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 "session.h"
  11 #include "sort.h"
  12 #include "util.h"
  13
  14 static int perf_session__open(struct perf_session *self, bool force)
  15 {
  16         struct stat input_stat;
  17
  18         if (!strcmp(self->filename, "-")) {
  19                 self->fd_pipe = true;
  20                 self->fd = STDIN_FILENO;
  21
  22                 if (perf_header__read(self, self->fd) < 0)
  23                         pr_err("incompatible file format");
  24
  25                 return 0;
  26         }
  27
  28         self->fd = open(self->filename, O_RDONLY);
  29         if (self->fd < 0) {
  30                 int err = errno;
  31
  32                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  33                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  34                         pr_err("  (try 'perf record' first)");
  35                 pr_err("\n");
  36                 return -errno;
  37         }
  38
  39         if (fstat(self->fd, &input_stat) < 0)
  40                 goto out_close;
  41
  42         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  43                 pr_err("file %s not owned by current user or root\n",
  44                        self->filename);
  45                 goto out_close;
  46         }
  47
  48         if (!input_stat.st_size) {
  49                 pr_info("zero-sized file (%s), nothing to do!\n",
  50                         self->filename);
  51                 goto out_close;
  52         }
  53
  54         if (perf_header__read(self, self->fd) < 0) {
  55                 pr_err("incompatible file format");
  56                 goto out_close;
  57         }
  58
  59         self->size = input_stat.st_size;
  60         return 0;
  61
  62 out_close:
  63         close(self->fd);
  64         self->fd = -1;
  65         return -1;
  66 }
  67
  68 static void perf_session__id_header_size(struct perf_session *session)
  69 {
  70        struct sample_data *data;
  71        u64 sample_type = session->sample_type;
  72        u16 size = 0;
  73
  74         if (!session->sample_id_all)
  75                 goto out;
  76
  77        if (sample_type & PERF_SAMPLE_TID)
  78                size += sizeof(data->tid) * 2;
  79
  80        if (sample_type & PERF_SAMPLE_TIME)
  81                size += sizeof(data->time);
  82
  83        if (sample_type & PERF_SAMPLE_ID)
  84                size += sizeof(data->id);
  85
  86        if (sample_type & PERF_SAMPLE_STREAM_ID)
  87                size += sizeof(data->stream_id);
  88
  89        if (sample_type & PERF_SAMPLE_CPU)
  90                size += sizeof(data->cpu) * 2;
  91 out:
  92        session->id_hdr_size = size;
  93 }
  94
  95 void perf_session__set_sample_id_all(struct perf_session *session, bool value)
  96 {
  97         session->sample_id_all = value;
  98         perf_session__id_header_size(session);
  99 }
 100
 101 void perf_session__set_sample_type(struct perf_session *session, u64 type)
 102 {
 103         session->sample_type = type;
 104 }
 105
 106 void perf_session__update_sample_type(struct perf_session *self)
 107 {
 108         self->sample_type = perf_header__sample_type(&self->header);
 109         self->sample_id_all = perf_header__sample_id_all(&self->header);
 110         perf_session__id_header_size(self);
 111 }
 112
 113 int perf_session__create_kernel_maps(struct perf_session *self)
 114 {
 115         int ret = machine__create_kernel_maps(&self->host_machine);
 116
 117         if (ret >= 0)
 118                 ret = machines__create_guest_kernel_maps(&self->machines);
 119         return ret;
 120 }
 121
 122 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 123 {
 124         machine__destroy_kernel_maps(&self->host_machine);
 125         machines__destroy_guest_kernel_maps(&self->machines);
 126 }
 127
 128 struct perf_session *perf_session__new(const char *filename, int mode, bool force, bool repipe)
 129 {
 130         size_t len = filename ? strlen(filename) + 1 : 0;
 131         struct perf_session *self = zalloc(sizeof(*self) + len);
 132
 133         if (self == NULL)
 134                 goto out;
 135
 136         if (perf_header__init(&self->header) < 0)
 137                 goto out_free;
 138
 139         memcpy(self->filename, filename, len);
 140         self->threads = RB_ROOT;
 141         INIT_LIST_HEAD(&self->dead_threads);
 142         self->hists_tree = RB_ROOT;
 143         self->last_match = NULL;
 144         /*
 145          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 146          * slices. On 32bit we use 32MB.
