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_samples.samples);
  79         INIT_LIST_HEAD(&session->ordered_samples.sample_cache);
  80         INIT_LIST_HEAD(&session->ordered_samples.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_samples && !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_samples = 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_samples)
 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 sample_queue {
 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_samples *os = &session->ordered_samples;
 457
 458         while (!list_empty(&os->to_free)) {
 459                 struct sample_queue *sq;
 460
 461                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 462                 list_del(&sq->list);
 463                 free(sq);
 464         }
 465 }
 466
 467 static int perf_session_deliver_event(struct perf_session *session,
 468                                       union perf_event *event,
 469                                       struct perf_sample *sample,
 470                                       struct perf_tool *tool,
 471                                       u64 file_offset);
 472
 473 static int flush_sample_queue(struct perf_session *s,
 474                        struct perf_tool *tool)
 475 {
 476         struct ordered_samples *os = &s->ordered_samples;
 477         struct list_head *head = &os->samples;
 478         struct sample_queue *tmp, *iter;
 479         struct perf_sample sample;
 480         u64 limit = os->next_flush;
 481         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 482         bool show_progress = limit == ULLONG_MAX;
 483         struct ui_progress prog;
 484         int ret;
 485
 486         if (!tool->ordered_samples || !limit)
 487                 return 0;
 488
 489         if (show_progress)
 490                 ui_progress__init(&prog, os->nr_samples, "Processing time ordered events...");
 491
 492         list_for_each_entry_safe(iter, tmp, head, list) {
 493                 if (session_done())
 494                         return 0;
 495
 496                 if (iter->timestamp > limit)
 497                         break;
 498
 499                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
 500                 if (ret)
 501                         pr_err("Can't parse sample, err = %d\n", ret);
 502                 else {
 503                         ret = perf_session_deliver_event(s, iter->event, &sample, tool,
 504                                                          iter->file_offset);
 505                         if (ret)
 506                                 return ret;
 507                 }
 508
 509                 os->last_flush = iter->timestamp;
 510                 list_del(&iter->list);
 511                 list_add(&iter->list, &os->sample_cache);
 512                 os->nr_samples--;
 513
 514                 if (show_progress)
 515                         ui_progress__update(&prog, 1);
 516         }
 517
 518         if (list_empty(head)) {
 519                 os->last_sample = NULL;
 520         } else if (last_ts <= limit) {
 521                 os->last_sample =
 522                         list_entry(head->prev, struct sample_queue, list);
 523         }
 524
 525         return 0;
 526 }
 527
 528 /*
 529  * When perf record finishes a pass on every buffers, it records this pseudo
 530  * event.
 531  * We record the max timestamp t found in the pass n.
 532  * Assuming these timestamps are monotonic across cpus, we know that if
 533  * a buffer still has events with timestamps below t, they will be all
 534  * available and then read in the pass n + 1.
 535  * Hence when we start to read the pass n + 2, we can safely flush every
 536  * events with timestamps below t.
 537  *
 538  *    ============ PASS n =================
 539  *       CPU 0         |   CPU 1
 540  *                     |
 541  *    cnt1 timestamps  |   cnt2 timestamps
 542  *          1          |         2
 543  *          2          |         3
 544  *          -          |         4  <--- max recorded
 545  *
 546  *    ============ PASS n + 1 ==============
 547  *       CPU 0         |   CPU 1
 548  *                     |
 549  *    cnt1 timestamps  |   cnt2 timestamps
 550  *          3          |         5
 551  *          4          |         6
 552  *          5          |         7 <---- max recorded
 553  *
 554  *      Flush every events below timestamp 4
 555  *
 556  *    ============ PASS n + 2 ==============
 557  *       CPU 0         |   CPU 1
 558  *                     |
 559  *    cnt1 timestamps  |   cnt2 timestamps
 560  *          6          |         8
 561  *          7          |         9
 562  *          -          |         10
 563  *
 564  *      Flush every events below timestamp 7
 565  *      etc...
 566  */
 567 static int process_finished_round(struct perf_tool *tool,
 568                                   union perf_event *event __maybe_unused,
 569                                   struct perf_session *session)
 570 {
 571         int ret = flush_sample_queue(session, tool);
 572         if (!ret)
 573                 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 574
 575         return ret;
 576 }
 577
 578 /* The queue is ordered by time */
 579 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 580 {
 581         struct ordered_samples *os = &s->ordered_samples;
 582         struct sample_queue *sample = os->last_sample;
 583         u64 timestamp = new->timestamp;
 584         struct list_head *p;
 585
 586         ++os->nr_samples;
 587         os->last_sample = new;
 588
 589         if (!sample) {
 590                 list_add(&new->list, &os->samples);
 591                 os->max_timestamp = timestamp;
 592                 return;
 593         }
 594
 595         /*
 596          * last_sample might point to some random place in the list as it's
 597          * the last queued event. We expect that the new event is close to
 598          * this.
