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 void perf_session__update_sample_type(struct perf_session *self)
  69 {
  70         self->sample_type = perf_header__sample_type(&self->header);
  71 }
  72
  73 int perf_session__create_kernel_maps(struct perf_session *self)
  74 {
  75         int ret = machine__create_kernel_maps(&self->host_machine);
  76
  77         if (ret >= 0)
  78                 ret = machines__create_guest_kernel_maps(&self->machines);
  79         return ret;
  80 }
  81
  82 static void perf_session__destroy_kernel_maps(struct perf_session *self)
  83 {
  84         machine__destroy_kernel_maps(&self->host_machine);
  85         machines__destroy_guest_kernel_maps(&self->machines);
  86 }
  87
  88 struct perf_session *perf_session__new(const char *filename, int mode, bool force, bool repipe)
  89 {
  90         size_t len = filename ? strlen(filename) + 1 : 0;
  91         struct perf_session *self = zalloc(sizeof(*self) + len);
  92
  93         if (self == NULL)
  94                 goto out;
  95
  96         if (perf_header__init(&self->header) < 0)
  97                 goto out_free;
  98
  99         memcpy(self->filename, filename, len);
 100         self->threads = RB_ROOT;
 101         INIT_LIST_HEAD(&self->dead_threads);
 102         self->hists_tree = RB_ROOT;
 103         self->last_match = NULL;
 104         self->mmap_window = 32;
 105         self->machines = RB_ROOT;
 106         self->repipe = repipe;
 107         INIT_LIST_HEAD(&self->ordered_samples.samples_head);
 108         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 109
 110         if (mode == O_RDONLY) {
 111                 if (perf_session__open(self, force) < 0)
 112                         goto out_delete;
 113         } else if (mode == O_WRONLY) {
 114                 /*
 115                  * In O_RDONLY mode this will be performed when reading the
 116                  * kernel MMAP event, in event__process_mmap().
 117                  */
 118                 if (perf_session__create_kernel_maps(self) < 0)
 119                         goto out_delete;
 120         }
 121
 122         perf_session__update_sample_type(self);
 123 out:
 124         return self;
 125 out_free:
 126         free(self);
 127         return NULL;
 128 out_delete:
 129         perf_session__delete(self);
 130         return NULL;
 131 }
 132
 133 static void perf_session__delete_dead_threads(struct perf_session *self)
 134 {
 135         struct thread *n, *t;
 136
 137         list_for_each_entry_safe(t, n, &self->dead_threads, node) {
 138                 list_del(&t->node);
 139                 thread__delete(t);
 140         }
 141 }
 142
 143 static void perf_session__delete_threads(struct perf_session *self)
 144 {
 145         struct rb_node *nd = rb_first(&self->threads);
 146
 147         while (nd) {
 148                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 149
 150                 rb_erase(&t->rb_node, &self->threads);
 151                 nd = rb_next(nd);
 152                 thread__delete(t);
 153         }
 154 }
 155
 156 void perf_session__delete(struct perf_session *self)
 157 {
 158         perf_header__exit(&self->header);
 159         perf_session__destroy_kernel_maps(self);
 160         perf_session__delete_dead_threads(self);
 161         perf_session__delete_threads(self);
 162         machine__exit(&self->host_machine);
 163         close(self->fd);
 164         free(self);
 165 }
 166
 167 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
 168 {
 169         self->last_match = NULL;
 170         rb_erase(&th->rb_node, &self->threads);
 171         /*
 172          * We may have references to this thread, for instance in some hist_entry
 173          * instances, so just move them to a separate list.