 147          */
 148 #if BITS_PER_LONG == 64
 149         self->mmap_window = ULLONG_MAX;
 150 #else
 151         self->mmap_window = 32 * 1024 * 1024ULL;
 152 #endif
 153         self->machines = RB_ROOT;
 154         self->repipe = repipe;
 155         INIT_LIST_HEAD(&self->ordered_samples.samples);
 156         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 157         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 158         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 159
 160         if (mode == O_RDONLY) {
 161                 if (perf_session__open(self, force) < 0)
 162                         goto out_delete;
 163         } else if (mode == O_WRONLY) {
 164                 /*
 165                  * In O_RDONLY mode this will be performed when reading the
 166                  * kernel MMAP event, in event__process_mmap().
 167                  */
 168                 if (perf_session__create_kernel_maps(self) < 0)
 169                         goto out_delete;
 170         }
 171
 172         perf_session__update_sample_type(self);
 173 out:
 174         return self;
 175 out_free:
 176         free(self);
 177         return NULL;
 178 out_delete:
 179         perf_session__delete(self);
 180         return NULL;
 181 }
 182
 183 static void perf_session__delete_dead_threads(struct perf_session *self)
 184 {
 185         struct thread *n, *t;
 186
 187         list_for_each_entry_safe(t, n, &self->dead_threads, node) {
 188                 list_del(&t->node);
 189                 thread__delete(t);
 190         }
 191 }
 192
 193 static void perf_session__delete_threads(struct perf_session *self)
 194 {
 195         struct rb_node *nd = rb_first(&self->threads);
 196
 197         while (nd) {
 198                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 199
 200                 rb_erase(&t->rb_node, &self->threads);
 201                 nd = rb_next(nd);
 202                 thread__delete(t);
 203         }
 204 }
 205
 206 void perf_session__delete(struct perf_session *self)
 207 {
 208         perf_header__exit(&self->header);
 209         perf_session__destroy_kernel_maps(self);
 210         perf_session__delete_dead_threads(self);
 211         perf_session__delete_threads(self);
 212         machine__exit(&self->host_machine);
 213         close(self->fd);
 214         free(self);
 215 }
 216
 217 void perf_session__remove_thread(struct perf_session *self, struct thread *th)
 218 {
 219         self->last_match = NULL;
 220         rb_erase(&th->rb_node, &self->threads);
 221         /*
 222          * We may have references to this thread, for instance in some hist_entry
 223          * instances, so just move them to a separate list.
 224          */
 225         list_add_tail(&th->node, &self->dead_threads);
 226 }
 227
 228 static bool symbol__match_parent_regex(struct symbol *sym)
 229 {
 230         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 231                 return 1;
 232
 233         return 0;
 234 }
 235
 236 struct map_symbol *perf_session__resolve_callchain(struct perf_session *self,
 237                                                    struct thread *thread,
 238                                                    struct ip_callchain *chain,
 239                                                    struct symbol **parent)
 240 {
 241         u8 cpumode = PERF_RECORD_MISC_USER;
 242         unsigned int i;
 243         struct map_symbol *syms = calloc(chain->nr, sizeof(*syms));
 244
 245         if (!syms)
 246                 return NULL;
 247
 248         for (i = 0; i < chain->nr; i++) {
 249                 u64 ip = chain->ips[i];
 250                 struct addr_location al;
 251
 252                 if (ip >= PERF_CONTEXT_MAX) {
 253                         switch (ip) {
 254                         case PERF_CONTEXT_HV:
 255                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 256                         case PERF_CONTEXT_KERNEL:
 257                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 258                         case PERF_CONTEXT_USER:
 259                                 cpumode = PERF_RECORD_MISC_USER;        break;
 260                         default:
 261                                 break;
 262                         }
 263                         continue;
 264                 }
 265
 266                 al.filtered = false;
 267                 thread__find_addr_location(thread, self, cpumode,
 268                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 269                 if (al.sym != NULL) {
 270                         if (sort__has_parent && !*parent &&
 271                             symbol__match_parent_regex(al.sym))
 272                                 *parent = al.sym;
 273                         if (!symbol_conf.use_callchain)
 274                                 break;
 275                         syms[i].map = al.map;
 276                         syms[i].sym = al.sym;
 277                 }
 278         }
 279
 280         return syms;
 281 }
 282
 283 static int process_event_synth_stub(event_t *event __used,
 284                                     struct perf_session *session __used)
 285 {
 286         dump_printf(": unhandled!\n");
