tools/perf/util/session.c

   1 #include <linux/kernel.h>
   2 #include <traceevent/event-parse.h>
   3
   4 #include <byteswap.h>
   5 #include <unistd.h>
   6 #include <sys/types.h>
   7 #include <sys/mman.h>
   8
   9 #include "evlist.h"
  10 #include "evsel.h"
  11 #include "session.h"
  12 #include "tool.h"
  13 #include "sort.h"
  14 #include "util.h"
  15 #include "cpumap.h"
  16 #include "perf_regs.h"
  17 #include "asm/bug.h"
  18
  19 static int perf_session__open(struct perf_session *session)
  20 {
  21         struct perf_data_file *file = session->file;
  22
  23         if (perf_session__read_header(session) < 0) {
  24                 pr_err("incompatible file format (rerun with -v to learn more)");
  25                 return -1;
  26         }
  27
  28         if (perf_data_file__is_pipe(file))
  29                 return 0;
  30
  31         if (!perf_evlist__valid_sample_type(session->evlist)) {
  32                 pr_err("non matching sample_type");
  33                 return -1;
  34         }
  35
  36         if (!perf_evlist__valid_sample_id_all(session->evlist)) {
  37                 pr_err("non matching sample_id_all");
  38                 return -1;
  39         }
  40
  41         if (!perf_evlist__valid_read_format(session->evlist)) {
  42                 pr_err("non matching read_format");
  43                 return -1;
  44         }
  45
  46         return 0;
  47 }
  48
  49 void perf_session__set_id_hdr_size(struct perf_session *session)
  50 {
  51         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
  52
  53         machines__set_id_hdr_size(&session->machines, id_hdr_size);
  54 }
  55
  56 int perf_session__create_kernel_maps(struct perf_session *session)
  57 {
  58         int ret = machine__create_kernel_maps(&session->machines.host);
  59
  60         if (ret >= 0)
  61                 ret = machines__create_guest_kernel_maps(&session->machines);
  62         return ret;
  63 }
  64
  65 static void perf_session__destroy_kernel_maps(struct perf_session *session)
  66 {
  67         machines__destroy_kernel_maps(&session->machines);
  68 }
  69
  70 struct perf_session *perf_session__new(struct perf_data_file *file,
  71                                        bool repipe, struct perf_tool *tool)
  72 {
  73         struct perf_session *session = zalloc(sizeof(*session));
  74
  75         if (!session)
  76                 goto out;
  77
  78         session->repipe = repipe;
  79         INIT_LIST_HEAD(&session->ordered_events.events);
  80         INIT_LIST_HEAD(&session->ordered_events.cache);
  81         INIT_LIST_HEAD(&session->ordered_events.to_free);
  82         session->ordered_events.max_alloc_size = (u64) -1;
  83         session->ordered_events.cur_alloc_size = 0;
  84         machines__init(&session->machines);
  85
  86         if (file) {
  87                 if (perf_data_file__open(file))
  88                         goto out_delete;
  89
  90                 session->file = file;
  91
  92                 if (perf_data_file__is_read(file)) {
  93                         if (perf_session__open(session) < 0)
  94                                 goto out_close;
  95
  96                         perf_session__set_id_hdr_size(session);
  97                 }
  98         }
  99
 100         if (!file || perf_data_file__is_write(file)) {
 101                 /*
 102                  * In O_RDONLY mode this will be performed when reading the
 103                  * kernel MMAP event, in perf_event__process_mmap().
 104                  */
 105                 if (perf_session__create_kernel_maps(session) < 0)
 106                         goto out_delete;
 107         }
 108
 109         if (tool && tool->ordering_requires_timestamps &&
 110             tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) {
 111                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 112                 tool->ordered_events = false;
 113         }
 114
 115         return session;
 116
 117  out_close:
 118         perf_data_file__close(file);
 119  out_delete:
 120         perf_session__delete(session);
 121  out:
 122         return NULL;
 123 }
 124
 125 static void perf_session__delete_dead_threads(struct perf_session *session)
 126 {
 127         machine__delete_dead_threads(&session->machines.host);
 128 }
 129
 130 static void perf_session__delete_threads(struct perf_session *session)
 131 {
 132         machine__delete_threads(&session->machines.host);
 133 }
 134
 135 static void perf_session_env__delete(struct perf_session_env *env)
 136 {
 137         zfree(&env->hostname);
 138         zfree(&env->os_release);
 139         zfree(&env->version);
 140         zfree(&env->arch);
 141         zfree(&env->cpu_desc);
 142         zfree(&env->cpuid);
 143
 144         zfree(&env->cmdline);
 145         zfree(&env->sibling_cores);
 146         zfree(&env->sibling_threads);
 147         zfree(&env->numa_nodes);
 148         zfree(&env->pmu_mappings);
 149 }
 150
 151 void perf_session__delete(struct perf_session *session)
 152 {
 153         perf_session__destroy_kernel_maps(session);
 154         perf_session__delete_dead_threads(session);
 155         perf_session__delete_threads(session);
 156         perf_session_env__delete(&session->header.env);
 157         machines__exit(&session->machines);
 158         if (session->file)
 159                 perf_data_file__close(session->file);
 160         free(session);
 161 }
 162
 163 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
 164                                                  __maybe_unused,
 165                                                  union perf_event *event
 166                                                  __maybe_unused,
 167                                                  struct perf_session *session
 168                                                 __maybe_unused)
 169 {
 170         dump_printf(": unhandled!\n");
 171         return 0;
 172 }
 173
 174 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
 175                                          union perf_event *event __maybe_unused,
 176                                          struct perf_evlist **pevlist
 177                                          __maybe_unused)
 178 {
 179         dump_printf(": unhandled!\n");
 180         return 0;
 181 }
 182
 183 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
 184                                      union perf_event *event __maybe_unused,
 185                                      struct perf_sample *sample __maybe_unused,
 186                                      struct perf_evsel *evsel __maybe_unused,
 187                                      struct machine *machine __maybe_unused)
 188 {
 189         dump_printf(": unhandled!\n");
 190         return 0;
 191 }
 192
 193 static int process_event_stub(struct perf_tool *tool __maybe_unused,
 194                               union perf_event *event __maybe_unused,
 195                               struct perf_sample *sample __maybe_unused,
 196                               struct machine *machine __maybe_unused)
 197 {
 198         dump_printf(": unhandled!\n");
 199         return 0;
 200 }
 201
