Pileus Git - ~andy/linux/blob - 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 "session.h"
  11 #include "sort.h"
  12 #include "util.h"
  13
  14 static int perf_session__open(struct perf_session *self, bool force)
  15 {
  16         struct stat input_stat;
  17
  18         if (!strcmp(self->filename, "-")) {
  19                 self->fd_pipe = true;
  20                 self->fd = STDIN_FILENO;
  21
  22                 if (perf_header__read(self, self->fd) < 0)
  23                         pr_err("incompatible file format");
  24
  25                 return 0;
  26         }
  27
  28         self->fd = open(self->filename, O_RDONLY);
  29         if (self->fd < 0) {
  30                 int err = errno;
  31
  32                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  33                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  34                         pr_err("  (try 'perf record' first)");
  35                 pr_err("\n");
  36                 return -errno;
  37         }
  38
  39         if (fstat(self->fd, &input_stat) < 0)
  40                 goto out_close;
  41
  42         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  43                 pr_err("file %s not owned by current user or root\n",
  44                        self->filename);
  45                 goto out_close;
  46         }
  47
  48         if (!input_stat.st_size) {
  49                 pr_info("zero-sized file (%s), nothing to do!\n",
  50                         self->filename);
  51                 goto out_close;
  52         }
  53
  54         if (perf_header__read(self, self->fd) < 0) {
  55                 pr_err("incompatible file format");
  56                 goto out_close;
  57         }
  58
  59         self->size = input_stat.st_size;
  60         return 0;
  61
  62 out_close:
  63         close(self->fd);
  64         self->fd = -1;
  65         return -1;
  66 }
  67
  68 static void perf_session__id_header_size(struct perf_session *session)
  69 {
  70        struct sample_data *data;
  71        u64 sample_type = session->sample_type;
  72        u16 size = 0;
  73
  74         if (!session->sample_id_all)
  75                 goto out;
  76
  77        if (sample_type & PERF_SAMPLE_TID)
  78                size += sizeof(data->tid) * 2;
  79
  80        if (sample_type & PERF_SAMPLE_TIME)
  81                size += sizeof(data->time);
  82
  83        if (sample_type & PERF_SAMPLE_ID)
  84                size += sizeof(data->id);
  85
  86        if (sample_type & PERF_SAMPLE_STREAM_ID)
  87                size += sizeof(data->stream_id);
  88
  89        if (sample_type & PERF_SAMPLE_CPU)
  90                size += sizeof(data->cpu) * 2;
  91 out:
  92        session->id_hdr_size = size;
  93 }
  94
  95 void perf_session__set_sample_id_all(struct perf_session *session, bool value)
  96 {
  97         session->sample_id_all = value;
  98         perf_session__id_header_size(session);
  99 }
 100
 101 void perf_session__set_sample_type(struct perf_session *session, u64 type)
 102 {
 103         session->sample_type = type;
 104 }
 105
 106 void perf_session__update_sample_type(struct perf_session *self)
 107 {
 108         self->sample_type = perf_header__sample_type(&self->header);
 109         self->sample_id_all = perf_header__sample_id_all(&self->header);
 110         perf_session__id_header_size(self);
 111 }
 112
 113 int perf_session__create_kernel_maps(struct perf_session *self)
 114 {
 115         int ret = machine__create_kernel_maps(&self->host_machine);
 116
 117         if (ret >= 0)
 118                 ret = machines__create_guest_kernel_maps(&self->machines);
 119         return ret;
 120 }
 121
 122 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 123 {
 124         machine__destroy_kernel_maps(&self->host_machine);
 125         machines__destroy_guest_kernel_maps(&self->machines);
 126 }
 127
 128 struct perf_session *perf_session__new(const char *filename, int mode,
 129                                        bool force, bool repipe,
 130                                        struct perf_event_ops *ops)
 131 {
 132         size_t len = filename ? strlen(filename) + 1 : 0;
 133         struct perf_session *self = zalloc(sizeof(*self) + len);
 134
 135         if (self == NULL)
 136                 goto out;
 137
 138         if (perf_header__init(&self->header) < 0)
 139                 goto out_free;
 140
 141         memcpy(self->filename, filename, len);
 142         self->threads = RB_ROOT;
 143         INIT_LIST_HEAD(&self->dead_threads);
 144         self->hists_tree = RB_ROOT;
 145         self->last_match = NULL;
 146         /*
 147          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 148          * slices. On 32bit we use 32MB.
