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, bool force)
  20 {
  21         struct stat input_stat;
  22
  23         if (!strcmp(self->filename, "-")) {
  24                 self->fd_pipe = true;
  25                 self->fd = STDIN_FILENO;
  26
  27                 if (perf_session__read_header(self) < 0)
  28                         pr_err("incompatible file format (rerun with -v to learn more)");
  29
  30                 return 0;
  31         }
  32
  33         self->fd = open(self->filename, O_RDONLY);
  34         if (self->fd < 0) {
  35                 int err = errno;
  36
  37                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  38                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  39                         pr_err("  (try 'perf record' first)");
  40                 pr_err("\n");
  41                 return -errno;
  42         }
  43
  44         if (fstat(self->fd, &input_stat) < 0)
  45                 goto out_close;
  46
  47         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  48                 pr_err("file %s not owned by current user or root\n",
  49                        self->filename);
  50                 goto out_close;
  51         }
  52
  53         if (!input_stat.st_size) {
  54                 pr_info("zero-sized file (%s), nothing to do!\n",
  55                         self->filename);
  56                 goto out_close;
  57         }
  58
  59         if (perf_session__read_header(self) < 0) {
  60                 pr_err("incompatible file format (rerun with -v to learn more)");
  61                 goto out_close;
  62         }
  63
  64         if (!perf_evlist__valid_sample_type(self->evlist)) {
  65                 pr_err("non matching sample_type");
  66                 goto out_close;
  67         }
  68
  69         if (!perf_evlist__valid_sample_id_all(self->evlist)) {
  70                 pr_err("non matching sample_id_all");
  71                 goto out_close;
  72         }
  73
  74         if (!perf_evlist__valid_read_format(self->evlist)) {
  75                 pr_err("non matching read_format");
  76                 goto out_close;
  77         }
  78
  79         self->size = input_stat.st_size;
  80         return 0;
  81
  82 out_close:
  83         close(self->fd);
  84         self->fd = -1;
  85         return -1;
  86 }
  87
  88 void perf_session__set_id_hdr_size(struct perf_session *session)
  89 {
  90         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
  91
  92         machines__set_id_hdr_size(&session->machines, id_hdr_size);
  93 }
  94
  95 int perf_session__create_kernel_maps(struct perf_session *self)
  96 {
  97         int ret = machine__create_kernel_maps(&self->machines.host);
  98
  99         if (ret >= 0)
 100                 ret = machines__create_guest_kernel_maps(&self->machines);
 101         return ret;
 102 }
 103
 104 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 105 {
 106         machines__destroy_kernel_maps(&self->machines);
 107 }
 108
 109 struct perf_session *perf_session__new(const char *filename, int mode,
 110                                        bool force, bool repipe,
 111                                        struct perf_tool *tool)
 112 {
 113         struct perf_session *self;
 114         struct stat st;
 115         size_t len;
 116
 117         if (!filename || !strlen(filename)) {
 118                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 119                         filename = "-";
 120                 else
 121                         filename = "perf.data";
 122         }
 123
 124         len = strlen(filename);
 125         self = zalloc(sizeof(*self) + len);
 126
 127         if (self == NULL)
 128                 goto out;
 129
 130         memcpy(self->filename, filename, len);
 131         self->repipe = repipe;
 132         INIT_LIST_HEAD(&self->ordered_samples.samples);
 133         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 134         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 135         machines__init(&self->machines);
 136
 137         if (mode == O_RDONLY) {
 138                 if (perf_session__open(self, force) < 0)
 139                         goto out_delete;
 140                 perf_session__set_id_hdr_size(self);
 141         } else if (mode == O_WRONLY) {
 142                 /*
 143                  * In O_RDONLY mode this will be performed when reading the
 144                  * kernel MMAP event, in perf_event__process_mmap().
 145                  */
 146                 if (perf_session__create_kernel_maps(self) < 0)
 147                         goto out_delete;
 148         }
 149
 150         if (tool && tool->ordering_requires_timestamps &&
 151             tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
 152                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 153                 tool->ordered_samples = false;
 154         }
 155
 156 out:
 157         return self;
 158 out_delete:
 159         perf_session__delete(self);
 160         return NULL;
 161 }
 162
 163 static void perf_session__delete_dead_threads(struct perf_session *session)
 164 {
 165         machine__delete_dead_threads(&session->machines.host);
 166 }
 167
 168 static void perf_session__delete_threads(struct perf_session *session)
 169 {
 170         machine__delete_threads(&session->machines.host);
 171 }
 172
 173 static void perf_session_env__delete(struct perf_session_env *env)
 174 {
 175         free(env->hostname);
 176         free(env->os_release);
 177         free(env->version);
 178         free(env->arch);
 179         free(env->cpu_desc);
 180         free(env->cpuid);
 181
 182         free(env->cmdline);
 183         free(env->sibling_cores);
 184         free(env->sibling_threads);
 185         free(env->numa_nodes);
 186         free(env->pmu_mappings);
 187 }
 188
 189 void perf_session__delete(struct perf_session *self)
 190 {
 191         perf_session__destroy_kernel_maps(self);
 192         perf_session__delete_dead_threads(self);
 193         perf_session__delete_threads(self);
 194         perf_session_env__delete(&self->header.env);
 195         machines__exit(&self->machines);
 196         close(self->fd);
 197         free(self);
 198         vdso__exit();
 199 }
 200
 201 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
 202                                                  __maybe_unused,
 203                                                  union perf_event *event
 204                                                  __maybe_unused,
 205                                                  struct perf_session *session
 206                                                 __maybe_unused)
 207 {
 208         dump_printf(": unhandled!\n");
 209         return 0;
 210 }
 211
 212 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
 213                                          union perf_event *event __maybe_unused,
 214                                          struct perf_evlist **pevlist
 215                                          __maybe_unused)
 216 {
 217         dump_printf(": unhandled!\n");
 218         return 0;
 219 }
