Pileus Git - ~andy/linux/blob - 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__read_evlist(struct perf_session *session)
  17 {
  18         int i, j;
  19
  20         session->evlist = perf_evlist__new(NULL, NULL);
  21         if (session->evlist == NULL)
  22                 return -ENOMEM;
  23
  24         for (i = 0; i < session->header.attrs; ++i) {
  25                 struct perf_header_attr *hattr = session->header.attr[i];
  26                 struct perf_evsel *evsel = perf_evsel__new(&hattr->attr, i);
  27
  28                 if (evsel == NULL)
  29                         goto out_delete_evlist;
  30                 /*
  31                  * Do it before so that if perf_evsel__alloc_id fails, this
  32                  * entry gets purged too at perf_evlist__delete().
  33                  */
  34                 perf_evlist__add(session->evlist, evsel);
  35                 /*
  36                  * We don't have the cpu and thread maps on the header, so
  37                  * for allocating the perf_sample_id table we fake 1 cpu and
  38                  * hattr->ids threads.
  39                  */
  40                 if (perf_evsel__alloc_id(evsel, 1, hattr->ids))
  41                         goto out_delete_evlist;
  42
  43                 for (j = 0; j < hattr->ids; ++j)
  44                         perf_evlist__id_hash(session->evlist, evsel, 0, j,
  45                                              hattr->id[j]);
  46         }
  47
  48         return 0;
  49
  50 out_delete_evlist:
  51         perf_evlist__delete(session->evlist);
  52         session->evlist = NULL;
  53         return -ENOMEM;
  54 }
  55
  56 static int perf_session__open(struct perf_session *self, bool force)
  57 {
  58         struct stat input_stat;
  59
  60         if (!strcmp(self->filename, "-")) {
  61                 self->fd_pipe = true;
  62                 self->fd = STDIN_FILENO;
  63
  64                 if (perf_header__read(self, self->fd) < 0)
  65                         pr_err("incompatible file format");
  66
  67                 return 0;
  68         }
  69
  70         self->fd = open(self->filename, O_RDONLY);
  71         if (self->fd < 0) {
  72                 int err = errno;
  73
  74                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  75                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  76                         pr_err("  (try 'perf record' first)");
  77                 pr_err("\n");
  78                 return -errno;
  79         }
  80
  81         if (fstat(self->fd, &input_stat) < 0)
  82                 goto out_close;
  83
  84         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  85                 pr_err("file %s not owned by current user or root\n",
  86                        self->filename);
  87                 goto out_close;
  88         }
  89
  90         if (!input_stat.st_size) {
  91                 pr_info("zero-sized file (%s), nothing to do!\n",
  92                         self->filename);
  93                 goto out_close;
  94         }
  95
  96         if (perf_header__read(self, self->fd) < 0) {
  97                 pr_err("incompatible file format");
  98                 goto out_close;
  99         }
 100
 101         if (perf_session__read_evlist(self) < 0) {
 102                 pr_err("Not enough memory to read the event selector list\n");
 103                 goto out_close;
 104         }
 105
 106         self->size = input_stat.st_size;
 107         return 0;
 108
 109 out_close:
 110         close(self->fd);
 111         self->fd = -1;
 112         return -1;
 113 }
 114
 115 static void perf_session__id_header_size(struct perf_session *session)
 116 {
 117        struct perf_sample *data;
 118        u64 sample_type = session->sample_type;
 119        u16 size = 0;
 120
 121         if (!session->sample_id_all)
 122                 goto out;
 123
 124        if (sample_type & PERF_SAMPLE_TID)
 125                size += sizeof(data->tid) * 2;
 126
 127        if (sample_type & PERF_SAMPLE_TIME)
 128                size += sizeof(data->time);
 129
 130        if (sample_type & PERF_SAMPLE_ID)
 131                size += sizeof(data->id);
 132
 133        if (sample_type & PERF_SAMPLE_STREAM_ID)
 134                size += sizeof(data->stream_id);
 135
 136        if (sample_type & PERF_SAMPLE_CPU)
 137                size += sizeof(data->cpu) * 2;
 138 out:
 139        session->id_hdr_size = size;
 140 }
 141
 142 void perf_session__set_sample_id_all(struct perf_session *session, bool value)
 143 {
 144         session->sample_id_all = value;
 145         perf_session__id_header_size(session);
 146 }
 147
 148 void perf_session__set_sample_type(struct perf_session *session, u64 type)
 149 {
 150         session->sample_type = type;
 151 }
 152
 153 void perf_session__update_sample_type(struct perf_session *self)
 154 {
 155         self->sample_type = perf_header__sample_type(&self->header);
 156         self->sample_id_all = perf_header__sample_id_all(&self->header);
 157         perf_session__id_header_size(self);
 158 }
 159
 160 int perf_session__create_kernel_maps(struct perf_session *self)
 161 {
 162         int ret = machine__create_kernel_maps(&self->host_machine);
 163
 164         if (ret >= 0)
 165                 ret = machines__create_guest_kernel_maps(&self->machines);
 166         return ret;
 167 }
 168
 169 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 170 {
 171         machine__destroy_kernel_maps(&self->host_machine);
 172         machines__destroy_guest_kernel_maps(&self->machines);
 173 }
 174
 175 struct perf_session *perf_session__new(const char *filename, int mode,
 176                                        bool force, bool repipe,
 177                                        struct perf_event_ops *ops)
 178 {
 179         size_t len = filename ? strlen(filename) + 1 : 0;
 180         struct perf_session *self = zalloc(sizeof(*self) + len);
 181
 182         if (self == NULL)
 183                 goto out;
 184
 185         if (perf_header__init(&self->header) < 0)
 186                 goto out_free;
 187
 188         memcpy(self->filename, filename, len);
 189         self->threads = RB_ROOT;
 190         INIT_LIST_HEAD(&self->dead_threads);
 191         self->last_match = NULL;
 192         /*
 193          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 194          * slices. On 32bit we use 32MB.
