tools/perf/util/session.c

   1 #include <linux/kernel.h>
   2 #include <traceevent/event-parse.h>
   3
   4 #include <byteswap.h>
   5 #include <unistd.h>
   6 #include <sys/types.h>
   7 #include <sys/mman.h>
   8
   9 #include "evlist.h"
  10 #include "evsel.h"
  11 #include "session.h"
  12 #include "tool.h"
  13 #include "sort.h"
  14 #include "util.h"
  15 #include "cpumap.h"
  16 #include "perf_regs.h"
  17 #include "vdso.h"
  18
  19 static int perf_session__open(struct perf_session *self)
  20 {
  21         struct perf_data_file *file = self->file;
  22
  23         if (perf_session__read_header(self) < 0) {
  24                 pr_err("incompatible file format (rerun with -v to learn more)");
  25                 return -1;
  26         }
  27
  28         if (perf_data_file__is_pipe(file))
  29                 return 0;
  30
  31         if (!perf_evlist__valid_sample_type(self->evlist)) {
  32                 pr_err("non matching sample_type");
  33                 return -1;
  34         }
  35
  36         if (!perf_evlist__valid_sample_id_all(self->evlist)) {
  37                 pr_err("non matching sample_id_all");
  38                 return -1;
  39         }
  40
  41         if (!perf_evlist__valid_read_format(self->evlist)) {
  42                 pr_err("non matching read_format");
  43                 return -1;
  44         }
  45
  46         return 0;
  47 }
  48
  49 void perf_session__set_id_hdr_size(struct perf_session *session)
  50 {
  51         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
  52
  53         machines__set_id_hdr_size(&session->machines, id_hdr_size);
  54 }
  55
  56 int perf_session__create_kernel_maps(struct perf_session *self)
  57 {
  58         int ret = machine__create_kernel_maps(&self->machines.host);
  59
  60         if (ret >= 0)
  61                 ret = machines__create_guest_kernel_maps(&self->machines);
  62         return ret;
  63 }
  64
  65 static void perf_session__destroy_kernel_maps(struct perf_session *self)
  66 {
  67         machines__destroy_kernel_maps(&self->machines);
  68 }
  69
  70 struct perf_session *perf_session__new(struct perf_data_file *file,
  71                                        bool repipe, struct perf_tool *tool)
  72 {
  73         struct perf_session *self;
  74
  75         self = zalloc(sizeof(*self));
  76         if (!self)
  77                 goto out;
  78
  79         self->repipe = repipe;
  80         INIT_LIST_HEAD(&self->ordered_samples.samples);
  81         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
  82         INIT_LIST_HEAD(&self->ordered_samples.to_free);
  83         machines__init(&self->machines);
  84
  85         if (file) {
  86                 if (perf_data_file__open(file))
  87                         goto out_delete;
  88
  89                 self->file = file;
  90
  91                 if (perf_data_file__is_read(file)) {
  92                         if (perf_session__open(self) < 0)
  93                                 goto out_close;
  94
  95                         perf_session__set_id_hdr_size(self);
  96                 }
  97         }
  98
  99         if (!file || perf_data_file__is_write(file)) {
 100                 /*
 101                  * In O_RDONLY mode this will be performed when reading the
 102                  * kernel MMAP event, in perf_event__process_mmap().
 103                  */
 104                 if (perf_session__create_kernel_maps(self) < 0)
 105                         goto out_delete;
 106         }
 107
 108         if (tool && tool->ordering_requires_timestamps &&
 109             tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
 110                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 111                 tool->ordered_samples = false;
 112         }
 113
 114         return self;
 115
 116  out_close:
 117         perf_data_file__close(file);
 118  out_delete:
 119         perf_session__delete(self);
 120  out:
 121         return NULL;
 122 }
 123
 124 static void perf_session__delete_dead_threads(struct perf_session *session)
 125 {
 126         machine__delete_dead_threads(&session->machines.host);
 127 }
 128
 129 static void perf_session__delete_threads(struct perf_session *session)
 130 {
 131         machine__delete_threads(&session->machines.host);
 132 }
 133
 134 static void perf_session_env__delete(struct perf_session_env *env)
 135 {
 136         free(env->hostname);
 137         free(env->os_release);
 138         free(env->version);
 139         free(env->arch);
 140         free(env->cpu_desc);
 141         free(env->cpuid);
 142
 143         free(env->cmdline);
 144         free(env->sibling_cores);
 145         free(env->sibling_threads);
 146         free(env->numa_nodes);
 147         free(env->pmu_mappings);
 148 }
 149
 150 void perf_session__delete(struct perf_session *self)
 151 {
 152         perf_session__destroy_kernel_maps(self);
 153         perf_session__delete_dead_threads(self);
 154         perf_session__delete_threads(self);
 155         perf_session_env__delete(&self->header.env);
 156         machines__exit(&self->machines);
 157         if (self->file)
 158                 perf_data_file__close(self->file);
 159         free(self);
 160         vdso__exit();
 161 }
 162
 163 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
 164                                                  __maybe_unused,
 165                                                  union perf_event *event
 166                                                  __maybe_unused,
 167                                                  struct perf_session *session
 168                                                 __maybe_unused)
 169 {
 170         dump_printf(": unhandled!\n");
 171         return 0;
 172 }
 173
 174 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
 175                                          union perf_event *event __maybe_unused,
 176                                          struct perf_evlist **pevlist
 177                                          __maybe_unused)
 178 {
 179         dump_printf(": unhandled!\n");
 180         return 0;
 181 }
 182
 183 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
 184                                      union perf_event *event __maybe_unused,
 185                                      struct perf_sample *sample __maybe_unused,
 186                                      struct perf_evsel *evsel __maybe_unused,
 187                                      struct machine *machine __maybe_unused)
 188 {
 189         dump_printf(": unhandled!\n");
 190         return 0;
 191 }
 192
 193 static int process_event_stub(struct perf_tool *tool __maybe_unused,
 194                               union perf_event *event __maybe_unused,
 195                               struct perf_sample *sample __maybe_unused,
 196                               struct machine *machine __maybe_unused)
 197 {
 198         dump_printf(": unhandled!\n");
 199         return 0;
 200 }
 201
 202 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
 203                                        union perf_event *event __maybe_unused,
 204                                        struct perf_session *perf_session
 205                                        __maybe_unused)
 206 {
 207         dump_printf(": unhandled!\n");
 208         return 0;
 209 }
 210