 287         return 0;
 288 }
 289
 290 static int process_event_stub(event_t *event __used,
 291                               struct sample_data *sample __used,
 292                               struct perf_session *session __used)
 293 {
 294         dump_printf(": unhandled!\n");
 295         return 0;
 296 }
 297
 298 static int process_finished_round_stub(event_t *event __used,
 299                                        struct perf_session *session __used,
 300                                        struct perf_event_ops *ops __used)
 301 {
 302         dump_printf(": unhandled!\n");
 303         return 0;
 304 }
 305
 306 static int process_finished_round(event_t *event,
 307                                   struct perf_session *session,
 308                                   struct perf_event_ops *ops);
 309
 310 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 311 {
 312         if (handler->sample == NULL)
 313                 handler->sample = process_event_stub;
 314         if (handler->mmap == NULL)
 315                 handler->mmap = process_event_stub;
 316         if (handler->comm == NULL)
 317                 handler->comm = process_event_stub;
 318         if (handler->fork == NULL)
 319                 handler->fork = process_event_stub;
 320         if (handler->exit == NULL)
 321                 handler->exit = process_event_stub;
 322         if (handler->lost == NULL)
 323                 handler->lost = event__process_lost;
 324         if (handler->read == NULL)
 325                 handler->read = process_event_stub;
 326         if (handler->throttle == NULL)
 327                 handler->throttle = process_event_stub;
 328         if (handler->unthrottle == NULL)
 329                 handler->unthrottle = process_event_stub;
 330         if (handler->attr == NULL)
 331                 handler->attr = process_event_synth_stub;
 332         if (handler->event_type == NULL)
 333                 handler->event_type = process_event_synth_stub;
 334         if (handler->tracing_data == NULL)
 335                 handler->tracing_data = process_event_synth_stub;
 336         if (handler->build_id == NULL)
 337                 handler->build_id = process_event_synth_stub;
 338         if (handler->finished_round == NULL) {
 339                 if (handler->ordered_samples)
 340                         handler->finished_round = process_finished_round;
 341                 else
 342                         handler->finished_round = process_finished_round_stub;
 343         }
 344 }
 345
 346 void mem_bswap_64(void *src, int byte_size)
 347 {
 348         u64 *m = src;
 349
 350         while (byte_size > 0) {
 351                 *m = bswap_64(*m);
 352                 byte_size -= sizeof(u64);
 353                 ++m;
 354         }
 355 }
 356
 357 static void event__all64_swap(event_t *self)
 358 {
 359         struct perf_event_header *hdr = &self->header;
 360         mem_bswap_64(hdr + 1, self->header.size - sizeof(*hdr));
 361 }
 362
 363 static void event__comm_swap(event_t *self)
 364 {
 365         self->comm.pid = bswap_32(self->comm.pid);
 366         self->comm.tid = bswap_32(self->comm.tid);
 367 }
 368
 369 static void event__mmap_swap(event_t *self)
 370 {
 371         self->mmap.pid   = bswap_32(self->mmap.pid);
 372         self->mmap.tid   = bswap_32(self->mmap.tid);
 373         self->mmap.start = bswap_64(self->mmap.start);
 374         self->mmap.len   = bswap_64(self->mmap.len);
 375         self->mmap.pgoff = bswap_64(self->mmap.pgoff);
 376 }
 377
 378 static void event__task_swap(event_t *self)
 379 {
 380         self->fork.pid  = bswap_32(self->fork.pid);
 381         self->fork.tid  = bswap_32(self->fork.tid);
 382         self->fork.ppid = bswap_32(self->fork.ppid);
 383         self->fork.ptid = bswap_32(self->fork.ptid);
 384         self->fork.time = bswap_64(self->fork.time);
 385 }
 386
 387 static void event__read_swap(event_t *self)
 388 {
 389         self->read.pid          = bswap_32(self->read.pid);
 390         self->read.tid          = bswap_32(self->read.tid);
 391         self->read.value        = bswap_64(self->read.value);
 392         self->read.time_enabled = bswap_64(self->read.time_enabled);
 393         self->read.time_running = bswap_64(self->read.time_running);
 394         self->read.id           = bswap_64(self->read.id);
 395 }
 396
 397 static void event__attr_swap(event_t *self)
 398 {
 399         size_t size;
 400
 401         self->attr.attr.type            = bswap_32(self->attr.attr.type);
 402         self->attr.attr.size            = bswap_32(self->attr.attr.size);
 403         self->attr.attr.config          = bswap_64(self->attr.attr.config);
 404         self->attr.attr.sample_period   = bswap_64(self->attr.attr.sample_period);
 405         self->attr.attr.sample_type     = bswap_64(self->attr.attr.sample_type);
 406         self->attr.attr.read_format     = bswap_64(self->attr.attr.read_format);
 407         self->attr.attr.wakeup_events   = bswap_32(self->attr.attr.wakeup_events);
 408         self->attr.attr.bp_type         = bswap_32(self->attr.attr.bp_type);
 409         self->attr.attr.bp_addr         = bswap_64(self->attr.attr.bp_addr);