 149          */
 150 #if BITS_PER_LONG == 64
 151         self->mmap_window = ULLONG_MAX;
 152 #else
 153         self->mmap_window = 32 * 1024 * 1024ULL;
 154 #endif
 155         self->machines = RB_ROOT;
 156         self->repipe = repipe;
 157         INIT_LIST_HEAD(&self->ordered_samples.samples);
 158         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 159         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 160         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 161
 162         if (mode == O_RDONLY) {
 163                 if (perf_session__open(self, force) < 0)
 164                         goto out_delete;
 165         } else if (mode == O_WRONLY) {
 166                 /*
 167                  * In O_RDONLY mode this will be performed when reading the
 168                  * kernel MMAP event, in event__process_mmap().
 169                  */
 170                 if (perf_session__create_kernel_maps(self) < 0)
 171                         goto out_delete;
 172         }
 173
 174         perf_session__update_sample_type(self);
 175
 176         if (ops && ops->ordering_requires_timestamps &&
 177             ops->ordered_samples && !self->sample_id_all) {
 178                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 179                 ops->ordered_samples = false;
 180         }
 181
 182 out:
 183         return self;
 184 out_free:
 185         free(self);
 186         return NULL;
 187 out_delete:
 188         perf_session__delete(self);
 189         return NULL;
 190 }
 191
 192 static void perf_session__delete_dead_threads(struct perf_session *self)
 193 {
 194         struct thread *n, *t;
 195
 196         list_for_each_entry_safe(t, n, &self->dead_threads, node) {
 197                 list_del(&t->node);
 198                 thread__delete(t);
 199         }
 200 }
 201
 202 static void perf_session__delete_threads(struct perf_session *self)
 203 {
 204         struct rb_node *nd = rb_first(&self->threads);
 205
 206         while (nd) {
 207                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 208
 209                 rb_erase(&t->rb_node, &self->threads);
 210                 nd = rb_next(nd);
 211                 thread__delete(t);
 212         }
 213 }
 214
 215 void perf_session__delete(struct perf_session *self)
 216 {
 217         perf_header__exit(&self->header);
 218         perf_session__destroy_kernel_maps(self);
 219         perf_session__delete_dead_threads(self);
 220         perf_session__delete_threads(self);
 221         machine__exit(&self->host_machine);
 222         close(self->fd);
 223         free(self);
 224 }
 225
 226 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
 227 {
 228         self->last_match = NULL;
 229         rb_erase(&th->rb_node, &self->threads);
 230         /*
 231          * We may have references to this thread, for instance in some hist_entry
 232          * instances, so just move them to a separate list.
 233          */
 234         list_add_tail(&th->node, &self->dead_threads);
 235 }
 236
 237 static bool symbol__match_parent_regex(struct symbol *sym)
 238 {
 239         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 240                 return 1;
 241
 242         return 0;
 243 }
 244
 245 int perf_session__resolve_callchain(struct perf_session *self,
 246                                     struct thread *thread,
 247                                     struct ip_callchain *chain,
 248                                     struct symbol **parent)
 249 {
 250         u8 cpumode = PERF_RECORD_MISC_USER;
 251         unsigned int i;
 252         int err;
 253
 254         callchain_cursor_reset(&self->callchain_cursor);
 255
 256         for (i = 0; i < chain->nr; i++) {
 257                 u64 ip = chain->ips[i];
 258                 struct addr_location al;
 259
 260                 if (ip >= PERF_CONTEXT_MAX) {
 261                         switch (ip) {
 262                         case PERF_CONTEXT_HV:
 263                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 264                         case PERF_CONTEXT_KERNEL:
 265                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 266                         case PERF_CONTEXT_USER:
 267                                 cpumode = PERF_RECORD_MISC_USER;        break;
 268                         default:
 269                                 break;
 270                         }
 271                         continue;
 272                 }
 273
 274                 al.filtered = false;
 275                 thread__find_addr_location(thread, self, cpumode,
 276                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 277                 if (al.sym != NULL) {
 278                         if (sort__has_parent && !*parent &&
 279                             symbol__match_parent_regex(al.sym))
 280                                 *parent = al.sym;
 281                         if (!symbol_conf.use_callchain)
 282                                 break;
 283                 }
 284
 285                 err = callchain_cursor_append(&self->callchain_cursor,
 286                                               ip, al.map, al.sym);
 287                 if (err)
 288                         return err;
 289         }
 290
 291         return 0;
 292 }
 293