 599          */
 600         if (sample->timestamp <= timestamp) {
 601                 while (sample->timestamp <= timestamp) {
 602                         p = sample->list.next;
 603                         if (p == &os->samples) {
 604                                 list_add_tail(&new->list, &os->samples);
 605                                 os->max_timestamp = timestamp;
 606                                 return;
 607                         }
 608                         sample = list_entry(p, struct sample_queue, list);
 609                 }
 610                 list_add_tail(&new->list, &sample->list);
 611         } else {
 612                 while (sample->timestamp > timestamp) {
 613                         p = sample->list.prev;
 614                         if (p == &os->samples) {
 615                                 list_add(&new->list, &os->samples);
 616                                 return;
 617                         }
 618                         sample = list_entry(p, struct sample_queue, list);
 619                 }
 620                 list_add(&new->list, &sample->list);
 621         }
 622 }
 623
 624 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 625
 626 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 627                                     struct perf_sample *sample, u64 file_offset)
 628 {
 629         struct ordered_samples *os = &s->ordered_samples;
 630         struct list_head *sc = &os->sample_cache;
 631         u64 timestamp = sample->time;
 632         struct sample_queue *new;
 633
 634         if (!timestamp || timestamp == ~0ULL)
 635                 return -ETIME;
 636
 637         if (timestamp < s->ordered_samples.last_flush) {
 638                 printf("Warning: Timestamp below last timeslice flush\n");
 639                 return -EINVAL;
 640         }
 641
 642         if (!list_empty(sc)) {
 643                 new = list_entry(sc->next, struct sample_queue, list);
 644                 list_del(&new->list);
 645         } else if (os->sample_buffer) {
 646                 new = os->sample_buffer + os->sample_buffer_idx;
 647                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 648                         os->sample_buffer = NULL;
 649         } else {
 650                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 651                 if (!os->sample_buffer)
 652                         return -ENOMEM;
 653                 list_add(&os->sample_buffer->list, &os->to_free);
 654                 os->sample_buffer_idx = 2;
 655                 new = os->sample_buffer + 1;
 656         }
 657
 658         new->timestamp = timestamp;
 659         new->file_offset = file_offset;
 660         new->event = event;
 661
 662         __queue_event(new, s);
 663
 664         return 0;
 665 }
 666
 667 static void callchain__printf(struct perf_sample *sample)
 668 {
 669         unsigned int i;
 670
 671         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 672
 673         for (i = 0; i < sample->callchain->nr; i++)
 674                 printf("..... %2d: %016" PRIx64 "\n",
 675                        i, sample->callchain->ips[i]);
 676 }
 677
 678 static void branch_stack__printf(struct perf_sample *sample)
 679 {
 680         uint64_t i;
 681
 682         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 683
 684         for (i = 0; i < sample->branch_stack->nr; i++)
 685                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 686                         i, sample->branch_stack->entries[i].from,
 687                         sample->branch_stack->entries[i].to);
 688 }
 689
 690 static void regs_dump__printf(u64 mask, u64 *regs)
 691 {
 692         unsigned rid, i = 0;
 693
 694         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 695                 u64 val = regs[i++];
 696
 697                 printf(".... %-5s 0x%" PRIx64 "\n",
 698                        perf_reg_name(rid), val);
 699         }
 700 }
 701
 702 static void regs_user__printf(struct perf_sample *sample)
 703 {
 704         struct regs_dump *user_regs = &sample->user_regs;
 705
 706         if (user_regs->regs) {
 707                 u64 mask = user_regs->mask;
 708                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
 709                 regs_dump__printf(mask, user_regs->regs);
 710         }
 711 }
 712
 713 static void stack_user__printf(struct stack_dump *dump)
 714 {
 715         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 716                dump->size, dump->offset);
 717 }
 718
 719 static void perf_session__print_tstamp(struct perf_session *session,
 720                                        union perf_event *event,
 721                                        struct perf_sample *sample)
 722 {
 723         u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
 724
 725         if (event->header.type != PERF_RECORD_SAMPLE &&
 726             !perf_evlist__sample_id_all(session->evlist)) {
 727                 fputs("-1 -1 ", stdout);
 728                 return;
 729         }
 730
 731         if ((sample_type & PERF_SAMPLE_CPU))
 732                 printf("%u ", sample->cpu);
 733
 734         if (sample_type & PERF_SAMPLE_TIME)
 735                 printf("%" PRIu64 " ", sample->time);
 736 }
 737
 738 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
 739 {
 740         printf("... sample_read:\n");
 741
 742         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 743                 printf("...... time enabled %016" PRIx64 "\n",
 744                        sample->read.time_enabled);
 745
 746         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 747                 printf("...... time running %016" PRIx64 "\n",
 748                        sample->read.time_running);
 749
 750         if (read_format & PERF_FORMAT_GROUP) {
 751                 u64 i;
 752
 753                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
 754
 755                 for (i = 0; i < sample->read.group.nr; i++) {
 756                         struct sample_read_value *value;
 757
 758                         value = &sample->read.group.values[i];
 759                         printf("..... id %016" PRIx64
 760                                ", value %016" PRIx64 "\n",
 761                                value->id, value->value);
 762                 }
 763         } else
 764                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
 765                         sample->read.one.id, sample->read.one.value);
 766 }
 767
 768 static void dump_event(struct perf_session *session, union perf_event *event,
 769                        u64 file_offset, struct perf_sample *sample)
 770 {
 771         if (!dump_trace)
 772                 return;
 773
 774         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 775                file_offset, event->header.size, event->header.type);
 776
 777         trace_event(event);