 410         self->attr.attr.bp_len          = bswap_64(self->attr.attr.bp_len);
 411
 412         size = self->header.size;
 413         size -= (void *)&self->attr.id - (void *)self;
 414         mem_bswap_64(self->attr.id, size);
 415 }
 416
 417 static void event__event_type_swap(event_t *self)
 418 {
 419         self->event_type.event_type.event_id =
 420                 bswap_64(self->event_type.event_type.event_id);
 421 }
 422
 423 static void event__tracing_data_swap(event_t *self)
 424 {
 425         self->tracing_data.size = bswap_32(self->tracing_data.size);
 426 }
 427
 428 typedef void (*event__swap_op)(event_t *self);
 429
 430 static event__swap_op event__swap_ops[] = {
 431         [PERF_RECORD_MMAP]   = event__mmap_swap,
 432         [PERF_RECORD_COMM]   = event__comm_swap,
 433         [PERF_RECORD_FORK]   = event__task_swap,
 434         [PERF_RECORD_EXIT]   = event__task_swap,
 435         [PERF_RECORD_LOST]   = event__all64_swap,
 436         [PERF_RECORD_READ]   = event__read_swap,
 437         [PERF_RECORD_SAMPLE] = event__all64_swap,
 438         [PERF_RECORD_HEADER_ATTR]   = event__attr_swap,
 439         [PERF_RECORD_HEADER_EVENT_TYPE]   = event__event_type_swap,
 440         [PERF_RECORD_HEADER_TRACING_DATA]   = event__tracing_data_swap,
 441         [PERF_RECORD_HEADER_BUILD_ID]   = NULL,
 442         [PERF_RECORD_HEADER_MAX]    = NULL,
 443 };
 444
 445 struct sample_queue {
 446         u64                     timestamp;
 447         u64                     file_offset;
 448         event_t                 *event;
 449         struct list_head        list;
 450 };
 451
 452 static void perf_session_free_sample_buffers(struct perf_session *session)
 453 {
 454         struct ordered_samples *os = &session->ordered_samples;
 455
 456         while (!list_empty(&os->to_free)) {
 457                 struct sample_queue *sq;
 458
 459                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 460                 list_del(&sq->list);
 461                 free(sq);
 462         }
 463 }
 464
 465 static int perf_session_deliver_event(struct perf_session *session,
 466                                       event_t *event,
 467                                       struct sample_data *sample,
 468                                       struct perf_event_ops *ops,
 469                                       u64 file_offset);
 470
 471 static void flush_sample_queue(struct perf_session *s,
 472                                struct perf_event_ops *ops)
 473 {
 474         struct ordered_samples *os = &s->ordered_samples;
 475         struct list_head *head = &os->samples;
 476         struct sample_queue *tmp, *iter;
 477         struct sample_data sample;
 478         u64 limit = os->next_flush;
 479         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 480
 481         if (!ops->ordered_samples || !limit)
 482                 return;
 483
 484         list_for_each_entry_safe(iter, tmp, head, list) {
 485                 if (iter->timestamp > limit)
 486                         break;
 487
 488                 event__parse_sample(iter->event, s, &sample);
 489                 perf_session_deliver_event(s, iter->event, &sample, ops,
 490                                            iter->file_offset);
 491
 492                 os->last_flush = iter->timestamp;
 493                 list_del(&iter->list);
 494                 list_add(&iter->list, &os->sample_cache);
 495         }
 496
 497         if (list_empty(head)) {
 498                 os->last_sample = NULL;
 499         } else if (last_ts <= limit) {
 500                 os->last_sample =
 501                         list_entry(head->prev, struct sample_queue, list);
 502         }
 503 }
 504
 505 /*
 506  * When perf record finishes a pass on every buffers, it records this pseudo
 507  * event.
 508  * We record the max timestamp t found in the pass n.
 509  * Assuming these timestamps are monotonic across cpus, we know that if
 510  * a buffer still has events with timestamps below t, they will be all
 511  * available and then read in the pass n + 1.
 512  * Hence when we start to read the pass n + 2, we can safely flush every
 513  * events with timestamps below t.
 514  *
 515  *    ============ PASS n =================
 516  *       CPU 0         |   CPU 1
 517  *                     |
 518  *    cnt1 timestamps  |   cnt2 timestamps
 519  *          1          |         2
 520  *          2          |         3
 521  *          -          |         4  <--- max recorded
 522  *
 523  *    ============ PASS n + 1 ==============
 524  *       CPU 0         |   CPU 1
 525  *                     |
 526  *    cnt1 timestamps  |   cnt2 timestamps
 527  *          3          |         5
 528  *          4          |         6
 529  *          5          |         7 <---- max recorded
 530  *
 531  *      Flush every events below timestamp 4
 532  *
 533  *    ============ PASS n + 2 ==============
 534  *       CPU 0         |   CPU 1
 535  *                     |
 536  *    cnt1 timestamps  |   cnt2 timestamps
 537  *          6          |         8
 538  *          7          |         9
 539  *          -          |         10
 540  *
 541  *      Flush every events below timestamp 7
 542  *      etc...
