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         if (!perf_evlist__valid_sample_type(self->evlist)) {
  62                 pr_err("non matching sample_type");
  63                 goto out_close;
  64         }
  65
  66         if (!perf_evlist__valid_sample_id_all(self->evlist)) {
  67                 pr_err("non matching sample_id_all");
  68                 goto out_close;
  69         }
  70
  71         self->size = input_stat.st_size;
  72         return 0;
  73
  74 out_close:
  75         close(self->fd);
  76         self->fd = -1;
  77         return -1;
  78 }
  79
  80 static void perf_session__id_header_size(struct perf_session *session)
  81 {
  82        struct perf_sample *data;
  83        u64 sample_type = session->sample_type;
  84        u16 size = 0;
  85
  86         if (!session->sample_id_all)
  87                 goto out;
  88
  89        if (sample_type & PERF_SAMPLE_TID)
  90                size += sizeof(data->tid) * 2;
  91
  92        if (sample_type & PERF_SAMPLE_TIME)
  93                size += sizeof(data->time);
  94
  95        if (sample_type & PERF_SAMPLE_ID)
  96                size += sizeof(data->id);
  97
  98        if (sample_type & PERF_SAMPLE_STREAM_ID)
  99                size += sizeof(data->stream_id);
 100
 101        if (sample_type & PERF_SAMPLE_CPU)
 102                size += sizeof(data->cpu) * 2;
 103 out:
 104        session->id_hdr_size = size;
 105 }
 106
 107 void perf_session__update_sample_type(struct perf_session *self)
 108 {
 109         self->sample_type = perf_evlist__sample_type(self->evlist);
 110         self->sample_size = __perf_evsel__sample_size(self->sample_type);
 111         self->sample_id_all = perf_evlist__sample_id_all(self->evlist);
 112         perf_session__id_header_size(self);
 113 }
 114
 115 int perf_session__create_kernel_maps(struct perf_session *self)
 116 {
 117         int ret = machine__create_kernel_maps(&self->host_machine);
 118
 119         if (ret >= 0)
 120                 ret = machines__create_guest_kernel_maps(&self->machines);
 121         return ret;
 122 }
 123
 124 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 125 {
 126         machine__destroy_kernel_maps(&self->host_machine);
 127         machines__destroy_guest_kernel_maps(&self->machines);
 128 }
 129
 130 struct perf_session *perf_session__new(const char *filename, int mode,
 131                                        bool force, bool repipe,
 132                                        struct perf_event_ops *ops)
 133 {
 134         size_t len = filename ? strlen(filename) + 1 : 0;
 135         struct perf_session *self = zalloc(sizeof(*self) + len);
 136
 137         if (self == NULL)
 138                 goto out;
 139
 140         memcpy(self->filename, filename, len);
 141         self->threads = RB_ROOT;
 142         INIT_LIST_HEAD(&self->dead_threads);
 143         self->last_match = NULL;
 144         /*
 145          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 146          * slices. On 32bit we use 32MB.
 147          */
 148 #if BITS_PER_LONG == 64
 149         self->mmap_window = ULLONG_MAX;
 150 #else
 151         self->mmap_window = 32 * 1024 * 1024ULL;
 152 #endif
 153         self->machines = RB_ROOT;
 154         self->repipe = repipe;
 155         INIT_LIST_HEAD(&self->ordered_samples.samples);
 156         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 157         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 158         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 159
 160         if (mode == O_RDONLY) {
 161                 if (perf_session__open(self, force) < 0)
 162                         goto out_delete;
 163                 perf_session__update_sample_type(self);
 164         } else if (mode == O_WRONLY) {
 165                 /*
 166                  * In O_RDONLY mode this will be performed when reading the
 167                  * kernel MMAP event, in perf_event__process_mmap().
 168                  */
 169                 if (perf_session__create_kernel_maps(self) < 0)
 170                         goto out_delete;
 171         }
 172
 173         if (ops && ops->ordering_requires_timestamps &&
 174             ops->ordered_samples && !self->sample_id_all) {
 175                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 176                 ops->ordered_samples = false;
 177         }
 178
 179 out:
 180         return self;
 181 out_delete:
 182         perf_session__delete(self);
 183         return NULL;
 184 }
 185
 186 static void perf_session__delete_dead_threads(struct perf_session *self)
 187 {
 188         struct thread *n, *t;
 189
 190         list_for_each_entry_safe(t, n, &self->dead_threads, node) {
 191                 list_del(&t->node);
 192                 thread__delete(t);
 193         }
 194 }
 195
 196 static void perf_session__delete_threads(struct perf_session *self)
 197 {
 198         struct rb_node *nd = rb_first(&self->threads);
 199
 200         while (nd) {
 201                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 202
 203                 rb_erase(&t->rb_node, &self->threads);
 204                 nd = rb_next(nd);
 205                 thread__delete(t);
 206         }
 207 }
 208
 209 void perf_session__delete(struct perf_session *self)
 210 {
 211         perf_session__destroy_kernel_maps(self);
 212         perf_session__delete_dead_threads(self);
 213         perf_session__delete_threads(self);
 214         machine__exit(&self->host_machine);
 215         close(self->fd);
 216         free(self);
 217 }
 218
 219 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
 220 {
 221         self->last_match = NULL;
 222         rb_erase(&th->rb_node, &self->threads);
 223         /*
 224          * We may have references to this thread, for instance in some hist_entry
 225          * instances, so just move them to a separate list.
