tools/perf/util/session.c

   1 #define _FILE_OFFSET_BITS 64
   2
   3 #include <linux/kernel.h>
   4
   5 #include <byteswap.h>
   6 #include <unistd.h>
   7 #include <sys/types.h>
   8 #include <sys/mman.h>
   9
  10 #include "evlist.h"
  11 #include "evsel.h"
  12 #include "session.h"
  13 #include "sort.h"
  14 #include "util.h"
  15
  16 static int perf_session__open(struct perf_session *self, bool force)
  17 {
  18         struct stat input_stat;
  19
  20         if (!strcmp(self->filename, "-")) {
  21                 self->fd_pipe = true;
  22                 self->fd = STDIN_FILENO;
  23
  24                 if (perf_session__read_header(self, self->fd) < 0)
  25                         pr_err("incompatible file format");
  26
  27                 return 0;
  28         }
  29
  30         self->fd = open(self->filename, O_RDONLY);
  31         if (self->fd < 0) {
  32                 int err = errno;
  33
  34                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  35                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  36                         pr_err("  (try 'perf record' first)");
  37                 pr_err("\n");
  38                 return -errno;
  39         }
  40
  41         if (fstat(self->fd, &input_stat) < 0)
  42                 goto out_close;
  43
  44         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  45                 pr_err("file %s not owned by current user or root\n",
  46                        self->filename);
  47                 goto out_close;
  48         }
  49
  50         if (!input_stat.st_size) {
  51                 pr_info("zero-sized file (%s), nothing to do!\n",
  52                         self->filename);
  53                 goto out_close;
  54         }
  55
  56         if (perf_session__read_header(self, self->fd) < 0) {
  57                 pr_err("incompatible file format");
  58                 goto out_close;
  59         }
  60
  61         self->size = input_stat.st_size;
  62         return 0;
  63
  64 out_close:
  65         close(self->fd);
  66         self->fd = -1;
  67         return -1;
  68 }
  69
  70 static void perf_session__id_header_size(struct perf_session *session)
  71 {
  72        struct perf_sample *data;
  73        u64 sample_type = session->sample_type;
  74        u16 size = 0;
  75
  76         if (!session->sample_id_all)
  77                 goto out;
  78
  79        if (sample_type & PERF_SAMPLE_TID)
  80                size += sizeof(data->tid) * 2;
  81
  82        if (sample_type & PERF_SAMPLE_TIME)
  83                size += sizeof(data->time);
  84
  85        if (sample_type & PERF_SAMPLE_ID)
  86                size += sizeof(data->id);
  87
  88        if (sample_type & PERF_SAMPLE_STREAM_ID)
  89                size += sizeof(data->stream_id);
  90
  91        if (sample_type & PERF_SAMPLE_CPU)
  92                size += sizeof(data->cpu) * 2;
  93 out:
  94        session->id_hdr_size = size;
  95 }
  96
  97 void perf_session__update_sample_type(struct perf_session *self)
  98 {
  99         self->sample_type = perf_evlist__sample_type(self->evlist);
 100         self->sample_size = perf_sample_size(self->sample_type);
 101         self->sample_id_all = perf_evlist__sample_id_all(self->evlist);
 102         perf_session__id_header_size(self);
 103 }
 104
 105 int perf_session__create_kernel_maps(struct perf_session *self)
 106 {
 107         int ret = machine__create_kernel_maps(&self->host_machine);
 108
 109         if (ret >= 0)
 110                 ret = machines__create_guest_kernel_maps(&self->machines);
 111         return ret;
 112 }
 113
 114 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 115 {
 116         machine__destroy_kernel_maps(&self->host_machine);
 117         machines__destroy_guest_kernel_maps(&self->machines);
 118 }
 119
 120 struct perf_session *perf_session__new(const char *filename, int mode,
 121                                        bool force, bool repipe,
 122                                        struct perf_event_ops *ops)
 123 {
 124         size_t len = filename ? strlen(filename) + 1 : 0;
 125         struct perf_session *self = zalloc(sizeof(*self) + len);
 126
 127         if (self == NULL)
 128                 goto out;
 129
 130         memcpy(self->filename, filename, len);
 131         self->threads = RB_ROOT;
 132         INIT_LIST_HEAD(&self->dead_threads);
 133         self->last_match = NULL;
 134         /*
 135          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 136          * slices. On 32bit we use 32MB.
 137          */
 138 #if BITS_PER_LONG == 64
 139         self->mmap_window = ULLONG_MAX;
 140 #else
 141         self->mmap_window = 32 * 1024 * 1024ULL;
 142 #endif
 143         self->machines = RB_ROOT;
 144         self->repipe = repipe;
 145         INIT_LIST_HEAD(&self->ordered_samples.samples);
 146         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 147         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 148         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 149
 150         if (mode == O_RDONLY) {
 151                 if (perf_session__open(self, force) < 0)
 152                         goto out_delete;
 153                 perf_session__update_sample_type(self);
 154         } else if (mode == O_WRONLY) {
 155                 /*
 156                  * In O_RDONLY mode this will be performed when reading the
 157                  * kernel MMAP event, in perf_event__process_mmap().
