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 "tool.h"
  14 #include "sort.h"
  15 #include "util.h"
  16 #include "cpumap.h"
  17 #include "event-parse.h"
  18 #include "perf_regs.h"
  19 #include "unwind.h"
  20
  21 static int perf_session__open(struct perf_session *self, bool force)
  22 {
  23         struct stat input_stat;
  24
  25         if (!strcmp(self->filename, "-")) {
  26                 self->fd_pipe = true;
  27                 self->fd = STDIN_FILENO;
  28
  29                 if (perf_session__read_header(self, self->fd) < 0)
  30                         pr_err("incompatible file format (rerun with -v to learn more)");
  31
  32                 return 0;
  33         }
  34
  35         self->fd = open(self->filename, O_RDONLY);
  36         if (self->fd < 0) {
  37                 int err = errno;
  38
  39                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  40                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  41                         pr_err("  (try 'perf record' first)");
  42                 pr_err("\n");
  43                 return -errno;
  44         }
  45
  46         if (fstat(self->fd, &input_stat) < 0)
  47                 goto out_close;
  48
  49         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  50                 pr_err("file %s not owned by current user or root\n",
  51                        self->filename);
  52                 goto out_close;
  53         }
  54
  55         if (!input_stat.st_size) {
  56                 pr_info("zero-sized file (%s), nothing to do!\n",
  57                         self->filename);
  58                 goto out_close;
  59         }
  60
  61         if (perf_session__read_header(self, self->fd) < 0) {
  62                 pr_err("incompatible file format (rerun with -v to learn more)");
  63                 goto out_close;
  64         }
  65
  66         if (!perf_evlist__valid_sample_type(self->evlist)) {
  67                 pr_err("non matching sample_type");
  68                 goto out_close;
  69         }
  70
  71         if (!perf_evlist__valid_sample_id_all(self->evlist)) {
  72                 pr_err("non matching sample_id_all");
  73                 goto out_close;
  74         }
  75
  76         self->size = input_stat.st_size;
  77         return 0;
  78
  79 out_close:
  80         close(self->fd);
  81         self->fd = -1;
  82         return -1;
  83 }
  84
  85 void perf_session__set_id_hdr_size(struct perf_session *session)
  86 {
  87         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
  88
  89         session->host_machine.id_hdr_size = id_hdr_size;
  90         machines__set_id_hdr_size(&session->machines, id_hdr_size);
  91 }
  92
  93 int perf_session__create_kernel_maps(struct perf_session *self)
  94 {
  95         int ret = machine__create_kernel_maps(&self->host_machine);
  96
  97         if (ret >= 0)
  98                 ret = machines__create_guest_kernel_maps(&self->machines);
  99         return ret;
 100 }
 101
 102 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 103 {
 104         machine__destroy_kernel_maps(&self->host_machine);
 105         machines__destroy_guest_kernel_maps(&self->machines);
 106 }
 107
 108 struct perf_session *perf_session__new(const char *filename, int mode,
 109                                        bool force, bool repipe,
 110                                        struct perf_tool *tool)
 111 {
 112         struct perf_session *self;
 113         struct stat st;
 114         size_t len;
 115
 116         if (!filename || !strlen(filename)) {
 117                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 118                         filename = "-";
 119                 else
 120                         filename = "perf.data";
 121         }
 122
 123         len = strlen(filename);
 124         self = zalloc(sizeof(*self) + len);
 125
 126         if (self == NULL)
 127                 goto out;
 128
 129         memcpy(self->filename, filename, len);
 130         /*
 131          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 132          * slices. On 32bit we use 32MB.
 133          */
 134 #if BITS_PER_LONG == 64
 135         self->mmap_window = ULLONG_MAX;
 136 #else
 137         self->mmap_window = 32 * 1024 * 1024ULL;
 138 #endif
 139         self->machines = RB_ROOT;
 140         self->repipe = repipe;
 141         INIT_LIST_HEAD(&self->ordered_samples.samples);
 142         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 143         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 144         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 145         hists__init(&self->hists);
 146
 147         if (mode == O_RDONLY) {
 148                 if (perf_session__open(self, force) < 0)
 149                         goto out_delete;
 150                 perf_session__set_id_hdr_size(self);
 151         } else if (mode == O_WRONLY) {
 152                 /*
 153                  * In O_RDONLY mode this will be performed when reading the
 154                  * kernel MMAP event, in perf_event__process_mmap().
