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
  18 static int perf_session__open(struct perf_session *session)
  19 {
  20         struct perf_data_file *file = session->file;
  21
  22         if (perf_session__read_header(session) < 0) {
  23                 pr_err("incompatible file format (rerun with -v to learn more)");
  24                 return -1;
  25         }
  26
  27         if (perf_data_file__is_pipe(file))
  28                 return 0;
  29
  30         if (!perf_evlist__valid_sample_type(session->evlist)) {
  31                 pr_err("non matching sample_type");
  32                 return -1;
  33         }
  34
  35         if (!perf_evlist__valid_sample_id_all(session->evlist)) {
  36                 pr_err("non matching sample_id_all");
  37                 return -1;
  38         }
  39
  40         if (!perf_evlist__valid_read_format(session->evlist)) {
  41                 pr_err("non matching read_format");
  42                 return -1;
  43         }
  44
  45         return 0;
  46 }
  47
  48 void perf_session__set_id_hdr_size(struct perf_session *session)
  49 {
  50         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
  51
  52         machines__set_id_hdr_size(&session->machines, id_hdr_size);
  53 }
  54
  55 int perf_session__create_kernel_maps(struct perf_session *session)
  56 {
  57         int ret = machine__create_kernel_maps(&session->machines.host);
  58
  59         if (ret >= 0)
  60                 ret = machines__create_guest_kernel_maps(&session->machines);
  61         return ret;
  62 }
  63
  64 static void perf_session__destroy_kernel_maps(struct perf_session *session)
  65 {
  66         machines__destroy_kernel_maps(&session->machines);
  67 }
  68
  69 struct perf_session *perf_session__new(struct perf_data_file *file,
  70                                        bool repipe, struct perf_tool *tool)
  71 {
  72         struct perf_session *session = zalloc(sizeof(*session));
  73
  74         if (!session)
  75                 goto out;
  76
  77         session->repipe = repipe;
  78         INIT_LIST_HEAD(&session->ordered_events.events);
  79         INIT_LIST_HEAD(&session->ordered_events.cache);
  80         INIT_LIST_HEAD(&session->ordered_events.to_free);
  81         machines__init(&session->machines);
  82
  83         if (file) {
  84                 if (perf_data_file__open(file))
  85                         goto out_delete;
  86
  87                 session->file = file;
  88
  89                 if (perf_data_file__is_read(file)) {
  90                         if (perf_session__open(session) < 0)
  91                                 goto out_close;
  92
  93                         perf_session__set_id_hdr_size(session);
  94                 }
  95         }
  96
  97         if (!file || perf_data_file__is_write(file)) {
  98                 /*
  99                  * In O_RDONLY mode this will be performed when reading the
 100                  * kernel MMAP event, in perf_event__process_mmap().
 101                  */
 102                 if (perf_session__create_kernel_maps(session) < 0)
 103                         goto out_delete;
 104         }
 105
 106         if (tool && tool->ordering_requires_timestamps &&
 107             tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) {
 108                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 109                 tool->ordered_events = false;
 110         }
 111
 112         return session;
 113
 114  out_close:
 115         perf_data_file__close(file);
 116  out_delete:
 117         perf_session__delete(session);
 118  out:
 119         return NULL;
 120 }
 121
 122 static void perf_session__delete_dead_threads(struct perf_session *session)
 123 {
 124         machine__delete_dead_threads(&session->machines.host);
 125 }
 126
 127 static void perf_session__delete_threads(struct perf_session *session)
 128 {
 129         machine__delete_threads(&session->machines.host);
 130 }
 131
 132 static void perf_session_env__delete(struct perf_session_env *env)
 133 {
 134         zfree(&env->hostname);
 135         zfree(&env->os_release);
 136         zfree(&env->version);
 137         zfree(&env->arch);
 138         zfree(&env->cpu_desc);
 139         zfree(&env->cpuid);
 140
 141         zfree(&env->cmdline);
 142         zfree(&env->sibling_cores);
 143         zfree(&env->sibling_threads);
 144         zfree(&env->numa_nodes);
 145         zfree(&env->pmu_mappings);
 146 }
 147
 148 void perf_session__delete(struct perf_session *session)
 149 {
 150         perf_session__destroy_kernel_maps(session);
 151         perf_session__delete_dead_threads(session);
 152         perf_session__delete_threads(session);
 153         perf_session_env__delete(&session->header.env);
 154         machines__exit(&session->machines);
 155         if (session->file)
 156                 perf_data_file__close(session->file);
 157         free(session);
 158 }
 159
 160 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
 161                                                  __maybe_unused,
 162                                                  union perf_event *event
 163                                                  __maybe_unused,
 164                                                  struct perf_session *session
 165                                                 __maybe_unused)
 166 {
 167         dump_printf(": unhandled!\n");
 168         return 0;
 169 }
 170
 171 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
 172                                          union perf_event *event __maybe_unused,
 173                                          struct perf_evlist **pevlist
 174                                          __maybe_unused)
 175 {
 176         dump_printf(": unhandled!\n");
 177         return 0;
 178 }
 179
 180 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
 181                                      union perf_event *event __maybe_unused,
 182                                      struct perf_sample *sample __maybe_unused,
 183                                      struct perf_evsel *evsel __maybe_unused,
 184                                      struct machine *machine __maybe_unused)
 185 {
 186         dump_printf(": unhandled!\n");
 187         return 0;
 188 }
 189
 190 static int process_event_stub(struct perf_tool *tool __maybe_unused,
 191                               union perf_event *event __maybe_unused,
 192                               struct perf_sample *sample __maybe_unused,
 193                               struct machine *machine __maybe_unused)
 194 {
 195         dump_printf(": unhandled!\n");
 196         return 0;
 197 }
 198
 199 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
 200                                        union perf_event *event __maybe_unused,
 201                                        struct perf_session *perf_session
 202                                        __maybe_unused)
