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 struct map_symbol *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         struct map_symbol *syms = calloc(chain->nr, sizeof(*syms));
 253
 254         if (!syms)
 255                 return NULL;
 256
 257         for (i = 0; i < chain->nr; i++) {
 258                 u64 ip = chain->ips[i];
 259                 struct addr_location al;
 260
 261                 if (ip >= PERF_CONTEXT_MAX) {
 262                         switch (ip) {
 263                         case PERF_CONTEXT_HV:
 264                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 265                         case PERF_CONTEXT_KERNEL:
 266                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 267                         case PERF_CONTEXT_USER:
 268                                 cpumode = PERF_RECORD_MISC_USER;        break;
 269                         default:
 270                                 break;
 271                         }
 272                         continue;
 273                 }
 274
 275                 al.filtered = false;
 276                 thread__find_addr_location(thread, self, cpumode,
 277                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 278                 if (al.sym != NULL) {
 279                         if (sort__has_parent && !*parent &&
 280                             symbol__match_parent_regex(al.sym))
 281                                 *parent = al.sym;
 282                         if (!symbol_conf.use_callchain)
 283                                 break;
 284                         syms[i].map = al.map;
 285                         syms[i].sym = al.sym;
 286                 }
 287         }
 288
 289         return syms;
 290 }
 291
 292 static int process_event_synth_stub(event_t *event __used,
 293                                     struct perf_session *session __used)
 294 {
 295         dump_printf(": unhandled!\n");
 296         return 0;
 297 }
 298
 299 static int process_event_stub(event_t *event __used,
 300                               struct sample_data *sample __used,
 301                               struct perf_session *session __used)
 302 {
 303         dump_printf(": unhandled!\n");
 304         return 0;
 305 }
 306
 307 static int process_finished_round_stub(event_t *event __used,
 308                                        struct perf_session *session __used,
 309                                        struct perf_event_ops *ops __used)
 310 {
 311         dump_printf(": unhandled!\n");
 312         return 0;
 313 }
 314
 315 static int process_finished_round(event_t *event,
 316                                   struct perf_session *session,
 317                                   struct perf_event_ops *ops);
 318
 319 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 320 {
 321         if (handler->sample == NULL)
 322                 handler->sample = process_event_stub;
 323         if (handler->mmap == NULL)
 324                 handler->mmap = process_event_stub;
 325         if (handler->comm == NULL)
 326                 handler->comm = process_event_stub;
 327         if (handler->fork == NULL)
 328                 handler->fork = process_event_stub;
 329         if (handler->exit == NULL)
 330                 handler->exit = process_event_stub;
 331         if (handler->lost == NULL)
 332                 handler->lost = event__process_lost;
 333         if (handler->read == NULL)
 334                 handler->read = process_event_stub;
 335         if (handler->throttle == NULL)
 336                 handler->throttle = process_event_stub;
 337         if (handler->unthrottle == NULL)
 338                 handler->unthrottle = process_event_stub;
 339         if (handler->attr == NULL)
 340                 handler->attr = process_event_synth_stub;
 341         if (handler->event_type == NULL)
 342                 handler->event_type = process_event_synth_stub;
 343         if (handler->tracing_data == NULL)
 344                 handler->tracing_data = process_event_synth_stub;
 345         if (handler->build_id == NULL)
 346                 handler->build_id = process_event_synth_stub;
 347         if (handler->finished_round == NULL) {
 348                 if (handler->ordered_samples)
 349                         handler->finished_round = process_finished_round;
 350                 else
 351                         handler->finished_round = process_finished_round_stub;
 352         }
 353 }
 354