 294 static int process_event_synth_stub(event_t *event __used,
 295                                     struct perf_session *session __used)
 296 {
 297         dump_printf(": unhandled!\n");
 298         return 0;
 299 }
 300
 301 static int process_event_stub(event_t *event __used,
 302                               struct sample_data *sample __used,
 303                               struct perf_session *session __used)
 304 {
 305         dump_printf(": unhandled!\n");
 306         return 0;
 307 }
 308
 309 static int process_finished_round_stub(event_t *event __used,
 310                                        struct perf_session *session __used,
 311                                        struct perf_event_ops *ops __used)
 312 {
 313         dump_printf(": unhandled!\n");
 314         return 0;
 315 }
 316
 317 static int process_finished_round(event_t *event,
 318                                   struct perf_session *session,
 319                                   struct perf_event_ops *ops);
 320
 321 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 322 {
 323         if (handler->sample == NULL)
 324                 handler->sample = process_event_stub;
 325         if (handler->mmap == NULL)
 326                 handler->mmap = process_event_stub;
 327         if (handler->comm == NULL)
 328                 handler->comm = process_event_stub;
 329         if (handler->fork == NULL)
 330                 handler->fork = process_event_stub;
 331         if (handler->exit == NULL)
 332                 handler->exit = process_event_stub;
 333         if (handler->lost == NULL)
 334                 handler->lost = event__process_lost;
 335         if (handler->read == NULL)
 336                 handler->read = process_event_stub;
 337         if (handler->throttle == NULL)
 338                 handler->throttle = process_event_stub;
 339         if (handler->unthrottle == NULL)
 340                 handler->unthrottle = process_event_stub;
 341         if (handler->attr == NULL)
 342                 handler->attr = process_event_synth_stub;
 343         if (handler->event_type == NULL)
 344                 handler->event_type = process_event_synth_stub;
 345         if (handler->tracing_data == NULL)
 346                 handler->tracing_data = process_event_synth_stub;
 347         if (handler->build_id == NULL)
 348                 handler->build_id = process_event_synth_stub;
 349         if (handler->finished_round == NULL) {
 350                 if (handler->ordered_samples)
 351                         handler->finished_round = process_finished_round;
 352                 else
 353                         handler->finished_round = process_finished_round_stub;
 354         }
 355 }
 356
 357 void mem_bswap_64(void *src, int byte_size)
 358 {
 359         u64 *m = src;
 360
 361         while (byte_size > 0) {
 362                 *m = bswap_64(*m);
 363                 byte_size -= sizeof(u64);
 364                 ++m;
 365         }
 366 }
 367
 368 static void event__all64_swap(event_t *self)
 369 {
 370         struct perf_event_header *hdr = &self->header;
 371         mem_bswap_64(hdr + 1, self->header.size - sizeof(*hdr));
 372 }
 373
 374 static void event__comm_swap(event_t *self)
 375 {
 376         self->comm.pid = bswap_32(self->comm.pid);
 377         self->comm.tid = bswap_32(self->comm.tid);
 378 }
 379
 380 static void event__mmap_swap(event_t *self)
 381 {
 382         self->mmap.pid   = bswap_32(self->mmap.pid);
 383         self->mmap.tid   = bswap_32(self->mmap.tid);
 384         self->mmap.start = bswap_64(self->mmap.start);
 385         self->mmap.len   = bswap_64(self->mmap.len);
 386         self->mmap.pgoff = bswap_64(self->mmap.pgoff);
 387 }
 388
 389 static void event__task_swap(event_t *self)
 390 {
 391         self->fork.pid  = bswap_32(self->fork.pid);
 392         self->fork.tid  = bswap_32(self->fork.tid);
 393         self->fork.ppid = bswap_32(self->fork.ppid);
 394         self->fork.ptid = bswap_32(self->fork.ptid);
 395         self->fork.time = bswap_64(self->fork.time);
 396 }
 397
 398 static void event__read_swap(event_t *self)
 399 {
 400         self->read.pid          = bswap_32(self->read.pid);
 401         self->read.tid          = bswap_32(self->read.tid);
 402         self->read.value        = bswap_64(self->read.value);
 403         self->read.time_enabled = bswap_64(self->read.time_enabled);
 404         self->read.time_running = bswap_64(self->read.time_running);
 405         self->read.id           = bswap_64(self->read.id);
 406 }
 407
 408 static void event__attr_swap(event_t *self)
 409 {
 410         size_t size;
 411
 412         self->attr.attr.type            = bswap_32(self->attr.attr.type);
 413         self->attr.attr.size            = bswap_32(self->attr.attr.size);
 414         self->attr.attr.config          = bswap_64(self->attr.attr.config);
 415         self->attr.attr.sample_period   = bswap_64(self->attr.attr.sample_period);
 416         self->attr.attr.sample_type     = bswap_64(self->attr.attr.sample_type);
 417         self->attr.attr.read_format     = bswap_64(self->attr.attr.read_format);
 418         self->attr.attr.wakeup_events   = bswap_32(self->attr.attr.wakeup_events);
 419         self->attr.attr.bp_type         = bswap_32(self->attr.attr.bp_type);
