tools/perf/util/session.c

   1 #define _FILE_OFFSET_BITS 64
   2
   3 #include <linux/kernel.h>
   4
   5 #include <byteswap.h>
   6 #include <unistd.h>
   7 #include <sys/types.h>
   8 #include <sys/mman.h>
   9
  10 #include "evlist.h"
  11 #include "evsel.h"
  12 #include "session.h"
  13 #include "sort.h"
  14 #include "util.h"
  15
  16 static int perf_session__open(struct perf_session *self, bool force)
  17 {
  18         struct stat input_stat;
  19
  20         if (!strcmp(self->filename, "-")) {
  21                 self->fd_pipe = true;
  22                 self->fd = STDIN_FILENO;
  23
  24                 if (perf_session__read_header(self, self->fd) < 0)
  25                         pr_err("incompatible file format");
  26
  27                 return 0;
  28         }
  29
  30         self->fd = open(self->filename, O_RDONLY);
  31         if (self->fd < 0) {
  32                 int err = errno;
  33
  34                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  35                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  36                         pr_err("  (try 'perf record' first)");
  37                 pr_err("\n");
  38                 return -errno;
  39         }
  40
  41         if (fstat(self->fd, &input_stat) < 0)
  42                 goto out_close;
  43
  44         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  45                 pr_err("file %s not owned by current user or root\n",
  46                        self->filename);
  47                 goto out_close;
  48         }
  49
  50         if (!input_stat.st_size) {
  51                 pr_info("zero-sized file (%s), nothing to do!\n",
  52                         self->filename);
  53                 goto out_close;
  54         }
  55
  56         if (perf_session__read_header(self, self->fd) < 0) {
  57                 pr_err("incompatible file format");
  58                 goto out_close;
  59         }
  60
  61         self->size = input_stat.st_size;
  62         return 0;
  63
  64 out_close:
  65         close(self->fd);
  66         self->fd = -1;
  67         return -1;
  68 }
  69
  70 static void perf_session__id_header_size(struct perf_session *session)
  71 {
  72        struct perf_sample *data;
  73        u64 sample_type = session->sample_type;
  74        u16 size = 0;
  75
  76         if (!session->sample_id_all)
  77                 goto out;
  78
  79        if (sample_type & PERF_SAMPLE_TID)
  80                size += sizeof(data->tid) * 2;
  81
  82        if (sample_type & PERF_SAMPLE_TIME)
  83                size += sizeof(data->time);
  84
  85        if (sample_type & PERF_SAMPLE_ID)
  86                size += sizeof(data->id);
  87
  88        if (sample_type & PERF_SAMPLE_STREAM_ID)
  89                size += sizeof(data->stream_id);
  90
  91        if (sample_type & PERF_SAMPLE_CPU)
  92                size += sizeof(data->cpu) * 2;
  93 out:
  94        session->id_hdr_size = size;
  95 }
  96
  97 void perf_session__update_sample_type(struct perf_session *self)
  98 {
  99         self->sample_type = perf_evlist__sample_type(self->evlist);
 100         self->sample_size = perf_sample_size(self->sample_type);
 101         self->sample_id_all = perf_evlist__sample_id_all(self->evlist);
 102         perf_session__id_header_size(self);
 103 }
 104
 105 int perf_session__create_kernel_maps(struct perf_session *self)
 106 {
 107         int ret = machine__create_kernel_maps(&self->host_machine);
 108
 109         if (ret >= 0)
 110                 ret = machines__create_guest_kernel_maps(&self->machines);
 111         return ret;
 112 }
 113
 114 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 115 {
 116         machine__destroy_kernel_maps(&self->host_machine);
 117         machines__destroy_guest_kernel_maps(&self->machines);
 118 }
 119
 120 struct perf_session *perf_session__new(const char *filename, int mode,
 121                                        bool force, bool repipe,
 122                                        struct perf_event_ops *ops)
 123 {
 124         size_t len = filename ? strlen(filename) + 1 : 0;
 125         struct perf_session *self = zalloc(sizeof(*self) + len);
 126
 127         if (self == NULL)
 128                 goto out;
 129
 130         memcpy(self->filename, filename, len);
 131         self->threads = RB_ROOT;
 132         INIT_LIST_HEAD(&self->dead_threads);
 133         self->last_match = NULL;
 134         /*
 135          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 136          * slices. On 32bit we use 32MB.
 137          */
 138 #if BITS_PER_LONG == 64
 139         self->mmap_window = ULLONG_MAX;
 140 #else
 141         self->mmap_window = 32 * 1024 * 1024ULL;
 142 #endif
 143         self->machines = RB_ROOT;
 144         self->repipe = repipe;
 145         INIT_LIST_HEAD(&self->ordered_samples.samples);
 146         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 147         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 148         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 149
 150         if (mode == O_RDONLY) {
 151                 if (perf_session__open(self, force) < 0)
 152                         goto out_delete;
 153                 perf_session__update_sample_type(self);
 154         } else if (mode == O_WRONLY) {
 155                 /*
 156                  * In O_RDONLY mode this will be performed when reading the
 157                  * kernel MMAP event, in perf_event__process_mmap().