 158                  */
 159                 if (perf_session__create_kernel_maps(self) < 0)
 160                         goto out_delete;
 161         }
 162
 163         if (ops && ops->ordering_requires_timestamps &&
 164             ops->ordered_samples && !self->sample_id_all) {
 165                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 166                 ops->ordered_samples = false;
 167         }
 168
 169 out:
 170         return self;
 171 out_delete:
 172         perf_session__delete(self);
 173         return NULL;
 174 }
 175
 176 static void perf_session__delete_dead_threads(struct perf_session *self)
 177 {
 178         struct thread *n, *t;
 179
 180         list_for_each_entry_safe(t, n, &self->dead_threads, node) {
 181                 list_del(&t->node);
 182                 thread__delete(t);
 183         }
 184 }
 185
 186 static void perf_session__delete_threads(struct perf_session *self)
 187 {
 188         struct rb_node *nd = rb_first(&self->threads);
 189
 190         while (nd) {
 191                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 192
 193                 rb_erase(&t->rb_node, &self->threads);
 194                 nd = rb_next(nd);
 195                 thread__delete(t);
 196         }
 197 }
 198
 199 void perf_session__delete(struct perf_session *self)
 200 {
 201         perf_session__destroy_kernel_maps(self);
 202         perf_session__delete_dead_threads(self);
 203         perf_session__delete_threads(self);
 204         machine__exit(&self->host_machine);
 205         close(self->fd);
 206         free(self);
 207 }
 208
 209 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
 210 {
 211         self->last_match = NULL;
 212         rb_erase(&th->rb_node, &self->threads);
 213         /*
 214          * We may have references to this thread, for instance in some hist_entry
 215          * instances, so just move them to a separate list.
 216          */
 217         list_add_tail(&th->node, &self->dead_threads);
 218 }
 219
 220 static bool symbol__match_parent_regex(struct symbol *sym)
 221 {
 222         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 223                 return 1;
 224
 225         return 0;
 226 }
 227
 228 int perf_session__resolve_callchain(struct perf_session *self,
 229                                     struct thread *thread,
 230                                     struct ip_callchain *chain,
 231                                     struct symbol **parent)
 232 {
 233         u8 cpumode = PERF_RECORD_MISC_USER;
 234         unsigned int i;
 235         int err;
 236
 237         callchain_cursor_reset(&self->callchain_cursor);
 238
 239         for (i = 0; i < chain->nr; i++) {
 240                 u64 ip = chain->ips[i];
 241                 struct addr_location al;
 242
 243                 if (ip >= PERF_CONTEXT_MAX) {
 244                         switch (ip) {
 245                         case PERF_CONTEXT_HV:
 246                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 247                         case PERF_CONTEXT_KERNEL:
 248                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 249                         case PERF_CONTEXT_USER:
 250                                 cpumode = PERF_RECORD_MISC_USER;        break;
 251                         default:
 252                                 break;
 253                         }
 254                         continue;
 255                 }
 256
 257                 al.filtered = false;
 258                 thread__find_addr_location(thread, self, cpumode,
 259                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 260                 if (al.sym != NULL) {
 261                         if (sort__has_parent && !*parent &&
 262                             symbol__match_parent_regex(al.sym))
 263                                 *parent = al.sym;
 264                         if (!symbol_conf.use_callchain)
 265                                 break;
 266                 }
 267
 268                 err = callchain_cursor_append(&self->callchain_cursor,
 269                                               ip, al.map, al.sym);
 270                 if (err)
 271                         return err;
 272         }
 273
 274         return 0;
 275 }
 276
 277 static int process_event_synth_stub(union perf_event *event __used,
 278                                     struct perf_session *session __used)
 279 {
 280         dump_printf(": unhandled!\n");
 281         return 0;
 282 }
 283
 284 static int process_event_sample_stub(union perf_event *event __used,
 285                                      struct perf_sample *sample __used,
 286                                      struct perf_evsel *evsel __used,
 287                                      struct perf_session *session __used)
 288 {
 289         dump_printf(": unhandled!\n");
 290         return 0;
 291 }
 292
 293 static int process_event_stub(union perf_event *event __used,
 294                               struct perf_sample *sample __used,
 295                               struct perf_session *session __used)
 296 {
 297         dump_printf(": unhandled!\n");
 298         return 0;
 299 }
 300
 301 static int process_finished_round_stub(union perf_event *event __used,
 302                                        struct perf_session *session __used,
 303                                        struct perf_event_ops *ops __used)
 304 {
 305         dump_printf(": unhandled!\n");
 306         return 0;
 307 }
 308
 309 static int process_finished_round(union perf_event *event,
 310                                   struct perf_session *session,
