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 #include "cpumap.h"
  16
  17 static int perf_session__open(struct perf_session *self, bool force)
  18 {
  19         struct stat input_stat;
  20
  21         if (!strcmp(self->filename, "-")) {
  22                 self->fd_pipe = true;
  23                 self->fd = STDIN_FILENO;
  24
  25                 if (perf_session__read_header(self, self->fd) < 0)
  26                         pr_err("incompatible file format");
  27
  28                 return 0;
  29         }
  30
  31         self->fd = open(self->filename, O_RDONLY);
  32         if (self->fd < 0) {
  33                 int err = errno;
  34
  35                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  36                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  37                         pr_err("  (try 'perf record' first)");
  38                 pr_err("\n");
  39                 return -errno;
  40         }
  41
  42         if (fstat(self->fd, &input_stat) < 0)
  43                 goto out_close;
  44
  45         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  46                 pr_err("file %s not owned by current user or root\n",
  47                        self->filename);
  48                 goto out_close;
  49         }
  50
  51         if (!input_stat.st_size) {
  52                 pr_info("zero-sized file (%s), nothing to do!\n",
  53                         self->filename);
  54                 goto out_close;
  55         }
  56
  57         if (perf_session__read_header(self, self->fd) < 0) {
  58                 pr_err("incompatible file format");
  59                 goto out_close;
  60         }
  61
  62         if (!perf_evlist__valid_sample_type(self->evlist)) {
  63                 pr_err("non matching sample_type");
  64                 goto out_close;
  65         }
  66
  67         if (!perf_evlist__valid_sample_id_all(self->evlist)) {
  68                 pr_err("non matching sample_id_all");
  69                 goto out_close;
  70         }
  71
  72         self->size = input_stat.st_size;
  73         return 0;
  74
  75 out_close:
  76         close(self->fd);
  77         self->fd = -1;
  78         return -1;
  79 }
  80
  81 void perf_session__update_sample_type(struct perf_session *self)
  82 {
  83         self->sample_type = perf_evlist__sample_type(self->evlist);
  84         self->sample_size = __perf_evsel__sample_size(self->sample_type);
  85         self->sample_id_all = perf_evlist__sample_id_all(self->evlist);
  86         self->id_hdr_size = perf_evlist__id_hdr_size(self->evlist);
  87 }
  88
  89 int perf_session__create_kernel_maps(struct perf_session *self)
  90 {
  91         int ret = machine__create_kernel_maps(&self->host_machine);
  92
  93         if (ret >= 0)
  94                 ret = machines__create_guest_kernel_maps(&self->machines);
  95         return ret;
  96 }
  97
  98 static void perf_session__destroy_kernel_maps(struct perf_session *self)
  99 {
 100         machine__destroy_kernel_maps(&self->host_machine);
 101         machines__destroy_guest_kernel_maps(&self->machines);
 102 }
 103
 104 struct perf_session *perf_session__new(const char *filename, int mode,
 105                                        bool force, bool repipe,
 106                                        struct perf_event_ops *ops)
 107 {
 108         size_t len = filename ? strlen(filename) + 1 : 0;
 109         struct perf_session *self = zalloc(sizeof(*self) + len);
 110
 111         if (self == NULL)
 112                 goto out;
 113
 114         memcpy(self->filename, filename, len);
 115         /*
 116          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 117          * slices. On 32bit we use 32MB.
 118          */
 119 #if BITS_PER_LONG == 64
 120         self->mmap_window = ULLONG_MAX;
 121 #else
 122         self->mmap_window = 32 * 1024 * 1024ULL;
 123 #endif
 124         self->machines = RB_ROOT;
 125         self->repipe = repipe;
 126         INIT_LIST_HEAD(&self->ordered_samples.samples);
 127         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 128         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 129         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 130
 131         if (mode == O_RDONLY) {
 132                 if (perf_session__open(self, force) < 0)
 133                         goto out_delete;
 134                 perf_session__update_sample_type(self);
 135         } else if (mode == O_WRONLY) {
 136                 /*
 137                  * In O_RDONLY mode this will be performed when reading the
 138                  * kernel MMAP event, in perf_event__process_mmap().
 139                  */
 140                 if (perf_session__create_kernel_maps(self) < 0)
 141                         goto out_delete;
 142         }
 143
 144         if (ops && ops->ordering_requires_timestamps &&
 145             ops->ordered_samples && !self->sample_id_all) {
 146                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 147                 ops->ordered_samples = false;
 148         }
 149
 150 out:
 151         return self;
 152 out_delete:
 153         perf_session__delete(self);
 154         return NULL;
 155 }
 156
 157 static void machine__delete_dead_threads(struct machine *machine)
 158 {
 159         struct thread *n, *t;
 160
 161         list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
 162                 list_del(&t->node);
 163                 thread__delete(t);
 164         }
 165 }
 166
 167 static void perf_session__delete_dead_threads(struct perf_session *session)
 168 {
 169         machine__delete_dead_threads(&session->host_machine);
 170 }
 171
 172 static void machine__delete_threads(struct machine *self)
 173 {
 174         struct rb_node *nd = rb_first(&self->threads);
 175
 176         while (nd) {
 177                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 178
 179                 rb_erase(&t->rb_node, &self->threads);
 180                 nd = rb_next(nd);
 181                 thread__delete(t);
 182         }
 183 }
 184
 185 static void perf_session__delete_threads(struct perf_session *session)
 186 {
 187         machine__delete_threads(&session->host_machine);
 188 }
 189
 190 void perf_session__delete(struct perf_session *self)
 191 {
 192         perf_session__destroy_kernel_maps(self);
 193         perf_session__delete_dead_threads(self);
 194         perf_session__delete_threads(self);
 195         machine__exit(&self->host_machine);
 196         close(self->fd);
 197         free(self);
 198 }
 199
 200 void machine__remove_thread(struct machine *self, struct thread *th)
 201 {
 202         self->last_match = NULL;
 203         rb_erase(&th->rb_node, &self->threads);
 204         /*
 205          * We may have references to this thread, for instance in some hist_entry
 206          * instances, so just move them to a separate list.
