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         event_t                 *event;
 448         struct list_head        list;
 449 };
 450
 451 static void perf_session_free_sample_buffers(struct perf_session *session)
 452 {
 453         struct ordered_samples *os = &session->ordered_samples;
 454
 455         while (!list_empty(&os->to_free)) {
 456                 struct sample_queue *sq;
 457
 458                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 459                 list_del(&sq->list);
 460                 free(sq);
 461         }
 462 }
 463
 464 static int perf_session_deliver_event(struct perf_session *session,
 465                                       event_t *event,
 466                                       struct sample_data *sample,
 467                                       struct perf_event_ops *ops);
 468
 469 static void flush_sample_queue(struct perf_session *s,
 470                                struct perf_event_ops *ops)
 471 {
 472         struct ordered_samples *os = &s->ordered_samples;
 473         struct list_head *head = &os->samples;
 474         struct sample_queue *tmp, *iter;
 475         struct sample_data sample;
 476         u64 limit = os->next_flush;
 477         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 478
 479         if (!ops->ordered_samples || !limit)
 480                 return;
 481
 482         list_for_each_entry_safe(iter, tmp, head, list) {
 483                 if (iter->timestamp > limit)
 484                         break;
 485
 486                 event__parse_sample(iter->event, s, &sample);
 487                 perf_session_deliver_event(s, iter->event, &sample, ops);
 488
 489                 os->last_flush = iter->timestamp;
 490                 list_del(&iter->list);
 491                 list_add(&iter->list, &os->sample_cache);
 492         }
 493
 494         if (list_empty(head)) {
 495                 os->last_sample = NULL;
 496         } else if (last_ts <= limit) {
 497                 os->last_sample =
 498                         list_entry(head->prev, struct sample_queue, list);
 499         }
 500 }
 501
 502 /*
 503  * When perf record finishes a pass on every buffers, it records this pseudo
 504  * event.
 505  * We record the max timestamp t found in the pass n.
 506  * Assuming these timestamps are monotonic across cpus, we know that if
 507  * a buffer still has events with timestamps below t, they will be all
 508  * available and then read in the pass n + 1.
 509  * Hence when we start to read the pass n + 2, we can safely flush every
 510  * events with timestamps below t.
 511  *
 512  *    ============ PASS n =================
 513  *       CPU 0         |   CPU 1
 514  *                     |
 515  *    cnt1 timestamps  |   cnt2 timestamps
 516  *          1          |         2
 517  *          2          |         3
 518  *          -          |         4  <--- max recorded
 519  *
 520  *    ============ PASS n + 1 ==============
 521  *       CPU 0         |   CPU 1
 522  *                     |
 523  *    cnt1 timestamps  |   cnt2 timestamps
 524  *          3          |         5
 525  *          4          |         6
 526  *          5          |         7 <---- max recorded
 527  *
 528  *      Flush every events below timestamp 4
 529  *
 530  *    ============ PASS n + 2 ==============
 531  *       CPU 0         |   CPU 1
 532  *                     |
 533  *    cnt1 timestamps  |   cnt2 timestamps
 534  *          6          |         8
 535  *          7          |         9
 536  *          -          |         10
 537  *
 538  *      Flush every events below timestamp 7
 539  *      etc...
 540  */
 541 static int process_finished_round(event_t *event __used,
 542                                   struct perf_session *session,
 543                                   struct perf_event_ops *ops)
 544 {
 545         flush_sample_queue(session, ops);
 546         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 547
 548         return 0;
 549 }
 550
 551 /* The queue is ordered by time */
 552 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 553 {
 554         struct ordered_samples *os = &s->ordered_samples;
 555         struct sample_queue *sample = os->last_sample;
 556         u64 timestamp = new->timestamp;
 557         struct list_head *p;
 558
 559         os->last_sample = new;
 560
 561         if (!sample) {
 562                 list_add(&new->list, &os->samples);
 563                 os->max_timestamp = timestamp;
 564                 return;
 565         }
 566
 567         /*
 568          * last_sample might point to some random place in the list as it's
 569          * the last queued event. We expect that the new event is close to
 570          * this.
