Pileus Git - ~andy/linux/blob - tools/perf/util/session.c

   1 #define _FILE_OFFSET_BITS 64
   2
   3 #include <linux/kernel.h>
   4
   5 #include <byteswap.h>
   6 #include <unistd.h>
   7 #include <sys/types.h>
   8 #include <sys/mman.h>
   9
  10 #include "evlist.h"
  11 #include "evsel.h"
  12 #include "session.h"
  13 #include "sort.h"
  14 #include "util.h"
  15
  16 static int perf_session__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_stub(union perf_event *event __used,
 284                               struct perf_sample *sample __used,
 285                               struct perf_session *session __used)
 286 {
 287         dump_printf(": unhandled!\n");
 288         return 0;
 289 }
 290
 291 static int process_finished_round_stub(union perf_event *event __used,
 292                                        struct perf_session *session __used,
 293                                        struct perf_event_ops *ops __used)
 294 {
 295         dump_printf(": unhandled!\n");
 296         return 0;
 297 }
 298
 299 static int process_finished_round(union perf_event *event,
 300                                   struct perf_session *session,
 301                                   struct perf_event_ops *ops);
 302
 303 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 304 {
 305         if (handler->sample == NULL)
 306                 handler->sample = process_event_stub;
 307         if (handler->mmap == NULL)
 308                 handler->mmap = process_event_stub;
 309         if (handler->comm == NULL)
 310                 handler->comm = process_event_stub;
 311         if (handler->fork == NULL)
 312                 handler->fork = process_event_stub;
 313         if (handler->exit == NULL)
 314                 handler->exit = process_event_stub;
 315         if (handler->lost == NULL)
 316                 handler->lost = perf_event__process_lost;
 317         if (handler->read == NULL)
 318                 handler->read = process_event_stub;
 319         if (handler->throttle == NULL)
 320                 handler->throttle = process_event_stub;
 321         if (handler->unthrottle == NULL)
 322                 handler->unthrottle = process_event_stub;
 323         if (handler->attr == NULL)
 324                 handler->attr = process_event_synth_stub;
 325         if (handler->event_type == NULL)
 326                 handler->event_type = process_event_synth_stub;
 327         if (handler->tracing_data == NULL)
 328                 handler->tracing_data = process_event_synth_stub;
 329         if (handler->build_id == NULL)
 330                 handler->build_id = process_event_synth_stub;
 331         if (handler->finished_round == NULL) {
 332                 if (handler->ordered_samples)
 333                         handler->finished_round = process_finished_round;
 334                 else
 335                         handler->finished_round = process_finished_round_stub;
 336         }
 337 }
 338
 339 void mem_bswap_64(void *src, int byte_size)
 340 {
 341         u64 *m = src;
 342
 343         while (byte_size > 0) {
 344                 *m = bswap_64(*m);
 345                 byte_size -= sizeof(u64);
 346                 ++m;
 347         }
 348 }
 349
 350 static void perf_event__all64_swap(union perf_event *event)
 351 {
 352         struct perf_event_header *hdr = &event->header;
 353         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 354 }
 355
 356 static void perf_event__comm_swap(union perf_event *event)
 357 {
 358         event->comm.pid = bswap_32(event->comm.pid);
 359         event->comm.tid = bswap_32(event->comm.tid);
 360 }
 361
 362 static void perf_event__mmap_swap(union perf_event *event)
 363 {
 364         event->mmap.pid   = bswap_32(event->mmap.pid);
 365         event->mmap.tid   = bswap_32(event->mmap.tid);
 366         event->mmap.start = bswap_64(event->mmap.start);
 367         event->mmap.len   = bswap_64(event->mmap.len);
 368         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 369 }
 370
 371 static void perf_event__task_swap(union perf_event *event)
 372 {
 373         event->fork.pid  = bswap_32(event->fork.pid);
 374         event->fork.tid  = bswap_32(event->fork.tid);
 375         event->fork.ppid = bswap_32(event->fork.ppid);
 376         event->fork.ptid = bswap_32(event->fork.ptid);
 377         event->fork.time = bswap_64(event->fork.time);
 378 }
 379
 380 static void perf_event__read_swap(union perf_event *event)
 381 {
 382         event->read.pid          = bswap_32(event->read.pid);
 383         event->read.tid          = bswap_32(event->read.tid);
 384         event->read.value        = bswap_64(event->read.value);
 385         event->read.time_enabled = bswap_64(event->read.time_enabled);
 386         event->read.time_running = bswap_64(event->read.time_running);
 387         event->read.id           = bswap_64(event->read.id);