 207          */
 208         list_add_tail(&th->node, &self->dead_threads);
 209 }
 210
 211 static bool symbol__match_parent_regex(struct symbol *sym)
 212 {
 213         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 214                 return 1;
 215
 216         return 0;
 217 }
 218
 219 int perf_session__resolve_callchain(struct perf_session *self,
 220                                     struct thread *thread,
 221                                     struct ip_callchain *chain,
 222                                     struct symbol **parent)
 223 {
 224         u8 cpumode = PERF_RECORD_MISC_USER;
 225         unsigned int i;
 226         int err;
 227
 228         callchain_cursor_reset(&self->callchain_cursor);
 229
 230         for (i = 0; i < chain->nr; i++) {
 231                 u64 ip;
 232                 struct addr_location al;
 233
 234                 if (callchain_param.order == ORDER_CALLEE)
 235                         ip = chain->ips[i];
 236                 else
 237                         ip = chain->ips[chain->nr - i - 1];
 238
 239                 if (ip >= PERF_CONTEXT_MAX) {
 240                         switch (ip) {
 241                         case PERF_CONTEXT_HV:
 242                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 243                         case PERF_CONTEXT_KERNEL:
 244                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 245                         case PERF_CONTEXT_USER:
 246                                 cpumode = PERF_RECORD_MISC_USER;        break;
 247                         default:
 248                                 break;
 249                         }
 250                         continue;
 251                 }
 252
 253                 al.filtered = false;
 254                 thread__find_addr_location(thread, self, cpumode,
 255                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 256                 if (al.sym != NULL) {
 257                         if (sort__has_parent && !*parent &&
 258                             symbol__match_parent_regex(al.sym))
 259                                 *parent = al.sym;
 260                         if (!symbol_conf.use_callchain)
 261                                 break;
 262                 }
 263
 264                 err = callchain_cursor_append(&self->callchain_cursor,
 265                                               ip, al.map, al.sym);
 266                 if (err)
 267                         return err;
 268         }
 269
 270         return 0;
 271 }
 272
 273 static int process_event_synth_stub(union perf_event *event __used,
 274                                     struct perf_session *session __used)
 275 {
 276         dump_printf(": unhandled!\n");
 277         return 0;
 278 }
 279
 280 static int process_event_sample_stub(union perf_event *event __used,
 281                                      struct perf_sample *sample __used,
 282                                      struct perf_evsel *evsel __used,
 283                                      struct perf_session *session __used)
 284 {
 285         dump_printf(": unhandled!\n");
 286         return 0;
 287 }
 288
 289 static int process_event_stub(union perf_event *event __used,
 290                               struct perf_sample *sample __used,
 291                               struct perf_session *session __used)
 292 {
 293         dump_printf(": unhandled!\n");
 294         return 0;
 295 }
 296
 297 static int process_finished_round_stub(union perf_event *event __used,
 298                                        struct perf_session *session __used,
 299                                        struct perf_event_ops *ops __used)
 300 {
 301         dump_printf(": unhandled!\n");
 302         return 0;
 303 }
 304
 305 static int process_finished_round(union perf_event *event,
 306                                   struct perf_session *session,
 307                                   struct perf_event_ops *ops);
 308
 309 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 310 {
 311         if (handler->sample == NULL)
 312                 handler->sample = process_event_sample_stub;
 313         if (handler->mmap == NULL)
 314                 handler->mmap = process_event_stub;
 315         if (handler->comm == NULL)
 316                 handler->comm = process_event_stub;
 317         if (handler->fork == NULL)
 318                 handler->fork = process_event_stub;
 319         if (handler->exit == NULL)
 320                 handler->exit = process_event_stub;
 321         if (handler->lost == NULL)
 322                 handler->lost = perf_event__process_lost;
 323         if (handler->read == NULL)
 324                 handler->read = process_event_stub;
 325         if (handler->throttle == NULL)
 326                 handler->throttle = process_event_stub;
 327         if (handler->unthrottle == NULL)
 328                 handler->unthrottle = process_event_stub;
 329         if (handler->attr == NULL)
 330                 handler->attr = process_event_synth_stub;
 331         if (handler->event_type == NULL)
 332                 handler->event_type = process_event_synth_stub;
 333         if (handler->tracing_data == NULL)
 334                 handler->tracing_data = process_event_synth_stub;
 335         if (handler->build_id == NULL)
 336                 handler->build_id = process_event_synth_stub;
 337         if (handler->finished_round == NULL) {
 338                 if (handler->ordered_samples)
 339                         handler->finished_round = process_finished_round;
 340                 else
 341                         handler->finished_round = process_finished_round_stub;
 342         }
 343 }
 344
 345 void mem_bswap_64(void *src, int byte_size)
 346 {
 347         u64 *m = src;
 348
 349         while (byte_size > 0) {
 350                 *m = bswap_64(*m);
 351                 byte_size -= sizeof(u64);
 352                 ++m;
 353         }
 354 }
 355
 356 static void perf_event__all64_swap(union perf_event *event)
 357 {
 358         struct perf_event_header *hdr = &event->header;
 359         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 360 }
 361
 362 static void perf_event__comm_swap(union perf_event *event)
 363 {
 364         event->comm.pid = bswap_32(event->comm.pid);
 365         event->comm.tid = bswap_32(event->comm.tid);
 366 }
 367
 368 static void perf_event__mmap_swap(union perf_event *event)
 369 {
 370         event->mmap.pid   = bswap_32(event->mmap.pid);