 571          */
 572         if (sample->timestamp <= timestamp) {
 573                 while (sample->timestamp <= timestamp) {
 574                         p = sample->list.next;
 575                         if (p == &os->samples) {
 576                                 list_add_tail(&new->list, &os->samples);
 577                                 os->max_timestamp = timestamp;
 578                                 return;
 579                         }
 580                         sample = list_entry(p, struct sample_queue, list);
 581                 }
 582                 list_add_tail(&new->list, &sample->list);
 583         } else {
 584                 while (sample->timestamp > timestamp) {
 585                         p = sample->list.prev;
 586                         if (p == &os->samples) {
 587                                 list_add(&new->list, &os->samples);
 588                                 return;
 589                         }
 590                         sample = list_entry(p, struct sample_queue, list);
 591                 }
 592                 list_add(&new->list, &sample->list);
 593         }
 594 }
 595
 596 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 597
 598 static int perf_session_queue_event(struct perf_session *s, event_t *event,
 599                                     struct sample_data *data)
 600 {
 601         struct ordered_samples *os = &s->ordered_samples;
 602         struct list_head *sc = &os->sample_cache;
 603         u64 timestamp = data->time;
 604         struct sample_queue *new;
 605
 606         if (!timestamp || timestamp == ~0ULL)
 607                 return -ETIME;
 608
 609         if (timestamp < s->ordered_samples.last_flush) {
 610                 printf("Warning: Timestamp below last timeslice flush\n");
 611                 return -EINVAL;
 612         }
 613
 614         if (!list_empty(sc)) {
 615                 new = list_entry(sc->next, struct sample_queue, list);
 616                 list_del(&new->list);
 617         } else if (os->sample_buffer) {
 618                 new = os->sample_buffer + os->sample_buffer_idx;
 619                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 620                         os->sample_buffer = NULL;
 621         } else {
 622                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 623                 if (!os->sample_buffer)
 624                         return -ENOMEM;
 625                 list_add(&os->sample_buffer->list, &os->to_free);
 626                 os->sample_buffer_idx = 2;
 627                 new = os->sample_buffer + 1;
 628         }
 629
 630         new->timestamp = timestamp;
 631         new->event = event;
 632
 633         __queue_event(new, s);
 634
 635         return 0;
 636 }
 637
 638 static void callchain__dump(struct sample_data *sample)
 639 {
 640         unsigned int i;
 641
 642         if (!dump_trace)
 643                 return;
 644
 645         printf("... chain: nr:%Lu\n", sample->callchain->nr);
 646
 647         for (i = 0; i < sample->callchain->nr; i++)
 648                 printf("..... %2d: %016Lx\n", i, sample->callchain->ips[i]);
 649 }
 650
 651 static void perf_session__print_tstamp(struct perf_session *session,
 652                                        event_t *event,
 653                                        struct sample_data *sample)
 654 {
 655         if (event->header.type != PERF_RECORD_SAMPLE &&
 656             !session->sample_id_all) {
 657                 fputs("-1 -1 ", stdout);
 658                 return;
 659         }
 660
 661         if ((session->sample_type & PERF_SAMPLE_CPU))
 662                 printf("%u ", sample->cpu);
 663
 664         if (session->sample_type & PERF_SAMPLE_TIME)
 665                 printf("%Lu ", sample->time);
 666 }
 667
 668 static void dump_event(struct perf_session *session, event_t *event,
 669                        u64 file_offset, struct sample_data *sample)
 670 {
 671         if (!dump_trace)
 672                 return;
 673
 674         dump_printf("\n%#Lx [%#x]: event: %d\n", file_offset,
 675                     event->header.size, event->header.type);
 676
 677         trace_event(event);
 678
 679         if (sample)
 680                 perf_session__print_tstamp(session, event, sample);
 681
 682         dump_printf("%#Lx [%#x]: PERF_RECORD_%s",
 683                     file_offset, event->header.size,
 684                     event__get_event_name(event->header.type));
 685 }
 686
 687 static void dump_sample(struct perf_session *session, event_t *event,
 688                         struct sample_data *sample)
 689 {
 690         dump_printf("(IP, %d): %d/%d: %#Lx period: %Ld\n", event->header.misc,
 691                     sample->pid, sample->tid, sample->ip, sample->period);
 692
 693         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 694                 callchain__dump(sample);
 695 }
 696
 697 static int perf_session_deliver_event(struct perf_session *session,
 698                                       event_t *event,
 699                                       struct sample_data *sample,
 700                                       struct perf_event_ops *ops)
 701 {
 702         switch (event->header.type) {
 703         case PERF_RECORD_SAMPLE:
 704                 return ops->sample(event, sample, session);