 158                  */
 159                 if (perf_session__create_kernel_maps(self) < 0)
 160                         goto out_delete;
 161         }
 162
 163         if (ops && ops->ordering_requires_timestamps &&
 164             ops->ordered_samples && !self->sample_id_all) {
 165                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 166                 ops->ordered_samples = false;
 167         }
 168
 169 out:
 170         return self;
 171 out_delete:
 172         perf_session__delete(self);
 173         return NULL;
 174 }
 175
 176 static void perf_session__delete_dead_threads(struct perf_session *self)
 177 {
 178         struct thread *n, *t;
 179
 180         list_for_each_entry_safe(t, n, &self->dead_threads, node) {
 181                 list_del(&t->node);
 182                 thread__delete(t);
 183         }
 184 }
 185
 186 static void perf_session__delete_threads(struct perf_session *self)
 187 {
 188         struct rb_node *nd = rb_first(&self->threads);
 189
 190         while (nd) {
 191                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 192
 193                 rb_erase(&t->rb_node, &self->threads);
 194                 nd = rb_next(nd);
 195                 thread__delete(t);
 196         }
 197 }
 198
 199 void perf_session__delete(struct perf_session *self)
 200 {
 201         perf_session__destroy_kernel_maps(self);
 202         perf_session__delete_dead_threads(self);
 203         perf_session__delete_threads(self);
 204         machine__exit(&self->host_machine);
 205         close(self->fd);
 206         free(self);
 207 }
 208
 209 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
 210 {
 211         self->last_match = NULL;
 212         rb_erase(&th->rb_node, &self->threads);
 213         /*
 214          * We may have references to this thread, for instance in some hist_entry
 215          * instances, so just move them to a separate list.
 216          */
 217         list_add_tail(&th->node, &self->dead_threads);
 218 }
 219
 220 static bool symbol__match_parent_regex(struct symbol *sym)
 221 {
 222         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 223                 return 1;
 224
 225         return 0;
 226 }
 227
 228 int perf_session__resolve_callchain(struct perf_session *self,
 229                                     struct thread *thread,
 230                                     struct ip_callchain *chain,
 231                                     struct symbol **parent)
 232 {
 233         u8 cpumode = PERF_RECORD_MISC_USER;
 234         unsigned int i;
 235         int err;
 236
 237         callchain_cursor_reset(&self->callchain_cursor);
 238
 239         for (i = 0; i < chain->nr; i++) {
 240                 u64 ip = chain->ips[i];
 241                 struct addr_location al;
 242
 243                 if (ip >= PERF_CONTEXT_MAX) {
 244                         switch (ip) {
 245                         case PERF_CONTEXT_HV:
 246                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 247                         case PERF_CONTEXT_KERNEL:
 248                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 249                         case PERF_CONTEXT_USER:
 250                                 cpumode = PERF_RECORD_MISC_USER;        break;
 251                         default:
 252                                 break;
 253                         }
 254                         continue;
 255                 }
 256
 257                 al.filtered = false;
 258                 thread__find_addr_location(thread, self, cpumode,
 259                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 260                 if (al.sym != NULL) {
 261                         if (sort__has_parent && !*parent &&
 262                             symbol__match_parent_regex(al.sym))
 263                                 *parent = al.sym;
 264                         if (!symbol_conf.use_callchain)
 265                                 break;
 266                 }
 267
 268                 err = callchain_cursor_append(&self->callchain_cursor,
 269                                               ip, al.map, al.sym);
 270                 if (err)
 271                         return err;
 272         }
 273
 274         return 0;
 275 }
 276
 277 static int process_event_synth_stub(union perf_event *event __used,
 278                                     struct perf_session *session __used)
 279 {
 280         dump_printf(": unhandled!\n");
 281         return 0;
 282 }
 283
 284 static int process_event_sample_stub(union perf_event *event __used,
 285                                      struct perf_sample *sample __used,
 286                                      struct perf_evsel *evsel __used,
 287                                      struct perf_session *session __used)
 288 {
 289         dump_printf(": unhandled!\n");
 290         return 0;
 291 }
 292
 293 static int process_event_stub(union perf_event *event __used,
 294                               struct perf_sample *sample __used,
 295                               struct perf_session *session __used)
 296 {
 297         dump_printf(": unhandled!\n");
 298         return 0;
 299 }
 300
 301 static int process_finished_round_stub(union perf_event *event __used,
 302                                        struct perf_session *session __used,
 303                                        struct perf_event_ops *ops __used)
 304 {
 305         dump_printf(": unhandled!\n");