 371         event->mmap.tid   = bswap_32(event->mmap.tid);
 372         event->mmap.start = bswap_64(event->mmap.start);
 373         event->mmap.len   = bswap_64(event->mmap.len);
 374         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 375 }
 376
 377 static void perf_event__task_swap(union perf_event *event)
 378 {
 379         event->fork.pid  = bswap_32(event->fork.pid);
 380         event->fork.tid  = bswap_32(event->fork.tid);
 381         event->fork.ppid = bswap_32(event->fork.ppid);
 382         event->fork.ptid = bswap_32(event->fork.ptid);
 383         event->fork.time = bswap_64(event->fork.time);
 384 }
 385
 386 static void perf_event__read_swap(union perf_event *event)
 387 {
 388         event->read.pid          = bswap_32(event->read.pid);
 389         event->read.tid          = bswap_32(event->read.tid);
 390         event->read.value        = bswap_64(event->read.value);
 391         event->read.time_enabled = bswap_64(event->read.time_enabled);
 392         event->read.time_running = bswap_64(event->read.time_running);
 393         event->read.id           = bswap_64(event->read.id);
 394 }
 395
 396 /* exported for swapping attributes in file header */
 397 void perf_event__attr_swap(struct perf_event_attr *attr)
 398 {
 399         attr->type              = bswap_32(attr->type);
 400         attr->size              = bswap_32(attr->size);
 401         attr->config            = bswap_64(attr->config);
 402         attr->sample_period     = bswap_64(attr->sample_period);
 403         attr->sample_type       = bswap_64(attr->sample_type);
 404         attr->read_format       = bswap_64(attr->read_format);
 405         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 406         attr->bp_type           = bswap_32(attr->bp_type);
 407         attr->bp_addr           = bswap_64(attr->bp_addr);
 408         attr->bp_len            = bswap_64(attr->bp_len);
 409 }
 410
 411 static void perf_event__hdr_attr_swap(union perf_event *event)
 412 {
 413         size_t size;
 414
 415         perf_event__attr_swap(&event->attr.attr);
 416
 417         size = event->header.size;
 418         size -= (void *)&event->attr.id - (void *)event;
 419         mem_bswap_64(event->attr.id, size);
 420 }
 421
 422 static void perf_event__event_type_swap(union perf_event *event)
 423 {
 424         event->event_type.event_type.event_id =
 425                 bswap_64(event->event_type.event_type.event_id);
 426 }
 427
 428 static void perf_event__tracing_data_swap(union perf_event *event)
 429 {
 430         event->tracing_data.size = bswap_32(event->tracing_data.size);
 431 }
 432
 433 typedef void (*perf_event__swap_op)(union perf_event *event);
 434
 435 static perf_event__swap_op perf_event__swap_ops[] = {
 436         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 437         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 438         [PERF_RECORD_FORK]                = perf_event__task_swap,
 439         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 440         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 441         [PERF_RECORD_READ]                = perf_event__read_swap,
 442         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 443         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 444         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 445         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 446         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 447         [PERF_RECORD_HEADER_MAX]          = NULL,
 448 };
 449
 450 struct sample_queue {
 451         u64                     timestamp;
 452         u64                     file_offset;
 453         union perf_event        *event;
 454         struct list_head        list;
 455 };
 456
 457 static void perf_session_free_sample_buffers(struct perf_session *session)
 458 {
 459         struct ordered_samples *os = &session->ordered_samples;
 460
 461         while (!list_empty(&os->to_free)) {
 462                 struct sample_queue *sq;
 463
 464                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 465                 list_del(&sq->list);
 466                 free(sq);
 467         }
 468 }
 469
 470 static int perf_session_deliver_event(struct perf_session *session,
 471                                       union perf_event *event,
 472                                       struct perf_sample *sample,
 473                                       struct perf_event_ops *ops,
 474                                       u64 file_offset);
 475
 476 static void flush_sample_queue(struct perf_session *s,
 477                                struct perf_event_ops *ops)
 478 {
 479         struct ordered_samples *os = &s->ordered_samples;
 480         struct list_head *head = &os->samples;
 481         struct sample_queue *tmp, *iter;
 482         struct perf_sample sample;
 483         u64 limit = os->next_flush;
 484         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 485         unsigned idx = 0, progress_next = os->nr_samples / 16;
 486         int ret;
 487
 488         if (!ops->ordered_samples || !limit)
 489                 return;
 490
 491         list_for_each_entry_safe(iter, tmp, head, list) {
 492                 if (iter->timestamp > limit)
 493                         break;
 494
 495                 ret = perf_session__parse_sample(s, iter->event, &sample);
 496                 if (ret)
 497                         pr_err("Can't parse sample, err = %d\n", ret);
 498                 else
 499                         perf_session_deliver_event(s, iter->event, &sample, ops,
 500                                                    iter->file_offset);
 501
 502                 os->last_flush = iter->timestamp;
 503                 list_del(&iter->list);
 504                 list_add(&iter->list, &os->sample_cache);
 505                 if (++idx >= progress_next) {
 506                         progress_next += os->nr_samples / 16;
 507                         ui_progress__update(idx, os->nr_samples,
 508                                             "Processing time ordered events...");
 509                 }
 510         }
 511
 512         if (list_empty(head)) {
 513                 os->last_sample = NULL;
 514         } else if (last_ts <= limit) {
 515                 os->last_sample =
 516                         list_entry(head->prev, struct sample_queue, list);
 517         }
 518
 519         os->nr_samples = 0;
 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->nr_samples;