 311                                   struct perf_event_ops *ops);
 312
 313 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 314 {
 315         if (handler->sample == NULL)
 316                 handler->sample = process_event_sample_stub;
 317         if (handler->mmap == NULL)
 318                 handler->mmap = process_event_stub;
 319         if (handler->comm == NULL)
 320                 handler->comm = process_event_stub;
 321         if (handler->fork == NULL)
 322                 handler->fork = process_event_stub;
 323         if (handler->exit == NULL)
 324                 handler->exit = process_event_stub;
 325         if (handler->lost == NULL)
 326                 handler->lost = perf_event__process_lost;
 327         if (handler->read == NULL)
 328                 handler->read = process_event_stub;
 329         if (handler->throttle == NULL)
 330                 handler->throttle = process_event_stub;
 331         if (handler->unthrottle == NULL)
 332                 handler->unthrottle = process_event_stub;
 333         if (handler->attr == NULL)
 334                 handler->attr = process_event_synth_stub;
 335         if (handler->event_type == NULL)
 336                 handler->event_type = process_event_synth_stub;
 337         if (handler->tracing_data == NULL)
 338                 handler->tracing_data = process_event_synth_stub;
 339         if (handler->build_id == NULL)
 340                 handler->build_id = process_event_synth_stub;
 341         if (handler->finished_round == NULL) {
 342                 if (handler->ordered_samples)
 343                         handler->finished_round = process_finished_round;
 344                 else
 345                         handler->finished_round = process_finished_round_stub;
 346         }
 347 }
 348
 349 void mem_bswap_64(void *src, int byte_size)
 350 {
 351         u64 *m = src;
 352
 353         while (byte_size > 0) {
 354                 *m = bswap_64(*m);
 355                 byte_size -= sizeof(u64);
 356                 ++m;
 357         }
 358 }
 359
 360 static void perf_event__all64_swap(union perf_event *event)
 361 {
 362         struct perf_event_header *hdr = &event->header;
 363         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 364 }
 365
 366 static void perf_event__comm_swap(union perf_event *event)
 367 {
 368         event->comm.pid = bswap_32(event->comm.pid);
 369         event->comm.tid = bswap_32(event->comm.tid);
 370 }
 371
 372 static void perf_event__mmap_swap(union perf_event *event)
 373 {
 374         event->mmap.pid   = bswap_32(event->mmap.pid);
 375         event->mmap.tid   = bswap_32(event->mmap.tid);
 376         event->mmap.start = bswap_64(event->mmap.start);
 377         event->mmap.len   = bswap_64(event->mmap.len);
 378         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 379 }
 380
 381 static void perf_event__task_swap(union perf_event *event)
 382 {
 383         event->fork.pid  = bswap_32(event->fork.pid);
 384         event->fork.tid  = bswap_32(event->fork.tid);
 385         event->fork.ppid = bswap_32(event->fork.ppid);
 386         event->fork.ptid = bswap_32(event->fork.ptid);
 387         event->fork.time = bswap_64(event->fork.time);
 388 }
 389
 390 static void perf_event__read_swap(union perf_event *event)
 391 {
 392         event->read.pid          = bswap_32(event->read.pid);
 393         event->read.tid          = bswap_32(event->read.tid);
 394         event->read.value        = bswap_64(event->read.value);
 395         event->read.time_enabled = bswap_64(event->read.time_enabled);
 396         event->read.time_running = bswap_64(event->read.time_running);
 397         event->read.id           = bswap_64(event->read.id);
 398 }
 399
 400 static void perf_event__attr_swap(union perf_event *event)
 401 {
 402         size_t size;
 403
 404         event->attr.attr.type           = bswap_32(event->attr.attr.type);
 405         event->attr.attr.size           = bswap_32(event->attr.attr.size);
 406         event->attr.attr.config         = bswap_64(event->attr.attr.config);
 407         event->attr.attr.sample_period  = bswap_64(event->attr.attr.sample_period);
 408         event->attr.attr.sample_type    = bswap_64(event->attr.attr.sample_type);
 409         event->attr.attr.read_format    = bswap_64(event->attr.attr.read_format);
 410         event->attr.attr.wakeup_events  = bswap_32(event->attr.attr.wakeup_events);
 411         event->attr.attr.bp_type        = bswap_32(event->attr.attr.bp_type);
 412         event->attr.attr.bp_addr        = bswap_64(event->attr.attr.bp_addr);
 413         event->attr.attr.bp_len         = bswap_64(event->attr.attr.bp_len);
 414
 415         size = event->header.size;
 416         size -= (void *)&event->attr.id - (void *)event;
 417         mem_bswap_64(event->attr.id, size);
 418 }
 419
 420 static void perf_event__event_type_swap(union perf_event *event)
 421 {
 422         event->event_type.event_type.event_id =
 423                 bswap_64(event->event_type.event_type.event_id);
 424 }
 425
 426 static void perf_event__tracing_data_swap(union perf_event *event)
 427 {
 428         event->tracing_data.size = bswap_32(event->tracing_data.size);
 429 }
 430
 431 typedef void (*perf_event__swap_op)(union perf_event *event);
 432
 433 static perf_event__swap_op perf_event__swap_ops[] = {
 434         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 435         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 436         [PERF_RECORD_FORK]                = perf_event__task_swap,
 437         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 438         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 439         [PERF_RECORD_READ]                = perf_event__read_swap,