 420         self->attr.attr.bp_addr         = bswap_64(self->attr.attr.bp_addr);
 421         self->attr.attr.bp_len          = bswap_64(self->attr.attr.bp_len);
 422
 423         size = self->header.size;
 424         size -= (void *)&self->attr.id - (void *)self;
 425         mem_bswap_64(self->attr.id, size);
 426 }
 427
 428 static void event__event_type_swap(event_t *self)
 429 {
 430         self->event_type.event_type.event_id =
 431                 bswap_64(self->event_type.event_type.event_id);
 432 }
 433
 434 static void event__tracing_data_swap(event_t *self)
 435 {
 436         self->tracing_data.size = bswap_32(self->tracing_data.size);
 437 }
 438
 439 typedef void (*event__swap_op)(event_t *self);
 440
 441 static event__swap_op event__swap_ops[] = {
 442         [PERF_RECORD_MMAP]   = event__mmap_swap,
 443         [PERF_RECORD_COMM]   = event__comm_swap,
 444         [PERF_RECORD_FORK]   = event__task_swap,
 445         [PERF_RECORD_EXIT]   = event__task_swap,
 446         [PERF_RECORD_LOST]   = event__all64_swap,
 447         [PERF_RECORD_READ]   = event__read_swap,
 448         [PERF_RECORD_SAMPLE] = event__all64_swap,
 449         [PERF_RECORD_HEADER_ATTR]   = event__attr_swap,
 450         [PERF_RECORD_HEADER_EVENT_TYPE]   = event__event_type_swap,
 451         [PERF_RECORD_HEADER_TRACING_DATA]   = event__tracing_data_swap,
 452         [PERF_RECORD_HEADER_BUILD_ID]   = NULL,
 453         [PERF_RECORD_HEADER_MAX]    = NULL,
 454 };
 455
 456 struct sample_queue {
 457         u64                     timestamp;
 458         u64                     file_offset;
 459         event_t                 *event;
 460         struct list_head        list;
 461 };
 462
 463 static void perf_session_free_sample_buffers(struct perf_session *session)
 464 {
 465         struct ordered_samples *os = &session->ordered_samples;
 466
 467         while (!list_empty(&os->to_free)) {
 468                 struct sample_queue *sq;
 469
 470                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 471                 list_del(&sq->list);
 472                 free(sq);
 473         }
 474 }
 475
 476 static int perf_session_deliver_event(struct perf_session *session,
 477                                       event_t *event,
 478                                       struct sample_data *sample,
 479                                       struct perf_event_ops *ops,
 480                                       u64 file_offset);
 481
 482 static void flush_sample_queue(struct perf_session *s,
 483                                struct perf_event_ops *ops)
 484 {
 485         struct ordered_samples *os = &s->ordered_samples;
 486         struct list_head *head = &os->samples;
 487         struct sample_queue *tmp, *iter;
 488         struct sample_data sample;
 489         u64 limit = os->next_flush;
 490         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 491
 492         if (!ops->ordered_samples || !limit)
 493                 return;
 494
 495         list_for_each_entry_safe(iter, tmp, head, list) {
 496                 if (iter->timestamp > limit)
 497                         break;
 498
 499                 event__parse_sample(iter->event, s, &sample);
 500                 perf_session_deliver_event(s, iter->event, &sample, ops,
 501                                            iter->file_offset);
 502
 503                 os->last_flush = iter->timestamp;
 504                 list_del(&iter->list);
 505                 list_add(&iter->list, &os->sample_cache);
 506         }
 507
 508         if (list_empty(head)) {
 509                 os->last_sample = NULL;
 510         } else if (last_ts <= limit) {
 511                 os->last_sample =
 512                         list_entry(head->prev, struct sample_queue, list);
 513         }
 514 }
 515
 516 /*
 517  * When perf record finishes a pass on every buffers, it records this pseudo
 518  * event.
 519  * We record the max timestamp t found in the pass n.
 520  * Assuming these timestamps are monotonic across cpus, we know that if
 521  * a buffer still has events with timestamps below t, they will be all
 522  * available and then read in the pass n + 1.
 523  * Hence when we start to read the pass n + 2, we can safely flush every
 524  * events with timestamps below t.
 525  *
 526  *    ============ PASS n =================
 527  *       CPU 0         |   CPU 1
 528  *                     |
 529  *    cnt1 timestamps  |   cnt2 timestamps
 530  *          1          |         2
 531  *          2          |         3
 532  *          -          |         4  <--- max recorded
 533  *
 534  *    ============ PASS n + 1 ==============
 535  *       CPU 0         |   CPU 1
 536  *                     |
 537  *    cnt1 timestamps  |   cnt2 timestamps
 538  *          3          |         5
 539  *          4          |         6
 540  *          5          |         7 <---- max recorded
 541  *
 542  *      Flush every events below timestamp 4
 543  *
 544  *    ============ PASS n + 2 ==============
 545  *       CPU 0         |   CPU 1
 546  *                     |
 547  *    cnt1 timestamps  |   cnt2 timestamps
 548  *          6          |         8
 549  *          7          |         9
 550  *          -          |         10
 551  *
 552  *      Flush every events below timestamp 7
 553  *      etc...