 202 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
 203                                        union perf_event *event __maybe_unused,
 204                                        struct perf_session *perf_session
 205                                        __maybe_unused)
 206 {
 207         dump_printf(": unhandled!\n");
 208         return 0;
 209 }
 210
 211 static int process_finished_round(struct perf_tool *tool,
 212                                   union perf_event *event,
 213                                   struct perf_session *session);
 214
 215 void perf_tool__fill_defaults(struct perf_tool *tool)
 216 {
 217         if (tool->sample == NULL)
 218                 tool->sample = process_event_sample_stub;
 219         if (tool->mmap == NULL)
 220                 tool->mmap = process_event_stub;
 221         if (tool->mmap2 == NULL)
 222                 tool->mmap2 = process_event_stub;
 223         if (tool->comm == NULL)
 224                 tool->comm = process_event_stub;
 225         if (tool->fork == NULL)
 226                 tool->fork = process_event_stub;
 227         if (tool->exit == NULL)
 228                 tool->exit = process_event_stub;
 229         if (tool->lost == NULL)
 230                 tool->lost = perf_event__process_lost;
 231         if (tool->read == NULL)
 232                 tool->read = process_event_sample_stub;
 233         if (tool->throttle == NULL)
 234                 tool->throttle = process_event_stub;
 235         if (tool->unthrottle == NULL)
 236                 tool->unthrottle = process_event_stub;
 237         if (tool->attr == NULL)
 238                 tool->attr = process_event_synth_attr_stub;
 239         if (tool->tracing_data == NULL)
 240                 tool->tracing_data = process_event_synth_tracing_data_stub;
 241         if (tool->build_id == NULL)
 242                 tool->build_id = process_finished_round_stub;
 243         if (tool->finished_round == NULL) {
 244                 if (tool->ordered_events)
 245                         tool->finished_round = process_finished_round;
 246                 else
 247                         tool->finished_round = process_finished_round_stub;
 248         }
 249 }
 250
 251 static void swap_sample_id_all(union perf_event *event, void *data)
 252 {
 253         void *end = (void *) event + event->header.size;
 254         int size = end - data;
 255
 256         BUG_ON(size % sizeof(u64));
 257         mem_bswap_64(data, size);
 258 }
 259
 260 static void perf_event__all64_swap(union perf_event *event,
 261                                    bool sample_id_all __maybe_unused)
 262 {
 263         struct perf_event_header *hdr = &event->header;
 264         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 265 }
 266
 267 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 268 {
 269         event->comm.pid = bswap_32(event->comm.pid);
 270         event->comm.tid = bswap_32(event->comm.tid);
 271
 272         if (sample_id_all) {
 273                 void *data = &event->comm.comm;
 274
 275                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 276                 swap_sample_id_all(event, data);
 277         }
 278 }
 279
 280 static void perf_event__mmap_swap(union perf_event *event,
 281                                   bool sample_id_all)
 282 {
 283         event->mmap.pid   = bswap_32(event->mmap.pid);
 284         event->mmap.tid   = bswap_32(event->mmap.tid);
 285         event->mmap.start = bswap_64(event->mmap.start);
 286         event->mmap.len   = bswap_64(event->mmap.len);
 287         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 288
 289         if (sample_id_all) {
 290                 void *data = &event->mmap.filename;
 291
 292                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 293                 swap_sample_id_all(event, data);
 294         }
 295 }
 296
 297 static void perf_event__mmap2_swap(union perf_event *event,
 298                                   bool sample_id_all)
 299 {
 300         event->mmap2.pid   = bswap_32(event->mmap2.pid);
 301         event->mmap2.tid   = bswap_32(event->mmap2.tid);
 302         event->mmap2.start = bswap_64(event->mmap2.start);
 303         event->mmap2.len   = bswap_64(event->mmap2.len);
 304         event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
 305         event->mmap2.maj   = bswap_32(event->mmap2.maj);
 306         event->mmap2.min   = bswap_32(event->mmap2.min);
 307         event->mmap2.ino   = bswap_64(event->mmap2.ino);
 308
 309         if (sample_id_all) {
 310                 void *data = &event->mmap2.filename;
 311
 312                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 313                 swap_sample_id_all(event, data);
 314         }
 315 }
 316 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 317 {
 318         event->fork.pid  = bswap_32(event->fork.pid);
 319         event->fork.tid  = bswap_32(event->fork.tid);
 320         event->fork.ppid = bswap_32(event->fork.ppid);
 321         event->fork.ptid = bswap_32(event->fork.ptid);
 322         event->fork.time = bswap_64(event->fork.time);
 323
 324         if (sample_id_all)
 325                 swap_sample_id_all(event, &event->fork + 1);
 326 }
 327
 328 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 329 {
 330         event->read.pid          = bswap_32(event->read.pid);
 331         event->read.tid          = bswap_32(event->read.tid);
 332         event->read.value        = bswap_64(event->read.value);
 333         event->read.time_enabled = bswap_64(event->read.time_enabled);
 334         event->read.time_running = bswap_64(event->read.time_running);
 335         event->read.id           = bswap_64(event->read.id);
 336
 337         if (sample_id_all)
 338                 swap_sample_id_all(event, &event->read + 1);
 339 }
 340
 341 static void perf_event__throttle_swap(union perf_event *event,
 342                                       bool sample_id_all)
 343 {
 344         event->throttle.time      = bswap_64(event->throttle.time);
 345         event->throttle.id        = bswap_64(event->throttle.id);
 346         event->throttle.stream_id = bswap_64(event->throttle.stream_id);
 347
 348         if (sample_id_all)
 349                 swap_sample_id_all(event, &event->throttle + 1);
 350 }
 351
 352 static u8 revbyte(u8 b)
 353 {
 354         int rev = (b >> 4) | ((b & 0xf) << 4);
 355         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 356         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 357         return (u8) rev;
 358 }
 359
 360 /*
 361  * XXX this is hack in attempt to carry flags bitfield
 362  * throught endian village. ABI says:
 363  *
 364  * Bit-fields are allocated from right to left (least to most significant)
 365  * on little-endian implementations and from left to right (most to least
 366  * significant) on big-endian implementations.
 367  *
 368  * The above seems to be byte specific, so we need to reverse each
 369  * byte of the bitfield. 'Internet' also says this might be implementation
 370  * specific and we probably need proper fix and carry perf_event_attr
 371  * bitfield flags in separate data file FEAT_ section. Thought this seems
 372  * to work for now.