 778
 779         if (sample)
 780                 perf_session__print_tstamp(session, event, sample);
 781
 782         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 783                event->header.size, perf_event__name(event->header.type));
 784 }
 785
 786 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 787                         struct perf_sample *sample)
 788 {
 789         u64 sample_type;
 790
 791         if (!dump_trace)
 792                 return;
 793
 794         printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 795                event->header.misc, sample->pid, sample->tid, sample->ip,
 796                sample->period, sample->addr);
 797
 798         sample_type = evsel->attr.sample_type;
 799
 800         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 801                 callchain__printf(sample);
 802
 803         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
 804                 branch_stack__printf(sample);
 805
 806         if (sample_type & PERF_SAMPLE_REGS_USER)
 807                 regs_user__printf(sample);
 808
 809         if (sample_type & PERF_SAMPLE_STACK_USER)
 810                 stack_user__printf(&sample->user_stack);
 811
 812         if (sample_type & PERF_SAMPLE_WEIGHT)
 813                 printf("... weight: %" PRIu64 "\n", sample->weight);
 814
 815         if (sample_type & PERF_SAMPLE_DATA_SRC)
 816                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
 817
 818         if (sample_type & PERF_SAMPLE_TRANSACTION)
 819                 printf("... transaction: %" PRIx64 "\n", sample->transaction);
 820
 821         if (sample_type & PERF_SAMPLE_READ)
 822                 sample_read__printf(sample, evsel->attr.read_format);
 823 }
 824
 825 static struct machine *
 826         perf_session__find_machine_for_cpumode(struct perf_session *session,
 827                                                union perf_event *event,
 828                                                struct perf_sample *sample)
 829 {
 830         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 831         struct machine *machine;
 832
 833         if (perf_guest &&
 834             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
 835              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
 836                 u32 pid;
 837
 838                 if (event->header.type == PERF_RECORD_MMAP
 839                     || event->header.type == PERF_RECORD_MMAP2)
 840                         pid = event->mmap.pid;
 841                 else
 842                         pid = sample->pid;
 843
 844                 machine = perf_session__find_machine(session, pid);
 845                 if (!machine)
 846                         machine = perf_session__findnew_machine(session,
 847                                                 DEFAULT_GUEST_KERNEL_ID);
 848                 return machine;
 849         }
 850
 851         return &session->machines.host;
 852 }
 853
 854 static int deliver_sample_value(struct perf_session *session,
 855                                 struct perf_tool *tool,
 856                                 union perf_event *event,
 857                                 struct perf_sample *sample,
 858                                 struct sample_read_value *v,
 859                                 struct machine *machine)
 860 {
 861         struct perf_sample_id *sid;
 862
 863         sid = perf_evlist__id2sid(session->evlist, v->id);
 864         if (sid) {
 865                 sample->id     = v->id;
 866                 sample->period = v->value - sid->period;
 867                 sid->period    = v->value;
 868         }
 869
 870         if (!sid || sid->evsel == NULL) {
 871                 ++session->stats.nr_unknown_id;
 872                 return 0;
 873         }
 874
 875         return tool->sample(tool, event, sample, sid->evsel, machine);
 876 }
 877
 878 static int deliver_sample_group(struct perf_session *session,
 879                                 struct perf_tool *tool,
 880                                 union  perf_event *event,
 881                                 struct perf_sample *sample,
 882                                 struct machine *machine)
 883 {
 884         int ret = -EINVAL;
 885         u64 i;
 886
 887         for (i = 0; i < sample->read.group.nr; i++) {
 888                 ret = deliver_sample_value(session, tool, event, sample,
 889                                            &sample->read.group.values[i],
 890                                            machine);
 891                 if (ret)
 892                         break;
 893         }
 894
 895         return ret;
 896 }
 897
 898 static int
 899 perf_session__deliver_sample(struct perf_session *session,
 900                              struct perf_tool *tool,
 901                              union  perf_event *event,
 902                              struct perf_sample *sample,
 903                              struct perf_evsel *evsel,
 904                              struct machine *machine)
 905 {
 906         /* We know evsel != NULL. */
 907         u64 sample_type = evsel->attr.sample_type;
 908         u64 read_format = evsel->attr.read_format;
 909
 910         /* Standard sample delievery. */
 911         if (!(sample_type & PERF_SAMPLE_READ))
 912                 return tool->sample(tool, event, sample, evsel, machine);
 913
 914         /* For PERF_SAMPLE_READ we have either single or group mode. */
 915         if (read_format & PERF_FORMAT_GROUP)
 916                 return deliver_sample_group(session, tool, event, sample,
 917                                             machine);
 918         else
 919                 return deliver_sample_value(session, tool, event, sample,
 920                                             &sample->read.one, machine);
 921 }
 922
 923 static int perf_session_deliver_event(struct perf_session *session,
 924                                       union perf_event *event,
 925                                       struct perf_sample *sample,
 926                                       struct perf_tool *tool,
 927                                       u64 file_offset)
 928 {
 929         struct perf_evsel *evsel;
 930         struct machine *machine;
 931
 932         dump_event(session, event, file_offset, sample);
 933
 934         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 935         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
 936                 /*
 937                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
 938                  * because the tools right now may apply filters, discarding
 939                  * some of the samples. For consistency, in the future we
 940                  * should have something like nr_filtered_samples and remove
 941                  * the sample->period from total_sample_period, etc, KISS for
 942                  * now tho.
