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 #include "cpumap.h"
  16
  17 static int perf_session__open(struct perf_session *self, bool force)
  18 {
  19         struct stat input_stat;
  20
  21         if (!strcmp(self->filename, "-")) {
  22                 self->fd_pipe = true;
  23                 self->fd = STDIN_FILENO;
  24
  25                 if (perf_session__read_header(self, self->fd) < 0)
  26                         pr_err("incompatible file format");
  27
  28                 return 0;
  29         }
  30
  31         self->fd = open(self->filename, O_RDONLY);
  32         if (self->fd < 0) {
  33                 int err = errno;
  34
  35                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  36                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  37                         pr_err("  (try 'perf record' first)");
  38                 pr_err("\n");
  39                 return -errno;
  40         }
  41
  42         if (fstat(self->fd, &input_stat) < 0)
  43                 goto out_close;
  44
  45         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  46                 pr_err("file %s not owned by current user or root\n",
  47                        self->filename);
  48                 goto out_close;
  49         }
  50
  51         if (!input_stat.st_size) {
  52                 pr_info("zero-sized file (%s), nothing to do!\n",
  53                         self->filename);
  54                 goto out_close;
  55         }
  56
  57         if (perf_session__read_header(self, self->fd) < 0) {
  58                 pr_err("incompatible file format");
  59                 goto out_close;
  60         }
  61
  62         if (!perf_evlist__valid_sample_type(self->evlist)) {
  63                 pr_err("non matching sample_type");
  64                 goto out_close;
  65         }
  66
  67         if (!perf_evlist__valid_sample_id_all(self->evlist)) {
  68                 pr_err("non matching sample_id_all");
  69                 goto out_close;
  70         }
  71
  72         self->size = input_stat.st_size;
  73         return 0;
  74
  75 out_close:
  76         close(self->fd);
  77         self->fd = -1;
  78         return -1;
  79 }
  80
  81 void perf_session__update_sample_type(struct perf_session *self)
  82 {
  83         self->sample_type = perf_evlist__sample_type(self->evlist);
  84         self->sample_size = __perf_evsel__sample_size(self->sample_type);
  85         self->sample_id_all = perf_evlist__sample_id_all(self->evlist);
  86         self->id_hdr_size = perf_evlist__id_hdr_size(self->evlist);
  87 }
  88
  89 int perf_session__create_kernel_maps(struct perf_session *self)
  90 {
  91         int ret = machine__create_kernel_maps(&self->host_machine);
  92
  93         if (ret >= 0)
  94                 ret = machines__create_guest_kernel_maps(&self->machines);
  95         return ret;
  96 }
  97
  98 static void perf_session__destroy_kernel_maps(struct perf_session *self)
  99 {
 100         machine__destroy_kernel_maps(&self->host_machine);
 101         machines__destroy_guest_kernel_maps(&self->machines);
 102 }
 103
 104 struct perf_session *perf_session__new(const char *filename, int mode,
 105                                        bool force, bool repipe,
 106                                        struct perf_event_ops *ops)
 107 {
 108         size_t len = filename ? strlen(filename) + 1 : 0;
 109         struct perf_session *self = zalloc(sizeof(*self) + len);
 110
 111         if (self == NULL)
 112                 goto out;
 113
 114         memcpy(self->filename, filename, len);
 115         /*
 116          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 117          * slices. On 32bit we use 32MB.
 118          */
 119 #if BITS_PER_LONG == 64
 120         self->mmap_window = ULLONG_MAX;
 121 #else
 122         self->mmap_window = 32 * 1024 * 1024ULL;
 123 #endif
 124         self->machines = RB_ROOT;
 125         self->repipe = repipe;
 126         INIT_LIST_HEAD(&self->ordered_samples.samples);
 127         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 128         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 129         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 130
 131         if (mode == O_RDONLY) {
 132                 if (perf_session__open(self, force) < 0)
 133                         goto out_delete;
 134                 perf_session__update_sample_type(self);
 135         } else if (mode == O_WRONLY) {
 136                 /*
 137                  * In O_RDONLY mode this will be performed when reading the
 138                  * kernel MMAP event, in perf_event__process_mmap().
 139                  */
 140                 if (perf_session__create_kernel_maps(self) < 0)
 141                         goto out_delete;
 142         }
 143
 144         if (ops && ops->ordering_requires_timestamps &&
 145             ops->ordered_samples && !self->sample_id_all) {
 146                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 147                 ops->ordered_samples = false;
 148         }
 149
 150 out:
 151         return self;
 152 out_delete:
 153         perf_session__delete(self);
 154         return NULL;
 155 }
 156
 157 static void machine__delete_dead_threads(struct machine *machine)
 158 {
 159         struct thread *n, *t;
 160
 161         list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
 162                 list_del(&t->node);
 163                 thread__delete(t);
 164         }
 165 }
 166
 167 static void perf_session__delete_dead_threads(struct perf_session *session)
 168 {
 169         machine__delete_dead_threads(&session->host_machine);
 170 }
 171
 172 static void machine__delete_threads(struct machine *self)
 173 {
 174         struct rb_node *nd = rb_first(&self->threads);
 175
 176         while (nd) {
 177                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 178
 179                 rb_erase(&t->rb_node, &self->threads);
 180                 nd = rb_next(nd);
 181                 thread__delete(t);
 182         }
 183 }
 184
 185 static void perf_session__delete_threads(struct perf_session *session)
 186 {
 187         machine__delete_threads(&session->host_machine);
 188 }
 189
 190 void perf_session__delete(struct perf_session *self)
 191 {
 192         perf_session__destroy_kernel_maps(self);
 193         perf_session__delete_dead_threads(self);
 194         perf_session__delete_threads(self);
 195         machine__exit(&self->host_machine);
 196         close(self->fd);
 197         free(self);
 198 }
 199
 200 void machine__remove_thread(struct machine *self, struct thread *th)
 201 {
 202         self->last_match = NULL;
 203         rb_erase(&th->rb_node, &self->threads);
 204         /*
 205          * We may have references to this thread, for instance in some hist_entry
 206          * instances, so just move them to a separate list.