 388 }
 389
 390 static void perf_event__attr_swap(union perf_event *event)
 391 {
 392         size_t size;
 393
 394         event->attr.attr.type           = bswap_32(event->attr.attr.type);
 395         event->attr.attr.size           = bswap_32(event->attr.attr.size);
 396         event->attr.attr.config         = bswap_64(event->attr.attr.config);
 397         event->attr.attr.sample_period  = bswap_64(event->attr.attr.sample_period);
 398         event->attr.attr.sample_type    = bswap_64(event->attr.attr.sample_type);
 399         event->attr.attr.read_format    = bswap_64(event->attr.attr.read_format);
 400         event->attr.attr.wakeup_events  = bswap_32(event->attr.attr.wakeup_events);
 401         event->attr.attr.bp_type        = bswap_32(event->attr.attr.bp_type);
 402         event->attr.attr.bp_addr        = bswap_64(event->attr.attr.bp_addr);
 403         event->attr.attr.bp_len         = bswap_64(event->attr.attr.bp_len);
 404
 405         size = event->header.size;
 406         size -= (void *)&event->attr.id - (void *)event;
 407         mem_bswap_64(event->attr.id, size);
 408 }
 409
 410 static void perf_event__event_type_swap(union perf_event *event)
 411 {
 412         event->event_type.event_type.event_id =
 413                 bswap_64(event->event_type.event_type.event_id);
 414 }
 415
 416 static void perf_event__tracing_data_swap(union perf_event *event)
 417 {
 418         event->tracing_data.size = bswap_32(event->tracing_data.size);
 419 }
 420
 421 typedef void (*perf_event__swap_op)(union perf_event *event);
 422
 423 static perf_event__swap_op perf_event__swap_ops[] = {
 424         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 425         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 426         [PERF_RECORD_FORK]                = perf_event__task_swap,
 427         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 428         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 429         [PERF_RECORD_READ]                = perf_event__read_swap,
 430         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 431         [PERF_RECORD_HEADER_ATTR]         = perf_event__attr_swap,
 432         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 433         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 434         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 435         [PERF_RECORD_HEADER_MAX]          = NULL,
 436 };
 437
 438 struct sample_queue {
 439         u64                     timestamp;
 440         u64                     file_offset;
 441         union perf_event        *event;
 442         struct list_head        list;
 443 };
 444
 445 static void perf_session_free_sample_buffers(struct perf_session *session)
 446 {
 447         struct ordered_samples *os = &session->ordered_samples;
 448
 449         while (!list_empty(&os->to_free)) {
 450                 struct sample_queue *sq;
 451
 452                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 453                 list_del(&sq->list);
 454                 free(sq);
 455         }
 456 }
 457
 458 static int perf_session_deliver_event(struct perf_session *session,
 459                                       union perf_event *event,
 460                                       struct perf_sample *sample,
 461                                       struct perf_event_ops *ops,
 462                                       u64 file_offset);
 463
 464 static void flush_sample_queue(struct perf_session *s,
 465                                struct perf_event_ops *ops)
 466 {
 467         struct ordered_samples *os = &s->ordered_samples;
 468         struct list_head *head = &os->samples;
 469         struct sample_queue *tmp, *iter;
 470         struct perf_sample sample;
 471         u64 limit = os->next_flush;
 472         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 473
 474         if (!ops->ordered_samples || !limit)
 475                 return;
 476
 477         list_for_each_entry_safe(iter, tmp, head, list) {
 478                 if (iter->timestamp > limit)
 479                         break;
 480
 481                 perf_session__parse_sample(s, iter->event, &sample);
 482                 perf_session_deliver_event(s, iter->event, &sample, ops,
 483                                            iter->file_offset);
 484
 485                 os->last_flush = iter->timestamp;
 486                 list_del(&iter->list);
 487                 list_add(&iter->list, &os->sample_cache);
 488         }
 489
 490         if (list_empty(head)) {
 491                 os->last_sample = NULL;
 492         } else if (last_ts <= limit) {
 493                 os->last_sample =
 494                         list_entry(head->prev, struct sample_queue, list);
 495         }
 496 }
 497
 498 /*
 499  * When perf record finishes a pass on every buffers, it records this pseudo
 500  * event.
