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 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->mmap2 == NULL)
 260                 tool->mmap2 = process_event_stub;
 261         if (tool->comm == NULL)
 262                 tool->comm = process_event_stub;
 263         if (tool->fork == NULL)
 264                 tool->fork = process_event_stub;
 265         if (tool->exit == NULL)
 266                 tool->exit = process_event_stub;
 267         if (tool->lost == NULL)
 268                 tool->lost = perf_event__process_lost;
 269         if (tool->read == NULL)
 270                 tool->read = process_event_sample_stub;
 271         if (tool->throttle == NULL)
 272                 tool->throttle = process_event_stub;
 273         if (tool->unthrottle == NULL)
 274                 tool->unthrottle = process_event_stub;
 275         if (tool->attr == NULL)
 276                 tool->attr = process_event_synth_attr_stub;
 277         if (tool->tracing_data == NULL)
 278                 tool->tracing_data = process_event_synth_tracing_data_stub;
 279         if (tool->build_id == NULL)
 280                 tool->build_id = process_finished_round_stub;
 281         if (tool->finished_round == NULL) {
 282                 if (tool->ordered_samples)
 283                         tool->finished_round = process_finished_round;
 284                 else
 285                         tool->finished_round = process_finished_round_stub;
 286         }
 287 }
 288
 289 void mem_bswap_32(void *src, int byte_size)
 290 {
 291         u32 *m = src;
 292         while (byte_size > 0) {
 293                 *m = bswap_32(*m);
 294                 byte_size -= sizeof(u32);
 295                 ++m;
 296         }
 297 }
 298
 299 void mem_bswap_64(void *src, int byte_size)
 300 {
 301         u64 *m = src;
 302
 303         while (byte_size > 0) {
 304                 *m = bswap_64(*m);
 305                 byte_size -= sizeof(u64);
 306                 ++m;
 307         }
 308 }
 309
 310 static void swap_sample_id_all(union perf_event *event, void *data)
 311 {
 312         void *end = (void *) event + event->header.size;
 313         int size = end - data;
 314
 315         BUG_ON(size % sizeof(u64));
 316         mem_bswap_64(data, size);
 317 }
 318
 319 static void perf_event__all64_swap(union perf_event *event,
 320                                    bool sample_id_all __maybe_unused)
 321 {
 322         struct perf_event_header *hdr = &event->header;
 323         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 324 }
 325
 326 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 327 {
 328         event->comm.pid = bswap_32(event->comm.pid);
 329         event->comm.tid = bswap_32(event->comm.tid);
 330
 331         if (sample_id_all) {
 332                 void *data = &event->comm.comm;
 333
 334                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 335                 swap_sample_id_all(event, data);
 336         }
 337 }
 338
 339 static void perf_event__mmap_swap(union perf_event *event,
 340                                   bool sample_id_all)
 341 {
 342         event->mmap.pid   = bswap_32(event->mmap.pid);
 343         event->mmap.tid   = bswap_32(event->mmap.tid);
 344         event->mmap.start = bswap_64(event->mmap.start);
 345         event->mmap.len   = bswap_64(event->mmap.len);
 346         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 347
 348         if (sample_id_all) {
 349                 void *data = &event->mmap.filename;
 350
 351                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 352                 swap_sample_id_all(event, data);
 353         }
 354 }
 355
 356 static void perf_event__mmap2_swap(union perf_event *event,
 357                                   bool sample_id_all)
 358 {
 359         event->mmap2.pid   = bswap_32(event->mmap2.pid);
 360         event->mmap2.tid   = bswap_32(event->mmap2.tid);
 361         event->mmap2.start = bswap_64(event->mmap2.start);
 362         event->mmap2.len   = bswap_64(event->mmap2.len);
 363         event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
 364         event->mmap2.maj   = bswap_32(event->mmap2.maj);
 365         event->mmap2.min   = bswap_32(event->mmap2.min);
 366         event->mmap2.ino   = bswap_64(event->mmap2.ino);
 367
 368         if (sample_id_all) {
 369                 void *data = &event->mmap2.filename;
 370
 371                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 372                 swap_sample_id_all(event, data);
 373         }
 374 }
 375 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 376 {
 377         event->fork.pid  = bswap_32(event->fork.pid);
 378         event->fork.tid  = bswap_32(event->fork.tid);
 379         event->fork.ppid = bswap_32(event->fork.ppid);
 380         event->fork.ptid = bswap_32(event->fork.ptid);
 381         event->fork.time = bswap_64(event->fork.time);
 382
 383         if (sample_id_all)
 384                 swap_sample_id_all(event, &event->fork + 1);
 385 }
 386
 387 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 388 {
 389         event->read.pid          = bswap_32(event->read.pid);
 390         event->read.tid          = bswap_32(event->read.tid);
 391         event->read.value        = bswap_64(event->read.value);
 392         event->read.time_enabled = bswap_64(event->read.time_enabled);
 393         event->read.time_running = bswap_64(event->read.time_running);
 394         event->read.id           = bswap_64(event->read.id);
 395
 396         if (sample_id_all)
 397                 swap_sample_id_all(event, &event->read + 1);
 398 }
 399
 400 static void perf_event__throttle_swap(union perf_event *event,
 401                                       bool sample_id_all)
 402 {
 403         event->throttle.time      = bswap_64(event->throttle.time);
 404         event->throttle.id        = bswap_64(event->throttle.id);
 405         event->throttle.stream_id = bswap_64(event->throttle.stream_id);
 406
 407         if (sample_id_all)
 408                 swap_sample_id_all(event, &event->throttle + 1);
 409 }
 410
 411 static u8 revbyte(u8 b)
 412 {
 413         int rev = (b >> 4) | ((b & 0xf) << 4);
 414         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 415         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 416         return (u8) rev;
 417 }
 418
 419 /*
 420  * XXX this is hack in attempt to carry flags bitfield
 421  * throught endian village. ABI says:
 422  *
 423  * Bit-fields are allocated from right to left (least to most significant)
 424  * on little-endian implementations and from left to right (most to least
 425  * significant) on big-endian implementations.