 554  */
 555 static int process_finished_round(event_t *event __used,
 556                                   struct perf_session *session,
 557                                   struct perf_event_ops *ops)
 558 {
 559         flush_sample_queue(session, ops);
 560         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 561
 562         return 0;
 563 }
 564
 565 /* The queue is ordered by time */
 566 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 567 {
 568         struct ordered_samples *os = &s->ordered_samples;
 569         struct sample_queue *sample = os->last_sample;
 570         u64 timestamp = new->timestamp;
 571         struct list_head *p;
 572
 573         os->last_sample = new;
 574
 575         if (!sample) {
 576                 list_add(&new->list, &os->samples);
 577                 os->max_timestamp = timestamp;
 578                 return;
 579         }
 580
 581         /*
 582          * last_sample might point to some random place in the list as it's
 583          * the last queued event. We expect that the new event is close to
 584          * this.
 585          */
 586         if (sample->timestamp <= timestamp) {
 587                 while (sample->timestamp <= timestamp) {
 588                         p = sample->list.next;
 589                         if (p == &os->samples) {
 590                                 list_add_tail(&new->list, &os->samples);
 591                                 os->max_timestamp = timestamp;
 592                                 return;
 593                         }
 594                         sample = list_entry(p, struct sample_queue, list);
 595                 }
 596                 list_add_tail(&new->list, &sample->list);
 597         } else {
 598                 while (sample->timestamp > timestamp) {
 599                         p = sample->list.prev;
 600                         if (p == &os->samples) {
 601                                 list_add(&new->list, &os->samples);
 602                                 return;
 603                         }
 604                         sample = list_entry(p, struct sample_queue, list);
 605                 }
 606                 list_add(&new->list, &sample->list);
 607         }
 608 }
 609
 610 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 611
 612 static int perf_session_queue_event(struct perf_session *s, event_t *event,
 613                                     struct sample_data *data, u64 file_offset)
 614 {
 615         struct ordered_samples *os = &s->ordered_samples;
 616         struct list_head *sc = &os->sample_cache;
 617         u64 timestamp = data->time;
 618         struct sample_queue *new;
 619
 620         if (!timestamp || timestamp == ~0ULL)
 621                 return -ETIME;
 622
 623         if (timestamp < s->ordered_samples.last_flush) {
 624                 printf("Warning: Timestamp below last timeslice flush\n");
 625                 return -EINVAL;
 626         }
 627
 628         if (!list_empty(sc)) {
 629                 new = list_entry(sc->next, struct sample_queue, list);
 630                 list_del(&new->list);
 631         } else if (os->sample_buffer) {
 632                 new = os->sample_buffer + os->sample_buffer_idx;
 633                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 634                         os->sample_buffer = NULL;
 635         } else {
 636                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 637                 if (!os->sample_buffer)
 638                         return -ENOMEM;
 639                 list_add(&os->sample_buffer->list, &os->to_free);
 640                 os->sample_buffer_idx = 2;
 641                 new = os->sample_buffer + 1;
 642         }
 643
 644         new->timestamp = timestamp;
 645         new->file_offset = file_offset;
 646         new->event = event;
 647
 648         __queue_event(new, s);
 649
 650         return 0;
 651 }
 652
 653 static void callchain__printf(struct sample_data *sample)
 654 {
 655         unsigned int i;
 656
 657         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 658
 659         for (i = 0; i < sample->callchain->nr; i++)
 660                 printf("..... %2d: %016" PRIx64 "\n",
 661                        i, sample->callchain->ips[i]);
 662 }
 663
 664 static void perf_session__print_tstamp(struct perf_session *session,
 665                                        event_t *event,
 666                                        struct sample_data *sample)
 667 {
 668         if (event->header.type != PERF_RECORD_SAMPLE &&
 669             !session->sample_id_all) {
 670                 fputs("-1 -1 ", stdout);
 671                 return;
 672         }
 673
 674         if ((session->sample_type & PERF_SAMPLE_CPU))
 675                 printf("%u ", sample->cpu);
 676
 677         if (session->sample_type & PERF_SAMPLE_TIME)
 678                 printf("%" PRIu64 " ", sample->time);
 679 }
 680
 681 static void dump_event(struct perf_session *session, event_t *event,
 682                        u64 file_offset, struct sample_data *sample)
 683 {