 195          */
 196 #if BITS_PER_LONG == 64
 197         self->mmap_window = ULLONG_MAX;
 198 #else
 199         self->mmap_window = 32 * 1024 * 1024ULL;
 200 #endif
 201         self->machines = RB_ROOT;
 202         self->repipe = repipe;
 203         INIT_LIST_HEAD(&self->ordered_samples.samples);
 204         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 205         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 206         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 207
 208         if (mode == O_RDONLY) {
 209                 if (perf_session__open(self, force) < 0)
 210                         goto out_delete;
 211         } else if (mode == O_WRONLY) {
 212                 /*
 213                  * In O_RDONLY mode this will be performed when reading the
 214                  * kernel MMAP event, in perf_event__process_mmap().
 215                  */
 216                 if (perf_session__create_kernel_maps(self) < 0)
 217                         goto out_delete;
 218         }
 219
 220         perf_session__update_sample_type(self);
 221
 222         if (ops && ops->ordering_requires_timestamps &&
 223             ops->ordered_samples && !self->sample_id_all) {
 224                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 225                 ops->ordered_samples = false;
 226         }
 227
 228 out:
 229         return self;
 230 out_free:
 231         free(self);
 232         return NULL;
 233 out_delete:
 234         perf_session__delete(self);
 235         return NULL;
 236 }
 237
 238 static void perf_session__delete_dead_threads(struct perf_session *self)
 239 {
 240         struct thread *n, *t;
 241
 242         list_for_each_entry_safe(t, n, &self->dead_threads, node) {
 243                 list_del(&t->node);
 244                 thread__delete(t);
 245         }
 246 }
 247
 248 static void perf_session__delete_threads(struct perf_session *self)
 249 {
 250         struct rb_node *nd = rb_first(&self->threads);
 251
 252         while (nd) {
 253                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 254
 255                 rb_erase(&t->rb_node, &self->threads);
 256                 nd = rb_next(nd);
 257                 thread__delete(t);
 258         }
 259 }
 260
 261 void perf_session__delete(struct perf_session *self)
 262 {
 263         perf_header__exit(&self->header);
 264         perf_session__destroy_kernel_maps(self);
 265         perf_session__delete_dead_threads(self);
 266         perf_session__delete_threads(self);
 267         machine__exit(&self->host_machine);
 268         close(self->fd);
 269         free(self);
 270 }
 271
 272 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
 273 {
 274         self->last_match = NULL;
 275         rb_erase(&th->rb_node, &self->threads);
 276         /*
 277          * We may have references to this thread, for instance in some hist_entry
 278          * instances, so just move them to a separate list.