 306         return 0;
 307 }
 308
 309 static int process_finished_round(union perf_event *event,
 310                                   struct perf_session *session,
 311                                   struct perf_event_ops *ops);
 312
 313 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 314 {
 315         if (handler->sample == NULL)
 316                 handler->sample = process_event_sample_stub;
 317         if (handler->mmap == NULL)
 318                 handler->mmap = process_event_stub;
 319         if (handler->comm == NULL)
 320                 handler->comm = process_event_stub;
 321         if (handler->fork == NULL)
 322                 handler->fork = process_event_stub;
 323         if (handler->exit == NULL)
 324                 handler->exit = process_event_stub;
 325         if (handler->lost == NULL)
 326                 handler->lost = perf_event__process_lost;
 327         if (handler->read == NULL)
 328                 handler->read = process_event_stub;
 329         if (handler->throttle == NULL)
 330                 handler->throttle = process_event_stub;
 331         if (handler->unthrottle == NULL)
 332                 handler->unthrottle = process_event_stub;
 333         if (handler->attr == NULL)
 334                 handler->attr = process_event_synth_stub;
 335         if (handler->event_type == NULL)
 336                 handler->event_type = process_event_synth_stub;
 337         if (handler->tracing_data == NULL)
 338                 handler->tracing_data = process_event_synth_stub;
 339         if (handler->build_id == NULL)
 340                 handler->build_id = process_event_synth_stub;
 341         if (handler->finished_round == NULL) {
 342                 if (handler->ordered_samples)
 343                         handler->finished_round = process_finished_round;
 344                 else
 345                         handler->finished_round = process_finished_round_stub;
 346         }
 347 }
 348
 349 void mem_bswap_64(void *src, int byte_size)
 350 {
 351         u64 *m = src;
 352
 353         while (byte_size > 0) {
 354                 *m = bswap_64(*m);
 355                 byte_size -= sizeof(u64);
 356                 ++m;
 357         }
 358 }
 359
 360 static void perf_event__all64_swap(union perf_event *event)
 361 {
 362         struct perf_event_header *hdr = &event->header;
 363         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 364 }
 365
 366 static void perf_event__comm_swap(union perf_event *event)
 367 {
 368         event->comm.pid = bswap_32(event->comm.pid);
 369         event->comm.tid = bswap_32(event->comm.tid);
 370 }
 371
 372 static void perf_event__mmap_swap(union perf_event *event)
 373 {
 374         event->mmap.pid   = bswap_32(event->mmap.pid);
 375         event->mmap.tid   = bswap_32(event->mmap.tid);
 376         event->mmap.start = bswap_64(event->mmap.start);
 377         event->mmap.len   = bswap_64(event->mmap.len);
 378         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 379 }
 380
 381 static void perf_event__task_swap(union perf_event *event)
 382 {
 383         event->fork.pid  = bswap_32(event->fork.pid);
 384         event->fork.tid  = bswap_32(event->fork.tid);
 385         event->fork.ppid = bswap_32(event->fork.ppid);
 386         event->fork.ptid = bswap_32(event->fork.ptid);
 387         event->fork.time = bswap_64(event->fork.time);
 388 }
 389
 390 static void perf_event__read_swap(union perf_event *event)
 391 {
 392         event->read.pid          = bswap_32(event->read.pid);
 393         event->read.tid          = bswap_32(event->read.tid);
 394         event->read.value        = bswap_64(event->read.value);
 395         event->read.time_enabled = bswap_64(event->read.time_enabled);
 396         event->read.time_running = bswap_64(event->read.time_running);
 397         event->read.id           = bswap_64(event->read.id);
 398 }
 399
 400 static void perf_event__attr_swap(union perf_event *event)
 401 {
 402         size_t size;
 403
 404         event->attr.attr.type           = bswap_32(event->attr.attr.type);
 405         event->attr.attr.size           = bswap_32(event->attr.attr.size);
 406         event->attr.attr.config         = bswap_64(event->attr.attr.config);
 407         event->attr.attr.sample_period  = bswap_64(event->attr.attr.sample_period);
 408         event->attr.attr.sample_type    = bswap_64(event->attr.attr.sample_type);
 409         event->attr.attr.read_format    = bswap_64(event->attr.attr.read_format);
 410         event->attr.attr.wakeup_events  = bswap_32(event->attr.attr.wakeup_events);
 411         event->attr.attr.bp_type        = bswap_32(event->attr.attr.bp_type);
 412         event->attr.attr.bp_addr        = bswap_64(event->attr.attr.bp_addr);
 413         event->attr.attr.bp_len         = bswap_64(event->attr.attr.bp_len);
 414
 415         size = event->header.size;
 416         size -= (void *)&event->attr.id - (void *)event;
 417         mem_bswap_64(event->attr.id, size);
 418 }
 419
 420 static void perf_event__event_type_swap(union perf_event *event)
 421 {
 422         event->event_type.event_type.event_id =
 423                 bswap_64(event->event_type.event_type.event_id);
 424 }
 425
 426 static void perf_event__tracing_data_swap(union perf_event *event)
 427 {
 428         event->tracing_data.size = bswap_32(event->tracing_data.size);
 429 }
 430
 431 typedef void (*perf_event__swap_op)(union perf_event *event);
 432
 433 static perf_event__swap_op perf_event__swap_ops[] = {