 174          */
 175         list_add_tail(&th->node, &self->dead_threads);
 176 }
 177
 178 static bool symbol__match_parent_regex(struct symbol *sym)
 179 {
 180         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 181                 return 1;
 182
 183         return 0;
 184 }
 185
 186 struct map_symbol *perf_session__resolve_callchain(struct perf_session *self,
 187                                                    struct thread *thread,
 188                                                    struct ip_callchain *chain,
 189                                                    struct symbol **parent)
 190 {
 191         u8 cpumode = PERF_RECORD_MISC_USER;
 192         unsigned int i;
 193         struct map_symbol *syms = calloc(chain->nr, sizeof(*syms));
 194
 195         if (!syms)
 196                 return NULL;
 197
 198         for (i = 0; i < chain->nr; i++) {
 199                 u64 ip = chain->ips[i];
 200                 struct addr_location al;
 201
 202                 if (ip >= PERF_CONTEXT_MAX) {
 203                         switch (ip) {
 204                         case PERF_CONTEXT_HV:
 205                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 206                         case PERF_CONTEXT_KERNEL:
 207                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 208                         case PERF_CONTEXT_USER:
 209                                 cpumode = PERF_RECORD_MISC_USER;        break;
 210                         default:
 211                                 break;
 212                         }
 213                         continue;
 214                 }
 215
 216                 al.filtered = false;
 217                 thread__find_addr_location(thread, self, cpumode,
 218                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 219                 if (al.sym != NULL) {
 220                         if (sort__has_parent && !*parent &&
 221                             symbol__match_parent_regex(al.sym))
 222                                 *parent = al.sym;
 223                         if (!symbol_conf.use_callchain)
 224                                 break;
 225                         syms[i].map = al.map;
 226                         syms[i].sym = al.sym;
 227                 }
 228         }
 229
 230         return syms;
 231 }
 232
 233 static int process_event_stub(event_t *event __used,
 234                               struct perf_session *session __used)
 235 {
 236         dump_printf(": unhandled!\n");
 237         return 0;
 238 }
 239
 240 static int process_finished_round_stub(event_t *event __used,
 241                                        struct perf_session *session __used,
 242                                        struct perf_event_ops *ops __used)
 243 {
 244         dump_printf(": unhandled!\n");
 245         return 0;
 246 }
 247
 248 static int process_finished_round(event_t *event,
 249                                   struct perf_session *session,
 250                                   struct perf_event_ops *ops);
 251
 252 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 253 {
 254         if (handler->sample == NULL)
 255                 handler->sample = process_event_stub;
 256         if (handler->mmap == NULL)
 257                 handler->mmap = process_event_stub;
 258         if (handler->comm == NULL)
 259                 handler->comm = process_event_stub;
 260         if (handler->fork == NULL)
 261                 handler->fork = process_event_stub;
 262         if (handler->exit == NULL)
 263                 handler->exit = process_event_stub;
 264         if (handler->lost == NULL)
 265                 handler->lost = process_event_stub;
 266         if (handler->read == NULL)
 267                 handler->read = process_event_stub;
 268         if (handler->throttle == NULL)
 269                 handler->throttle = process_event_stub;
 270         if (handler->unthrottle == NULL)
 271                 handler->unthrottle = process_event_stub;
 272         if (handler->attr == NULL)
 273                 handler->attr = process_event_stub;
 274         if (handler->event_type == NULL)
 275                 handler->event_type = process_event_stub;
 276         if (handler->tracing_data == NULL)
 277                 handler->tracing_data = process_event_stub;
 278         if (handler->build_id == NULL)
 279                 handler->build_id = process_event_stub;
 280         if (handler->finished_round == NULL) {
 281                 if (handler->ordered_samples)
 282                         handler->finished_round = process_finished_round;
 283                 else
 284                         handler->finished_round = process_finished_round_stub;
 285         }
 286 }
 287
 288 void mem_bswap_64(void *src, int byte_size)
 289 {
 290         u64 *m = src;
 291
 292         while (byte_size > 0) {
 293                 *m = bswap_64(*m);
 294                 byte_size -= sizeof(u64);
 295                 ++m;
 296         }
 297 }
 298