 501  * We record the max timestamp t found in the pass n.
 502  * Assuming these timestamps are monotonic across cpus, we know that if
 503  * a buffer still has events with timestamps below t, they will be all
 504  * available and then read in the pass n + 1.
 505  * Hence when we start to read the pass n + 2, we can safely flush every
 506  * events with timestamps below t.
 507  *
 508  *    ============ PASS n =================
 509  *       CPU 0         |   CPU 1
 510  *                     |
 511  *    cnt1 timestamps  |   cnt2 timestamps
 512  *          1          |         2
 513  *          2          |         3
 514  *          -          |         4  <--- max recorded
 515  *
 516  *    ============ PASS n + 1 ==============
 517  *       CPU 0         |   CPU 1
 518  *                     |
 519  *    cnt1 timestamps  |   cnt2 timestamps
 520  *          3          |         5
 521  *          4          |         6
 522  *          5          |         7 <---- max recorded
 523  *
 524  *      Flush every events below timestamp 4
 525  *
 526  *    ============ PASS n + 2 ==============
 527  *       CPU 0         |   CPU 1
 528  *                     |
 529  *    cnt1 timestamps  |   cnt2 timestamps
 530  *          6          |         8
 531  *          7          |         9
 532  *          -          |         10
 533  *
 534  *      Flush every events below timestamp 7
 535  *      etc...
 536  */
 537 static int process_finished_round(union perf_event *event __used,
 538                                   struct perf_session *session,
 539                                   struct perf_event_ops *ops)
 540 {
 541         flush_sample_queue(session, ops);
 542         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 543
 544         return 0;
 545 }
 546
 547 /* The queue is ordered by time */
 548 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 549 {
 550         struct ordered_samples *os = &s->ordered_samples;
 551         struct sample_queue *sample = os->last_sample;
 552         u64 timestamp = new->timestamp;
 553         struct list_head *p;
 554
 555         os->last_sample = new;
 556
 557         if (!sample) {
 558                 list_add(&new->list, &os->samples);
 559                 os->max_timestamp = timestamp;
 560                 return;
 561         }
 562
 563         /*
 564          * last_sample might point to some random place in the list as it's
 565          * the last queued event. We expect that the new event is close to
 566          * this.
 567          */
 568         if (sample->timestamp <= timestamp) {
 569                 while (sample->timestamp <= timestamp) {
 570                         p = sample->list.next;
 571                         if (p == &os->samples) {
 572                                 list_add_tail(&new->list, &os->samples);
 573                                 os->max_timestamp = timestamp;
 574                                 return;
 575                         }
 576                         sample = list_entry(p, struct sample_queue, list);
 577                 }
 578                 list_add_tail(&new->list, &sample->list);
 579         } else {
 580                 while (sample->timestamp > timestamp) {
 581                         p = sample->list.prev;
 582                         if (p == &os->samples) {
 583                                 list_add(&new->list, &os->samples);
 584                                 return;
 585                         }
 586                         sample = list_entry(p, struct sample_queue, list);
 587                 }
 588                 list_add(&new->list, &sample->list);
 589         }
 590 }
 591
 592 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 593
 594 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 595                                     struct perf_sample *sample, u64 file_offset)
 596 {
 597         struct ordered_samples *os = &s->ordered_samples;
 598         struct list_head *sc = &os->sample_cache;
 599         u64 timestamp = sample->time;
 600         struct sample_queue *new;
 601
 602         if (!timestamp || timestamp == ~0ULL)
 603                 return -ETIME;
 604
 605         if (timestamp < s->ordered_samples.last_flush) {
 606                 printf("Warning: Timestamp below last timeslice flush\n");
 607                 return -EINVAL;
 608         }
 609
 610         if (!list_empty(sc)) {
 611                 new = list_entry(sc->next, struct sample_queue, list);
 612                 list_del(&new->list);
 613         } else if (os->sample_buffer) {
 614                 new = os->sample_buffer + os->sample_buffer_idx;
 615                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 616                         os->sample_buffer = NULL;
 617         } else {
 618                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 619                 if (!os->sample_buffer)
 620                         return -ENOMEM;
 621                 list_add(&os->sample_buffer->list, &os->to_free);
 622                 os->sample_buffer_idx = 2;
 623                 new = os->sample_buffer + 1;
 624         }
 625