 434         [PERF_RECORD_MMAP]                = perf_event__mmap_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_SAMPLE]              = perf_event__all64_swap,
 441         [PERF_RECORD_HEADER_ATTR]         = perf_event__attr_swap,
 442         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 443         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 444         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 445         [PERF_RECORD_HEADER_MAX]          = NULL,
 446 };
 447
 448 struct sample_queue {
 449         u64                     timestamp;
 450         u64                     file_offset;
 451         union perf_event        *event;
 452         struct list_head        list;
 453 };
 454
 455 static void perf_session_free_sample_buffers(struct perf_session *session)
 456 {
 457         struct ordered_samples *os = &session->ordered_samples;
 458
 459         while (!list_empty(&os->to_free)) {
 460                 struct sample_queue *sq;
 461
 462                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 463                 list_del(&sq->list);
 464                 free(sq);
 465         }
 466 }
 467
 468 static int perf_session_deliver_event(struct perf_session *session,
 469                                       union perf_event *event,
 470                                       struct perf_sample *sample,
 471                                       struct perf_event_ops *ops,
 472                                       u64 file_offset);
 473
 474 static void flush_sample_queue(struct perf_session *s,
 475                                struct perf_event_ops *ops)
 476 {
 477         struct ordered_samples *os = &s->ordered_samples;
 478         struct list_head *head = &os->samples;
 479         struct sample_queue *tmp, *iter;
 480         struct perf_sample sample;
 481         u64 limit = os->next_flush;
 482         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 483
 484         if (!ops->ordered_samples || !limit)
 485                 return;
 486
 487         list_for_each_entry_safe(iter, tmp, head, list) {
 488                 if (iter->timestamp > limit)
 489                         break;
 490
 491                 perf_session__parse_sample(s, iter->event, &sample);
 492                 perf_session_deliver_event(s, iter->event, &sample, ops,
 493                                            iter->file_offset);
 494
 495                 os->last_flush = iter->timestamp;
 496                 list_del(&iter->list);
 497                 list_add(&iter->list, &os->sample_cache);
 498         }
 499
 500         if (list_empty(head)) {
 501                 os->last_sample = NULL;
 502         } else if (last_ts <= limit) {
 503                 os->last_sample =
 504                         list_entry(head->prev, struct sample_queue, list);
 505         }
 506 }
 507
 508 /*
 509  * When perf record finishes a pass on every buffers, it records this pseudo
 510  * event.
 511  * We record the max timestamp t found in the pass n.
 512  * Assuming these timestamps are monotonic across cpus, we know that if
 513  * a buffer still has events with timestamps below t, they will be all
 514  * available and then read in the pass n + 1.
 515  * Hence when we start to read the pass n + 2, we can safely flush every
 516  * events with timestamps below t.
 517  *
 518  *    ============ PASS n =================
 519  *       CPU 0         |   CPU 1
 520  *                     |
 521  *    cnt1 timestamps  |   cnt2 timestamps
 522  *          1          |         2
 523  *          2          |         3
 524  *          -          |         4  <--- max recorded
 525  *
 526  *    ============ PASS n + 1 ==============
 527  *       CPU 0         |   CPU 1
 528  *                     |
 529  *    cnt1 timestamps  |   cnt2 timestamps
 530  *          3          |         5
 531  *          4          |         6
 532  *          5          |         7 <---- max recorded
 533  *
 534  *      Flush every events below timestamp 4
 535  *
 536  *    ============ PASS n + 2 ==============
 537  *       CPU 0         |   CPU 1
 538  *                     |
 539  *    cnt1 timestamps  |   cnt2 timestamps
 540  *          6          |         8
 541  *          7          |         9
 542  *          -          |         10
 543  *
 544  *      Flush every events below timestamp 7
 545  *      etc...
 546  */
 547 static int process_finished_round(union perf_event *event __used,
 548                                   struct perf_session *session,
 549                                   struct perf_event_ops *ops)
 550 {
 551         flush_sample_queue(session, ops);
 552         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 553
 554         return 0;
 555 }
 556
 557 /* The queue is ordered by time */
 558 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 559 {
 560         struct ordered_samples *os = &s->ordered_samples;
 561         struct sample_queue *sample = os->last_sample;
 562         u64 timestamp = new->timestamp;
 563         struct list_head *p;
 564
 565         os->last_sample = new;
 566
 567         if (!sample) {
 568                 list_add(&new->list, &os->samples);
 569                 os->max_timestamp = timestamp;
 570                 return;
 571         }
 572
 573         /*
 574          * last_sample might point to some random place in the list as it's
 575          * the last queued event. We expect that the new event is close to
 576          * this.