 299 static void event__all64_swap(event_t *self)
 300 {
 301         struct perf_event_header *hdr = &self->header;
 302         mem_bswap_64(hdr + 1, self->header.size - sizeof(*hdr));
 303 }
 304
 305 static void event__comm_swap(event_t *self)
 306 {
 307         self->comm.pid = bswap_32(self->comm.pid);
 308         self->comm.tid = bswap_32(self->comm.tid);
 309 }
 310
 311 static void event__mmap_swap(event_t *self)
 312 {
 313         self->mmap.pid   = bswap_32(self->mmap.pid);
 314         self->mmap.tid   = bswap_32(self->mmap.tid);
 315         self->mmap.start = bswap_64(self->mmap.start);
 316         self->mmap.len   = bswap_64(self->mmap.len);
 317         self->mmap.pgoff = bswap_64(self->mmap.pgoff);
 318 }
 319
 320 static void event__task_swap(event_t *self)
 321 {
 322         self->fork.pid  = bswap_32(self->fork.pid);
 323         self->fork.tid  = bswap_32(self->fork.tid);
 324         self->fork.ppid = bswap_32(self->fork.ppid);
 325         self->fork.ptid = bswap_32(self->fork.ptid);
 326         self->fork.time = bswap_64(self->fork.time);
 327 }
 328
 329 static void event__read_swap(event_t *self)
 330 {
 331         self->read.pid          = bswap_32(self->read.pid);
 332         self->read.tid          = bswap_32(self->read.tid);
 333         self->read.value        = bswap_64(self->read.value);
 334         self->read.time_enabled = bswap_64(self->read.time_enabled);
 335         self->read.time_running = bswap_64(self->read.time_running);
 336         self->read.id           = bswap_64(self->read.id);
 337 }
 338
 339 static void event__attr_swap(event_t *self)
 340 {
 341         size_t size;
 342
 343         self->attr.attr.type            = bswap_32(self->attr.attr.type);
 344         self->attr.attr.size            = bswap_32(self->attr.attr.size);
 345         self->attr.attr.config          = bswap_64(self->attr.attr.config);
 346         self->attr.attr.sample_period   = bswap_64(self->attr.attr.sample_period);
 347         self->attr.attr.sample_type     = bswap_64(self->attr.attr.sample_type);
 348         self->attr.attr.read_format     = bswap_64(self->attr.attr.read_format);
 349         self->attr.attr.wakeup_events   = bswap_32(self->attr.attr.wakeup_events);
 350         self->attr.attr.bp_type         = bswap_32(self->attr.attr.bp_type);
 351         self->attr.attr.bp_addr         = bswap_64(self->attr.attr.bp_addr);
 352         self->attr.attr.bp_len          = bswap_64(self->attr.attr.bp_len);
 353
 354         size = self->header.size;
 355         size -= (void *)&self->attr.id - (void *)self;
 356         mem_bswap_64(self->attr.id, size);
 357 }
 358
 359 static void event__event_type_swap(event_t *self)
 360 {
 361         self->event_type.event_type.event_id =
 362                 bswap_64(self->event_type.event_type.event_id);
 363 }
 364
 365 static void event__tracing_data_swap(event_t *self)
 366 {
 367         self->tracing_data.size = bswap_32(self->tracing_data.size);
 368 }
 369
 370 typedef void (*event__swap_op)(event_t *self);
 371
 372 static event__swap_op event__swap_ops[] = {
 373         [PERF_RECORD_MMAP]   = event__mmap_swap,
 374         [PERF_RECORD_COMM]   = event__comm_swap,
 375         [PERF_RECORD_FORK]   = event__task_swap,
 376         [PERF_RECORD_EXIT]   = event__task_swap,
 377         [PERF_RECORD_LOST]   = event__all64_swap,
 378         [PERF_RECORD_READ]   = event__read_swap,
 379         [PERF_RECORD_SAMPLE] = event__all64_swap,
 380         [PERF_RECORD_HEADER_ATTR]   = event__attr_swap,
 381         [PERF_RECORD_HEADER_EVENT_TYPE]   = event__event_type_swap,
 382         [PERF_RECORD_HEADER_TRACING_DATA]   = event__tracing_data_swap,
 383         [PERF_RECORD_HEADER_BUILD_ID]   = NULL,
 384         [PERF_RECORD_HEADER_MAX]    = NULL,
 385 };
 386
 387 struct sample_queue {
 388         u64                     timestamp;
 389         struct sample_event     *event;
 390         struct list_head        list;
 391 };
 392
 393 static void flush_sample_queue(struct perf_session *s,
 394                                struct perf_event_ops *ops)
 395 {
 396         struct list_head *head = &s->ordered_samples.samples_head;
 397         u64 limit = s->ordered_samples.next_flush;
 398         struct sample_queue *tmp, *iter;
 399
 400         if (!ops->ordered_samples || !limit)
 401                 return;
 402
 403         list_for_each_entry_safe(iter, tmp, head, list) {
 404                 if (iter->timestamp > limit)
 405                         return;
 406
 407                 if (iter == s->ordered_samples.last_inserted)
 408                         s->ordered_samples.last_inserted = NULL;
 409
 410                 ops->sample((event_t *)iter->event, s);
 411
 412                 s->ordered_samples.last_flush = iter->timestamp;
 413                 list_del(&iter->list);
 414                 free(iter->event);
 415                 free(iter);
 416         }
 417 }
 418
 419 /*
 420  * When perf record finishes a pass on every buffers, it records this pseudo
 421  * event.