 543  */
 544 static int process_finished_round(event_t *event __used,
 545                                   struct perf_session *session,
 546                                   struct perf_event_ops *ops)
 547 {
 548         flush_sample_queue(session, ops);
 549         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 550
 551         return 0;
 552 }
 553
 554 /* The queue is ordered by time */
 555 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 556 {
 557         struct ordered_samples *os = &s->ordered_samples;
 558         struct sample_queue *sample = os->last_sample;
 559         u64 timestamp = new->timestamp;
 560         struct list_head *p;
 561
 562         os->last_sample = new;
 563
 564         if (!sample) {
 565                 list_add(&new->list, &os->samples);
 566                 os->max_timestamp = timestamp;
 567                 return;
 568         }
 569
 570         /*
 571          * last_sample might point to some random place in the list as it's
 572          * the last queued event. We expect that the new event is close to
 573          * this.
 574          */
 575         if (sample->timestamp <= timestamp) {
 576                 while (sample->timestamp <= timestamp) {
 577                         p = sample->list.next;
 578                         if (p == &os->samples) {
 579                                 list_add_tail(&new->list, &os->samples);
 580                                 os->max_timestamp = timestamp;
 581                                 return;
 582                         }
 583                         sample = list_entry(p, struct sample_queue, list);
 584                 }
 585                 list_add_tail(&new->list, &sample->list);
 586         } else {
 587                 while (sample->timestamp > timestamp) {
 588                         p = sample->list.prev;
 589                         if (p == &os->samples) {
 590                                 list_add(&new->list, &os->samples);
 591                                 return;
 592                         }
 593                         sample = list_entry(p, struct sample_queue, list);
 594                 }
 595                 list_add(&new->list, &sample->list);
 596         }
 597 }
 598
 599 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 600
 601 static int perf_session_queue_event(struct perf_session *s, event_t *event,
 602                                     struct sample_data *data, u64 file_offset)
 603 {
 604         struct ordered_samples *os = &s->ordered_samples;
 605         struct list_head *sc = &os->sample_cache;
 606         u64 timestamp = data->time;
 607         struct sample_queue *new;
 608
 609         if (!timestamp || timestamp == ~0ULL)
 610                 return -ETIME;
 611
 612         if (timestamp < s->ordered_samples.last_flush) {
 613                 printf("Warning: Timestamp below last timeslice flush\n");
 614                 return -EINVAL;
 615         }
 616
 617         if (!list_empty(sc)) {
 618                 new = list_entry(sc->next, struct sample_queue, list);
 619                 list_del(&new->list);
 620         } else if (os->sample_buffer) {
 621                 new = os->sample_buffer + os->sample_buffer_idx;
 622                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 623                         os->sample_buffer = NULL;
 624         } else {
 625                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 626                 if (!os->sample_buffer)
 627                         return -ENOMEM;
 628                 list_add(&os->sample_buffer->list, &os->to_free);
 629                 os->sample_buffer_idx = 2;
 630                 new = os->sample_buffer + 1;
 631         }
 632
 633         new->timestamp = timestamp;
 634         new->file_offset = file_offset;
 635         new->event = event;
 636
 637         __queue_event(new, s);
 638
 639         return 0;
 640 }
 641
 642 static void callchain__dump(struct sample_data *sample)
 643 {
 644         unsigned int i;
 645
 646         if (!dump_trace)
 647                 return;
 648
 649         printf("... chain: nr:%Lu\n", sample->callchain->nr);
 650
 651         for (i = 0; i < sample->callchain->nr; i++)
 652                 printf("..... %2d: %016Lx\n", i, sample->callchain->ips[i]);
 653 }
 654
 655 static void perf_session__print_tstamp(struct perf_session *session,
 656                                        event_t *event,
 657                                        struct sample_data *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("%Lu ", sample->time);
 670 }
 671
 672 static void dump_event(struct perf_session *session, event_t *event,
 673                        u64 file_offset, struct sample_data *sample)
 674 {
 675         if (!dump_trace)
 676                 return;
 677
 678         dump_printf("\n%#Lx [%#x]: event: %d\n", file_offset,
 679                     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         dump_printf("%#Lx [%#x]: PERF_RECORD_%s",
 687                     file_offset, event->header.size,
 688                     event__get_event_name(event->header.type));
 689 }
 690
 691 static void dump_sample(struct perf_session *session, event_t *event,
 692                         struct sample_data *sample)
 693 {
 694         dump_printf("(IP, %d): %d/%d: %#Lx period: %Ld\n", event->header.misc,
 695                     sample->pid, sample->tid, sample->ip, sample->period);
 696
 697         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 698                 callchain__dump(sample);
 699 }
 700
 701 static int perf_session_deliver_event(struct perf_session *session,
 702                                       event_t *event,