 355 void mem_bswap_64(void *src, int byte_size)
 356 {
 357         u64 *m = src;
 358
 359         while (byte_size > 0) {
 360                 *m = bswap_64(*m);
 361                 byte_size -= sizeof(u64);
 362                 ++m;
 363         }
 364 }
 365
 366 static void event__all64_swap(event_t *self)
 367 {
 368         struct perf_event_header *hdr = &self->header;
 369         mem_bswap_64(hdr + 1, self->header.size - sizeof(*hdr));
 370 }
 371
 372 static void event__comm_swap(event_t *self)
 373 {
 374         self->comm.pid = bswap_32(self->comm.pid);
 375         self->comm.tid = bswap_32(self->comm.tid);
 376 }
 377
 378 static void event__mmap_swap(event_t *self)
 379 {
 380         self->mmap.pid   = bswap_32(self->mmap.pid);
 381         self->mmap.tid   = bswap_32(self->mmap.tid);
 382         self->mmap.start = bswap_64(self->mmap.start);
 383         self->mmap.len   = bswap_64(self->mmap.len);
 384         self->mmap.pgoff = bswap_64(self->mmap.pgoff);
 385 }
 386
 387 static void event__task_swap(event_t *self)
 388 {
 389         self->fork.pid  = bswap_32(self->fork.pid);
 390         self->fork.tid  = bswap_32(self->fork.tid);
 391         self->fork.ppid = bswap_32(self->fork.ppid);
 392         self->fork.ptid = bswap_32(self->fork.ptid);
 393         self->fork.time = bswap_64(self->fork.time);
 394 }
 395
 396 static void event__read_swap(event_t *self)
 397 {
 398         self->read.pid          = bswap_32(self->read.pid);
 399         self->read.tid          = bswap_32(self->read.tid);
 400         self->read.value        = bswap_64(self->read.value);
 401         self->read.time_enabled = bswap_64(self->read.time_enabled);
 402         self->read.time_running = bswap_64(self->read.time_running);
 403         self->read.id           = bswap_64(self->read.id);
 404 }
 405
 406 static void event__attr_swap(event_t *self)
 407 {
 408         size_t size;
 409
 410         self->attr.attr.type            = bswap_32(self->attr.attr.type);
 411         self->attr.attr.size            = bswap_32(self->attr.attr.size);
 412         self->attr.attr.config          = bswap_64(self->attr.attr.config);
 413         self->attr.attr.sample_period   = bswap_64(self->attr.attr.sample_period);
 414         self->attr.attr.sample_type     = bswap_64(self->attr.attr.sample_type);
 415         self->attr.attr.read_format     = bswap_64(self->attr.attr.read_format);
 416         self->attr.attr.wakeup_events   = bswap_32(self->attr.attr.wakeup_events);
 417         self->attr.attr.bp_type         = bswap_32(self->attr.attr.bp_type);
 418         self->attr.attr.bp_addr         = bswap_64(self->attr.attr.bp_addr);
 419         self->attr.attr.bp_len          = bswap_64(self->attr.attr.bp_len);
 420
 421         size = self->header.size;
 422         size -= (void *)&self->attr.id - (void *)self;
 423         mem_bswap_64(self->attr.id, size);
 424 }
 425
 426 static void event__event_type_swap(event_t *self)
 427 {
 428         self->event_type.event_type.event_id =
 429                 bswap_64(self->event_type.event_type.event_id);
 430 }
 431
 432 static void event__tracing_data_swap(event_t *self)
 433 {
 434         self->tracing_data.size = bswap_32(self->tracing_data.size);
 435 }
 436
 437 typedef void (*event__swap_op)(event_t *self);
 438
 439 static event__swap_op event__swap_ops[] = {
 440         [PERF_RECORD_MMAP]   = event__mmap_swap,
 441         [PERF_RECORD_COMM]   = event__comm_swap,
 442         [PERF_RECORD_FORK]   = event__task_swap,
 443         [PERF_RECORD_EXIT]   = event__task_swap,
 444         [PERF_RECORD_LOST]   = event__all64_swap,
 445         [PERF_RECORD_READ]   = event__read_swap,
 446         [PERF_RECORD_SAMPLE] = event__all64_swap,
 447         [PERF_RECORD_HEADER_ATTR]   = event__attr_swap,
 448         [PERF_RECORD_HEADER_EVENT_TYPE]   = event__event_type_swap,
 449         [PERF_RECORD_HEADER_TRACING_DATA]   = event__tracing_data_swap,
 450         [PERF_RECORD_HEADER_BUILD_ID]   = NULL,
 451         [PERF_RECORD_HEADER_MAX]    = NULL,
 452 };
 453
 454 struct sample_queue {
 455         u64                     timestamp;
 456         u64                     file_offset;
 457         event_t                 *event;
 458         struct list_head        list;
 459 };
 460
 461 static void perf_session_free_sample_buffers(struct perf_session *session)
 462 {
 463         struct ordered_samples *os = &session->ordered_samples;
 464
 465         while (!list_empty(&os->to_free)) {
 466                 struct sample_queue *sq;