 203 {
 204         dump_printf(": unhandled!\n");
 205         return 0;
 206 }
 207
 208 static int process_finished_round(struct perf_tool *tool,
 209                                   union perf_event *event,
 210                                   struct perf_session *session);
 211
 212 void perf_tool__fill_defaults(struct perf_tool *tool)
 213 {
 214         if (tool->sample == NULL)
 215                 tool->sample = process_event_sample_stub;
 216         if (tool->mmap == NULL)
 217                 tool->mmap = process_event_stub;
 218         if (tool->mmap2 == NULL)
 219                 tool->mmap2 = process_event_stub;
 220         if (tool->comm == NULL)
 221                 tool->comm = process_event_stub;
 222         if (tool->fork == NULL)
 223                 tool->fork = process_event_stub;
 224         if (tool->exit == NULL)
 225                 tool->exit = process_event_stub;
 226         if (tool->lost == NULL)
 227                 tool->lost = perf_event__process_lost;
 228         if (tool->read == NULL)
 229                 tool->read = process_event_sample_stub;
 230         if (tool->throttle == NULL)
 231                 tool->throttle = process_event_stub;
 232         if (tool->unthrottle == NULL)
 233                 tool->unthrottle = process_event_stub;
 234         if (tool->attr == NULL)
 235                 tool->attr = process_event_synth_attr_stub;
 236         if (tool->tracing_data == NULL)
 237                 tool->tracing_data = process_event_synth_tracing_data_stub;
 238         if (tool->build_id == NULL)
 239                 tool->build_id = process_finished_round_stub;
 240         if (tool->finished_round == NULL) {
 241                 if (tool->ordered_events)
 242                         tool->finished_round = process_finished_round;
 243                 else
 244                         tool->finished_round = process_finished_round_stub;
 245         }
 246 }
 247
 248 static void swap_sample_id_all(union perf_event *event, void *data)
 249 {
 250         void *end = (void *) event + event->header.size;
 251         int size = end - data;
 252
 253         BUG_ON(size % sizeof(u64));
 254         mem_bswap_64(data, size);
 255 }
 256
 257 static void perf_event__all64_swap(union perf_event *event,
 258                                    bool sample_id_all __maybe_unused)
 259 {
 260         struct perf_event_header *hdr = &event->header;
 261         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 262 }
 263
 264 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 265 {
 266         event->comm.pid = bswap_32(event->comm.pid);
 267         event->comm.tid = bswap_32(event->comm.tid);
 268
 269         if (sample_id_all) {
 270                 void *data = &event->comm.comm;
 271
 272                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 273                 swap_sample_id_all(event, data);
 274         }
 275 }
 276
 277 static void perf_event__mmap_swap(union perf_event *event,
 278                                   bool sample_id_all)
 279 {
 280         event->mmap.pid   = bswap_32(event->mmap.pid);
 281         event->mmap.tid   = bswap_32(event->mmap.tid);
 282         event->mmap.start = bswap_64(event->mmap.start);
 283         event->mmap.len   = bswap_64(event->mmap.len);
 284         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 285
 286         if (sample_id_all) {
 287                 void *data = &event->mmap.filename;
 288
 289                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 290                 swap_sample_id_all(event, data);
 291         }
 292 }
 293
 294 static void perf_event__mmap2_swap(union perf_event *event,
 295                                   bool sample_id_all)
 296 {
 297         event->mmap2.pid   = bswap_32(event->mmap2.pid);
 298         event->mmap2.tid   = bswap_32(event->mmap2.tid);
 299         event->mmap2.start = bswap_64(event->mmap2.start);
 300         event->mmap2.len   = bswap_64(event->mmap2.len);
 301         event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
 302         event->mmap2.maj   = bswap_32(event->mmap2.maj);
 303         event->mmap2.min   = bswap_32(event->mmap2.min);
 304         event->mmap2.ino   = bswap_64(event->mmap2.ino);
 305
 306         if (sample_id_all) {
 307                 void *data = &event->mmap2.filename;
 308
 309                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 310                 swap_sample_id_all(event, data);
 311         }
 312 }
 313 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 314 {
 315         event->fork.pid  = bswap_32(event->fork.pid);
 316         event->fork.tid  = bswap_32(event->fork.tid);
 317         event->fork.ppid = bswap_32(event->fork.ppid);
 318         event->fork.ptid = bswap_32(event->fork.ptid);
 319         event->fork.time = bswap_64(event->fork.time);
 320
 321         if (sample_id_all)
 322                 swap_sample_id_all(event, &event->fork + 1);
 323 }
 324
 325 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 326 {
 327         event->read.pid          = bswap_32(event->read.pid);
 328         event->read.tid          = bswap_32(event->read.tid);
 329         event->read.value        = bswap_64(event->read.value);
 330         event->read.time_enabled = bswap_64(event->read.time_enabled);
 331         event->read.time_running = bswap_64(event->read.time_running);
 332         event->read.id           = bswap_64(event->read.id);
 333
 334         if (sample_id_all)
 335                 swap_sample_id_all(event, &event->read + 1);
 336 }
 337
 338 static void perf_event__throttle_swap(union perf_event *event,
 339                                       bool sample_id_all)
 340 {
 341         event->throttle.time      = bswap_64(event->throttle.time);
 342         event->throttle.id        = bswap_64(event->throttle.id);
 343         event->throttle.stream_id = bswap_64(event->throttle.stream_id);
 344
 345         if (sample_id_all)
 346                 swap_sample_id_all(event, &event->throttle + 1);
 347 }
 348
 349 static u8 revbyte(u8 b)
 350 {
 351         int rev = (b >> 4) | ((b & 0xf) << 4);
 352         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 353         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 354         return (u8) rev;
 355 }
 356
 357 /*
 358  * XXX this is hack in attempt to carry flags bitfield
 359  * throught endian village. ABI says:
 360  *
 361  * Bit-fields are allocated from right to left (least to most significant)
 362  * on little-endian implementations and from left to right (most to least
 363  * significant) on big-endian implementations.
 364  *
 365  * The above seems to be byte specific, so we need to reverse each
 366  * byte of the bitfield. 'Internet' also says this might be implementation
 367  * specific and we probably need proper fix and carry perf_event_attr
 368  * bitfield flags in separate data file FEAT_ section. Thought this seems
 369  * to work for now.