 703                                       struct sample_data *sample,
 704                                       struct perf_event_ops *ops,
 705                                       u64 file_offset)
 706 {
 707         dump_event(session, event, file_offset, sample);
 708
 709         switch (event->header.type) {
 710         case PERF_RECORD_SAMPLE:
 711                 dump_sample(session, event, sample);
 712                 return ops->sample(event, sample, session);
 713         case PERF_RECORD_MMAP:
 714                 return ops->mmap(event, sample, session);
 715         case PERF_RECORD_COMM:
 716                 return ops->comm(event, sample, session);
 717         case PERF_RECORD_FORK:
 718                 return ops->fork(event, sample, session);
 719         case PERF_RECORD_EXIT:
 720                 return ops->exit(event, sample, session);
 721         case PERF_RECORD_LOST:
 722                 return ops->lost(event, sample, session);
 723         case PERF_RECORD_READ:
 724                 return ops->read(event, sample, session);
 725         case PERF_RECORD_THROTTLE:
 726                 return ops->throttle(event, sample, session);
 727         case PERF_RECORD_UNTHROTTLE:
 728                 return ops->unthrottle(event, sample, session);
 729         default:
 730                 ++session->hists.stats.nr_unknown_events;
 731                 return -1;
 732         }
 733 }
 734
 735 static int perf_session__preprocess_sample(struct perf_session *session,
 736                                            event_t *event, struct sample_data *sample)
 737 {
 738         if (event->header.type != PERF_RECORD_SAMPLE ||
 739             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 740                 return 0;
 741
 742         if (!ip_callchain__valid(sample->callchain, event)) {
 743                 pr_debug("call-chain problem with event, skipping it.\n");
 744                 ++session->hists.stats.nr_invalid_chains;
 745                 session->hists.stats.total_invalid_chains += sample->period;
 746                 return -EINVAL;
 747         }
 748         return 0;
 749 }
 750
 751 static int perf_session__process_event(struct perf_session *session,
 752                                        event_t *event,
 753                                        struct perf_event_ops *ops,
 754                                        u64 file_offset)
 755 {
 756         struct sample_data sample;
 757         int ret;
 758
 759         if (session->header.needs_swap && event__swap_ops[event->header.type])
 760                 event__swap_ops[event->header.type](event);
 761
 762         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 763                 return -EINVAL;
 764
 765         hists__inc_nr_events(&session->hists, event->header.type);
 766
 767         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 768                 dump_event(session, event, file_offset, NULL);
 769
 770         /* These events are processed right away */
 771         switch (event->header.type) {
 772         case PERF_RECORD_HEADER_ATTR:
 773                 return ops->attr(event, session);
 774         case PERF_RECORD_HEADER_EVENT_TYPE:
 775                 return ops->event_type(event, session);
 776         case PERF_RECORD_HEADER_TRACING_DATA:
 777                 /* setup for reading amidst mmap */
 778                 lseek(session->fd, file_offset, SEEK_SET);
 779                 return ops->tracing_data(event, session);
 780         case PERF_RECORD_HEADER_BUILD_ID:
 781                 return ops->build_id(event, session);
 782         case PERF_RECORD_FINISHED_ROUND:
 783                 return ops->finished_round(event, session, ops);
 784         default:
 785                 break;
 786         }
 787
 788         /*
 789          * For all kernel events we get the sample data
 790          */
 791         event__parse_sample(event, session, &sample);
 792
 793         /* Preprocess sample records - precheck callchains */
 794         if (perf_session__preprocess_sample(session, event, &sample))
 795                 return 0;
 796
 797         if (ops->ordered_samples) {
 798                 ret = perf_session_queue_event(session, event, &sample,
 799                                                file_offset);
 800                 if (ret != -ETIME)
 801                         return ret;
 802         }
 803
 804         return perf_session_deliver_event(session, event, &sample, ops,
 805                                           file_offset);
 806 }
 807
 808 void perf_event_header__bswap(struct perf_event_header *self)
 809 {
 810         self->type = bswap_32(self->type);
 811         self->misc = bswap_16(self->misc);
 812         self->size = bswap_16(self->size);
 813 }
 814
 815 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 816 {
 817         struct thread *thread = perf_session__findnew(self, 0);
 818
 819         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 820                 pr_err("problem inserting idle task.\n");
 821                 thread = NULL;
 822         }
 823
 824         return thread;
 825 }
 826
 827 int do_read(int fd, void *buf, size_t size)
 828 {
 829         void *buf_start = buf;
 830
 831         while (size) {
 832                 int ret = read(fd, buf, size);
 833
 834                 if (ret <= 0)
 835                         return ret;
 836
 837                 size -= ret;
 838                 buf += ret;
 839         }
 840
 841         return buf - buf_start;
 842 }
 843
 844 #define session_done()  (*(volatile int *)(&session_done))
 845 volatile int session_done;
 846
 847 static int __perf_session__process_pipe_events(struct perf_session *self,