 684         if (!dump_trace)
 685                 return;
 686
 687         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 688                file_offset, event->header.size, event->header.type);
 689
 690         trace_event(event);
 691
 692         if (sample)
 693                 perf_session__print_tstamp(session, event, sample);
 694
 695         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 696                event->header.size, event__get_event_name(event->header.type));
 697 }
 698
 699 static void dump_sample(struct perf_session *session, event_t *event,
 700                         struct sample_data *sample)
 701 {
 702         if (!dump_trace)
 703                 return;
 704
 705         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 "\n",
 706                event->header.misc, sample->pid, sample->tid, sample->ip,
 707                sample->period);
 708
 709         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 710                 callchain__printf(sample);
 711 }
 712
 713 static int perf_session_deliver_event(struct perf_session *session,
 714                                       event_t *event,
 715                                       struct sample_data *sample,
 716                                       struct perf_event_ops *ops,
 717                                       u64 file_offset)
 718 {
 719         dump_event(session, event, file_offset, sample);
 720
 721         switch (event->header.type) {
 722         case PERF_RECORD_SAMPLE:
 723                 dump_sample(session, event, sample);
 724                 return ops->sample(event, sample, session);
 725         case PERF_RECORD_MMAP:
 726                 return ops->mmap(event, sample, session);
 727         case PERF_RECORD_COMM:
 728                 return ops->comm(event, sample, session);
 729         case PERF_RECORD_FORK:
 730                 return ops->fork(event, sample, session);
 731         case PERF_RECORD_EXIT:
 732                 return ops->exit(event, sample, session);
 733         case PERF_RECORD_LOST:
 734                 return ops->lost(event, sample, session);
 735         case PERF_RECORD_READ:
 736                 return ops->read(event, sample, session);
 737         case PERF_RECORD_THROTTLE:
 738                 return ops->throttle(event, sample, session);
 739         case PERF_RECORD_UNTHROTTLE:
 740                 return ops->unthrottle(event, sample, session);
 741         default:
 742                 ++session->hists.stats.nr_unknown_events;
 743                 return -1;
 744         }
 745 }
 746
 747 static int perf_session__preprocess_sample(struct perf_session *session,
 748                                            event_t *event, struct sample_data *sample)
 749 {
 750         if (event->header.type != PERF_RECORD_SAMPLE ||
 751             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 752                 return 0;
 753
 754         if (!ip_callchain__valid(sample->callchain, event)) {
 755                 pr_debug("call-chain problem with event, skipping it.\n");
 756                 ++session->hists.stats.nr_invalid_chains;
 757                 session->hists.stats.total_invalid_chains += sample->period;
 758                 return -EINVAL;
 759         }
 760         return 0;
 761 }
 762
 763 static int perf_session__process_user_event(struct perf_session *session, event_t *event,
 764                                             struct perf_event_ops *ops, u64 file_offset)
 765 {
 766         dump_event(session, event, file_offset, NULL);
 767
 768         /* These events are processed right away */
 769         switch (event->header.type) {
 770         case PERF_RECORD_HEADER_ATTR:
 771                 return ops->attr(event, session);
 772         case PERF_RECORD_HEADER_EVENT_TYPE:
 773                 return ops->event_type(event, session);
 774         case PERF_RECORD_HEADER_TRACING_DATA:
 775                 /* setup for reading amidst mmap */
 776                 lseek(session->fd, file_offset, SEEK_SET);
 777                 return ops->tracing_data(event, session);
 778         case PERF_RECORD_HEADER_BUILD_ID:
 779                 return ops->build_id(event, session);
 780         case PERF_RECORD_FINISHED_ROUND:
 781                 return ops->finished_round(event, session, ops);
 782         default:
 783                 return -EINVAL;
 784         }
 785 }
 786
 787 static int perf_session__process_event(struct perf_session *session,
 788                                        event_t *event,
 789                                        struct perf_event_ops *ops,
 790                                        u64 file_offset)
 791 {
 792         struct sample_data sample;
 793         int ret;
 794
 795         if (session->header.needs_swap && event__swap_ops[event->header.type])
 796                 event__swap_ops[event->header.type](event);
 797
 798         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 799                 return -EINVAL;
 800
 801         hists__inc_nr_events(&session->hists, event->header.type);
 802