 426  *
 427  * The above seems to be byte specific, so we need to reverse each
 428  * byte of the bitfield. 'Internet' also says this might be implementation
 429  * specific and we probably need proper fix and carry perf_event_attr
 430  * bitfield flags in separate data file FEAT_ section. Thought this seems
 431  * to work for now.
 432  */
 433 static void swap_bitfield(u8 *p, unsigned len)
 434 {
 435         unsigned i;
 436
 437         for (i = 0; i < len; i++) {
 438                 *p = revbyte(*p);
 439                 p++;
 440         }
 441 }
 442
 443 /* exported for swapping attributes in file header */
 444 void perf_event__attr_swap(struct perf_event_attr *attr)
 445 {
 446         attr->type              = bswap_32(attr->type);
 447         attr->size              = bswap_32(attr->size);
 448         attr->config            = bswap_64(attr->config);
 449         attr->sample_period     = bswap_64(attr->sample_period);
 450         attr->sample_type       = bswap_64(attr->sample_type);
 451         attr->read_format       = bswap_64(attr->read_format);
 452         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 453         attr->bp_type           = bswap_32(attr->bp_type);
 454         attr->bp_addr           = bswap_64(attr->bp_addr);
 455         attr->bp_len            = bswap_64(attr->bp_len);
 456
 457         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 458 }
 459
 460 static void perf_event__hdr_attr_swap(union perf_event *event,
 461                                       bool sample_id_all __maybe_unused)
 462 {
 463         size_t size;
 464
 465         perf_event__attr_swap(&event->attr.attr);
 466
 467         size = event->header.size;
 468         size -= (void *)&event->attr.id - (void *)event;
 469         mem_bswap_64(event->attr.id, size);
 470 }
 471
 472 static void perf_event__event_type_swap(union perf_event *event,
 473                                         bool sample_id_all __maybe_unused)
 474 {
 475         event->event_type.event_type.event_id =
 476                 bswap_64(event->event_type.event_type.event_id);
 477 }
 478
 479 static void perf_event__tracing_data_swap(union perf_event *event,
 480                                           bool sample_id_all __maybe_unused)
 481 {
 482         event->tracing_data.size = bswap_32(event->tracing_data.size);
 483 }
 484
 485 typedef void (*perf_event__swap_op)(union perf_event *event,
 486                                     bool sample_id_all);
 487
 488 static perf_event__swap_op perf_event__swap_ops[] = {
 489         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 490         [PERF_RECORD_MMAP2]               = perf_event__mmap2_swap,
 491         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 492         [PERF_RECORD_FORK]                = perf_event__task_swap,
 493         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 494         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 495         [PERF_RECORD_READ]                = perf_event__read_swap,
 496         [PERF_RECORD_THROTTLE]            = perf_event__throttle_swap,
 497         [PERF_RECORD_UNTHROTTLE]          = perf_event__throttle_swap,
 498         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 499         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 500         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 501         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 502         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 503         [PERF_RECORD_HEADER_MAX]          = NULL,
 504 };
 505
 506 struct sample_queue {
 507         u64                     timestamp;
 508         u64                     file_offset;
 509         union perf_event        *event;
 510         struct list_head        list;
 511 };
 512
 513 static void perf_session_free_sample_buffers(struct perf_session *session)
 514 {
 515         struct ordered_samples *os = &session->ordered_samples;
 516
 517         while (!list_empty(&os->to_free)) {
 518                 struct sample_queue *sq;
 519
 520                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 521                 list_del(&sq->list);
 522                 free(sq);
 523         }
 524 }
 525
 526 static int perf_session_deliver_event(struct perf_session *session,
 527                                       union perf_event *event,
 528                                       struct perf_sample *sample,
 529                                       struct perf_tool *tool,
 530                                       u64 file_offset);
 531
 532 static int flush_sample_queue(struct perf_session *s,
 533                        struct perf_tool *tool)
 534 {
 535         struct ordered_samples *os = &s->ordered_samples;
 536         struct list_head *head = &os->samples;
 537         struct sample_queue *tmp, *iter;
 538         struct perf_sample sample;
 539         u64 limit = os->next_flush;
 540         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 541         unsigned idx = 0, progress_next = os->nr_samples / 16;
 542         bool show_progress = limit == ULLONG_MAX;
 543         int ret;
 544
 545         if (!tool->ordered_samples || !limit)
 546                 return 0;
 547
 548         list_for_each_entry_safe(iter, tmp, head, list) {
 549                 if (session_done())
 550                         return 0;
 551
 552                 if (iter->timestamp > limit)
 553                         break;
 554
 555                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
 556                 if (ret)
 557                         pr_err("Can't parse sample, err = %d\n", ret);
 558                 else {
 559                         ret = perf_session_deliver_event(s, iter->event, &sample, tool,
 560                                                          iter->file_offset);
 561                         if (ret)
 562                                 return ret;
 563                 }
 564
 565                 os->last_flush = iter->timestamp;
 566                 list_del(&iter->list);
 567                 list_add(&iter->list, &os->sample_cache);
 568                 if (show_progress && (++idx >= progress_next)) {
 569                         progress_next += os->nr_samples / 16;
 570                         ui_progress__update(idx, os->nr_samples,
 571                                             "Processing time ordered events...");
 572                 }
 573         }
 574
 575         if (list_empty(head)) {
 576                 os->last_sample = NULL;
 577         } else if (last_ts <= limit) {
 578                 os->last_sample =
 579                         list_entry(head->prev, struct sample_queue, list);
 580         }
 581
 582         os->nr_samples = 0;
 583
 584         return 0;
 585 }
 586
 587 /*
 588  * When perf record finishes a pass on every buffers, it records this pseudo
 589  * event.