 207          */
 208         list_add_tail(&th->node, &self->dead_threads);
 209 }
 210
 211 static bool symbol__match_parent_regex(struct symbol *sym)
 212 {
 213         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 214                 return 1;
 215
 216         return 0;
 217 }
 218
 219 int perf_session__resolve_callchain(struct perf_session *self,
 220                                     struct thread *thread,
 221                                     struct ip_callchain *chain,
 222                                     struct symbol **parent)
 223 {
 224         u8 cpumode = PERF_RECORD_MISC_USER;
 225         unsigned int i;
 226         int err;
 227
 228         callchain_cursor_reset(&self->callchain_cursor);
 229
 230         for (i = 0; i < chain->nr; i++) {
 231                 u64 ip;
 232                 struct addr_location al;
 233
 234                 if (callchain_param.order == ORDER_CALLEE)
 235                         ip = chain->ips[i];
 236                 else
 237                         ip = chain->ips[chain->nr - i - 1];
 238
 239                 if (ip >= PERF_CONTEXT_MAX) {
 240                         switch (ip) {
 241                         case PERF_CONTEXT_HV:
 242                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 243                         case PERF_CONTEXT_KERNEL:
 244                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 245                         case PERF_CONTEXT_USER:
 246                                 cpumode = PERF_RECORD_MISC_USER;        break;
 247                         default:
 248                                 break;
 249                         }
 250                         continue;
 251                 }
 252
 253                 al.filtered = false;
 254                 thread__find_addr_location(thread, self, cpumode,
 255                                 MAP__FUNCTION, thread->pid, ip, &al, NULL);
 256                 if (al.sym != NULL) {
 257                         if (sort__has_parent && !*parent &&
 258                             symbol__match_parent_regex(al.sym))
 259                                 *parent = al.sym;
 260                         if (!symbol_conf.use_callchain)
 261                                 break;
 262                 }
 263
 264                 err = callchain_cursor_append(&self->callchain_cursor,
 265                                               ip, al.map, al.sym);
 266                 if (err)
 267                         return err;
 268         }
 269
 270         return 0;
 271 }
 272
 273 static int process_event_synth_stub(union perf_event *event __used,
 274                                     struct perf_session *session __used)
 275 {
 276         dump_printf(": unhandled!\n");
 277         return 0;
 278 }
 279
 280 static int process_event_synth_attr_stub(union perf_event *event __used,
 281                                          struct perf_evlist **pevlist __used)
 282 {
 283         dump_printf(": unhandled!\n");
 284         return 0;
 285 }
 286
 287 static int process_event_sample_stub(union perf_event *event __used,
 288                                      struct perf_sample *sample __used,
 289                                      struct perf_evsel *evsel __used,
 290                                      struct perf_session *session __used)
 291 {
 292         dump_printf(": unhandled!\n");
 293         return 0;
 294 }
 295
 296 static int process_event_stub(union perf_event *event __used,
 297                               struct perf_sample *sample __used,
 298                               struct perf_session *session __used)
 299 {
 300         dump_printf(": unhandled!\n");
 301         return 0;
 302 }
 303
 304 static int process_finished_round_stub(union perf_event *event __used,
 305                                        struct perf_session *session __used,
 306                                        struct perf_event_ops *ops __used)
 307 {
 308         dump_printf(": unhandled!\n");
 309         return 0;
 310 }
 311
 312 static int process_finished_round(union perf_event *event,
 313                                   struct perf_session *session,
 314                                   struct perf_event_ops *ops);
 315
 316 static void perf_event_ops__fill_defaults(struct perf_event_ops *handler)
 317 {
 318         if (handler->sample == NULL)
 319                 handler->sample = process_event_sample_stub;
 320         if (handler->mmap == NULL)
 321                 handler->mmap = process_event_stub;
 322         if (handler->comm == NULL)
 323                 handler->comm = process_event_stub;
 324         if (handler->fork == NULL)
 325                 handler->fork = process_event_stub;
 326         if (handler->exit == NULL)
 327                 handler->exit = process_event_stub;
 328         if (handler->lost == NULL)
 329                 handler->lost = perf_event__process_lost;
 330         if (handler->read == NULL)
 331                 handler->read = process_event_stub;
 332         if (handler->throttle == NULL)
 333                 handler->throttle = process_event_stub;
 334         if (handler->unthrottle == NULL)
 335                 handler->unthrottle = process_event_stub;
 336         if (handler->attr == NULL)
 337                 handler->attr = process_event_synth_attr_stub;
 338         if (handler->event_type == NULL)
 339                 handler->event_type = process_event_synth_stub;
 340         if (handler->tracing_data == NULL)
 341                 handler->tracing_data = process_event_synth_stub;
 342         if (handler->build_id == NULL)
 343                 handler->build_id = process_event_synth_stub;
 344         if (handler->finished_round == NULL) {
 345                 if (handler->ordered_samples)
 346                         handler->finished_round = process_finished_round;
 347                 else
 348                         handler->finished_round = process_finished_round_stub;
 349         }
 350 }
 351
 352 void mem_bswap_64(void *src, int byte_size)
 353 {
 354         u64 *m = src;
 355
 356         while (byte_size > 0) {
 357                 *m = bswap_64(*m);
 358                 byte_size -= sizeof(u64);
 359                 ++m;
 360         }
 361 }
 362
 363 static void perf_event__all64_swap(union perf_event *event)
 364 {
 365         struct perf_event_header *hdr = &event->header;
 366         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 367 }
 368
 369 static void perf_event__comm_swap(union perf_event *event)
 370 {
 371         event->comm.pid = bswap_32(event->comm.pid);
 372         event->comm.tid = bswap_32(event->comm.tid);
 373 }
 374