 626         new->timestamp = timestamp;
 627         new->file_offset = file_offset;
 628         new->event = event;
 629
 630         __queue_event(new, s);
 631
 632         return 0;
 633 }
 634
 635 static void callchain__printf(struct perf_sample *sample)
 636 {
 637         unsigned int i;
 638
 639         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 640
 641         for (i = 0; i < sample->callchain->nr; i++)
 642                 printf("..... %2d: %016" PRIx64 "\n",
 643                        i, sample->callchain->ips[i]);
 644 }
 645
 646 static void perf_session__print_tstamp(struct perf_session *session,
 647                                        union perf_event *event,
 648                                        struct perf_sample *sample)
 649 {
 650         if (event->header.type != PERF_RECORD_SAMPLE &&
 651             !session->sample_id_all) {
 652                 fputs("-1 -1 ", stdout);
 653                 return;
 654         }
 655
 656         if ((session->sample_type & PERF_SAMPLE_CPU))
 657                 printf("%u ", sample->cpu);
 658
 659         if (session->sample_type & PERF_SAMPLE_TIME)
 660                 printf("%" PRIu64 " ", sample->time);
 661 }
 662
 663 static void dump_event(struct perf_session *session, union perf_event *event,
 664                        u64 file_offset, struct perf_sample *sample)
 665 {
 666         if (!dump_trace)
 667                 return;
 668
 669         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 670                file_offset, event->header.size, event->header.type);
 671
 672         trace_event(event);
 673
 674         if (sample)
 675                 perf_session__print_tstamp(session, event, sample);
 676
 677         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 678                event->header.size, perf_event__name(event->header.type));
 679 }
 680
 681 static void dump_sample(struct perf_session *session, union perf_event *event,
 682                         struct perf_sample *sample)
 683 {
 684         if (!dump_trace)
 685                 return;
 686
 687         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 "\n",
 688                event->header.misc, sample->pid, sample->tid, sample->ip,
 689                sample->period);
 690
 691         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 692                 callchain__printf(sample);
 693 }
 694
 695 static int perf_session_deliver_event(struct perf_session *session,
 696                                       union perf_event *event,
 697                                       struct perf_sample *sample,
 698                                       struct perf_event_ops *ops,
 699                                       u64 file_offset)
 700 {
 701         dump_event(session, event, file_offset, sample);
 702
 703         switch (event->header.type) {
 704         case PERF_RECORD_SAMPLE:
 705                 dump_sample(session, event, sample);
 706                 return ops->sample(event, sample, session);
 707         case PERF_RECORD_MMAP:
 708                 return ops->mmap(event, sample, session);
 709         case PERF_RECORD_COMM:
 710                 return ops->comm(event, sample, session);
 711         case PERF_RECORD_FORK:
 712                 return ops->fork(event, sample, session);
 713         case PERF_RECORD_EXIT:
 714                 return ops->exit(event, sample, session);
 715         case PERF_RECORD_LOST:
 716                 return ops->lost(event, sample, session);
 717         case PERF_RECORD_READ:
 718                 return ops->read(event, sample, session);
 719         case PERF_RECORD_THROTTLE:
 720                 return ops->throttle(event, sample, session);
 721         case PERF_RECORD_UNTHROTTLE:
 722                 return ops->unthrottle(event, sample, session);
 723         default:
 724                 ++session->hists.stats.nr_unknown_events;
 725                 return -1;
 726         }
 727 }
 728
 729 static int perf_session__preprocess_sample(struct perf_session *session,
 730                                            union perf_event *event, struct perf_sample *sample)
 731 {
 732         if (event->header.type != PERF_RECORD_SAMPLE ||
 733             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 734                 return 0;
 735
 736         if (!ip_callchain__valid(sample->callchain, event)) {
 737                 pr_debug("call-chain problem with event, skipping it.\n");
 738                 ++session->hists.stats.nr_invalid_chains;
 739                 session->hists.stats.total_invalid_chains += sample->period;
 740                 return -EINVAL;
 741         }
 742         return 0;
 743 }
 744
 745 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
 746                                             struct perf_event_ops *ops, u64 file_offset)
 747 {
 748         dump_event(session, event, file_offset, NULL);
 749
 750         /* These events are processed right away */
 751         switch (event->header.type) {
 752         case PERF_RECORD_HEADER_ATTR:
 753                 return ops->attr(event, session);