 211 static int process_finished_round(struct perf_tool *tool,
 212                                   union perf_event *event,
 213                                   struct perf_session *session);
 214
 215 void perf_tool__fill_defaults(struct perf_tool *tool)
 216 {
 217         if (tool->sample == NULL)
 218                 tool->sample = process_event_sample_stub;
 219         if (tool->mmap == NULL)
 220                 tool->mmap = process_event_stub;
 221         if (tool->mmap2 == NULL)
 222                 tool->mmap2 = process_event_stub;
 223         if (tool->comm == NULL)
 224                 tool->comm = process_event_stub;
 225         if (tool->fork == NULL)
 226                 tool->fork = process_event_stub;
 227         if (tool->exit == NULL)
 228                 tool->exit = process_event_stub;
 229         if (tool->lost == NULL)
 230                 tool->lost = perf_event__process_lost;
 231         if (tool->read == NULL)
 232                 tool->read = process_event_sample_stub;
 233         if (tool->throttle == NULL)
 234                 tool->throttle = process_event_stub;
 235         if (tool->unthrottle == NULL)
 236                 tool->unthrottle = process_event_stub;
 237         if (tool->attr == NULL)
 238                 tool->attr = process_event_synth_attr_stub;
 239         if (tool->tracing_data == NULL)
 240                 tool->tracing_data = process_event_synth_tracing_data_stub;
 241         if (tool->build_id == NULL)
 242                 tool->build_id = process_finished_round_stub;
 243         if (tool->finished_round == NULL) {
 244                 if (tool->ordered_samples)
 245                         tool->finished_round = process_finished_round;
 246                 else
 247                         tool->finished_round = process_finished_round_stub;
 248         }
 249 }
 250
 251 void mem_bswap_32(void *src, int byte_size)
 252 {
 253         u32 *m = src;
 254         while (byte_size > 0) {
 255                 *m = bswap_32(*m);
 256                 byte_size -= sizeof(u32);
 257                 ++m;
 258         }
 259 }
 260
 261 void mem_bswap_64(void *src, int byte_size)
 262 {
 263         u64 *m = src;
 264
 265         while (byte_size > 0) {
 266                 *m = bswap_64(*m);
 267                 byte_size -= sizeof(u64);
 268                 ++m;
 269         }
 270 }
 271
 272 static void swap_sample_id_all(union perf_event *event, void *data)
 273 {
 274         void *end = (void *) event + event->header.size;
 275         int size = end - data;
 276
 277         BUG_ON(size % sizeof(u64));
 278         mem_bswap_64(data, size);
 279 }
 280
 281 static void perf_event__all64_swap(union perf_event *event,
 282                                    bool sample_id_all __maybe_unused)
 283 {
 284         struct perf_event_header *hdr = &event->header;
 285         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 286 }
 287
 288 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 289 {
 290         event->comm.pid = bswap_32(event->comm.pid);
 291         event->comm.tid = bswap_32(event->comm.tid);
 292
 293         if (sample_id_all) {
 294                 void *data = &event->comm.comm;
 295
 296                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 297                 swap_sample_id_all(event, data);
 298         }
 299 }
 300
 301 static void perf_event__mmap_swap(union perf_event *event,
 302                                   bool sample_id_all)
 303 {
 304         event->mmap.pid   = bswap_32(event->mmap.pid);
 305         event->mmap.tid   = bswap_32(event->mmap.tid);
 306         event->mmap.start = bswap_64(event->mmap.start);
 307         event->mmap.len   = bswap_64(event->mmap.len);
 308         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 309
 310         if (sample_id_all) {
 311                 void *data = &event->mmap.filename;
 312
 313                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 314                 swap_sample_id_all(event, data);
 315         }
 316 }
 317
 318 static void perf_event__mmap2_swap(union perf_event *event,
 319                                   bool sample_id_all)
 320 {
 321         event->mmap2.pid   = bswap_32(event->mmap2.pid);
 322         event->mmap2.tid   = bswap_32(event->mmap2.tid);
 323         event->mmap2.start = bswap_64(event->mmap2.start);
 324         event->mmap2.len   = bswap_64(event->mmap2.len);
 325         event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
 326         event->mmap2.maj   = bswap_32(event->mmap2.maj);
 327         event->mmap2.min   = bswap_32(event->mmap2.min);
 328         event->mmap2.ino   = bswap_64(event->mmap2.ino);
 329
 330         if (sample_id_all) {
 331                 void *data = &event->mmap2.filename;
 332
 333                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 334                 swap_sample_id_all(event, data);
 335         }
 336 }
 337 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 338 {
 339         event->fork.pid  = bswap_32(event->fork.pid);
 340         event->fork.tid  = bswap_32(event->fork.tid);
 341         event->fork.ppid = bswap_32(event->fork.ppid);
 342         event->fork.ptid = bswap_32(event->fork.ptid);
 343         event->fork.time = bswap_64(event->fork.time);
 344
 345         if (sample_id_all)
 346                 swap_sample_id_all(event, &event->fork + 1);
 347 }
 348
 349 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 350 {
 351         event->read.pid          = bswap_32(event->read.pid);
 352         event->read.tid          = bswap_32(event->read.tid);
 353         event->read.value        = bswap_64(event->read.value);
 354         event->read.time_enabled = bswap_64(event->read.time_enabled);
 355         event->read.time_running = bswap_64(event->read.time_running);
 356         event->read.id           = bswap_64(event->read.id);
 357
 358         if (sample_id_all)
 359                 swap_sample_id_all(event, &event->read + 1);
 360 }
 361
 362 static void perf_event__throttle_swap(union perf_event *event,
 363                                       bool sample_id_all)
 364 {
 365         event->throttle.time      = bswap_64(event->throttle.time);
 366         event->throttle.id        = bswap_64(event->throttle.id);
 367         event->throttle.stream_id = bswap_64(event->throttle.stream_id);
 368
 369         if (sample_id_all)
 370                 swap_sample_id_all(event, &event->throttle + 1);
 371 }
 372
 373 static u8 revbyte(u8 b)
 374 {
 375         int rev = (b >> 4) | ((b & 0xf) << 4);
 376         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 377         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 378         return (u8) rev;
 379 }
 380
 381 /*
 382  * XXX this is hack in attempt to carry flags bitfield
 383  * throught endian village. ABI says:
 384  *
 385  * Bit-fields are allocated from right to left (least to most significant)
 386  * on little-endian implementations and from left to right (most to least
 387  * significant) on big-endian implementations.
 388  *
 389  * The above seems to be byte specific, so we need to reverse each
 390  * byte of the bitfield. 'Internet' also says this might be implementation
 391  * specific and we probably need proper fix and carry perf_event_attr
 392  * bitfield flags in separate data file FEAT_ section. Thought this seems
 393  * to work for now.
