size_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
1205 struct 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
1217 void 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
1293 int 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 }
1329
1330 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1331                                 bool full)
1332 {
1333         struct stat st;
1334         int ret;
1335
1336         if (session == NULL || fp == NULL)
1337                 return;
1338
1339         ret = fstat(session->fd, &st);
1340         if (ret == -1)
1341                 return;
1342
1343         fprintf(fp, "# ========\n");
1344         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1345         perf_header__fprintf_info(session, fp, full);
1346         fprintf(fp, "# ========\n#\n");
1347 }