 279          */
 280         list_add_tail(&th->node, &self->dead_threads);
 281 }
 282
 283 static bool symbol__match_parent_regex(struct symbol *sym)
 284 {
 285         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 286                 return 1;
 287
 288         return 0;
 289 }
 290
 291 int perf_session__resolve_callchain(struct perf_session *self,
 292                                     struct thread *thread,
 293                                     struct ip_callchain *chain,
 294                                     struct symbol **parent)
 295 {
 296         u8 cpumode = PERF_RECORD_MISC_USER;
 297         unsigned int i;
 298         int err;
 299
 300         callchain_cursor_reset(&self->callchain_cursor);
 301
 302         for (i = 0; i < chain->nr; i++) {
 303                 u64 ip = chain->ips[i];
 304                 struct addr_location al;
 305
 306                 if (ip >= PERF_CONTEXT_MAX) {
 307                         switch (ip) {
 308                         case PERF_CONTEXT_HV:
 309                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 310                         case PERF_CONTEXT_KERNEL:
 311                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 312                         case PERF_CONTEXT_USER:
 313                                 cpumode = PERF_RECORD_MISC_USER;        break;
 314                         default:
 315                                 break;
 316                         }
 317                         continue;
 318                 }
 319
 320                 al.filtered = false;
 321                 thread__find_addr_location(thread, self, cpumode,
 322                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 323                 if (al.sym != NULL) {
 324                         if (sort__has_parent && !*parent &&
 325                             symbol__match_parent_regex(al.sym))
 326                                 *parent = al.sym;
 327                         if (!symbol_conf.use_callchain)
 328                                 break;
 329                 }
 330
 331                 err = callchain_cursor_append(&self->callchain_cursor,
 332                                               ip, al.map, al.sym);
 333                 if (err)
 334                         return err;
 335         }
 336
 337         return 0;
 338 }
 339
 340 static int process_event_synth_stub(union perf_event *event __used,
 341                                     struct perf_session *session __used)
 342 {
 343         dump_printf(": unhandled!\n");
 344         return 0;
 345 }
 346
 347 static int process_event_stub(union perf_event *event __used,
 348                               struct perf_sample *sample __used,
 349                               struct perf_session *session __used)
 350 {
 351         dump_printf(": unhandled!\n");
 352         return 0;
 353 }
 354
 355 static int process_finished_round_stub(union perf_event *event __used,
 356                                        struct perf_session *session __used,
 357                                        struct perf_event_ops *ops __used)
 358 {
 359         dump_printf(": unhandled!\n");
 360         return 0;
 361 }
 362
 363 static int process_finished_round(union perf_event *event,
 364                                   struct perf_session *session,
 365                                   struct perf_event_ops *ops);
 366
 367 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 368 {
 369         if (handler->sample == NULL)
 370                 handler->sample = process_event_stub;
 371         if (handler->mmap == NULL)
 372                 handler->mmap = process_event_stub;
 373         if (handler->comm == NULL)
 374                 handler->comm = process_event_stub;
 375         if (handler->fork == NULL)
 376                 handler->fork = process_event_stub;
 377         if (handler->exit == NULL)
 378                 handler->exit = process_event_stub;
 379         if (handler->lost == NULL)
 380                 handler->lost = perf_event__process_lost;
 381         if (handler->read == NULL)
 382                 handler->read = process_event_stub;
 383         if (handler->throttle == NULL)
 384                 handler->throttle = process_event_stub;
 385         if (handler->unthrottle == NULL)
 386                 handler->unthrottle = process_event_stub;
 387         if (handler->attr == NULL)
 388                 handler->attr = process_event_synth_stub;
 389         if (handler->event_type == NULL)
 390                 handler->event_type = process_event_synth_stub;
 391         if (handler->tracing_data == NULL)
 392                 handler->tracing_data = process_event_synth_stub;
 393         if (handler->build_id == NULL)
 394                 handler->build_id = process_event_synth_stub;
 395         if (handler->finished_round == NULL) {
 396                 if (handler->ordered_samples)
 397                         handler->finished_round = process_finished_round;
 398                 else
 399                         handler->finished_round = process_finished_round_stub;
 400         }
 401 }
 402
 403 void mem_bswap_64(void *src, int byte_size)
 404 {
 405         u64 *m = src;
 406
 407         while (byte_size > 0) {
 408                 *m = bswap_64(*m);
 409                 byte_size -= sizeof(u64);
 410                 ++m;
 411         }
 412 }
 413
 414 static void perf_event__all64_swap(union perf_event *event)
 415 {
 416         struct perf_event_header *hdr = &event->header;
 417         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 418 }
 419
 420 static void perf_event__comm_swap(union perf_event *event)
 421 {
 422         event->comm.pid = bswap_32(event->comm.pid);
 423         event->comm.tid = bswap_32(event->comm.tid);
 424 }
 425
 426 static void perf_event__mmap_swap(union perf_event *event)
 427 {
 428         event->mmap.pid   = bswap_32(event->mmap.pid);
 429         event->mmap.tid   = bswap_32(event->mmap.tid);
 430         event->mmap.start = bswap_64(event->mmap.start);
 431         event->mmap.len   = bswap_64(event->mmap.len);
 432         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 433 }