 370  */
 371 static void swap_bitfield(u8 *p, unsigned len)
 372 {
 373         unsigned i;
 374
 375         for (i = 0; i < len; i++) {
 376                 *p = revbyte(*p);
 377                 p++;
 378         }
 379 }
 380
 381 /* exported for swapping attributes in file header */
 382 void perf_event__attr_swap(struct perf_event_attr *attr)
 383 {
 384         attr->type              = bswap_32(attr->type);
 385         attr->size              = bswap_32(attr->size);
 386         attr->config            = bswap_64(attr->config);
 387         attr->sample_period     = bswap_64(attr->sample_period);
 388         attr->sample_type       = bswap_64(attr->sample_type);
 389         attr->read_format       = bswap_64(attr->read_format);
 390         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 391         attr->bp_type           = bswap_32(attr->bp_type);
 392         attr->bp_addr           = bswap_64(attr->bp_addr);
 393         attr->bp_len            = bswap_64(attr->bp_len);
 394         attr->branch_sample_type = bswap_64(attr->branch_sample_type);
 395         attr->sample_regs_user   = bswap_64(attr->sample_regs_user);
 396         attr->sample_stack_user  = bswap_32(attr->sample_stack_user);
 397
 398         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 399 }
 400
 401 static void perf_event__hdr_attr_swap(union perf_event *event,
 402                                       bool sample_id_all __maybe_unused)
 403 {
 404         size_t size;
 405
 406         perf_event__attr_swap(&event->attr.attr);
 407
 408         size = event->header.size;
 409         size -= (void *)&event->attr.id - (void *)event;
 410         mem_bswap_64(event->attr.id, size);
 411 }
 412
 413 static void perf_event__event_type_swap(union perf_event *event,
 414                                         bool sample_id_all __maybe_unused)
 415 {
 416         event->event_type.event_type.event_id =
 417                 bswap_64(event->event_type.event_type.event_id);
 418 }
 419
 420 static void perf_event__tracing_data_swap(union perf_event *event,
 421                                           bool sample_id_all __maybe_unused)
 422 {
 423         event->tracing_data.size = bswap_32(event->tracing_data.size);
 424 }
 425
 426 typedef void (*perf_event__swap_op)(union perf_event *event,
 427                                     bool sample_id_all);
 428
 429 static perf_event__swap_op perf_event__swap_ops[] = {
 430         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 431         [PERF_RECORD_MMAP2]               = perf_event__mmap2_swap,
 432         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 433         [PERF_RECORD_FORK]                = perf_event__task_swap,
 434         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 435         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 436         [PERF_RECORD_READ]                = perf_event__read_swap,
 437         [PERF_RECORD_THROTTLE]            = perf_event__throttle_swap,
 438         [PERF_RECORD_UNTHROTTLE]          = perf_event__throttle_swap,
 439         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 440         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 441         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 442         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 443         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 444         [PERF_RECORD_HEADER_MAX]          = NULL,
 445 };
 446
 447 struct ordered_event {
 448         u64                     timestamp;
 449         u64                     file_offset;
 450         union perf_event        *event;
 451         struct list_head        list;
 452 };
 453
 454 static void perf_session_free_sample_buffers(struct perf_session *session)
 455 {
 456         struct ordered_events *oe = &session->ordered_events;
 457
 458         while (!list_empty(&oe->to_free)) {
 459                 struct ordered_event *event;
 460
 461                 event = list_entry(oe->to_free.next, struct ordered_event, list);
 462                 list_del(&event->list);
 463                 free(event);
 464         }
 465 }
 466
 467 /* The queue is ordered by time */
 468 static void queue_event(struct ordered_events *oe, struct ordered_event *new)
 469 {
 470         struct ordered_event *last = oe->last;
 471         u64 timestamp = new->timestamp;
 472         struct list_head *p;
 473
 474         ++oe->nr_events;
 475         oe->last = new;
 476
 477         if (!last) {
 478                 list_add(&new->list, &oe->events);
 479                 oe->max_timestamp = timestamp;
 480                 return;
 481         }
 482
 483         /*
 484          * last event might point to some random place in the list as it's
 485          * the last queued event. We expect that the new event is close to
 486          * this.
 487          */
 488         if (last->timestamp <= timestamp) {
 489                 while (last->timestamp <= timestamp) {
 490                         p = last->list.next;
 491                         if (p == &oe->events) {
 492                                 list_add_tail(&new->list, &oe->events);
 493                                 oe->max_timestamp = timestamp;
 494                                 return;
 495                         }
 496                         last = list_entry(p, struct ordered_event, list);
 497                 }
 498                 list_add_tail(&new->list, &last->list);
 499         } else {
 500                 while (last->timestamp > timestamp) {
 501                         p = last->list.prev;
 502                         if (p == &oe->events) {
 503                                 list_add(&new->list, &oe->events);
 504                                 return;
 505                         }
 506                         last = list_entry(p, struct ordered_event, list);
 507                 }
 508                 list_add(&new->list, &last->list);
 509         }
 510 }
 511
 512 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct ordered_event))
 513 static struct ordered_event *alloc_event(struct ordered_events *oe)
 514 {
 515         struct list_head *cache = &oe->cache;
 516         struct ordered_event *new;
 517
 518         if (!list_empty(cache)) {
 519                 new = list_entry(cache->next, struct ordered_event, list);
 520                 list_del(&new->list);
 521         } else if (oe->buffer) {
 522                 new = oe->buffer + oe->buffer_idx;
 523                 if (++oe->buffer_idx == MAX_SAMPLE_BUFFER)
 524                         oe->buffer = NULL;
 525         } else {
 526                 oe->buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 527                 if (!oe->buffer)
 528                         return NULL;
 529                 list_add(&oe->buffer->list, &oe->to_free);
 530
 531                 /* First entry is abused to maintain the to_free list. */
 532                 oe->buffer_idx = 2;
 533                 new = oe->buffer + 1;
 534         }
 535
 536         return new;
 537 }
 538
 539 static struct ordered_event *
 540 ordered_events__new(struct ordered_events *oe, u64 timestamp)
 541 {
 542         struct ordered_event *new;
 543
 544         new = alloc_event(oe);
 545         if (new) {
 546                 new->timestamp = timestamp;
 547                 queue_event(oe, new);
 548         }
 549
 550         return new;
 551 }
 552
 553 static void
 554 ordered_events__delete(struct ordered_events *oe, struct ordered_event *event)
 555 {
 556         list_del(&event->list);
 557         list_add(&event->list, &oe->cache);
 558         oe->nr_events--;
 559 }
 560
 561 static int perf_session_deliver_event(struct perf_session *session,
 562                                       union perf_event *event,
 563                                       struct perf_sample *sample,
 564                                       struct perf_tool *tool,
 565                                       u64 file_offset);
 566
 567 static int ordered_events__flush(struct perf_session *s,
 568                                  struct perf_tool *tool)
 569 {
 570         struct ordered_events *oe = &s->ordered_events;
 571         struct list_head *head = &oe->events;
 572         struct ordered_event *tmp, *iter;
 573         struct perf_sample sample;
 574         u64 limit = oe->next_flush;
 575         u64 last_ts = oe->last ? oe->last->timestamp : 0ULL;
 576         bool show_progress = limit == ULLONG_MAX;
 577         struct ui_progress prog;
 578         int ret;
 579
 580         if (!tool->ordered_events || !limit)
 581                 return 0;
 582
 583         if (show_progress)
 584                 ui_progress__init(&prog, oe->nr_events, "Processing time ordered events...");
 585
 586         list_for_each_entry_safe(iter, tmp, head, list) {
 587                 if (session_done())
 588                         return 0;
 589
 590                 if (iter->timestamp > limit)
 591                         break;
 592
 593                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
 594                 if (ret)
 595                         pr_err("Can't parse sample, err = %d\n", ret);
 596                 else {
 597                         ret = perf_session_deliver_event(s, iter->event, &sample, tool,
 598                                                          iter->file_offset);
 599                         if (ret)
 600                                 return ret;
 601                 }
 602
 603                 ordered_events__delete(oe, iter);
 604                 oe->last_flush = iter->timestamp;
 605
 606                 if (show_progress)
 607                         ui_progress__update(&prog, 1);
 608         }
 609
 610         if (list_empty(head))
 611                 oe->last = NULL;
 612         else if (last_ts <= limit)
 613                 oe->last = list_entry(head->prev, struct ordered_event, list);
 614
 615         return 0;
 616 }
 617
 618 /*
 619  * When perf record finishes a pass on every buffers, it records this pseudo
 620  * event.