 226          */
 227         list_add_tail(&th->node, &self->dead_threads);
 228 }
 229
 230 static bool symbol__match_parent_regex(struct symbol *sym)
 231 {
 232         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 233                 return 1;
 234
 235         return 0;
 236 }
 237
 238 int perf_session__resolve_callchain(struct perf_session *self,
 239                                     struct thread *thread,
 240                                     struct ip_callchain *chain,
 241                                     struct symbol **parent)
 242 {
 243         u8 cpumode = PERF_RECORD_MISC_USER;
 244         unsigned int i;
 245         int err;
 246
 247         callchain_cursor_reset(&self->callchain_cursor);
 248
 249         for (i = 0; i < chain->nr; i++) {
 250                 u64 ip = chain->ips[i];
 251                 struct addr_location al;
 252
 253                 if (ip >= PERF_CONTEXT_MAX) {
 254                         switch (ip) {
 255                         case PERF_CONTEXT_HV:
 256                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 257                         case PERF_CONTEXT_KERNEL:
 258                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 259                         case PERF_CONTEXT_USER:
 260                                 cpumode = PERF_RECORD_MISC_USER;        break;
 261                         default:
 262                                 break;
 263                         }
 264                         continue;
 265                 }
 266
 267                 al.filtered = false;
 268                 thread__find_addr_location(thread, self, cpumode,
 269                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 270                 if (al.sym != NULL) {
 271                         if (sort__has_parent && !*parent &&
 272                             symbol__match_parent_regex(al.sym))
 273                                 *parent = al.sym;
 274                         if (!symbol_conf.use_callchain)
 275                                 break;
 276                 }
 277
 278                 err = callchain_cursor_append(&self->callchain_cursor,
 279                                               ip, al.map, al.sym);
 280                 if (err)
 281                         return err;
 282         }
 283
 284         return 0;
 285 }
 286
 287 static int process_event_synth_stub(union perf_event *event __used,
 288                                     struct perf_session *session __used)
 289 {
 290         dump_printf(": unhandled!\n");
 291         return 0;
 292 }
 293
 294 static int process_event_sample_stub(union perf_event *event __used,
 295                                      struct perf_sample *sample __used,
 296                                      struct perf_evsel *evsel __used,
 297                                      struct perf_session *session __used)
 298 {
 299         dump_printf(": unhandled!\n");
 300         return 0;
 301 }
 302
 303 static int process_event_stub(union perf_event *event __used,
 304                               struct perf_sample *sample __used,
 305                               struct perf_session *session __used)
 306 {
 307         dump_printf(": unhandled!\n");
 308         return 0;
 309 }
 310
 311 static int process_finished_round_stub(union perf_event *event __used,
 312                                        struct perf_session *session __used,
 313                                        struct perf_event_ops *ops __used)
 314 {
 315         dump_printf(": unhandled!\n");
 316         return 0;
 317 }
 318
 319 static int process_finished_round(union perf_event *event,
 320                                   struct perf_session *session,
 321                                   struct perf_event_ops *ops);
 322
 323 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 324 {
 325         if (handler->sample == NULL)
 326                 handler->sample = process_event_sample_stub;
 327         if (handler->mmap == NULL)
 328                 handler->mmap = process_event_stub;
 329         if (handler->comm == NULL)
 330                 handler->comm = process_event_stub;
 331         if (handler->fork == NULL)
 332                 handler->fork = process_event_stub;
 333         if (handler->exit == NULL)
 334                 handler->exit = process_event_stub;
 335         if (handler->lost == NULL)
 336                 handler->lost = perf_event__process_lost;
 337         if (handler->read == NULL)
 338                 handler->read = process_event_stub;
 339         if (handler->throttle == NULL)
 340                 handler->throttle = process_event_stub;
 341         if (handler->unthrottle == NULL)
 342                 handler->unthrottle = process_event_stub;
 343         if (handler->attr == NULL)
 344                 handler->attr = process_event_synth_stub;
 345         if (handler->event_type == NULL)
 346                 handler->event_type = process_event_synth_stub;
 347         if (handler->tracing_data == NULL)
 348                 handler->tracing_data = process_event_synth_stub;
 349         if (handler->build_id == NULL)
 350                 handler->build_id = process_event_synth_stub;
 351         if (handler->finished_round == NULL) {
 352                 if (handler->ordered_samples)
 353                         handler->finished_round = process_finished_round;
 354                 else
 355                         handler->finished_round = process_finished_round_stub;
 356         }
 357 }
 358
 359 void mem_bswap_64(void *src, int byte_size)
 360 {
 361         u64 *m = src;
 362
 363         while (byte_size > 0) {
 364                 *m = bswap_64(*m);
 365                 byte_size -= sizeof(u64);
 366                 ++m;
 367         }
 368 }
 369
 370 static void perf_event__all64_swap(union perf_event *event)
 371 {
 372         struct perf_event_header *hdr = &event->header;
 373         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 374 }
 375
 376 static void perf_event__comm_swap(union perf_event *event)
 377 {
 378         event->comm.pid = bswap_32(event->comm.pid);
 379         event->comm.tid = bswap_32(event->comm.tid);
 380 }
 381
 382 static void perf_event__mmap_swap(union perf_event *event)