 705         case PERF_RECORD_MMAP:
 706                 return ops->mmap(event, sample, session);
 707         case PERF_RECORD_COMM:
 708                 return ops->comm(event, sample, session);
 709         case PERF_RECORD_FORK:
 710                 return ops->fork(event, sample, session);
 711         case PERF_RECORD_EXIT:
 712                 return ops->exit(event, sample, session);
 713         case PERF_RECORD_LOST:
 714                 return ops->lost(event, sample, session);
 715         case PERF_RECORD_READ:
 716                 return ops->read(event, sample, session);
 717         case PERF_RECORD_THROTTLE:
 718                 return ops->throttle(event, sample, session);
 719         case PERF_RECORD_UNTHROTTLE:
 720                 return ops->unthrottle(event, sample, session);
 721         default:
 722                 ++session->hists.stats.nr_unknown_events;
 723                 return -1;
 724         }
 725 }
 726
 727 static int perf_session__process_event(struct perf_session *session,
 728                                        event_t *event,
 729                                        struct perf_event_ops *ops,
 730                                        u64 file_offset)
 731 {
 732         struct sample_data sample;
 733         int ret;
 734
 735         if (session->header.needs_swap && event__swap_ops[event->header.type])
 736                 event__swap_ops[event->header.type](event);
 737
 738         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 739                 return -EINVAL;
 740
 741         hists__inc_nr_events(&session->hists, event->header.type);
 742
 743         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 744                 dump_event(session, event, file_offset, NULL);
 745         else {
 746                 event__parse_sample(event, session, &sample);
 747                 dump_event(session, event, file_offset, &sample);
 748         }
 749
 750         /* These events are processed right away */
 751         switch (event->header.type) {
 752         case PERF_RECORD_SAMPLE:
 753                 if (session->sample_type & PERF_SAMPLE_CALLCHAIN) {
 754                         if (!ip_callchain__valid(sample.callchain, event)) {
 755                                 pr_debug("call-chain problem with event, "
 756                                          "skipping it.\n");
 757                                 ++session->hists.stats.nr_invalid_chains;
 758                                 session->hists.stats.total_invalid_chains +=
 759                                         sample.period;
 760                                 return 0;
 761                         }
 762                 }
 763                 dump_sample(session, event, &sample);
 764                 break;
 765
 766         case PERF_RECORD_HEADER_ATTR:
 767                 return ops->attr(event, session);
 768         case PERF_RECORD_HEADER_EVENT_TYPE:
 769                 return ops->event_type(event, session);
 770         case PERF_RECORD_HEADER_TRACING_DATA:
 771                 /* setup for reading amidst mmap */
 772                 lseek(session->fd, file_offset, SEEK_SET);
 773                 return ops->tracing_data(event, session);
 774         case PERF_RECORD_HEADER_BUILD_ID:
 775                 return ops->build_id(event, session);
 776         case PERF_RECORD_FINISHED_ROUND:
 777                 return ops->finished_round(event, session, ops);
 778         default:
 779                 break;
 780         }
 781
 782         if (ops->ordered_samples) {
 783                 ret = perf_session_queue_event(session, event, &sample);
 784                 if (ret != -ETIME)
 785                         return ret;
 786         }
 787
 788         return perf_session_deliver_event(session, event, &sample, ops);
 789 }
 790
 791 void perf_event_header__bswap(struct perf_event_header *self)
 792 {
 793         self->type = bswap_32(self->type);
 794         self->misc = bswap_16(self->misc);
 795         self->size = bswap_16(self->size);
 796 }
 797
 798 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 799 {
 800         struct thread *thread = perf_session__findnew(self, 0);
 801
 802         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 803                 pr_err("problem inserting idle task.\n");
 804                 thread = NULL;
 805         }
 806
 807         return thread;
 808 }
 809
 810 int do_read(int fd, void *buf, size_t size)
 811 {
 812         void *buf_start = buf;
 813
 814         while (size) {
 815                 int ret = read(fd, buf, size);
 816
 817                 if (ret <= 0)
 818                         return ret;
 819
 820                 size -= ret;
 821                 buf += ret;
 822         }
 823
 824         return buf - buf_start;
 825 }
 826
 827 #define session_done()  (*(volatile int *)(&session_done))
 828 volatile int session_done;
 829
 830 static int __perf_session__process_pipe_events(struct perf_session *self,
 831                                                struct perf_event_ops *ops)
 832 {
 833         event_t event;
 834         uint32_t size;
 835         int skip = 0;
 836         u64 head;
 837         int err;
 838         void *p;
 839
 840         perf_event_ops__fill_defaults(ops);
 841
 842         head = 0;
 843 more:
 844         err = do_read(self->fd, &event, sizeof(struct perf_event_header));
 845         if (err <= 0) {
 846                 if (err == 0)
 847                         goto done;
 848
 849                 pr_err("failed to read event header\n");
 850                 goto out_err;
 851         }
 852
 853         if (self->header.needs_swap)
 854                 perf_event_header__bswap(&event.header);
 855
 856         size = event.header.size;
 857         if (size == 0)
 858                 size = 8;
 859
 860         p = &event;
 861         p += sizeof(struct perf_event_header);
 862
 863         if (size - sizeof(struct perf_event_header)) {
 864                 err = do_read(self->fd, p,
 865                               size - sizeof(struct perf_event_header));
 866                 if (err <= 0) {
 867                         if (err == 0) {
 868                                 pr_err("unexpected end of event stream\n");
 869                                 goto done;
 870                         }
 871
 872                         pr_err("failed to read event data\n");
 873                         goto out_err;
 874                 }
 875         }
 876
 877         if (size == 0 ||
 878             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 879                 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
 880                             head, event.header.size, event.header.type);
 881                 /*
 882                  * assume we lost track of the stream, check alignment, and
 883                  * increment a single u64 in the hope to catch on again 'soon'.
 884                  */
 885                 if (unlikely(head & 7))
 886                         head &= ~7ULL;
 887
 888                 size = 8;
 889         }
 890
 891         head += size;
 892
 893         if (skip > 0)
 894                 head += skip;
 895
 896         if (!session_done())
 897                 goto more;
 898 done:
 899         err = 0;
 900 out_err:
 901         perf_session_free_sample_buffers(self);
 902         return err;
 903 }
 904
 905 int __perf_session__process_events(struct perf_session *session,
 906                                    u64 data_offset, u64 data_size,
 907                                    u64 file_size, struct perf_event_ops *ops)
 908 {
 909         u64 head, page_offset, file_offset, file_pos, progress_next;
 910         int err, mmap_prot, mmap_flags, map_idx = 0;
 911         struct ui_progress *progress;
 912         size_t  page_size, mmap_size;
 913         char *buf, *mmaps[8];
 914         event_t *event;
 915         uint32_t size;
 916
 917         perf_event_ops__fill_defaults(ops);
 918
 919         page_size = sysconf(_SC_PAGESIZE);
 920
 921         page_offset = page_size * (data_offset / page_size);
 922         file_offset = page_offset;
 923         head = data_offset - page_offset;
 924
 925         if (data_offset + data_size < file_size)
 926                 file_size = data_offset + data_size;
 927
 928         progress_next = file_size / 16;
 929         progress = ui_progress__new("Processing events...", file_size);
 930         if (progress == NULL)
 931                 return -1;
 932
 933         mmap_size = session->mmap_window;
 934         if (mmap_size > file_size)
 935                 mmap_size = file_size;
 936
 937         memset(mmaps, 0, sizeof(mmaps));
 938
 939         mmap_prot  = PROT_READ;
 940         mmap_flags = MAP_SHARED;
 941
 942         if (session->header.needs_swap) {
 943                 mmap_prot  |= PROT_WRITE;
 944                 mmap_flags = MAP_PRIVATE;
 945         }
 946 remap:
 947         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
 948                    file_offset);
 949         if (buf == MAP_FAILED) {
 950                 pr_err("failed to mmap file\n");
 951                 err = -errno;
 952                 goto out_err;
 953         }
 954         mmaps[map_idx] = buf;
 955         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
 956         file_pos = file_offset + head;
 957
 958 more:
 959         event = (event_t *)(buf + head);
 960
 961         if (session->header.needs_swap)
 962                 perf_event_header__bswap(&event->header);
 963         size = event->header.size;
 964         if (size == 0)
 965                 size = 8;
 966
 967         if (head + event->header.size >= mmap_size) {
 968                 if (mmaps[map_idx]) {
 969                         munmap(mmaps[map_idx], mmap_size);
 970                         mmaps[map_idx] = NULL;
 971                 }
 972
 973                 page_offset = page_size * (head / page_size);
 974                 file_offset += page_offset;
 975                 head -= page_offset;
 976                 goto remap;
 977         }
 978
 979         size = event->header.size;
 980
 981         if (size == 0 ||
 982             perf_session__process_event(session, event, ops, file_pos) < 0) {
 983                 dump_printf("%#Lx [%#x]: skipping unknown header type: %d\n",
 984                             file_offset + head, event->header.size,
 985                             event->header.type);
 986                 /*
 987                  * assume we lost track of the stream, check alignment, and
 988                  * increment a single u64 in the hope to catch on again 'soon'.