 440         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 441         [PERF_RECORD_HEADER_ATTR]         = perf_event__attr_swap,
 442         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 443         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 444         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 445         [PERF_RECORD_HEADER_MAX]          = NULL,
 446 };
 447
 448 struct sample_queue {
 449         u64                     timestamp;
 450         u64                     file_offset;
 451         union perf_event        *event;
 452         struct list_head        list;
 453 };
 454
 455 static void perf_session_free_sample_buffers(struct perf_session *session)
 456 {
 457         struct ordered_samples *os = &session->ordered_samples;
 458
 459         while (!list_empty(&os->to_free)) {
 460                 struct sample_queue *sq;
 461
 462                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 463                 list_del(&sq->list);
 464                 free(sq);
 465         }
 466 }
 467
 468 static int perf_session_deliver_event(struct perf_session *session,
 469                                       union perf_event *event,
 470                                       struct perf_sample *sample,
 471                                       struct perf_event_ops *ops,
 472                                       u64 file_offset);
 473
 474 static void flush_sample_queue(struct perf_session *s,
 475                                struct perf_event_ops *ops)
 476 {
 477         struct ordered_samples *os = &s->ordered_samples;
 478         struct list_head *head = &os->samples;
 479         struct sample_queue *tmp, *iter;
 480         struct perf_sample sample;
 481         u64 limit = os->next_flush;
 482         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 483         int ret;
 484
 485         if (!ops->ordered_samples || !limit)
 486                 return;
 487
 488         list_for_each_entry_safe(iter, tmp, head, list) {
 489                 if (iter->timestamp > limit)
 490                         break;
 491
 492                 ret = perf_session__parse_sample(s, iter->event, &sample);
 493                 if (ret)
 494                         pr_err("Can't parse sample, err = %d\n", ret);
 495                 else
 496                         perf_session_deliver_event(s, iter->event, &sample, ops,
 497                                                    iter->file_offset);
 498
 499                 os->last_flush = iter->timestamp;
 500                 list_del(&iter->list);
 501                 list_add(&iter->list, &os->sample_cache);
 502         }
 503
 504         if (list_empty(head)) {
 505                 os->last_sample = NULL;
 506         } else if (last_ts <= limit) {
 507                 os->last_sample =
 508                         list_entry(head->prev, struct sample_queue, list);
 509         }
 510 }
 511
 512 /*
 513  * When perf record finishes a pass on every buffers, it records this pseudo
 514  * event.
 515  * We record the max timestamp t found in the pass n.
 516  * Assuming these timestamps are monotonic across cpus, we know that if
 517  * a buffer still has events with timestamps below t, they will be all
 518  * available and then read in the pass n + 1.
 519  * Hence when we start to read the pass n + 2, we can safely flush every
 520  * events with timestamps below t.
 521  *
 522  *    ============ PASS n =================
 523  *       CPU 0         |   CPU 1
 524  *                     |
 525  *    cnt1 timestamps  |   cnt2 timestamps
 526  *          1          |         2
 527  *          2          |         3
 528  *          -          |         4  <--- max recorded
 529  *
 530  *    ============ PASS n + 1 ==============
 531  *       CPU 0         |   CPU 1
 532  *                     |
 533  *    cnt1 timestamps  |   cnt2 timestamps
 534  *          3          |         5
 535  *          4          |         6
 536  *          5          |         7 <---- max recorded
 537  *
 538  *      Flush every events below timestamp 4
 539  *
 540  *    ============ PASS n + 2 ==============
 541  *       CPU 0         |   CPU 1
 542  *                     |
 543  *    cnt1 timestamps  |   cnt2 timestamps
 544  *          6          |         8
 545  *          7          |         9
 546  *          -          |         10
 547  *
 548  *      Flush every events below timestamp 7
 549  *      etc...
 550  */
 551 static int process_finished_round(union perf_event *event __used,
 552                                   struct perf_session *session,
 553                                   struct perf_event_ops *ops)
 554 {
 555         flush_sample_queue(session, ops);
 556         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 557
 558         return 0;
 559 }
 560
 561 /* The queue is ordered by time */
 562 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 563 {
 564         struct ordered_samples *os = &s->ordered_samples;
 565         struct sample_queue *sample = os->last_sample;
 566         u64 timestamp = new->timestamp;
 567         struct list_head *p;
 568
 569         os->last_sample = new;
 570
 571         if (!sample) {
 572                 list_add(&new->list, &os->samples);
 573                 os->max_timestamp = timestamp;
 574                 return;
 575         }
 576
 577         /*
 578          * last_sample might point to some random place in the list as it's
 579          * the last queued event. We expect that the new event is close to
 580          * this.