 803         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 804                 return perf_session__process_user_event(session, event, ops, file_offset);
 805
 806         /*
 807          * For all kernel events we get the sample data
 808          */
 809         event__parse_sample(event, session, &sample);
 810
 811         /* Preprocess sample records - precheck callchains */
 812         if (perf_session__preprocess_sample(session, event, &sample))
 813                 return 0;
 814
 815         if (ops->ordered_samples) {
 816                 ret = perf_session_queue_event(session, event, &sample,
 817                                                file_offset);
 818                 if (ret != -ETIME)
 819                         return ret;
 820         }
 821
 822         return perf_session_deliver_event(session, event, &sample, ops,
 823                                           file_offset);
 824 }
 825
 826 void perf_event_header__bswap(struct perf_event_header *self)
 827 {
 828         self->type = bswap_32(self->type);
 829         self->misc = bswap_16(self->misc);
 830         self->size = bswap_16(self->size);
 831 }
 832
 833 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 834 {
 835         struct thread *thread = perf_session__findnew(self, 0);
 836
 837         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 838                 pr_err("problem inserting idle task.\n");
 839                 thread = NULL;
 840         }
 841
 842         return thread;
 843 }
 844
 845 static void perf_session__warn_about_errors(const struct perf_session *session,
 846                                             const struct perf_event_ops *ops)
 847 {
 848         if (ops->lost == event__process_lost &&
 849             session->hists.stats.total_lost != 0) {
 850                 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64
 851                             "!\n\nCheck IO/CPU overload!\n\n",
 852                             session->hists.stats.total_period,
 853                             session->hists.stats.total_lost);
 854         }
 855
 856         if (session->hists.stats.nr_unknown_events != 0) {
 857                 ui__warning("Found %u unknown events!\n\n"
 858                             "Is this an older tool processing a perf.data "
 859                             "file generated by a more recent tool?\n\n"
 860                             "If that is not the case, consider "
 861                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 862                             session->hists.stats.nr_unknown_events);
 863         }
 864
 865         if (session->hists.stats.nr_invalid_chains != 0) {
 866                 ui__warning("Found invalid callchains!\n\n"
 867                             "%u out of %u events were discarded for this reason.\n\n"
 868                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 869                             session->hists.stats.nr_invalid_chains,
 870                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 871         }
 872 }
 873
 874 #define session_done()  (*(volatile int *)(&session_done))
 875 volatile int session_done;
 876
 877 static int __perf_session__process_pipe_events(struct perf_session *self,
 878                                                struct perf_event_ops *ops)
 879 {
 880         event_t event;
 881         uint32_t size;
 882         int skip = 0;
 883         u64 head;
 884         int err;
 885         void *p;
 886
 887         perf_event_ops__fill_defaults(ops);
 888
 889         head = 0;
 890 more:
 891         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 892         if (err <= 0) {
 893                 if (err == 0)
 894                         goto done;
 895
 896                 pr_err("failed to read event header\n");
 897                 goto out_err;
 898         }
 899
 900         if (self->header.needs_swap)
 901                 perf_event_header__bswap(&event.header);
 902
 903         size = event.header.size;
 904         if (size == 0)
 905                 size = 8;
 906
 907         p = &event;
 908         p += sizeof(struct perf_event_header);
 909
 910         if (size - sizeof(struct perf_event_header)) {
 911                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 912                 if (err <= 0) {
 913                         if (err == 0) {
 914                                 pr_err("unexpected end of event stream\n");
 915                                 goto done;
 916                         }
 917
 918                         pr_err("failed to read event data\n");
 919                         goto out_err;
 920                 }
 921         }
 922
 923         if (size == 0 ||
 924             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 925                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 926                             head, event.header.size, event.header.type);
 927                 /*
 928                  * assume we lost track of the stream, check alignment, and
 929                  * increment a single u64 in the hope to catch on again 'soon'.