 373  */
 374 static void swap_bitfield(u8 *p, unsigned len)
 375 {
 376         unsigned i;
 377
 378         for (i = 0; i < len; i++) {
 379                 *p = revbyte(*p);
 380                 p++;
 381         }
 382 }
 383
 384 /* exported for swapping attributes in file header */
 385 void perf_event__attr_swap(struct perf_event_attr *attr)
 386 {
 387         attr->type              = bswap_32(attr->type);
 388         attr->size              = bswap_32(attr->size);
 389         attr->config            = bswap_64(attr->config);
 390         attr->sample_period     = bswap_64(attr->sample_period);
 391         attr->sample_type       = bswap_64(attr->sample_type);
 392         attr->read_format       = bswap_64(attr->read_format);
 393         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 394         attr->bp_type           = bswap_32(attr->bp_type);
 395         attr->bp_addr           = bswap_64(attr->bp_addr);
 396         attr->bp_len            = bswap_64(attr->bp_len);
 397         attr->branch_sample_type = bswap_64(attr->branch_sample_type);
 398         attr->sample_regs_user   = bswap_64(attr->sample_regs_user);
 399         attr->sample_stack_user  = bswap_32(attr->sample_stack_user);
 400
 401         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 402 }
 403
 404 static void perf_event__hdr_attr_swap(union perf_event *event,
 405                                       bool sample_id_all __maybe_unused)
 406 {
 407         size_t size;
 408
 409         perf_event__attr_swap(&event->attr.attr);
 410
 411         size = event->header.size;
 412         size -= (void *)&event->attr.id - (void *)event;
 413         mem_bswap_64(event->attr.id, size);
 414 }
 415
 416 static void perf_event__event_type_swap(union perf_event *event,
 417                                         bool sample_id_all __maybe_unused)
 418 {
 419         event->event_type.event_type.event_id =
 420                 bswap_64(event->event_type.event_type.event_id);
 421 }
 422
 423 static void perf_event__tracing_data_swap(union perf_event *event,
 424                                           bool sample_id_all __maybe_unused)
 425 {
 426         event->tracing_data.size = bswap_32(event->tracing_data.size);
 427 }
 428
 429 typedef void (*perf_event__swap_op)(union perf_event *event,
 430                                     bool sample_id_all);
 431
 432 static perf_event__swap_op perf_event__swap_ops[] = {
 433         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 434         [PERF_RECORD_MMAP2]               = perf_event__mmap2_swap,
 435         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 436         [PERF_RECORD_FORK]                = perf_event__task_swap,
 437         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 438         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 439         [PERF_RECORD_READ]                = perf_event__read_swap,
 440         [PERF_RECORD_THROTTLE]            = perf_event__throttle_swap,
 441         [PERF_RECORD_UNTHROTTLE]          = perf_event__throttle_swap,
 442         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 443         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 444         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 445         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 446         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 447         [PERF_RECORD_HEADER_MAX]          = NULL,
 448 };
 449
 450 struct ordered_event {
 451         u64                     timestamp;
 452         u64                     file_offset;
 453         union perf_event        *event;
 454         struct list_head        list;
 455 };
 456
 457 enum oe_flush {
 458         OE_FLUSH__FINAL,
 459         OE_FLUSH__ROUND,
 460         OE_FLUSH__HALF,
 461 };
 462
 463 static void perf_session_free_sample_buffers(struct perf_session *session)
 464 {
 465         struct ordered_events *oe = &session->ordered_events;
 466
 467         while (!list_empty(&oe->to_free)) {
 468                 struct ordered_event *event;
 469
 470                 event = list_entry(oe->to_free.next, struct ordered_event, list);
 471                 list_del(&event->list);
 472                 free(event);
 473         }
 474 }
 475
 476 /* The queue is ordered by time */
 477 static void queue_event(struct ordered_events *oe, struct ordered_event *new)
 478 {
 479         struct ordered_event *last = oe->last;
 480         u64 timestamp = new->timestamp;
 481         struct list_head *p;
 482
 483         ++oe->nr_events;
 484         oe->last = new;
 485
 486         if (!last) {
 487                 list_add(&new->list, &oe->events);
 488                 oe->max_timestamp = timestamp;
 489                 return;
 490         }
 491
 492         /*
 493          * last event might point to some random place in the list as it's
 494          * the last queued event. We expect that the new event is close to
 495          * this.
 496          */
 497         if (last->timestamp <= timestamp) {
 498                 while (last->timestamp <= timestamp) {
 499                         p = last->list.next;
 500                         if (p == &oe->events) {
 501                                 list_add_tail(&new->list, &oe->events);
 502                                 oe->max_timestamp = timestamp;
 503                                 return;
 504                         }
 505                         last = list_entry(p, struct ordered_event, list);
 506                 }
 507                 list_add_tail(&new->list, &last->list);
 508         } else {
 509                 while (last->timestamp > timestamp) {
 510                         p = last->list.prev;
 511                         if (p == &oe->events) {
 512                                 list_add(&new->list, &oe->events);
 513                                 return;
 514                         }
 515                         last = list_entry(p, struct ordered_event, list);
 516                 }
 517                 list_add(&new->list, &last->list);
 518         }
 519 }
 520
 521 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct ordered_event))
 522 static struct ordered_event *alloc_event(struct ordered_events *oe)
 523 {
 524         struct list_head *cache = &oe->cache;
 525         struct ordered_event *new = NULL;
 526
 527         if (!list_empty(cache)) {
 528                 new = list_entry(cache->next, struct ordered_event, list);
 529                 list_del(&new->list);
 530         } else if (oe->buffer) {
 531                 new = oe->buffer + oe->buffer_idx;
 532                 if (++oe->buffer_idx == MAX_SAMPLE_BUFFER)
 533                         oe->buffer = NULL;
 534         } else if (oe->cur_alloc_size < oe->max_alloc_size) {
 535                 size_t size = MAX_SAMPLE_BUFFER * sizeof(*new);
 536
 537                 oe->buffer = malloc(size);
 538                 if (!oe->buffer)
 539                         return NULL;
 540
 541                 oe->cur_alloc_size += size;
 542                 list_add(&oe->buffer->list, &oe->to_free);
 543
 544                 /* First entry is abused to maintain the to_free list. */
 545                 oe->buffer_idx = 2;
 546                 new = oe->buffer + 1;
 547         }
 548
 549         return new;
 550 }
 551
 552 static struct ordered_event *
 553 ordered_events__new(struct ordered_events *oe, u64 timestamp)
 554 {
 555         struct ordered_event *new;
 556
 557         new = alloc_event(oe);
 558         if (new) {
 559                 new->timestamp = timestamp;
 560                 queue_event(oe, new);
 561         }
 562
 563         return new;
 564 }
 565
 566 static void
 567 ordered_events__delete(struct ordered_events *oe, struct ordered_event *event)
 568 {
 569         list_del(&event->list);
 570         list_add(&event->list, &oe->cache);
 571         oe->nr_events--;
 572 }
 573
 574 static int __ordered_events__flush(struct perf_session *s,
 575                                    struct perf_tool *tool)
 576 {
 577         struct ordered_events *oe = &s->ordered_events;
 578         struct list_head *head = &oe->events;
 579         struct ordered_event *tmp, *iter;
 580         struct perf_sample sample;
 581         u64 limit = oe->next_flush;
 582         u64 last_ts = oe->last ? oe->last->timestamp : 0ULL;
 583         bool show_progress = limit == ULLONG_MAX;
 584         struct ui_progress prog;
 585         int ret;
 586
 587         if (!tool->ordered_events || !limit)
 588                 return 0;
 589
 590         if (show_progress)
 591                 ui_progress__init(&prog, oe->nr_events, "Processing time ordered events...");
 592
 593         list_for_each_entry_safe(iter, tmp, head, list) {
 594                 if (session_done())
 595                         return 0;
 596
 597                 if (iter->timestamp > limit)
 598                         break;
 599
 600                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
 601                 if (ret)
 602                         pr_err("Can't parse sample, err = %d\n", ret);
 603                 else {
 604                         ret = perf_session__deliver_event(s, iter->event, &sample, tool,
 605                                                           iter->file_offset);
 606                         if (ret)
 607                                 return ret;
 608                 }
 609
 610                 ordered_events__delete(oe, iter);
 611                 oe->last_flush = iter->timestamp;
 612
 613                 if (show_progress)
 614                         ui_progress__update(&prog, 1);
 615         }
 616
 617         if (list_empty(head))
 618                 oe->last = NULL;
 619         else if (last_ts <= limit)
 620                 oe->last = list_entry(head->prev, struct ordered_event, list);
 621
 622         return 0;
 623 }
 624
 625 static int ordered_events__flush(struct perf_session *s, struct perf_tool *tool,
 626                                  enum oe_flush how)
 627 {
 628         struct ordered_events *oe = &s->ordered_events;
 629         int err;
 630
 631         switch (how) {
 632         case OE_FLUSH__FINAL:
 633                 oe->next_flush = ULLONG_MAX;
 634                 break;
 635
 636         case OE_FLUSH__HALF:
 637         {
 638                 struct ordered_event *first, *last;
 639                 struct list_head *head = &oe->events;
 640
 641                 first = list_entry(head->next, struct ordered_event, list);
 642                 last = oe->last;
 643
 644                 /* Warn if we are called before any event got allocated. */
 645                 if (WARN_ONCE(!last || list_empty(head), "empty queue"))
 646                         return 0;
 647
 648                 oe->next_flush  = first->timestamp;
 649                 oe->next_flush += (last->timestamp - first->timestamp) / 2;
 650                 break;
 651         }
 652
 653         case OE_FLUSH__ROUND:
 654         default:
 655                 break;
 656         };
 657
 658         err = __ordered_events__flush(s, tool);
 659
 660         if (!err) {
 661                 if (how == OE_FLUSH__ROUND)
 662                         oe->next_flush = oe->max_timestamp;
 663         }
 664
 665         return err;
 666 }
 667
 668 /*
 669  * When perf record finishes a pass on every buffers, it records this pseudo
 670  * event.