 943                  *
 944                  * Also testing against NULL allows us to handle files without
 945                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
 946                  * future probably it'll be a good idea to restrict event
 947                  * processing via perf_session to files with both set.
 948                  */
 949                 hists__inc_nr_events(&evsel->hists, event->header.type);
 950         }
 951
 952         machine = perf_session__find_machine_for_cpumode(session, event,
 953                                                          sample);
 954
 955         switch (event->header.type) {
 956         case PERF_RECORD_SAMPLE:
 957                 dump_sample(evsel, event, sample);
 958                 if (evsel == NULL) {
 959                         ++session->stats.nr_unknown_id;
 960                         return 0;
 961                 }
 962                 if (machine == NULL) {
 963                         ++session->stats.nr_unprocessable_samples;
 964                         return 0;
 965                 }
 966                 return perf_session__deliver_sample(session, tool, event,
 967                                                     sample, evsel, machine);
 968         case PERF_RECORD_MMAP:
 969                 return tool->mmap(tool, event, sample, machine);
 970         case PERF_RECORD_MMAP2:
 971                 return tool->mmap2(tool, event, sample, machine);
 972         case PERF_RECORD_COMM:
 973                 return tool->comm(tool, event, sample, machine);
 974         case PERF_RECORD_FORK:
 975                 return tool->fork(tool, event, sample, machine);
 976         case PERF_RECORD_EXIT:
 977                 return tool->exit(tool, event, sample, machine);
 978         case PERF_RECORD_LOST:
 979                 if (tool->lost == perf_event__process_lost)
 980                         session->stats.total_lost += event->lost.lost;
 981                 return tool->lost(tool, event, sample, machine);
 982         case PERF_RECORD_READ:
 983                 return tool->read(tool, event, sample, evsel, machine);
 984         case PERF_RECORD_THROTTLE:
 985                 return tool->throttle(tool, event, sample, machine);
 986         case PERF_RECORD_UNTHROTTLE:
 987                 return tool->unthrottle(tool, event, sample, machine);
 988         default:
 989                 ++session->stats.nr_unknown_events;
 990                 return -1;
 991         }
 992 }
 993
 994 static s64 perf_session__process_user_event(struct perf_session *session,
 995                                             union perf_event *event,
 996                                             struct perf_tool *tool,
 997                                             u64 file_offset)
 998 {
 999         int fd = perf_data_file__fd(session->file);
1000         int err;
1001
1002         dump_event(session, event, file_offset, NULL);
1003
1004         /* These events are processed right away */
1005         switch (event->header.type) {
1006         case PERF_RECORD_HEADER_ATTR:
1007                 err = tool->attr(tool, event, &session->evlist);
1008                 if (err == 0)
1009                         perf_session__set_id_hdr_size(session);
1010                 return err;
1011         case PERF_RECORD_HEADER_EVENT_TYPE:
1012                 /*
1013                  * Depreceated, but we need to handle it for sake
1014                  * of old data files create in pipe mode.