 467
 468                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 469                 list_del(&sq->list);
 470                 free(sq);
 471         }
 472 }
 473
 474 static int perf_session_deliver_event(struct perf_session *session,
 475                                       event_t *event,
 476                                       struct sample_data *sample,
 477                                       struct perf_event_ops *ops,
 478                                       u64 file_offset);
 479
 480 static void flush_sample_queue(struct perf_session *s,
 481                                struct perf_event_ops *ops)
 482 {
 483         struct ordered_samples *os = &s->ordered_samples;
 484         struct list_head *head = &os->samples;
 485         struct sample_queue *tmp, *iter;
 486         struct sample_data sample;
 487         u64 limit = os->next_flush;
 488         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 489
 490         if (!ops->ordered_samples || !limit)
 491                 return;
 492
 493         list_for_each_entry_safe(iter, tmp, head, list) {
 494                 if (iter->timestamp > limit)
 495                         break;
 496
 497                 event__parse_sample(iter->event, s, &sample);
 498                 perf_session_deliver_event(s, iter->event, &sample, ops,
 499                                            iter->file_offset);
 500
 501                 os->last_flush = iter->timestamp;
 502                 list_del(&iter->list);
 503                 list_add(&iter->list, &os->sample_cache);
 504         }
 505
 506         if (list_empty(head)) {
 507                 os->last_sample = NULL;
 508         } else if (last_ts <= limit) {
 509                 os->last_sample =
 510                         list_entry(head->prev, struct sample_queue, list);
 511         }
 512 }
 513
 514 /*
 515  * When perf record finishes a pass on every buffers, it records this pseudo
 516  * event.
 517  * We record the max timestamp t found in the pass n.
 518  * Assuming these timestamps are monotonic across cpus, we know that if
 519  * a buffer still has events with timestamps below t, they will be all
 520  * available and then read in the pass n + 1.
 521  * Hence when we start to read the pass n + 2, we can safely flush every
 522  * events with timestamps below t.
 523  *
 524  *    ============ PASS n =================
 525  *       CPU 0         |   CPU 1
 526  *                     |
 527  *    cnt1 timestamps  |   cnt2 timestamps
 528  *          1          |         2
 529  *          2          |         3
 530  *          -          |         4  <--- max recorded
 531  *
 532  *    ============ PASS n + 1 ==============
 533  *       CPU 0         |   CPU 1
 534  *                     |
 535  *    cnt1 timestamps  |   cnt2 timestamps
 536  *          3          |         5
 537  *          4          |         6
 538  *          5          |         7 <---- max recorded
 539  *
 540  *      Flush every events below timestamp 4
 541  *
 542  *    ============ PASS n + 2 ==============
 543  *       CPU 0         |   CPU 1
 544  *                     |
 545  *    cnt1 timestamps  |   cnt2 timestamps
 546  *          6          |         8
 547  *          7          |         9
 548  *          -          |         10
 549  *
 550  *      Flush every events below timestamp 7
 551  *      etc...
 552  */
 553 static int process_finished_round(event_t *event __used,
 554                                   struct perf_session *session,
 555                                   struct perf_event_ops *ops)
 556 {
 557         flush_sample_queue(session, ops);
 558         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 559
 560         return 0;
 561 }
 562
 563 /* The queue is ordered by time */
 564 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 565 {
 566         struct ordered_samples *os = &s->ordered_samples;
 567         struct sample_queue *sample = os->last_sample;
 568         u64 timestamp = new->timestamp;
 569         struct list_head *p;
 570
 571         os->last_sample = new;
 572
 573         if (!sample) {
 574                 list_add(&new->list, &os->samples);
 575                 os->max_timestamp = timestamp;
 576                 return;
 577         }
 578
 579         /*
 580          * last_sample might point to some random place in the list as it's
 581          * the last queued event. We expect that the new event is close to
 582          * this.