 671  * We record the max timestamp t found in the pass n.
 672  * Assuming these timestamps are monotonic across cpus, we know that if
 673  * a buffer still has events with timestamps below t, they will be all
 674  * available and then read in the pass n + 1.
 675  * Hence when we start to read the pass n + 2, we can safely flush every
 676  * events with timestamps below t.
 677  *
 678  *    ============ PASS n =================
 679  *       CPU 0         |   CPU 1
 680  *                     |
 681  *    cnt1 timestamps  |   cnt2 timestamps
 682  *          1          |         2
 683  *          2          |         3
 684  *          -          |         4  <--- max recorded
 685  *
 686  *    ============ PASS n + 1 ==============
 687  *       CPU 0         |   CPU 1
 688  *                     |
 689  *    cnt1 timestamps  |   cnt2 timestamps
 690  *          3          |         5
 691  *          4          |         6
 692  *          5          |         7 <---- max recorded
 693  *
 694  *      Flush every events below timestamp 4
 695  *
 696  *    ============ PASS n + 2 ==============
 697  *       CPU 0         |   CPU 1
 698  *                     |
 699  *    cnt1 timestamps  |   cnt2 timestamps
 700  *          6          |         8
 701  *          7          |         9
 702  *          -          |         10
 703  *
 704  *      Flush every events below timestamp 7
 705  *      etc...
 706  */
 707 static int process_finished_round(struct perf_tool *tool,
 708                                   union perf_event *event __maybe_unused,
 709                                   struct perf_session *session)
 710 {
 711         return ordered_events__flush(session, tool, OE_FLUSH__ROUND);
 712 }
 713
 714 int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 715                              struct perf_tool *tool, struct perf_sample *sample,
 716                              u64 file_offset)
 717 {
 718         struct ordered_events *oe = &s->ordered_events;
 719         u64 timestamp = sample->time;
 720         struct ordered_event *new;
 721
 722         if (!timestamp || timestamp == ~0ULL)
 723                 return -ETIME;
 724
 725         if (timestamp < s->ordered_events.last_flush) {
 726                 printf("Warning: Timestamp below last timeslice flush\n");
 727                 return -EINVAL;
 728         }
 729
 730         new = ordered_events__new(oe, timestamp);
 731         if (!new) {
 732                 ordered_events__flush(s, tool, OE_FLUSH__HALF);
 733                 new = ordered_events__new(oe, timestamp);
 734         }
 735
 736         if (!new)
 737                 return -ENOMEM;
 738
 739         new->file_offset = file_offset;
 740         new->event = event;
 741         return 0;
 742 }
 743
 744 static void callchain__printf(struct perf_sample *sample)
 745 {
 746         unsigned int i;
 747
 748         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 749
 750         for (i = 0; i < sample->callchain->nr; i++)
 751                 printf("..... %2d: %016" PRIx64 "\n",
 752                        i, sample->callchain->ips[i]);
 753 }
 754
 755 static void branch_stack__printf(struct perf_sample *sample)
 756 {
 757         uint64_t i;
 758
 759         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 760
 761         for (i = 0; i < sample->branch_stack->nr; i++)
 762                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 763                         i, sample->branch_stack->entries[i].from,
 764                         sample->branch_stack->entries[i].to);
 765 }
 766
 767 static void regs_dump__printf(u64 mask, u64 *regs)
 768 {
 769         unsigned rid, i = 0;
 770
 771         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 772                 u64 val = regs[i++];
 773
 774                 printf(".... %-5s 0x%" PRIx64 "\n",
 775                        perf_reg_name(rid), val);
 776         }
 777 }
 778
 779 static void regs_user__printf(struct perf_sample *sample)
 780 {
 781         struct regs_dump *user_regs = &sample->user_regs;
 782
 783         if (user_regs->regs) {
 784                 u64 mask = user_regs->mask;
 785                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
 786                 regs_dump__printf(mask, user_regs->regs);
 787         }
 788 }
 789
 790 static void stack_user__printf(struct stack_dump *dump)
 791 {
 792         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 793                dump->size, dump->offset);
 794 }
 795
 796 static void perf_session__print_tstamp(struct perf_session *session,
 797                                        union perf_event *event,
 798                                        struct perf_sample *sample)
 799 {
 800         u64 sample_type = __perf_evlist__combined_sample_type(session->evlist);
 801
 802         if (event->header.type != PERF_RECORD_SAMPLE &&
 803             !perf_evlist__sample_id_all(session->evlist)) {
 804                 fputs("-1 -1 ", stdout);
 805                 return;
 806         }
 807
 808         if ((sample_type & PERF_SAMPLE_CPU))
 809                 printf("%u ", sample->cpu);
 810
 811         if (sample_type & PERF_SAMPLE_TIME)
 812                 printf("%" PRIu64 " ", sample->time);
 813 }
 814
 815 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
 816 {
 817         printf("... sample_read:\n");
 818
 819         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 820                 printf("...... time enabled %016" PRIx64 "\n",
 821                        sample->read.time_enabled);
 822
 823         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 824                 printf("...... time running %016" PRIx64 "\n",
 825                        sample->read.time_running);
 826
 827         if (read_format & PERF_FORMAT_GROUP) {
 828                 u64 i;
 829
 830                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
 831
 832                 for (i = 0; i < sample->read.group.nr; i++) {
 833                         struct sample_read_value *value;
 834
 835                         value = &sample->read.group.values[i];
 836                         printf("..... id %016" PRIx64
 837                                ", value %016" PRIx64 "\n",
 838                                value->id, value->value);
 839                 }
 840         } else
 841                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
 842                         sample->read.one.id, sample->read.one.value);
 843 }
 844
 845 static void dump_event(struct perf_session *session, union perf_event *event,
 846                        u64 file_offset, struct perf_sample *sample)
 847 {
 848         if (!dump_trace)
 849                 return;
 850
 851         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 852                file_offset, event->header.size, event->header.type);
 853
 854         trace_event(event);
 855
 856         if (sample)
 857                 perf_session__print_tstamp(session, event, sample);
 858
 859         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 860                event->header.size, perf_event__name(event->header.type));
 861 }
 862
 863 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 864                         struct perf_sample *sample)
 865 {
 866         u64 sample_type;
 867
 868         if (!dump_trace)
 869                 return;
 870
 871         printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 872                event->header.misc, sample->pid, sample->tid, sample->ip,
 873                sample->period, sample->addr);
 874
 875         sample_type = evsel->attr.sample_type;
 876
 877         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 878                 callchain__printf(sample);
 879
 880         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
 881                 branch_stack__printf(sample);
 882
 883         if (sample_type & PERF_SAMPLE_REGS_USER)
 884                 regs_user__printf(sample);
 885
 886         if (sample_type & PERF_SAMPLE_STACK_USER)
 887                 stack_user__printf(&sample->user_stack);
 888
 889         if (sample_type & PERF_SAMPLE_WEIGHT)
 890                 printf("... weight: %" PRIu64 "\n", sample->weight);
 891
 892         if (sample_type & PERF_SAMPLE_DATA_SRC)
 893                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
 894
 895         if (sample_type & PERF_SAMPLE_TRANSACTION)
 896                 printf("... transaction: %" PRIx64 "\n", sample->transaction);
 897
 898         if (sample_type & PERF_SAMPLE_READ)
 899                 sample_read__printf(sample, evsel->attr.read_format);