 590  * We record the max timestamp t found in the pass n.
 591  * Assuming these timestamps are monotonic across cpus, we know that if
 592  * a buffer still has events with timestamps below t, they will be all
 593  * available and then read in the pass n + 1.
 594  * Hence when we start to read the pass n + 2, we can safely flush every
 595  * events with timestamps below t.
 596  *
 597  *    ============ PASS n =================
 598  *       CPU 0         |   CPU 1
 599  *                     |
 600  *    cnt1 timestamps  |   cnt2 timestamps
 601  *          1          |         2
 602  *          2          |         3
 603  *          -          |         4  <--- max recorded
 604  *
 605  *    ============ PASS n + 1 ==============
 606  *       CPU 0         |   CPU 1
 607  *                     |
 608  *    cnt1 timestamps  |   cnt2 timestamps
 609  *          3          |         5
 610  *          4          |         6
 611  *          5          |         7 <---- max recorded
 612  *
 613  *      Flush every events below timestamp 4
 614  *
 615  *    ============ PASS n + 2 ==============
 616  *       CPU 0         |   CPU 1
 617  *                     |
 618  *    cnt1 timestamps  |   cnt2 timestamps
 619  *          6          |         8
 620  *          7          |         9
 621  *          -          |         10
 622  *
 623  *      Flush every events below timestamp 7
 624  *      etc...
 625  */
 626 static int process_finished_round(struct perf_tool *tool,
 627                                   union perf_event *event __maybe_unused,
 628                                   struct perf_session *session)
 629 {
 630         int ret = flush_sample_queue(session, tool);
 631         if (!ret)
 632                 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 633
 634         return ret;
 635 }
 636
 637 /* The queue is ordered by time */
 638 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 639 {
 640         struct ordered_samples *os = &s->ordered_samples;
 641         struct sample_queue *sample = os->last_sample;
 642         u64 timestamp = new->timestamp;
 643         struct list_head *p;
 644
 645         ++os->nr_samples;
 646         os->last_sample = new;
 647
 648         if (!sample) {
 649                 list_add(&new->list, &os->samples);
 650                 os->max_timestamp = timestamp;
 651                 return;
 652         }
 653
 654         /*
 655          * last_sample might point to some random place in the list as it's
 656          * the last queued event. We expect that the new event is close to
 657          * this.
 658          */
 659         if (sample->timestamp <= timestamp) {
 660                 while (sample->timestamp <= timestamp) {
 661                         p = sample->list.next;
 662                         if (p == &os->samples) {
 663                                 list_add_tail(&new->list, &os->samples);
 664                                 os->max_timestamp = timestamp;
 665                                 return;
 666                         }
 667                         sample = list_entry(p, struct sample_queue, list);
 668                 }
 669                 list_add_tail(&new->list, &sample->list);
 670         } else {
 671                 while (sample->timestamp > timestamp) {
 672                         p = sample->list.prev;
 673                         if (p == &os->samples) {
 674                                 list_add(&new->list, &os->samples);
 675                                 return;
 676                         }
 677                         sample = list_entry(p, struct sample_queue, list);
 678                 }
 679                 list_add(&new->list, &sample->list);
 680         }
 681 }
 682
 683 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 684
 685 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 686                                     struct perf_sample *sample, u64 file_offset)
 687 {
 688         struct ordered_samples *os = &s->ordered_samples;
 689         struct list_head *sc = &os->sample_cache;
 690         u64 timestamp = sample->time;
 691         struct sample_queue *new;
 692
 693         if (!timestamp || timestamp == ~0ULL)
 694                 return -ETIME;
 695
 696         if (timestamp < s->ordered_samples.last_flush) {
 697                 printf("Warning: Timestamp below last timeslice flush\n");
 698                 return -EINVAL;
 699         }
 700
 701         if (!list_empty(sc)) {
 702                 new = list_entry(sc->next, struct sample_queue, list);
 703                 list_del(&new->list);
 704         } else if (os->sample_buffer) {
 705                 new = os->sample_buffer + os->sample_buffer_idx;
 706                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 707                         os->sample_buffer = NULL;
 708         } else {
 709                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 710                 if (!os->sample_buffer)
 711                         return -ENOMEM;
 712                 list_add(&os->sample_buffer->list, &os->to_free);
 713                 os->sample_buffer_idx = 2;
 714                 new = os->sample_buffer + 1;
 715         }
 716
 717         new->timestamp = timestamp;
 718         new->file_offset = file_offset;
 719         new->event = event;
 720
 721         __queue_event(new, s);
 722
 723         return 0;
 724 }
 725
 726 static void callchain__printf(struct perf_sample *sample)
 727 {
 728         unsigned int i;
 729
 730         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 731
 732         for (i = 0; i < sample->callchain->nr; i++)
 733                 printf("..... %2d: %016" PRIx64 "\n",
 734                        i, sample->callchain->ips[i]);
 735 }
 736
 737 static void branch_stack__printf(struct perf_sample *sample)
 738 {
 739         uint64_t i;
 740
 741         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 742
 743         for (i = 0; i < sample->branch_stack->nr; i++)
 744                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 745                         i, sample->branch_stack->entries[i].from,
 746                         sample->branch_stack->entries[i].to);
 747 }
 748
 749 static void regs_dump__printf(u64 mask, u64 *regs)
 750 {
 751         unsigned rid, i = 0;
 752
 753         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 754                 u64 val = regs[i++];
 755
 756                 printf(".... %-5s 0x%" PRIx64 "\n",
 757                        perf_reg_name(rid), val);
 758         }
 759 }
 760
 761 static void regs_user__printf(struct perf_sample *sample, u64 mask)
 762 {
 763         struct regs_dump *user_regs = &sample->user_regs;
 764
 765         if (user_regs->regs) {
 766                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
 767                 regs_dump__printf(mask, user_regs->regs);
 768         }
 769 }
 770
 771 static void stack_user__printf(struct stack_dump *dump)
 772 {
 773         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 774                dump->size, dump->offset);
 775 }
 776
 777 static void perf_session__print_tstamp(struct perf_session *session,
 778                                        union perf_event *event,
 779                                        struct perf_sample *sample)
 780 {
 781         u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
 782
 783         if (event->header.type != PERF_RECORD_SAMPLE &&
 784             !perf_evlist__sample_id_all(session->evlist)) {
 785                 fputs("-1 -1 ", stdout);
 786                 return;
 787         }
 788
 789         if ((sample_type & PERF_SAMPLE_CPU))
 790                 printf("%u ", sample->cpu);
 791
 792         if (sample_type & PERF_SAMPLE_TIME)
 793                 printf("%" PRIu64 " ", sample->time);
 794 }
 795
 796 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
 797 {
 798         printf("... sample_read:\n");
 799
 800         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 801                 printf("...... time enabled %016" PRIx64 "\n",