 220
 221 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
 222                                      union perf_event *event __maybe_unused,
 223                                      struct perf_sample *sample __maybe_unused,
 224                                      struct perf_evsel *evsel __maybe_unused,
 225                                      struct machine *machine __maybe_unused)
 226 {
 227         dump_printf(": unhandled!\n");
 228         return 0;
 229 }
 230
 231 static int process_event_stub(struct perf_tool *tool __maybe_unused,
 232                               union perf_event *event __maybe_unused,
 233                               struct perf_sample *sample __maybe_unused,
 234                               struct machine *machine __maybe_unused)
 235 {
 236         dump_printf(": unhandled!\n");
 237         return 0;
 238 }
 239
 240 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
 241                                        union perf_event *event __maybe_unused,
 242                                        struct perf_session *perf_session
 243                                        __maybe_unused)
 244 {
 245         dump_printf(": unhandled!\n");
 246         return 0;
 247 }
 248
 249 static int process_finished_round(struct perf_tool *tool,
 250                                   union perf_event *event,
 251                                   struct perf_session *session);
 252
 253 static void perf_tool__fill_defaults(struct perf_tool *tool)
 254 {
 255         if (tool->sample == NULL)
 256                 tool->sample = process_event_sample_stub;
 257         if (tool->mmap == NULL)
 258                 tool->mmap = process_event_stub;
 259         if (tool->comm == NULL)
 260                 tool->comm = process_event_stub;
 261         if (tool->fork == NULL)
 262                 tool->fork = process_event_stub;
 263         if (tool->exit == NULL)
 264                 tool->exit = process_event_stub;
 265         if (tool->lost == NULL)
 266                 tool->lost = perf_event__process_lost;
 267         if (tool->read == NULL)
 268                 tool->read = process_event_sample_stub;
 269         if (tool->throttle == NULL)
 270                 tool->throttle = process_event_stub;
 271         if (tool->unthrottle == NULL)
 272                 tool->unthrottle = process_event_stub;
 273         if (tool->attr == NULL)
 274                 tool->attr = process_event_synth_attr_stub;
 275         if (tool->tracing_data == NULL)
 276                 tool->tracing_data = process_event_synth_tracing_data_stub;
 277         if (tool->build_id == NULL)
 278                 tool->build_id = process_finished_round_stub;
 279         if (tool->finished_round == NULL) {
 280                 if (tool->ordered_samples)
 281                         tool->finished_round = process_finished_round;
 282                 else
 283                         tool->finished_round = process_finished_round_stub;
 284         }
 285 }
 286
 287 void mem_bswap_32(void *src, int byte_size)
 288 {
 289         u32 *m = src;
 290         while (byte_size > 0) {
 291                 *m = bswap_32(*m);
 292                 byte_size -= sizeof(u32);
 293                 ++m;
 294         }
 295 }
 296
 297 void mem_bswap_64(void *src, int byte_size)
 298 {
 299         u64 *m = src;
 300
 301         while (byte_size > 0) {
 302                 *m = bswap_64(*m);
 303                 byte_size -= sizeof(u64);
 304                 ++m;
 305         }
 306 }
 307
 308 static void swap_sample_id_all(union perf_event *event, void *data)
 309 {
 310         void *end = (void *) event + event->header.size;
 311         int size = end - data;
 312
 313         BUG_ON(size % sizeof(u64));
 314         mem_bswap_64(data, size);
 315 }
 316
 317 static void perf_event__all64_swap(union perf_event *event,
 318                                    bool sample_id_all __maybe_unused)
 319 {
 320         struct perf_event_header *hdr = &event->header;
 321         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 322 }
 323
 324 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 325 {
 326         event->comm.pid = bswap_32(event->comm.pid);
 327         event->comm.tid = bswap_32(event->comm.tid);
 328
 329         if (sample_id_all) {
 330                 void *data = &event->comm.comm;
 331
 332                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 333                 swap_sample_id_all(event, data);
 334         }
 335 }
 336
 337 static void perf_event__mmap_swap(union perf_event *event,
 338                                   bool sample_id_all)
 339 {
 340         event->mmap.pid   = bswap_32(event->mmap.pid);
 341         event->mmap.tid   = bswap_32(event->mmap.tid);
 342         event->mmap.start = bswap_64(event->mmap.start);
 343         event->mmap.len   = bswap_64(event->mmap.len);
 344         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 345
 346         if (sample_id_all) {
 347                 void *data = &event->mmap.filename;
 348
 349                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 350                 swap_sample_id_all(event, data);
 351         }
 352 }
 353
 354 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 355 {
 356         event->fork.pid  = bswap_32(event->fork.pid);
 357         event->fork.tid  = bswap_32(event->fork.tid);
 358         event->fork.ppid = bswap_32(event->fork.ppid);
 359         event->fork.ptid = bswap_32(event->fork.ptid);
 360         event->fork.time = bswap_64(event->fork.time);
 361
 362         if (sample_id_all)
 363                 swap_sample_id_all(event, &event->fork + 1);
 364 }
 365
 366 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 367 {
 368         event->read.pid          = bswap_32(event->read.pid);
 369         event->read.tid          = bswap_32(event->read.tid);
 370         event->read.value        = bswap_64(event->read.value);
 371         event->read.time_enabled = bswap_64(event->read.time_enabled);
 372         event->read.time_running = bswap_64(event->read.time_running);
 373         event->read.id           = bswap_64(event->read.id);
 374
 375         if (sample_id_all)
 376                 swap_sample_id_all(event, &event->read + 1);
 377 }
 378
 379 static u8 revbyte(u8 b)
 380 {
 381         int rev = (b >> 4) | ((b & 0xf) << 4);
 382         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 383         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 384         return (u8) rev;
 385 }
 386
 387 /*
 388  * XXX this is hack in attempt to carry flags bitfield
 389  * throught endian village. ABI says:
 390  *
 391  * Bit-fields are allocated from right to left (least to most significant)
 392  * on little-endian implementations and from left to right (most to least
 393  * significant) on big-endian implementations.
 394  *
 395  * The above seems to be byte specific, so we need to reverse each
 396  * byte of the bitfield. 'Internet' also says this might be implementation
 397  * specific and we probably need proper fix and carry perf_event_attr
 398  * bitfield flags in separate data file FEAT_ section. Thought this seems
 399  * to work for now.