 621  * We record the max timestamp t found in the pass n.
 622  * Assuming these timestamps are monotonic across cpus, we know that if
 623  * a buffer still has events with timestamps below t, they will be all
 624  * available and then read in the pass n + 1.
 625  * Hence when we start to read the pass n + 2, we can safely flush every
 626  * events with timestamps below t.
 627  *
 628  *    ============ PASS n =================
 629  *       CPU 0         |   CPU 1
 630  *                     |
 631  *    cnt1 timestamps  |   cnt2 timestamps
 632  *          1          |         2
 633  *          2          |         3
 634  *          -          |         4  <--- max recorded
 635  *
 636  *    ============ PASS n + 1 ==============
 637  *       CPU 0         |   CPU 1
 638  *                     |
 639  *    cnt1 timestamps  |   cnt2 timestamps
 640  *          3          |         5
 641  *          4          |         6
 642  *          5          |         7 <---- max recorded
 643  *
 644  *      Flush every events below timestamp 4
 645  *
 646  *    ============ PASS n + 2 ==============
 647  *       CPU 0         |   CPU 1
 648  *                     |
 649  *    cnt1 timestamps  |   cnt2 timestamps
 650  *          6          |         8
 651  *          7          |         9
 652  *          -          |         10
 653  *
 654  *      Flush every events below timestamp 7
 655  *      etc...
 656  */
 657 static int process_finished_round(struct perf_tool *tool,
 658                                   union perf_event *event __maybe_unused,
 659                                   struct perf_session *session)
 660 {
 661         int ret = ordered_events__flush(session, tool);
 662         if (!ret)
 663                 session->ordered_events.next_flush = session->ordered_events.max_timestamp;
 664
 665         return ret;
 666 }
 667
 668 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 669                                     struct perf_sample *sample, u64 file_offset)
 670 {
 671         struct ordered_events *oe = &s->ordered_events;
 672         u64 timestamp = sample->time;
 673         struct ordered_event *new;
 674
 675         if (!timestamp || timestamp == ~0ULL)
 676                 return -ETIME;
 677
 678         if (timestamp < s->ordered_events.last_flush) {
 679                 printf("Warning: Timestamp below last timeslice flush\n");
 680                 return -EINVAL;
 681         }
 682
 683         new = ordered_events__new(oe, timestamp);
 684         if (!new)
 685                 return -ENOMEM;
 686
 687         new->file_offset = file_offset;
 688         new->event = event;
 689         return 0;
 690 }
 691
 692 static void callchain__printf(struct perf_sample *sample)
 693 {
 694         unsigned int i;
 695
 696         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 697
 698         for (i = 0; i < sample->callchain->nr; i++)
 699                 printf("..... %2d: %016" PRIx64 "\n",
 700                        i, sample->callchain->ips[i]);
 701 }
 702
 703 static void branch_stack__printf(struct perf_sample *sample)
 704 {
 705         uint64_t i;
 706
 707         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 708
 709         for (i = 0; i < sample->branch_stack->nr; i++)
 710                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 711                         i, sample->branch_stack->entries[i].from,
 712                         sample->branch_stack->entries[i].to);
 713 }
 714
 715 static void regs_dump__printf(u64 mask, u64 *regs)
 716 {
 717         unsigned rid, i = 0;
 718
 719         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 720                 u64 val = regs[i++];
 721
 722                 printf(".... %-5s 0x%" PRIx64 "\n",
 723                        perf_reg_name(rid), val);
 724         }
 725 }
 726
 727 static void regs_user__printf(struct perf_sample *sample)
 728 {
 729         struct regs_dump *user_regs = &sample->user_regs;
 730
 731         if (user_regs->regs) {
 732                 u64 mask = user_regs->mask;
 733                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
 734                 regs_dump__printf(mask, user_regs->regs);
 735         }
 736 }
 737
 738 static void stack_user__printf(struct stack_dump *dump)
 739 {
 740         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 741                dump->size, dump->offset);
 742 }
 743
 744 static void perf_session__print_tstamp(struct perf_session *session,
 745                                        union perf_event *event,
 746                                        struct perf_sample *sample)
 747 {
 748         u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
 749
 750         if (event->header.type != PERF_RECORD_SAMPLE &&
 751             !perf_evlist__sample_id_all(session->evlist)) {
 752                 fputs("-1 -1 ", stdout);
 753                 return;
 754         }
 755
 756         if ((sample_type & PERF_SAMPLE_CPU))
 757                 printf("%u ", sample->cpu);
 758
 759         if (sample_type & PERF_SAMPLE_TIME)
 760                 printf("%" PRIu64 " ", sample->time);
 761 }
 762
 763 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
 764 {
 765         printf("... sample_read:\n");
 766
 767         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 768                 printf("...... time enabled %016" PRIx64 "\n",
 769                        sample->read.time_enabled);
 770
 771         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 772                 printf("...... time running %016" PRIx64 "\n",
 773                        sample->read.time_running);
 774
 775         if (read_format & PERF_FORMAT_GROUP) {
 776                 u64 i;
 777
 778                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
 779
 780                 for (i = 0; i < sample->read.group.nr; i++) {
 781                         struct sample_read_value *value;
 782
 783                         value = &sample->read.group.values[i];
 784                         printf("..... id %016" PRIx64
 785                                ", value %016" PRIx64 "\n",
 786                                value->id, value->value);
 787                 }
 788         } else
 789                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
 790                         sample->read.one.id, sample->read.one.value);
 791 }
 792
 793 static void dump_event(struct perf_session *session, union perf_event *event,
 794                        u64 file_offset, struct perf_sample *sample)
 795 {
 796         if (!dump_trace)
 797                 return;
 798
 799         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 800                file_offset, event->header.size, event->header.type);
 801
 802         trace_event(event);
 803
 804         if (sample)
 805                 perf_session__print_tstamp(session, event, sample);
 806
 807         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 808                event->header.size, perf_event__name(event->header.type));
 809 }
 810
 811 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 812                         struct perf_sample *sample)
 813 {
 814         u64 sample_type;
 815
 816         if (!dump_trace)
 817                 return;
 818
 819         printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 820                event->header.misc, sample->pid, sample->tid, sample->ip,
 821                sample->period, sample->addr);
 822
 823         sample_type = evsel->attr.sample_type;
 824
 825         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 826                 callchain__printf(sample);
 827
 828         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
 829                 branch_stack__printf(sample);
 830
 831         if (sample_type & PERF_SAMPLE_REGS_USER)
 832                 regs_user__printf(sample);
 833
 834         if (sample_type & PERF_SAMPLE_STACK_USER)
 835                 stack_user__printf(&sample->user_stack);
 836
 837         if (sample_type & PERF_SAMPLE_WEIGHT)
 838                 printf("... weight: %" PRIu64 "\n", sample->weight);
 839
 840         if (sample_type & PERF_SAMPLE_DATA_SRC)
 841                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
 842
 843         if (sample_type & PERF_SAMPLE_TRANSACTION)
 844                 printf("... transaction: %" PRIx64 "\n", sample->transaction);
 845
 846         if (sample_type & PERF_SAMPLE_READ)
 847                 sample_read__printf(sample, evsel->attr.read_format);
 848 }
 849
 850 static struct machine *