 802                        sample->read.time_enabled);
 803
 804         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 805                 printf("...... time running %016" PRIx64 "\n",
 806                        sample->read.time_running);
 807
 808         if (read_format & PERF_FORMAT_GROUP) {
 809                 u64 i;
 810
 811                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
 812
 813                 for (i = 0; i < sample->read.group.nr; i++) {
 814                         struct sample_read_value *value;
 815
 816                         value = &sample->read.group.values[i];
 817                         printf("..... id %016" PRIx64
 818                                ", value %016" PRIx64 "\n",
 819                                value->id, value->value);
 820                 }
 821         } else
 822                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
 823                         sample->read.one.id, sample->read.one.value);
 824 }
 825
 826 static void dump_event(struct perf_session *session, union perf_event *event,
 827                        u64 file_offset, struct perf_sample *sample)
 828 {
 829         if (!dump_trace)
 830                 return;
 831
 832         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 833                file_offset, event->header.size, event->header.type);
 834
 835         trace_event(event);
 836
 837         if (sample)
 838                 perf_session__print_tstamp(session, event, sample);
 839
 840         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 841                event->header.size, perf_event__name(event->header.type));
 842 }
 843
 844 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 845                         struct perf_sample *sample)
 846 {
 847         u64 sample_type;
 848
 849         if (!dump_trace)
 850                 return;
 851
 852         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 853                event->header.misc, sample->pid, sample->tid, sample->ip,
 854                sample->period, sample->addr);
 855
 856         sample_type = evsel->attr.sample_type;
 857
 858         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 859                 callchain__printf(sample);
 860
 861         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
 862                 branch_stack__printf(sample);
 863
 864         if (sample_type & PERF_SAMPLE_REGS_USER)
 865                 regs_user__printf(sample, evsel->attr.sample_regs_user);
 866
 867         if (sample_type & PERF_SAMPLE_STACK_USER)
 868                 stack_user__printf(&sample->user_stack);
 869
 870         if (sample_type & PERF_SAMPLE_WEIGHT)
 871                 printf("... weight: %" PRIu64 "\n", sample->weight);
 872
 873         if (sample_type & PERF_SAMPLE_DATA_SRC)
 874                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
 875
 876         if (sample_type & PERF_SAMPLE_TRANSACTION)
 877                 printf("... transaction: %" PRIx64 "\n", sample->transaction);
 878
 879         if (sample_type & PERF_SAMPLE_READ)
 880                 sample_read__printf(sample, evsel->attr.read_format);
 881 }
 882
 883 static struct machine *
 884         perf_session__find_machine_for_cpumode(struct perf_session *session,
 885                                                union perf_event *event,
 886                                                struct perf_sample *sample)
 887 {
 888         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 889
 890         if (perf_guest &&
 891             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
 892              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
 893                 u32 pid;
 894
 895                 if (event->header.type == PERF_RECORD_MMAP
 896                     || event->header.type == PERF_RECORD_MMAP2)
 897                         pid = event->mmap.pid;
 898                 else
 899                         pid = sample->pid;
 900
 901                 return perf_session__findnew_machine(session, pid);
 902         }
 903
 904         return &session->machines.host;
 905 }
 906
 907 static int deliver_sample_value(struct perf_session *session,
 908                                 struct perf_tool *tool,
 909                                 union perf_event *event,
 910                                 struct perf_sample *sample,
 911                                 struct sample_read_value *v,
 912                                 struct machine *machine)
 913 {
 914         struct perf_sample_id *sid;
 915
 916         sid = perf_evlist__id2sid(session->evlist, v->id);
 917         if (sid) {
 918                 sample->id     = v->id;
 919                 sample->period = v->value - sid->period;
 920                 sid->period    = v->value;
 921         }
 922
 923         if (!sid || sid->evsel == NULL) {
 924                 ++session->stats.nr_unknown_id;
 925                 return 0;
 926         }
 927
 928         return tool->sample(tool, event, sample, sid->evsel, machine);
 929 }
 930
 931 static int deliver_sample_group(struct perf_session *session,
 932                                 struct perf_tool *tool,
 933                                 union  perf_event *event,
 934                                 struct perf_sample *sample,
 935                                 struct machine *machine)
 936 {
 937         int ret = -EINVAL;
 938         u64 i;
 939
 940         for (i = 0; i < sample->read.group.nr; i++) {
 941                 ret = deliver_sample_value(session, tool, event, sample,
 942                                            &sample->read.group.values[i],
 943                                            machine);
 944                 if (ret)
 945                         break;
 946         }
 947
 948         return ret;
 949 }
 950
 951 static int
 952 perf_session__deliver_sample(struct perf_session *session,
 953                              struct perf_tool *tool,
 954                              union  perf_event *event,
 955                              struct perf_sample *sample,
 956                              struct perf_evsel *evsel,
 957                              struct machine *machine)
 958 {
 959         /* We know evsel != NULL. */
 960         u64 sample_type = evsel->attr.sample_type;
 961         u64 read_format = evsel->attr.read_format;
 962
 963         /* Standard sample delievery. */
 964         if (!(sample_type & PERF_SAMPLE_READ))
 965                 return tool->sample(tool, event, sample, evsel, machine);
 966
 967         /* For PERF_SAMPLE_READ we have either single or group mode. */
 968         if (read_format & PERF_FORMAT_GROUP)
 969                 return deliver_sample_group(session, tool, event, sample,
 970                                             machine);
 971         else
 972                 return deliver_sample_value(session, tool, event, sample,
 973                                             &sample->read.one, machine);
 974 }
 975
 976 static int perf_session_deliver_event(struct perf_session *session,
 977                                       union perf_event *event,
 978                                       struct perf_sample *sample,
 979                                       struct perf_tool *tool,
 980                                       u64 file_offset)
 981 {
 982         struct perf_evsel *evsel;
 983         struct machine *machine;
 984
 985         dump_event(session, event, file_offset, sample);
 986
 987         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 988         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
 989                 /*
 990                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
 991                  * because the tools right now may apply filters, discarding
 992                  * some of the samples. For consistency, in the future we
 993                  * should have something like nr_filtered_samples and remove
 994                  * the sample->period from total_sample_period, etc, KISS for
 995                  * now tho.