 383 {
 384         event->mmap.pid   = bswap_32(event->mmap.pid);
 385         event->mmap.tid   = bswap_32(event->mmap.tid);
 386         event->mmap.start = bswap_64(event->mmap.start);
 387         event->mmap.len   = bswap_64(event->mmap.len);
 388         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 389 }
 390
 391 static void perf_event__task_swap(union perf_event *event)
 392 {
 393         event->fork.pid  = bswap_32(event->fork.pid);
 394         event->fork.tid  = bswap_32(event->fork.tid);
 395         event->fork.ppid = bswap_32(event->fork.ppid);
 396         event->fork.ptid = bswap_32(event->fork.ptid);
 397         event->fork.time = bswap_64(event->fork.time);
 398 }
 399
 400 static void perf_event__read_swap(union perf_event *event)
 401 {
 402         event->read.pid          = bswap_32(event->read.pid);
 403         event->read.tid          = bswap_32(event->read.tid);
 404         event->read.value        = bswap_64(event->read.value);
 405         event->read.time_enabled = bswap_64(event->read.time_enabled);
 406         event->read.time_running = bswap_64(event->read.time_running);
 407         event->read.id           = bswap_64(event->read.id);
 408 }
 409
 410 static void perf_event__attr_swap(union perf_event *event)
 411 {
 412         size_t size;
 413
 414         event->attr.attr.type           = bswap_32(event->attr.attr.type);
 415         event->attr.attr.size           = bswap_32(event->attr.attr.size);
 416         event->attr.attr.config         = bswap_64(event->attr.attr.config);
 417         event->attr.attr.sample_period  = bswap_64(event->attr.attr.sample_period);
 418         event->attr.attr.sample_type    = bswap_64(event->attr.attr.sample_type);
 419         event->attr.attr.read_format    = bswap_64(event->attr.attr.read_format);
 420         event->attr.attr.wakeup_events  = bswap_32(event->attr.attr.wakeup_events);
 421         event->attr.attr.bp_type        = bswap_32(event->attr.attr.bp_type);
 422         event->attr.attr.bp_addr        = bswap_64(event->attr.attr.bp_addr);
 423         event->attr.attr.bp_len         = bswap_64(event->attr.attr.bp_len);
 424
 425         size = event->header.size;
 426         size -= (void *)&event->attr.id - (void *)event;
 427         mem_bswap_64(event->attr.id, size);
 428 }
 429
 430 static void perf_event__event_type_swap(union perf_event *event)
 431 {
 432         event->event_type.event_type.event_id =
 433                 bswap_64(event->event_type.event_type.event_id);
 434 }
 435
 436 static void perf_event__tracing_data_swap(union perf_event *event)
 437 {
 438         event->tracing_data.size = bswap_32(event->tracing_data.size);
 439 }
 440
 441 typedef void (*perf_event__swap_op)(union perf_event *event);
 442
 443 static perf_event__swap_op perf_event__swap_ops[] = {
 444         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 445         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 446         [PERF_RECORD_FORK]                = perf_event__task_swap,
 447         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 448         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 449         [PERF_RECORD_READ]                = perf_event__read_swap,
 450         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 451         [PERF_RECORD_HEADER_ATTR]         = perf_event__attr_swap,
 452         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 453         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 454         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 455         [PERF_RECORD_HEADER_MAX]          = NULL,
 456 };
 457
 458 struct sample_queue {
 459         u64                     timestamp;
 460         u64                     file_offset;
 461         union perf_event        *event;
 462         struct list_head        list;
 463 };
 464
 465 static void perf_session_free_sample_buffers(struct perf_session *session)
 466 {
 467         struct ordered_samples *os = &session->ordered_samples;
 468
 469         while (!list_empty(&os->to_free)) {
 470                 struct sample_queue *sq;
 471
 472                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 473                 list_del(&sq->list);
 474                 free(sq);
 475         }
 476 }
 477
 478 static int perf_session_deliver_event(struct perf_session *session,
 479                                       union perf_event *event,
 480                                       struct perf_sample *sample,
 481                                       struct perf_event_ops *ops,
 482                                       u64 file_offset);
 483
 484 static void flush_sample_queue(struct perf_session *s,
 485                                struct perf_event_ops *ops)
 486 {
 487         struct ordered_samples *os = &s->ordered_samples;
 488         struct list_head *head = &os->samples;
 489         struct sample_queue *tmp, *iter;
 490         struct perf_sample sample;
 491         u64 limit = os->next_flush;
 492         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 493         int ret;
 494
 495         if (!ops->ordered_samples || !limit)
 496                 return;
 497
 498         list_for_each_entry_safe(iter, tmp, head, list) {
 499                 if (iter->timestamp > limit)
 500                         break;
 501
 502                 ret = perf_session__parse_sample(s, iter->event, &sample);
 503                 if (ret)
 504                         pr_err("Can't parse sample, err = %d\n", ret);
 505                 else
 506                         perf_session_deliver_event(s, iter->event, &sample, ops,
 507                                                    iter->file_offset);
 508
 509                 os->last_flush = iter->timestamp;
 510                 list_del(&iter->list);
 511                 list_add(&iter->list, &os->sample_cache);
 512         }
 513
 514         if (list_empty(head)) {
 515                 os->last_sample = NULL;
 516         } else if (last_ts <= limit) {
 517                 os->last_sample =
 518                         list_entry(head->prev, struct sample_queue, list);
 519         }
 520 }
 521
 522 /*
 523  * When perf record finishes a pass on every buffers, it records this pseudo
 524  * event.