 155                  */
 156                 if (perf_session__create_kernel_maps(self) < 0)
 157                         goto out_delete;
 158         }
 159
 160         if (tool && tool->ordering_requires_timestamps &&
 161             tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
 162                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 163                 tool->ordered_samples = false;
 164         }
 165
 166 out:
 167         return self;
 168 out_delete:
 169         perf_session__delete(self);
 170         return NULL;
 171 }
 172
 173 static void machine__delete_dead_threads(struct machine *machine)
 174 {
 175         struct thread *n, *t;
 176
 177         list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
 178                 list_del(&t->node);
 179                 thread__delete(t);
 180         }
 181 }
 182
 183 static void perf_session__delete_dead_threads(struct perf_session *session)
 184 {
 185         machine__delete_dead_threads(&session->host_machine);
 186 }
 187
 188 static void machine__delete_threads(struct machine *self)
 189 {
 190         struct rb_node *nd = rb_first(&self->threads);
 191
 192         while (nd) {
 193                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 194
 195                 rb_erase(&t->rb_node, &self->threads);
 196                 nd = rb_next(nd);
 197                 thread__delete(t);
 198         }
 199 }
 200
 201 static void perf_session__delete_threads(struct perf_session *session)
 202 {
 203         machine__delete_threads(&session->host_machine);
 204 }
 205
 206 void perf_session__delete(struct perf_session *self)
 207 {
 208         perf_session__destroy_kernel_maps(self);
 209         perf_session__delete_dead_threads(self);
 210         perf_session__delete_threads(self);
 211         machine__exit(&self->host_machine);
 212         close(self->fd);
 213         free(self);
 214 }
 215
 216 void machine__remove_thread(struct machine *self, struct thread *th)
 217 {
 218         self->last_match = NULL;
 219         rb_erase(&th->rb_node, &self->threads);
 220         /*
 221          * We may have references to this thread, for instance in some hist_entry
 222          * instances, so just move them to a separate list.
 223          */
 224         list_add_tail(&th->node, &self->dead_threads);
 225 }
 226
 227 static bool symbol__match_parent_regex(struct symbol *sym)
 228 {
 229         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 230                 return 1;
 231
 232         return 0;
 233 }
 234
 235 static const u8 cpumodes[] = {
 236         PERF_RECORD_MISC_USER,
 237         PERF_RECORD_MISC_KERNEL,
 238         PERF_RECORD_MISC_GUEST_USER,
 239         PERF_RECORD_MISC_GUEST_KERNEL
 240 };
 241 #define NCPUMODES (sizeof(cpumodes)/sizeof(u8))
 242
 243 static void ip__resolve_ams(struct machine *self, struct thread *thread,
 244                             struct addr_map_symbol *ams,
 245                             u64 ip)
 246 {
 247         struct addr_location al;
 248         size_t i;
 249         u8 m;
 250
 251         memset(&al, 0, sizeof(al));
 252
 253         for (i = 0; i < NCPUMODES; i++) {
 254                 m = cpumodes[i];
 255                 /*
 256                  * We cannot use the header.misc hint to determine whether a
 257                  * branch stack address is user, kernel, guest, hypervisor.
 258                  * Branches may straddle the kernel/user/hypervisor boundaries.
 259                  * Thus, we have to try consecutively until we find a match
 260                  * or else, the symbol is unknown
 261                  */
 262                 thread__find_addr_location(thread, self, m, MAP__FUNCTION,
 263                                 ip, &al, NULL);
 264                 if (al.sym)
 265                         goto found;
 266         }
 267 found:
 268         ams->addr = ip;
 269         ams->al_addr = al.addr;
 270         ams->sym = al.sym;
 271         ams->map = al.map;
 272 }
 273
 274 struct branch_info *machine__resolve_bstack(struct machine *self,
 275                                             struct thread *thr,
 276                                             struct branch_stack *bs)
 277 {
 278         struct branch_info *bi;
 279         unsigned int i;
 280
 281         bi = calloc(bs->nr, sizeof(struct branch_info));
 282         if (!bi)
 283                 return NULL;
 284
 285         for (i = 0; i < bs->nr; i++) {
 286                 ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to);
 287                 ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from);
 288                 bi[i].flags = bs->entries[i].flags;
 289         }
 290         return bi;
 291 }
 292
 293 static int machine__resolve_callchain_sample(struct machine *machine,
 294                                              struct thread *thread,
 295                                              struct ip_callchain *chain,
 296                                              struct symbol **parent)
 297
 298 {
 299         u8 cpumode = PERF_RECORD_MISC_USER;
 300         unsigned int i;
 301         int err;
 302
 303         callchain_cursor_reset(&callchain_cursor);
 304
 305         if (chain->nr > PERF_MAX_STACK_DEPTH) {
 306                 pr_warning("corrupted callchain. skipping...\n");
 307                 return 0;
 308         }
 309
 310         for (i = 0; i < chain->nr; i++) {
 311                 u64 ip;
 312                 struct addr_location al;
 313
 314                 if (callchain_param.order == ORDER_CALLEE)
 315                         ip = chain->ips[i];
 316                 else
 317                         ip = chain->ips[chain->nr - i - 1];
 318
 319                 if (ip >= PERF_CONTEXT_MAX) {
 320                         switch (ip) {
 321                         case PERF_CONTEXT_HV:
 322                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;
 323                                 break;
 324                         case PERF_CONTEXT_KERNEL:
 325                                 cpumode = PERF_RECORD_MISC_KERNEL;
 326                                 break;
 327                         case PERF_CONTEXT_USER:
 328                                 cpumode = PERF_RECORD_MISC_USER;
 329                                 break;
 330                         default:
 331                                 pr_debug("invalid callchain context: "
 332                                          "%"PRId64"\n", (s64) ip);
 333                                 /*
 334                                  * It seems the callchain is corrupted.