 930                  */
 931                 if (unlikely(head & 7))
 932                         head &= ~7ULL;
 933
 934                 size = 8;
 935         }
 936
 937         head += size;
 938
 939         if (skip > 0)
 940                 head += skip;
 941
 942         if (!session_done())
 943                 goto more;
 944 done:
 945         err = 0;
 946 out_err:
 947         perf_session__warn_about_errors(self, ops);
 948         perf_session_free_sample_buffers(self);
 949         return err;
 950 }
 951
 952 int __perf_session__process_events(struct perf_session *session,
 953                                    u64 data_offset, u64 data_size,
 954                                    u64 file_size, struct perf_event_ops *ops)
 955 {
 956         u64 head, page_offset, file_offset, file_pos, progress_next;
 957         int err, mmap_prot, mmap_flags, map_idx = 0;
 958         struct ui_progress *progress;
 959         size_t  page_size, mmap_size;
 960         char *buf, *mmaps[8];
 961         event_t *event;
 962         uint32_t size;
 963
 964         perf_event_ops__fill_defaults(ops);
 965
 966         page_size = sysconf(_SC_PAGESIZE);
 967
 968         page_offset = page_size * (data_offset / page_size);
 969         file_offset = page_offset;
 970         head = data_offset - page_offset;
 971
 972         if (data_offset + data_size < file_size)
 973                 file_size = data_offset + data_size;
 974
 975         progress_next = file_size / 16;
 976         progress = ui_progress__new("Processing events...", file_size);
 977         if (progress == NULL)
 978                 return -1;
 979
 980         mmap_size = session->mmap_window;
 981         if (mmap_size > file_size)
 982                 mmap_size = file_size;
 983
 984         memset(mmaps, 0, sizeof(mmaps));
 985
 986         mmap_prot  = PROT_READ;
 987         mmap_flags = MAP_SHARED;
 988
 989         if (session->header.needs_swap) {
 990                 mmap_prot  |= PROT_WRITE;
 991                 mmap_flags = MAP_PRIVATE;
 992         }
 993 remap:
 994         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
 995                    file_offset);
 996         if (buf == MAP_FAILED) {
 997                 pr_err("failed to mmap file\n");
 998                 err = -errno;
 999                 goto out_err;
1000         }
1001         mmaps[map_idx] = buf;
1002         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1003         file_pos = file_offset + head;
1004
1005 more:
1006         event = (event_t *)(buf + head);
1007
1008         if (session->header.needs_swap)
1009                 perf_event_header__bswap(&event->header);
1010         size = event->header.size;
1011         if (size == 0)
1012                 size = 8;
1013
1014         if (head + event->header.size > mmap_size) {
1015                 if (mmaps[map_idx]) {
1016                         munmap(mmaps[map_idx], mmap_size);
1017                         mmaps[map_idx] = NULL;
1018                 }
1019
1020                 page_offset = page_size * (head / page_size);
1021                 file_offset += page_offset;
1022                 head -= page_offset;
1023                 goto remap;
1024         }
1025
1026         size = event->header.size;
1027
1028         if (size == 0 ||
1029             perf_session__process_event(session, event, ops, file_pos) < 0) {
1030                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1031                             file_offset + head, event->header.size,
1032                             event->header.type);
1033                 /*
1034                  * assume we lost track of the stream, check alignment, and
1035                  * increment a single u64 in the hope to catch on again 'soon'.
1036                  */
1037                 if (unlikely(head & 7))
1038                         head &= ~7ULL;
1039
1040                 size = 8;
1041         }
1042
1043         head += size;
1044         file_pos += size;
1045
1046         if (file_pos >= progress_next) {
1047                 progress_next += file_size / 16;
1048                 ui_progress__update(progress, file_pos);
1049         }
1050
1051         if (file_pos < file_size)
1052                 goto more;
1053
1054         err = 0;
1055         /* do the final flush for ordered samples */
1056         session->ordered_samples.next_flush = ULLONG_MAX;
1057         flush_sample_queue(session, ops);
1058 out_err:
1059         ui_progress__delete(progress);
1060         perf_session__warn_about_errors(session, ops);
1061         perf_session_free_sample_buffers(session);
1062         return err;
1063 }
1064
1065 int perf_session__process_events(struct perf_session *self,
1066                                  struct perf_event_ops *ops)
1067 {
1068         int err;
1069
1070         if (perf_session__register_idle_thread(self) == NULL)
1071                 return -ENOMEM;
1072
1073         if (!self->fd_pipe)
1074                 err = __perf_session__process_events(self,
1075                                                      self->header.data_offset,
1076                                                      self->header.data_size,
1077                                                      self->size, ops);
1078         else
1079                 err = __perf_session__process_pipe_events(self, ops);
1080
1081         return err;
1082 }
1083
1084 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1085 {
1086         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1087                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1088                 return false;
1089         }
1090
1091         return true;
1092 }
1093
1094 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1095                                              const char *symbol_name,
1096                                              u64 addr)
1097 {
1098         char *bracket;
1099         enum map_type i;
1100         struct ref_reloc_sym *ref;
1101
1102         ref = zalloc(sizeof(struct ref_reloc_sym));
1103         if (ref == NULL)
1104                 return -ENOMEM;
1105
1106         ref->name = strdup(symbol_name);
1107         if (ref->name == NULL) {
1108                 free(ref);
1109                 return -ENOMEM;
1110         }
1111
1112         bracket = strchr(ref->name, ']');
1113         if (bracket)
1114                 *bracket = '\0';
1115
1116         ref->addr = addr;
1117
1118         for (i = 0; i < MAP__NR_TYPES; ++i) {
1119                 struct kmap *kmap = map__kmap(maps[i]);
1120                 kmap->ref_reloc_sym = ref;
1121         }
1122
1123         return 0;
1124 }
1125
1126 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1127 {
1128         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1129                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1130                machines__fprintf_dsos(&self->machines, fp);
1131 }
1132
1133 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1134                                           bool with_hits)
1135 {
1136         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1137         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1138 }