1015                  */
1016                 return 0;
1017         case PERF_RECORD_HEADER_TRACING_DATA:
1018                 /* setup for reading amidst mmap */
1019                 lseek(fd, file_offset, SEEK_SET);
1020                 return tool->tracing_data(tool, event, session);
1021         case PERF_RECORD_HEADER_BUILD_ID:
1022                 return tool->build_id(tool, event, session);
1023         case PERF_RECORD_FINISHED_ROUND:
1024                 return tool->finished_round(tool, event, session);
1025         default:
1026                 return -EINVAL;
1027         }
1028 }
1029
1030 static void event_swap(union perf_event *event, bool sample_id_all)
1031 {
1032         perf_event__swap_op swap;
1033
1034         swap = perf_event__swap_ops[event->header.type];
1035         if (swap)
1036                 swap(event, sample_id_all);
1037 }
1038
1039 static s64 perf_session__process_event(struct perf_session *session,
1040                                        union perf_event *event,
1041                                        struct perf_tool *tool,
1042                                        u64 file_offset)
1043 {
1044         struct perf_sample sample;
1045         int ret;
1046
1047         if (session->header.needs_swap)
1048                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1049
1050         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1051                 return -EINVAL;
1052
1053         events_stats__inc(&session->stats, event->header.type);
1054
1055         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1056                 return perf_session__process_user_event(session, event, tool, file_offset);
1057
1058         /*
1059          * For all kernel events we get the sample data
1060          */
1061         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1062         if (ret)
1063                 return ret;
1064
1065         if (tool->ordered_samples) {
1066                 ret = perf_session_queue_event(session, event, &sample,
1067                                                file_offset);
1068                 if (ret != -ETIME)
1069                         return ret;
1070         }
1071
1072         return perf_session_deliver_event(session, event, &sample, tool,
1073                                           file_offset);
1074 }
1075
1076 void perf_event_header__bswap(struct perf_event_header *hdr)
1077 {
1078         hdr->type = bswap_32(hdr->type);
1079         hdr->misc = bswap_16(hdr->misc);
1080         hdr->size = bswap_16(hdr->size);
1081 }
1082
1083 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1084 {
1085         return machine__findnew_thread(&session->machines.host, -1, pid);
1086 }
1087
1088 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1089 {
1090         struct thread *thread;
1091
1092         thread = machine__findnew_thread(&session->machines.host, 0, 0);
1093         if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1094                 pr_err("problem inserting idle task.\n");
1095                 thread = NULL;
1096         }
1097
1098         return thread;
1099 }
1100
1101 static void perf_session__warn_about_errors(const struct perf_session *session,
1102                                             const struct perf_tool *tool)
1103 {
1104         if (tool->lost == perf_event__process_lost &&
1105             session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1106                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1107                             "Check IO/CPU overload!\n\n",
1108                             session->stats.nr_events[0],
1109                             session->stats.nr_events[PERF_RECORD_LOST]);
1110         }
1111
1112         if (session->stats.nr_unknown_events != 0) {
1113                 ui__warning("Found %u unknown events!\n\n"
1114                             "Is this an older tool processing a perf.data "
1115                             "file generated by a more recent tool?\n\n"
1116                             "If that is not the case, consider "
1117                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1118                             session->stats.nr_unknown_events);
1119         }
1120
1121         if (session->stats.nr_unknown_id != 0) {
1122                 ui__warning("%u samples with id not present in the header\n",
1123                             session->stats.nr_unknown_id);
1124         }
1125
1126         if (session->stats.nr_invalid_chains != 0) {
1127                 ui__warning("Found invalid callchains!\n\n"
1128                             "%u out of %u events were discarded for this reason.\n\n"
1129                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1130                             session->stats.nr_invalid_chains,
1131                             session->stats.nr_events[PERF_RECORD_SAMPLE]);
1132         }
1133
1134         if (session->stats.nr_unprocessable_samples != 0) {
1135                 ui__warning("%u unprocessable samples recorded.\n"
1136                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1137                             session->stats.nr_unprocessable_samples);
1138         }
1139 }
1140
1141 volatile int session_done;
1142
1143 static int __perf_session__process_pipe_events(struct perf_session *session,
1144                                                struct perf_tool *tool)
1145 {
1146         int fd = perf_data_file__fd(session->file);
1147         union perf_event *event;
1148         uint32_t size, cur_size = 0;
1149         void *buf = NULL;
1150         s64 skip = 0;