 394  */
 395 static void swap_bitfield(u8 *p, unsigned len)
 396 {
 397         unsigned i;
 398
 399         for (i = 0; i < len; i++) {
 400                 *p = revbyte(*p);
 401                 p++;
 402         }
 403 }
 404
 405 /* exported for swapping attributes in file header */
 406 void perf_event__attr_swap(struct perf_event_attr *attr)
 407 {
 408         attr->type              = bswap_32(attr->type);
 409         attr->size              = bswap_32(attr->size);
 410         attr->config            = bswap_64(attr->config);
 411         attr->sample_period     = bswap_64(attr->sample_period);
 412         attr->sample_type       = bswap_64(attr->sample_type);
 413         attr->read_format       = bswap_64(attr->read_format);
 414         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 415         attr->bp_type           = bswap_32(attr->bp_type);
 416         attr->bp_addr           = bswap_64(attr->bp_addr);
 417         attr->bp_len            = bswap_64(attr->bp_len);
 418         attr->branch_sample_type = bswap_64(attr->branch_sample_type);
 419         attr->sample_regs_user   = bswap_64(attr->sample_regs_user);
 420         attr->sample_stack_user  = bswap_32(attr->sample_stack_user);
 421
 422         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 423 }
 424
 425 static void perf_event__hdr_attr_swap(union perf_event *event,
 426                                       bool sample_id_all __maybe_unused)
 427 {
 428         size_t size;
 429
 430         perf_event__attr_swap(&event->attr.attr);
 431
 432         size = event->header.size;
 433         size -= (void *)&event->attr.id - (void *)event;
 434         mem_bswap_64(event->attr.id, size);
 435 }
 436
 437 static void perf_event__event_type_swap(union perf_event *event,
 438                                         bool sample_id_all __maybe_unused)
 439 {
 440         event->event_type.event_type.event_id =
 441                 bswap_64(event->event_type.event_type.event_id);
 442 }
 443
 444 static void perf_event__tracing_data_swap(union perf_event *event,
 445                                           bool sample_id_all __maybe_unused)
 446 {
 447         event->tracing_data.size = bswap_32(event->tracing_data.size);
 448 }
 449
 450 typedef void (*perf_event__swap_op)(union perf_event *event,
 451                                     bool sample_id_all);
 452
 453 static perf_event__swap_op perf_event__swap_ops[] = {
 454         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 455         [PERF_RECORD_MMAP2]               = perf_event__mmap2_swap,
 456         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 457         [PERF_RECORD_FORK]                = perf_event__task_swap,
 458         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 459         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 460         [PERF_RECORD_READ]                = perf_event__read_swap,
 461         [PERF_RECORD_THROTTLE]            = perf_event__throttle_swap,
 462         [PERF_RECORD_UNTHROTTLE]          = perf_event__throttle_swap,
 463         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 464         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 465         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 466         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 467         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 468         [PERF_RECORD_HEADER_MAX]          = NULL,
 469 };
 470
 471 struct sample_queue {
 472         u64                     timestamp;
 473         u64                     file_offset;
 474         union perf_event        *event;
 475         struct list_head        list;
 476 };
 477
 478 static void perf_session_free_sample_buffers(struct perf_session *session)
 479 {
 480         struct ordered_samples *os = &session->ordered_samples;
 481
 482         while (!list_empty(&os->to_free)) {
 483                 struct sample_queue *sq;
 484
 485                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 486                 list_del(&sq->list);
 487                 free(sq);
 488         }
 489 }
 490
 491 static int perf_session_deliver_event(struct perf_session *session,
 492                                       union perf_event *event,
 493                                       struct perf_sample *sample,
 494                                       struct perf_tool *tool,
 495                                       u64 file_offset);
 496
 497 static int flush_sample_queue(struct perf_session *s,
 498                        struct perf_tool *tool)
 499 {
 500         struct ordered_samples *os = &s->ordered_samples;
 501         struct list_head *head = &os->samples;
 502         struct sample_queue *tmp, *iter;
 503         struct perf_sample sample;
 504         u64 limit = os->next_flush;
 505         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 506         bool show_progress = limit == ULLONG_MAX;
 507         struct ui_progress prog;
 508         int ret;
 509
 510         if (!tool->ordered_samples || !limit)
 511                 return 0;
 512
 513         if (show_progress)
 514                 ui_progress__init(&prog, os->nr_samples, "Processing time ordered events...");
 515
 516         list_for_each_entry_safe(iter, tmp, head, list) {
 517                 if (session_done())
 518                         return 0;
 519
 520                 if (iter->timestamp > limit)
 521                         break;
 522
 523                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
 524                 if (ret)
 525                         pr_err("Can't parse sample, err = %d\n", ret);
 526                 else {
 527                         ret = perf_session_deliver_event(s, iter->event, &sample, tool,
 528                                                          iter->file_offset);
 529                         if (ret)
 530                                 return ret;
 531                 }
 532
 533                 os->last_flush = iter->timestamp;
 534                 list_del(&iter->list);
 535                 list_add(&iter->list, &os->sample_cache);
 536
 537                 if (show_progress)
 538                         ui_progress__update(&prog, 1);
 539         }
 540
 541         if (list_empty(head)) {
 542                 os->last_sample = NULL;
 543         } else if (last_ts <= limit) {
 544                 os->last_sample =
 545                         list_entry(head->prev, struct sample_queue, list);
 546         }
 547
 548         os->nr_samples = 0;
 549
 550         return 0;
 551 }
 552
 553 /*
 554  * When perf record finishes a pass on every buffers, it records this pseudo
 555  * event.
 556  * We record the max timestamp t found in the pass n.
 557  * Assuming these timestamps are monotonic across cpus, we know that if
 558  * a buffer still has events with timestamps below t, they will be all
 559  * available and then read in the pass n + 1.
 560  * Hence when we start to read the pass n + 2, we can safely flush every
 561  * events with timestamps below t.
 562  *
 563  *    ============ PASS n =================
 564  *       CPU 0         |   CPU 1
 565  *                     |
 566  *    cnt1 timestamps  |   cnt2 timestamps
 567  *          1          |         2
 568  *          2          |         3
 569  *          -          |         4  <--- max recorded
 570  *
 571  *    ============ PASS n + 1 ==============
 572  *       CPU 0         |   CPU 1
 573  *                     |
 574  *    cnt1 timestamps  |   cnt2 timestamps
 575  *          3          |         5
 576  *          4          |         6
 577  *          5          |         7 <---- max recorded
 578  *
 579  *      Flush every events below timestamp 4
 580  *
 581  *    ============ PASS n + 2 ==============
 582  *       CPU 0         |   CPU 1
 583  *                     |
 584  *    cnt1 timestamps  |   cnt2 timestamps
 585  *          6          |         8
 586  *          7          |         9
 587  *          -          |         10
 588  *
 589  *      Flush every events below timestamp 7
 590  *      etc...