 848                                                struct perf_event_ops *ops)
 849 {
 850         event_t event;
 851         uint32_t size;
 852         int skip = 0;
 853         u64 head;
 854         int err;
 855         void *p;
 856
 857         perf_event_ops__fill_defaults(ops);
 858
 859         head = 0;
 860 more:
 861         err = do_read(self->fd, &event, sizeof(struct perf_event_header));
 862         if (err <= 0) {
 863                 if (err == 0)
 864                         goto done;
 865
 866                 pr_err("failed to read event header\n");
 867                 goto out_err;
 868         }
 869
 870         if (self->header.needs_swap)
 871                 perf_event_header__bswap(&event.header);
 872
 873         size = event.header.size;
 874         if (size == 0)
 875                 size = 8;
 876
 877         p = &event;
 878         p += sizeof(struct perf_event_header);
 879
 880         if (size - sizeof(struct perf_event_header)) {
 881                 err = do_read(self->fd, p,
 882                               size - sizeof(struct perf_event_header));
 883                 if (err <= 0) {
 884                         if (err == 0) {
 885                                 pr_err("unexpected end of event stream\n");
 886                                 goto done;
 887                         }
 888
 889                         pr_err("failed to read event data\n");
 890                         goto out_err;
 891                 }
 892         }
 893
 894         if (size == 0 ||
 895             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 896                 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
 897                             head, event.header.size, event.header.type);
 898                 /*
 899                  * assume we lost track of the stream, check alignment, and
 900                  * increment a single u64 in the hope to catch on again 'soon'.
 901                  */
 902                 if (unlikely(head & 7))
 903                         head &= ~7ULL;
 904
 905                 size = 8;
 906         }
 907
 908         head += size;
 909
 910         if (skip > 0)
 911                 head += skip;
 912
 913         if (!session_done())
 914                 goto more;
 915 done:
 916         err = 0;
 917 out_err:
 918         perf_session_free_sample_buffers(self);
 919         return err;
 920 }
 921
 922 int __perf_session__process_events(struct perf_session *session,
 923                                    u64 data_offset, u64 data_size,
 924                                    u64 file_size, struct perf_event_ops *ops)
 925 {
 926         u64 head, page_offset, file_offset, file_pos, progress_next;
 927         int err, mmap_prot, mmap_flags, map_idx = 0;
 928         struct ui_progress *progress;
 929         size_t  page_size, mmap_size;
 930         char *buf, *mmaps[8];
 931         event_t *event;
 932         uint32_t size;
 933
 934         perf_event_ops__fill_defaults(ops);
 935
 936         page_size = sysconf(_SC_PAGESIZE);
 937
 938         page_offset = page_size * (data_offset / page_size);
 939         file_offset = page_offset;
 940         head = data_offset - page_offset;
 941
 942         if (data_offset + data_size < file_size)
 943                 file_size = data_offset + data_size;
 944
 945         progress_next = file_size / 16;
 946         progress = ui_progress__new("Processing events...", file_size);
 947         if (progress == NULL)
 948                 return -1;
 949
 950         mmap_size = session->mmap_window;
 951         if (mmap_size > file_size)
 952                 mmap_size = file_size;
 953
 954         memset(mmaps, 0, sizeof(mmaps));
 955
 956         mmap_prot  = PROT_READ;
 957         mmap_flags = MAP_SHARED;
 958
 959         if (session->header.needs_swap) {
 960                 mmap_prot  |= PROT_WRITE;
 961                 mmap_flags = MAP_PRIVATE;
 962         }
 963 remap:
 964         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
 965                    file_offset);
 966         if (buf == MAP_FAILED) {
 967                 pr_err("failed to mmap file\n");
 968                 err = -errno;
 969                 goto out_err;
 970         }
 971         mmaps[map_idx] = buf;
 972         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
 973         file_pos = file_offset + head;
 974
 975 more:
 976         event = (event_t *)(buf + head);
 977
 978         if (session->header.needs_swap)
 979                 perf_event_header__bswap(&event->header);
 980         size = event->header.size;
 981         if (size == 0)
 982                 size = 8;
 983
 984         if (head + event->header.size >= mmap_size) {
 985                 if (mmaps[map_idx]) {
 986                         munmap(mmaps[map_idx], mmap_size);
 987                         mmaps[map_idx] = NULL;
 988                 }
 989
 990                 page_offset = page_size * (head / page_size);
 991                 file_offset += page_offset;
 992                 head -= page_offset;
 993                 goto remap;
 994         }
 995
 996         size = event->header.size;
 997
 998         if (size == 0 ||
 999             perf_session__process_event(session, event, ops, file_pos) < 0) {
1000                 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
1001                             file_offset + head, event->header.size,
1002                             event->header.type);
1003                 /*
1004                  * assume we lost track of the stream, check alignment, and
1005                  * increment a single u64 in the hope to catch on again 'soon'.