 375 static void perf_event__mmap_swap(union perf_event *event)
 376 {
 377         event->mmap.pid   = bswap_32(event->mmap.pid);
 378         event->mmap.tid   = bswap_32(event->mmap.tid);
 379         event->mmap.start = bswap_64(event->mmap.start);
 380         event->mmap.len   = bswap_64(event->mmap.len);
 381         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 382 }
 383
 384 static void perf_event__task_swap(union perf_event *event)
 385 {
 386         event->fork.pid  = bswap_32(event->fork.pid);
 387         event->fork.tid  = bswap_32(event->fork.tid);
 388         event->fork.ppid = bswap_32(event->fork.ppid);
 389         event->fork.ptid = bswap_32(event->fork.ptid);
 390         event->fork.time = bswap_64(event->fork.time);
 391 }
 392
 393 static void perf_event__read_swap(union perf_event *event)
 394 {
 395         event->read.pid          = bswap_32(event->read.pid);
 396         event->read.tid          = bswap_32(event->read.tid);
 397         event->read.value        = bswap_64(event->read.value);
 398         event->read.time_enabled = bswap_64(event->read.time_enabled);
 399         event->read.time_running = bswap_64(event->read.time_running);
 400         event->read.id           = bswap_64(event->read.id);
 401 }
 402
 403 /* exported for swapping attributes in file header */
 404 void perf_event__attr_swap(struct perf_event_attr *attr)
 405 {
 406         attr->type              = bswap_32(attr->type);
 407         attr->size              = bswap_32(attr->size);
 408         attr->config            = bswap_64(attr->config);
 409         attr->sample_period     = bswap_64(attr->sample_period);
 410         attr->sample_type       = bswap_64(attr->sample_type);
 411         attr->read_format       = bswap_64(attr->read_format);
 412         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 413         attr->bp_type           = bswap_32(attr->bp_type);
 414         attr->bp_addr           = bswap_64(attr->bp_addr);
 415         attr->bp_len            = bswap_64(attr->bp_len);
 416 }
 417
 418 static void perf_event__hdr_attr_swap(union perf_event *event)
 419 {
 420         size_t size;
 421
 422         perf_event__attr_swap(&event->attr.attr);
 423
 424         size = event->header.size;
 425         size -= (void *)&event->attr.id - (void *)event;
 426         mem_bswap_64(event->attr.id, size);
 427 }
 428
 429 static void perf_event__event_type_swap(union perf_event *event)
 430 {
 431         event->event_type.event_type.event_id =
 432                 bswap_64(event->event_type.event_type.event_id);
 433 }
 434
 435 static void perf_event__tracing_data_swap(union perf_event *event)
 436 {
 437         event->tracing_data.size = bswap_32(event->tracing_data.size);
 438 }
 439
 440 typedef void (*perf_event__swap_op)(union perf_event *event);
 441
 442 static perf_event__swap_op perf_event__swap_ops[] = {
 443         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 444         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 445         [PERF_RECORD_FORK]                = perf_event__task_swap,
 446         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 447         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 448         [PERF_RECORD_READ]                = perf_event__read_swap,
 449         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 450         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 451         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 452         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 453         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 454         [PERF_RECORD_HEADER_MAX]          = NULL,
 455 };
 456
 457 struct sample_queue {
 458         u64                     timestamp;
 459         u64                     file_offset;
 460         union perf_event        *event;
 461         struct list_head        list;
 462 };
 463
 464 static void perf_session_free_sample_buffers(struct perf_session *session)
 465 {
 466         struct ordered_samples *os = &session->ordered_samples;
 467
 468         while (!list_empty(&os->to_free)) {
 469                 struct sample_queue *sq;
 470
 471                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 472                 list_del(&sq->list);
 473                 free(sq);
 474         }
 475 }
 476
 477 static int perf_session_deliver_event(struct perf_session *session,
 478                                       union perf_event *event,
 479                                       struct perf_sample *sample,
 480                                       struct perf_event_ops *ops,
 481                                       u64 file_offset);
 482
 483 static void flush_sample_queue(struct perf_session *s,
 484                                struct perf_event_ops *ops)
 485 {
 486         struct ordered_samples *os = &s->ordered_samples;
 487         struct list_head *head = &os->samples;
 488         struct sample_queue *tmp, *iter;
 489         struct perf_sample sample;
 490         u64 limit = os->next_flush;
 491         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 492         unsigned idx = 0, progress_next = os->nr_samples / 16;
 493         int ret;
 494
 495         if (!ops->ordered_samples || !limit)
 496                 return;
 497
 498         list_for_each_entry_safe(iter, tmp, head, list) {
 499                 if (iter->timestamp > limit)
 500                         break;
 501
 502                 ret = perf_session__parse_sample(s, iter->event, &sample);
 503                 if (ret)
 504                         pr_err("Can't parse sample, err = %d\n", ret);
 505                 else
 506                         perf_session_deliver_event(s, iter->event, &sample, ops,
 507                                                    iter->file_offset);
 508
 509                 os->last_flush = iter->timestamp;
 510                 list_del(&iter->list);
 511                 list_add(&iter->list, &os->sample_cache);
 512                 if (++idx >= progress_next) {
 513                         progress_next += os->nr_samples / 16;
 514                         ui_progress__update(idx, os->nr_samples,
 515                                             "Processing time ordered events...");
 516                 }
 517         }
 518
 519         if (list_empty(head)) {
 520                 os->last_sample = NULL;
 521         } else if (last_ts <= limit) {
 522                 os->last_sample =
 523                         list_entry(head->prev, struct sample_queue, list);
 524         }
 525
 526         os->nr_samples = 0;
 527 }
 528
 529 /*
 530  * When perf record finishes a pass on every buffers, it records this pseudo
 531  * event.