 996                  *
 997                  * Also testing against NULL allows us to handle files without
 998                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
 999                  * future probably it'll be a good idea to restrict event
1000                  * processing via perf_session to files with both set.
1001                  */
1002                 hists__inc_nr_events(&evsel->hists, event->header.type);
1003         }
1004
1005         machine = perf_session__find_machine_for_cpumode(session, event,
1006                                                          sample);
1007
1008         switch (event->header.type) {
1009         case PERF_RECORD_SAMPLE:
1010                 dump_sample(evsel, event, sample);
1011                 if (evsel == NULL) {
1012                         ++session->stats.nr_unknown_id;
1013                         return 0;
1014                 }
1015                 if (machine == NULL) {
1016                         ++session->stats.nr_unprocessable_samples;
1017                         return 0;
1018                 }
1019                 return perf_session__deliver_sample(session, tool, event,
1020                                                     sample, evsel, machine);
1021         case PERF_RECORD_MMAP:
1022                 return tool->mmap(tool, event, sample, machine);
1023         case PERF_RECORD_MMAP2:
1024                 return tool->mmap2(tool, event, sample, machine);
1025         case PERF_RECORD_COMM:
1026                 return tool->comm(tool, event, sample, machine);
1027         case PERF_RECORD_FORK:
1028                 return tool->fork(tool, event, sample, machine);
1029         case PERF_RECORD_EXIT:
1030                 return tool->exit(tool, event, sample, machine);
1031         case PERF_RECORD_LOST:
1032                 if (tool->lost == perf_event__process_lost)
1033                         session->stats.total_lost += event->lost.lost;
1034                 return tool->lost(tool, event, sample, machine);
1035         case PERF_RECORD_READ:
1036                 return tool->read(tool, event, sample, evsel, machine);
1037         case PERF_RECORD_THROTTLE:
1038                 return tool->throttle(tool, event, sample, machine);
1039         case PERF_RECORD_UNTHROTTLE:
1040                 return tool->unthrottle(tool, event, sample, machine);
1041         default:
1042                 ++session->stats.nr_unknown_events;
1043                 return -1;
1044         }
1045 }
1046
1047 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1048                                             struct perf_tool *tool, u64 file_offset)
1049 {
1050         int err;
1051
1052         dump_event(session, event, file_offset, NULL);
1053
1054         /* These events are processed right away */
1055         switch (event->header.type) {
1056         case PERF_RECORD_HEADER_ATTR:
1057                 err = tool->attr(tool, event, &session->evlist);
1058                 if (err == 0)
1059                         perf_session__set_id_hdr_size(session);
1060                 return err;
1061         case PERF_RECORD_HEADER_TRACING_DATA:
1062                 /* setup for reading amidst mmap */
1063                 lseek(session->fd, file_offset, SEEK_SET);
1064                 return tool->tracing_data(tool, event, session);
1065         case PERF_RECORD_HEADER_BUILD_ID:
1066                 return tool->build_id(tool, event, session);
1067         case PERF_RECORD_FINISHED_ROUND:
1068                 return tool->finished_round(tool, event, session);
1069         default:
1070                 return -EINVAL;
1071         }
1072 }
1073
1074 static void event_swap(union perf_event *event, bool sample_id_all)
1075 {
1076         perf_event__swap_op swap;
1077
1078         swap = perf_event__swap_ops[event->header.type];
1079         if (swap)
1080                 swap(event, sample_id_all);
1081 }
1082
1083 static int perf_session__process_event(struct perf_session *session,
1084                                        union perf_event *event,
1085                                        struct perf_tool *tool,
1086                                        u64 file_offset)
1087 {
1088         struct perf_sample sample;
1089         int ret;
1090
1091         if (session->header.needs_swap)
1092                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1093
1094         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1095                 return -EINVAL;
1096
1097         events_stats__inc(&session->stats, event->header.type);
1098
1099         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1100                 return perf_session__process_user_event(session, event, tool, file_offset);
1101
1102         /*
1103          * For all kernel events we get the sample data
1104          */
1105         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1106         if (ret)
1107                 return ret;
1108
1109         if (tool->ordered_samples) {
1110                 ret = perf_session_queue_event(session, event, &sample,
1111                                                file_offset);
1112                 if (ret != -ETIME)
1113                         return ret;
1114         }
1115
1116         return perf_session_deliver_event(session, event, &sample, tool,
1117                                           file_offset);
1118 }
1119
1120 void perf_event_header__bswap(struct perf_event_header *self)
1121 {
1122         self->type = bswap_32(self->type);
1123         self->misc = bswap_16(self->misc);
1124         self->size = bswap_16(self->size);
1125 }
1126
1127 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1128 {
1129         return machine__findnew_thread(&session->machines.host, 0, pid);
1130 }
1131
1132 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1133 {
1134         struct thread *thread = perf_session__findnew(self, 0);
1135
1136         if (thread == NULL || thread__set_comm(thread, "swapper")) {
1137                 pr_err("problem inserting idle task.\n");
1138                 thread = NULL;
1139         }
1140
1141         return thread;
1142 }
1143
1144 static void perf_session__warn_about_errors(const struct perf_session *session,