 851         perf_session__find_machine_for_cpumode(struct perf_session *session,
 852                                                union perf_event *event,
 853                                                struct perf_sample *sample)
 854 {
 855         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 856         struct machine *machine;
 857
 858         if (perf_guest &&
 859             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
 860              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
 861                 u32 pid;
 862
 863                 if (event->header.type == PERF_RECORD_MMAP
 864                     || event->header.type == PERF_RECORD_MMAP2)
 865                         pid = event->mmap.pid;
 866                 else
 867                         pid = sample->pid;
 868
 869                 machine = perf_session__find_machine(session, pid);
 870                 if (!machine)
 871                         machine = perf_session__findnew_machine(session,
 872                                                 DEFAULT_GUEST_KERNEL_ID);
 873                 return machine;
 874         }
 875
 876         return &session->machines.host;
 877 }
 878
 879 static int deliver_sample_value(struct perf_session *session,
 880                                 struct perf_tool *tool,
 881                                 union perf_event *event,
 882                                 struct perf_sample *sample,
 883                                 struct sample_read_value *v,
 884                                 struct machine *machine)
 885 {
 886         struct perf_sample_id *sid;
 887
 888         sid = perf_evlist__id2sid(session->evlist, v->id);
 889         if (sid) {
 890                 sample->id     = v->id;
 891                 sample->period = v->value - sid->period;
 892                 sid->period    = v->value;
 893         }
 894
 895         if (!sid || sid->evsel == NULL) {
 896                 ++session->stats.nr_unknown_id;
 897                 return 0;
 898         }
 899
 900         return tool->sample(tool, event, sample, sid->evsel, machine);
 901 }
 902
 903 static int deliver_sample_group(struct perf_session *session,
 904                                 struct perf_tool *tool,
 905                                 union  perf_event *event,
 906                                 struct perf_sample *sample,
 907                                 struct machine *machine)
 908 {
 909         int ret = -EINVAL;
 910         u64 i;
 911
 912         for (i = 0; i < sample->read.group.nr; i++) {
 913                 ret = deliver_sample_value(session, tool, event, sample,
 914                                            &sample->read.group.values[i],
 915                                            machine);
 916                 if (ret)
 917                         break;
 918         }
 919
 920         return ret;
 921 }
 922
 923 static int
 924 perf_session__deliver_sample(struct perf_session *session,
 925                              struct perf_tool *tool,
 926                              union  perf_event *event,
 927                              struct perf_sample *sample,
 928                              struct perf_evsel *evsel,
 929                              struct machine *machine)
 930 {
 931         /* We know evsel != NULL. */
 932         u64 sample_type = evsel->attr.sample_type;
 933         u64 read_format = evsel->attr.read_format;
 934
 935         /* Standard sample delievery. */
 936         if (!(sample_type & PERF_SAMPLE_READ))
 937                 return tool->sample(tool, event, sample, evsel, machine);
 938
 939         /* For PERF_SAMPLE_READ we have either single or group mode. */
 940         if (read_format & PERF_FORMAT_GROUP)
 941                 return deliver_sample_group(session, tool, event, sample,
 942                                             machine);
 943         else
 944                 return deliver_sample_value(session, tool, event, sample,
 945                                             &sample->read.one, machine);
 946 }
 947
 948 static int perf_session_deliver_event(struct perf_session *session,
 949                                       union perf_event *event,
 950                                       struct perf_sample *sample,
 951                                       struct perf_tool *tool,
 952                                       u64 file_offset)
 953 {
 954         struct perf_evsel *evsel;
 955         struct machine *machine;
 956
 957         dump_event(session, event, file_offset, sample);
 958
 959         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 960         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
 961                 /*
 962                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
 963                  * because the tools right now may apply filters, discarding
 964                  * some of the samples. For consistency, in the future we
 965                  * should have something like nr_filtered_samples and remove
 966                  * the sample->period from total_sample_period, etc, KISS for
 967                  * now tho.
 968                  *
 969                  * Also testing against NULL allows us to handle files without
 970                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
 971                  * future probably it'll be a good idea to restrict event
 972                  * processing via perf_session to files with both set.
 973                  */
 974                 hists__inc_nr_events(&evsel->hists, event->header.type);
 975         }
 976
 977         machine = perf_session__find_machine_for_cpumode(session, event,
 978                                                          sample);
 979
 980         switch (event->header.type) {
 981         case PERF_RECORD_SAMPLE:
 982                 dump_sample(evsel, event, sample);
 983                 if (evsel == NULL) {
 984                         ++session->stats.nr_unknown_id;
 985                         return 0;
 986                 }
 987                 if (machine == NULL) {
 988                         ++session->stats.nr_unprocessable_samples;
 989                         return 0;
 990                 }
 991                 return perf_session__deliver_sample(session, tool, event,
 992                                                     sample, evsel, machine);
 993         case PERF_RECORD_MMAP:
 994                 return tool->mmap(tool, event, sample, machine);
 995         case PERF_RECORD_MMAP2:
 996                 return tool->mmap2(tool, event, sample, machine);
 997         case PERF_RECORD_COMM:
 998                 return tool->comm(tool, event, sample, machine);
 999         case PERF_RECORD_FORK:
1000                 return tool->fork(tool, event, sample, machine);
1001         case PERF_RECORD_EXIT:
1002                 return tool->exit(tool, event, sample, machine);
1003         case PERF_RECORD_LOST:
1004                 if (tool->lost == perf_event__process_lost)
1005                         session->stats.total_lost += event->lost.lost;
1006                 return tool->lost(tool, event, sample, machine);
1007         case PERF_RECORD_READ:
1008                 return tool->read(tool, event, sample, evsel, machine);
1009         case PERF_RECORD_THROTTLE:
1010                 return tool->throttle(tool, event, sample, machine);
1011         case PERF_RECORD_UNTHROTTLE:
1012                 return tool->unthrottle(tool, event, sample, machine);
1013         default:
1014                 ++session->stats.nr_unknown_events;
1015                 return -1;
1016         }
1017 }
1018
1019 static s64 perf_session__process_user_event(struct perf_session *session,
1020                                             union perf_event *event,
1021                                             struct perf_tool *tool,
1022                                             u64 file_offset)
1023 {
1024         int fd = perf_data_file__fd(session->file);
1025         int err;
1026
1027         dump_event(session, event, file_offset, NULL);
1028
1029         /* These events are processed right away */
1030         switch (event->header.type) {
1031         case PERF_RECORD_HEADER_ATTR:
1032                 err = tool->attr(tool, event, &session->evlist);
1033                 if (err == 0)
1034                         perf_session__set_id_hdr_size(session);
1035                 return err;
1036         case PERF_RECORD_HEADER_EVENT_TYPE:
1037                 /*
1038                  * Depreceated, but we need to handle it for sake
1039                  * of old data files create in pipe mode.