 525  * We record the max timestamp t found in the pass n.
 526  * Assuming these timestamps are monotonic across cpus, we know that if
 527  * a buffer still has events with timestamps below t, they will be all
 528  * available and then read in the pass n + 1.
 529  * Hence when we start to read the pass n + 2, we can safely flush every
 530  * events with timestamps below t.
 531  *
 532  *    ============ PASS n =================
 533  *       CPU 0         |   CPU 1
 534  *                     |
 535  *    cnt1 timestamps  |   cnt2 timestamps
 536  *          1          |         2
 537  *          2          |         3
 538  *          -          |         4  <--- max recorded
 539  *
 540  *    ============ PASS n + 1 ==============
 541  *       CPU 0         |   CPU 1
 542  *                     |
 543  *    cnt1 timestamps  |   cnt2 timestamps
 544  *          3          |         5
 545  *          4          |         6
 546  *          5          |         7 <---- max recorded
 547  *
 548  *      Flush every events below timestamp 4
 549  *
 550  *    ============ PASS n + 2 ==============
 551  *       CPU 0         |   CPU 1
 552  *                     |
 553  *    cnt1 timestamps  |   cnt2 timestamps
 554  *          6          |         8
 555  *          7          |         9
 556  *          -          |         10
 557  *
 558  *      Flush every events below timestamp 7
 559  *      etc...
 560  */
 561 static int process_finished_round(union perf_event *event __used,
 562                                   struct perf_session *session,
 563                                   struct perf_event_ops *ops)
 564 {
 565         flush_sample_queue(session, ops);
 566         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 567
 568         return 0;
 569 }
 570
 571 /* The queue is ordered by time */
 572 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 573 {
 574         struct ordered_samples *os = &s->ordered_samples;
 575         struct sample_queue *sample = os->last_sample;
 576         u64 timestamp = new->timestamp;
 577         struct list_head *p;
 578
 579         os->last_sample = new;
 580
 581         if (!sample) {
 582                 list_add(&new->list, &os->samples);
 583                 os->max_timestamp = timestamp;
 584                 return;
 585         }
 586
 587         /*
 588          * last_sample might point to some random place in the list as it's
 589          * the last queued event. We expect that the new event is close to
 590          * this.