 583          */
 584         if (sample->timestamp <= timestamp) {
 585                 while (sample->timestamp <= timestamp) {
 586                         p = sample->list.next;
 587                         if (p == &os->samples) {
 588                                 list_add_tail(&new->list, &os->samples);
 589                                 os->max_timestamp = timestamp;
 590                                 return;
 591                         }
 592                         sample = list_entry(p, struct sample_queue, list);
 593                 }
 594                 list_add_tail(&new->list, &sample->list);
 595         } else {
 596                 while (sample->timestamp > timestamp) {
 597                         p = sample->list.prev;
 598                         if (p == &os->samples) {
 599                                 list_add(&new->list, &os->samples);
 600                                 return;
 601                         }
 602                         sample = list_entry(p, struct sample_queue, list);
 603                 }
 604                 list_add(&new->list, &sample->list);
 605         }
 606 }
 607
 608 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 609
 610 static int perf_session_queue_event(struct perf_session *s, event_t *event,
 611                                     struct sample_data *data, u64 file_offset)
 612 {
 613         struct ordered_samples *os = &s->ordered_samples;
 614         struct list_head *sc = &os->sample_cache;
 615         u64 timestamp = data->time;
 616         struct sample_queue *new;
 617
 618         if (!timestamp || timestamp == ~0ULL)
 619                 return -ETIME;
 620
 621         if (timestamp < s->ordered_samples.last_flush) {
 622                 printf("Warning: Timestamp below last timeslice flush\n");
 623                 return -EINVAL;
 624         }
 625
 626         if (!list_empty(sc)) {
 627                 new = list_entry(sc->next, struct sample_queue, list);
 628                 list_del(&new->list);
 629         } else if (os->sample_buffer) {
 630                 new = os->sample_buffer + os->sample_buffer_idx;
 631                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 632                         os->sample_buffer = NULL;
 633         } else {
 634                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 635                 if (!os->sample_buffer)
 636                         return -ENOMEM;
 637                 list_add(&os->sample_buffer->list, &os->to_free);
 638                 os->sample_buffer_idx = 2;
 639                 new = os->sample_buffer + 1;
 640         }
 641
 642         new->timestamp = timestamp;
 643         new->file_offset = file_offset;
 644         new->event = event;
 645
 646         __queue_event(new, s);
 647
 648         return 0;
 649 }
 650
 651 static void callchain__printf(struct sample_data *sample)
 652 {
 653         unsigned int i;
 654
 655         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 656
 657         for (i = 0; i < sample->callchain->nr; i++)
 658                 printf("..... %2d: %016" PRIx64 "\n",
 659                        i, sample->callchain->ips[i]);
 660 }
 661
 662 static void perf_session__print_tstamp(struct perf_session *session,
 663                                        event_t *event,
 664                                        struct sample_data *sample)
 665 {
 666         if (event->header.type != PERF_RECORD_SAMPLE &&
 667             !session->sample_id_all) {
 668                 fputs("-1 -1 ", stdout);
 669                 return;
 670         }
 671
 672         if ((session->sample_type & PERF_SAMPLE_CPU))
 673                 printf("%u ", sample->cpu);
 674
 675         if (session->sample_type & PERF_SAMPLE_TIME)
 676                 printf("%" PRIu64 " ", sample->time);
 677 }
 678
 679 static void dump_event(struct perf_session *session, event_t *event,
 680                        u64 file_offset, struct sample_data *sample)
 681 {
 682         if (!dump_trace)
 683                 return;
 684
 685         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 686                file_offset, event->header.size, event->header.type);
 687
 688         trace_event(event);
 689
 690         if (sample)
 691                 perf_session__print_tstamp(session, event, sample);
 692
 693         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 694                event->header.size, event__get_event_name(event->header.type));
 695 }
 696
 697 static void dump_sample(struct perf_session *session, event_t *event,
 698                         struct sample_data *sample)
 699 {
 700         if (!dump_trace)
 701                 return;
 702
 703         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 "\n",
 704                event->header.misc, sample->pid, sample->tid, sample->ip,
 705                sample->period);
 706
 707         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 708                 callchain__printf(sample);
 709 }
 710
 711 static int perf_session_deliver_event(struct perf_session *session,
 712                                       event_t *event,
 713                                       struct sample_data *sample,
 714                                       struct perf_event_ops *ops,
 715                                       u64 file_offset)