1151         u64 head;
1152         ssize_t err;
1153         void *p;
1154
1155         perf_tool__fill_defaults(tool);
1156
1157         head = 0;
1158         cur_size = sizeof(union perf_event);
1159
1160         buf = malloc(cur_size);
1161         if (!buf)
1162                 return -errno;
1163 more:
1164         event = buf;
1165         err = readn(fd, event, sizeof(struct perf_event_header));
1166         if (err <= 0) {
1167                 if (err == 0)
1168                         goto done;
1169
1170                 pr_err("failed to read event header\n");
1171                 goto out_err;
1172         }
1173
1174         if (session->header.needs_swap)
1175                 perf_event_header__bswap(&event->header);
1176
1177         size = event->header.size;
1178         if (size < sizeof(struct perf_event_header)) {
1179                 pr_err("bad event header size\n");
1180                 goto out_err;
1181         }
1182
1183         if (size > cur_size) {
1184                 void *new = realloc(buf, size);
1185                 if (!new) {
1186                         pr_err("failed to allocate memory to read event\n");
1187                         goto out_err;
1188                 }
1189                 buf = new;
1190                 cur_size = size;
1191                 event = buf;
1192         }
1193         p = event;
1194         p += sizeof(struct perf_event_header);
1195
1196         if (size - sizeof(struct perf_event_header)) {
1197                 err = readn(fd, p, size - sizeof(struct perf_event_header));
1198                 if (err <= 0) {
1199                         if (err == 0) {
1200                                 pr_err("unexpected end of event stream\n");
1201                                 goto done;
1202                         }
1203
1204                         pr_err("failed to read event data\n");
1205                         goto out_err;
1206                 }
1207         }
1208
1209         if ((skip = perf_session__process_event(session, event, tool, head)) < 0) {
1210                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1211                        head, event->header.size, event->header.type);
1212                 err = -EINVAL;
1213                 goto out_err;
1214         }
1215
1216         head += size;
1217
1218         if (skip > 0)
1219                 head += skip;
1220
1221         if (!session_done())
1222                 goto more;
1223 done:
1224         /* do the final flush for ordered samples */
1225         session->ordered_samples.next_flush = ULLONG_MAX;
1226         err = flush_sample_queue(session, tool);
1227 out_err:
1228         free(buf);
1229         perf_session__warn_about_errors(session, tool);
1230         perf_session_free_sample_buffers(session);
1231         return err;
1232 }
1233
1234 static union perf_event *
1235 fetch_mmaped_event(struct perf_session *session,
1236                    u64 head, size_t mmap_size, char *buf)
1237 {
1238         union perf_event *event;
1239
1240         /*
1241          * Ensure we have enough space remaining to read
1242          * the size of the event in the headers.
1243          */
1244         if (head + sizeof(event->header) > mmap_size)
1245                 return NULL;
1246
1247         event = (union perf_event *)(buf + head);
1248
1249         if (session->header.needs_swap)
1250                 perf_event_header__bswap(&event->header);
1251
1252         if (head + event->header.size > mmap_size) {
1253                 /* We're not fetching the event so swap back again */
1254                 if (session->header.needs_swap)
1255                         perf_event_header__bswap(&event->header);
1256                 return NULL;
1257         }
1258
1259         return event;
1260 }
1261
1262 /*
1263  * On 64bit we can mmap the data file in one go. No need for tiny mmap
1264  * slices. On 32bit we use 32MB.
1265  */
1266 #if BITS_PER_LONG == 64
1267 #define MMAP_SIZE ULLONG_MAX
1268 #define NUM_MMAPS 1
1269 #else
1270 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1271 #define NUM_MMAPS 128
1272 #endif
1273
1274 int __perf_session__process_events(struct perf_session *session,
1275                                    u64 data_offset, u64 data_size,
1276                                    u64 file_size, struct perf_tool *tool)
1277 {
1278         int fd = perf_data_file__fd(session->file);
1279         u64 head, page_offset, file_offset, file_pos, size;
1280         int err, mmap_prot, mmap_flags, map_idx = 0;
1281         size_t  mmap_size;
1282         char *buf, *mmaps[NUM_MMAPS];
1283         union perf_event *event;
1284         struct ui_progress prog;
1285         s64 skip;
1286
1287         perf_tool__fill_defaults(tool);
1288
1289         page_offset = page_size * (data_offset / page_size);
1290         file_offset = page_offset;
1291         head = data_offset - page_offset;
1292
1293         if (data_size && (data_offset + data_size < file_size))
1294                 file_size = data_offset + data_size;
1295
1296         ui_progress__init(&prog, file_size, "Processing events...");
1297
1298         mmap_size = MMAP_SIZE;
1299         if (mmap_size > file_size) {
1300                 mmap_size = file_size;
1301                 session->one_mmap = true;
1302         }
1303
1304         memset(mmaps, 0, sizeof(mmaps));
1305
1306         mmap_prot  = PROT_READ;
1307         mmap_flags = MAP_SHARED;
1308
1309         if (session->header.needs_swap) {
1310                 mmap_prot  |= PROT_WRITE;
1311                 mmap_flags = MAP_PRIVATE;
1312         }
1313 remap:
1314         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1315                    file_offset);
1316         if (buf == MAP_FAILED) {
1317                 pr_err("failed to mmap file\n");
1318                 err = -errno;
1319                 goto out_err;
1320         }
1321         mmaps[map_idx] = buf;
1322         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1323         file_pos = file_offset + head;
1324         if (session->one_mmap) {
1325                 session->one_mmap_addr = buf;
1326                 session->one_mmap_offset = file_offset;
1327         }
1328
1329 more:
1330         event = fetch_mmaped_event(session, head, mmap_size, buf);
1331         if (!event) {
1332                 if (mmaps[map_idx]) {
1333                         munmap(mmaps[map_idx], mmap_size);
1334                         mmaps[map_idx] = NULL;
1335                 }
1336
1337                 page_offset = page_size * (head / page_size);
1338                 file_offset += page_offset;
1339                 head -= page_offset;
1340                 goto remap;
1341         }
1342
1343         size = event->header.size;
1344
1345         if (size < sizeof(struct perf_event_header) ||
1346             (skip = perf_session__process_event(session, event, tool, file_pos))
1347                                                                         < 0) {
1348                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1349                        file_offset + head, event->header.size,
1350                        event->header.type);
1351                 err = -EINVAL;
1352                 goto out_err;
1353         }
1354
1355         if (skip)
1356                 size += skip;
1357
1358         head += size;
1359         file_pos += size;
1360
1361         ui_progress__update(&prog, size);
1362
1363         if (session_done())
1364                 goto out;
1365
1366         if (file_pos < file_size)
1367                 goto more;
1368
1369 out:
1370         /* do the final flush for ordered samples */
1371         session->ordered_samples.next_flush = ULLONG_MAX;
1372         err = flush_sample_queue(session, tool);
1373 out_err:
1374         ui_progress__finish();
1375         perf_session__warn_about_errors(session, tool);
1376         perf_session_free_sample_buffers(session);
1377         session->one_mmap = false;
1378         return err;
1379 }
1380
1381 int perf_session__process_events(struct perf_session *session,
1382                                  struct perf_tool *tool)
1383 {
1384         u64 size = perf_data_file__size(session->file);
1385         int err;
1386
1387         if (perf_session__register_idle_thread(session) == NULL)
1388                 return -ENOMEM;
1389
1390         if (!perf_data_file__is_pipe(session->file))
1391                 err = __perf_session__process_events(session,
1392                                                      session->header.data_offset,
1393                                                      session->header.data_size,
1394                                                      size, tool);
1395         else
1396                 err = __perf_session__process_pipe_events(session, tool);
1397
1398         return err;
1399 }
1400
1401 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1402 {
1403         struct perf_evsel *evsel;
1404
1405         evlist__for_each(session->evlist, evsel) {
1406                 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1407                         return true;
1408         }
1409
1410         pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1411         return false;
1412 }
1413
1414 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1415                                      const char *symbol_name, u64 addr)
1416 {
1417         char *bracket;
1418         enum map_type i;
1419         struct ref_reloc_sym *ref;
1420
1421         ref = zalloc(sizeof(struct ref_reloc_sym));
1422         if (ref == NULL)
1423                 return -ENOMEM;
1424
1425         ref->name = strdup(symbol_name);
1426         if (ref->name == NULL) {
1427                 free(ref);
1428                 return -ENOMEM;
1429         }
1430
1431         bracket = strchr(ref->name, ']');
1432         if (bracket)
1433                 *bracket = '\0';
1434
1435         ref->addr = addr;
1436
1437         for (i = 0; i < MAP__NR_TYPES; ++i) {
1438                 struct kmap *kmap = map__kmap(maps[i]);
1439                 kmap->ref_reloc_sym = ref;
1440         }
1441
1442         return 0;
1443 }
1444
1445 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1446 {
1447         return machines__fprintf_dsos(&session->machines, fp);
1448 }
1449
1450 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1451                                           bool (skip)(struct dso *dso, int parm), int parm)
1452 {
1453         return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1454 }
1455
1456 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1457 {
1458         struct perf_evsel *pos;
1459         size_t ret = fprintf(fp, "Aggregated stats:\n");
1460
1461         ret += events_stats__fprintf(&session->stats, fp);
1462
1463         evlist__for_each(session->evlist, pos) {
1464                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1465                 ret += events_stats__fprintf(&pos->hists.stats, fp);
1466         }
1467
1468         return ret;
1469 }
1470
1471 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1472 {
1473         /*
1474          * FIXME: Here we have to actually print all the machines in this
1475          * session, not just the host...