 422  * We record the max timestamp t found in the pass n.
 423  * Assuming these timestamps are monotonic across cpus, we know that if
 424  * a buffer still has events with timestamps below t, they will be all
 425  * available and then read in the pass n + 1.
 426  * Hence when we start to read the pass n + 2, we can safely flush every
 427  * events with timestamps below t.
 428  *
 429  *    ============ PASS n =================
 430  *       CPU 0         |   CPU 1
 431  *                     |
 432  *    cnt1 timestamps  |   cnt2 timestamps
 433  *          1          |         2
 434  *          2          |         3
 435  *          -          |         4  <--- max recorded
 436  *
 437  *    ============ PASS n + 1 ==============
 438  *       CPU 0         |   CPU 1
 439  *                     |
 440  *    cnt1 timestamps  |   cnt2 timestamps
 441  *          3          |         5
 442  *          4          |         6
 443  *          5          |         7 <---- max recorded
 444  *
 445  *      Flush every events below timestamp 4
 446  *
 447  *    ============ PASS n + 2 ==============
 448  *       CPU 0         |   CPU 1
 449  *                     |
 450  *    cnt1 timestamps  |   cnt2 timestamps
 451  *          6          |         8
 452  *          7          |         9
 453  *          -          |         10
 454  *
 455  *      Flush every events below timestamp 7
 456  *      etc...
 457  */
 458 static int process_finished_round(event_t *event __used,
 459                                   struct perf_session *session,
 460                                   struct perf_event_ops *ops)
 461 {
 462         flush_sample_queue(session, ops);
 463         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 464
 465         return 0;
 466 }
 467
 468 static void __queue_sample_end(struct sample_queue *new, struct list_head *head)
 469 {
 470         struct sample_queue *iter;
 471
 472         list_for_each_entry_reverse(iter, head, list) {
 473                 if (iter->timestamp < new->timestamp) {
 474                         list_add(&new->list, &iter->list);
 475                         return;
 476                 }
 477         }
 478
 479         list_add(&new->list, head);
 480 }
 481
 482 static void __queue_sample_before(struct sample_queue *new,
 483                                   struct sample_queue *iter,
 484                                   struct list_head *head)
 485 {
 486         list_for_each_entry_continue_reverse(iter, head, list) {
 487                 if (iter->timestamp < new->timestamp) {
 488                         list_add(&new->list, &iter->list);
 489                         return;
 490                 }
 491         }
 492
 493         list_add(&new->list, head);
 494 }
 495
 496 static void __queue_sample_after(struct sample_queue *new,
 497                                  struct sample_queue *iter,
 498                                  struct list_head *head)
 499 {
 500         list_for_each_entry_continue(iter, head, list) {
 501                 if (iter->timestamp > new->timestamp) {
 502                         list_add_tail(&new->list, &iter->list);
 503                         return;
 504                 }
 505         }
 506         list_add_tail(&new->list, head);
 507 }
 508
 509 /* The queue is ordered by time */
 510 static void __queue_sample_event(struct sample_queue *new,
 511                                  struct perf_session *s)
 512 {
 513         struct sample_queue *last_inserted = s->ordered_samples.last_inserted;
 514         struct list_head *head = &s->ordered_samples.samples_head;
 515
 516
 517         if (!last_inserted) {
 518                 __queue_sample_end(new, head);
 519                 return;
 520         }
 521
 522         /*
 523          * Most of the time the current event has a timestamp
 524          * very close to the last event inserted, unless we just switched
 525          * to another event buffer. Having a sorting based on a list and
 526          * on the last inserted event that is close to the current one is
 527          * probably more efficient than an rbtree based sorting.