 577          */
 578         if (sample->timestamp <= timestamp) {
 579                 while (sample->timestamp <= timestamp) {
 580                         p = sample->list.next;
 581                         if (p == &os->samples) {
 582                                 list_add_tail(&new->list, &os->samples);
 583                                 os->max_timestamp = timestamp;
 584                                 return;
 585                         }
 586                         sample = list_entry(p, struct sample_queue, list);
 587                 }
 588                 list_add_tail(&new->list, &sample->list);
 589         } else {
 590                 while (sample->timestamp > timestamp) {
 591                         p = sample->list.prev;
 592                         if (p == &os->samples) {
 593                                 list_add(&new->list, &os->samples);
 594                                 return;
 595                         }
 596                         sample = list_entry(p, struct sample_queue, list);
 597                 }
 598                 list_add(&new->list, &sample->list);
 599         }
 600 }
 601
 602 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 603
 604 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 605                                     struct perf_sample *sample, u64 file_offset)
 606 {
 607         struct ordered_samples *os = &s->ordered_samples;
 608         struct list_head *sc = &os->sample_cache;
 609         u64 timestamp = sample->time;
 610         struct sample_queue *new;
 611
 612         if (!timestamp || timestamp == ~0ULL)
 613                 return -ETIME;
 614
 615         if (timestamp < s->ordered_samples.last_flush) {
 616                 printf("Warning: Timestamp below last timeslice flush\n");
 617                 return -EINVAL;
 618         }
 619
 620         if (!list_empty(sc)) {
 621                 new = list_entry(sc->next, struct sample_queue, list);
 622                 list_del(&new->list);
 623         } else if (os->sample_buffer) {
 624                 new = os->sample_buffer + os->sample_buffer_idx;
 625                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 626                         os->sample_buffer = NULL;
 627         } else {
 628                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 629                 if (!os->sample_buffer)
 630                         return -ENOMEM;
 631                 list_add(&os->sample_buffer->list, &os->to_free);
 632                 os->sample_buffer_idx = 2;
 633                 new = os->sample_buffer + 1;
 634         }
 635
 636         new->timestamp = timestamp;
 637         new->file_offset = file_offset;
 638         new->event = event;
 639
 640         __queue_event(new, s);
 641
 642         return 0;
 643 }
 644
 645 static void callchain__printf(struct perf_sample *sample)
 646 {
 647         unsigned int i;
 648
 649         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 650
 651         for (i = 0; i < sample->callchain->nr; i++)
 652                 printf("..... %2d: %016" PRIx64 "\n",
 653                        i, sample->callchain->ips[i]);
 654 }
 655
 656 static void perf_session__print_tstamp(struct perf_session *session,
 657                                        union perf_event *event,
 658                                        struct perf_sample *sample)
 659 {
 660         if (event->header.type != PERF_RECORD_SAMPLE &&
 661             !session->sample_id_all) {
 662                 fputs("-1 -1 ", stdout);
 663                 return;
 664         }
 665
 666         if ((session->sample_type & PERF_SAMPLE_CPU))
 667                 printf("%u ", sample->cpu);
 668
 669         if (session->sample_type & PERF_SAMPLE_TIME)
 670                 printf("%" PRIu64 " ", sample->time);
 671 }
 672
 673 static void dump_event(struct perf_session *session, union perf_event *event,
 674                        u64 file_offset, struct perf_sample *sample)
 675 {
 676         if (!dump_trace)
 677                 return;
 678
 679         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 680                file_offset, event->header.size, event->header.type);
 681
 682         trace_event(event);
 683
 684         if (sample)
 685                 perf_session__print_tstamp(session, event, sample);
 686
 687         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 688                event->header.size, perf_event__name(event->header.type));
 689 }
 690
 691 static void dump_sample(struct perf_session *session, union perf_event *event,
 692                         struct perf_sample *sample)
 693 {
 694         if (!dump_trace)
 695                 return;
 696
 697         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 "\n",
 698                event->header.misc, sample->pid, sample->tid, sample->ip,
 699                sample->period);
 700
 701         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 702                 callchain__printf(sample);
 703 }
 704
 705 static int perf_session_deliver_event(struct perf_session *session,
 706                                       union perf_event *event,
 707                                       struct perf_sample *sample,
 708                                       struct perf_event_ops *ops,
 709                                       u64 file_offset)
 710 {
 711         struct perf_evsel *evsel;
 712
 713         dump_event(session, event, file_offset, sample);
 714
 715         switch (event->header.type) {
 716         case PERF_RECORD_SAMPLE:
 717                 dump_sample(session, event, sample);
 718                 evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 719                 if (evsel == NULL) {
 720                         ++session->hists.stats.nr_unknown_id;
 721                         return -1;
 722                 }
 723                 return ops->sample(event, sample, evsel, session);
 724         case PERF_RECORD_MMAP:
 725                 return ops->mmap(event, sample, session);
 726         case PERF_RECORD_COMM:
 727                 return ops->comm(event, sample, session);
 728         case PERF_RECORD_FORK:
 729                 return ops->fork(event, sample, session);
 730         case PERF_RECORD_EXIT:
 731                 return ops->exit(event, sample, session);
 732         case PERF_RECORD_LOST:
 733                 return ops->lost(event, sample, session);
 734         case PERF_RECORD_READ:
 735                 return ops->read(event, sample, session);
 736         case PERF_RECORD_THROTTLE:
 737                 return ops->throttle(event, sample, session);
 738         case PERF_RECORD_UNTHROTTLE:
 739                 return ops->unthrottle(event, sample, session);
 740         default:
 741                 ++session->hists.stats.nr_unknown_events;
 742                 return -1;
 743         }
 744 }
 745
 746 static int perf_session__preprocess_sample(struct perf_session *session,
 747                                            union perf_event *event, struct perf_sample *sample)
 748 {
 749         if (event->header.type != PERF_RECORD_SAMPLE ||
 750             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 751                 return 0;
 752
 753         if (!ip_callchain__valid(sample->callchain, event)) {
 754                 pr_debug("call-chain problem with event, skipping it.\n");