 591          */
 592         if (sample->timestamp <= timestamp) {
 593                 while (sample->timestamp <= timestamp) {
 594                         p = sample->list.next;
 595                         if (p == &os->samples) {
 596                                 list_add_tail(&new->list, &os->samples);
 597                                 os->max_timestamp = timestamp;
 598                                 return;
 599                         }
 600                         sample = list_entry(p, struct sample_queue, list);
 601                 }
 602                 list_add_tail(&new->list, &sample->list);
 603         } else {
 604                 while (sample->timestamp > timestamp) {
 605                         p = sample->list.prev;
 606                         if (p == &os->samples) {
 607                                 list_add(&new->list, &os->samples);
 608                                 return;
 609                         }
 610                         sample = list_entry(p, struct sample_queue, list);
 611                 }
 612                 list_add(&new->list, &sample->list);
 613         }
 614 }
 615
 616 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 617
 618 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 619                                     struct perf_sample *sample, u64 file_offset)
 620 {
 621         struct ordered_samples *os = &s->ordered_samples;
 622         struct list_head *sc = &os->sample_cache;
 623         u64 timestamp = sample->time;
 624         struct sample_queue *new;
 625
 626         if (!timestamp || timestamp == ~0ULL)
 627                 return -ETIME;
 628
 629         if (timestamp < s->ordered_samples.last_flush) {
 630                 printf("Warning: Timestamp below last timeslice flush\n");
 631                 return -EINVAL;
 632         }
 633
 634         if (!list_empty(sc)) {
 635                 new = list_entry(sc->next, struct sample_queue, list);
 636                 list_del(&new->list);
 637         } else if (os->sample_buffer) {
 638                 new = os->sample_buffer + os->sample_buffer_idx;
 639                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 640                         os->sample_buffer = NULL;
 641         } else {
 642                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 643                 if (!os->sample_buffer)
 644                         return -ENOMEM;
 645                 list_add(&os->sample_buffer->list, &os->to_free);
 646                 os->sample_buffer_idx = 2;
 647                 new = os->sample_buffer + 1;
 648         }
 649
 650         new->timestamp = timestamp;
 651         new->file_offset = file_offset;
 652         new->event = event;
 653
 654         __queue_event(new, s);
 655
 656         return 0;
 657 }
 658
 659 static void callchain__printf(struct perf_sample *sample)
 660 {
 661         unsigned int i;
 662
 663         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 664
 665         for (i = 0; i < sample->callchain->nr; i++)
 666                 printf("..... %2d: %016" PRIx64 "\n",
 667                        i, sample->callchain->ips[i]);
 668 }
 669
 670 static void perf_session__print_tstamp(struct perf_session *session,
 671                                        union perf_event *event,
 672                                        struct perf_sample *sample)
 673 {
 674         if (event->header.type != PERF_RECORD_SAMPLE &&
 675             !session->sample_id_all) {
 676                 fputs("-1 -1 ", stdout);
 677                 return;
 678         }
 679
 680         if ((session->sample_type & PERF_SAMPLE_CPU))
 681                 printf("%u ", sample->cpu);
 682
 683         if (session->sample_type & PERF_SAMPLE_TIME)
 684                 printf("%" PRIu64 " ", sample->time);
 685 }
 686
 687 static void dump_event(struct perf_session *session, union perf_event *event,
 688                        u64 file_offset, struct perf_sample *sample)
 689 {
 690         if (!dump_trace)
 691                 return;
 692
 693         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 694                file_offset, event->header.size, event->header.type);
 695
 696         trace_event(event);
 697
 698         if (sample)
 699                 perf_session__print_tstamp(session, event, sample);
 700
 701         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 702                event->header.size, perf_event__name(event->header.type));
 703 }
 704
 705 static void dump_sample(struct perf_session *session, union perf_event *event,
 706                         struct perf_sample *sample)
 707 {
 708         if (!dump_trace)
 709                 return;
 710
 711         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 "\n",
 712                event->header.misc, sample->pid, sample->tid, sample->ip,
 713                sample->period);
 714
 715         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 716                 callchain__printf(sample);
 717 }
 718
 719 static int perf_session_deliver_event(struct perf_session *session,
 720                                       union perf_event *event,
 721                                       struct perf_sample *sample,
 722                                       struct perf_event_ops *ops,
 723                                       u64 file_offset)
 724 {
 725         struct perf_evsel *evsel;
 726
 727         dump_event(session, event, file_offset, sample);
 728
 729         switch (event->header.type) {
 730         case PERF_RECORD_SAMPLE:
 731                 dump_sample(session, event, sample);
 732                 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 733                 if (evsel == NULL) {
 734                         ++session->hists.stats.nr_unknown_id;
 735                         return -1;
 736                 }
 737                 return ops->sample(event, sample, evsel, session);
 738         case PERF_RECORD_MMAP:
 739                 return ops->mmap(event, sample, session);
 740         case PERF_RECORD_COMM:
 741                 return ops->comm(event, sample, session);
 742         case PERF_RECORD_FORK:
 743                 return ops->fork(event, sample, session);
 744         case PERF_RECORD_EXIT:
 745                 return ops->exit(event, sample, session);
 746         case PERF_RECORD_LOST:
 747                 return ops->lost(event, sample, session);
 748         case PERF_RECORD_READ:
 749                 return ops->read(event, sample, session);
 750         case PERF_RECORD_THROTTLE:
 751                 return ops->throttle(event, sample, session);
 752         case PERF_RECORD_UNTHROTTLE:
 753                 return ops->unthrottle(event, sample, session);
 754         default:
 755                 ++session->hists.stats.nr_unknown_events;
 756                 return -1;
 757         }
 758 }
 759
 760 static int perf_session__preprocess_sample(struct perf_session *session,
 761                                            union perf_event *event, struct perf_sample *sample)
 762 {
 763         if (event->header.type != PERF_RECORD_SAMPLE ||
 764             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 765                 return 0;
 766
 767         if (!ip_callchain__valid(sample->callchain, event)) {
 768                 pr_debug("call-chain problem with event, skipping it.\n");
 769                 ++session->hists.stats.nr_invalid_chains;