1476          */
1477         return machine__fprintf(&session->machines.host, fp);
1478 }
1479
1480 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1481                                               unsigned int type)
1482 {
1483         struct perf_evsel *pos;
1484
1485         evlist__for_each(session->evlist, pos) {
1486                 if (pos->attr.type == type)
1487                         return pos;
1488         }
1489         return NULL;
1490 }
1491
1492 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1493                           struct addr_location *al,
1494                           unsigned int print_opts, unsigned int stack_depth)
1495 {
1496         struct callchain_cursor_node *node;
1497         int print_ip = print_opts & PRINT_IP_OPT_IP;
1498         int print_sym = print_opts & PRINT_IP_OPT_SYM;
1499         int print_dso = print_opts & PRINT_IP_OPT_DSO;
1500         int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1501         int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1502         int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1503         char s = print_oneline ? ' ' : '\t';
1504
1505         if (symbol_conf.use_callchain && sample->callchain) {
1506                 struct addr_location node_al;
1507
1508                 if (machine__resolve_callchain(al->machine, evsel, al->thread,
1509                                                sample, NULL, NULL,
1510                                                PERF_MAX_STACK_DEPTH) != 0) {
1511                         if (verbose)
1512                                 error("Failed to resolve callchain. Skipping\n");
1513                         return;
1514                 }
1515                 callchain_cursor_commit(&callchain_cursor);
1516
1517                 if (print_symoffset)
1518                         node_al = *al;
1519
1520                 while (stack_depth) {
1521                         u64 addr = 0;
1522
1523                         node = callchain_cursor_current(&callchain_cursor);
1524                         if (!node)
1525                                 break;
1526
1527                         if (node->sym && node->sym->ignore)
1528                                 goto next;
1529
1530                         if (print_ip)
1531                                 printf("%c%16" PRIx64, s, node->ip);
1532
1533                         if (node->map)
1534                                 addr = node->map->map_ip(node->map, node->ip);
1535
1536                         if (print_sym) {
1537                                 printf(" ");
1538                                 if (print_symoffset) {
1539                                         node_al.addr = addr;
1540                                         node_al.map  = node->map;
1541                                         symbol__fprintf_symname_offs(node->sym, &node_al, stdout);
1542                                 } else
1543                                         symbol__fprintf_symname(node->sym, stdout);
1544                         }
1545
1546                         if (print_dso) {
1547                                 printf(" (");
1548                                 map__fprintf_dsoname(node->map, stdout);
1549                                 printf(")");
1550                         }
1551
1552                         if (print_srcline)
1553                                 map__fprintf_srcline(node->map, addr, "\n  ",
1554                                                      stdout);
1555
1556                         if (!print_oneline)
1557                                 printf("\n");
1558
1559                         stack_depth--;
1560 next:
1561                         callchain_cursor_advance(&callchain_cursor);
1562                 }
1563
1564         } else {
1565                 if (al->sym && al->sym->ignore)
1566                         return;
1567
1568                 if (print_ip)
1569                         printf("%16" PRIx64, sample->ip);
1570
1571                 if (print_sym) {
1572                         printf(" ");
1573                         if (print_symoffset)
1574                                 symbol__fprintf_symname_offs(al->sym, al,
1575                                                              stdout);
1576                         else
1577                                 symbol__fprintf_symname(al->sym, stdout);
1578                 }
1579
1580                 if (print_dso) {
1581                         printf(" (");
1582                         map__fprintf_dsoname(al->map, stdout);
1583                         printf(")");
1584                 }
1585
1586                 if (print_srcline)
1587                         map__fprintf_srcline(al->map, al->addr, "\n  ", stdout);
1588         }
1589 }
1590
1591 int perf_session__cpu_bitmap(struct perf_session *session,
1592                              const char *cpu_list, unsigned long *cpu_bitmap)
1593 {
1594         int i, err = -1;
1595         struct cpu_map *map;
1596
1597         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1598                 struct perf_evsel *evsel;
1599
1600                 evsel = perf_session__find_first_evtype(session, i);
1601                 if (!evsel)
1602                         continue;
1603
1604                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1605                         pr_err("File does not contain CPU events. "
1606                                "Remove -c option to proceed.\n");
1607                         return -1;
1608                 }
1609         }
1610
1611         map = cpu_map__new(cpu_list);
1612         if (map == NULL) {
1613                 pr_err("Invalid cpu_list\n");
1614                 return -1;
1615         }
1616
1617         for (i = 0; i < map->nr; i++) {
1618                 int cpu = map->map[i];
1619
1620                 if (cpu >= MAX_NR_CPUS) {
1621                         pr_err("Requested CPU %d too large. "
1622                                "Consider raising MAX_NR_CPUS\n", cpu);
1623                         goto out_delete_map;
1624                 }
1625
1626                 set_bit(cpu, cpu_bitmap);
1627         }
1628
1629         err = 0;
1630
1631 out_delete_map:
1632         cpu_map__delete(map);
1633         return err;
1634 }
1635
1636 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1637                                 bool full)
1638 {
1639         struct stat st;
1640         int fd, ret;
1641
1642         if (session == NULL || fp == NULL)
1643                 return;
1644
1645         fd = perf_data_file__fd(session->file);
1646
1647         ret = fstat(fd, &st);
1648         if (ret == -1)
1649                 return;
1650
1651         fprintf(fp, "# ========\n");
1652         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1653         perf_header__fprintf_info(session, fp, full);
1654         fprintf(fp, "# ========\n#\n");
1655 }
1656
1657
1658 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1659                                              const struct perf_evsel_str_handler *assocs,
1660                                              size_t nr_assocs)
1661 {
1662         struct perf_evsel *evsel;
1663         size_t i;
1664         int err;
1665
1666         for (i = 0; i < nr_assocs; i++) {
1667                 /*
1668                  * Adding a handler for an event not in the session,
1669                  * just ignore it.
1670                  */
1671                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1672                 if (evsel == NULL)
1673                         continue;
1674
1675                 err = -EEXIST;
1676                 if (evsel->handler != NULL)
1677                         goto out;
1678                 evsel->handler = assocs[i].handler;
1679         }
1680
1681         err = 0;
1682 out:
1683         return err;
1684 }