 400  */
 401 static void swap_bitfield(u8 *p, unsigned len)
 402 {
 403         unsigned i;
 404
 405         for (i = 0; i < len; i++) {
 406                 *p = revbyte(*p);
 407                 p++;
 408         }
 409 }
 410
 411 /* exported for swapping attributes in file header */
 412 void perf_event__attr_swap(struct perf_event_attr *attr)
 413 {
 414         attr->type              = bswap_32(attr->type);
 415         attr->size              = bswap_32(attr->size);
 416         attr->config            = bswap_64(attr->config);
 417         attr->sample_period     = bswap_64(attr->sample_period);
 418         attr->sample_type       = bswap_64(attr->sample_type);
 419         attr->read_format       = bswap_64(attr->read_format);
 420         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 421         attr->bp_type           = bswap_32(attr->bp_type);
 422         attr->bp_addr           = bswap_64(attr->bp_addr);
 423         attr->bp_len            = bswap_64(attr->bp_len);
 424
 425         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 426 }
 427
 428 static void perf_event__hdr_attr_swap(union perf_event *event,
 429                                       bool sample_id_all __maybe_unused)
 430 {
 431         size_t size;
 432
 433         perf_event__attr_swap(&event->attr.attr);
 434
 435         size = event->header.size;
 436         size -= (void *)&event->attr.id - (void *)event;
 437         mem_bswap_64(event->attr.id, size);
 438 }
 439
 440 static void perf_event__event_type_swap(union perf_event *event,
 441                                         bool sample_id_all __maybe_unused)
 442 {
 443         event->event_type.event_type.event_id =
 444                 bswap_64(event->event_type.event_type.event_id);
 445 }
 446
 447 static void perf_event__tracing_data_swap(union perf_event *event,
 448                                           bool sample_id_all __maybe_unused)
 449 {
 450         event->tracing_data.size = bswap_32(event->tracing_data.size);
 451 }
 452
 453 typedef void (*perf_event__swap_op)(union perf_event *event,
 454                                     bool sample_id_all);
 455
 456 static perf_event__swap_op perf_event__swap_ops[] = {
 457         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 458         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 459         [PERF_RECORD_FORK]                = perf_event__task_swap,
 460         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 461         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 462         [PERF_RECORD_READ]                = perf_event__read_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         unsigned idx = 0, progress_next = os->nr_samples / 16;
 507         int ret;
 508
 509         if (!tool->ordered_samples || !limit)
 510                 return 0;
 511
 512         list_for_each_entry_safe(iter, tmp, head, list) {
 513                 if (iter->timestamp > limit)
 514                         break;
 515
 516                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
 517                 if (ret)
 518                         pr_err("Can't parse sample, err = %d\n", ret);
 519                 else {
 520                         ret = perf_session_deliver_event(s, iter->event, &sample, tool,
 521                                                          iter->file_offset);
 522                         if (ret)
 523                                 return ret;
 524                 }
 525
 526                 os->last_flush = iter->timestamp;
 527                 list_del(&iter->list);
 528                 list_add(&iter->list, &os->sample_cache);
 529                 if (++idx >= progress_next) {
 530                         progress_next += os->nr_samples / 16;
 531                         ui_progress__update(idx, os->nr_samples,
 532                                             "Processing time ordered events...");
 533                 }
 534         }
 535
 536         if (list_empty(head)) {
 537                 os->last_sample = NULL;
 538         } else if (last_ts <= limit) {
 539                 os->last_sample =
 540                         list_entry(head->prev, struct sample_queue, list);
 541         }
 542
 543         os->nr_samples = 0;
 544
 545         return 0;
 546 }
 547
 548 /*
 549  * When perf record finishes a pass on every buffers, it records this pseudo
 550  * event.
 551  * We record the max timestamp t found in the pass n.
 552  * Assuming these timestamps are monotonic across cpus, we know that if
 553  * a buffer still has events with timestamps below t, they will be all
 554  * available and then read in the pass n + 1.
 555  * Hence when we start to read the pass n + 2, we can safely flush every
 556  * events with timestamps below t.
 557  *
 558  *    ============ PASS n =================
 559  *       CPU 0         |   CPU 1
 560  *                     |
 561  *    cnt1 timestamps  |   cnt2 timestamps
 562  *          1          |         2
 563  *          2          |         3
 564  *          -          |         4  <--- max recorded
 565  *
 566  *    ============ PASS n + 1 ==============
 567  *       CPU 0         |   CPU 1
 568  *                     |
 569  *    cnt1 timestamps  |   cnt2 timestamps
 570  *          3          |         5
 571  *          4          |         6
 572  *          5          |         7 <---- max recorded
 573  *
 574  *      Flush every events below timestamp 4
 575  *
 576  *    ============ PASS n + 2 ==============
 577  *       CPU 0         |   CPU 1
 578  *                     |
 579  *    cnt1 timestamps  |   cnt2 timestamps
 580  *          6          |         8
 581  *          7          |         9
 582  *          -          |         10
 583  *
 584  *      Flush every events below timestamp 7
 585  *      etc...
 586  */
 587 static int process_finished_round(struct perf_tool *tool,
 588                                   union perf_event *event __maybe_unused,
 589                                   struct perf_session *session)
 590 {
 591         int ret = flush_sample_queue(session, tool);
 592         if (!ret)
 593                 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 594
 595         return ret;
 596 }
 597
 598 /* The queue is ordered by time */
 599 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 600 {
 601         struct ordered_samples *os = &s->ordered_samples;
 602         struct sample_queue *sample = os->last_sample;
 603         u64 timestamp = new->timestamp;
 604         struct list_head *p;
 605
 606         ++os->nr_samples;
 607         os->last_sample = new;
 608
 609         if (!sample) {
 610                 list_add(&new->list, &os->samples);
 611                 os->max_timestamp = timestamp;
 612                 return;
 613         }
 614
 615         /*
 616          * last_sample might point to some random place in the list as it's
 617          * the last queued event. We expect that the new event is close to
 618          * this.