1145                                             const struct perf_tool *tool)
1146 {
1147         if (tool->lost == perf_event__process_lost &&
1148             session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1149                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1150                             "Check IO/CPU overload!\n\n",
1151                             session->stats.nr_events[0],
1152                             session->stats.nr_events[PERF_RECORD_LOST]);
1153         }
1154
1155         if (session->stats.nr_unknown_events != 0) {
1156                 ui__warning("Found %u unknown events!\n\n"
1157                             "Is this an older tool processing a perf.data "
1158                             "file generated by a more recent tool?\n\n"
1159                             "If that is not the case, consider "
1160                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1161                             session->stats.nr_unknown_events);
1162         }
1163
1164         if (session->stats.nr_unknown_id != 0) {
1165                 ui__warning("%u samples with id not present in the header\n",
1166                             session->stats.nr_unknown_id);
1167         }
1168
1169         if (session->stats.nr_invalid_chains != 0) {
1170                 ui__warning("Found invalid callchains!\n\n"
1171                             "%u out of %u events were discarded for this reason.\n\n"
1172                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1173                             session->stats.nr_invalid_chains,
1174                             session->stats.nr_events[PERF_RECORD_SAMPLE]);
1175         }
1176
1177         if (session->stats.nr_unprocessable_samples != 0) {
1178                 ui__warning("%u unprocessable samples recorded.\n"
1179                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1180                             session->stats.nr_unprocessable_samples);
1181         }
1182 }
1183
1184 volatile int session_done;
1185
1186 static int __perf_session__process_pipe_events(struct perf_session *self,
1187                                                struct perf_tool *tool)
1188 {
1189         union perf_event *event;
1190         uint32_t size, cur_size = 0;
1191         void *buf = NULL;
1192         int skip = 0;
1193         u64 head;
1194         int err;
1195         void *p;
1196
1197         perf_tool__fill_defaults(tool);
1198
1199         head = 0;
1200         cur_size = sizeof(union perf_event);
1201
1202         buf = malloc(cur_size);
1203         if (!buf)
1204                 return -errno;
1205 more:
1206         event = buf;
1207         err = readn(self->fd, event, sizeof(struct perf_event_header));
1208         if (err <= 0) {
1209                 if (err == 0)
1210                         goto done;
1211
1212                 pr_err("failed to read event header\n");
1213                 goto out_err;
1214         }
1215
1216         if (self->header.needs_swap)
1217                 perf_event_header__bswap(&event->header);
1218
1219         size = event->header.size;
1220         if (size < sizeof(struct perf_event_header)) {
1221                 pr_err("bad event header size\n");
1222                 goto out_err;
1223         }
1224
1225         if (size > cur_size) {
1226                 void *new = realloc(buf, size);
1227                 if (!new) {
1228                         pr_err("failed to allocate memory to read event\n");
1229                         goto out_err;
1230                 }
1231                 buf = new;
1232                 cur_size = size;
1233                 event = buf;
1234         }
1235         p = event;
1236         p += sizeof(struct perf_event_header);
1237
1238         if (size - sizeof(struct perf_event_header)) {
1239                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1240                 if (err <= 0) {
1241                         if (err == 0) {
1242                                 pr_err("unexpected end of event stream\n");
1243                                 goto done;
1244                         }
1245
1246                         pr_err("failed to read event data\n");
1247                         goto out_err;
1248                 }
1249         }
1250
1251         if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1252                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1253                        head, event->header.size, event->header.type);
1254                 err = -EINVAL;
1255                 goto out_err;
1256         }
1257
1258         head += size;
1259
1260         if (skip > 0)
1261                 head += skip;
1262
1263         if (!session_done())
1264                 goto more;
1265 done:
1266         /* do the final flush for ordered samples */
1267         self->ordered_samples.next_flush = ULLONG_MAX;
1268         err = flush_sample_queue(self, tool);
1269 out_err:
1270         free(buf);
1271         perf_session__warn_about_errors(self, tool);
1272         perf_session_free_sample_buffers(self);
1273         return err;
1274 }
1275
1276 static union perf_event *
1277 fetch_mmaped_event(struct perf_session *session,
1278                    u64 head, size_t mmap_size, char *buf)
1279 {
1280         union perf_event *event;
1281
1282         /*
1283          * Ensure we have enough space remaining to read
1284          * the size of the event in the headers.
1285          */
1286         if (head + sizeof(event->header) > mmap_size)
1287                 return NULL;
1288
1289         event = (union perf_event *)(buf + head);
1290
1291         if (session->header.needs_swap)
1292                 perf_event_header__bswap(&event->header);
1293
1294         if (head + event->header.size > mmap_size) {
1295                 /* We're not fetching the event so swap back again */
1296                 if (session->header.needs_swap)
1297                         perf_event_header__bswap(&event->header);
1298                 return NULL;
1299         }
1300
1301         return event;
1302 }
1303
1304 /*
1305  * On 64bit we can mmap the data file in one go. No need for tiny mmap
1306  * slices. On 32bit we use 32MB.