 900 }
 901
 902 static struct machine *
 903         perf_session__find_machine_for_cpumode(struct perf_session *session,
 904                                                union perf_event *event,
 905                                                struct perf_sample *sample)
 906 {
 907         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 908         struct machine *machine;
 909
 910         if (perf_guest &&
 911             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
 912              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
 913                 u32 pid;
 914
 915                 if (event->header.type == PERF_RECORD_MMAP
 916                     || event->header.type == PERF_RECORD_MMAP2)
 917                         pid = event->mmap.pid;
 918                 else
 919                         pid = sample->pid;
 920
 921                 machine = perf_session__find_machine(session, pid);
 922                 if (!machine)
 923                         machine = perf_session__findnew_machine(session,
 924                                                 DEFAULT_GUEST_KERNEL_ID);
 925                 return machine;
 926         }
 927
 928         return &session->machines.host;
 929 }
 930
 931 static int deliver_sample_value(struct perf_session *session,
 932                                 struct perf_tool *tool,
 933                                 union perf_event *event,
 934                                 struct perf_sample *sample,
 935                                 struct sample_read_value *v,
 936                                 struct machine *machine)
 937 {
 938         struct perf_sample_id *sid;
 939
 940         sid = perf_evlist__id2sid(session->evlist, v->id);
 941         if (sid) {
 942                 sample->id     = v->id;
 943                 sample->period = v->value - sid->period;
 944                 sid->period    = v->value;
 945         }
 946
 947         if (!sid || sid->evsel == NULL) {
 948                 ++session->stats.nr_unknown_id;
 949                 return 0;
 950         }
 951
 952         return tool->sample(tool, event, sample, sid->evsel, machine);
 953 }
 954
 955 static int deliver_sample_group(struct perf_session *session,
 956                                 struct perf_tool *tool,
 957                                 union  perf_event *event,
 958                                 struct perf_sample *sample,
 959                                 struct machine *machine)
 960 {
 961         int ret = -EINVAL;
 962         u64 i;
 963
 964         for (i = 0; i < sample->read.group.nr; i++) {
 965                 ret = deliver_sample_value(session, tool, event, sample,
 966                                            &sample->read.group.values[i],
 967                                            machine);
 968                 if (ret)
 969                         break;
 970         }
 971
 972         return ret;
 973 }
 974
 975 static int
 976 perf_session__deliver_sample(struct perf_session *session,
 977                              struct perf_tool *tool,
 978                              union  perf_event *event,
 979                              struct perf_sample *sample,
 980                              struct perf_evsel *evsel,
 981                              struct machine *machine)
 982 {
 983         /* We know evsel != NULL. */
 984         u64 sample_type = evsel->attr.sample_type;
 985         u64 read_format = evsel->attr.read_format;
 986
 987         /* Standard sample delievery. */
 988         if (!(sample_type & PERF_SAMPLE_READ))
 989                 return tool->sample(tool, event, sample, evsel, machine);
 990
 991         /* For PERF_SAMPLE_READ we have either single or group mode. */
 992         if (read_format & PERF_FORMAT_GROUP)
 993                 return deliver_sample_group(session, tool, event, sample,
 994                                             machine);
 995         else
 996                 return deliver_sample_value(session, tool, event, sample,
 997                                             &sample->read.one, machine);
 998 }
 999
1000 int perf_session__deliver_event(struct perf_session *session,
1001                                 union perf_event *event,
1002                                 struct perf_sample *sample,
1003                                 struct perf_tool *tool, u64 file_offset)
1004 {
1005         struct perf_evsel *evsel;
1006         struct machine *machine;
1007
1008         dump_event(session, event, file_offset, sample);
1009
1010         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
1011         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
1012                 /*
1013                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
1014                  * because the tools right now may apply filters, discarding
1015                  * some of the samples. For consistency, in the future we
1016                  * should have something like nr_filtered_samples and remove
1017                  * the sample->period from total_sample_period, etc, KISS for
1018                  * now tho.
1019                  *
1020                  * Also testing against NULL allows us to handle files without
1021                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
1022                  * future probably it'll be a good idea to restrict event
1023                  * processing via perf_session to files with both set.
1024                  */
1025                 hists__inc_nr_events(&evsel->hists, event->header.type);
1026         }
1027
1028         machine = perf_session__find_machine_for_cpumode(session, event,
1029                                                          sample);
1030
1031         switch (event->header.type) {
1032         case PERF_RECORD_SAMPLE:
1033                 dump_sample(evsel, event, sample);
1034                 if (evsel == NULL) {
1035                         ++session->stats.nr_unknown_id;
1036                         return 0;
1037                 }
1038                 if (machine == NULL) {
1039                         ++session->stats.nr_unprocessable_samples;
1040                         return 0;
1041                 }
1042                 return perf_session__deliver_sample(session, tool, event,
1043                                                     sample, evsel, machine);
1044         case PERF_RECORD_MMAP:
1045                 return tool->mmap(tool, event, sample, machine);
1046         case PERF_RECORD_MMAP2:
1047                 return tool->mmap2(tool, event, sample, machine);
1048         case PERF_RECORD_COMM:
1049                 return tool->comm(tool, event, sample, machine);
1050         case PERF_RECORD_FORK:
1051                 return tool->fork(tool, event, sample, machine);
1052         case PERF_RECORD_EXIT:
1053                 return tool->exit(tool, event, sample, machine);
1054         case PERF_RECORD_LOST:
1055                 if (tool->lost == perf_event__process_lost)
1056                         session->stats.total_lost += event->lost.lost;
1057                 return tool->lost(tool, event, sample, machine);
1058         case PERF_RECORD_READ:
1059                 return tool->read(tool, event, sample, evsel, machine);
1060         case PERF_RECORD_THROTTLE:
1061                 return tool->throttle(tool, event, sample, machine);
1062         case PERF_RECORD_UNTHROTTLE:
1063                 return tool->unthrottle(tool, event, sample, machine);
1064         default:
1065                 ++session->stats.nr_unknown_events;
1066                 return -1;
1067         }
1068 }
1069
1070 static s64 perf_session__process_user_event(struct perf_session *session,
1071                                             union perf_event *event,
1072                                             struct perf_tool *tool,
1073                                             u64 file_offset)
1074 {
1075         int fd = perf_data_file__fd(session->file);
1076         int err;
1077
1078         dump_event(session, event, file_offset, NULL);
1079
1080         /* These events are processed right away */
1081         switch (event->header.type) {
1082         case PERF_RECORD_HEADER_ATTR:
1083                 err = tool->attr(tool, event, &session->evlist);
1084                 if (err == 0)
1085                         perf_session__set_id_hdr_size(session);
1086                 return err;
1087         case PERF_RECORD_HEADER_EVENT_TYPE:
1088                 /*
1089                  * Depreceated, but we need to handle it for sake
1090                  * of old data files create in pipe mode.