 591  */
 592 static int process_finished_round(struct perf_tool *tool,
 593                                   union perf_event *event __maybe_unused,
 594                                   struct perf_session *session)
 595 {
 596         int ret = flush_sample_queue(session, tool);
 597         if (!ret)
 598                 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 599
 600         return ret;
 601 }
 602
 603 /* The queue is ordered by time */
 604 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 605 {
 606         struct ordered_samples *os = &s->ordered_samples;
 607         struct sample_queue *sample = os->last_sample;
 608         u64 timestamp = new->timestamp;
 609         struct list_head *p;
 610
 611         ++os->nr_samples;
 612         os->last_sample = new;
 613
 614         if (!sample) {
 615                 list_add(&new->list, &os->samples);
 616                 os->max_timestamp = timestamp;
 617                 return;
 618         }
 619
 620         /*
 621          * last_sample might point to some random place in the list as it's
 622          * the last queued event. We expect that the new event is close to
 623          * this.
 624          */
 625         if (sample->timestamp <= timestamp) {
 626                 while (sample->timestamp <= timestamp) {
 627                         p = sample->list.next;
 628                         if (p == &os->samples) {
 629                                 list_add_tail(&new->list, &os->samples);
 630                                 os->max_timestamp = timestamp;
 631                                 return;
 632                         }
 633                         sample = list_entry(p, struct sample_queue, list);
 634                 }
 635                 list_add_tail(&new->list, &sample->list);
 636         } else {
 637                 while (sample->timestamp > timestamp) {
 638                         p = sample->list.prev;
 639                         if (p == &os->samples) {
 640                                 list_add(&new->list, &os->samples);
 641                                 return;
 642                         }
 643                         sample = list_entry(p, struct sample_queue, list);
 644                 }
 645                 list_add(&new->list, &sample->list);
 646         }
 647 }
 648
 649 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 650
 651 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 652                                     struct perf_sample *sample, u64 file_offset)
 653 {
 654         struct ordered_samples *os = &s->ordered_samples;
 655         struct list_head *sc = &os->sample_cache;
 656         u64 timestamp = sample->time;
 657         struct sample_queue *new;
 658
 659         if (!timestamp || timestamp == ~0ULL)
 660                 return -ETIME;
 661
 662         if (timestamp < s->ordered_samples.last_flush) {
 663                 printf("Warning: Timestamp below last timeslice flush\n");
 664                 return -EINVAL;
 665         }
 666
 667         if (!list_empty(sc)) {
 668                 new = list_entry(sc->next, struct sample_queue, list);
 669                 list_del(&new->list);
 670         } else if (os->sample_buffer) {
 671                 new = os->sample_buffer + os->sample_buffer_idx;
 672                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 673                         os->sample_buffer = NULL;
 674         } else {
 675                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 676                 if (!os->sample_buffer)
 677                         return -ENOMEM;
 678                 list_add(&os->sample_buffer->list, &os->to_free);
 679                 os->sample_buffer_idx = 2;
 680                 new = os->sample_buffer + 1;
 681         }
 682
 683         new->timestamp = timestamp;
 684         new->file_offset = file_offset;
 685         new->event = event;
 686
 687         __queue_event(new, s);
 688
 689         return 0;
 690 }
 691
 692 static void callchain__printf(struct perf_sample *sample)
 693 {
 694         unsigned int i;
 695
 696         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 697
 698         for (i = 0; i < sample->callchain->nr; i++)
 699                 printf("..... %2d: %016" PRIx64 "\n",
 700                        i, sample->callchain->ips[i]);
 701 }
 702
 703 static void branch_stack__printf(struct perf_sample *sample)
 704 {
 705         uint64_t i;
 706
 707         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 708
 709         for (i = 0; i < sample->branch_stack->nr; i++)
 710                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 711                         i, sample->branch_stack->entries[i].from,
 712                         sample->branch_stack->entries[i].to);
 713 }
 714
 715 static void regs_dump__printf(u64 mask, u64 *regs)
 716 {
 717         unsigned rid, i = 0;
 718
 719         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 720                 u64 val = regs[i++];
 721
 722                 printf(".... %-5s 0x%" PRIx64 "\n",
 723                        perf_reg_name(rid), val);
 724         }
 725 }
 726
 727 static void regs_user__printf(struct perf_sample *sample, u64 mask)
 728 {
 729         struct regs_dump *user_regs = &sample->user_regs;
 730
 731         if (user_regs->regs) {
 732                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
 733                 regs_dump__printf(mask, user_regs->regs);
 734         }
 735 }
 736
 737 static void stack_user__printf(struct stack_dump *dump)
 738 {
 739         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 740                dump->size, dump->offset);
 741 }
 742
 743 static void perf_session__print_tstamp(struct perf_session *session,
 744                                        union perf_event *event,
 745                                        struct perf_sample *sample)
 746 {
 747         u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
 748
 749         if (event->header.type != PERF_RECORD_SAMPLE &&
 750             !perf_evlist__sample_id_all(session->evlist)) {
 751                 fputs("-1 -1 ", stdout);
 752                 return;
 753         }
 754
 755         if ((sample_type & PERF_SAMPLE_CPU))
 756                 printf("%u ", sample->cpu);
 757
 758         if (sample_type & PERF_SAMPLE_TIME)
 759                 printf("%" PRIu64 " ", sample->time);
 760 }
 761
 762 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
 763 {
 764         printf("... sample_read:\n");
 765
 766         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 767                 printf("...... time enabled %016" PRIx64 "\n",
 768                        sample->read.time_enabled);
 769
 770         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 771                 printf("...... time running %016" PRIx64 "\n",
 772                        sample->read.time_running);
 773
 774         if (read_format & PERF_FORMAT_GROUP) {
 775                 u64 i;
 776
 777                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
 778
 779                 for (i = 0; i < sample->read.group.nr; i++) {
 780                         struct sample_read_value *value;
 781
 782                         value = &sample->read.group.values[i];
 783                         printf("..... id %016" PRIx64
 784                                ", value %016" PRIx64 "\n",
 785                                value->id, value->value);
 786                 }
 787         } else