 581          */
 582         if (sample->timestamp <= timestamp) {
 583                 while (sample->timestamp <= timestamp) {
 584                         p = sample->list.next;
 585                         if (p == &os->samples) {
 586                                 list_add_tail(&new->list, &os->samples);
 587                                 os->max_timestamp = timestamp;
 588                                 return;
 589                         }
 590                         sample = list_entry(p, struct sample_queue, list);
 591                 }
 592                 list_add_tail(&new->list, &sample->list);
 593         } else {
 594                 while (sample->timestamp > timestamp) {
 595                         p = sample->list.prev;
 596                         if (p == &os->samples) {
 597                                 list_add(&new->list, &os->samples);
 598                                 return;
 599                         }
 600                         sample = list_entry(p, struct sample_queue, list);
 601                 }
 602                 list_add(&new->list, &sample->list);
 603         }
 604 }
 605
 606 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 607
 608 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 609                                     struct perf_sample *sample, u64 file_offset)
 610 {
 611         struct ordered_samples *os = &s->ordered_samples;
 612         struct list_head *sc = &os->sample_cache;
 613         u64 timestamp = sample->time;
 614         struct sample_queue *new;
 615
 616         if (!timestamp || timestamp == ~0ULL)
 617                 return -ETIME;
 618
 619         if (timestamp < s->ordered_samples.last_flush) {
 620                 printf("Warning: Timestamp below last timeslice flush\n");
 621                 return -EINVAL;
 622         }
 623
 624         if (!list_empty(sc)) {
 625                 new = list_entry(sc->next, struct sample_queue, list);
 626                 list_del(&new->list);
 627         } else if (os->sample_buffer) {
 628                 new = os->sample_buffer + os->sample_buffer_idx;
 629                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 630                         os->sample_buffer = NULL;
 631         } else {
 632                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 633                 if (!os->sample_buffer)
 634                         return -ENOMEM;
 635                 list_add(&os->sample_buffer->list, &os->to_free);
 636                 os->sample_buffer_idx = 2;
 637                 new = os->sample_buffer + 1;
 638         }
 639
 640         new->timestamp = timestamp;
 641         new->file_offset = file_offset;
 642         new->event = event;
 643
 644         __queue_event(new, s);
 645
 646         return 0;
 647 }
 648
 649 static void callchain__printf(struct perf_sample *sample)
 650 {
 651         unsigned int i;
 652
 653         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 654
 655         for (i = 0; i < sample->callchain->nr; i++)
 656                 printf("..... %2d: %016" PRIx64 "\n",
 657                        i, sample->callchain->ips[i]);
 658 }
 659
 660 static void perf_session__print_tstamp(struct perf_session *session,
 661                                        union perf_event *event,
 662                                        struct perf_sample *sample)
 663 {
 664         if (event->header.type != PERF_RECORD_SAMPLE &&
 665             !session->sample_id_all) {
 666                 fputs("-1 -1 ", stdout);
 667                 return;
 668         }
 669
 670         if ((session->sample_type & PERF_SAMPLE_CPU))
 671                 printf("%u ", sample->cpu);
 672
 673         if (session->sample_type & PERF_SAMPLE_TIME)
 674                 printf("%" PRIu64 " ", sample->time);
 675 }
 676
 677 static void dump_event(struct perf_session *session, union perf_event *event,
 678                        u64 file_offset, struct perf_sample *sample)
 679 {
 680         if (!dump_trace)
 681                 return;
 682
 683         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 684                file_offset, event->header.size, event->header.type);
 685
 686         trace_event(event);
 687
 688         if (sample)
 689                 perf_session__print_tstamp(session, event, sample);
 690
 691         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 692                event->header.size, perf_event__name(event->header.type));
 693 }
 694
 695 static void dump_sample(struct perf_session *session, union perf_event *event,
 696                         struct perf_sample *sample)
 697 {
 698         if (!dump_trace)
 699                 return;
 700
 701         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 "\n",
 702                event->header.misc, sample->pid, sample->tid, sample->ip,
 703                sample->period);
 704
 705         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 706                 callchain__printf(sample);
 707 }
 708
 709 static int perf_session_deliver_event(struct perf_session *session,
 710                                       union perf_event *event,
 711                                       struct perf_sample *sample,
 712                                       struct perf_event_ops *ops,
 713                                       u64 file_offset)
 714 {
 715         struct perf_evsel *evsel;
 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                 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 723                 if (evsel == NULL) {
 724                         ++session->hists.stats.nr_unknown_id;
 725                         return -1;
 726                 }
 727                 return ops->sample(event, sample, evsel, session);
 728         case PERF_RECORD_MMAP:
 729                 return ops->mmap(event, sample, session);
 730         case PERF_RECORD_COMM:
 731                 return ops->comm(event, sample, session);
 732         case PERF_RECORD_FORK:
 733                 return ops->fork(event, sample, session);
 734         case PERF_RECORD_EXIT:
 735                 return ops->exit(event, sample, session);
 736         case PERF_RECORD_LOST:
 737                 return ops->lost(event, sample, session);
 738         case PERF_RECORD_READ:
 739                 return ops->read(event, sample, session);
 740         case PERF_RECORD_THROTTLE:
 741                 return ops->throttle(event, sample, session);
 742         case PERF_RECORD_UNTHROTTLE:
 743                 return ops->unthrottle(event, sample, session);
 744         default:
 745                 ++session->hists.stats.nr_unknown_events;
 746                 return -1;
 747         }
 748 }
 749
 750 static int perf_session__preprocess_sample(struct perf_session *session,
 751                                            union perf_event *event, struct perf_sample *sample)
 752 {
 753         if (event->header.type != PERF_RECORD_SAMPLE ||
 754             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 755                 return 0;
 756
 757         if (!ip_callchain__valid(sample->callchain, event)) {
 758                 pr_debug("call-chain problem with event, skipping it.\n");
 759                 ++session->hists.stats.nr_invalid_chains;
 760                 session->hists.stats.total_invalid_chains += sample->period;