 580         os->last_sample = new;
 581
 582         if (!sample) {
 583                 list_add(&new->list, &os->samples);
 584                 os->max_timestamp = timestamp;
 585                 return;
 586         }
 587
 588         /*
 589          * last_sample might point to some random place in the list as it's
 590          * the last queued event. We expect that the new event is close to
 591          * this.
 592          */
 593         if (sample->timestamp <= timestamp) {
 594                 while (sample->timestamp <= timestamp) {
 595                         p = sample->list.next;
 596                         if (p == &os->samples) {
 597                                 list_add_tail(&new->list, &os->samples);
 598                                 os->max_timestamp = timestamp;
 599                                 return;
 600                         }
 601                         sample = list_entry(p, struct sample_queue, list);
 602                 }
 603                 list_add_tail(&new->list, &sample->list);
 604         } else {
 605                 while (sample->timestamp > timestamp) {
 606                         p = sample->list.prev;
 607                         if (p == &os->samples) {
 608                                 list_add(&new->list, &os->samples);
 609                                 return;
 610                         }
 611                         sample = list_entry(p, struct sample_queue, list);
 612                 }
 613                 list_add(&new->list, &sample->list);
 614         }
 615 }
 616
 617 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 618
 619 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 620                                     struct perf_sample *sample, u64 file_offset)
 621 {
 622         struct ordered_samples *os = &s->ordered_samples;
 623         struct list_head *sc = &os->sample_cache;
 624         u64 timestamp = sample->time;
 625         struct sample_queue *new;
 626
 627         if (!timestamp || timestamp == ~0ULL)
 628                 return -ETIME;
 629
 630         if (timestamp < s->ordered_samples.last_flush) {
 631                 printf("Warning: Timestamp below last timeslice flush\n");
 632                 return -EINVAL;
 633         }
 634
 635         if (!list_empty(sc)) {
 636                 new = list_entry(sc->next, struct sample_queue, list);
 637                 list_del(&new->list);
 638         } else if (os->sample_buffer) {
 639                 new = os->sample_buffer + os->sample_buffer_idx;
 640                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 641                         os->sample_buffer = NULL;
 642         } else {
 643                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 644                 if (!os->sample_buffer)
 645                         return -ENOMEM;
 646                 list_add(&os->sample_buffer->list, &os->to_free);
 647                 os->sample_buffer_idx = 2;
 648                 new = os->sample_buffer + 1;
 649         }
 650
 651         new->timestamp = timestamp;
 652         new->file_offset = file_offset;
 653         new->event = event;
 654
 655         __queue_event(new, s);
 656
 657         return 0;
 658 }
 659
 660 static void callchain__printf(struct perf_sample *sample)
 661 {
 662         unsigned int i;
 663
 664         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 665
 666         for (i = 0; i < sample->callchain->nr; i++)
 667                 printf("..... %2d: %016" PRIx64 "\n",
 668                        i, sample->callchain->ips[i]);
 669 }
 670
 671 static void perf_session__print_tstamp(struct perf_session *session,
 672                                        union perf_event *event,
 673                                        struct perf_sample *sample)
 674 {
 675         if (event->header.type != PERF_RECORD_SAMPLE &&
 676             !session->sample_id_all) {
 677                 fputs("-1 -1 ", stdout);
 678                 return;
 679         }
 680
 681         if ((session->sample_type & PERF_SAMPLE_CPU))
 682                 printf("%u ", sample->cpu);
 683
 684         if (session->sample_type & PERF_SAMPLE_TIME)
 685                 printf("%" PRIu64 " ", sample->time);
 686 }
 687
 688 static void dump_event(struct perf_session *session, union perf_event *event,
 689                        u64 file_offset, struct perf_sample *sample)
 690 {
 691         if (!dump_trace)
 692                 return;
 693
 694         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 695                file_offset, event->header.size, event->header.type);
 696
 697         trace_event(event);
 698
 699         if (sample)
 700                 perf_session__print_tstamp(session, event, sample);
 701
 702         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 703                event->header.size, perf_event__name(event->header.type));
 704 }
 705
 706 static void dump_sample(struct perf_session *session, union perf_event *event,
 707                         struct perf_sample *sample)
 708 {
 709         if (!dump_trace)
 710                 return;
 711
 712         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 713                event->header.misc, sample->pid, sample->tid, sample->ip,
 714                sample->period, sample->addr);
 715
 716         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 717                 callchain__printf(sample);
 718 }
 719
 720 static int perf_session_deliver_event(struct perf_session *session,
 721                                       union perf_event *event,
 722                                       struct perf_sample *sample,
 723                                       struct perf_event_ops *ops,
 724                                       u64 file_offset)
 725 {
 726         struct perf_evsel *evsel;
 727
 728         dump_event(session, event, file_offset, sample);
 729
 730         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 731         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
 732                 /*
 733                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
 734                  * because the tools right now may apply filters, discarding
 735                  * some of the samples. For consistency, in the future we
 736                  * should have something like nr_filtered_samples and remove
 737                  * the sample->period from total_sample_period, etc, KISS for
 738                  * now tho.