 788                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
 789                         sample->read.one.id, sample->read.one.value);
 790 }
 791
 792 static void dump_event(struct perf_session *session, union perf_event *event,
 793                        u64 file_offset, struct perf_sample *sample)
 794 {
 795         if (!dump_trace)
 796                 return;
 797
 798         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 799                file_offset, event->header.size, event->header.type);
 800
 801         trace_event(event);
 802
 803         if (sample)
 804                 perf_session__print_tstamp(session, event, sample);
 805
 806         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 807                event->header.size, perf_event__name(event->header.type));
 808 }
 809
 810 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 811                         struct perf_sample *sample)
 812 {
 813         u64 sample_type;
 814
 815         if (!dump_trace)
 816                 return;
 817
 818         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 819                event->header.misc, sample->pid, sample->tid, sample->ip,
 820                sample->period, sample->addr);
 821
 822         sample_type = evsel->attr.sample_type;
 823
 824         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 825                 callchain__printf(sample);
 826
 827         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
 828                 branch_stack__printf(sample);
 829
 830         if (sample_type & PERF_SAMPLE_REGS_USER)
 831                 regs_user__printf(sample, evsel->attr.sample_regs_user);
 832
 833         if (sample_type & PERF_SAMPLE_STACK_USER)
 834                 stack_user__printf(&sample->user_stack);
 835
 836         if (sample_type & PERF_SAMPLE_WEIGHT)
 837                 printf("... weight: %" PRIu64 "\n", sample->weight);
 838
 839         if (sample_type & PERF_SAMPLE_DATA_SRC)
 840                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
 841
 842         if (sample_type & PERF_SAMPLE_TRANSACTION)
 843                 printf("... transaction: %" PRIx64 "\n", sample->transaction);
 844
 845         if (sample_type & PERF_SAMPLE_READ)
 846                 sample_read__printf(sample, evsel->attr.read_format);
 847 }
 848
 849 static struct machine *
 850         perf_session__find_machine_for_cpumode(struct perf_session *session,
 851                                                union perf_event *event,
 852                                                struct perf_sample *sample)
 853 {
 854         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 855
 856         if (perf_guest &&
 857             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
 858              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
 859                 u32 pid;
 860
 861                 if (event->header.type == PERF_RECORD_MMAP
 862                     || event->header.type == PERF_RECORD_MMAP2)
 863                         pid = event->mmap.pid;
 864                 else
 865                         pid = sample->pid;
 866
 867                 return perf_session__findnew_machine(session, pid);
 868         }
 869
 870         return &session->machines.host;
 871 }
 872
 873 static int deliver_sample_value(struct perf_session *session,
 874                                 struct perf_tool *tool,
 875                                 union perf_event *event,
 876                                 struct perf_sample *sample,
 877                                 struct sample_read_value *v,
 878                                 struct machine *machine)
 879 {
 880         struct perf_sample_id *sid;
 881
 882         sid = perf_evlist__id2sid(session->evlist, v->id);
 883         if (sid) {
 884                 sample->id     = v->id;
 885                 sample->period = v->value - sid->period;
 886                 sid->period    = v->value;
 887         }
 888
 889         if (!sid || sid->evsel == NULL) {
 890                 ++session->stats.nr_unknown_id;
 891                 return 0;
 892         }
 893
 894         return tool->sample(tool, event, sample, sid->evsel, machine);
 895 }
 896
 897 static int deliver_sample_group(struct perf_session *session,
 898                                 struct perf_tool *tool,
 899                                 union  perf_event *event,
 900                                 struct perf_sample *sample,
 901                                 struct machine *machine)
 902 {
 903         int ret = -EINVAL;
 904         u64 i;
 905
 906         for (i = 0; i < sample->read.group.nr; i++) {
 907                 ret = deliver_sample_value(session, tool, event, sample,
 908                                            &sample->read.group.values[i],
 909                                            machine);
 910                 if (ret)
 911                         break;
 912         }
 913
 914         return ret;
 915 }
 916
 917 static int
 918 perf_session__deliver_sample(struct perf_session *session,
 919                              struct perf_tool *tool,
 920                              union  perf_event *event,
 921                              struct perf_sample *sample,
 922                              struct perf_evsel *evsel,
 923                              struct machine *machine)
 924 {
 925         /* We know evsel != NULL. */
 926         u64 sample_type = evsel->attr.sample_type;
 927         u64 read_format = evsel->attr.read_format;
 928
 929         /* Standard sample delievery. */
 930         if (!(sample_type & PERF_SAMPLE_READ))
 931                 return tool->sample(tool, event, sample, evsel, machine);
 932
 933         /* For PERF_SAMPLE_READ we have either single or group mode. */
 934         if (read_format & PERF_FORMAT_GROUP)
 935                 return deliver_sample_group(session, tool, event, sample,
 936                                             machine);
 937         else
 938                 return deliver_sample_value(session, tool, event, sample,
 939                                             &sample->read.one, machine);
 940 }
 941
 942 static int perf_session_deliver_event(struct perf_session *session,
 943                                       union perf_event *event,
 944                                       struct perf_sample *sample,
 945                                       struct perf_tool *tool,
 946                                       u64 file_offset)
 947 {
 948         struct perf_evsel *evsel;
 949         struct machine *machine;
 950
 951         dump_event(session, event, file_offset, sample);
 952
 953         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 954         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
 955                 /*
 956                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
 957                  * because the tools right now may apply filters, discarding
 958                  * some of the samples. For consistency, in the future we
 959                  * should have something like nr_filtered_samples and remove
 960                  * the sample->period from total_sample_period, etc, KISS for
 961                  * now tho.
 962                  *
 963                  * Also testing against NULL allows us to handle files without
 964                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
 965                  * future probably it'll be a good idea to restrict event
 966                  * processing via perf_session to files with both set.