 434
 435 static void perf_event__task_swap(union perf_event *event)
 436 {
 437         event->fork.pid  = bswap_32(event->fork.pid);
 438         event->fork.tid  = bswap_32(event->fork.tid);
 439         event->fork.ppid = bswap_32(event->fork.ppid);
 440         event->fork.ptid = bswap_32(event->fork.ptid);
 441         event->fork.time = bswap_64(event->fork.time);
 442 }
 443
 444 static void perf_event__read_swap(union perf_event *event)
 445 {
 446         event->read.pid          = bswap_32(event->read.pid);
 447         event->read.tid          = bswap_32(event->read.tid);
 448         event->read.value        = bswap_64(event->read.value);
 449         event->read.time_enabled = bswap_64(event->read.time_enabled);
 450         event->read.time_running = bswap_64(event->read.time_running);
 451         event->read.id           = bswap_64(event->read.id);
 452 }
 453
 454 static void perf_event__attr_swap(union perf_event *event)
 455 {
 456         size_t size;
 457
 458         event->attr.attr.type           = bswap_32(event->attr.attr.type);
 459         event->attr.attr.size           = bswap_32(event->attr.attr.size);
 460         event->attr.attr.config         = bswap_64(event->attr.attr.config);
 461         event->attr.attr.sample_period  = bswap_64(event->attr.attr.sample_period);
 462         event->attr.attr.sample_type    = bswap_64(event->attr.attr.sample_type);
 463         event->attr.attr.read_format    = bswap_64(event->attr.attr.read_format);
 464         event->attr.attr.wakeup_events  = bswap_32(event->attr.attr.wakeup_events);
 465         event->attr.attr.bp_type        = bswap_32(event->attr.attr.bp_type);
 466         event->attr.attr.bp_addr        = bswap_64(event->attr.attr.bp_addr);
 467         event->attr.attr.bp_len         = bswap_64(event->attr.attr.bp_len);
 468
 469         size = event->header.size;
 470         size -= (void *)&event->attr.id - (void *)event;
 471         mem_bswap_64(event->attr.id, size);
 472 }
 473
 474 static void perf_event__event_type_swap(union perf_event *event)
 475 {
 476         event->event_type.event_type.event_id =
 477                 bswap_64(event->event_type.event_type.event_id);
 478 }
 479
 480 static void perf_event__tracing_data_swap(union perf_event *event)
 481 {
 482         event->tracing_data.size = bswap_32(event->tracing_data.size);
 483 }
 484
 485 typedef void (*perf_event__swap_op)(union perf_event *event);
 486
 487 static perf_event__swap_op perf_event__swap_ops[] = {
 488         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 489         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 490         [PERF_RECORD_FORK]                = perf_event__task_swap,
 491         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 492         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 493         [PERF_RECORD_READ]                = perf_event__read_swap,
 494         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 495         [PERF_RECORD_HEADER_ATTR]         = perf_event__attr_swap,
 496         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 497         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 498         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 499         [PERF_RECORD_HEADER_MAX]          = NULL,
 500 };
 501
 502 struct sample_queue {
 503         u64                     timestamp;
 504         u64                     file_offset;
 505         union perf_event        *event;
 506         struct list_head        list;
 507 };
 508
 509 static void perf_session_free_sample_buffers(struct perf_session *session)
 510 {
 511         struct ordered_samples *os = &session->ordered_samples;
 512
 513         while (!list_empty(&os->to_free)) {
 514                 struct sample_queue *sq;
 515
 516                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 517                 list_del(&sq->list);
 518                 free(sq);
 519         }
 520 }
 521
 522 static int perf_session_deliver_event(struct perf_session *session,
 523                                       union perf_event *event,
 524                                       struct perf_sample *sample,
 525                                       struct perf_event_ops *ops,
 526                                       u64 file_offset);
 527
 528 static void flush_sample_queue(struct perf_session *s,
 529                                struct perf_event_ops *ops)
 530 {
 531         struct ordered_samples *os = &s->ordered_samples;
 532         struct list_head *head = &os->samples;
 533         struct sample_queue *tmp, *iter;
 534         struct perf_sample sample;
 535         u64 limit = os->next_flush;
 536         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 537
 538         if (!ops->ordered_samples || !limit)
 539                 return;
 540
 541         list_for_each_entry_safe(iter, tmp, head, list) {
 542                 if (iter->timestamp > limit)
 543                         break;
 544
 545                 perf_session__parse_sample(s, iter->event, &sample);
 546                 perf_session_deliver_event(s, iter->event, &sample, ops,
 547                                            iter->file_offset);
 548
 549                 os->last_flush = iter->timestamp;
 550                 list_del(&iter->list);
 551                 list_add(&iter->list, &os->sample_cache);
 552         }
 553
 554         if (list_empty(head)) {
 555                 os->last_sample = NULL;
 556         } else if (last_ts <= limit) {
 557                 os->last_sample =
 558                         list_entry(head->prev, struct sample_queue, list);
 559         }
 560 }
 561
 562 /*
 563  * When perf record finishes a pass on every buffers, it records this pseudo
 564  * event.
 565  * We record the max timestamp t found in the pass n.
 566  * Assuming these timestamps are monotonic across cpus, we know that if
 567  * a buffer still has events with timestamps below t, they will be all
 568  * available and then read in the pass n + 1.
 569  * Hence when we start to read the pass n + 2, we can safely flush every
 570  * events with timestamps below t.