 761                 return -EINVAL;
 762         }
 763         return 0;
 764 }
 765
 766 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
 767                                             struct perf_event_ops *ops, u64 file_offset)
 768 {
 769         dump_event(session, event, file_offset, NULL);
 770
 771         /* These events are processed right away */
 772         switch (event->header.type) {
 773         case PERF_RECORD_HEADER_ATTR:
 774                 return ops->attr(event, session);
 775         case PERF_RECORD_HEADER_EVENT_TYPE:
 776                 return ops->event_type(event, session);
 777         case PERF_RECORD_HEADER_TRACING_DATA:
 778                 /* setup for reading amidst mmap */
 779                 lseek(session->fd, file_offset, SEEK_SET);
 780                 return ops->tracing_data(event, session);
 781         case PERF_RECORD_HEADER_BUILD_ID:
 782                 return ops->build_id(event, session);
 783         case PERF_RECORD_FINISHED_ROUND:
 784                 return ops->finished_round(event, session, ops);
 785         default:
 786                 return -EINVAL;
 787         }
 788 }
 789
 790 static int perf_session__process_event(struct perf_session *session,
 791                                        union perf_event *event,
 792                                        struct perf_event_ops *ops,
 793                                        u64 file_offset)
 794 {
 795         struct perf_sample sample;
 796         int ret;
 797
 798         if (session->header.needs_swap &&
 799             perf_event__swap_ops[event->header.type])
 800                 perf_event__swap_ops[event->header.type](event);
 801
 802         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 803                 return -EINVAL;
 804
 805         hists__inc_nr_events(&session->hists, event->header.type);
 806
 807         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 808                 return perf_session__process_user_event(session, event, ops, file_offset);
 809
 810         /*
 811          * For all kernel events we get the sample data
 812          */
 813         ret = perf_session__parse_sample(session, event, &sample);
 814         if (ret)
 815                 return ret;
 816
 817         /* Preprocess sample records - precheck callchains */
 818         if (perf_session__preprocess_sample(session, event, &sample))
 819                 return 0;
 820
 821         if (ops->ordered_samples) {
 822                 ret = perf_session_queue_event(session, event, &sample,
 823                                                file_offset);
 824                 if (ret != -ETIME)
 825                         return ret;
 826         }
 827
 828         return perf_session_deliver_event(session, event, &sample, ops,
 829                                           file_offset);
 830 }
 831
 832 void perf_event_header__bswap(struct perf_event_header *self)
 833 {
 834         self->type = bswap_32(self->type);
 835         self->misc = bswap_16(self->misc);
 836         self->size = bswap_16(self->size);
 837 }
 838
 839 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 840 {
 841         struct thread *thread = perf_session__findnew(self, 0);
 842
 843         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 844                 pr_err("problem inserting idle task.\n");
 845                 thread = NULL;
 846         }
 847
 848         return thread;
 849 }
 850
 851 static void perf_session__warn_about_errors(const struct perf_session *session,
 852                                             const struct perf_event_ops *ops)
 853 {
 854         if (ops->lost == perf_event__process_lost &&
 855             session->hists.stats.total_lost != 0) {
 856                 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64
 857                             "!\n\nCheck IO/CPU overload!\n\n",
 858                             session->hists.stats.total_period,
 859                             session->hists.stats.total_lost);
 860         }
 861
 862         if (session->hists.stats.nr_unknown_events != 0) {
 863                 ui__warning("Found %u unknown events!\n\n"
 864                             "Is this an older tool processing a perf.data "
 865                             "file generated by a more recent tool?\n\n"
 866                             "If that is not the case, consider "
 867                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 868                             session->hists.stats.nr_unknown_events);
 869         }
 870
 871         if (session->hists.stats.nr_unknown_id != 0) {
 872                 ui__warning("%u samples with id not present in the header\n",
 873                             session->hists.stats.nr_unknown_id);
 874         }
 875
 876         if (session->hists.stats.nr_invalid_chains != 0) {
 877                 ui__warning("Found invalid callchains!\n\n"
 878                             "%u out of %u events were discarded for this reason.\n\n"
 879                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 880                             session->hists.stats.nr_invalid_chains,
 881                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 882         }
 883 }
 884
 885 #define session_done()  (*(volatile int *)(&session_done))
 886 volatile int session_done;
 887
 888 static int __perf_session__process_pipe_events(struct perf_session *self,
 889                                                struct perf_event_ops *ops)
 890 {
 891         union perf_event event;
 892         uint32_t size;
 893         int skip = 0;
 894         u64 head;
 895         int err;
 896         void *p;
 897
 898         perf_event_ops__fill_defaults(ops);
 899
 900         head = 0;
 901 more:
 902         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 903         if (err <= 0) {
 904                 if (err == 0)
 905                         goto done;
 906
 907                 pr_err("failed to read event header\n");
 908                 goto out_err;
 909         }
 910
 911         if (self->header.needs_swap)
 912                 perf_event_header__bswap(&event.header);
 913
 914         size = event.header.size;
 915         if (size == 0)
 916                 size = 8;
 917
 918         p = &event;
 919         p += sizeof(struct perf_event_header);
 920
 921         if (size - sizeof(struct perf_event_header)) {
 922                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 923                 if (err <= 0) {
 924                         if (err == 0) {
 925                                 pr_err("unexpected end of event stream\n");
 926                                 goto done;
 927                         }
 928
 929                         pr_err("failed to read event data\n");
 930                         goto out_err;
 931                 }
 932         }
 933
 934         if (size == 0 ||
 935             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 936                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 937                             head, event.header.size, event.header.type);
 938                 /*
 939                  * assume we lost track of the stream, check alignment, and
 940                  * increment a single u64 in the hope to catch on again 'soon'.