 619          */
 620         if (sample->timestamp <= timestamp) {
 621                 while (sample->timestamp <= timestamp) {
 622                         p = sample->list.next;
 623                         if (p == &os->samples) {
 624                                 list_add_tail(&new->list, &os->samples);
 625                                 os->max_timestamp = timestamp;
 626                                 return;
 627                         }
 628                         sample = list_entry(p, struct sample_queue, list);
 629                 }
 630                 list_add_tail(&new->list, &sample->list);
 631         } else {
 632                 while (sample->timestamp > timestamp) {
 633                         p = sample->list.prev;
 634                         if (p == &os->samples) {
 635                                 list_add(&new->list, &os->samples);
 636                                 return;
 637                         }
 638                         sample = list_entry(p, struct sample_queue, list);
 639                 }
 640                 list_add(&new->list, &sample->list);
 641         }
 642 }
 643
 644 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 645
 646 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 647                                     struct perf_sample *sample, u64 file_offset)
 648 {
 649         struct ordered_samples *os = &s->ordered_samples;
 650         struct list_head *sc = &os->sample_cache;
 651         u64 timestamp = sample->time;
 652         struct sample_queue *new;
 653
 654         if (!timestamp || timestamp == ~0ULL)
 655                 return -ETIME;
 656
 657         if (timestamp < s->ordered_samples.last_flush) {
 658                 printf("Warning: Timestamp below last timeslice flush\n");
 659                 return -EINVAL;
 660         }
 661
 662         if (!list_empty(sc)) {
 663                 new = list_entry(sc->next, struct sample_queue, list);
 664                 list_del(&new->list);
 665         } else if (os->sample_buffer) {
 666                 new = os->sample_buffer + os->sample_buffer_idx;
 667                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 668                         os->sample_buffer = NULL;
 669         } else {
 670                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 671                 if (!os->sample_buffer)
 672                         return -ENOMEM;
 673                 list_add(&os->sample_buffer->list, &os->to_free);
 674                 os->sample_buffer_idx = 2;
 675                 new = os->sample_buffer + 1;
 676         }
 677
 678         new->timestamp = timestamp;
 679         new->file_offset = file_offset;
 680         new->event = event;
 681
 682         __queue_event(new, s);
 683
 684         return 0;
 685 }
 686
 687 static void callchain__printf(struct perf_sample *sample)
 688 {
 689         unsigned int i;
 690
 691         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 692
 693         for (i = 0; i < sample->callchain->nr; i++)
 694                 printf("..... %2d: %016" PRIx64 "\n",
 695                        i, sample->callchain->ips[i]);
 696 }
 697
 698 static void branch_stack__printf(struct perf_sample *sample)
 699 {
 700         uint64_t i;
 701
 702         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 703
 704         for (i = 0; i < sample->branch_stack->nr; i++)
 705                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 706                         i, sample->branch_stack->entries[i].from,
 707                         sample->branch_stack->entries[i].to);
 708 }
 709
 710 static void regs_dump__printf(u64 mask, u64 *regs)
 711 {
 712         unsigned rid, i = 0;
 713
 714         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 715                 u64 val = regs[i++];
 716
 717                 printf(".... %-5s 0x%" PRIx64 "\n",
 718                        perf_reg_name(rid), val);
 719         }
 720 }
 721
 722 static void regs_user__printf(struct perf_sample *sample, u64 mask)
 723 {
 724         struct regs_dump *user_regs = &sample->user_regs;
 725
 726         if (user_regs->regs) {
 727                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
 728                 regs_dump__printf(mask, user_regs->regs);
 729         }
 730 }
 731
 732 static void stack_user__printf(struct stack_dump *dump)
 733 {
 734         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 735                dump->size, dump->offset);
 736 }
 737
 738 static void perf_session__print_tstamp(struct perf_session *session,
 739                                        union perf_event *event,
 740                                        struct perf_sample *sample)
 741 {
 742         u64 sample_type = perf_evlist__sample_type(session->evlist);
 743
 744         if (event->header.type != PERF_RECORD_SAMPLE &&
 745             !perf_evlist__sample_id_all(session->evlist)) {
 746                 fputs("-1 -1 ", stdout);
 747                 return;
 748         }
 749
 750         if ((sample_type & PERF_SAMPLE_CPU))
 751                 printf("%u ", sample->cpu);
 752
 753         if (sample_type & PERF_SAMPLE_TIME)
 754                 printf("%" PRIu64 " ", sample->time);
 755 }
 756
 757 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
 758 {
 759         printf("... sample_read:\n");
 760
 761         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 762                 printf("...... time enabled %016" PRIx64 "\n",
 763                        sample->read.time_enabled);
 764
 765         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 766                 printf("...... time running %016" PRIx64 "\n",
 767                        sample->read.time_running);
 768
 769         if (read_format & PERF_FORMAT_GROUP) {
 770                 u64 i;
 771
 772                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
 773
 774                 for (i = 0; i < sample->read.group.nr; i++) {
 775                         struct sample_read_value *value;
 776
 777                         value = &sample->read.group.values[i];
 778                         printf("..... id %016" PRIx64
 779                                ", value %016" PRIx64 "\n",
 780                                value->id, value->value);
 781                 }
 782         } else
 783                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
 784                         sample->read.one.id, sample->read.one.value);
 785 }
 786
 787 static void dump_event(struct perf_session *session, union perf_event *event,
 788                        u64 file_offset, struct perf_sample *sample)
 789 {
 790         if (!dump_trace)
 791                 return;
 792
 793         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 794                file_offset, event->header.size, event->header.type);
 795
 796         trace_event(event);
 797
 798         if (sample)
 799                 perf_session__print_tstamp(session, event, sample);
 800
 801         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 802                event->header.size, perf_event__name(event->header.type));
 803 }
 804
 805 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 806                         struct perf_sample *sample)
 807 {
 808         u64 sample_type;
 809
 810         if (!dump_trace)
 811                 return;
 812
 813         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 814                event->header.misc, sample->pid, sample->tid, sample->ip,
 815                sample->period, sample->addr);
 816
 817         sample_type = evsel->attr.sample_type;
 818
 819         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 820                 callchain__printf(sample);
 821
 822         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
 823                 branch_stack__printf(sample);
 824
 825         if (sample_type & PERF_SAMPLE_REGS_USER)
 826                 regs_user__printf(sample, evsel->attr.sample_regs_user);
 827
 828         if (sample_type & PERF_SAMPLE_STACK_USER)
 829                 stack_user__printf(&sample->user_stack);
 830
 831         if (sample_type & PERF_SAMPLE_WEIGHT)
 832                 printf("... weight: %" PRIu64 "\n", sample->weight);
 833
 834         if (sample_type & PERF_SAMPLE_DATA_SRC)
 835                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
 836
 837         if (sample_type & PERF_SAMPLE_READ)
 838                 sample_read__printf(sample, evsel->attr.read_format);
 839 }