 755                 ++session->hists.stats.nr_invalid_chains;
 756                 session->hists.stats.total_invalid_chains += sample->period;
 757                 return -EINVAL;
 758         }
 759         return 0;
 760 }
 761
 762 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
 763                                             struct perf_event_ops *ops, u64 file_offset)
 764 {
 765         dump_event(session, event, file_offset, NULL);
 766
 767         /* These events are processed right away */
 768         switch (event->header.type) {
 769         case PERF_RECORD_HEADER_ATTR:
 770                 return ops->attr(event, session);
 771         case PERF_RECORD_HEADER_EVENT_TYPE:
 772                 return ops->event_type(event, session);
 773         case PERF_RECORD_HEADER_TRACING_DATA:
 774                 /* setup for reading amidst mmap */
 775                 lseek(session->fd, file_offset, SEEK_SET);
 776                 return ops->tracing_data(event, session);
 777         case PERF_RECORD_HEADER_BUILD_ID:
 778                 return ops->build_id(event, session);
 779         case PERF_RECORD_FINISHED_ROUND:
 780                 return ops->finished_round(event, session, ops);
 781         default:
 782                 return -EINVAL;
 783         }
 784 }
 785
 786 static int perf_session__process_event(struct perf_session *session,
 787                                        union perf_event *event,
 788                                        struct perf_event_ops *ops,
 789                                        u64 file_offset)
 790 {
 791         struct perf_sample sample;
 792         int ret;
 793
 794         if (session->header.needs_swap &&
 795             perf_event__swap_ops[event->header.type])
 796                 perf_event__swap_ops[event->header.type](event);
 797
 798         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 799                 return -EINVAL;
 800
 801         hists__inc_nr_events(&session->hists, event->header.type);
 802
 803         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 804                 return perf_session__process_user_event(session, event, ops, file_offset);
 805
 806         /*
 807          * For all kernel events we get the sample data
 808          */
 809         perf_session__parse_sample(session, event, &sample);
 810
 811         /* Preprocess sample records - precheck callchains */
 812         if (perf_session__preprocess_sample(session, event, &sample))
 813                 return 0;
 814
 815         if (ops->ordered_samples) {
 816                 ret = perf_session_queue_event(session, event, &sample,
 817                                                file_offset);
 818                 if (ret != -ETIME)
 819                         return ret;
 820         }
 821
 822         return perf_session_deliver_event(session, event, &sample, ops,
 823                                           file_offset);
 824 }
 825
 826 void perf_event_header__bswap(struct perf_event_header *self)
 827 {
 828         self->type = bswap_32(self->type);
 829         self->misc = bswap_16(self->misc);
 830         self->size = bswap_16(self->size);
 831 }
 832
 833 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 834 {
 835         struct thread *thread = perf_session__findnew(self, 0);
 836
 837         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 838                 pr_err("problem inserting idle task.\n");
 839                 thread = NULL;
 840         }
 841
 842         return thread;
 843 }
 844
 845 static void perf_session__warn_about_errors(const struct perf_session *session,
 846                                             const struct perf_event_ops *ops)
 847 {
 848         if (ops->lost == perf_event__process_lost &&
 849             session->hists.stats.total_lost != 0) {
 850                 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64
 851                             "!\n\nCheck IO/CPU overload!\n\n",
 852                             session->hists.stats.total_period,
 853                             session->hists.stats.total_lost);
 854         }
 855
 856         if (session->hists.stats.nr_unknown_events != 0) {
 857                 ui__warning("Found %u unknown events!\n\n"
 858                             "Is this an older tool processing a perf.data "
 859                             "file generated by a more recent tool?\n\n"
 860                             "If that is not the case, consider "
 861                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 862                             session->hists.stats.nr_unknown_events);
 863         }
 864
 865         if (session->hists.stats.nr_unknown_id != 0) {
 866                 ui__warning("%u samples with id not present in the header\n",
 867                             session->hists.stats.nr_unknown_id);
 868         }
 869
 870         if (session->hists.stats.nr_invalid_chains != 0) {
 871                 ui__warning("Found invalid callchains!\n\n"
 872                             "%u out of %u events were discarded for this reason.\n\n"
 873                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 874                             session->hists.stats.nr_invalid_chains,
 875                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 876         }
 877 }
 878
 879 #define session_done()  (*(volatile int *)(&session_done))
 880 volatile int session_done;
 881
 882 static int __perf_session__process_pipe_events(struct perf_session *self,
 883                                                struct perf_event_ops *ops)
 884 {
 885         union perf_event event;
 886         uint32_t size;
 887         int skip = 0;
 888         u64 head;
 889         int err;
 890         void *p;
 891
 892         perf_event_ops__fill_defaults(ops);
 893
 894         head = 0;
 895 more:
 896         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 897         if (err <= 0) {
 898                 if (err == 0)
 899                         goto done;
 900
 901                 pr_err("failed to read event header\n");
 902                 goto out_err;
 903         }
 904
 905         if (self->header.needs_swap)
 906                 perf_event_header__bswap(&event.header);
 907
 908         size = event.header.size;
 909         if (size == 0)
 910                 size = 8;
 911
 912         p = &event;
 913         p += sizeof(struct perf_event_header);
 914
 915         if (size - sizeof(struct perf_event_header)) {
 916                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 917                 if (err <= 0) {
 918                         if (err == 0) {
 919                                 pr_err("unexpected end of event stream\n");
 920                                 goto done;
 921                         }
 922
 923                         pr_err("failed to read event data\n");
 924                         goto out_err;
 925                 }
 926         }
 927
 928         if (size == 0 ||
 929             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 930                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 931                             head, event.header.size, event.header.type);
 932                 /*
 933                  * assume we lost track of the stream, check alignment, and
 934                  * increment a single u64 in the hope to catch on again 'soon'.