 967                  */
 968                 hists__inc_nr_events(&evsel->hists, event->header.type);
 969         }
 970
 971         machine = perf_session__find_machine_for_cpumode(session, event,
 972                                                          sample);
 973
 974         switch (event->header.type) {
 975         case PERF_RECORD_SAMPLE:
 976                 dump_sample(evsel, event, sample);
 977                 if (evsel == NULL) {
 978                         ++session->stats.nr_unknown_id;
 979                         return 0;
 980                 }
 981                 if (machine == NULL) {
 982                         ++session->stats.nr_unprocessable_samples;
 983                         return 0;
 984                 }
 985                 return perf_session__deliver_sample(session, tool, event,
 986                                                     sample, evsel, machine);
 987         case PERF_RECORD_MMAP:
 988                 return tool->mmap(tool, event, sample, machine);
 989         case PERF_RECORD_MMAP2:
 990                 return tool->mmap2(tool, event, sample, machine);
 991         case PERF_RECORD_COMM:
 992                 return tool->comm(tool, event, sample, machine);
 993         case PERF_RECORD_FORK:
 994                 return tool->fork(tool, event, sample, machine);
 995         case PERF_RECORD_EXIT:
 996                 return tool->exit(tool, event, sample, machine);
 997         case PERF_RECORD_LOST:
 998                 if (tool->lost == perf_event__process_lost)
 999                         session->stats.total_lost += event->lost.lost;
1000                 return tool->lost(tool, event, sample, machine);
1001         case PERF_RECORD_READ:
1002                 return tool->read(tool, event, sample, evsel, machine);
1003         case PERF_RECORD_THROTTLE:
1004                 return tool->throttle(tool, event, sample, machine);
1005         case PERF_RECORD_UNTHROTTLE:
1006                 return tool->unthrottle(tool, event, sample, machine);
1007         default:
1008                 ++session->stats.nr_unknown_events;
1009                 return -1;
1010         }
1011 }
1012
1013 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1014                                             struct perf_tool *tool, u64 file_offset)
1015 {
1016         int fd = perf_data_file__fd(session->file);
1017         int err;
1018
1019         dump_event(session, event, file_offset, NULL);
1020
1021         /* These events are processed right away */
1022         switch (event->header.type) {
1023         case PERF_RECORD_HEADER_ATTR:
1024                 err = tool->attr(tool, event, &session->evlist);
1025                 if (err == 0)
1026                         perf_session__set_id_hdr_size(session);
1027                 return err;
1028         case PERF_RECORD_HEADER_TRACING_DATA:
1029                 /* setup for reading amidst mmap */
1030                 lseek(fd, file_offset, SEEK_SET);
1031                 return tool->tracing_data(tool, event, session);
1032         case PERF_RECORD_HEADER_BUILD_ID:
1033                 return tool->build_id(tool, event, session);
1034         case PERF_RECORD_FINISHED_ROUND:
1035                 return tool->finished_round(tool, event, session);
1036         default:
1037                 return -EINVAL;
1038         }
1039 }
1040
1041 static void event_swap(union perf_event *event, bool sample_id_all)
1042 {
1043         perf_event__swap_op swap;
1044
1045         swap = perf_event__swap_ops[event->header.type];
1046         if (swap)
1047                 swap(event, sample_id_all);
1048 }
1049
1050 static int perf_session__process_event(struct perf_session *session,
1051                                        union perf_event *event,
1052                                        struct perf_tool *tool,
1053                                        u64 file_offset)
1054 {
1055         struct perf_sample sample;
1056         int ret;
1057
1058         if (session->header.needs_swap)
1059                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1060
1061         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1062                 return -EINVAL;
1063
1064         events_stats__inc(&session->stats, event->header.type);
1065
1066         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1067                 return perf_session__process_user_event(session, event, tool, file_offset);
1068
1069         /*
1070          * For all kernel events we get the sample data
1071          */
1072         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1073         if (ret)
1074                 return ret;
1075
1076         if (tool->ordered_samples) {
1077                 ret = perf_session_queue_event(session, event, &sample,
1078                                                file_offset);
1079                 if (ret != -ETIME)
1080                         return ret;
1081         }
1082
1083         return perf_session_deliver_event(session, event, &sample, tool,
1084                                           file_offset);
1085 }
1086
1087 void perf_event_header__bswap(struct perf_event_header *self)
1088 {
1089         self->type = bswap_32(self->type);
1090         self->misc = bswap_16(self->misc);
1091         self->size = bswap_16(self->size);
1092 }
1093
1094 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1095 {
1096         return machine__findnew_thread(&session->machines.host, 0, pid);
1097 }
1098
1099 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1100 {
1101         struct thread *thread = perf_session__findnew(self, 0);
1102
1103         if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1104                 pr_err("problem inserting idle task.\n");
1105                 thread = NULL;
1106         }
1107
1108         return thread;
1109 }
1110
1111 static void perf_session__warn_about_errors(const struct perf_session *session,
1112                                             const struct perf_tool *tool)
1113 {
1114         if (tool->lost == perf_event__process_lost &&
1115             session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1116                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1117                             "Check IO/CPU overload!\n\n",
1118                             session->stats.nr_events[0],
1119                             session->stats.nr_events[PERF_RECORD_LOST]);
1120         }
1121
1122         if (session->stats.nr_unknown_events != 0) {
1123                 ui__warning("Found %u unknown events!\n\n"
1124                             "Is this an older tool processing a perf.data "
1125                             "file generated by a more recent tool?\n\n"
1126                             "If that is not the case, consider "
1127                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1128                             session->stats.nr_unknown_events);
1129         }
1130
1131         if (session->stats.nr_unknown_id != 0) {
1132                 ui__warning("%u samples with id not present in the header\n",
1133                             session->stats.nr_unknown_id);
1134         }
1135
1136         if (session->stats.nr_invalid_chains != 0) {
1137                 ui__warning("Found invalid callchains!\n\n"
1138                             "%u out of %u events were discarded for this reason.\n\n"
1139                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1140                             session->stats.nr_invalid_chains,
1141                             session->stats.nr_events[PERF_RECORD_SAMPLE]);
1142         }
1143
1144         if (session->stats.nr_unprocessable_samples != 0) {
1145                 ui__warning("%u unprocessable samples recorded.\n"
1146                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1147                             session->stats.nr_unprocessable_samples);
1148         }
1149 }
1150
1151 volatile int session_done;
1152
1153 static int __perf_session__process_pipe_events(struct perf_session *self,
1154                                                struct perf_tool *tool)
1155 {
1156         int fd = perf_data_file__fd(self->file);
1157         union perf_event *event;
1158         uint32_t size, cur_size = 0;
1159         void *buf = NULL;
1160         int skip = 0;
1161         u64 head;
1162         int err;
1163         void *p;
1164
1165         perf_tool__fill_defaults(tool);
1166
1167         head = 0;
1168         cur_size = sizeof(union perf_event);
1169
1170         buf = malloc(cur_size);
1171         if (!buf)
1172                 return -errno;
1173 more:
1174         event = buf;
1175         err = readn(fd, event, sizeof(struct perf_event_header));
1176         if (err <= 0) {
1177                 if (err == 0)
1178                         goto done;
1179
1180                 pr_err("failed to read event header\n");
1181                 goto out_err;
1182         }
1183
1184         if (self->header.needs_swap)
1185                 perf_event_header__bswap(&event->header);
1186
1187         size = event->header.size;
1188         if (size < sizeof(struct perf_event_header)) {
1189                 pr_err("bad event header size\n");
1190                 goto out_err;
1191         }
1192
1193         if (size > cur_size) {
1194                 void *new = realloc(buf, size);
1195                 if (!new) {
1196                         pr_err("failed to allocate memory to read event\n");
1197                         goto out_err;
1198                 }
1199                 buf = new;
1200                 cur_size = size;
1201                 event = buf;
1202         }
1203         p = event;
1204         p += sizeof(struct perf_event_header);
1205
1206         if (size - sizeof(struct perf_event_header)) {
1207                 err = readn(fd, p, size - sizeof(struct perf_event_header));
1208                 if (err <= 0) {
1209                         if (err == 0) {
1210                                 pr_err("unexpected end of event stream\n");
1211                                 goto done;
1212                         }
1213
1214                         pr_err("failed to read event data\n");
1215                         goto out_err;
1216                 }
1217         }
1218
1219         if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1220                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1221                        head, event->header.size, event->header.type);
1222                 err = -EINVAL;
1223                 goto out_err;
1224         }
1225
1226         head += size;
1227
1228         if (skip > 0)
1229                 head += skip;
1230
1231         if (!session_done())
1232                 goto more;
1233 done:
1234         /* do the final flush for ordered samples */
1235         self->ordered_samples.next_flush = ULLONG_MAX;
1236         err = flush_sample_queue(self, tool);
1237 out_err:
1238         free(buf);
1239         perf_session__warn_about_errors(self, tool);
1240         perf_session_free_sample_buffers(self);
1241         return err;
1242 }
1243
1244 static union perf_event *
1245 fetch_mmaped_event(struct perf_session *session,
1246                    u64 head, size_t mmap_size, char *buf)
1247 {
1248         union perf_event *event;
1249
1250         /*
1251          * Ensure we have enough space remaining to read
1252          * the size of the event in the headers.