 840
 841 static struct machine *
 842         perf_session__find_machine_for_cpumode(struct perf_session *session,
 843                                                union perf_event *event)
 844 {
 845         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 846
 847         if (perf_guest &&
 848             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
 849              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
 850                 u32 pid;
 851
 852                 if (event->header.type == PERF_RECORD_MMAP)
 853                         pid = event->mmap.pid;
 854                 else
 855                         pid = event->ip.pid;
 856
 857                 return perf_session__findnew_machine(session, pid);
 858         }
 859
 860         return &session->machines.host;
 861 }
 862
 863 static int deliver_sample_value(struct perf_session *session,
 864                                 struct perf_tool *tool,
 865                                 union perf_event *event,
 866                                 struct perf_sample *sample,
 867                                 struct sample_read_value *v,
 868                                 struct machine *machine)
 869 {
 870         struct perf_sample_id *sid;
 871
 872         sid = perf_evlist__id2sid(session->evlist, v->id);
 873         if (sid) {
 874                 sample->id     = v->id;
 875                 sample->period = v->value - sid->period;
 876                 sid->period    = v->value;
 877         }
 878
 879         if (!sid || sid->evsel == NULL) {
 880                 ++session->stats.nr_unknown_id;
 881                 return 0;
 882         }
 883
 884         return tool->sample(tool, event, sample, sid->evsel, machine);
 885 }
 886
 887 static int deliver_sample_group(struct perf_session *session,
 888                                 struct perf_tool *tool,
 889                                 union  perf_event *event,
 890                                 struct perf_sample *sample,
 891                                 struct machine *machine)
 892 {
 893         int ret = -EINVAL;
 894         u64 i;
 895
 896         for (i = 0; i < sample->read.group.nr; i++) {
 897                 ret = deliver_sample_value(session, tool, event, sample,
 898                                            &sample->read.group.values[i],
 899                                            machine);
 900                 if (ret)
 901                         break;
 902         }
 903
 904         return ret;
 905 }
 906
 907 static int
 908 perf_session__deliver_sample(struct perf_session *session,
 909                              struct perf_tool *tool,
 910                              union  perf_event *event,
 911                              struct perf_sample *sample,
 912                              struct perf_evsel *evsel,
 913                              struct machine *machine)
 914 {
 915         /* We know evsel != NULL. */
 916         u64 sample_type = evsel->attr.sample_type;
 917         u64 read_format = evsel->attr.read_format;
 918
 919         /* Standard sample delievery. */
 920         if (!(sample_type & PERF_SAMPLE_READ))
 921                 return tool->sample(tool, event, sample, evsel, machine);
 922
 923         /* For PERF_SAMPLE_READ we have either single or group mode. */
 924         if (read_format & PERF_FORMAT_GROUP)
 925                 return deliver_sample_group(session, tool, event, sample,
 926                                             machine);
 927         else
 928                 return deliver_sample_value(session, tool, event, sample,
 929                                             &sample->read.one, machine);
 930 }
 931
 932 static int perf_session_deliver_event(struct perf_session *session,
 933                                       union perf_event *event,
 934                                       struct perf_sample *sample,
 935                                       struct perf_tool *tool,
 936                                       u64 file_offset)
 937 {
 938         struct perf_evsel *evsel;
 939         struct machine *machine;
 940
 941         dump_event(session, event, file_offset, sample);
 942
 943         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 944         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
 945                 /*
 946                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
 947                  * because the tools right now may apply filters, discarding
 948                  * some of the samples. For consistency, in the future we
 949                  * should have something like nr_filtered_samples and remove
 950                  * the sample->period from total_sample_period, etc, KISS for
 951                  * now tho.
 952                  *
 953                  * Also testing against NULL allows us to handle files without
 954                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
 955                  * future probably it'll be a good idea to restrict event
 956                  * processing via perf_session to files with both set.
 957                  */
 958                 hists__inc_nr_events(&evsel->hists, event->header.type);
 959         }
 960
 961         machine = perf_session__find_machine_for_cpumode(session, event);
 962
 963         switch (event->header.type) {
 964         case PERF_RECORD_SAMPLE:
 965                 dump_sample(evsel, event, sample);
 966                 if (evsel == NULL) {
 967                         ++session->stats.nr_unknown_id;
 968                         return 0;
 969                 }
 970                 if (machine == NULL) {
 971                         ++session->stats.nr_unprocessable_samples;
 972                         return 0;
 973                 }
 974                 return perf_session__deliver_sample(session, tool, event,
 975                                                     sample, evsel, machine);
 976         case PERF_RECORD_MMAP:
 977                 return tool->mmap(tool, event, sample, machine);
 978         case PERF_RECORD_COMM:
 979                 return tool->comm(tool, event, sample, machine);
 980         case PERF_RECORD_FORK:
 981                 return tool->fork(tool, event, sample, machine);
 982         case PERF_RECORD_EXIT:
 983                 return tool->exit(tool, event, sample, machine);
 984         case PERF_RECORD_LOST:
 985                 if (tool->lost == perf_event__process_lost)
 986                         session->stats.total_lost += event->lost.lost;
 987                 return tool->lost(tool, event, sample, machine);
 988         case PERF_RECORD_READ:
 989                 return tool->read(tool, event, sample, evsel, machine);
 990         case PERF_RECORD_THROTTLE:
 991                 return tool->throttle(tool, event, sample, machine);
 992         case PERF_RECORD_UNTHROTTLE:
 993                 return tool->unthrottle(tool, event, sample, machine);
 994         default:
 995                 ++session->stats.nr_unknown_events;
 996                 return -1;
 997         }
 998 }
 999
1000 static int perf_session__preprocess_sample(struct perf_session *session,
1001                                            union perf_event *event, struct perf_sample *sample)
1002 {
1003         if (event->header.type != PERF_RECORD_SAMPLE ||
1004             !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN))
1005                 return 0;
1006
1007         if (!ip_callchain__valid(sample->callchain, event)) {
1008                 pr_debug("call-chain problem with event, skipping it.\n");
1009                 ++session->stats.nr_invalid_chains;
1010                 session->stats.total_invalid_chains += sample->period;
1011                 return -EINVAL;
1012         }
1013         return 0;
1014 }
1015
1016 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1017                                             struct perf_tool *tool, u64 file_offset)
1018 {
1019         int err;
1020
1021         dump_event(session, event, file_offset, NULL);
1022
1023         /* These events are processed right away */
1024         switch (event->header.type) {
1025         case PERF_RECORD_HEADER_ATTR:
1026                 err = tool->attr(tool, event, &session->evlist);
1027                 if (err == 0)
1028                         perf_session__set_id_hdr_size(session);
1029                 return err;
1030         case PERF_RECORD_HEADER_TRACING_DATA:
1031                 /* setup for reading amidst mmap */