1040                  */
1041                 return 0;
1042         case PERF_RECORD_HEADER_TRACING_DATA:
1043                 /* setup for reading amidst mmap */
1044                 lseek(fd, file_offset, SEEK_SET);
1045                 return tool->tracing_data(tool, event, session);
1046         case PERF_RECORD_HEADER_BUILD_ID:
1047                 return tool->build_id(tool, event, session);
1048         case PERF_RECORD_FINISHED_ROUND:
1049                 return tool->finished_round(tool, event, session);
1050         default:
1051                 return -EINVAL;
1052         }
1053 }
1054
1055 static void event_swap(union perf_event *event, bool sample_id_all)
1056 {
1057         perf_event__swap_op swap;
1058
1059         swap = perf_event__swap_ops[event->header.type];
1060         if (swap)
1061                 swap(event, sample_id_all);
1062 }
1063
1064 static s64 perf_session__process_event(struct perf_session *session,
1065                                        union perf_event *event,
1066                                        struct perf_tool *tool,
1067                                        u64 file_offset)
1068 {
1069         struct perf_sample sample;
1070         int ret;
1071
1072         if (session->header.needs_swap)
1073                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1074
1075         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1076                 return -EINVAL;
1077
1078         events_stats__inc(&session->stats, event->header.type);
1079
1080         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1081                 return perf_session__process_user_event(session, event, tool, file_offset);
1082
1083         /*
1084          * For all kernel events we get the sample data
1085          */
1086         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1087         if (ret)
1088                 return ret;
1089
1090         if (tool->ordered_events) {
1091                 ret = perf_session_queue_event(session, event, &sample,
1092                                                file_offset);
1093                 if (ret != -ETIME)
1094                         return ret;
1095         }
1096
1097         return perf_session_deliver_event(session, event, &sample, tool,
1098                                           file_offset);
1099 }
1100
1101 void perf_event_header__bswap(struct perf_event_header *hdr)
1102 {
1103         hdr->type = bswap_32(hdr->type);
1104         hdr->misc = bswap_16(hdr->misc);
1105         hdr->size = bswap_16(hdr->size);
1106 }
1107
1108 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1109 {
1110         return machine__findnew_thread(&session->machines.host, -1, pid);
1111 }
1112
1113 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1114 {
1115         struct thread *thread;
1116
1117         thread = machine__findnew_thread(&session->machines.host, 0, 0);
1118         if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1119                 pr_err("problem inserting idle task.\n");
1120                 thread = NULL;
1121         }
1122
1123         return thread;
1124 }
1125
1126 static void perf_session__warn_about_errors(const struct perf_session *session,
1127                                             const struct perf_tool *tool)
1128 {
1129         if (tool->lost == perf_event__process_lost &&
1130             session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1131                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1132                             "Check IO/CPU overload!\n\n",
1133                             session->stats.nr_events[0],
1134                             session->stats.nr_events[PERF_RECORD_LOST]);
1135         }
1136
1137         if (session->stats.nr_unknown_events != 0) {
1138                 ui__warning("Found %u unknown events!\n\n"
1139                             "Is this an older tool processing a perf.data "
1140                             "file generated by a more recent tool?\n\n"
1141                             "If that is not the case, consider "
1142                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1143                             session->stats.nr_unknown_events);
1144         }
1145
1146         if (session->stats.nr_unknown_id != 0) {
1147                 ui__warning("%u samples with id not present in the header\n",
1148                             session->stats.nr_unknown_id);
1149         }
1150
1151         if (session->stats.nr_invalid_chains != 0) {
1152                 ui__warning("Found invalid callchains!\n\n"
1153                             "%u out of %u events were discarded for this reason.\n\n"
1154                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1155                             session->stats.nr_invalid_chains,
1156                             session->stats.nr_events[PERF_RECORD_SAMPLE]);
1157         }
1158
1159         if (session->stats.nr_unprocessable_samples != 0) {
1160                 ui__warning("%u unprocessable samples recorded.\n"
1161                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1162                             session->stats.nr_unprocessable_samples);
1163         }
1164 }
1165
1166 volatile int session_done;
1167
1168 static int __perf_session__process_pipe_events(struct perf_session *session,
1169                                                struct perf_tool *tool)
1170 {
1171         int fd = perf_data_file__fd(session->file);
1172         union perf_event *event;
1173         uint32_t size, cur_size = 0;
1174         void *buf = NULL;
1175         s64 skip = 0;
1176         u64 head;
1177         ssize_t err;
1178         void *p;
1179
1180         perf_tool__fill_defaults(tool);
1181
1182         head = 0;
1183         cur_size = sizeof(union perf_event);
1184
1185         buf = malloc(cur_size);
1186         if (!buf)
1187                 return -errno;
1188 more:
1189         event = buf;
1190         err = readn(fd, event, sizeof(struct perf_event_header));
1191         if (err <= 0) {
1192                 if (err == 0)
1193                         goto done;
1194
1195                 pr_err("failed to read event header\n");
1196                 goto out_err;
1197         }
1198
1199         if (session->header.needs_swap)
1200                 perf_event_header__bswap(&event->header);
1201
1202         size = event->header.size;
1203         if (size < sizeof(struct perf_event_header)) {
1204                 pr_err("bad event header size\n");
1205                 goto out_err;
1206         }
1207
1208         if (size > cur_size) {
1209                 void *new = realloc(buf, size);
1210                 if (!new) {
1211                         pr_err("failed to allocate memory to read event\n");
1212                         goto out_err;
1213                 }
1214                 buf = new;
1215                 cur_size = size;
1216                 event = buf;
1217         }
1218         p = event;
1219         p += sizeof(struct perf_event_header);
1220
1221         if (size - sizeof(struct perf_event_header)) {
1222                 err = readn(fd, p, size - sizeof(struct perf_event_header));
1223                 if (err <= 0) {
1224                         if (err == 0) {
1225                                 pr_err("unexpected end of event stream\n");
1226                                 goto done;
1227                         }
1228
1229                         pr_err("failed to read event data\n");
1230                         goto out_err;
1231                 }
1232         }
1233
1234         if ((skip = perf_session__process_event(session, event, tool, head)) < 0) {
1235                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1236                        head, event->header.size, event->header.type);
1237                 err = -EINVAL;
1238                 goto out_err;
1239         }
1240
1241         head += size;
1242
1243         if (skip > 0)
1244                 head += skip;
1245
1246         if (!session_done())
1247                 goto more;
1248 done:
1249         /* do the final flush for ordered samples */
1250         session->ordered_events.next_flush = ULLONG_MAX;
1251         err = ordered_events__flush(session, tool);
1252 out_err:
1253         free(buf);
1254         perf_session__warn_about_errors(session, tool);
1255         perf_session_free_sample_buffers(session);
1256         return err;
1257 }
1258
1259 static union perf_event *
1260 fetch_mmaped_event(struct perf_session *session,
1261                    u64 head, size_t mmap_size, char *buf)
1262 {
1263         union perf_event *event;
1264
1265         /*
1266          * Ensure we have enough space remaining to read
1267          * the size of the event in the headers.