1307  */
1308 #if BITS_PER_LONG == 64
1309 #define MMAP_SIZE ULLONG_MAX
1310 #define NUM_MMAPS 1
1311 #else
1312 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1313 #define NUM_MMAPS 128
1314 #endif
1315
1316 int __perf_session__process_events(struct perf_session *session,
1317                                    u64 data_offset, u64 data_size,
1318                                    u64 file_size, struct perf_tool *tool)
1319 {
1320         u64 head, page_offset, file_offset, file_pos, progress_next;
1321         int err, mmap_prot, mmap_flags, map_idx = 0;
1322         size_t  mmap_size;
1323         char *buf, *mmaps[NUM_MMAPS];
1324         union perf_event *event;
1325         uint32_t size;
1326
1327         perf_tool__fill_defaults(tool);
1328
1329         page_offset = page_size * (data_offset / page_size);
1330         file_offset = page_offset;
1331         head = data_offset - page_offset;
1332
1333         if (data_size && (data_offset + data_size < file_size))
1334                 file_size = data_offset + data_size;
1335
1336         progress_next = file_size / 16;
1337
1338         mmap_size = MMAP_SIZE;
1339         if (mmap_size > file_size)
1340                 mmap_size = file_size;
1341
1342         memset(mmaps, 0, sizeof(mmaps));
1343
1344         mmap_prot  = PROT_READ;
1345         mmap_flags = MAP_SHARED;
1346
1347         if (session->header.needs_swap) {
1348                 mmap_prot  |= PROT_WRITE;
1349                 mmap_flags = MAP_PRIVATE;
1350         }
1351 remap:
1352         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1353                    file_offset);
1354         if (buf == MAP_FAILED) {
1355                 pr_err("failed to mmap file\n");
1356                 err = -errno;
1357                 goto out_err;
1358         }
1359         mmaps[map_idx] = buf;
1360         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1361         file_pos = file_offset + head;
1362
1363 more:
1364         event = fetch_mmaped_event(session, head, mmap_size, buf);
1365         if (!event) {
1366                 if (mmaps[map_idx]) {
1367                         munmap(mmaps[map_idx], mmap_size);
1368                         mmaps[map_idx] = NULL;
1369                 }
1370
1371                 page_offset = page_size * (head / page_size);
1372                 file_offset += page_offset;
1373                 head -= page_offset;
1374                 goto remap;
1375         }
1376
1377         size = event->header.size;
1378
1379         if (size < sizeof(struct perf_event_header) ||
1380             perf_session__process_event(session, event, tool, file_pos) < 0) {
1381                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1382                        file_offset + head, event->header.size,
1383                        event->header.type);
1384                 err = -EINVAL;
1385                 goto out_err;
1386         }
1387
1388         head += size;
1389         file_pos += size;
1390
1391         if (file_pos >= progress_next) {
1392                 progress_next += file_size / 16;
1393                 ui_progress__update(file_pos, file_size,
1394                                     "Processing events...");
1395         }
1396
1397         if (session_done())
1398                 goto out;
1399
1400         if (file_pos < file_size)
1401                 goto more;
1402
1403 out:
1404         /* do the final flush for ordered samples */
1405         session->ordered_samples.next_flush = ULLONG_MAX;
1406         err = flush_sample_queue(session, tool);
1407 out_err:
1408         ui_progress__finish();
1409         perf_session__warn_about_errors(session, tool);
1410         perf_session_free_sample_buffers(session);
1411         return err;
1412 }
1413
1414 int perf_session__process_events(struct perf_session *self,
1415                                  struct perf_tool *tool)
1416 {
1417         int err;
1418
1419         if (perf_session__register_idle_thread(self) == NULL)
1420                 return -ENOMEM;
1421
1422         if (!self->fd_pipe)
1423                 err = __perf_session__process_events(self,
1424                                                      self->header.data_offset,
1425                                                      self->header.data_size,
1426                                                      self->size, tool);
1427         else
1428                 err = __perf_session__process_pipe_events(self, tool);
1429
1430         return err;
1431 }
1432
1433 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1434 {
1435         struct perf_evsel *evsel;
1436
1437         list_for_each_entry(evsel, &session->evlist->entries, node) {
1438                 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1439                         return true;
1440         }
1441
1442         pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1443         return false;
1444 }
1445
1446 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1447                                      const char *symbol_name, u64 addr)
1448 {
1449         char *bracket;
1450         enum map_type i;
1451         struct ref_reloc_sym *ref;
1452
1453         ref = zalloc(sizeof(struct ref_reloc_sym));
1454         if (ref == NULL)
1455                 return -ENOMEM;
1456
1457         ref->name = strdup(symbol_name);
1458         if (ref->name == NULL) {
1459                 free(ref);
1460                 return -ENOMEM;
1461         }
1462
1463         bracket = strchr(ref->name, ']');
1464         if (bracket)
1465                 *bracket = '\0';
1466
1467         ref->addr = addr;
1468
1469         for (i = 0; i < MAP__NR_TYPES; ++i) {
1470                 struct kmap *kmap = map__kmap(maps[i]);
1471                 kmap->ref_reloc_sym = ref;
1472         }
1473
1474         return 0;
1475 }
1476
1477 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1478 {
1479         return machines__fprintf_dsos(&self->machines, fp);
1480 }
1481
1482 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1483                                           bool (skip)(struct dso *dso, int parm), int parm)
1484 {
1485         return machines__fprintf_dsos_buildid(&self->machines, fp, skip, parm);
1486 }
1487
1488 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1489 {
1490         struct perf_evsel *pos;
1491         size_t ret = fprintf(fp, "Aggregated stats:\n");
1492
1493         ret += events_stats__fprintf(&session->stats, fp);
1494
1495         list_for_each_entry(pos, &session->evlist->entries, node) {
1496                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1497                 ret += events_stats__fprintf(&pos->hists.stats, fp);
1498         }
1499
1500         return ret;
1501 }
1502
1503 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1504 {
1505         /*
1506          * FIXME: Here we have to actually print all the machines in this
1507          * session, not just the host...