1032                 lseek(session->fd, file_offset, SEEK_SET);
1033                 return tool->tracing_data(tool, event, session);
1034         case PERF_RECORD_HEADER_BUILD_ID:
1035                 return tool->build_id(tool, event, session);
1036         case PERF_RECORD_FINISHED_ROUND:
1037                 return tool->finished_round(tool, event, session);
1038         default:
1039                 return -EINVAL;
1040         }
1041 }
1042
1043 static void event_swap(union perf_event *event, bool sample_id_all)
1044 {
1045         perf_event__swap_op swap;
1046
1047         swap = perf_event__swap_ops[event->header.type];
1048         if (swap)
1049                 swap(event, sample_id_all);
1050 }
1051
1052 static int perf_session__process_event(struct perf_session *session,
1053                                        union perf_event *event,
1054                                        struct perf_tool *tool,
1055                                        u64 file_offset)
1056 {
1057         struct perf_sample sample;
1058         int ret;
1059
1060         if (session->header.needs_swap)
1061                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1062
1063         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1064                 return -EINVAL;
1065
1066         events_stats__inc(&session->stats, event->header.type);
1067
1068         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1069                 return perf_session__process_user_event(session, event, tool, file_offset);
1070
1071         /*
1072          * For all kernel events we get the sample data
1073          */
1074         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1075         if (ret)
1076                 return ret;
1077
1078         /* Preprocess sample records - precheck callchains */
1079         if (perf_session__preprocess_sample(session, event, &sample))
1080                 return 0;
1081
1082         if (tool->ordered_samples) {
1083                 ret = perf_session_queue_event(session, event, &sample,
1084                                                file_offset);
1085                 if (ret != -ETIME)
1086                         return ret;
1087         }
1088
1089         return perf_session_deliver_event(session, event, &sample, tool,
1090                                           file_offset);
1091 }
1092
1093 void perf_event_header__bswap(struct perf_event_header *self)
1094 {
1095         self->type = bswap_32(self->type);
1096         self->misc = bswap_16(self->misc);
1097         self->size = bswap_16(self->size);
1098 }
1099
1100 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1101 {
1102         return machine__findnew_thread(&session->machines.host, pid);
1103 }
1104
1105 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1106 {
1107         struct thread *thread = perf_session__findnew(self, 0);
1108
1109         if (thread == NULL || thread__set_comm(thread, "swapper")) {
1110                 pr_err("problem inserting idle task.\n");
1111                 thread = NULL;
1112         }
1113
1114         return thread;
1115 }
1116
1117 static void perf_session__warn_about_errors(const struct perf_session *session,
1118                                             const struct perf_tool *tool)
1119 {
1120         if (tool->lost == perf_event__process_lost &&
1121             session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1122                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1123                             "Check IO/CPU overload!\n\n",
1124                             session->stats.nr_events[0],
1125                             session->stats.nr_events[PERF_RECORD_LOST]);
1126         }
1127
1128         if (session->stats.nr_unknown_events != 0) {
1129                 ui__warning("Found %u unknown events!\n\n"
1130                             "Is this an older tool processing a perf.data "
1131                             "file generated by a more recent tool?\n\n"
1132                             "If that is not the case, consider "
1133                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1134                             session->stats.nr_unknown_events);
1135         }
1136
1137         if (session->stats.nr_unknown_id != 0) {
1138                 ui__warning("%u samples with id not present in the header\n",
1139                             session->stats.nr_unknown_id);
1140         }
1141
1142         if (session->stats.nr_invalid_chains != 0) {
1143                 ui__warning("Found invalid callchains!\n\n"
1144                             "%u out of %u events were discarded for this reason.\n\n"
1145                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1146                             session->stats.nr_invalid_chains,
1147                             session->stats.nr_events[PERF_RECORD_SAMPLE]);
1148         }
1149
1150         if (session->stats.nr_unprocessable_samples != 0) {
1151                 ui__warning("%u unprocessable samples recorded.\n"
1152                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1153                             session->stats.nr_unprocessable_samples);
1154         }
1155 }
1156
1157 #define session_done()  (*(volatile int *)(&session_done))
1158 volatile int session_done;
1159
1160 static int __perf_session__process_pipe_events(struct perf_session *self,
1161                                                struct perf_tool *tool)
1162 {
1163         union perf_event *event;
1164         uint32_t size, cur_size = 0;
1165         void *buf = NULL;
1166         int skip = 0;
1167         u64 head;
1168         int err;
1169         void *p;
1170
1171         perf_tool__fill_defaults(tool);
1172
1173         head = 0;
1174         cur_size = sizeof(union perf_event);
1175
1176         buf = malloc(cur_size);
1177         if (!buf)
1178                 return -errno;
1179 more:
1180         event = buf;
1181         err = readn(self->fd, event, sizeof(struct perf_event_header));
1182         if (err <= 0) {
1183                 if (err == 0)
1184                         goto done;
1185
1186                 pr_err("failed to read event header\n");
1187                 goto out_err;
1188         }
1189
1190         if (self->header.needs_swap)
1191                 perf_event_header__bswap(&event->header);
1192
1193         size = event->header.size;
1194         if (size < sizeof(struct perf_event_header)) {
1195                 pr_err("bad event header size\n");
1196                 goto out_err;
1197         }
1198
1199         if (size > cur_size) {
1200                 void *new = realloc(buf, size);
1201                 if (!new) {
1202                         pr_err("failed to allocate memory to read event\n");
1203                         goto out_err;
1204                 }
1205                 buf = new;
1206                 cur_size = size;
1207                 event = buf;
1208         }
1209         p = event;
1210         p += sizeof(struct perf_event_header);
1211
1212         if (size - sizeof(struct perf_event_header)) {
1213                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1214                 if (err <= 0) {
1215                         if (err == 0) {
1216                                 pr_err("unexpected end of event stream\n");
1217                                 goto done;
1218                         }
1219
1220                         pr_err("failed to read event data\n");
1221                         goto out_err;
1222                 }
1223         }
1224
1225         if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1226                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1227                        head, event->header.size, event->header.type);
1228                 err = -EINVAL;
1229                 goto out_err;
1230         }
1231
1232         head += size;
1233
1234         if (skip > 0)
1235                 head += skip;
1236
1237         if (!session_done())
1238                 goto more;
1239 done:
1240         err = 0;
1241 out_err:
1242         free(buf);
1243         perf_session__warn_about_errors(self, tool);
1244         perf_session_free_sample_buffers(self);
1245         return err;
1246 }
1247
1248 static union perf_event *
1249 fetch_mmaped_event(struct perf_session *session,
1250                    u64 head, size_t mmap_size, char *buf)
1251 {
1252         union perf_event *event;
1253
1254         /*
1255          * Ensure we have enough space remaining to read
1256          * the size of the event in the headers.