 571  *
 572  *    ============ PASS n =================
 573  *       CPU 0         |   CPU 1
 574  *                     |
 575  *    cnt1 timestamps  |   cnt2 timestamps
 576  *          1          |         2
 577  *          2          |         3
 578  *          -          |         4  <--- max recorded
 579  *
 580  *    ============ PASS n + 1 ==============
 581  *       CPU 0         |   CPU 1
 582  *                     |
 583  *    cnt1 timestamps  |   cnt2 timestamps
 584  *          3          |         5
 585  *          4          |         6
 586  *          5          |         7 <---- max recorded
 587  *
 588  *      Flush every events below timestamp 4
 589  *
 590  *    ============ PASS n + 2 ==============
 591  *       CPU 0         |   CPU 1
 592  *                     |
 593  *    cnt1 timestamps  |   cnt2 timestamps
 594  *          6          |         8
 595  *          7          |         9
 596  *          -          |         10
 597  *
 598  *      Flush every events below timestamp 7
 599  *      etc...
 600  */
 601 static int process_finished_round(union perf_event *event __used,
 602                                   struct perf_session *session,
 603                                   struct perf_event_ops *ops)
 604 {
 605         flush_sample_queue(session, ops);
 606         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 607
 608         return 0;
 609 }
 610
 611 /* The queue is ordered by time */
 612 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 613 {
 614         struct ordered_samples *os = &s->ordered_samples;
 615         struct sample_queue *sample = os->last_sample;
 616         u64 timestamp = new->timestamp;
 617         struct list_head *p;
 618
 619         os->last_sample = new;
 620
 621         if (!sample) {
 622                 list_add(&new->list, &os->samples);
 623                 os->max_timestamp = timestamp;
 624                 return;
 625         }
 626
 627         /*
 628          * last_sample might point to some random place in the list as it's
 629          * the last queued event. We expect that the new event is close to
 630          * this.
 631          */
 632         if (sample->timestamp <= timestamp) {
 633                 while (sample->timestamp <= timestamp) {
 634                         p = sample->list.next;
 635                         if (p == &os->samples) {
 636                                 list_add_tail(&new->list, &os->samples);
 637                                 os->max_timestamp = timestamp;
 638                                 return;
 639                         }
 640                         sample = list_entry(p, struct sample_queue, list);
 641                 }
 642                 list_add_tail(&new->list, &sample->list);
 643         } else {
 644                 while (sample->timestamp > timestamp) {
 645                         p = sample->list.prev;
 646                         if (p == &os->samples) {
 647                                 list_add(&new->list, &os->samples);
 648                                 return;
 649                         }
 650                         sample = list_entry(p, struct sample_queue, list);
 651                 }
 652                 list_add(&new->list, &sample->list);
 653         }
 654 }
 655
 656 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 657
 658 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 659                                     struct perf_sample *sample, u64 file_offset)
 660 {
 661         struct ordered_samples *os = &s->ordered_samples;
 662         struct list_head *sc = &os->sample_cache;
 663         u64 timestamp = sample->time;
 664         struct sample_queue *new;
 665
 666         if (!timestamp || timestamp == ~0ULL)
 667                 return -ETIME;
 668
 669         if (timestamp < s->ordered_samples.last_flush) {
 670                 printf("Warning: Timestamp below last timeslice flush\n");
 671                 return -EINVAL;
 672         }
 673
 674         if (!list_empty(sc)) {
 675                 new = list_entry(sc->next, struct sample_queue, list);
 676                 list_del(&new->list);
 677         } else if (os->sample_buffer) {
 678                 new = os->sample_buffer + os->sample_buffer_idx;
 679                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 680                         os->sample_buffer = NULL;
 681         } else {
 682                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 683                 if (!os->sample_buffer)
 684                         return -ENOMEM;
 685                 list_add(&os->sample_buffer->list, &os->to_free);
 686                 os->sample_buffer_idx = 2;
 687                 new = os->sample_buffer + 1;
 688         }
 689
 690         new->timestamp = timestamp;
 691         new->file_offset = file_offset;
 692         new->event = event;
 693
 694         __queue_event(new, s);
 695
 696         return 0;
 697 }
 698
 699 static void callchain__printf(struct perf_sample *sample)
 700 {
 701         unsigned int i;
 702
 703         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 704
 705         for (i = 0; i < sample->callchain->nr; i++)
 706                 printf("..... %2d: %016" PRIx64 "\n",
 707                        i, sample->callchain->ips[i]);
 708 }
 709
 710 static void perf_session__print_tstamp(struct perf_session *session,
 711                                        union perf_event *event,
 712                                        struct perf_sample *sample)
 713 {
 714         if (event->header.type != PERF_RECORD_SAMPLE &&
 715             !session->sample_id_all) {
 716                 fputs("-1 -1 ", stdout);
 717                 return;
 718         }
 719
 720         if ((session->sample_type & PERF_SAMPLE_CPU))
 721                 printf("%u ", sample->cpu);
 722
 723         if (session->sample_type & PERF_SAMPLE_TIME)
 724                 printf("%" PRIu64 " ", sample->time);
 725 }
 726
 727 static void dump_event(struct perf_session *session, union perf_event *event,