 335                                  * Discard all.
 336                                  */
 337                                 callchain_cursor_reset(&callchain_cursor);
 338                                 return 0;
 339                         }
 340                         continue;
 341                 }
 342
 343                 al.filtered = false;
 344                 thread__find_addr_location(thread, machine, cpumode,
 345                                            MAP__FUNCTION, ip, &al, NULL);
 346                 if (al.sym != NULL) {
 347                         if (sort__has_parent && !*parent &&
 348                             symbol__match_parent_regex(al.sym))
 349                                 *parent = al.sym;
 350                         if (!symbol_conf.use_callchain)
 351                                 break;
 352                 }
 353
 354                 err = callchain_cursor_append(&callchain_cursor,
 355                                               ip, al.map, al.sym);
 356                 if (err)
 357                         return err;
 358         }
 359
 360         return 0;
 361 }
 362
 363 static int unwind_entry(struct unwind_entry *entry, void *arg)
 364 {
 365         struct callchain_cursor *cursor = arg;
 366         return callchain_cursor_append(cursor, entry->ip,
 367                                        entry->map, entry->sym);
 368 }
 369
 370 int machine__resolve_callchain(struct machine *machine,
 371                                struct perf_evsel *evsel,
 372                                struct thread *thread,
 373                                struct perf_sample *sample,
 374                                struct symbol **parent)
 375
 376 {
 377         int ret;
 378
 379         callchain_cursor_reset(&callchain_cursor);
 380
 381         ret = machine__resolve_callchain_sample(machine, thread,
 382                                                 sample->callchain, parent);
 383         if (ret)
 384                 return ret;
 385
 386         /* Can we do dwarf post unwind? */
 387         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
 388               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
 389                 return 0;
 390
 391         return unwind__get_entries(unwind_entry, &callchain_cursor, machine,
 392                                    thread, evsel->attr.sample_regs_user,
 393                                    sample);
 394
 395 }
 396
 397 static int process_event_synth_tracing_data_stub(union perf_event *event __used,
 398                                                  struct perf_session *session __used)
 399 {
 400         dump_printf(": unhandled!\n");
 401         return 0;
 402 }
 403
 404 static int process_event_synth_attr_stub(union perf_event *event __used,
 405                                          struct perf_evlist **pevlist __used)
 406 {
 407         dump_printf(": unhandled!\n");
 408         return 0;
 409 }
 410
 411 static int process_event_sample_stub(struct perf_tool *tool __used,
 412                                      union perf_event *event __used,
 413                                      struct perf_sample *sample __used,
 414                                      struct perf_evsel *evsel __used,
 415                                      struct machine *machine __used)
 416 {
 417         dump_printf(": unhandled!\n");
 418         return 0;
 419 }
 420
 421 static int process_event_stub(struct perf_tool *tool __used,
 422                               union perf_event *event __used,
 423                               struct perf_sample *sample __used,
 424                               struct machine *machine __used)
 425 {
 426         dump_printf(": unhandled!\n");
 427         return 0;
 428 }
 429
 430 static int process_finished_round_stub(struct perf_tool *tool __used,
 431                                        union perf_event *event __used,
 432                                        struct perf_session *perf_session __used)
 433 {
 434         dump_printf(": unhandled!\n");
 435         return 0;
 436 }
 437
 438 static int process_event_type_stub(struct perf_tool *tool __used,
 439                                    union perf_event *event __used)
 440 {
 441         dump_printf(": unhandled!\n");
 442         return 0;
 443 }
 444
 445 static int process_finished_round(struct perf_tool *tool,
 446                                   union perf_event *event,
 447                                   struct perf_session *session);
 448
 449 static void perf_tool__fill_defaults(struct perf_tool *tool)
 450 {
 451         if (tool->sample == NULL)
 452                 tool->sample = process_event_sample_stub;
 453         if (tool->mmap == NULL)