 739                  *
 740                  * Also testing against NULL allows us to handle files without
 741                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
 742                  * future probably it'll be a good idea to restrict event
 743                  * processing via perf_session to files with both set.
 744                  */
 745                 hists__inc_nr_events(&evsel->hists, event->header.type);
 746         }
 747
 748         switch (event->header.type) {
 749         case PERF_RECORD_SAMPLE:
 750                 dump_sample(session, event, sample);
 751                 if (evsel == NULL) {
 752                         ++session->hists.stats.nr_unknown_id;
 753                         return -1;
 754                 }
 755                 return ops->sample(event, sample, evsel, session);
 756         case PERF_RECORD_MMAP:
 757                 return ops->mmap(event, sample, session);
 758         case PERF_RECORD_COMM:
 759                 return ops->comm(event, sample, session);
 760         case PERF_RECORD_FORK:
 761                 return ops->fork(event, sample, session);
 762         case PERF_RECORD_EXIT:
 763                 return ops->exit(event, sample, session);
 764         case PERF_RECORD_LOST:
 765                 return ops->lost(event, sample, session);
 766         case PERF_RECORD_READ:
 767                 return ops->read(event, sample, session);
 768         case PERF_RECORD_THROTTLE:
 769                 return ops->throttle(event, sample, session);
 770         case PERF_RECORD_UNTHROTTLE:
 771                 return ops->unthrottle(event, sample, session);
 772         default:
 773                 ++session->hists.stats.nr_unknown_events;
 774                 return -1;
 775         }
 776 }
 777
 778 static int perf_session__preprocess_sample(struct perf_session *session,
 779                                            union perf_event *event, struct perf_sample *sample)
 780 {
 781         if (event->header.type != PERF_RECORD_SAMPLE ||
 782             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 783                 return 0;
 784
 785         if (!ip_callchain__valid(sample->callchain, event)) {
 786                 pr_debug("call-chain problem with event, skipping it.\n");
 787                 ++session->hists.stats.nr_invalid_chains;
 788                 session->hists.stats.total_invalid_chains += sample->period;
 789                 return -EINVAL;
 790         }
 791         return 0;
 792 }
 793
 794 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
 795                                             struct perf_event_ops *ops, u64 file_offset)
 796 {
 797         dump_event(session, event, file_offset, NULL);
 798
 799         /* These events are processed right away */
 800         switch (event->header.type) {
 801         case PERF_RECORD_HEADER_ATTR:
 802                 return ops->attr(event, session);
 803         case PERF_RECORD_HEADER_EVENT_TYPE:
 804                 return ops->event_type(event, session);
 805         case PERF_RECORD_HEADER_TRACING_DATA:
 806                 /* setup for reading amidst mmap */
 807                 lseek(session->fd, file_offset, SEEK_SET);
 808                 return ops->tracing_data(event, session);
 809         case PERF_RECORD_HEADER_BUILD_ID:
 810                 return ops->build_id(event, session);
 811         case PERF_RECORD_FINISHED_ROUND:
 812                 return ops->finished_round(event, session, ops);
 813         default:
 814                 return -EINVAL;
 815         }
 816 }
 817
 818 static int perf_session__process_event(struct perf_session *session,
 819                                        union perf_event *event,
 820                                        struct perf_event_ops *ops,
 821                                        u64 file_offset)
 822 {
 823         struct perf_sample sample;
 824         int ret;
 825
 826         if (session->header.needs_swap &&
 827             perf_event__swap_ops[event->header.type])
 828                 perf_event__swap_ops[event->header.type](event);
 829
 830         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 831                 return -EINVAL;
 832
 833         hists__inc_nr_events(&session->hists, event->header.type);
 834
 835         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 836                 return perf_session__process_user_event(session, event, ops, file_offset);
 837
 838         /*
 839          * For all kernel events we get the sample data
 840          */
 841         ret = perf_session__parse_sample(session, event, &sample);
 842         if (ret)
 843                 return ret;
 844
 845         /* Preprocess sample records - precheck callchains */
 846         if (perf_session__preprocess_sample(session, event, &sample))
 847                 return 0;
 848
 849         if (ops->ordered_samples) {
 850                 ret = perf_session_queue_event(session, event, &sample,
 851                                                file_offset);
 852                 if (ret != -ETIME)
 853                         return ret;
 854         }
 855
 856         return perf_session_deliver_event(session, event, &sample, ops,
 857                                           file_offset);
 858 }
 859
 860 void perf_event_header__bswap(struct perf_event_header *self)
 861 {
 862         self->type = bswap_32(self->type);