1091                  */
1092                 return 0;
1093         case PERF_RECORD_HEADER_TRACING_DATA:
1094                 /* setup for reading amidst mmap */
1095                 lseek(fd, file_offset, SEEK_SET);
1096                 return tool->tracing_data(tool, event, session);
1097         case PERF_RECORD_HEADER_BUILD_ID:
1098                 return tool->build_id(tool, event, session);
1099         case PERF_RECORD_FINISHED_ROUND:
1100                 return tool->finished_round(tool, event, session);
1101         default:
1102                 return -EINVAL;
1103         }
1104 }
1105
1106 static void event_swap(union perf_event *event, bool sample_id_all)
1107 {
1108         perf_event__swap_op swap;
1109
1110         swap = perf_event__swap_ops[event->header.type];
1111         if (swap)
1112                 swap(event, sample_id_all);
1113 }
1114
1115 static s64 perf_session__process_event(struct perf_session *session,
1116                                        union perf_event *event,
1117                                        struct perf_tool *tool,
1118                                        u64 file_offset)
1119 {
1120         struct perf_sample sample;
1121         int ret;
1122
1123         if (session->header.needs_swap)
1124                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1125
1126         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1127                 return -EINVAL;
1128
1129         events_stats__inc(&session->stats, event->header.type);
1130
1131         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1132                 return perf_session__process_user_event(session, event, tool, file_offset);
1133
1134         /*
1135          * For all kernel events we get the sample data
1136          */
1137         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1138         if (ret)
1139                 return ret;
1140
1141         if (tool->ordered_events) {
1142                 ret = perf_session_queue_event(session, event, tool, &sample,
1143                                                file_offset);
1144                 if (ret != -ETIME)
1145                         return ret;
1146         }
1147
1148         return perf_session__deliver_event(session, event, &sample, tool,
1149                                            file_offset);
1150 }
1151
1152 void perf_event_header__bswap(struct perf_event_header *hdr)
1153 {
1154         hdr->type = bswap_32(hdr->type);
1155         hdr->misc = bswap_16(hdr->misc);
1156         hdr->size = bswap_16(hdr->size);
1157 }
1158
1159 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1160 {
1161         return machine__findnew_thread(&session->machines.host, -1, pid);
1162 }
1163
1164 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1165 {
1166         struct thread *thread;
1167
1168         thread = machine__findnew_thread(&session->machines.host, 0, 0);
1169         if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1170                 pr_err("problem inserting idle task.\n");
1171                 thread = NULL;
1172         }
1173
1174         return thread;
1175 }
1176
1177 static void perf_session__warn_about_errors(const struct perf_session *session,
1178                                             const struct perf_tool *tool)
1179 {
1180         if (tool->lost == perf_event__process_lost &&
1181             session->stats.nr_events[PERF_RECORD_LOST] != 0) {
1182                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1183                             "Check IO/CPU overload!\n\n",
1184                             session->stats.nr_events[0],
1185                             session->stats.nr_events[PERF_RECORD_LOST]);
1186         }
1187
1188         if (session->stats.nr_unknown_events != 0) {
1189                 ui__warning("Found %u unknown events!\n\n"
1190                             "Is this an older tool processing a perf.data "
1191                             "file generated by a more recent tool?\n\n"
1192                             "If that is not the case, consider "
1193                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1194                             session->stats.nr_unknown_events);
1195         }
1196
1197         if (session->stats.nr_unknown_id != 0) {
1198                 ui__warning("%u samples with id not present in the header\n",
1199                             session->stats.nr_unknown_id);
1200         }
1201
1202         if (session->stats.nr_invalid_chains != 0) {
1203                 ui__warning("Found invalid callchains!\n\n"
1204                             "%u out of %u events were discarded for this reason.\n\n"
1205                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1206                             session->stats.nr_invalid_chains,
1207                             session->stats.nr_events[PERF_RECORD_SAMPLE]);
1208         }
1209
1210         if (session->stats.nr_unprocessable_samples != 0) {
1211                 ui__warning("%u unprocessable samples recorded.\n"
1212                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1213                             session->stats.nr_unprocessable_samples);
1214         }
1215 }
1216
1217 volatile int session_done;
1218
1219 static int __perf_session__process_pipe_events(struct perf_session *session,
1220                                                struct perf_tool *tool)
1221 {
1222         int fd = perf_data_file__fd(session->file);
1223         union perf_event *event;
1224         uint32_t size, cur_size = 0;
1225         void *buf = NULL;
1226         s64 skip = 0;
1227         u64 head;
1228         ssize_t err;
1229         void *p;
1230
1231         perf_tool__fill_defaults(tool);
1232
1233         head = 0;
1234         cur_size = sizeof(union perf_event);
1235
1236         buf = malloc(cur_size);
1237         if (!buf)
1238                 return -errno;
1239 more:
1240         event = buf;
1241         err = readn(fd, event, sizeof(struct perf_event_header));
1242         if (err <= 0) {
1243                 if (err == 0)
1244                         goto done;
1245
1246                 pr_err("failed to read event header\n");
1247                 goto out_err;
1248         }
1249
1250         if (session->header.needs_swap)
1251                 perf_event_header__bswap(&event->header);
1252
1253         size = event->header.size;
1254         if (size < sizeof(struct perf_event_header)) {
1255                 pr_err("bad event header size\n");
1256                 goto out_err;
1257         }
1258
1259         if (size > cur_size) {
1260                 void *new = realloc(buf, size);
1261                 if (!new) {
1262                         pr_err("failed to allocate memory to read event\n");
1263                         goto out_err;
1264                 }
1265                 buf = new;
1266                 cur_size = size;
1267                 event = buf;
1268         }
1269         p = event;
1270         p += sizeof(struct perf_event_header);
1271
1272         if (size - sizeof(struct perf_event_header)) {
1273                 err = readn(fd, p, size - sizeof(struct perf_event_header));
1274                 if (err <= 0) {
1275                         if (err == 0) {
1276                                 pr_err("unexpected end of event stream\n");
1277                                 goto done;
1278                         }
1279
1280                         pr_err("failed to read event data\n");
1281                         goto out_err;
1282                 }
1283         }
1284
1285         if ((skip = perf_session__process_event(session, event, tool, head)) < 0) {
1286                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1287                        head, event->header.size, event->header.type);
1288                 err = -EINVAL;
1289                 goto out_err;
1290         }
1291
1292         head += size;
1293
1294         if (skip > 0)
1295                 head += skip;
1296
1297         if (!session_done())
1298                 goto more;
1299 done:
1300         /* do the final flush for ordered samples */
1301         err = ordered_events__flush(session, tool, OE_FLUSH__FINAL);
1302 out_err:
1303         free(buf);
1304         perf_session__warn_about_errors(session, tool);
1305         perf_session_free_sample_buffers(session);
1306         return err;
1307 }
1308
1309 static union perf_event *
1310 fetch_mmaped_event(struct perf_session *session,
1311                    u64 head, size_t mmap_size, char *buf)
1312 {
1313         union perf_event *event;
1314
1315         /*
1316          * Ensure we have enough space remaining to read
1317          * the size of the event in the headers.
