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