 935                  */
 936                 if (unlikely(head & 7))
 937                         head &= ~7ULL;
 938
 939                 size = 8;
 940         }
 941
 942         head += size;
 943
 944         if (skip > 0)
 945                 head += skip;
 946
 947         if (!session_done())
 948                 goto more;
 949 done:
 950         err = 0;
 951 out_err:
 952         perf_session__warn_about_errors(self, ops);
 953         perf_session_free_sample_buffers(self);
 954         return err;
 955 }
 956
 957 int __perf_session__process_events(struct perf_session *session,
 958                                    u64 data_offset, u64 data_size,
 959                                    u64 file_size, struct perf_event_ops *ops)
 960 {
 961         u64 head, page_offset, file_offset, file_pos, progress_next;
 962         int err, mmap_prot, mmap_flags, map_idx = 0;
 963         struct ui_progress *progress;
 964         size_t  page_size, mmap_size;
 965         char *buf, *mmaps[8];
 966         union perf_event *event;
 967         uint32_t size;
 968
 969         perf_event_ops__fill_defaults(ops);
 970
 971         page_size = sysconf(_SC_PAGESIZE);
 972
 973         page_offset = page_size * (data_offset / page_size);
 974         file_offset = page_offset;
 975         head = data_offset - page_offset;
 976
 977         if (data_offset + data_size < file_size)
 978                 file_size = data_offset + data_size;
 979
 980         progress_next = file_size / 16;
 981         progress = ui_progress__new("Processing events...", file_size);
 982         if (progress == NULL)
 983                 return -1;
 984
 985         mmap_size = session->mmap_window;
 986         if (mmap_size > file_size)
 987                 mmap_size = file_size;
 988
 989         memset(mmaps, 0, sizeof(mmaps));
 990
 991         mmap_prot  = PROT_READ;
 992         mmap_flags = MAP_SHARED;
 993
 994         if (session->header.needs_swap) {
 995                 mmap_prot  |= PROT_WRITE;
 996                 mmap_flags = MAP_PRIVATE;
 997         }
 998 remap:
 999         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1000                    file_offset);
1001         if (buf == MAP_FAILED) {
1002                 pr_err("failed to mmap file\n");
1003                 err = -errno;
1004                 goto out_err;
1005         }
1006         mmaps[map_idx] = buf;
1007         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1008         file_pos = file_offset + head;
1009
1010 more:
1011         /*
1012          * Ensure we have enough space remaining to read
1013          * the size of the event in the headers.
1014          */
1015         if (head + sizeof(event->header) > mmap_size)
1016                 goto remap;
1017
1018         event = (union perf_event *)(buf + head);
1019
1020         if (session->header.needs_swap)
1021                 perf_event_header__bswap(&event->header);
1022
1023         if (head + event->header.size > mmap_size) {
1024                 if (mmaps[map_idx]) {
1025                         munmap(mmaps[map_idx], mmap_size);
1026                         mmaps[map_idx] = NULL;
1027                 }
1028
1029                 page_offset = page_size * (head / page_size);
1030                 file_offset += page_offset;
1031                 head -= page_offset;
1032                 goto remap;
1033         }
1034
1035         size = event->header.size;
1036
1037         if (size == 0 ||
1038             perf_session__process_event(session, event, ops, file_pos) < 0) {
1039                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1040                             file_offset + head, event->header.size,
1041                             event->header.type);
1042                 /*
1043                  * assume we lost track of the stream, check alignment, and
1044                  * increment a single u64 in the hope to catch on again 'soon'.