1257          */
1258         if (head + sizeof(event->header) > mmap_size)
1259                 return NULL;
1260
1261         event = (union perf_event *)(buf + head);
1262
1263         if (session->header.needs_swap)
1264                 perf_event_header__bswap(&event->header);
1265
1266         if (head + event->header.size > mmap_size) {
1267                 /* We're not fetching the event so swap back again */
1268                 if (session->header.needs_swap)
1269                         perf_event_header__bswap(&event->header);
1270                 return NULL;
1271         }
1272
1273         return event;
1274 }
1275
1276 /*
1277  * On 64bit we can mmap the data file in one go. No need for tiny mmap
1278  * slices. On 32bit we use 32MB.
1279  */
1280 #if BITS_PER_LONG == 64
1281 #define MMAP_SIZE ULLONG_MAX
1282 #define NUM_MMAPS 1
1283 #else
1284 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1285 #define NUM_MMAPS 128
1286 #endif
1287
1288 int __perf_session__process_events(struct perf_session *session,
1289                                    u64 data_offset, u64 data_size,
1290                                    u64 file_size, struct perf_tool *tool)
1291 {
1292         u64 head, page_offset, file_offset, file_pos, progress_next;
1293         int err, mmap_prot, mmap_flags, map_idx = 0;
1294         size_t  mmap_size;
1295         char *buf, *mmaps[NUM_MMAPS];
1296         union perf_event *event;
1297         uint32_t size;
1298
1299         perf_tool__fill_defaults(tool);
1300
1301         page_offset = page_size * (data_offset / page_size);
1302         file_offset = page_offset;
1303         head = data_offset - page_offset;
1304
1305         if (data_offset + data_size < file_size)
1306                 file_size = data_offset + data_size;
1307
1308         progress_next = file_size / 16;
1309
1310         mmap_size = MMAP_SIZE;
1311         if (mmap_size > file_size)
1312                 mmap_size = file_size;
1313
1314         memset(mmaps, 0, sizeof(mmaps));
1315
1316         mmap_prot  = PROT_READ;
1317         mmap_flags = MAP_SHARED;
1318
1319         if (session->header.needs_swap) {
1320                 mmap_prot  |= PROT_WRITE;
1321                 mmap_flags = MAP_PRIVATE;
1322         }
1323 remap:
1324         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1325                    file_offset);
1326         if (buf == MAP_FAILED) {
1327                 pr_err("failed to mmap file\n");
1328                 err = -errno;
1329                 goto out_err;
1330         }
1331         mmaps[map_idx] = buf;
1332         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1333         file_pos = file_offset + head;
1334
1335 more:
1336         event = fetch_mmaped_event(session, head, mmap_size, buf);
1337         if (!event) {
1338                 if (mmaps[map_idx]) {
1339                         munmap(mmaps[map_idx], mmap_size);
1340                         mmaps[map_idx] = NULL;
1341                 }
1342
1343                 page_offset = page_size * (head / page_size);
1344                 file_offset += page_offset;
1345                 head -= page_offset;
1346                 goto remap;
1347         }
1348
1349         size = event->header.size;
1350
1351         if (size < sizeof(struct perf_event_header) ||
1352             perf_session__process_event(session, event, tool, file_pos) < 0) {
1353                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1354                        file_offset + head, event->header.size,
1355                        event->header.type);
1356                 err = -EINVAL;
1357                 goto out_err;
1358         }
1359
1360         head += size;
1361         file_pos += size;
1362
1363         if (file_pos >= progress_next) {
1364                 progress_next += file_size / 16;
1365                 ui_progress__update(file_pos, file_size,
1366                                     "Processing events...");
1367         }
1368
1369         if (file_pos < file_size)
1370                 goto more;
1371
1372         err = 0;
1373         /* do the final flush for ordered samples */
1374         session->ordered_samples.next_flush = ULLONG_MAX;
1375         err = flush_sample_queue(session, tool);
1376 out_err:
1377         ui_progress__finish();
1378         perf_session__warn_about_errors(session, tool);
1379         perf_session_free_sample_buffers(session);
1380         return err;
1381 }
1382
1383 int perf_session__process_events(struct perf_session *self,
1384                                  struct perf_tool *tool)
1385 {
1386         int err;
1387
1388         if (perf_session__register_idle_thread(self) == NULL)
1389                 return -ENOMEM;
1390
1391         if (!self->fd_pipe)
1392                 err = __perf_session__process_events(self,
1393                                                      self->header.data_offset,
1394                                                      self->header.data_size,
1395                                                      self->size, tool);
1396         else
1397                 err = __perf_session__process_pipe_events(self, tool);
1398
1399         return err;
1400 }
1401
1402 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1403 {
1404         if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) {
1405                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1406                 return false;
1407         }
1408
1409         return true;
1410 }
1411
1412 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1413                                      const char *symbol_name, u64 addr)
1414 {
1415         char *bracket;
1416         enum map_type i;
1417         struct ref_reloc_sym *ref;
1418
1419         ref = zalloc(sizeof(struct ref_reloc_sym));
1420         if (ref == NULL)
1421                 return -ENOMEM;
1422
1423         ref->name = strdup(symbol_name);
1424         if (ref->name == NULL) {
1425                 free(ref);
1426                 return -ENOMEM;
1427         }
1428
1429         bracket = strchr(ref->name, ']');
1430         if (bracket)
1431                 *bracket = '\0';
1432
1433         ref->addr = addr;
1434
1435         for (i = 0; i < MAP__NR_TYPES; ++i) {
1436                 struct kmap *kmap = map__kmap(maps[i]);
1437                 kmap->ref_reloc_sym = ref;
1438         }
1439
1440         return 0;
1441 }
1442
1443 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1444 {
1445         return machines__fprintf_dsos(&self->machines, fp);
1446 }
1447
1448 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1449                                           bool (skip)(struct dso *dso, int parm), int parm)
1450 {
1451         return machines__fprintf_dsos_buildid(&self->machines, fp, skip, parm);
1452 }
1453
1454 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1455 {
1456         struct perf_evsel *pos;
1457         size_t ret = fprintf(fp, "Aggregated stats:\n");
1458
1459         ret += events_stats__fprintf(&session->stats, fp);
1460
1461         list_for_each_entry(pos, &session->evlist->entries, node) {
1462                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1463                 ret += events_stats__fprintf(&pos->hists.stats, fp);
1464         }
1465
1466         return ret;
1467 }
1468
1469 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1470 {
1471         /*
1472          * FIXME: Here we have to actually print all the machines in this
1473          * session, not just the host...