 941                  */
 942                 if (unlikely(head & 7))
 943                         head &= ~7ULL;
 944
 945                 size = 8;
 946         }
 947
 948         head += size;
 949
 950         if (skip > 0)
 951                 head += skip;
 952
 953         if (!session_done())
 954                 goto more;
 955 done:
 956         err = 0;
 957 out_err:
 958         perf_session__warn_about_errors(self, ops);
 959         perf_session_free_sample_buffers(self);
 960         return err;
 961 }
 962
 963 static union perf_event *
 964 fetch_mmaped_event(struct perf_session *session,
 965                    u64 head, size_t mmap_size, char *buf)
 966 {
 967         union perf_event *event;
 968
 969         /*
 970          * Ensure we have enough space remaining to read
 971          * the size of the event in the headers.
 972          */
 973         if (head + sizeof(event->header) > mmap_size)
 974                 return NULL;
 975
 976         event = (union perf_event *)(buf + head);
 977
 978         if (session->header.needs_swap)
 979                 perf_event_header__bswap(&event->header);
 980
 981         if (head + event->header.size > mmap_size)
 982                 return NULL;
 983
 984         return event;
 985 }
 986
 987 int __perf_session__process_events(struct perf_session *session,
 988                                    u64 data_offset, u64 data_size,
 989                                    u64 file_size, struct perf_event_ops *ops)
 990 {
 991         u64 head, page_offset, file_offset, file_pos, progress_next;
 992         int err, mmap_prot, mmap_flags, map_idx = 0;
 993         struct ui_progress *progress;
 994         size_t  page_size, mmap_size;
 995         char *buf, *mmaps[8];
 996         union perf_event *event;
 997         uint32_t size;
 998
 999         perf_event_ops__fill_defaults(ops);
1000
1001         page_size = sysconf(_SC_PAGESIZE);
1002
1003         page_offset = page_size * (data_offset / page_size);
1004         file_offset = page_offset;
1005         head = data_offset - page_offset;
1006
1007         if (data_offset + data_size < file_size)
1008                 file_size = data_offset + data_size;
1009
1010         progress_next = file_size / 16;
1011         progress = ui_progress__new("Processing events...", file_size);
1012         if (progress == NULL)
1013                 return -1;
1014
1015         mmap_size = session->mmap_window;
1016         if (mmap_size > file_size)
1017                 mmap_size = file_size;
1018
1019         memset(mmaps, 0, sizeof(mmaps));
1020
1021         mmap_prot  = PROT_READ;
1022         mmap_flags = MAP_SHARED;
1023
1024         if (session->header.needs_swap) {
1025                 mmap_prot  |= PROT_WRITE;
1026                 mmap_flags = MAP_PRIVATE;
1027         }
1028 remap:
1029         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1030                    file_offset);
1031         if (buf == MAP_FAILED) {
1032                 pr_err("failed to mmap file\n");
1033                 err = -errno;
1034                 goto out_err;
1035         }
1036         mmaps[map_idx] = buf;
1037         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1038         file_pos = file_offset + head;
1039
1040 more:
1041         event = fetch_mmaped_event(session, head, mmap_size, buf);
1042         if (!event) {
1043                 if (mmaps[map_idx]) {
1044                         munmap(mmaps[map_idx], mmap_size);
1045                         mmaps[map_idx] = NULL;
1046                 }
1047
1048                 page_offset = page_size * (head / page_size);
1049                 file_offset += page_offset;
1050                 head -= page_offset;
1051                 goto remap;
1052         }
1053
1054         size = event->header.size;
1055
1056         if (size == 0 ||
1057             perf_session__process_event(session, event, ops, file_pos) < 0) {
1058                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1059                             file_offset + head, event->header.size,
1060                             event->header.type);
1061                 /*
1062                  * assume we lost track of the stream, check alignment, and
1063                  * increment a single u64 in the hope to catch on again 'soon'.