 528          */
 529         if (last_inserted->timestamp >= new->timestamp)
 530                 __queue_sample_before(new, last_inserted, head);
 531         else
 532                 __queue_sample_after(new, last_inserted, head);
 533 }
 534
 535 static int queue_sample_event(event_t *event, struct sample_data *data,
 536                               struct perf_session *s)
 537 {
 538         u64 timestamp = data->time;
 539         struct sample_queue *new;
 540
 541
 542         if (timestamp < s->ordered_samples.last_flush) {
 543                 printf("Warning: Timestamp below last timeslice flush\n");
 544                 return -EINVAL;
 545         }
 546
 547         new = malloc(sizeof(*new));
 548         if (!new)
 549                 return -ENOMEM;
 550
 551         new->timestamp = timestamp;
 552
 553         new->event = malloc(event->header.size);
 554         if (!new->event) {
 555                 free(new);
 556                 return -ENOMEM;
 557         }
 558
 559         memcpy(new->event, event, event->header.size);
 560
 561         __queue_sample_event(new, s);
 562         s->ordered_samples.last_inserted = new;
 563
 564         if (new->timestamp > s->ordered_samples.max_timestamp)
 565                 s->ordered_samples.max_timestamp = new->timestamp;
 566
 567         return 0;
 568 }
 569
 570 static int perf_session__process_sample(event_t *event, struct perf_session *s,
 571                                         struct perf_event_ops *ops)
 572 {
 573         struct sample_data data;
 574
 575         if (!ops->ordered_samples)
 576                 return ops->sample(event, s);
 577
 578         bzero(&data, sizeof(struct sample_data));
 579         event__parse_sample(event, s->sample_type, &data);
 580
 581         queue_sample_event(event, &data, s);
 582
 583         return 0;
 584 }
 585
 586 static int perf_session__process_event(struct perf_session *self,
 587                                        event_t *event,
 588                                        struct perf_event_ops *ops,
 589                                        u64 offset, u64 head)
 590 {
 591         trace_event(event);
 592
 593         if (event->header.type < PERF_RECORD_HEADER_MAX) {
 594                 dump_printf("%#Lx [%#x]: PERF_RECORD_%s",
 595                             offset + head, event->header.size,
 596                             event__name[event->header.type]);
 597                 hists__inc_nr_events(&self->hists, event->header.type);
 598         }
 599
 600         if (self->header.needs_swap && event__swap_ops[event->header.type])
 601                 event__swap_ops[event->header.type](event);
 602
 603         switch (event->header.type) {
 604         case PERF_RECORD_SAMPLE:
 605                 return perf_session__process_sample(event, self, ops);
 606         case PERF_RECORD_MMAP:
 607                 return ops->mmap(event, self);
 608         case PERF_RECORD_COMM:
 609                 return ops->comm(event, self);
 610         case PERF_RECORD_FORK:
 611                 return ops->fork(event, self);
 612         case PERF_RECORD_EXIT:
 613                 return ops->exit(event, self);
 614         case PERF_RECORD_LOST:
 615                 return ops->lost(event, self);
 616         case PERF_RECORD_READ:
 617                 return ops->read(event, self);
 618         case PERF_RECORD_THROTTLE:
 619                 return ops->throttle(event, self);
 620         case PERF_RECORD_UNTHROTTLE:
 621                 return ops->unthrottle(event, self);
 622         case PERF_RECORD_HEADER_ATTR:
 623                 return ops->attr(event, self);
 624         case PERF_RECORD_HEADER_EVENT_TYPE:
 625                 return ops->event_type(event, self);
 626         case PERF_RECORD_HEADER_TRACING_DATA:
 627                 /* setup for reading amidst mmap */
 628                 lseek(self->fd, offset + head, SEEK_SET);
 629                 return ops->tracing_data(event, self);
 630         case PERF_RECORD_HEADER_BUILD_ID:
 631                 return ops->build_id(event, self);
 632         case PERF_RECORD_FINISHED_ROUND:
 633                 return ops->finished_round(event, self, ops);
 634         default:
 635                 ++self->hists.stats.nr_unknown_events;
 636                 return -1;
 637         }
 638 }
 639
 640 void perf_event_header__bswap(struct perf_event_header *self)
 641 {
 642         self->type = bswap_32(self->type);
 643         self->misc = bswap_16(self->misc);
 644         self->size = bswap_16(self->size);
 645 }
 646
 647 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 648 {
 649         struct thread *thread = perf_session__findnew(self, 0);
 650