 532  * We record the max timestamp t found in the pass n.
 533  * Assuming these timestamps are monotonic across cpus, we know that if
 534  * a buffer still has events with timestamps below t, they will be all
 535  * available and then read in the pass n + 1.
 536  * Hence when we start to read the pass n + 2, we can safely flush every
 537  * events with timestamps below t.
 538  *
 539  *    ============ PASS n =================
 540  *       CPU 0         |   CPU 1
 541  *                     |
 542  *    cnt1 timestamps  |   cnt2 timestamps
 543  *          1          |         2
 544  *          2          |         3
 545  *          -          |         4  <--- max recorded
 546  *
 547  *    ============ PASS n + 1 ==============
 548  *       CPU 0         |   CPU 1
 549  *                     |
 550  *    cnt1 timestamps  |   cnt2 timestamps
 551  *          3          |         5
 552  *          4          |         6
 553  *          5          |         7 <---- max recorded
 554  *
 555  *      Flush every events below timestamp 4
 556  *
 557  *    ============ PASS n + 2 ==============
 558  *       CPU 0         |   CPU 1
 559  *                     |
 560  *    cnt1 timestamps  |   cnt2 timestamps
 561  *          6          |         8
 562  *          7          |         9
 563  *          -          |         10
 564  *
 565  *      Flush every events below timestamp 7
 566  *      etc...
 567  */
 568 static int process_finished_round(union perf_event *event __used,
 569                                   struct perf_session *session,
 570                                   struct perf_event_ops *ops)
 571 {
 572         flush_sample_queue(session, ops);
 573         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 574
 575         return 0;
 576 }
 577
 578 /* The queue is ordered by time */
 579 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 580 {
 581         struct ordered_samples *os = &s->ordered_samples;
 582         struct sample_queue *sample = os->last_sample;
 583         u64 timestamp = new->timestamp;
 584         struct list_head *p;
 585
 586         ++os->nr_samples;
 587         os->last_sample = new;
 588
 589         if (!sample) {
 590                 list_add(&new->list, &os->samples);
 591                 os->max_timestamp = timestamp;
 592                 return;
 593         }
 594
 595         /*
 596          * last_sample might point to some random place in the list as it's
 597          * the last queued event. We expect that the new event is close to
 598          * this.
 599          */
 600         if (sample->timestamp <= timestamp) {
 601                 while (sample->timestamp <= timestamp) {
 602                         p = sample->list.next;
 603                         if (p == &os->samples) {
 604                                 list_add_tail(&new->list, &os->samples);
 605                                 os->max_timestamp = timestamp;
 606                                 return;
 607                         }
 608                         sample = list_entry(p, struct sample_queue, list);
 609                 }
 610                 list_add_tail(&new->list, &sample->list);
 611         } else {
 612                 while (sample->timestamp > timestamp) {
 613                         p = sample->list.prev;
 614                         if (p == &os->samples) {
 615                                 list_add(&new->list, &os->samples);
 616                                 return;
 617                         }
 618                         sample = list_entry(p, struct sample_queue, list);
 619                 }
 620                 list_add(&new->list, &sample->list);
 621         }
 622 }
 623
 624 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 625
 626 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 627                                     struct perf_sample *sample, u64 file_offset)
 628 {
 629         struct ordered_samples *os = &s->ordered_samples;
 630         struct list_head *sc = &os->sample_cache;
 631         u64 timestamp = sample->time;
 632         struct sample_queue *new;
 633
 634         if (!timestamp || timestamp == ~0ULL)
 635                 return -ETIME;
 636
 637         if (timestamp < s->ordered_samples.last_flush) {
 638                 printf("Warning: Timestamp below last timeslice flush\n");
 639                 return -EINVAL;
 640         }
 641
 642         if (!list_empty(sc)) {
 643                 new = list_entry(sc->next, struct sample_queue, list);
 644                 list_del(&new->list);
 645         } else if (os->sample_buffer) {
 646                 new = os->sample_buffer + os->sample_buffer_idx;
 647                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 648                         os->sample_buffer = NULL;
 649         } else {
 650                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 651                 if (!os->sample_buffer)
 652                         return -ENOMEM;
 653                 list_add(&os->sample_buffer->list, &os->to_free);
 654                 os->sample_buffer_idx = 2;
 655                 new = os->sample_buffer + 1;
 656         }
 657
 658         new->timestamp = timestamp;
 659         new->file_offset = file_offset;
 660         new->event = event;
 661
 662         __queue_event(new, s);
 663
 664         return 0;
 665 }
 666
 667 static void callchain__printf(struct perf_sample *sample)
 668 {
 669         unsigned int i;
 670
 671         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 672
 673         for (i = 0; i < sample->callchain->nr; i++)
 674                 printf("..... %2d: %016" PRIx64 "\n",
 675                        i, sample->callchain->ips[i]);
 676 }
 677
 678 static void perf_session__print_tstamp(struct perf_session *session,
 679                                        union perf_event *event,
 680                                        struct perf_sample *sample)
 681 {
 682         if (event->header.type != PERF_RECORD_SAMPLE &&
 683             !session->sample_id_all) {
 684                 fputs("-1 -1 ", stdout);
 685                 return;
 686         }
 687
 688         if ((session->sample_type & PERF_SAMPLE_CPU))
 689                 printf("%u ", sample->cpu);
 690
 691         if (session->sample_type & PERF_SAMPLE_TIME)
 692                 printf("%" PRIu64 " ", sample->time);
 693 }
 694
 695 static void dump_event(struct perf_session *session, union perf_event *event,
 696                        u64 file_offset, struct perf_sample *sample)