1268          */
1269         if (head + sizeof(event->header) > mmap_size)
1270                 return NULL;
1271
1272         event = (union perf_event *)(buf + head);
1273
1274         if (session->header.needs_swap)
1275                 perf_event_header__bswap(&event->header);
1276
1277         if (head + event->header.size > mmap_size) {
1278                 /* We're not fetching the event so swap back again */
1279                 if (session->header.needs_swap)
1280                         perf_event_header__bswap(&event->header);
1281                 return NULL;
1282         }
1283
1284         return event;
1285 }
1286
1287 /*
1288  * On 64bit we can mmap the data file in one go. No need for tiny mmap
1289  * slices. On 32bit we use 32MB.
1290  */
1291 #if BITS_PER_LONG == 64
1292 #define MMAP_SIZE ULLONG_MAX
1293 #define NUM_MMAPS 1
1294 #else
1295 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1296 #define NUM_MMAPS 128
1297 #endif
1298
1299 int __perf_session__process_events(struct perf_session *session,
1300                                    u64 data_offset, u64 data_size,
1301                                    u64 file_size, struct perf_tool *tool)
1302 {
1303         int fd = perf_data_file__fd(session->file);
1304         u64 head, page_offset, file_offset, file_pos, size;
1305         int err, mmap_prot, mmap_flags, map_idx = 0;
1306         size_t  mmap_size;
1307         char *buf, *mmaps[NUM_MMAPS];
1308         union perf_event *event;
1309         struct ui_progress prog;
1310         s64 skip;
1311
1312         perf_tool__fill_defaults(tool);
1313
1314         page_offset = page_size * (data_offset / page_size);
1315         file_offset = page_offset;
1316         head = data_offset - page_offset;
1317
1318         if (data_size && (data_offset + data_size < file_size))
1319                 file_size = data_offset + data_size;
1320
1321         ui_progress__init(&prog, file_size, "Processing events...");
1322
1323         mmap_size = MMAP_SIZE;
1324         if (mmap_size > file_size) {
1325                 mmap_size = file_size;
1326                 session->one_mmap = true;
1327         }
1328
1329         memset(mmaps, 0, sizeof(mmaps));
1330
1331         mmap_prot  = PROT_READ;
1332         mmap_flags = MAP_SHARED;
1333
1334         if (session->header.needs_swap) {
1335                 mmap_prot  |= PROT_WRITE;
1336                 mmap_flags = MAP_PRIVATE;
1337         }
1338 remap:
1339         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1340                    file_offset);
1341         if (buf == MAP_FAILED) {
1342                 pr_err("failed to mmap file\n");
1343                 err = -errno;
1344                 goto out_err;
1345         }
1346         mmaps[map_idx] = buf;
1347         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1348         file_pos = file_offset + head;
1349         if (session->one_mmap) {
1350                 session->one_mmap_addr = buf;
1351                 session->one_mmap_offset = file_offset;
1352         }
1353
1354 more:
1355         event = fetch_mmaped_event(session, head, mmap_size, buf);
1356         if (!event) {
1357                 if (mmaps[map_idx]) {
1358                         munmap(mmaps[map_idx], mmap_size);
1359                         mmaps[map_idx] = NULL;
1360                 }
1361
1362                 page_offset = page_size * (head / page_size);
1363                 file_offset += page_offset;
1364                 head -= page_offset;
1365                 goto remap;
1366         }
1367
1368         size = event->header.size;
1369
1370         if (size < sizeof(struct perf_event_header) ||
1371             (skip = perf_session__process_event(session, event, tool, file_pos))
1372                                                                         < 0) {
1373                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1374                        file_offset + head, event->header.size,
1375                        event->header.type);
1376                 err = -EINVAL;
1377                 goto out_err;
1378         }
1379
1380         if (skip)
1381                 size += skip;
1382
1383         head += size;
1384         file_pos += size;
1385
1386         ui_progress__update(&prog, size);
1387
1388         if (session_done())
1389                 goto out;
1390
1391         if (file_pos < file_size)
1392                 goto more;
1393
1394 out:
1395         /* do the final flush for ordered samples */
1396         session->ordered_events.next_flush = ULLONG_MAX;
1397         err = ordered_events__flush(session, tool);
1398 out_err:
1399         ui_progress__finish();
1400         perf_session__warn_about_errors(session, tool);
1401         perf_session_free_sample_buffers(session);
1402         session->one_mmap = false;
1403         return err;
1404 }
1405
1406 int perf_session__process_events(struct perf_session *session,
1407                                  struct perf_tool *tool)
1408 {
1409         u64 size = perf_data_file__size(session->file);
1410         int err;
1411
1412         if (perf_session__register_idle_thread(session) == NULL)
1413                 return -ENOMEM;
1414
1415         if (!perf_data_file__is_pipe(session->file))
1416                 err = __perf_session__process_events(session,
1417                                                      session->header.data_offset,
1418                                                      session->header.data_size,
1419                                                      size, tool);
1420         else
1421                 err = __perf_session__process_pipe_events(session, tool);
1422
1423         return err;
1424 }
1425
1426 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1427 {
1428         struct perf_evsel *evsel;
1429
1430         evlist__for_each(session->evlist, evsel) {
1431                 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1432                         return true;
1433         }
1434
1435         pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1436         return false;
1437 }
1438
1439 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1440                                      const char *symbol_name, u64 addr)
1441 {
1442         char *bracket;
1443         enum map_type i;
1444         struct ref_reloc_sym *ref;
1445
1446         ref = zalloc(sizeof(struct ref_reloc_sym));
1447         if (ref == NULL)
1448                 return -ENOMEM;
1449
1450         ref->name = strdup(symbol_name);
1451         if (ref->name == NULL) {
1452                 free(ref);
1453                 return -ENOMEM;
1454         }
1455
1456         bracket = strchr(ref->name, ']');
1457         if (bracket)
1458                 *bracket = '\0';
1459
1460         ref->addr = addr;
1461
1462         for (i = 0; i < MAP__NR_TYPES; ++i) {
1463                 struct kmap *kmap = map__kmap(maps[i]);
1464                 kmap->ref_reloc_sym = ref;
1465         }
1466
1467         return 0;
1468 }
1469
1470 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1471 {
1472         return machines__fprintf_dsos(&session->machines, fp);
1473 }
1474
1475 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1476                                           bool (skip)(struct dso *dso, int parm), int parm)
1477 {
1478         return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1479 }
1480
1481 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1482 {
1483         struct perf_evsel *pos;
1484         size_t ret = fprintf(fp, "Aggregated stats:\n");
1485
1486         ret += events_stats__fprintf(&session->stats, fp);
1487
1488         evlist__for_each(session->evlist, pos) {
1489                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1490                 ret += events_stats__fprintf(&pos->hists.stats, fp);
1491         }
1492
1493         return ret;
1494 }
1495
1496 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1497 {
1498         /*
1499          * FIXME: Here we have to actually print all the machines in this
1500          * session, not just the host...