1474          */
1475         return machine__fprintf(&session->machines.host, fp);
1476 }
1477
1478 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1479                                               unsigned int type)
1480 {
1481         struct perf_evsel *pos;
1482
1483         list_for_each_entry(pos, &session->evlist->entries, node) {
1484                 if (pos->attr.type == type)
1485                         return pos;
1486         }
1487         return NULL;
1488 }
1489
1490 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1491                           struct perf_sample *sample, struct machine *machine,
1492                           int print_sym, int print_dso, int print_symoffset)
1493 {
1494         struct addr_location al;
1495         struct callchain_cursor_node *node;
1496
1497         if (perf_event__preprocess_sample(event, machine, &al, sample,
1498                                           NULL) < 0) {
1499                 error("problem processing %d event, skipping it.\n",
1500                         event->header.type);
1501                 return;
1502         }
1503
1504         if (symbol_conf.use_callchain && sample->callchain) {
1505
1506                 if (machine__resolve_callchain(machine, evsel, al.thread,
1507                                                sample, NULL, NULL) != 0) {
1508                         if (verbose)
1509                                 error("Failed to resolve callchain. Skipping\n");
1510                         return;
1511                 }
1512                 callchain_cursor_commit(&callchain_cursor);
1513
1514                 while (1) {
1515                         node = callchain_cursor_current(&callchain_cursor);
1516                         if (!node)
1517                                 break;
1518
1519                         printf("\t%16" PRIx64, node->ip);
1520                         if (print_sym) {
1521                                 printf(" ");
1522                                 symbol__fprintf_symname(node->sym, stdout);
1523                         }
1524                         if (print_dso) {
1525                                 printf(" (");
1526                                 map__fprintf_dsoname(node->map, stdout);
1527                                 printf(")");
1528                         }
1529                         printf("\n");
1530
1531                         callchain_cursor_advance(&callchain_cursor);
1532                 }
1533
1534         } else {
1535                 printf("%16" PRIx64, sample->ip);
1536                 if (print_sym) {
1537                         printf(" ");
1538                         if (print_symoffset)
1539                                 symbol__fprintf_symname_offs(al.sym, &al,
1540                                                              stdout);
1541                         else
1542                                 symbol__fprintf_symname(al.sym, stdout);
1543                 }
1544
1545                 if (print_dso) {
1546                         printf(" (");
1547                         map__fprintf_dsoname(al.map, stdout);
1548                         printf(")");
1549                 }
1550         }
1551 }
1552
1553 int perf_session__cpu_bitmap(struct perf_session *session,
1554                              const char *cpu_list, unsigned long *cpu_bitmap)
1555 {
1556         int i;
1557         struct cpu_map *map;
1558
1559         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1560                 struct perf_evsel *evsel;
1561
1562                 evsel = perf_session__find_first_evtype(session, i);
1563                 if (!evsel)
1564                         continue;
1565
1566                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1567                         pr_err("File does not contain CPU events. "
1568                                "Remove -c option to proceed.\n");
1569                         return -1;
1570                 }
1571         }
1572
1573         map = cpu_map__new(cpu_list);
1574         if (map == NULL) {
1575                 pr_err("Invalid cpu_list\n");
1576                 return -1;
1577         }
1578
1579         for (i = 0; i < map->nr; i++) {
1580                 int cpu = map->map[i];
1581
1582                 if (cpu >= MAX_NR_CPUS) {
1583                         pr_err("Requested CPU %d too large. "
1584                                "Consider raising MAX_NR_CPUS\n", cpu);
1585                         return -1;
1586                 }
1587
1588                 set_bit(cpu, cpu_bitmap);
1589         }
1590
1591         return 0;
1592 }
1593
1594 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1595                                 bool full)
1596 {
1597         struct stat st;
1598         int ret;
1599
1600         if (session == NULL || fp == NULL)
1601                 return;
1602
1603         ret = fstat(session->fd, &st);
1604         if (ret == -1)
1605                 return;
1606
1607         fprintf(fp, "# ========\n");
1608         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1609         perf_header__fprintf_info(session, fp, full);
1610         fprintf(fp, "# ========\n#\n");
1611 }
1612
1613
1614 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1615                                              const struct perf_evsel_str_handler *assocs,
1616                                              size_t nr_assocs)
1617 {
1618         struct perf_evlist *evlist = session->evlist;
1619         struct event_format *format;
1620         struct perf_evsel *evsel;
1621         char *tracepoint, *name;
1622         size_t i;
1623         int err;
1624
1625         for (i = 0; i < nr_assocs; i++) {
1626                 err = -ENOMEM;
1627                 tracepoint = strdup(assocs[i].name);
1628                 if (tracepoint == NULL)
1629                         goto out;
1630
1631                 err = -ENOENT;
1632                 name = strchr(tracepoint, ':');
1633                 if (name == NULL)
1634                         goto out_free;
1635
1636                 *name++ = '\0';
1637                 format = pevent_find_event_by_name(session->pevent,
1638                                                    tracepoint, name);
1639                 if (format == NULL) {
1640                         /*
1641                          * Adding a handler for an event not in the session,
1642                          * just ignore it.
1643                          */
1644                         goto next;
1645                 }
1646
1647                 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1648                 if (evsel == NULL)
1649                         goto next;
1650
1651                 err = -EEXIST;
1652                 if (evsel->handler.func != NULL)
1653                         goto out_free;
1654                 evsel->handler.func = assocs[i].handler;
1655 next:
1656                 free(tracepoint);
1657         }
1658
1659         err = 0;
1660 out:
1661         return err;
1662
1663 out_free:
1664         free(tracepoint);
1665         goto out;
1666 }