 697 {
 698         if (!dump_trace)
 699                 return;
 700
 701         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 702                file_offset, event->header.size, event->header.type);
 703
 704         trace_event(event);
 705
 706         if (sample)
 707                 perf_session__print_tstamp(session, event, sample);
 708
 709         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 710                event->header.size, perf_event__name(event->header.type));
 711 }
 712
 713 static void dump_sample(struct perf_session *session, union perf_event *event,
 714                         struct perf_sample *sample)
 715 {
 716         if (!dump_trace)
 717                 return;
 718
 719         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 720                event->header.misc, sample->pid, sample->tid, sample->ip,
 721                sample->period, sample->addr);
 722
 723         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 724                 callchain__printf(sample);
 725 }
 726
 727 static int perf_session_deliver_event(struct perf_session *session,
 728                                       union perf_event *event,
 729                                       struct perf_sample *sample,
 730                                       struct perf_event_ops *ops,
 731                                       u64 file_offset)
 732 {
 733         struct perf_evsel *evsel;
 734
 735         dump_event(session, event, file_offset, sample);
 736
 737         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 738         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
 739                 /*
 740                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
 741                  * because the tools right now may apply filters, discarding
 742                  * some of the samples. For consistency, in the future we
 743                  * should have something like nr_filtered_samples and remove
 744                  * the sample->period from total_sample_period, etc, KISS for
 745                  * now tho.
 746                  *
 747                  * Also testing against NULL allows us to handle files without
 748                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
 749                  * future probably it'll be a good idea to restrict event
 750                  * processing via perf_session to files with both set.
 751                  */
 752                 hists__inc_nr_events(&evsel->hists, event->header.type);
 753         }
 754
 755         switch (event->header.type) {
 756         case PERF_RECORD_SAMPLE:
 757                 dump_sample(session, event, sample);
 758                 if (evsel == NULL) {
 759                         ++session->hists.stats.nr_unknown_id;
 760                         return -1;
 761                 }
 762                 return ops->sample(event, sample, evsel, session);
 763         case PERF_RECORD_MMAP:
 764                 return ops->mmap(event, sample, session);
 765         case PERF_RECORD_COMM:
 766                 return ops->comm(event, sample, session);
 767         case PERF_RECORD_FORK:
 768                 return ops->fork(event, sample, session);
 769         case PERF_RECORD_EXIT:
 770                 return ops->exit(event, sample, session);
 771         case PERF_RECORD_LOST:
 772                 return ops->lost(event, sample, session);
 773         case PERF_RECORD_READ:
 774                 return ops->read(event, sample, session);
 775         case PERF_RECORD_THROTTLE:
 776                 return ops->throttle(event, sample, session);
 777         case PERF_RECORD_UNTHROTTLE:
 778                 return ops->unthrottle(event, sample, session);
 779         default:
 780                 ++session->hists.stats.nr_unknown_events;
 781                 return -1;
 782         }
 783 }
 784
 785 static int perf_session__preprocess_sample(struct perf_session *session,
 786                                            union perf_event *event, struct perf_sample *sample)
 787 {
 788         if (event->header.type != PERF_RECORD_SAMPLE ||
 789             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
 790                 return 0;
 791
 792         if (!ip_callchain__valid(sample->callchain, event)) {
 793                 pr_debug("call-chain problem with event, skipping it.\n");
 794                 ++session->hists.stats.nr_invalid_chains;
 795                 session->hists.stats.total_invalid_chains += sample->period;
 796                 return -EINVAL;
 797         }
 798         return 0;
 799 }
 800
 801 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
 802                                             struct perf_event_ops *ops, u64 file_offset)
 803 {
 804         int err;
 805
 806         dump_event(session, event, file_offset, NULL);
 807
 808         /* These events are processed right away */
 809         switch (event->header.type) {
 810         case PERF_RECORD_HEADER_ATTR:
 811                 err = ops->attr(event, &session->evlist);
 812                 if (err == 0)
 813                         perf_session__update_sample_type(session);
 814                 return err;
 815         case PERF_RECORD_HEADER_EVENT_TYPE:
 816                 return ops->event_type(event, session);
 817         case PERF_RECORD_HEADER_TRACING_DATA:
 818                 /* setup for reading amidst mmap */
 819                 lseek(session->fd, file_offset, SEEK_SET);
 820                 return ops->tracing_data(event, session);
 821         case PERF_RECORD_HEADER_BUILD_ID:
 822                 return ops->build_id(event, session);
 823         case PERF_RECORD_FINISHED_ROUND:
 824                 return ops->finished_round(event, session, ops);
 825         default:
 826                 return -EINVAL;
 827         }
 828 }
 829
 830 static int perf_session__process_event(struct perf_session *session,
 831                                        union perf_event *event,
 832                                        struct perf_event_ops *ops,
 833                                        u64 file_offset)
 834 {
 835         struct perf_sample sample;
 836         int ret;
 837
 838         if (session->header.needs_swap &&
 839             perf_event__swap_ops[event->header.type])
 840                 perf_event__swap_ops[event->header.type](event);
 841
 842         if (event->header.type >= PERF_RECORD_HEADER_MAX)
 843                 return -EINVAL;
 844
 845         hists__inc_nr_events(&session->hists, event->header.type);