 716 {
 717         dump_event(session, event, file_offset, sample);
 718
 719         switch (event->header.type) {
 720         case PERF_RECORD_SAMPLE:
 721                 dump_sample(session, event, sample);
 722                 return ops->sample(event, sample, session);
 723         case PERF_RECORD_MMAP:
 724                 return ops->mmap(event, sample, session);
 725         case PERF_RECORD_COMM:
 726                 return ops->comm(event, sample, session);
 727         case PERF_RECORD_FORK:
 728                 return ops->fork(event, sample, session);
 729         case PERF_RECORD_EXIT:
 730                 return ops->exit(event, sample, session);
 731         case PERF_RECORD_LOST:
 732                 return ops->lost(event, sample, session);
 733         case PERF_RECORD_READ:
 734                 return ops->read(event, sample, session);
 735         case PERF_RECORD_THROTTLE:
 736                 return ops->throttle(event, sample, session);
 737         case PERF_RECORD_UNTHROTTLE:
 738                 return ops->unthrottle(event, sample, session);
 739         default:
 740                 ++session->hists.stats.nr_unknown_events;
 741                 return -1;
 742         }
 743 }
 744
 745 static int perf_session__preprocess_sample(struct perf_session *session,
 746                                            event_t *event, struct sample_data *sample)
 747 {
 748         if (event->header.type != PERF_RECORD_SAMPLE ||
 749             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 750                 return 0;
 751
 752         if (!ip_callchain__valid(sample->callchain, event)) {
 753                 pr_debug("call-chain problem with event, skipping it.\n");
 754                 ++session->hists.stats.nr_invalid_chains;
 755                 session->hists.stats.total_invalid_chains += sample->period;
 756                 return -EINVAL;
 757         }
 758         return 0;
 759 }
 760
 761 static int perf_session__process_user_event(struct perf_session *session, event_t *event,
 762                                             struct perf_event_ops *ops, u64 file_offset)
 763 {
 764         dump_event(session, event, file_offset, NULL);
 765
 766         /* These events are processed right away */
 767         switch (event->header.type) {
 768         case PERF_RECORD_HEADER_ATTR:
 769                 return ops->attr(event, session);
 770         case PERF_RECORD_HEADER_EVENT_TYPE:
 771                 return ops->event_type(event, session);
 772         case PERF_RECORD_HEADER_TRACING_DATA:
 773                 /* setup for reading amidst mmap */
 774                 lseek(session->fd, file_offset, SEEK_SET);
 775                 return ops->tracing_data(event, session);
 776         case PERF_RECORD_HEADER_BUILD_ID:
 777                 return ops->build_id(event, session);
 778         case PERF_RECORD_FINISHED_ROUND:
 779                 return ops->finished_round(event, session, ops);
 780         default:
 781                 return -EINVAL;
 782         }
 783 }
 784
 785 static int perf_session__process_event(struct perf_session *session,
 786                                        event_t *event,
 787                                        struct perf_event_ops *ops,
 788                                        u64 file_offset)
 789 {
 790         struct sample_data sample;
 791         int ret;
 792
 793         if (session->header.needs_swap && event__swap_ops[event->header.type])
 794                 event__swap_ops[event->header.type](event);
 795
 796         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 797                 return -EINVAL;
 798
 799         hists__inc_nr_events(&session->hists, event->header.type);
 800
 801         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 802                 return perf_session__process_user_event(session, event, ops, file_offset);
 803
 804         /*
 805          * For all kernel events we get the sample data
 806          */
 807         event__parse_sample(event, session, &sample);
 808
 809         /* Preprocess sample records - precheck callchains */
 810         if (perf_session__preprocess_sample(session, event, &sample))
 811                 return 0;
 812
 813         if (ops->ordered_samples) {
 814                 ret = perf_session_queue_event(session, event, &sample,
 815                                                file_offset);
 816                 if (ret != -ETIME)
 817                         return ret;
 818         }
 819
 820         return perf_session_deliver_event(session, event, &sample, ops,
 821                                           file_offset);
 822 }
 823
 824 void perf_event_header__bswap(struct perf_event_header *self)
 825 {
 826         self->type = bswap_32(self->type);
 827         self->misc = bswap_16(self->misc);
 828         self->size = bswap_16(self->size);
 829 }
 830
 831 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 832 {
 833         struct thread *thread = perf_session__findnew(self, 0);
 834
 835         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 836                 pr_err("problem inserting idle task.\n");
 837                 thread = NULL;
 838         }
 839
 840         return thread;
 841 }
 842
 843 static void perf_session__warn_about_errors(const struct perf_session *session,