 728                        u64 file_offset, struct perf_sample *sample)
 729 {
 730         if (!dump_trace)
 731                 return;
 732
 733         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 734                file_offset, event->header.size, event->header.type);
 735
 736         trace_event(event);
 737
 738         if (sample)
 739                 perf_session__print_tstamp(session, event, sample);
 740
 741         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 742                event->header.size, perf_event__name(event->header.type));
 743 }
 744
 745 static void dump_sample(struct perf_session *session, union perf_event *event,
 746                         struct perf_sample *sample)
 747 {
 748         if (!dump_trace)
 749                 return;
 750
 751         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 "\n",
 752                event->header.misc, sample->pid, sample->tid, sample->ip,
 753                sample->period);
 754
 755         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 756                 callchain__printf(sample);
 757 }
 758
 759 static int perf_session_deliver_event(struct perf_session *session,
 760                                       union perf_event *event,
 761                                       struct perf_sample *sample,
 762                                       struct perf_event_ops *ops,
 763                                       u64 file_offset)
 764 {
 765         dump_event(session, event, file_offset, sample);
 766
 767         switch (event->header.type) {
 768         case PERF_RECORD_SAMPLE:
 769                 dump_sample(session, event, sample);
 770                 return ops->sample(event, sample, session);
 771         case PERF_RECORD_MMAP:
 772                 return ops->mmap(event, sample, session);
 773         case PERF_RECORD_COMM:
 774                 return ops->comm(event, sample, session);
 775         case PERF_RECORD_FORK:
 776                 return ops->fork(event, sample, session);
 777         case PERF_RECORD_EXIT:
 778                 return ops->exit(event, sample, session);
 779         case PERF_RECORD_LOST:
 780                 return ops->lost(event, sample, session);
 781         case PERF_RECORD_READ:
 782                 return ops->read(event, sample, session);
 783         case PERF_RECORD_THROTTLE:
 784                 return ops->throttle(event, sample, session);
 785         case PERF_RECORD_UNTHROTTLE:
 786                 return ops->unthrottle(event, sample, session);
 787         default:
 788                 ++session->hists.stats.nr_unknown_events;
 789                 return -1;
 790         }
 791 }
 792
 793 static int perf_session__preprocess_sample(struct perf_session *session,
 794                                            union perf_event *event, struct perf_sample *sample)
 795 {
 796         if (event->header.type != PERF_RECORD_SAMPLE ||
 797             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 798                 return 0;
 799
 800         if (!ip_callchain__valid(sample->callchain, event)) {
 801                 pr_debug("call-chain problem with event, skipping it.\n");
 802                 ++session->hists.stats.nr_invalid_chains;
 803                 session->hists.stats.total_invalid_chains += sample->period;
 804                 return -EINVAL;
 805         }
 806         return 0;
 807 }
 808
 809 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
 810                                             struct perf_event_ops *ops, u64 file_offset)
 811 {
 812         dump_event(session, event, file_offset, NULL);
 813
 814         /* These events are processed right away */
 815         switch (event->header.type) {
 816         case PERF_RECORD_HEADER_ATTR:
 817                 return ops->attr(event, session);
 818         case PERF_RECORD_HEADER_EVENT_TYPE:
 819                 return ops->event_type(event, session);
 820         case PERF_RECORD_HEADER_TRACING_DATA:
 821                 /* setup for reading amidst mmap */
 822                 lseek(session->fd, file_offset, SEEK_SET);
 823                 return ops->tracing_data(event, session);
 824         case PERF_RECORD_HEADER_BUILD_ID:
 825                 return ops->build_id(event, session);
 826         case PERF_RECORD_FINISHED_ROUND:
 827                 return ops->finished_round(event, session, ops);
 828         default:
 829                 return -EINVAL;
 830         }
 831 }
 832
 833 static int perf_session__process_event(struct perf_session *session,
 834                                        union perf_event *event,
 835                                        struct perf_event_ops *ops,
 836                                        u64 file_offset)
 837 {
 838         struct perf_sample sample;
 839         int ret;
 840
 841         if (session->header.needs_swap &&
 842             perf_event__swap_ops[event->header.type])
 843                 perf_event__swap_ops[event->header.type](event);
 844
 845         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 846                 return -EINVAL;
 847
 848         hists__inc_nr_events(&session->hists, event->header.type);
 849
 850         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 851                 return perf_session__process_user_event(session, event, ops, file_offset);
 852
 853         /*
 854          * For all kernel events we get the sample data
 855          */
 856         perf_session__parse_sample(session, event, &sample);
 857
 858         /* Preprocess sample records - precheck callchains */
 859         if (perf_session__preprocess_sample(session, event, &sample))
 860                 return 0;
 861
 862         if (ops->ordered_samples) {
 863                 ret = perf_session_queue_event(session, event, &sample,
 864                                                file_offset);
 865                 if (ret != -ETIME)
 866                         return ret;
 867         }
 868