1253          */
1254         if (head + sizeof(event->header) > mmap_size)
1255                 return NULL;
1256
1257         event = (union perf_event *)(buf + head);
1258
1259         if (session->header.needs_swap)
1260                 perf_event_header__bswap(&event->header);
1261
1262         if (head + event->header.size > mmap_size) {
1263                 /* We're not fetching the event so swap back again */
1264                 if (session->header.needs_swap)
1265                         perf_event_header__bswap(&event->header);
1266                 return NULL;
1267         }
1268
1269         return event;
1270 }
1271
1272 /*
1273  * On 64bit we can mmap the data file in one go. No need for tiny mmap
1274  * slices. On 32bit we use 32MB.
1275  */
1276 #if BITS_PER_LONG == 64
1277 #define MMAP_SIZE ULLONG_MAX
1278 #define NUM_MMAPS 1
1279 #else
1280 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1281 #define NUM_MMAPS 128
1282 #endif
1283
1284 int __perf_session__process_events(struct perf_session *session,
1285                                    u64 data_offset, u64 data_size,
1286                                    u64 file_size, struct perf_tool *tool)
1287 {
1288         int fd = perf_data_file__fd(session->file);
1289         u64 head, page_offset, file_offset, file_pos;
1290         int err, mmap_prot, mmap_flags, map_idx = 0;
1291         size_t  mmap_size;
1292         char *buf, *mmaps[NUM_MMAPS];
1293         union perf_event *event;
1294         uint32_t size;
1295         struct ui_progress prog;
1296
1297         perf_tool__fill_defaults(tool);
1298
1299         page_offset = page_size * (data_offset / page_size);
1300         file_offset = page_offset;
1301         head = data_offset - page_offset;
1302
1303         if (data_size && (data_offset + data_size < file_size))
1304                 file_size = data_offset + data_size;
1305
1306         ui_progress__init(&prog, file_size, "Processing events...");
1307
1308         mmap_size = MMAP_SIZE;
1309         if (mmap_size > file_size)
1310                 mmap_size = file_size;
1311
1312         memset(mmaps, 0, sizeof(mmaps));
1313
1314         mmap_prot  = PROT_READ;
1315         mmap_flags = MAP_SHARED;
1316
1317         if (session->header.needs_swap) {
1318                 mmap_prot  |= PROT_WRITE;
1319                 mmap_flags = MAP_PRIVATE;
1320         }
1321 remap:
1322         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1323                    file_offset);
1324         if (buf == MAP_FAILED) {
1325                 pr_err("failed to mmap file\n");
1326                 err = -errno;
1327                 goto out_err;
1328         }
1329         mmaps[map_idx] = buf;
1330         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1331         file_pos = file_offset + head;
1332
1333 more:
1334         event = fetch_mmaped_event(session, head, mmap_size, buf);
1335         if (!event) {
1336                 if (mmaps[map_idx]) {
1337                         munmap(mmaps[map_idx], mmap_size);
1338                         mmaps[map_idx] = NULL;
1339                 }
1340
1341                 page_offset = page_size * (head / page_size);
1342                 file_offset += page_offset;
1343                 head -= page_offset;
1344                 goto remap;
1345         }
1346
1347         size = event->header.size;
1348
1349         if (size < sizeof(struct perf_event_header) ||
1350             perf_session__process_event(session, event, tool, file_pos) < 0) {
1351                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1352                        file_offset + head, event->header.size,
1353                        event->header.type);
1354                 err = -EINVAL;
1355                 goto out_err;
1356         }
1357
1358         head += size;
1359         file_pos += size;
1360
1361         ui_progress__update(&prog, size);
1362
1363         if (session_done())
1364                 goto out;
1365
1366         if (file_pos < file_size)
1367                 goto more;
1368
1369 out:
1370         /* do the final flush for ordered samples */
1371         session->ordered_samples.next_flush = ULLONG_MAX;
1372         err = flush_sample_queue(session, tool);
1373 out_err:
1374         ui_progress__finish();
1375         perf_session__warn_about_errors(session, tool);
1376         perf_session_free_sample_buffers(session);
1377         return err;
1378 }
1379
1380 int perf_session__process_events(struct perf_session *self,
1381                                  struct perf_tool *tool)
1382 {
1383         u64 size = perf_data_file__size(self->file);
1384         int err;
1385
1386         if (perf_session__register_idle_thread(self) == NULL)
1387                 return -ENOMEM;
1388
1389         if (!perf_data_file__is_pipe(self->file))
1390                 err = __perf_session__process_events(self,
1391                                                      self->header.data_offset,
1392                                                      self->header.data_size,
1393                                                      size, tool);
1394         else
1395                 err = __perf_session__process_pipe_events(self, tool);
1396
1397         return err;
1398 }
1399
1400 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1401 {
1402         struct perf_evsel *evsel;
1403
1404         list_for_each_entry(evsel, &session->evlist->entries, node) {
1405                 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1406                         return true;
1407         }
1408
1409         pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1410         return false;
1411 }
1412
1413 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1414                                      const char *symbol_name, u64 addr)
1415 {
1416         char *bracket;
1417         enum map_type i;
1418         struct ref_reloc_sym *ref;
1419
1420         ref = zalloc(sizeof(struct ref_reloc_sym));
1421         if (ref == NULL)
1422                 return -ENOMEM;
1423
1424         ref->name = strdup(symbol_name);
1425         if (ref->name == NULL) {
1426                 free(ref);
1427                 return -ENOMEM;
1428         }
1429
1430         bracket = strchr(ref->name, ']');
1431         if (bracket)
1432                 *bracket = '\0';
1433
1434         ref->addr = addr;
1435
1436         for (i = 0; i < MAP__NR_TYPES; ++i) {
1437                 struct kmap *kmap = map__kmap(maps[i]);
1438                 kmap->ref_reloc_sym = ref;
1439         }
1440
1441         return 0;
1442 }
1443
1444 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1445 {
1446         return machines__fprintf_dsos(&self->machines, fp);
1447 }
1448
1449 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1450                                           bool (skip)(struct dso *dso, int parm), int parm)
1451 {
1452         return machines__fprintf_dsos_buildid(&self->machines, fp, skip, parm);
1453 }
1454
1455 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1456 {
1457         struct perf_evsel *pos;
1458         size_t ret = fprintf(fp, "Aggregated stats:\n");
1459
1460         ret += events_stats__fprintf(&session->stats, fp);
1461
1462         list_for_each_entry(pos, &session->evlist->entries, node) {
1463                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1464                 ret += events_stats__fprintf(&pos->hists.stats, fp);
1465         }
1466
1467         return ret;
1468 }
1469
1470 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1471 {
1472         /*
1473          * FIXME: Here we have to actually print all the machines in this
1474          * session, not just the host...