 844                                             const struct perf_event_ops *ops)
 845 {
 846         if (ops->lost == event__process_lost &&
 847             session->hists.stats.total_lost != 0) {
 848                 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64
 849                             "!\n\nCheck IO/CPU overload!\n\n",
 850                             session->hists.stats.total_period,
 851                             session->hists.stats.total_lost);
 852         }
 853
 854         if (session->hists.stats.nr_unknown_events != 0) {
 855                 ui__warning("Found %u unknown events!\n\n"
 856                             "Is this an older tool processing a perf.data "
 857                             "file generated by a more recent tool?\n\n"
 858                             "If that is not the case, consider "
 859                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 860                             session->hists.stats.nr_unknown_events);
 861         }
 862
 863         if (session->hists.stats.nr_invalid_chains != 0) {
 864                 ui__warning("Found invalid callchains!\n\n"
 865                             "%u out of %u events were discarded for this reason.\n\n"
 866                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 867                             session->hists.stats.nr_invalid_chains,
 868                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 869         }
 870 }
 871
 872 #define session_done()  (*(volatile int *)(&session_done))
 873 volatile int session_done;
 874
 875 static int __perf_session__process_pipe_events(struct perf_session *self,
 876                                                struct perf_event_ops *ops)
 877 {
 878         event_t event;
 879         uint32_t size;
 880         int skip = 0;
 881         u64 head;
 882         int err;
 883         void *p;
 884
 885         perf_event_ops__fill_defaults(ops);
 886
 887         head = 0;
 888 more:
 889         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 890         if (err <= 0) {
 891                 if (err == 0)
 892                         goto done;
 893
 894                 pr_err("failed to read event header\n");
 895                 goto out_err;
 896         }
 897
 898         if (self->header.needs_swap)
 899                 perf_event_header__bswap(&event.header);
 900
 901         size = event.header.size;
 902         if (size == 0)
 903                 size = 8;
 904
 905         p = &event;
 906         p += sizeof(struct perf_event_header);
 907
 908         if (size - sizeof(struct perf_event_header)) {
 909                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 910                 if (err <= 0) {
 911                         if (err == 0) {
 912                                 pr_err("unexpected end of event stream\n");
 913                                 goto done;
 914                         }
 915
 916                         pr_err("failed to read event data\n");
 917                         goto out_err;
 918                 }
 919         }
 920
 921         if (size == 0 ||
 922             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 923                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 924                             head, event.header.size, event.header.type);
 925                 /*
 926                  * assume we lost track of the stream, check alignment, and
 927                  * increment a single u64 in the hope to catch on again 'soon'.
 928                  */
 929                 if (unlikely(head & 7))
 930                         head &= ~7ULL;
 931
 932                 size = 8;
 933         }
 934
 935         head += size;
 936
 937         if (skip > 0)
 938                 head += skip;
 939
 940         if (!session_done())
 941                 goto more;
 942 done:
 943         err = 0;
 944 out_err:
 945         perf_session__warn_about_errors(self, ops);
 946         perf_session_free_sample_buffers(self);
 947         return err;
 948 }
 949
 950 int __perf_session__process_events(struct perf_session *session,
 951                                    u64 data_offset, u64 data_size,
 952                                    u64 file_size, struct perf_event_ops *ops)
 953 {
 954         u64 head, page_offset, file_offset, file_pos, progress_next;
 955         int err, mmap_prot, mmap_flags, map_idx = 0;
 956         struct ui_progress *progress;
 957         size_t  page_size, mmap_size;
 958         char *buf, *mmaps[8];
 959         event_t *event;
 960         uint32_t size;
 961
 962         perf_event_ops__fill_defaults(ops);
 963
 964         page_size = sysconf(_SC_PAGESIZE);
 965
 966         page_offset = page_size * (data_offset / page_size);
 967         file_offset = page_offset;
 968         head = data_offset - page_offset;
 969
 970         if (data_offset + data_size < file_size)
 971                 file_size = data_offset + data_size;
 972
 973         progress_next = file_size / 16;
 974         progress = ui_progress__new("Processing events...", file_size);
 975         if (progress == NULL)
 976                 return -1;
 977
 978         mmap_size = session->mmap_window;
 979         if (mmap_size > file_size)
 980                 mmap_size = file_size;
 981
 982         memset(mmaps, 0, sizeof(mmaps));
 983
 984         mmap_prot  = PROT_READ;