 863         self->misc = bswap_16(self->misc);
 864         self->size = bswap_16(self->size);
 865 }
 866
 867 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
 868 {
 869         return machine__findnew_thread(&session->host_machine, pid);
 870 }
 871
 872 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 873 {
 874         struct thread *thread = perf_session__findnew(self, 0);
 875
 876         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 877                 pr_err("problem inserting idle task.\n");
 878                 thread = NULL;
 879         }
 880
 881         return thread;
 882 }
 883
 884 static void perf_session__warn_about_errors(const struct perf_session *session,
 885                                             const struct perf_event_ops *ops)
 886 {
 887         if (ops->lost == perf_event__process_lost &&
 888             session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
 889                 ui__warning("Processed %d events and lost %d chunks!\n\n"
 890                             "Check IO/CPU overload!\n\n",
 891                             session->hists.stats.nr_events[0],
 892                             session->hists.stats.nr_events[PERF_RECORD_LOST]);
 893         }
 894
 895         if (session->hists.stats.nr_unknown_events != 0) {
 896                 ui__warning("Found %u unknown events!\n\n"
 897                             "Is this an older tool processing a perf.data "
 898                             "file generated by a more recent tool?\n\n"
 899                             "If that is not the case, consider "
 900                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 901                             session->hists.stats.nr_unknown_events);
 902         }
 903
 904         if (session->hists.stats.nr_unknown_id != 0) {
 905                 ui__warning("%u samples with id not present in the header\n",
 906                             session->hists.stats.nr_unknown_id);
 907         }
 908
 909         if (session->hists.stats.nr_invalid_chains != 0) {
 910                 ui__warning("Found invalid callchains!\n\n"
 911                             "%u out of %u events were discarded for this reason.\n\n"
 912                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 913                             session->hists.stats.nr_invalid_chains,
 914                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 915         }
 916 }
 917
 918 #define session_done()  (*(volatile int *)(&session_done))
 919 volatile int session_done;
 920
 921 static int __perf_session__process_pipe_events(struct perf_session *self,
 922                                                struct perf_event_ops *ops)
 923 {
 924         union perf_event event;
 925         uint32_t size;
 926         int skip = 0;
 927         u64 head;
 928         int err;
 929         void *p;
 930
 931         perf_event_ops__fill_defaults(ops);
 932
 933         head = 0;
 934 more:
 935         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 936         if (err <= 0) {
 937                 if (err == 0)
 938                         goto done;
 939
 940                 pr_err("failed to read event header\n");
 941                 goto out_err;
 942         }
 943
 944         if (self->header.needs_swap)
 945                 perf_event_header__bswap(&event.header);
 946
 947         size = event.header.size;
 948         if (size == 0)
 949                 size = 8;
 950
 951         p = &event;
 952         p += sizeof(struct perf_event_header);
 953
 954         if (size - sizeof(struct perf_event_header)) {
 955                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 956                 if (err <= 0) {
 957                         if (err == 0) {
 958                                 pr_err("unexpected end of event stream\n");
 959                                 goto done;
 960                         }
 961
 962                         pr_err("failed to read event data\n");
 963                         goto out_err;
 964                 }
 965         }
 966
 967         if (size == 0 ||
 968             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 969                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 970                             head, event.header.size, event.header.type);
 971                 /*
 972                  * assume we lost track of the stream, check alignment, and
 973                  * increment a single u64 in the hope to catch on again 'soon'.
 974                  */
 975                 if (unlikely(head & 7))
 976                         head &= ~7ULL;
 977
 978                 size = 8;
 979         }
 980
 981         head += size;
 982
 983         if (skip > 0)
 984                 head += skip;
 985
 986         if (!session_done())
 987                 goto more;
 988 done:
 989         err = 0;
 990 out_err:
 991         perf_session__warn_about_errors(self, ops);
 992         perf_session_free_sample_buffers(self);
 993         return err;
 994 }
 995
 996 static union perf_event *
 997 fetch_mmaped_event(struct perf_session *session,
 998                    u64 head, size_t mmap_size, char *buf)
 999 {
1000         union perf_event *event;
1001
1002         /*
1003          * Ensure we have enough space remaining to read
1004          * the size of the event in the headers.