1475          */
1476         return machine__fprintf(&session->machines.host, fp);
1477 }
1478
1479 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1480                                               unsigned int type)
1481 {
1482         struct perf_evsel *pos;
1483
1484         list_for_each_entry(pos, &session->evlist->entries, node) {
1485                 if (pos->attr.type == type)
1486                         return pos;
1487         }
1488         return NULL;
1489 }
1490
1491 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1492                           struct perf_sample *sample, struct machine *machine,
1493                           unsigned int print_opts, unsigned int stack_depth)
1494 {
1495         struct addr_location al;
1496         struct callchain_cursor_node *node;
1497         int print_ip = print_opts & PRINT_IP_OPT_IP;
1498         int print_sym = print_opts & PRINT_IP_OPT_SYM;
1499         int print_dso = print_opts & PRINT_IP_OPT_DSO;
1500         int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1501         int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1502         char s = print_oneline ? ' ' : '\t';
1503
1504         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
1505                 error("problem processing %d event, skipping it.\n",
1506                         event->header.type);
1507                 return;
1508         }
1509
1510         if (symbol_conf.use_callchain && sample->callchain) {
1511
1512                 if (machine__resolve_callchain(machine, evsel, al.thread,
1513                                                sample, NULL, NULL,
1514                                                PERF_MAX_STACK_DEPTH) != 0) {
1515                         if (verbose)
1516                                 error("Failed to resolve callchain. Skipping\n");
1517                         return;
1518                 }
1519                 callchain_cursor_commit(&callchain_cursor);
1520
1521                 while (stack_depth) {
1522                         node = callchain_cursor_current(&callchain_cursor);
1523                         if (!node)
1524                                 break;
1525
1526                         if (print_ip)
1527                                 printf("%c%16" PRIx64, s, node->ip);
1528
1529                         if (print_sym) {
1530                                 printf(" ");
1531                                 if (print_symoffset) {
1532                                         al.addr = node->ip;
1533                                         al.map  = node->map;
1534                                         symbol__fprintf_symname_offs(node->sym, &al, stdout);
1535                                 } else
1536                                         symbol__fprintf_symname(node->sym, stdout);
1537                         }
1538
1539                         if (print_dso) {
1540                                 printf(" (");
1541                                 map__fprintf_dsoname(node->map, stdout);
1542                                 printf(")");
1543                         }
1544
1545                         if (!print_oneline)
1546                                 printf("\n");
1547
1548                         callchain_cursor_advance(&callchain_cursor);
1549
1550                         stack_depth--;
1551                 }
1552
1553         } else {
1554                 if (print_ip)
1555                         printf("%16" PRIx64, sample->ip);
1556
1557                 if (print_sym) {
1558                         printf(" ");
1559                         if (print_symoffset)
1560                                 symbol__fprintf_symname_offs(al.sym, &al,
1561                                                              stdout);
1562                         else
1563                                 symbol__fprintf_symname(al.sym, stdout);
1564                 }
1565
1566                 if (print_dso) {
1567                         printf(" (");
1568                         map__fprintf_dsoname(al.map, stdout);
1569                         printf(")");
1570                 }
1571         }
1572 }
1573
1574 int perf_session__cpu_bitmap(struct perf_session *session,
1575                              const char *cpu_list, unsigned long *cpu_bitmap)
1576 {
1577         int i;
1578         struct cpu_map *map;
1579
1580         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1581                 struct perf_evsel *evsel;
1582
1583                 evsel = perf_session__find_first_evtype(session, i);
1584                 if (!evsel)
1585                         continue;
1586
1587                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1588                         pr_err("File does not contain CPU events. "
1589                                "Remove -c option to proceed.\n");
1590                         return -1;
1591                 }
1592         }
1593
1594         map = cpu_map__new(cpu_list);
1595         if (map == NULL) {
1596                 pr_err("Invalid cpu_list\n");
1597                 return -1;
1598         }
1599
1600         for (i = 0; i < map->nr; i++) {
1601                 int cpu = map->map[i];
1602
1603                 if (cpu >= MAX_NR_CPUS) {
1604                         pr_err("Requested CPU %d too large. "
1605                                "Consider raising MAX_NR_CPUS\n", cpu);
1606                         return -1;
1607                 }
1608
1609                 set_bit(cpu, cpu_bitmap);
1610         }
1611
1612         return 0;
1613 }
1614
1615 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1616                                 bool full)
1617 {
1618         int fd = perf_data_file__fd(session->file);
1619         struct stat st;
1620         int ret;
1621
1622         if (session == NULL || fp == NULL)
1623                 return;
1624
1625         ret = fstat(fd, &st);
1626         if (ret == -1)
1627                 return;
1628
1629         fprintf(fp, "# ========\n");
1630         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1631         perf_header__fprintf_info(session, fp, full);
1632         fprintf(fp, "# ========\n#\n");
1633 }
1634
1635
1636 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1637                                              const struct perf_evsel_str_handler *assocs,
1638                                              size_t nr_assocs)
1639 {
1640         struct perf_evsel *evsel;
1641         size_t i;
1642         int err;
1643
1644         for (i = 0; i < nr_assocs; i++) {
1645                 /*
1646                  * Adding a handler for an event not in the session,
1647                  * just ignore it.
1648                  */
1649                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1650                 if (evsel == NULL)
1651                         continue;
1652
1653                 err = -EEXIST;
1654                 if (evsel->handler.func != NULL)
1655                         goto out;
1656                 evsel->handler.func = assocs[i].handler;
1657         }
1658
1659         err = 0;
1660 out:
1661         return err;
1662 }