 846
 847         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
 848                 return perf_session__process_user_event(session, event, ops, file_offset);
 849
 850         /*
 851          * For all kernel events we get the sample data
 852          */
 853         ret = perf_session__parse_sample(session, event, &sample);
 854         if (ret)
 855                 return ret;
 856
 857         /* Preprocess sample records - precheck callchains */
 858         if (perf_session__preprocess_sample(session, event, &sample))
 859                 return 0;
 860
 861         if (ops->ordered_samples) {
 862                 ret = perf_session_queue_event(session, event, &sample,
 863                                                file_offset);
 864                 if (ret != -ETIME)
 865                         return ret;
 866         }
 867
 868         return perf_session_deliver_event(session, event, &sample, ops,
 869                                           file_offset);
 870 }
 871
 872 void perf_event_header__bswap(struct perf_event_header *self)
 873 {
 874         self->type = bswap_32(self->type);
 875         self->misc = bswap_16(self->misc);
 876         self->size = bswap_16(self->size);
 877 }
 878
 879 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
 880 {
 881         return machine__findnew_thread(&session->host_machine, pid);
 882 }
 883
 884 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
 885 {
 886         struct thread *thread = perf_session__findnew(self, 0);
 887
 888         if (thread == NULL || thread__set_comm(thread, "swapper")) {
 889                 pr_err("problem inserting idle task.\n");
 890                 thread = NULL;
 891         }
 892
 893         return thread;
 894 }
 895
 896 static void perf_session__warn_about_errors(const struct perf_session *session,
 897                                             const struct perf_event_ops *ops)
 898 {
 899         if (ops->lost == perf_event__process_lost &&
 900             session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
 901                 ui__warning("Processed %d events and lost %d chunks!\n\n"
 902                             "Check IO/CPU overload!\n\n",
 903                             session->hists.stats.nr_events[0],
 904                             session->hists.stats.nr_events[PERF_RECORD_LOST]);
 905         }
 906
 907         if (session->hists.stats.nr_unknown_events != 0) {
 908                 ui__warning("Found %u unknown events!\n\n"
 909                             "Is this an older tool processing a perf.data "
 910                             "file generated by a more recent tool?\n\n"
 911                             "If that is not the case, consider "
 912                             "reporting to linux-kernel@vger.kernel.org.\n\n",
 913                             session->hists.stats.nr_unknown_events);
 914         }
 915
 916         if (session->hists.stats.nr_unknown_id != 0) {
 917                 ui__warning("%u samples with id not present in the header\n",
 918                             session->hists.stats.nr_unknown_id);
 919         }
 920
 921         if (session->hists.stats.nr_invalid_chains != 0) {
 922                 ui__warning("Found invalid callchains!\n\n"
 923                             "%u out of %u events were discarded for this reason.\n\n"
 924                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
 925                             session->hists.stats.nr_invalid_chains,
 926                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
 927         }
 928 }
 929
 930 #define session_done()  (*(volatile int *)(&session_done))
 931 volatile int session_done;
 932
 933 static int __perf_session__process_pipe_events(struct perf_session *self,
 934                                                struct perf_event_ops *ops)
 935 {
 936         union perf_event event;
 937         uint32_t size;
 938         int skip = 0;
 939         u64 head;
 940         int err;
 941         void *p;
 942
 943         perf_event_ops__fill_defaults(ops);
 944
 945         head = 0;
 946 more:
 947         err = readn(self->fd, &event, sizeof(struct perf_event_header));
 948         if (err <= 0) {
 949                 if (err == 0)
 950                         goto done;
 951
 952                 pr_err("failed to read event header\n");
 953                 goto out_err;
 954         }
 955
 956         if (self->header.needs_swap)
 957                 perf_event_header__bswap(&event.header);
 958
 959         size = event.header.size;
 960         if (size == 0)
 961                 size = 8;
 962
 963         p = &event;
 964         p += sizeof(struct perf_event_header);
 965
 966         if (size - sizeof(struct perf_event_header)) {
 967                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
 968                 if (err <= 0) {
 969                         if (err == 0) {
 970                                 pr_err("unexpected end of event stream\n");
 971                                 goto done;
 972                         }
 973
 974                         pr_err("failed to read event data\n");
 975                         goto out_err;
 976                 }
 977         }
 978
 979         if (size == 0 ||
 980             (skip = perf_session__process_event(self, &event, ops, head)) < 0) {
 981                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
 982                             head, event.header.size, event.header.type);
 983                 /*
 984                  * assume we lost track of the stream, check alignment, and
 985                  * increment a single u64 in the hope to catch on again 'soon'.
 986                  */
 987                 if (unlikely(head & 7))
 988                         head &= ~7ULL;
 989
 990                 size = 8;
 991         }
 992
 993         head += size;
 994
 995         if (skip > 0)
 996                 head += skip;
 997
 998         if (!session_done())
 999                 goto more;
1000 done:
1001         err = 0;
1002 out_err:
1003         perf_session__warn_about_errors(self, ops);
1004         perf_session_free_sample_buffers(self);
1005         return err;
1006 }
1007
1008 static union perf_event *
1009 fetch_mmaped_event(struct perf_session *session,
1010                    u64 head, size_t mmap_size, char *buf)
1011 {
1012         union perf_event *event;
1013
1014         /*
1015          * Ensure we have enough space remaining to read
1016          * the size of the event in the headers.