 770                 session->hists.stats.total_invalid_chains += sample->period;
 771                 return -EINVAL;
 772         }
 773         return 0;
 774 }
 775
 776 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
 777                                             struct perf_event_ops *ops, u64 file_offset)
 778 {
 779         dump_event(session, event, file_offset, NULL);
 780
 781         /* These events are processed right away */
 782         switch (event->header.type) {
 783         case PERF_RECORD_HEADER_ATTR:
 784                 return ops->attr(event, session);
 785         case PERF_RECORD_HEADER_EVENT_TYPE:
 786                 return ops->event_type(event, session);
 787         case PERF_RECORD_HEADER_TRACING_DATA:
 788                 /* setup for reading amidst mmap */
 789                 lseek(session->fd, file_offset, SEEK_SET);
 790                 return ops->tracing_data(event, session);
 791         case PERF_RECORD_HEADER_BUILD_ID:
 792                 return ops->build_id(event, session);
 793         case PERF_RECORD_FINISHED_ROUND:
 794                 return ops->finished_round(event, session, ops);
 795         default:
 796                 return -EINVAL;
 797         }
 798 }
 799
 800 static int perf_session__process_event(struct perf_session *session,
 801                                        union perf_event *event,
 802                                        struct perf_event_ops *ops,
 803                                        u64 file_offset)
 804 {
 805         struct perf_sample sample;
 806         int ret;
 807
 808         if (session->header.needs_swap &&
 809             perf_event__swap_ops[event->header.type])
 810                 perf_event__swap_ops[event->header.type](event);
 811
 812         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 813                 return -EINVAL;
 814
 815         hists__inc_nr_events(&session->hists, event->header.type);
 816
 817         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 818                 return perf_session__process_user_event(session, event, ops, file_offset);
 819
 820         /*
 821          * For all kernel events we get the sample data
 822          */
 823         ret = perf_session__parse_sample(session, event, &sample);
 824         if (ret)
 825                 return ret;
 826
 827         /* Preprocess sample records - precheck callchains */
 828         if (perf_session__preprocess_sample(session, event, &sample))
 829                 return 0;
 830
 831         if (ops->ordered_samples) {
 832                 ret = perf_session_queue_event(session, event, &sample,
 833                                                file_offset);
 834                 if (ret != -ETIME)
 835                         return ret;
 836         }
 837
 838         return perf_session_deliver_event(session, event, &sample, ops,
 839                                           file_offset);
 840 }
 841
 842 void perf_event_header__bswap(struct perf_event_header *self)
 843 {
 844         self->type = bswap_32(self->type);
 845         self->misc = bswap_16(self->misc);
 846         self->size = bswap_16(self->size);
 847 }
 848
 849 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 850 {
 851         struct thread *thread = perf_session__findnew(self, 0);
 852
 853         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 854                 pr_err("problem inserting idle task.\n");
 855                 thread = NULL;
 856         }
 857
 858         return thread;
 859 }
 860
 861 static void perf_session__warn_about_errors(const struct perf_session *session,
 862                                             const struct perf_event_ops *ops)
 863 {
 864         if (ops->lost == perf_event__process_lost &&
 865             session->hists.stats.total_lost != 0) {
 866                 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64
 867                             "!\n\nCheck IO/CPU overload!\n\n",
 868                             session->hists.stats.total_period,
 869                             session->hists.stats.total_lost);
 870         }
 871
 872         if (session->hists.stats.nr_unknown_events != 0) {
 873                 ui__warning("Found %u unknown events!\n\n"
 874                             "Is this an older tool processing a perf.data "
 875                             "file generated by a more recent tool?\n\n"
 876                             "If that is not the case, consider "
 877                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 878                             session->hists.stats.nr_unknown_events);
 879         }
 880
 881         if (session->hists.stats.nr_unknown_id != 0) {
 882                 ui__warning("%u samples with id not present in the header\n",
 883                             session->hists.stats.nr_unknown_id);
 884         }
 885
 886         if (session->hists.stats.nr_invalid_chains != 0) {
 887                 ui__warning("Found invalid callchains!\n\n"
 888                             "%u out of %u events were discarded for this reason.\n\n"
 889                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 890                             session->hists.stats.nr_invalid_chains,
 891                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 892         }
 893 }
 894
 895 #define session_done()  (*(volatile int *)(&session_done))
 896 volatile int session_done;
 897
 898 static int __perf_session__process_pipe_events(struct perf_session *self,
 899                                                struct perf_event_ops *ops)
 900 {
 901         union perf_event event;
 902         uint32_t size;
 903         int skip = 0;
 904         u64 head;
 905         int err;
 906         void *p;
 907
 908         perf_event_ops__fill_defaults(ops);
 909
 910         head = 0;
 911 more:
 912         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 913         if (err <= 0) {
 914                 if (err == 0)
 915                         goto done;
 916
 917                 pr_err("failed to read event header\n");
 918                 goto out_err;
 919         }
 920
 921         if (self->header.needs_swap)
 922                 perf_event_header__bswap(&event.header);
 923
 924         size = event.header.size;
 925         if (size == 0)
 926                 size = 8;
 927
 928         p = &event;
 929         p += sizeof(struct perf_event_header);
 930
 931         if (size - sizeof(struct perf_event_header)) {
 932                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 933                 if (err <= 0) {
 934                         if (err == 0) {
 935                                 pr_err("unexpected end of event stream\n");
 936                                 goto done;
 937                         }
 938
 939                         pr_err("failed to read event data\n");
 940                         goto out_err;
 941                 }
 942         }
 943
 944         if (size == 0 ||
 945             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 946                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 947                             head, event.header.size, event.header.type);
 948                 /*
 949                  * assume we lost track of the stream, check alignment, and
 950                  * increment a single u64 in the hope to catch on again 'soon'.