 869         return perf_session_deliver_event(session, event, &sample, ops,
 870                                           file_offset);
 871 }
 872
 873 void perf_event_header__bswap(struct perf_event_header *self)
 874 {
 875         self->type = bswap_32(self->type);
 876         self->misc = bswap_16(self->misc);
 877         self->size = bswap_16(self->size);
 878 }
 879
 880 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 881 {
 882         struct thread *thread = perf_session__findnew(self, 0);
 883
 884         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 885                 pr_err("problem inserting idle task.\n");
 886                 thread = NULL;
 887         }
 888
 889         return thread;
 890 }
 891
 892 static void perf_session__warn_about_errors(const struct perf_session *session,
 893                                             const struct perf_event_ops *ops)
 894 {
 895         if (ops->lost == perf_event__process_lost &&
 896             session->hists.stats.total_lost != 0) {
 897                 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64
 898                             "!\n\nCheck IO/CPU overload!\n\n",
 899                             session->hists.stats.total_period,
 900                             session->hists.stats.total_lost);
 901         }
 902
 903         if (session->hists.stats.nr_unknown_events != 0) {
 904                 ui__warning("Found %u unknown events!\n\n"
 905                             "Is this an older tool processing a perf.data "
 906                             "file generated by a more recent tool?\n\n"
 907                             "If that is not the case, consider "
 908                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 909                             session->hists.stats.nr_unknown_events);
 910         }
 911
 912         if (session->hists.stats.nr_invalid_chains != 0) {
 913                 ui__warning("Found invalid callchains!\n\n"
 914                             "%u out of %u events were discarded for this reason.\n\n"
 915                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 916                             session->hists.stats.nr_invalid_chains,
 917                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 918         }
 919 }
 920
 921 #define session_done()  (*(volatile int *)(&session_done))
 922 volatile int session_done;
 923
 924 static int __perf_session__process_pipe_events(struct perf_session *self,
 925                                                struct perf_event_ops *ops)
 926 {
 927         union perf_event event;
 928         uint32_t size;
 929         int skip = 0;
 930         u64 head;
 931         int err;
 932         void *p;
 933
 934         perf_event_ops__fill_defaults(ops);
 935
 936         head = 0;
 937 more:
 938         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 939         if (err <= 0) {
 940                 if (err == 0)
 941                         goto done;
 942
 943                 pr_err("failed to read event header\n");
 944                 goto out_err;
 945         }
 946
 947         if (self->header.needs_swap)
 948                 perf_event_header__bswap(&event.header);
 949
 950         size = event.header.size;
 951         if (size == 0)
 952                 size = 8;
 953
 954         p = &event;
 955         p += sizeof(struct perf_event_header);
 956
 957         if (size - sizeof(struct perf_event_header)) {
 958                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 959                 if (err <= 0) {
 960                         if (err == 0) {
 961                                 pr_err("unexpected end of event stream\n");
 962                                 goto done;
 963                         }
 964
 965                         pr_err("failed to read event data\n");
 966                         goto out_err;
 967                 }
 968         }
 969
 970         if (size == 0 ||
 971             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 972                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 973                             head, event.header.size, event.header.type);
 974                 /*
 975                  * assume we lost track of the stream, check alignment, and
 976                  * increment a single u64 in the hope to catch on again 'soon'.
 977                  */
 978                 if (unlikely(head & 7))
 979                         head &= ~7ULL;
 980
 981                 size = 8;
 982         }
 983
 984         head += size;
 985
 986         if (skip > 0)
 987                 head += skip;
 988
 989         if (!session_done())
 990                 goto more;
 991 done:
 992         err = 0;
 993 out_err:
 994         perf_session__warn_about_errors(self, ops);
 995         perf_session_free_sample_buffers(self);
 996         return err;
 997 }
 998
 999 int __perf_session__process_events(struct perf_session *session,
1000                                    u64 data_offset, u64 data_size,
1001                                    u64 file_size, struct perf_event_ops *ops)
1002 {
1003         u64 head, page_offset, file_offset, file_pos, progress_next;
1004         int err, mmap_prot, mmap_flags, map_idx = 0;
1005         struct ui_progress *progress;
1006         size_t  page_size, mmap_size;
1007         char *buf, *mmaps[8];
1008         union perf_event *event;
1009         uint32_t size;
1010
1011         perf_event_ops__fill_defaults(ops);
1012
1013         page_size = sysconf(_SC_PAGESIZE);
1014
1015         page_offset = page_size * (data_offset / page_size);
1016         file_offset = page_offset;
1017         head = data_offset - page_offset;
1018
1019         if (data_offset + data_size < file_size)
1020                 file_size = data_offset + data_size;
1021
1022         progress_next = file_size / 16;
1023         progress = ui_progress__new("Processing events...", file_size);
1024         if (progress == NULL)
1025                 return -1;
1026
1027         mmap_size = session->mmap_window;