1318          */
1319         if (head + sizeof(event->header) > mmap_size)
1320                 return NULL;
1321
1322         event = (union perf_event *)(buf + head);
1323
1324         if (session->header.needs_swap)
1325                 perf_event_header__bswap(&event->header);
1326
1327         if (head + event->header.size > mmap_size) {
1328                 /* We're not fetching the event so swap back again */
1329                 if (session->header.needs_swap)
1330                         perf_event_header__bswap(&event->header);
1331                 return NULL;
1332         }
1333
1334         return event;
1335 }
1336
1337 /*
1338  * On 64bit we can mmap the data file in one go. No need for tiny mmap
1339  * slices. On 32bit we use 32MB.
1340  */
1341 #if BITS_PER_LONG == 64
1342 #define MMAP_SIZE ULLONG_MAX
1343 #define NUM_MMAPS 1
1344 #else
1345 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1346 #define NUM_MMAPS 128
1347 #endif
1348
1349 int __perf_session__process_events(struct perf_session *session,
1350                                    u64 data_offset, u64 data_size,
1351                                    u64 file_size, struct perf_tool *tool)
1352 {
1353         int fd = perf_data_file__fd(session->file);
1354         u64 head, page_offset, file_offset, file_pos, size;
1355         int err, mmap_prot, mmap_flags, map_idx = 0;
1356         size_t  mmap_size;
1357         char *buf, *mmaps[NUM_MMAPS];
1358         union perf_event *event;
1359         struct ui_progress prog;
1360         s64 skip;
1361
1362         perf_tool__fill_defaults(tool);
1363
1364         page_offset = page_size * (data_offset / page_size);
1365         file_offset = page_offset;
1366         head = data_offset - page_offset;
1367
1368         if (data_size && (data_offset + data_size < file_size))
1369                 file_size = data_offset + data_size;
1370
1371         ui_progress__init(&prog, file_size, "Processing events...");
1372
1373         mmap_size = MMAP_SIZE;
1374         if (mmap_size > file_size) {
1375                 mmap_size = file_size;
1376                 session->one_mmap = true;
1377         }
1378
1379         memset(mmaps, 0, sizeof(mmaps));
1380
1381         mmap_prot  = PROT_READ;
1382         mmap_flags = MAP_SHARED;
1383
1384         if (session->header.needs_swap) {
1385                 mmap_prot  |= PROT_WRITE;
1386                 mmap_flags = MAP_PRIVATE;
1387         }
1388 remap:
1389         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1390                    file_offset);
1391         if (buf == MAP_FAILED) {
1392                 pr_err("failed to mmap file\n");
1393                 err = -errno;
1394                 goto out_err;
1395         }
1396         mmaps[map_idx] = buf;
1397         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1398         file_pos = file_offset + head;
1399         if (session->one_mmap) {
1400                 session->one_mmap_addr = buf;
1401                 session->one_mmap_offset = file_offset;
1402         }
1403
1404 more:
1405         event = fetch_mmaped_event(session, head, mmap_size, buf);
1406         if (!event) {
1407                 if (mmaps[map_idx]) {
1408                         munmap(mmaps[map_idx], mmap_size);
1409                         mmaps[map_idx] = NULL;
1410                 }
1411
1412                 page_offset = page_size * (head / page_size);
1413                 file_offset += page_offset;
1414                 head -= page_offset;
1415                 goto remap;
1416         }
1417
1418         size = event->header.size;
1419
1420         if (size < sizeof(struct perf_event_header) ||
1421             (skip = perf_session__process_event(session, event, tool, file_pos))
1422                                                                         < 0) {
1423                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1424                        file_offset + head, event->header.size,
1425                        event->header.type);
1426                 err = -EINVAL;
1427                 goto out_err;
1428         }
1429
1430         if (skip)
1431                 size += skip;
1432
1433         head += size;
1434         file_pos += size;
1435
1436         ui_progress__update(&prog, size);
1437
1438         if (session_done())
1439                 goto out;
1440
1441         if (file_pos < file_size)
1442                 goto more;
1443
1444 out:
1445         /* do the final flush for ordered samples */
1446         err = ordered_events__flush(session, tool, OE_FLUSH__FINAL);
1447 out_err:
1448         ui_progress__finish();
1449         perf_session__warn_about_errors(session, tool);
1450         perf_session_free_sample_buffers(session);
1451         session->one_mmap = false;
1452         return err;
1453 }
1454
1455 int perf_session__process_events(struct perf_session *session,
1456                                  struct perf_tool *tool)
1457 {
1458         u64 size = perf_data_file__size(session->file);
1459         int err;
1460
1461         if (perf_session__register_idle_thread(session) == NULL)
1462                 return -ENOMEM;
1463
1464         if (!perf_data_file__is_pipe(session->file))
1465                 err = __perf_session__process_events(session,
1466                                                      session->header.data_offset,
1467                                                      session->header.data_size,
1468                                                      size, tool);
1469         else
1470                 err = __perf_session__process_pipe_events(session, tool);
1471
1472         return err;
1473 }
1474
1475 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1476 {
1477         struct perf_evsel *evsel;
1478
1479         evlist__for_each(session->evlist, evsel) {
1480                 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1481                         return true;
1482         }
1483
1484         pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1485         return false;
1486 }
1487
1488 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1489                                      const char *symbol_name, u64 addr)
1490 {
1491         char *bracket;
1492         enum map_type i;
1493         struct ref_reloc_sym *ref;
1494
1495         ref = zalloc(sizeof(struct ref_reloc_sym));
1496         if (ref == NULL)
1497                 return -ENOMEM;
1498
1499         ref->name = strdup(symbol_name);
1500         if (ref->name == NULL) {
1501                 free(ref);
1502                 return -ENOMEM;
1503         }
1504
1505         bracket = strchr(ref->name, ']');
1506         if (bracket)
1507                 *bracket = '\0';
1508
1509         ref->addr = addr;
1510
1511         for (i = 0; i < MAP__NR_TYPES; ++i) {
1512                 struct kmap *kmap = map__kmap(maps[i]);
1513                 kmap->ref_reloc_sym = ref;
1514         }
1515
1516         return 0;
1517 }
1518
1519 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1520 {
1521         return machines__fprintf_dsos(&session->machines, fp);
1522 }
1523
1524 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1525                                           bool (skip)(struct dso *dso, int parm), int parm)
1526 {
1527         return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1528 }
1529
1530 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1531 {
1532         struct perf_evsel *pos;
1533         size_t ret = fprintf(fp, "Aggregated stats:\n");
1534
1535         ret += events_stats__fprintf(&session->stats, fp);
1536
1537         evlist__for_each(session->evlist, pos) {
1538                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1539                 ret += events_stats__fprintf(&pos->hists.stats, fp);
1540         }
1541
1542         return ret;
1543 }
1544
1545 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1546 {
1547         /*
1548          * FIXME: Here we have to actually print all the machines in this
1549          * session, not just the host...