1508          */
1509         return machine__fprintf(&session->machines.host, fp);
1510 }
1511
1512 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1513                                               unsigned int type)
1514 {
1515         struct perf_evsel *pos;
1516
1517         list_for_each_entry(pos, &session->evlist->entries, node) {
1518                 if (pos->attr.type == type)
1519                         return pos;
1520         }
1521         return NULL;
1522 }
1523
1524 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1525                           struct perf_sample *sample, struct machine *machine,
1526                           unsigned int print_opts, unsigned int stack_depth)
1527 {
1528         struct addr_location al;
1529         struct callchain_cursor_node *node;
1530         int print_ip = print_opts & PRINT_IP_OPT_IP;
1531         int print_sym = print_opts & PRINT_IP_OPT_SYM;
1532         int print_dso = print_opts & PRINT_IP_OPT_DSO;
1533         int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1534         int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1535         char s = print_oneline ? ' ' : '\t';
1536
1537         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
1538                 error("problem processing %d event, skipping it.\n",
1539                         event->header.type);
1540                 return;
1541         }
1542
1543         if (symbol_conf.use_callchain && sample->callchain) {
1544
1545                 if (machine__resolve_callchain(machine, evsel, al.thread,
1546                                                sample, NULL, NULL) != 0) {
1547                         if (verbose)
1548                                 error("Failed to resolve callchain. Skipping\n");
1549                         return;
1550                 }
1551                 callchain_cursor_commit(&callchain_cursor);
1552
1553                 while (stack_depth) {
1554                         node = callchain_cursor_current(&callchain_cursor);
1555                         if (!node)
1556                                 break;
1557
1558                         if (print_ip)
1559                                 printf("%c%16" PRIx64, s, node->ip);
1560
1561                         if (print_sym) {
1562                                 printf(" ");
1563                                 if (print_symoffset) {
1564                                         al.addr = node->ip;
1565                                         al.map  = node->map;
1566                                         symbol__fprintf_symname_offs(node->sym, &al, stdout);
1567                                 } else
1568                                         symbol__fprintf_symname(node->sym, stdout);
1569                         }
1570
1571                         if (print_dso) {
1572                                 printf(" (");
1573                                 map__fprintf_dsoname(node->map, stdout);
1574                                 printf(")");
1575                         }
1576
1577                         if (!print_oneline)
1578                                 printf("\n");
1579
1580                         callchain_cursor_advance(&callchain_cursor);
1581
1582                         stack_depth--;
1583                 }
1584
1585         } else {
1586                 if (print_ip)
1587                         printf("%16" PRIx64, sample->ip);
1588
1589                 if (print_sym) {
1590                         printf(" ");
1591                         if (print_symoffset)
1592                                 symbol__fprintf_symname_offs(al.sym, &al,
1593                                                              stdout);
1594                         else
1595                                 symbol__fprintf_symname(al.sym, stdout);
1596                 }
1597
1598                 if (print_dso) {
1599                         printf(" (");
1600                         map__fprintf_dsoname(al.map, stdout);
1601                         printf(")");
1602                 }
1603         }
1604 }
1605
1606 int perf_session__cpu_bitmap(struct perf_session *session,
1607                              const char *cpu_list, unsigned long *cpu_bitmap)
1608 {
1609         int i;
1610         struct cpu_map *map;
1611
1612         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1613                 struct perf_evsel *evsel;
1614
1615                 evsel = perf_session__find_first_evtype(session, i);
1616                 if (!evsel)
1617                         continue;
1618
1619                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1620                         pr_err("File does not contain CPU events. "
1621                                "Remove -c option to proceed.\n");
1622                         return -1;
1623                 }
1624         }
1625
1626         map = cpu_map__new(cpu_list);
1627         if (map == NULL) {
1628                 pr_err("Invalid cpu_list\n");
1629                 return -1;
1630         }
1631
1632         for (i = 0; i < map->nr; i++) {
1633                 int cpu = map->map[i];
1634
1635                 if (cpu >= MAX_NR_CPUS) {
1636                         pr_err("Requested CPU %d too large. "
1637                                "Consider raising MAX_NR_CPUS\n", cpu);
1638                         return -1;
1639                 }
1640
1641                 set_bit(cpu, cpu_bitmap);
1642         }
1643
1644         return 0;
1645 }
1646
1647 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1648                                 bool full)
1649 {
1650         struct stat st;
1651         int ret;
1652
1653         if (session == NULL || fp == NULL)
1654                 return;
1655
1656         ret = fstat(session->fd, &st);
1657         if (ret == -1)
1658                 return;
1659
1660         fprintf(fp, "# ========\n");
1661         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1662         perf_header__fprintf_info(session, fp, full);
1663         fprintf(fp, "# ========\n#\n");
1664 }
1665
1666
1667 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1668                                              const struct perf_evsel_str_handler *assocs,
1669                                              size_t nr_assocs)
1670 {
1671         struct perf_evsel *evsel;
1672         size_t i;
1673         int err;
1674
1675         for (i = 0; i < nr_assocs; i++) {
1676                 /*
1677                  * Adding a handler for an event not in the session,
1678                  * just ignore it.
1679                  */
1680                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1681                 if (evsel == NULL)
1682                         continue;
1683
1684                 err = -EEXIST;
1685                 if (evsel->handler.func != NULL)
1686                         goto out;
1687                 evsel->handler.func = assocs[i].handler;
1688         }
1689
1690         err = 0;
1691 out:
1692         return err;
1693 }