 754         case PERF_RECORD_HEADER_EVENT_TYPE:
 755                 return ops->event_type(event, session);
 756         case PERF_RECORD_HEADER_TRACING_DATA:
 757                 /* setup for reading amidst mmap */
 758                 lseek(session->fd, file_offset, SEEK_SET);
 759                 return ops->tracing_data(event, session);
 760         case PERF_RECORD_HEADER_BUILD_ID:
 761                 return ops->build_id(event, session);
 762         case PERF_RECORD_FINISHED_ROUND:
 763                 return ops->finished_round(event, session, ops);
 764         default:
 765                 return -EINVAL;
 766         }
 767 }
 768
 769 static int perf_session__process_event(struct perf_session *session,
 770                                        union perf_event *event,
 771                                        struct perf_event_ops *ops,
 772                                        u64 file_offset)
 773 {
 774         struct perf_sample sample;
 775         int ret;
 776
 777         if (session->header.needs_swap &&
 778             perf_event__swap_ops[event->header.type])
 779                 perf_event__swap_ops[event->header.type](event);
 780
 781         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 782                 return -EINVAL;
 783
 784         hists__inc_nr_events(&session->hists, event->header.type);
 785
 786         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 787                 return perf_session__process_user_event(session, event, ops, file_offset);
 788
 789         /*
 790          * For all kernel events we get the sample data
 791          */
 792         perf_session__parse_sample(session, event, &sample);
 793
 794         /* Preprocess sample records - precheck callchains */
 795         if (perf_session__preprocess_sample(session, event, &sample))
 796                 return 0;
 797
 798         if (ops->ordered_samples) {
 799                 ret = perf_session_queue_event(session, event, &sample,
 800                                                file_offset);
 801                 if (ret != -ETIME)
 802                         return ret;
 803         }
 804
 805         return perf_session_deliver_event(session, event, &sample, ops,
 806                                           file_offset);
 807 }
 808
 809 void perf_event_header__bswap(struct perf_event_header *self)
 810 {
 811         self->type = bswap_32(self->type);
 812         self->misc = bswap_16(self->misc);
 813         self->size = bswap_16(self->size);
 814 }
 815
 816 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 817 {
 818         struct thread *thread = perf_session__findnew(self, 0);
 819
 820         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 821                 pr_err("problem inserting idle task.\n");
 822                 thread = NULL;
 823         }
 824
 825         return thread;
 826 }
 827
 828 static void perf_session__warn_about_errors(const struct perf_session *session,
 829                                             const struct perf_event_ops *ops)
 830 {
 831         if (ops->lost == perf_event__process_lost &&
 832             session->hists.stats.total_lost != 0) {
 833                 ui__warning("Processed %" PRIu64 " events and LOST %" PRIu64
 834                             "!\n\nCheck IO/CPU overload!\n\n",
 835                             session->hists.stats.total_period,
 836                             session->hists.stats.total_lost);
 837         }
 838
 839         if (session->hists.stats.nr_unknown_events != 0) {
 840                 ui__warning("Found %u unknown events!\n\n"
 841                             "Is this an older tool processing a perf.data "
 842                             "file generated by a more recent tool?\n\n"
 843                             "If that is not the case, consider "
 844                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 845                             session->hists.stats.nr_unknown_events);
 846         }
 847
 848         if (session->hists.stats.nr_invalid_chains != 0) {
 849                 ui__warning("Found invalid callchains!\n\n"
 850                             "%u out of %u events were discarded for this reason.\n\n"
 851                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 852                             session->hists.stats.nr_invalid_chains,
 853                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 854         }
 855 }
 856
 857 #define session_done()  (*(volatile int *)(&session_done))
 858 volatile int session_done;
 859
 860 static int __perf_session__process_pipe_events(struct perf_session *self,
 861                                                struct perf_event_ops *ops)
 862 {
 863         union perf_event event;
 864         uint32_t size;
 865         int skip = 0;
 866         u64 head;
 867         int err;
 868         void *p;
 869
 870         perf_event_ops__fill_defaults(ops);
 871
 872         head = 0;
 873 more:
 874         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 875         if (err <= 0) {
 876                 if (err == 0)
 877                         goto done;
 878