 651         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 652                 pr_err("problem inserting idle task.\n");
 653                 thread = NULL;
 654         }
 655
 656         return thread;
 657 }
 658
 659 int do_read(int fd, void *buf, size_t size)
 660 {
 661         void *buf_start = buf;
 662
 663         while (size) {
 664                 int ret = read(fd, buf, size);
 665
 666                 if (ret <= 0)
 667                         return ret;
 668
 669                 size -= ret;
 670                 buf += ret;
 671         }
 672
 673         return buf - buf_start;
 674 }
 675
 676 #define session_done()  (*(volatile int *)(&session_done))
 677 volatile int session_done;
 678
 679 static int __perf_session__process_pipe_events(struct perf_session *self,
 680                                                struct perf_event_ops *ops)
 681 {
 682         event_t event;
 683         uint32_t size;
 684         int skip = 0;
 685         u64 head;
 686         int err;
 687         void *p;
 688
 689         perf_event_ops__fill_defaults(ops);
 690
 691         head = 0;
 692 more:
 693         err = do_read(self->fd, &event, sizeof(struct perf_event_header));
 694         if (err <= 0) {
 695                 if (err == 0)
 696                         goto done;
 697
 698                 pr_err("failed to read event header\n");
 699                 goto out_err;
 700         }
 701
 702         if (self->header.needs_swap)
 703                 perf_event_header__bswap(&event.header);
 704
 705         size = event.header.size;
 706         if (size == 0)
 707                 size = 8;
 708
 709         p = &event;
 710         p += sizeof(struct perf_event_header);
 711
 712         if (size - sizeof(struct perf_event_header)) {
 713                 err = do_read(self->fd, p,
 714                               size - sizeof(struct perf_event_header));
 715                 if (err <= 0) {
 716                         if (err == 0) {
 717                                 pr_err("unexpected end of event stream\n");
 718                                 goto done;
 719                         }
 720
 721                         pr_err("failed to read event data\n");
 722                         goto out_err;
 723                 }
 724         }
 725
 726         if (size == 0 ||
 727             (skip = perf_session__process_event(self, &event, ops,
 728                                                 0, head)) < 0) {
 729                 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
 730                             head, event.header.size, event.header.type);
 731                 /*
 732                  * assume we lost track of the stream, check alignment, and
 733                  * increment a single u64 in the hope to catch on again 'soon'.
 734                  */
 735                 if (unlikely(head & 7))
 736                         head &= ~7ULL;
 737
 738                 size = 8;
 739         }
 740
 741         head += size;
 742
 743         dump_printf("\n%#Lx [%#x]: event: %d\n",
 744                     head, event.header.size, event.header.type);
 745
 746         if (skip > 0)
 747                 head += skip;
 748
 749         if (!session_done())
 750                 goto more;
 751 done:
 752         err = 0;
 753 out_err:
 754         return err;
 755 }
 756
 757 int __perf_session__process_events(struct perf_session *self,
 758                                    u64 data_offset, u64 data_size,
 759                                    u64 file_size, struct perf_event_ops *ops)
 760 {
 761         int err, mmap_prot, mmap_flags;
 762         u64 head, shift;
 763         u64 offset = 0;
 764         size_t  page_size;
 765         event_t *event;
 766         uint32_t size;
 767         char *buf;
 768         struct ui_progress *progress = ui_progress__new("Processing events...",
 769                                                         self->size);
 770         if (progress == NULL)
 771                 return -1;
 772
 773         perf_event_ops__fill_defaults(ops);
 774
 775         page_size = sysconf(_SC_PAGESIZE);
 776
 777         head = data_offset;
 778         shift = page_size * (head / page_size);
 779         offset += shift;
 780         head -= shift;
 781
 782         mmap_prot  = PROT_READ;
 783         mmap_flags = MAP_SHARED;
 784
 785         if (self->header.needs_swap) {
 786                 mmap_prot  |= PROT_WRITE;
 787                 mmap_flags = MAP_PRIVATE;
 788         }
 789 remap:
 790         buf = mmap(NULL, page_size * self->mmap_window, mmap_prot,
 791                    mmap_flags, self->fd, offset);
 792         if (buf == MAP_FAILED) {
 793                 pr_err("failed to mmap file\n");
 794                 err = -errno;