 951                  */
 952                 if (unlikely(head & 7))
 953                         head &= ~7ULL;
 954
 955                 size = 8;
 956         }
 957
 958         head += size;
 959
 960         if (skip > 0)
 961                 head += skip;
 962
 963         if (!session_done())
 964                 goto more;
 965 done:
 966         err = 0;
 967 out_err:
 968         perf_session__warn_about_errors(self, ops);
 969         perf_session_free_sample_buffers(self);
 970         return err;
 971 }
 972
 973 static union perf_event *
 974 fetch_mmaped_event(struct perf_session *session,
 975                    u64 head, size_t mmap_size, char *buf)
 976 {
 977         union perf_event *event;
 978
 979         /*
 980          * Ensure we have enough space remaining to read
 981          * the size of the event in the headers.
 982          */
 983         if (head + sizeof(event->header) > mmap_size)
 984                 return NULL;
 985
 986         event = (union perf_event *)(buf + head);
 987
 988         if (session->header.needs_swap)
 989                 perf_event_header__bswap(&event->header);
 990
 991         if (head + event->header.size > mmap_size)
 992                 return NULL;
 993
 994         return event;
 995 }
 996
 997 int __perf_session__process_events(struct perf_session *session,
 998                                    u64 data_offset, u64 data_size,
 999                                    u64 file_size, struct perf_event_ops *ops)
1000 {
1001         u64 head, page_offset, file_offset, file_pos, progress_next;
1002         int err, mmap_prot, mmap_flags, map_idx = 0;
1003         struct ui_progress *progress;
1004         size_t  page_size, mmap_size;
1005         char *buf, *mmaps[8];
1006         union perf_event *event;
1007         uint32_t size;
1008
1009         perf_event_ops__fill_defaults(ops);
1010
1011         page_size = sysconf(_SC_PAGESIZE);
1012
1013         page_offset = page_size * (data_offset / page_size);
1014         file_offset = page_offset;
1015         head = data_offset - page_offset;
1016
1017         if (data_offset + data_size < file_size)
1018                 file_size = data_offset + data_size;
1019
1020         progress_next = file_size / 16;
1021         progress = ui_progress__new("Processing events...", file_size);
1022         if (progress == NULL)
1023                 return -1;
1024
1025         mmap_size = session->mmap_window;
1026         if (mmap_size > file_size)
1027                 mmap_size = file_size;
1028
1029         memset(mmaps, 0, sizeof(mmaps));
1030
1031         mmap_prot  = PROT_READ;
1032         mmap_flags = MAP_SHARED;
1033
1034         if (session->header.needs_swap) {
1035                 mmap_prot  |= PROT_WRITE;
1036                 mmap_flags = MAP_PRIVATE;
1037         }
1038 remap:
1039         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1040                    file_offset);
1041         if (buf == MAP_FAILED) {
1042                 pr_err("failed to mmap file\n");
1043                 err = -errno;
1044                 goto out_err;
1045         }
1046         mmaps[map_idx] = buf;
1047         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1048         file_pos = file_offset + head;
1049
1050 more:
1051         event = fetch_mmaped_event(session, head, mmap_size, buf);
1052         if (!event) {
1053                 if (mmaps[map_idx]) {
1054                         munmap(mmaps[map_idx], mmap_size);
1055                         mmaps[map_idx] = NULL;
1056                 }
1057
1058                 page_offset = page_size * (head / page_size);
1059                 file_offset += page_offset;
1060                 head -= page_offset;
1061                 goto remap;
1062         }
1063
1064         size = event->header.size;
1065
1066         if (size == 0 ||
1067             perf_session__process_event(session, event, ops, file_pos) < 0) {
1068                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1069                             file_offset + head, event->header.size,
1070                             event->header.type);
1071                 /*
1072                  * assume we lost track of the stream, check alignment, and
1073                  * increment a single u64 in the hope to catch on again 'soon'.