1017          */
1018         if (head + sizeof(event->header) > mmap_size)
1019                 return NULL;
1020
1021         event = (union perf_event *)(buf + head);
1022
1023         if (session->header.needs_swap)
1024                 perf_event_header__bswap(&event->header);
1025
1026         if (head + event->header.size > mmap_size)
1027                 return NULL;
1028
1029         return event;
1030 }
1031
1032 int __perf_session__process_events(struct perf_session *session,
1033                                    u64 data_offset, u64 data_size,
1034                                    u64 file_size, struct perf_event_ops *ops)
1035 {
1036         u64 head, page_offset, file_offset, file_pos, progress_next;
1037         int err, mmap_prot, mmap_flags, map_idx = 0;
1038         size_t  page_size, mmap_size;
1039         char *buf, *mmaps[8];
1040         union perf_event *event;
1041         uint32_t size;
1042
1043         perf_event_ops__fill_defaults(ops);
1044
1045         page_size = sysconf(_SC_PAGESIZE);
1046
1047         page_offset = page_size * (data_offset / page_size);
1048         file_offset = page_offset;
1049         head = data_offset - page_offset;
1050
1051         if (data_offset + data_size < file_size)
1052                 file_size = data_offset + data_size;
1053
1054         progress_next = file_size / 16;
1055
1056         mmap_size = session->mmap_window;
1057         if (mmap_size > file_size)
1058                 mmap_size = file_size;
1059
1060         memset(mmaps, 0, sizeof(mmaps));
1061
1062         mmap_prot  = PROT_READ;
1063         mmap_flags = MAP_SHARED;
1064
1065         if (session->header.needs_swap) {
1066                 mmap_prot  |= PROT_WRITE;
1067                 mmap_flags = MAP_PRIVATE;
1068         }
1069 remap:
1070         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1071                    file_offset);
1072         if (buf == MAP_FAILED) {
1073                 pr_err("failed to mmap file\n");
1074                 err = -errno;
1075                 goto out_err;
1076         }
1077         mmaps[map_idx] = buf;
1078         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1079         file_pos = file_offset + head;
1080
1081 more:
1082         event = fetch_mmaped_event(session, head, mmap_size, buf);
1083         if (!event) {
1084                 if (mmaps[map_idx]) {
1085                         munmap(mmaps[map_idx], mmap_size);
1086                         mmaps[map_idx] = NULL;
1087                 }
1088
1089                 page_offset = page_size * (head / page_size);
1090                 file_offset += page_offset;
1091                 head -= page_offset;
1092                 goto remap;
1093         }
1094
1095         size = event->header.size;
1096
1097         if (size == 0 ||
1098             perf_session__process_event(session, event, ops, file_pos) < 0) {
1099                 dump_printf("%#" PRIx64 " [%#x]: skipping unknown header type: %d\n",
1100                             file_offset + head, event->header.size,
1101                             event->header.type);
1102                 /*
1103                  * assume we lost track of the stream, check alignment, and
1104                  * increment a single u64 in the hope to catch on again 'soon'.
1105                  */
1106                 if (unlikely(head & 7))
1107                         head &= ~7ULL;
1108
1109                 size = 8;
1110         }
1111
1112         head += size;
1113         file_pos += size;
1114
1115         if (file_pos >= progress_next) {
1116                 progress_next += file_size / 16;
1117                 ui_progress__update(file_pos, file_size,
1118                                     "Processing events...");
1119         }
1120
1121         if (file_pos < file_size)
1122                 goto more;
1123
1124         err = 0;
1125         /* do the final flush for ordered samples */
1126         session->ordered_samples.next_flush = ULLONG_MAX;
1127         flush_sample_queue(session, ops);
1128 out_err:
1129         perf_session__warn_about_errors(session, ops);
1130         perf_session_free_sample_buffers(session);
1131         return err;
1132 }
1133
1134 int perf_session__process_events(struct perf_session *self,
1135                                  struct perf_event_ops *ops)
1136 {
1137         int err;
1138
1139         if (perf_session__register_idle_thread(self) == NULL)
1140                 return -ENOMEM;
1141
1142         if (!self->fd_pipe)
1143                 err = __perf_session__process_events(self,
1144                                                      self->header.data_offset,
1145                                                      self->header.data_size,
1146                                                      self->size, ops);
1147         else
1148                 err = __perf_session__process_pipe_events(self, ops);
1149
1150         return err;
1151 }
1152
1153 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1154 {
1155         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1156                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1157                 return false;
1158         }
1159
1160         return true;
1161 }
1162
1163 int perf_session__set_kallsyms_ref_reloc_sym(struct map **maps,
1164                                              const char *symbol_name,
1165                                              u64 addr)
1166 {
1167         char *bracket;
1168         enum map_type i;
1169         struct ref_reloc_sym *ref;
1170
1171         ref = zalloc(sizeof(struct ref_reloc_sym));
1172         if (ref == NULL)
1173                 return -ENOMEM;
1174
1175         ref->name = strdup(symbol_name);
1176         if (ref->name == NULL) {
1177                 free(ref);
1178                 return -ENOMEM;
1179         }
1180
1181         bracket = strchr(ref->name, ']');
1182         if (bracket)
1183                 *bracket = '\0';
1184
1185         ref->addr = addr;
1186
1187         for (i = 0; i < MAP__NR_TYPES; ++i) {
1188                 struct kmap *kmap = map__kmap(maps[i]);
1189                 kmap->ref_reloc_sym = ref;
1190         }
1191
1192         return 0;
1193 }
1194
1195 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1196 {
1197         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1198                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1199                machines__fprintf_dsos(&self->machines, fp);
1200 }
1201
1202 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1203                                           bool with_hits)
1204 {
1205         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1206         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1207 }
1208
1209 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1210 {
1211         struct perf_evsel *pos;
1212         size_t ret = fprintf(fp, "Aggregated stats:\n");
1213
1214         ret += hists__fprintf_nr_events(&session->hists, fp);
1215
1216         list_for_each_entry(pos, &session->evlist->entries, node) {
1217                 ret += fprintf(fp, "%s stats:\n", event_name(pos));
1218                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1219         }
1220
1221         return ret;
1222 }
1223
1224 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1225 {
1226         /*
1227          * FIXME: Here we have to actually print all the machines in this
1228          * session, not just the host...