 795                 goto out_err;
 796         }
 797
 798 more:
 799         event = (event_t *)(buf + head);
 800         ui_progress__update(progress, offset);
 801
 802         if (self->header.needs_swap)
 803                 perf_event_header__bswap(&event->header);
 804         size = event->header.size;
 805         if (size == 0)
 806                 size = 8;
 807
 808         if (head + event->header.size >= page_size * self->mmap_window) {
 809                 int munmap_ret;
 810
 811                 shift = page_size * (head / page_size);
 812
 813                 munmap_ret = munmap(buf, page_size * self->mmap_window);
 814                 assert(munmap_ret == 0);
 815
 816                 offset += shift;
 817                 head -= shift;
 818                 goto remap;
 819         }
 820
 821         size = event->header.size;
 822
 823         dump_printf("\n%#Lx [%#x]: event: %d\n",
 824                     offset + head, event->header.size, event->header.type);
 825
 826         if (size == 0 ||
 827             perf_session__process_event(self, event, ops, offset, head) < 0) {
 828                 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
 829                             offset + head, event->header.size,
 830                             event->header.type);
 831                 /*
 832                  * assume we lost track of the stream, check alignment, and
 833                  * increment a single u64 in the hope to catch on again 'soon'.
 834                  */
 835                 if (unlikely(head & 7))
 836                         head &= ~7ULL;
 837
 838                 size = 8;
 839         }
 840
 841         head += size;
 842
 843         if (offset + head >= data_offset + data_size)
 844                 goto done;
 845
 846         if (offset + head < file_size)
 847                 goto more;
 848 done:
 849         err = 0;
 850         /* do the final flush for ordered samples */
 851         self->ordered_samples.next_flush = ULLONG_MAX;
 852         flush_sample_queue(self, ops);
 853 out_err:
 854         ui_progress__delete(progress);
 855         return err;
 856 }
 857
 858 int perf_session__process_events(struct perf_session *self,
 859                                  struct perf_event_ops *ops)
 860 {
 861         int err;
 862
 863         if (perf_session__register_idle_thread(self) == NULL)
 864                 return -ENOMEM;
 865
 866         if (!self->fd_pipe)
 867                 err = __perf_session__process_events(self,
 868                                                      self->header.data_offset,
 869                                                      self->header.data_size,
 870                                                      self->size, ops);
 871         else
 872                 err = __perf_session__process_pipe_events(self, ops);
 873
 874         return err;
 875 }
 876
 877 bool perf_session__has_traces(struct perf_session *self, const char *msg)
 878 {
 879         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
 880                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
 881                 return false;
 882         }
 883
 884         return true;
 885 }
 886
 887 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
 888                                              const char *symbol_name,
 889                                              u64 addr)
 890 {
 891         char *bracket;
 892         enum map_type i;
 893         struct ref_reloc_sym *ref;
 894
 895         ref = zalloc(sizeof(struct ref_reloc_sym));
 896         if (ref == NULL)
 897                 return -ENOMEM;
 898
 899         ref->name = strdup(symbol_name);
 900         if (ref->name == NULL) {
 901                 free(ref);
 902                 return -ENOMEM;
 903         }
 904
 905         bracket = strchr(ref->name, ']');
 906         if (bracket)
 907                 *bracket = '\0';
 908
 909         ref->addr = addr;
 910
 911         for (i = 0; i < MAP__NR_TYPES; ++i) {
 912                 struct kmap *kmap = map__kmap(maps[i]);
 913                 kmap->ref_reloc_sym = ref;
 914         }
 915
 916         return 0;
 917 }
 918
 919 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
 920 {
 921         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
 922                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
 923                machines__fprintf_dsos(&self->machines, fp);
 924 }
 925
 926 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
 927                                           bool with_hits)
 928 {
 929         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
 930         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
 931 }