 989                  */
 990                 if (unlikely(head & 7))
 991                         head &= ~7ULL;
 992
 993                 size = 8;
 994         }
 995
 996         head += size;
 997         file_pos += size;
 998
 999         if (file_pos >= progress_next) {
1000                 progress_next += file_size / 16;
1001                 ui_progress__update(progress, file_pos);
1002         }
1003
1004         if (file_pos < file_size)
1005                 goto more;
1006
1007         err = 0;
1008         /* do the final flush for ordered samples */
1009         session->ordered_samples.next_flush = ULLONG_MAX;
1010         flush_sample_queue(session, ops);
1011 out_err:
1012         ui_progress__delete(progress);
1013
1014         if (ops->lost == event__process_lost &&
1015             session->hists.stats.total_lost != 0) {
1016                 ui__warning("Processed %Lu events and LOST %Lu!\n\n"
1017                             "Check IO/CPU overload!\n\n",
1018                             session->hists.stats.total_period,
1019                             session->hists.stats.total_lost);
1020         }
1021
1022         if (session->hists.stats.nr_unknown_events != 0) {
1023                 ui__warning("Found %u unknown events!\n\n"
1024                             "Is this an older tool processing a perf.data "
1025                             "file generated by a more recent tool?\n\n"
1026                             "If that is not the case, consider "
1027                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1028                             session->hists.stats.nr_unknown_events);
1029         }
1030
1031         if (session->hists.stats.nr_invalid_chains != 0) {
1032                 ui__warning("Found invalid callchains!\n\n"
1033                             "%u out of %u events were discarded for this reason.\n\n"
1034                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1035                             session->hists.stats.nr_invalid_chains,
1036                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1037         }
1038
1039         perf_session_free_sample_buffers(session);
1040         return err;
1041 }
1042
1043 int perf_session__process_events(struct perf_session *self,
1044                                  struct perf_event_ops *ops)
1045 {
1046         int err;
1047
1048         if (perf_session__register_idle_thread(self) == NULL)
1049                 return -ENOMEM;
1050
1051         if (!self->fd_pipe)
1052                 err = __perf_session__process_events(self,
1053                                                      self->header.data_offset,
1054                                                      self->header.data_size,
1055                                                      self->size, ops);
1056         else
1057                 err = __perf_session__process_pipe_events(self, ops);
1058
1059         return err;
1060 }
1061
1062 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1063 {
1064         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1065                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1066                 return false;
1067         }
1068
1069         return true;
1070 }
1071
1072 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1073                                              const char *symbol_name,
1074                                              u64 addr)
1075 {
1076         char *bracket;
1077         enum map_type i;
1078         struct ref_reloc_sym *ref;
1079
1080         ref = zalloc(sizeof(struct ref_reloc_sym));
1081         if (ref == NULL)
1082                 return -ENOMEM;
1083
1084         ref->name = strdup(symbol_name);
1085         if (ref->name == NULL) {
1086                 free(ref);
1087                 return -ENOMEM;
1088         }
1089
1090         bracket = strchr(ref->name, ']');
1091         if (bracket)
1092                 *bracket = '\0';
1093
1094         ref->addr = addr;
1095
1096         for (i = 0; i < MAP__NR_TYPES; ++i) {
1097                 struct kmap *kmap = map__kmap(maps[i]);
1098                 kmap->ref_reloc_sym = ref;
1099         }
1100
1101         return 0;
1102 }
1103
1104 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1105 {
1106         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1107                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1108                machines__fprintf_dsos(&self->machines, fp);
1109 }
1110
1111 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1112                                           bool with_hits)
1113 {
1114         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1115         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1116 }