 985         mmap_flags = MAP_SHARED;
 986
 987         if (session->header.needs_swap) {
 988                 mmap_prot  |= PROT_WRITE;
 989                 mmap_flags = MAP_PRIVATE;
 990         }
 991 remap:
 992         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
 993                    file_offset);
 994         if (buf == MAP_FAILED) {
 995                 pr_err("failed to mmap file\n");
 996                 err = -errno;
 997                 goto out_err;
 998         }
 999         mmaps[map_idx] = buf;
1000         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1001         file_pos = file_offset + head;
1002
1003 more:
1004         event = (event_t *)(buf + head);
1005
1006         if (session->header.needs_swap)
1007                 perf_event_header__bswap(&event->header);
1008         size = event->header.size;
1009         if (size == 0)
1010                 size = 8;
1011
1012         if (head + event->header.size > mmap_size) {
1013                 if (mmaps[map_idx]) {
1014                         munmap(mmaps[map_idx], mmap_size);
1015                         mmaps[map_idx] = NULL;
1016                 }
1017
1018                 page_offset = page_size * (head / page_size);
1019                 file_offset += page_offset;
1020                 head -= page_offset;
1021                 goto remap;
1022         }
1023
1024         size = event->header.size;
1025
1026         if (size == 0 ||
1027             perf_session__process_event(session, event, ops, file_pos) < 0) {
1028                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1029                             file_offset + head, event->header.size,
1030                             event->header.type);
1031                 /*
1032                  * assume we lost track of the stream, check alignment, and
1033                  * increment a single u64 in the hope to catch on again 'soon'.
1034                  */
1035                 if (unlikely(head & 7))
1036                         head &= ~7ULL;
1037
1038                 size = 8;
1039         }
1040
1041         head += size;
1042         file_pos += size;
1043
1044         if (file_pos >= progress_next) {
1045                 progress_next += file_size / 16;
1046                 ui_progress__update(progress, file_pos);
1047         }
1048
1049         if (file_pos < file_size)
1050                 goto more;
1051
1052         err = 0;
1053         /* do the final flush for ordered samples */
1054         session->ordered_samples.next_flush = ULLONG_MAX;
1055         flush_sample_queue(session, ops);
1056 out_err:
1057         ui_progress__delete(progress);
1058         perf_session__warn_about_errors(session, ops);
1059         perf_session_free_sample_buffers(session);
1060         return err;
1061 }
1062
1063 int perf_session__process_events(struct perf_session *self,
1064                                  struct perf_event_ops *ops)
1065 {
1066         int err;
1067
1068         if (perf_session__register_idle_thread(self) == NULL)
1069                 return -ENOMEM;
1070
1071         if (!self->fd_pipe)
1072                 err = __perf_session__process_events(self,
1073                                                      self->header.data_offset,
1074                                                      self->header.data_size,
1075                                                      self->size, ops);
1076         else
1077                 err = __perf_session__process_pipe_events(self, ops);
1078
1079         return err;
1080 }
1081
1082 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1083 {
1084         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1085                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1086                 return false;
1087         }
1088
1089         return true;
1090 }
1091
1092 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1093                                              const char *symbol_name,
1094                                              u64 addr)
1095 {
1096         char *bracket;
1097         enum map_type i;
1098         struct ref_reloc_sym *ref;
1099
1100         ref = zalloc(sizeof(struct ref_reloc_sym));
1101         if (ref == NULL)
1102                 return -ENOMEM;
1103
1104         ref->name = strdup(symbol_name);
1105         if (ref->name == NULL) {
1106                 free(ref);
1107                 return -ENOMEM;
1108         }
1109
1110         bracket = strchr(ref->name, ']');
1111         if (bracket)
1112                 *bracket = '\0';
1113
1114         ref->addr = addr;
1115
1116         for (i = 0; i < MAP__NR_TYPES; ++i) {
1117                 struct kmap *kmap = map__kmap(maps[i]);
1118                 kmap->ref_reloc_sym = ref;
1119         }
1120
1121         return 0;
1122 }
1123
1124 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1125 {
1126         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1127                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1128                machines__fprintf_dsos(&self->machines, fp);
1129 }
1130
1131 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1132                                           bool with_hits)
1133 {
1134         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1135         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1136 }