1550          */
1551         return machine__fprintf(&session->machines.host, fp);
1552 }
1553
1554 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1555                                               unsigned int type)
1556 {
1557         struct perf_evsel *pos;
1558
1559         evlist__for_each(session->evlist, pos) {
1560                 if (pos->attr.type == type)
1561                         return pos;
1562         }
1563         return NULL;
1564 }
1565
1566 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1567                           struct addr_location *al,
1568                           unsigned int print_opts, unsigned int stack_depth)
1569 {
1570         struct callchain_cursor_node *node;
1571         int print_ip = print_opts & PRINT_IP_OPT_IP;
1572         int print_sym = print_opts & PRINT_IP_OPT_SYM;
1573         int print_dso = print_opts & PRINT_IP_OPT_DSO;
1574         int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1575         int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1576         int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1577         char s = print_oneline ? ' ' : '\t';
1578
1579         if (symbol_conf.use_callchain && sample->callchain) {
1580                 struct addr_location node_al;
1581
1582                 if (machine__resolve_callchain(al->machine, evsel, al->thread,
1583                                                sample, NULL, NULL,
1584                                                PERF_MAX_STACK_DEPTH) != 0) {
1585                         if (verbose)
1586                                 error("Failed to resolve callchain. Skipping\n");
1587                         return;
1588                 }
1589                 callchain_cursor_commit(&callchain_cursor);
1590
1591                 if (print_symoffset)
1592                         node_al = *al;
1593
1594                 while (stack_depth) {
1595                         u64 addr = 0;
1596
1597                         node = callchain_cursor_current(&callchain_cursor);
1598                         if (!node)
1599                                 break;
1600
1601                         if (node->sym && node->sym->ignore)
1602                                 goto next;
1603
1604                         if (print_ip)
1605                                 printf("%c%16" PRIx64, s, node->ip);
1606
1607                         if (node->map)
1608                                 addr = node->map->map_ip(node->map, node->ip);
1609
1610                         if (print_sym) {
1611                                 printf(" ");
1612                                 if (print_symoffset) {
1613                                         node_al.addr = addr;
1614                                         node_al.map  = node->map;
1615                                         symbol__fprintf_symname_offs(node->sym, &node_al, stdout);
1616                                 } else
1617                                         symbol__fprintf_symname(node->sym, stdout);
1618                         }
1619
1620                         if (print_dso) {
1621                                 printf(" (");
1622                                 map__fprintf_dsoname(node->map, stdout);
1623                                 printf(")");
1624                         }
1625
1626                         if (print_srcline)
1627                                 map__fprintf_srcline(node->map, addr, "\n  ",
1628                                                      stdout);
1629
1630                         if (!print_oneline)
1631                                 printf("\n");
1632
1633                         stack_depth--;
1634 next:
1635                         callchain_cursor_advance(&callchain_cursor);
1636                 }
1637
1638         } else {
1639                 if (al->sym && al->sym->ignore)
1640                         return;
1641
1642                 if (print_ip)
1643                         printf("%16" PRIx64, sample->ip);
1644
1645                 if (print_sym) {
1646                         printf(" ");
1647                         if (print_symoffset)
1648                                 symbol__fprintf_symname_offs(al->sym, al,
1649                                                              stdout);
1650                         else
1651                                 symbol__fprintf_symname(al->sym, stdout);
1652                 }
1653
1654                 if (print_dso) {
1655                         printf(" (");
1656                         map__fprintf_dsoname(al->map, stdout);
1657                         printf(")");
1658                 }
1659
1660                 if (print_srcline)
1661                         map__fprintf_srcline(al->map, al->addr, "\n  ", stdout);
1662         }
1663 }
1664
1665 int perf_session__cpu_bitmap(struct perf_session *session,
1666                              const char *cpu_list, unsigned long *cpu_bitmap)
1667 {
1668         int i, err = -1;
1669         struct cpu_map *map;
1670
1671         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1672                 struct perf_evsel *evsel;
1673
1674                 evsel = perf_session__find_first_evtype(session, i);
1675                 if (!evsel)
1676                         continue;
1677
1678                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1679                         pr_err("File does not contain CPU events. "
1680                                "Remove -c option to proceed.\n");
1681                         return -1;
1682                 }
1683         }
1684
1685         map = cpu_map__new(cpu_list);
1686         if (map == NULL) {
1687                 pr_err("Invalid cpu_list\n");
1688                 return -1;
1689         }
1690
1691         for (i = 0; i < map->nr; i++) {
1692                 int cpu = map->map[i];
1693
1694                 if (cpu >= MAX_NR_CPUS) {
1695                         pr_err("Requested CPU %d too large. "
1696                                "Consider raising MAX_NR_CPUS\n", cpu);
1697                         goto out_delete_map;
1698                 }
1699
1700                 set_bit(cpu, cpu_bitmap);
1701         }
1702
1703         err = 0;
1704
1705 out_delete_map:
1706         cpu_map__delete(map);
1707         return err;
1708 }
1709
1710 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1711                                 bool full)
1712 {
1713         struct stat st;
1714         int fd, ret;
1715
1716         if (session == NULL || fp == NULL)
1717                 return;
1718
1719         fd = perf_data_file__fd(session->file);
1720
1721         ret = fstat(fd, &st);
1722         if (ret == -1)
1723                 return;
1724
1725         fprintf(fp, "# ========\n");
1726         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1727         perf_header__fprintf_info(session, fp, full);
1728         fprintf(fp, "# ========\n#\n");
1729 }
1730
1731
1732 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1733                                              const struct perf_evsel_str_handler *assocs,
1734                                              size_t nr_assocs)
1735 {
1736         struct perf_evsel *evsel;
1737         size_t i;
1738         int err;
1739
1740         for (i = 0; i < nr_assocs; i++) {
1741                 /*
1742                  * Adding a handler for an event not in the session,
1743                  * just ignore it.
1744                  */
1745                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1746                 if (evsel == NULL)
1747                         continue;
1748
1749                 err = -EEXIST;
1750                 if (evsel->handler != NULL)
1751                         goto out;
1752                 evsel->handler = assocs[i].handler;
1753         }
1754
1755         err = 0;
1756 out:
1757         return err;
1758 }