1501          */
1502         return machine__fprintf(&session->machines.host, fp);
1503 }
1504
1505 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1506                                               unsigned int type)
1507 {
1508         struct perf_evsel *pos;
1509
1510         evlist__for_each(session->evlist, pos) {
1511                 if (pos->attr.type == type)
1512                         return pos;
1513         }
1514         return NULL;
1515 }
1516
1517 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1518                           struct addr_location *al,
1519                           unsigned int print_opts, unsigned int stack_depth)
1520 {
1521         struct callchain_cursor_node *node;
1522         int print_ip = print_opts & PRINT_IP_OPT_IP;
1523         int print_sym = print_opts & PRINT_IP_OPT_SYM;
1524         int print_dso = print_opts & PRINT_IP_OPT_DSO;
1525         int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1526         int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1527         int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1528         char s = print_oneline ? ' ' : '\t';
1529
1530         if (symbol_conf.use_callchain && sample->callchain) {
1531                 struct addr_location node_al;
1532
1533                 if (machine__resolve_callchain(al->machine, evsel, al->thread,
1534                                                sample, NULL, NULL,
1535                                                PERF_MAX_STACK_DEPTH) != 0) {
1536                         if (verbose)
1537                                 error("Failed to resolve callchain. Skipping\n");
1538                         return;
1539                 }
1540                 callchain_cursor_commit(&callchain_cursor);
1541
1542                 if (print_symoffset)
1543                         node_al = *al;
1544
1545                 while (stack_depth) {
1546                         u64 addr = 0;
1547
1548                         node = callchain_cursor_current(&callchain_cursor);
1549                         if (!node)
1550                                 break;
1551
1552                         if (node->sym && node->sym->ignore)
1553                                 goto next;
1554
1555                         if (print_ip)
1556                                 printf("%c%16" PRIx64, s, node->ip);
1557
1558                         if (node->map)
1559                                 addr = node->map->map_ip(node->map, node->ip);
1560
1561                         if (print_sym) {
1562                                 printf(" ");
1563                                 if (print_symoffset) {
1564                                         node_al.addr = addr;
1565                                         node_al.map  = node->map;
1566                                         symbol__fprintf_symname_offs(node->sym, &node_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_srcline)
1578                                 map__fprintf_srcline(node->map, addr, "\n  ",
1579                                                      stdout);
1580
1581                         if (!print_oneline)
1582                                 printf("\n");
1583
1584                         stack_depth--;
1585 next:
1586                         callchain_cursor_advance(&callchain_cursor);
1587                 }
1588
1589         } else {
1590                 if (al->sym && al->sym->ignore)
1591                         return;
1592
1593                 if (print_ip)
1594                         printf("%16" PRIx64, sample->ip);
1595
1596                 if (print_sym) {
1597                         printf(" ");
1598                         if (print_symoffset)
1599                                 symbol__fprintf_symname_offs(al->sym, al,
1600                                                              stdout);
1601                         else
1602                                 symbol__fprintf_symname(al->sym, stdout);
1603                 }
1604
1605                 if (print_dso) {
1606                         printf(" (");
1607                         map__fprintf_dsoname(al->map, stdout);
1608                         printf(")");
1609                 }
1610
1611                 if (print_srcline)
1612                         map__fprintf_srcline(al->map, al->addr, "\n  ", stdout);
1613         }
1614 }
1615
1616 int perf_session__cpu_bitmap(struct perf_session *session,
1617                              const char *cpu_list, unsigned long *cpu_bitmap)
1618 {
1619         int i, err = -1;
1620         struct cpu_map *map;
1621
1622         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1623                 struct perf_evsel *evsel;
1624
1625                 evsel = perf_session__find_first_evtype(session, i);
1626                 if (!evsel)
1627                         continue;
1628
1629                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1630                         pr_err("File does not contain CPU events. "
1631                                "Remove -c option to proceed.\n");
1632                         return -1;
1633                 }
1634         }
1635
1636         map = cpu_map__new(cpu_list);
1637         if (map == NULL) {
1638                 pr_err("Invalid cpu_list\n");
1639                 return -1;
1640         }
1641
1642         for (i = 0; i < map->nr; i++) {
1643                 int cpu = map->map[i];
1644
1645                 if (cpu >= MAX_NR_CPUS) {
1646                         pr_err("Requested CPU %d too large. "
1647                                "Consider raising MAX_NR_CPUS\n", cpu);
1648                         goto out_delete_map;
1649                 }
1650
1651                 set_bit(cpu, cpu_bitmap);
1652         }
1653
1654         err = 0;
1655
1656 out_delete_map:
1657         cpu_map__delete(map);
1658         return err;
1659 }
1660
1661 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1662                                 bool full)
1663 {
1664         struct stat st;
1665         int fd, ret;
1666
1667         if (session == NULL || fp == NULL)
1668                 return;
1669
1670         fd = perf_data_file__fd(session->file);
1671
1672         ret = fstat(fd, &st);
1673         if (ret == -1)
1674                 return;
1675
1676         fprintf(fp, "# ========\n");
1677         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1678         perf_header__fprintf_info(session, fp, full);
1679         fprintf(fp, "# ========\n#\n");
1680 }
1681
1682
1683 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1684                                              const struct perf_evsel_str_handler *assocs,
1685                                              size_t nr_assocs)
1686 {
1687         struct perf_evsel *evsel;
1688         size_t i;
1689         int err;
1690
1691         for (i = 0; i < nr_assocs; i++) {
1692                 /*
1693                  * Adding a handler for an event not in the session,
1694                  * just ignore it.
1695                  */
1696                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1697                 if (evsel == NULL)
1698                         continue;
1699
1700                 err = -EEXIST;
1701                 if (evsel->handler != NULL)
1702                         goto out;
1703                 evsel->handler = assocs[i].handler;
1704         }
1705
1706         err = 0;
1707 out:
1708         return err;
1709 }