1074                  */
1075                 if (unlikely(head & 7))
1076                         head &= ~7ULL;
1077
1078                 size = 8;
1079         }
1080
1081         head += size;
1082         file_pos += size;
1083
1084         if (file_pos >= progress_next) {
1085                 progress_next += file_size / 16;
1086                 ui_progress__update(progress, file_pos);
1087         }
1088
1089         if (file_pos < file_size)
1090                 goto more;
1091
1092         err = 0;
1093         /* do the final flush for ordered samples */
1094         session->ordered_samples.next_flush = ULLONG_MAX;
1095         flush_sample_queue(session, ops);
1096 out_err:
1097         ui_progress__delete(progress);
1098         perf_session__warn_about_errors(session, ops);
1099         perf_session_free_sample_buffers(session);
1100         return err;
1101 }
1102
1103 int perf_session__process_events(struct perf_session *self,
1104                                  struct perf_event_ops *ops)
1105 {
1106         int err;
1107
1108         if (perf_session__register_idle_thread(self) == NULL)
1109                 return -ENOMEM;
1110
1111         if (!self->fd_pipe)
1112                 err = __perf_session__process_events(self,
1113                                                      self->header.data_offset,
1114                                                      self->header.data_size,
1115                                                      self->size, ops);
1116         else
1117                 err = __perf_session__process_pipe_events(self, ops);
1118
1119         return err;
1120 }
1121
1122 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1123 {
1124         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1125                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1126                 return false;
1127         }
1128
1129         return true;
1130 }
1131
1132 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1133                                              const char *symbol_name,
1134                                              u64 addr)
1135 {
1136         char *bracket;
1137         enum map_type i;
1138         struct ref_reloc_sym *ref;
1139
1140         ref = zalloc(sizeof(struct ref_reloc_sym));
1141         if (ref == NULL)
1142                 return -ENOMEM;
1143
1144         ref->name = strdup(symbol_name);
1145         if (ref->name == NULL) {
1146                 free(ref);
1147                 return -ENOMEM;
1148         }
1149
1150         bracket = strchr(ref->name, ']');
1151         if (bracket)
1152                 *bracket = '\0';
1153
1154         ref->addr = addr;
1155
1156         for (i = 0; i < MAP__NR_TYPES; ++i) {
1157                 struct kmap *kmap = map__kmap(maps[i]);
1158                 kmap->ref_reloc_sym = ref;
1159         }
1160
1161         return 0;
1162 }
1163
1164 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1165 {
1166         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1167                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1168                machines__fprintf_dsos(&self->machines, fp);
1169 }
1170
1171 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1172                                           bool with_hits)
1173 {
1174         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1175         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1176 }
1177
1178 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1179 {
1180         struct perf_evsel *pos;
1181         size_t ret = fprintf(fp, "Aggregated stats:\n");
1182
1183         ret += hists__fprintf_nr_events(&session->hists, fp);
1184
1185         list_for_each_entry(pos, &session->evlist->entries, node) {
1186                 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1187                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1188         }
1189
1190         return ret;
1191 }
1192
1193 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1194                                               unsigned int type)
1195 {
1196         struct perf_evsel *pos;
1197
1198         list_for_each_entry(pos, &session->evlist->entries, node) {
1199                 if (pos->attr.type == type)
1200                         return pos;
1201         }
1202         return NULL;
1203 }
1204
1205 void perf_session__print_symbols(union perf_event *event,
1206                                 struct perf_sample *sample,
1207                                 struct perf_session *session)
1208 {
1209         struct addr_location al;
1210         const char *symname, *dsoname;
1211         struct callchain_cursor *cursor = &session->callchain_cursor;
1212         struct callchain_cursor_node *node;
1213
1214         if (perf_event__preprocess_sample(event, session, &al, sample,
1215                                           NULL) < 0) {
1216                 error("problem processing %d event, skipping it.\n",
1217                         event->header.type);
1218                 return;
1219         }
1220
1221         if (symbol_conf.use_callchain && sample->callchain) {
1222
1223                 if (perf_session__resolve_callchain(session, al.thread,
1224                                                 sample->callchain, NULL) != 0) {
1225                         if (verbose)
1226                                 error("Failed to resolve callchain. Skipping\n");
1227                         return;
1228                 }
1229                 callchain_cursor_commit(cursor);
1230
1231                 while (1) {
1232                         node = callchain_cursor_current(cursor);
1233                         if (!node)
1234                                 break;
1235
1236                         if (node->sym && node->sym->name)
1237                                 symname = node->sym->name;
1238                         else
1239                                 symname = "";
1240
1241                         if (node->map && node->map->dso && node->map->dso->name)
1242                                 dsoname = node->map->dso->name;
1243                         else
1244                                 dsoname = "";
1245
1246                         printf("\t%16" PRIx64 " %s (%s)\n", node->ip, symname, dsoname);
1247
1248                         callchain_cursor_advance(cursor);
1249                 }
1250
1251         } else {
1252                 if (al.sym && al.sym->name)
1253                         symname = al.sym->name;
1254                 else
1255                         symname = "";
1256
1257                 if (al.map && al.map->dso && al.map->dso->name)
1258                         dsoname = al.map->dso->name;
1259                 else
1260                         dsoname = "";
1261
1262                 printf("%16" PRIx64 " %s (%s)", al.addr, symname, dsoname);
1263         }
1264 }