1006                  */
1007                 if (unlikely(head & 7))
1008                         head &= ~7ULL;
1009
1010                 size = 8;
1011         }
1012
1013         head += size;
1014         file_pos += size;
1015
1016         if (file_pos >= progress_next) {
1017                 progress_next += file_size / 16;
1018                 ui_progress__update(progress, file_pos);
1019         }
1020
1021         if (file_pos < file_size)
1022                 goto more;
1023
1024         err = 0;
1025         /* do the final flush for ordered samples */
1026         session->ordered_samples.next_flush = ULLONG_MAX;
1027         flush_sample_queue(session, ops);
1028 out_err:
1029         ui_progress__delete(progress);
1030
1031         if (ops->lost == event__process_lost &&
1032             session->hists.stats.total_lost != 0) {
1033                 ui__warning("Processed %Lu events and LOST %Lu!\n\n"
1034                             "Check IO/CPU overload!\n\n",
1035                             session->hists.stats.total_period,
1036                             session->hists.stats.total_lost);
1037         }
1038
1039         if (session->hists.stats.nr_unknown_events != 0) {
1040                 ui__warning("Found %u unknown events!\n\n"
1041                             "Is this an older tool processing a perf.data "
1042                             "file generated by a more recent tool?\n\n"
1043                             "If that is not the case, consider "
1044                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1045                             session->hists.stats.nr_unknown_events);
1046         }
1047
1048         if (session->hists.stats.nr_invalid_chains != 0) {
1049                 ui__warning("Found invalid callchains!\n\n"
1050                             "%u out of %u events were discarded for this reason.\n\n"
1051                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1052                             session->hists.stats.nr_invalid_chains,
1053                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1054         }
1055
1056         perf_session_free_sample_buffers(session);
1057         return err;
1058 }
1059
1060 int perf_session__process_events(struct perf_session *self,
1061                                  struct perf_event_ops *ops)
1062 {
1063         int err;
1064
1065         if (perf_session__register_idle_thread(self) == NULL)
1066                 return -ENOMEM;
1067
1068         if (!self->fd_pipe)
1069                 err = __perf_session__process_events(self,
1070                                                      self->header.data_offset,
1071                                                      self->header.data_size,
1072                                                      self->size, ops);
1073         else
1074                 err = __perf_session__process_pipe_events(self, ops);
1075
1076         return err;
1077 }
1078
1079 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1080 {
1081         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1082                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1083                 return false;
1084         }
1085
1086         return true;
1087 }
1088
1089 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1090                                              const char *symbol_name,
1091                                              u64 addr)
1092 {
1093         char *bracket;
1094         enum map_type i;
1095         struct ref_reloc_sym *ref;
1096
1097         ref = zalloc(sizeof(struct ref_reloc_sym));
1098         if (ref == NULL)
1099                 return -ENOMEM;
1100
1101         ref->name = strdup(symbol_name);
1102         if (ref->name == NULL) {
1103                 free(ref);
1104                 return -ENOMEM;
1105         }
1106
1107         bracket = strchr(ref->name, ']');
1108         if (bracket)
1109                 *bracket = '\0';
1110
1111         ref->addr = addr;
1112
1113         for (i = 0; i < MAP__NR_TYPES; ++i) {
1114                 struct kmap *kmap = map__kmap(maps[i]);
1115                 kmap->ref_reloc_sym = ref;
1116         }
1117
1118         return 0;
1119 }
1120
1121 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1122 {
1123         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1124                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1125                machines__fprintf_dsos(&self->machines, fp);
1126 }
1127
1128 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1129                                           bool with_hits)
1130 {
1131         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1132         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1133 }