1028         if (mmap_size > file_size)
1029                 mmap_size = file_size;
1030
1031         memset(mmaps, 0, sizeof(mmaps));
1032
1033         mmap_prot  = PROT_READ;
1034         mmap_flags = MAP_SHARED;
1035
1036         if (session->header.needs_swap) {
1037                 mmap_prot  |= PROT_WRITE;
1038                 mmap_flags = MAP_PRIVATE;
1039         }
1040 remap:
1041         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1042                    file_offset);
1043         if (buf == MAP_FAILED) {
1044                 pr_err("failed to mmap file\n");
1045                 err = -errno;
1046                 goto out_err;
1047         }
1048         mmaps[map_idx] = buf;
1049         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1050         file_pos = file_offset + head;
1051
1052 more:
1053         event = (union perf_event *)(buf + head);
1054
1055         if (session->header.needs_swap)
1056                 perf_event_header__bswap(&event->header);
1057         size = event->header.size;
1058         if (size == 0)
1059                 size = 8;
1060
1061         if (head + event->header.size > mmap_size) {
1062                 if (mmaps[map_idx]) {
1063                         munmap(mmaps[map_idx], mmap_size);
1064                         mmaps[map_idx] = NULL;
1065                 }
1066
1067                 page_offset = page_size * (head / page_size);
1068                 file_offset += page_offset;
1069                 head -= page_offset;
1070                 goto remap;
1071         }
1072
1073         size = event->header.size;
1074
1075         if (size == 0 ||
1076             perf_session__process_event(session, event, ops, file_pos) < 0) {
1077                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1078                             file_offset + head, event->header.size,
1079                             event->header.type);
1080                 /*
1081                  * assume we lost track of the stream, check alignment, and
1082                  * increment a single u64 in the hope to catch on again 'soon'.
1083                  */
1084                 if (unlikely(head & 7))
1085                         head &= ~7ULL;
1086
1087                 size = 8;
1088         }
1089
1090         head += size;
1091         file_pos += size;
1092
1093         if (file_pos >= progress_next) {
1094                 progress_next += file_size / 16;
1095                 ui_progress__update(progress, file_pos);
1096         }
1097
1098         if (file_pos < file_size)
1099                 goto more;
1100
1101         err = 0;
1102         /* do the final flush for ordered samples */
1103         session->ordered_samples.next_flush = ULLONG_MAX;
1104         flush_sample_queue(session, ops);
1105 out_err:
1106         ui_progress__delete(progress);
1107         perf_session__warn_about_errors(session, ops);
1108         perf_session_free_sample_buffers(session);
1109         return err;
1110 }
1111
1112 int perf_session__process_events(struct perf_session *self,
1113                                  struct perf_event_ops *ops)
1114 {
1115         int err;
1116
1117         if (perf_session__register_idle_thread(self) == NULL)
1118                 return -ENOMEM;
1119
1120         if (!self->fd_pipe)
1121                 err = __perf_session__process_events(self,
1122                                                      self->header.data_offset,
1123                                                      self->header.data_size,
1124                                                      self->size, ops);
1125         else
1126                 err = __perf_session__process_pipe_events(self, ops);
1127
1128         return err;
1129 }
1130
1131 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1132 {
1133         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1134                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1135                 return false;
1136         }
1137
1138         return true;
1139 }
1140
1141 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1142                                              const char *symbol_name,
1143                                              u64 addr)
1144 {
1145         char *bracket;
1146         enum map_type i;
1147         struct ref_reloc_sym *ref;
1148
1149         ref = zalloc(sizeof(struct ref_reloc_sym));
1150         if (ref == NULL)
1151                 return -ENOMEM;
1152
1153         ref->name = strdup(symbol_name);
1154         if (ref->name == NULL) {
1155                 free(ref);
1156                 return -ENOMEM;
1157         }
1158
1159         bracket = strchr(ref->name, ']');
1160         if (bracket)
1161                 *bracket = '\0';
1162
1163         ref->addr = addr;
1164
1165         for (i = 0; i < MAP__NR_TYPES; ++i) {
1166                 struct kmap *kmap = map__kmap(maps[i]);
1167                 kmap->ref_reloc_sym = ref;
1168         }
1169
1170         return 0;
1171 }
1172
1173 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1174 {
1175         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1176                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1177                machines__fprintf_dsos(&self->machines, fp);
1178 }
1179
1180 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1181                                           bool with_hits)
1182 {
1183         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1184         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1185 }
1186
1187 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1188 {
1189         struct perf_evsel *pos;
1190         size_t ret = fprintf(fp, "Aggregated stats:\n");
1191
1192         ret += hists__fprintf_nr_events(&session->hists, fp);
1193
1194         list_for_each_entry(pos, &session->evlist->entries, node) {
1195                 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1196                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1197         }
1198
1199         return ret;
1200 }