 879                 pr_err("failed to read event header\n");
 880                 goto out_err;
 881         }
 882
 883         if (self->header.needs_swap)
 884                 perf_event_header__bswap(&event.header);
 885
 886         size = event.header.size;
 887         if (size == 0)
 888                 size = 8;
 889
 890         p = &event;
 891         p += sizeof(struct perf_event_header);
 892
 893         if (size - sizeof(struct perf_event_header)) {
 894                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 895                 if (err <= 0) {
 896                         if (err == 0) {
 897                                 pr_err("unexpected end of event stream\n");
 898                                 goto done;
 899                         }
 900
 901                         pr_err("failed to read event data\n");
 902                         goto out_err;
 903                 }
 904         }
 905
 906         if (size == 0 ||
 907             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 908                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 909                             head, event.header.size, event.header.type);
 910                 /*
 911                  * assume we lost track of the stream, check alignment, and
 912                  * increment a single u64 in the hope to catch on again 'soon'.
 913                  */
 914                 if (unlikely(head & 7))
 915                         head &= ~7ULL;
 916
 917                 size = 8;
 918         }
 919
 920         head += size;
 921
 922         if (skip > 0)
 923                 head += skip;
 924
 925         if (!session_done())
 926                 goto more;
 927 done:
 928         err = 0;
 929 out_err:
 930         perf_session__warn_about_errors(self, ops);
 931         perf_session_free_sample_buffers(self);
 932         return err;
 933 }
 934
 935 int __perf_session__process_events(struct perf_session *session,
 936                                    u64 data_offset, u64 data_size,
 937                                    u64 file_size, struct perf_event_ops *ops)
 938 {
 939         u64 head, page_offset, file_offset, file_pos, progress_next;
 940         int err, mmap_prot, mmap_flags, map_idx = 0;
 941         struct ui_progress *progress;
 942         size_t  page_size, mmap_size;
 943         char *buf, *mmaps[8];
 944         union perf_event *event;
 945         uint32_t size;
 946
 947         perf_event_ops__fill_defaults(ops);
 948
 949         page_size = sysconf(_SC_PAGESIZE);
 950
 951         page_offset = page_size * (data_offset / page_size);
 952         file_offset = page_offset;
 953         head = data_offset - page_offset;
 954
 955         if (data_offset + data_size < file_size)
 956                 file_size = data_offset + data_size;
 957
 958         progress_next = file_size / 16;
 959         progress = ui_progress__new("Processing events...", file_size);
 960         if (progress == NULL)
 961                 return -1;
 962
 963         mmap_size = session->mmap_window;
 964         if (mmap_size > file_size)
 965                 mmap_size = file_size;
 966
 967         memset(mmaps, 0, sizeof(mmaps));
 968
 969         mmap_prot  = PROT_READ;
 970         mmap_flags = MAP_SHARED;
 971
 972         if (session->header.needs_swap) {
 973                 mmap_prot  |= PROT_WRITE;
 974                 mmap_flags = MAP_PRIVATE;
 975         }
 976 remap:
 977         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
 978                    file_offset);
 979         if (buf == MAP_FAILED) {
 980                 pr_err("failed to mmap file\n");
 981                 err = -errno;
 982                 goto out_err;
 983         }
 984         mmaps[map_idx] = buf;
 985         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
 986         file_pos = file_offset + head;
 987
 988 more:
 989         event = (union perf_event *)(buf + head);
 990
 991         if (session->header.needs_swap)
 992                 perf_event_header__bswap(&event->header);
 993         size = event->header.size;
 994         if (size == 0)
 995                 size = 8;
 996
 997         if (head + event->header.size > mmap_size) {
 998                 if (mmaps[map_idx]) {
 999                         munmap(mmaps[map_idx], mmap_size);
1000                         mmaps[map_idx] = NULL;
1001                 }
1002
1003                 page_offset = page_size * (head / page_size);
1004                 file_offset += page_offset;
1005                 head -= page_offset;
1006                 goto remap;
1007         }
1008
1009         size = event->header.size;
1010
1011         if (size == 0 ||
1012             perf_session__process_event(session, event, ops, file_pos) < 0) {
1013                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1014                             file_offset + head, event->header.size,
1015                             event->header.type);
1016                 /*
1017                  * assume we lost track of the stream, check alignment, and
1018                  * increment a single u64 in the hope to catch on again 'soon'.