1005          */
1006         if (head + sizeof(event->header) > mmap_size)
1007                 return NULL;
1008
1009         event = (union perf_event *)(buf + head);
1010
1011         if (session->header.needs_swap)
1012                 perf_event_header__bswap(&event->header);
1013
1014         if (head + event->header.size > mmap_size)
1015                 return NULL;
1016
1017         return event;
1018 }
1019
1020 int __perf_session__process_events(struct perf_session *session,
1021                                    u64 data_offset, u64 data_size,
1022                                    u64 file_size, struct perf_event_ops *ops)
1023 {
1024         u64 head, page_offset, file_offset, file_pos, progress_next;
1025         int err, mmap_prot, mmap_flags, map_idx = 0;
1026         size_t  page_size, mmap_size;
1027         char *buf, *mmaps[8];
1028         union perf_event *event;
1029         uint32_t size;
1030
1031         perf_event_ops__fill_defaults(ops);
1032
1033         page_size = sysconf(_SC_PAGESIZE);
1034
1035         page_offset = page_size * (data_offset / page_size);
1036         file_offset = page_offset;
1037         head = data_offset - page_offset;
1038
1039         if (data_offset + data_size < file_size)
1040                 file_size = data_offset + data_size;
1041
1042         progress_next = file_size / 16;
1043
1044         mmap_size = session->mmap_window;
1045         if (mmap_size > file_size)
1046                 mmap_size = file_size;
1047
1048         memset(mmaps, 0, sizeof(mmaps));
1049
1050         mmap_prot  = PROT_READ;
1051         mmap_flags = MAP_SHARED;
1052
1053         if (session->header.needs_swap) {
1054                 mmap_prot  |= PROT_WRITE;
1055                 mmap_flags = MAP_PRIVATE;
1056         }
1057 remap:
1058         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1059                    file_offset);
1060         if (buf == MAP_FAILED) {
1061                 pr_err("failed to mmap file\n");
1062                 err = -errno;
1063                 goto out_err;
1064         }
1065         mmaps[map_idx] = buf;
1066         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1067         file_pos = file_offset + head;
1068
1069 more:
1070         event = fetch_mmaped_event(session, head, mmap_size, buf);
1071         if (!event) {
1072                 if (mmaps[map_idx]) {
1073                         munmap(mmaps[map_idx], mmap_size);
1074                         mmaps[map_idx] = NULL;
1075                 }
1076
1077                 page_offset = page_size * (head / page_size);
1078                 file_offset += page_offset;
1079                 head -= page_offset;
1080                 goto remap;
1081         }
1082
1083         size = event->header.size;
1084
1085         if (size == 0 ||
1086             perf_session__process_event(session, event, ops, file_pos) < 0) {
1087                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1088                             file_offset + head, event->header.size,
1089                             event->header.type);
1090                 /*
1091                  * assume we lost track of the stream, check alignment, and
1092                  * increment a single u64 in the hope to catch on again 'soon'.
1093                  */
1094                 if (unlikely(head & 7))
1095                         head &= ~7ULL;
1096
1097                 size = 8;
1098         }
1099
1100         head += size;
1101         file_pos += size;
1102
1103         if (file_pos >= progress_next) {
1104                 progress_next += file_size / 16;
1105                 ui_progress__update(file_pos, file_size,
1106                                     "Processing events...");
1107         }
1108
1109         if (file_pos < file_size)
1110                 goto more;
1111
1112         err = 0;
1113         /* do the final flush for ordered samples */
1114         session->ordered_samples.next_flush = ULLONG_MAX;
1115         flush_sample_queue(session, ops);
1116 out_err:
1117         perf_session__warn_about_errors(session, ops);
1118         perf_session_free_sample_buffers(session);
1119         return err;
1120 }
1121
1122 int perf_session__process_events(struct perf_session *self,
1123                                  struct perf_event_ops *ops)
1124 {
1125         int err;
1126
1127         if (perf_session__register_idle_thread(self) == NULL)
1128                 return -ENOMEM;
1129
1130         if (!self->fd_pipe)
1131                 err = __perf_session__process_events(self,
1132                                                      self->header.data_offset,
1133                                                      self->header.data_size,
1134                                                      self->size, ops);
1135         else
1136                 err = __perf_session__process_pipe_events(self, ops);
1137
1138         return err;
1139 }
1140
1141 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1142 {
1143         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1144                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1145                 return false;
1146         }
1147
1148         return true;
1149 }
1150
1151 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1152                                              const char *symbol_name,
1153                                              u64 addr)
1154 {
1155         char *bracket;
1156         enum map_type i;
1157         struct ref_reloc_sym *ref;
1158
1159         ref = zalloc(sizeof(struct ref_reloc_sym));
1160         if (ref == NULL)
1161                 return -ENOMEM;
1162
1163         ref->name = strdup(symbol_name);
1164         if (ref->name == NULL) {
1165                 free(ref);
1166                 return -ENOMEM;
1167         }
1168
1169         bracket = strchr(ref->name, ']');
1170         if (bracket)
1171                 *bracket = '\0';
1172
1173         ref->addr = addr;
1174
1175         for (i = 0; i < MAP__NR_TYPES; ++i) {
1176                 struct kmap *kmap = map__kmap(maps[i]);
1177                 kmap->ref_reloc_sym = ref;
1178         }
1179
1180         return 0;
1181 }
1182
1183 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1184 {
1185         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1186                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1187                machines__fprintf_dsos(&self->machines, fp);
1188 }
1189
1190 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1191                                           bool with_hits)
1192 {
1193         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1194         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1195 }
1196
1197 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1198 {
1199         struct perf_evsel *pos;
1200         size_t ret = fprintf(fp, "Aggregated stats:\n");
1201
1202         ret += hists__fprintf_nr_events(&session->hists, fp);
1203
1204         list_for_each_entry(pos, &session->evlist->entries, node) {
1205                 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1206                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1207         }
1208
1209         return ret;
1210 }
1211
1212 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1213 {
1214         /*
1215          * FIXME: Here we have to actually print all the machines in this
1216          * session, not just the host...