1064                  */
1065                 if (unlikely(head & 7))
1066                         head &= ~7ULL;
1067
1068                 size = 8;
1069         }
1070
1071         head += size;
1072         file_pos += size;
1073
1074         if (file_pos >= progress_next) {
1075                 progress_next += file_size / 16;
1076                 ui_progress__update(progress, file_pos);
1077         }
1078
1079         if (file_pos < file_size)
1080                 goto more;
1081
1082         err = 0;
1083         /* do the final flush for ordered samples */
1084         session->ordered_samples.next_flush = ULLONG_MAX;
1085         flush_sample_queue(session, ops);
1086 out_err:
1087         ui_progress__delete(progress);
1088         perf_session__warn_about_errors(session, ops);
1089         perf_session_free_sample_buffers(session);
1090         return err;
1091 }
1092
1093 int perf_session__process_events(struct perf_session *self,
1094                                  struct perf_event_ops *ops)
1095 {
1096         int err;
1097
1098         if (perf_session__register_idle_thread(self) == NULL)
1099                 return -ENOMEM;
1100
1101         if (!self->fd_pipe)
1102                 err = __perf_session__process_events(self,
1103                                                      self->header.data_offset,
1104                                                      self->header.data_size,
1105                                                      self->size, ops);
1106         else
1107                 err = __perf_session__process_pipe_events(self, ops);
1108
1109         return err;
1110 }
1111
1112 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1113 {
1114         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1115                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1116                 return false;
1117         }
1118
1119         return true;
1120 }
1121
1122 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1123                                              const char *symbol_name,
1124                                              u64 addr)
1125 {
1126         char *bracket;
1127         enum map_type i;
1128         struct ref_reloc_sym *ref;
1129
1130         ref = zalloc(sizeof(struct ref_reloc_sym));
1131         if (ref == NULL)
1132                 return -ENOMEM;
1133
1134         ref->name = strdup(symbol_name);
1135         if (ref->name == NULL) {
1136                 free(ref);
1137                 return -ENOMEM;
1138         }
1139
1140         bracket = strchr(ref->name, ']');
1141         if (bracket)
1142                 *bracket = '\0';
1143
1144         ref->addr = addr;
1145
1146         for (i = 0; i < MAP__NR_TYPES; ++i) {
1147                 struct kmap *kmap = map__kmap(maps[i]);
1148                 kmap->ref_reloc_sym = ref;
1149         }
1150
1151         return 0;
1152 }
1153
1154 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1155 {
1156         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1157                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1158                machines__fprintf_dsos(&self->machines, fp);
1159 }
1160
1161 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1162                                           bool with_hits)
1163 {
1164         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1165         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1166 }
1167
1168 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1169 {
1170         struct perf_evsel *pos;
1171         size_t ret = fprintf(fp, "Aggregated stats:\n");
1172
1173         ret += hists__fprintf_nr_events(&session->hists, fp);
1174
1175         list_for_each_entry(pos, &session->evlist->entries, node) {
1176                 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1177                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1178         }
1179
1180         return ret;
1181 }
1182
1183 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1184                                               unsigned int type)
1185 {
1186         struct perf_evsel *pos;
1187
1188         list_for_each_entry(pos, &session->evlist->entries, node) {
1189                 if (pos->attr.type == type)
1190                         return pos;
1191         }
1192         return NULL;
1193 }
1194
1195 void perf_session__print_symbols(union perf_event *event,
1196                                 struct perf_sample *sample,
1197                                 struct perf_session *session)
1198 {
1199         struct addr_location al;
1200         const char *symname, *dsoname;
1201         struct callchain_cursor *cursor = &session->callchain_cursor;
1202         struct callchain_cursor_node *node;
1203
1204         if (perf_event__preprocess_sample(event, session, &al, sample,
1205                                           NULL) < 0) {
1206                 error("problem processing %d event, skipping it.\n",
1207                         event->header.type);
1208                 return;
1209         }
1210
1211         if (symbol_conf.use_callchain && sample->callchain) {
1212
1213                 if (perf_session__resolve_callchain(session, al.thread,
1214                                                 sample->callchain, NULL) != 0) {
1215                         if (verbose)
1216                                 error("Failed to resolve callchain. Skipping\n");
1217                         return;
1218                 }
1219                 callchain_cursor_commit(cursor);
1220
1221                 while (1) {
1222                         node = callchain_cursor_current(cursor);
1223                         if (!node)
1224                                 break;
1225
1226                         if (node->sym && node->sym->name)
1227                                 symname = node->sym->name;
1228                         else
1229                                 symname = "";
1230
1231                         if (node->map && node->map->dso && node->map->dso->name)
1232                                 dsoname = node->map->dso->name;
1233                         else
1234                                 dsoname = "";
1235
1236                         printf("\t%16" PRIx64 " %s (%s)\n", node->ip, symname, dsoname);
1237
1238                         callchain_cursor_advance(cursor);
1239                 }
1240
1241         } else {
1242                 if (al.sym && al.sym->name)
1243                         symname = al.sym->name;
1244                 else
1245                         symname = "";
1246
1247                 if (al.map && al.map->dso && al.map->dso->name)
1248                         dsoname = al.map->dso->name;
1249                 else
1250                         dsoname = "";
1251
1252                 printf("%16" PRIx64 " %s (%s)", al.addr, symname, dsoname);
1253         }
1254 }