1019                  */
1020                 if (unlikely(head & 7))
1021                         head &= ~7ULL;
1022
1023                 size = 8;
1024         }
1025
1026         head += size;
1027         file_pos += size;
1028
1029         if (file_pos >= progress_next) {
1030                 progress_next += file_size / 16;
1031                 ui_progress__update(progress, file_pos);
1032         }
1033
1034         if (file_pos < file_size)
1035                 goto more;
1036
1037         err = 0;
1038         /* do the final flush for ordered samples */
1039         session->ordered_samples.next_flush = ULLONG_MAX;
1040         flush_sample_queue(session, ops);
1041 out_err:
1042         ui_progress__delete(progress);
1043         perf_session__warn_about_errors(session, ops);
1044         perf_session_free_sample_buffers(session);
1045         return err;
1046 }
1047
1048 int perf_session__process_events(struct perf_session *self,
1049                                  struct perf_event_ops *ops)
1050 {
1051         int err;
1052
1053         if (perf_session__register_idle_thread(self) == NULL)
1054                 return -ENOMEM;
1055
1056         if (!self->fd_pipe)
1057                 err = __perf_session__process_events(self,
1058                                                      self->header.data_offset,
1059                                                      self->header.data_size,
1060                                                      self->size, ops);
1061         else
1062                 err = __perf_session__process_pipe_events(self, ops);
1063
1064         return err;
1065 }
1066
1067 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1068 {
1069         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1070                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1071                 return false;
1072         }
1073
1074         return true;
1075 }
1076
1077 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1078                                              const char *symbol_name,
1079                                              u64 addr)
1080 {
1081         char *bracket;
1082         enum map_type i;
1083         struct ref_reloc_sym *ref;
1084
1085         ref = zalloc(sizeof(struct ref_reloc_sym));
1086         if (ref == NULL)
1087                 return -ENOMEM;
1088
1089         ref->name = strdup(symbol_name);
1090         if (ref->name == NULL) {
1091                 free(ref);
1092                 return -ENOMEM;
1093         }
1094
1095         bracket = strchr(ref->name, ']');
1096         if (bracket)
1097                 *bracket = '\0';
1098
1099         ref->addr = addr;
1100
1101         for (i = 0; i < MAP__NR_TYPES; ++i) {
1102                 struct kmap *kmap = map__kmap(maps[i]);
1103                 kmap->ref_reloc_sym = ref;
1104         }
1105
1106         return 0;
1107 }
1108
1109 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1110 {
1111         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1112                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1113                machines__fprintf_dsos(&self->machines, fp);
1114 }
1115
1116 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1117                                           bool with_hits)
1118 {
1119         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1120         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1121 }
1122
1123 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1124 {
1125         struct perf_evsel *pos;
1126         size_t ret = fprintf(fp, "Aggregated stats:\n");
1127
1128         ret += hists__fprintf_nr_events(&session->hists, fp);
1129
1130         list_for_each_entry(pos, &session->evlist->entries, node) {
1131                 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1132                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1133         }
1134
1135         return ret;
1136 }