1217          */
1218         return machine__fprintf(&session->host_machine, fp);
1219 }
1220
1221 void perf_session__remove_thread(struct perf_session *session,
1222                                  struct thread *th)
1223 {
1224         /*
1225          * FIXME: This one makes no sense, we need to remove the thread from
1226          * the machine it belongs to, perf_session can have many machines, so
1227          * doing it always on ->host_machine is wrong.  Fix when auditing all
1228          * the 'perf kvm' code.
1229          */
1230         machine__remove_thread(&session->host_machine, th);
1231 }
1232
1233 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1234                                               unsigned int type)
1235 {
1236         struct perf_evsel *pos;
1237
1238         list_for_each_entry(pos, &session->evlist->entries, node) {
1239                 if (pos->attr.type == type)
1240                         return pos;
1241         }
1242         return NULL;
1243 }
1244
1245 void perf_session__print_ip(union perf_event *event,
1246                             struct perf_sample *sample,
1247                             struct perf_session *session,
1248                             int print_sym, int print_dso)
1249 {
1250         struct addr_location al;
1251         const char *symname, *dsoname;
1252         struct callchain_cursor *cursor = &session->callchain_cursor;
1253         struct callchain_cursor_node *node;
1254
1255         if (perf_event__preprocess_sample(event, session, &al, sample,
1256                                           NULL) < 0) {
1257                 error("problem processing %d event, skipping it.\n",
1258                         event->header.type);
1259                 return;
1260         }
1261
1262         if (symbol_conf.use_callchain && sample->callchain) {
1263
1264                 if (perf_session__resolve_callchain(session, al.thread,
1265                                                 sample->callchain, NULL) != 0) {
1266                         if (verbose)
1267                                 error("Failed to resolve callchain. Skipping\n");
1268                         return;
1269                 }
1270                 callchain_cursor_commit(cursor);
1271
1272                 while (1) {
1273                         node = callchain_cursor_current(cursor);
1274                         if (!node)
1275                                 break;
1276
1277                         printf("\t%16" PRIx64, node->ip);
1278                         if (print_sym) {
1279                                 if (node->sym && node->sym->name)
1280                                         symname = node->sym->name;
1281                                 else
1282                                         symname = "";
1283
1284                                 printf(" %s", symname);
1285                         }
1286                         if (print_dso) {
1287                                 if (node->map && node->map->dso && node->map->dso->name)
1288                                         dsoname = node->map->dso->name;
1289                                 else
1290                                         dsoname = "";
1291
1292                                 printf(" (%s)", dsoname);
1293                         }
1294                         printf("\n");
1295
1296                         callchain_cursor_advance(cursor);
1297                 }
1298
1299         } else {
1300                 printf("%16" PRIx64, sample->ip);
1301                 if (print_sym) {
1302                         if (al.sym && al.sym->name)
1303                                 symname = al.sym->name;
1304                         else
1305                                 symname = "";
1306
1307                         printf(" %s", symname);
1308                 }
1309
1310                 if (print_dso) {
1311                         if (al.map && al.map->dso && al.map->dso->name)
1312                                 dsoname = al.map->dso->name;
1313                         else
1314                                 dsoname = "";
1315
1316                         printf(" (%s)", dsoname);
1317                 }
1318         }
1319 }
1320
1321 int perf_session__cpu_bitmap(struct perf_session *session,
1322                              const char *cpu_list, unsigned long *cpu_bitmap)
1323 {
1324         int i;
1325         struct cpu_map *map;
1326
1327         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1328                 struct perf_evsel *evsel;
1329
1330                 evsel = perf_session__find_first_evtype(session, i);
1331                 if (!evsel)
1332                         continue;
1333
1334                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1335                         pr_err("File does not contain CPU events. "
1336                                "Remove -c option to proceed.\n");
1337                         return -1;
1338                 }
1339         }
1340
1341         map = cpu_map__new(cpu_list);
1342
1343         for (i = 0; i < map->nr; i++) {
1344                 int cpu = map->map[i];
1345
1346                 if (cpu >= MAX_NR_CPUS) {
1347                         pr_err("Requested CPU %d too large. "
1348                                "Consider raising MAX_NR_CPUS\n", cpu);
1349                         return -1;
1350                 }
1351
1352                 set_bit(cpu, cpu_bitmap);
1353         }
1354
1355         return 0;
1356 }
1357
1358 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1359                                 bool full)
1360 {
1361         struct stat st;
1362         int ret;
1363
1364         if (session == NULL || fp == NULL)
1365                 return;
1366
1367         ret = fstat(session->fd, &st);
1368         if (ret == -1)
1369                 return;
1370
1371         fprintf(fp, "# ========\n");
1372         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1373         perf_header__fprintf_info(session, fp, full);
1374         fprintf(fp, "# ========\n#\n");
1375 }