1229          */
1230         return machine__fprintf(&session->host_machine, fp);
1231 }
1232
1233 void perf_session__remove_thread(struct perf_session *session,
1234                                  struct thread *th)
1235 {
1236         /*
1237          * FIXME: This one makes no sense, we need to remove the thread from
1238          * the machine it belongs to, perf_session can have many machines, so
1239          * doing it always on ->host_machine is wrong.  Fix when auditing all
1240          * the 'perf kvm' code.
1241          */
1242         machine__remove_thread(&session->host_machine, th);
1243 }
1244
1245 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1246                                               unsigned int type)
1247 {
1248         struct perf_evsel *pos;
1249
1250         list_for_each_entry(pos, &session->evlist->entries, node) {
1251                 if (pos->attr.type == type)
1252                         return pos;
1253         }
1254         return NULL;
1255 }
1256
1257 void perf_session__print_ip(union perf_event *event,
1258                             struct perf_sample *sample,
1259                             struct perf_session *session,
1260                             int print_sym, int print_dso)
1261 {
1262         struct addr_location al;
1263         const char *symname, *dsoname;
1264         struct callchain_cursor *cursor = &session->callchain_cursor;
1265         struct callchain_cursor_node *node;
1266
1267         if (perf_event__preprocess_sample(event, session, &al, sample,
1268                                           NULL) < 0) {
1269                 error("problem processing %d event, skipping it.\n",
1270                         event->header.type);
1271                 return;
1272         }
1273
1274         if (symbol_conf.use_callchain && sample->callchain) {
1275
1276                 if (perf_session__resolve_callchain(session, al.thread,
1277                                                 sample->callchain, NULL) != 0) {
1278                         if (verbose)
1279                                 error("Failed to resolve callchain. Skipping\n");
1280                         return;
1281                 }
1282                 callchain_cursor_commit(cursor);
1283
1284                 while (1) {
1285                         node = callchain_cursor_current(cursor);
1286                         if (!node)
1287                                 break;
1288
1289                         printf("\t%16" PRIx64, node->ip);
1290                         if (print_sym) {
1291                                 if (node->sym && node->sym->name)
1292                                         symname = node->sym->name;
1293                                 else
1294                                         symname = "";
1295
1296                                 printf(" %s", symname);
1297                         }
1298                         if (print_dso) {
1299                                 if (node->map && node->map->dso && node->map->dso->name)
1300                                         dsoname = node->map->dso->name;
1301                                 else
1302                                         dsoname = "";
1303
1304                                 printf(" (%s)", dsoname);
1305                         }
1306                         printf("\n");
1307
1308                         callchain_cursor_advance(cursor);
1309                 }
1310
1311         } else {
1312                 printf("%16" PRIx64, sample->ip);
1313                 if (print_sym) {
1314                         if (al.sym && al.sym->name)
1315                                 symname = al.sym->name;
1316                         else
1317                                 symname = "";
1318
1319                         printf(" %s", symname);
1320                 }
1321
1322                 if (print_dso) {
1323                         if (al.map && al.map->dso && al.map->dso->name)
1324                                 dsoname = al.map->dso->name;
1325                         else
1326                                 dsoname = "";
1327
1328                         printf(" (%s)", dsoname);
1329                 }
1330         }
1331 }
1332
1333 int perf_session__cpu_bitmap(struct perf_session *session,
1334                              const char *cpu_list, unsigned long *cpu_bitmap)
1335 {
1336         int i;
1337         struct cpu_map *map;
1338
1339         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1340                 struct perf_evsel *evsel;
1341
1342                 evsel = perf_session__find_first_evtype(session, i);
1343                 if (!evsel)
1344                         continue;
1345
1346                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1347                         pr_err("File does not contain CPU events. "
1348                                "Remove -c option to proceed.\n");
1349                         return -1;
1350                 }
1351         }
1352
1353         map = cpu_map__new(cpu_list);
1354
1355         for (i = 0; i < map->nr; i++) {
1356                 int cpu = map->map[i];
1357
1358                 if (cpu >= MAX_NR_CPUS) {
1359                         pr_err("Requested CPU %d too large. "
1360                                "Consider raising MAX_NR_CPUS\n", cpu);
1361                         return -1;
1362                 }
1363
1364                 set_bit(cpu, cpu_bitmap);
1365         }
1366
1367         return 0;
1368 }
1369
1370 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1371                                 bool full)
1372 {
1373         struct stat st;
1374         int ret;
1375
1376         if (session == NULL || fp == NULL)
1377                 return;
1378
1379         ret = fstat(session->fd, &st);
1380         if (ret == -1)
1381                 return;
1382
1383         fprintf(fp, "# ========\n");
1384         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1385         perf_header__fprintf_info(session, fp, full);
1386         fprintf(fp, "# ========\n#\n");
1387 }