1045                  */
1046                 if (unlikely(head & 7))
1047                         head &= ~7ULL;
1048
1049                 size = 8;
1050         }
1051
1052         head += size;
1053         file_pos += size;
1054
1055         if (file_pos >= progress_next) {
1056                 progress_next += file_size / 16;
1057                 ui_progress__update(progress, file_pos);
1058         }
1059
1060         if (file_pos < file_size)
1061                 goto more;
1062
1063         err = 0;
1064         /* do the final flush for ordered samples */
1065         session->ordered_samples.next_flush = ULLONG_MAX;
1066         flush_sample_queue(session, ops);
1067 out_err:
1068         ui_progress__delete(progress);
1069         perf_session__warn_about_errors(session, ops);
1070         perf_session_free_sample_buffers(session);
1071         return err;
1072 }
1073
1074 int perf_session__process_events(struct perf_session *self,
1075                                  struct perf_event_ops *ops)
1076 {
1077         int err;
1078
1079         if (perf_session__register_idle_thread(self) == NULL)
1080                 return -ENOMEM;
1081
1082         if (!self->fd_pipe)
1083                 err = __perf_session__process_events(self,
1084                                                      self->header.data_offset,
1085                                                      self->header.data_size,
1086                                                      self->size, ops);
1087         else
1088                 err = __perf_session__process_pipe_events(self, ops);
1089
1090         return err;
1091 }
1092
1093 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1094 {
1095         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1096                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1097                 return false;
1098         }
1099
1100         return true;
1101 }
1102
1103 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1104                                              const char *symbol_name,
1105                                              u64 addr)
1106 {
1107         char *bracket;
1108         enum map_type i;
1109         struct ref_reloc_sym *ref;
1110
1111         ref = zalloc(sizeof(struct ref_reloc_sym));
1112         if (ref == NULL)
1113                 return -ENOMEM;
1114
1115         ref->name = strdup(symbol_name);
1116         if (ref->name == NULL) {
1117                 free(ref);
1118                 return -ENOMEM;
1119         }
1120
1121         bracket = strchr(ref->name, ']');
1122         if (bracket)
1123                 *bracket = '\0';
1124
1125         ref->addr = addr;
1126
1127         for (i = 0; i < MAP__NR_TYPES; ++i) {
1128                 struct kmap *kmap = map__kmap(maps[i]);
1129                 kmap->ref_reloc_sym = ref;
1130         }
1131
1132         return 0;
1133 }
1134
1135 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1136 {
1137         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1138                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1139                machines__fprintf_dsos(&self->machines, fp);
1140 }
1141
1142 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1143                                           bool with_hits)
1144 {
1145         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1146         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1147 }
1148
1149 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1150 {
1151         struct perf_evsel *pos;
1152         size_t ret = fprintf(fp, "Aggregated stats:\n");
1153
1154         ret += hists__fprintf_nr_events(&session->hists, fp);
1155
1156         list_for_each_entry(pos, &session->evlist->entries, node) {
1157                 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1158                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1159         }
1160
1161         return ret;
1162 }
1163
1164 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1165                                               unsigned int type)
1166 {
1167         struct perf_evsel *pos;
1168
1169         list_for_each_entry(pos, &session->evlist->entries, node) {
1170                 if (pos->attr.type == type)
1171                         return pos;
1172         }
1173         return NULL;
1174 }
1175
1176 void perf_session__print_symbols(union perf_event *event,
1177                                 struct perf_sample *sample,
1178                                 struct perf_session *session)
1179 {
1180         struct addr_location al;
1181         const char *symname, *dsoname;
1182         struct callchain_cursor *cursor = &session->callchain_cursor;
1183         struct callchain_cursor_node *node;
1184
1185         if (perf_event__preprocess_sample(event, session, &al, sample,
1186                                           NULL) < 0) {
1187                 error("problem processing %d event, skipping it.\n",
1188                         event->header.type);
1189                 return;
1190         }
1191
1192         if (symbol_conf.use_callchain && sample->callchain) {
1193
1194                 if (perf_session__resolve_callchain(session, al.thread,
1195                                                 sample->callchain, NULL) != 0) {
1196                         if (verbose)
1197                                 error("Failed to resolve callchain. Skipping\n");
1198                         return;
1199                 }
1200                 callchain_cursor_commit(cursor);
1201
1202                 while (1) {
1203                         node = callchain_cursor_current(cursor);
1204                         if (!node)
1205                                 break;
1206
1207                         if (node->sym && node->sym->name)
1208                                 symname = node->sym->name;
1209                         else
1210                                 symname = "";
1211
1212                         if (node->map && node->map->dso && node->map->dso->name)
1213                                 dsoname = node->map->dso->name;
1214                         else
1215                                 dsoname = "";
1216
1217                         printf("\t%16" PRIx64 " %s (%s)\n", node->ip, symname, dsoname);
1218
1219                         callchain_cursor_advance(cursor);
1220                 }
1221
1222         } else {
1223                 if (al.sym && al.sym->name)
1224                         symname = al.sym->name;
1225                 else
1226                         symname = "";
1227
1228                 if (al.map && al.map->dso && al.map->dso->name)
1229                         dsoname = al.map->dso->name;
1230                 else
1231                         dsoname = "";
1232
1233                 printf("%16" PRIx64 " %s (%s)", al.addr, symname, dsoname);
1234         }
1235 }