 454                 tool->mmap = process_event_stub;
 455         if (tool->comm == NULL)
 456                 tool->comm = process_event_stub;
 457         if (tool->fork == NULL)
 458                 tool->fork = process_event_stub;
 459         if (tool->exit == NULL)
 460                 tool->exit = process_event_stub;
 461         if (tool->lost == NULL)
 462                 tool->lost = perf_event__process_lost;
 463         if (tool->read == NULL)
 464                 tool->read = process_event_sample_stub;
 465         if (tool->throttle == NULL)
 466                 tool->throttle = process_event_stub;
 467         if (tool->unthrottle == NULL)
 468                 tool->unthrottle = process_event_stub;
 469         if (tool->attr == NULL)
 470                 tool->attr = process_event_synth_attr_stub;
 471         if (tool->event_type == NULL)
 472                 tool->event_type = process_event_type_stub;
 473         if (tool->tracing_data == NULL)
 474                 tool->tracing_data = process_event_synth_tracing_data_stub;
 475         if (tool->build_id == NULL)
 476                 tool->build_id = process_finished_round_stub;
 477         if (tool->finished_round == NULL) {
 478                 if (tool->ordered_samples)
 479                         tool->finished_round = process_finished_round;
 480                 else
 481                         tool->finished_round = process_finished_round_stub;
 482         }
 483 }
 484
 485 void mem_bswap_32(void *src, int byte_size)
 486 {
 487         u32 *m = src;
 488         while (byte_size > 0) {
 489                 *m = bswap_32(*m);
 490                 byte_size -= sizeof(u32);
 491                 ++m;
 492         }
 493 }
 494
 495 void mem_bswap_64(void *src, int byte_size)
 496 {
 497         u64 *m = src;
 498
 499         while (byte_size > 0) {
 500                 *m = bswap_64(*m);
 501                 byte_size -= sizeof(u64);
 502                 ++m;
 503         }
 504 }
 505
 506 static void swap_sample_id_all(union perf_event *event, void *data)
 507 {
 508         void *end = (void *) event + event->header.size;
 509         int size = end - data;
 510
 511         BUG_ON(size % sizeof(u64));
 512         mem_bswap_64(data, size);
 513 }
 514
 515 static void perf_event__all64_swap(union perf_event *event,
 516                                    bool sample_id_all __used)
 517 {
 518         struct perf_event_header *hdr = &event->header;
 519         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 520 }
 521
 522 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 523 {
 524         event->comm.pid = bswap_32(event->comm.pid);
 525         event->comm.tid = bswap_32(event->comm.tid);
 526
 527         if (sample_id_all) {
 528                 void *data = &event->comm.comm;
 529
 530                 data += ALIGN(strlen(data) + 1, sizeof(u64));
 531                 swap_sample_id_all(event, data);
 532         }
 533 }
 534
 535 static void perf_event__mmap_swap(union perf_event *event,
 536                                   bool sample_id_all)
 537 {
 538         event->mmap.pid   = bswap_32(event->mmap.pid);
 539         event->mmap.tid   = bswap_32(event->mmap.tid);
 540         event->mmap.start = bswap_64(event->mmap.start);
 541         event->mmap.len   = bswap_64(event->mmap.len);
 542         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 543
 544         if (sample_id_all) {
 545                 void *data = &event->mmap.filename;
 546
 547                 data += ALIGN(strlen(data) + 1, sizeof(u64));
 548                 swap_sample_id_all(event, data);
 549         }
 550 }
 551
 552 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 553 {
 554         event->fork.pid  = bswap_32(event->fork.pid);
 555         event->fork.tid  = bswap_32(event->fork.tid);
 556         event->fork.ppid = bswap_32(event->fork.ppid);
 557         event->fork.ptid = bswap_32(event->fork.ptid);
 558         event->fork.time = bswap_64(event->fork.time);
 559
 560         if (sample_id_all)
 561                 swap_sample_id_all(event, &event->fork + 1);
 562 }
 563
 564 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 565 {
 566         event->read.pid          = bswap_32(event->read.pid);
 567         event->read.tid          = bswap_32(event->read.tid);
 568         event->read.value        = bswap_64(event->read.value);
 569         event->read.time_enabled = bswap_64(event->read.time_enabled);
 570         event->read.time_running = bswap_64(event->read.time_running);
 571         event->read.id           = bswap_64(event->read.id);
 572
 573         if (sample_id_all)
 574                 swap_sample_id_all(event, &event->read + 1);
 575 }
 576
 577 static u8 revbyte(u8 b)
 578 {
 579         int rev = (b >> 4) | ((b & 0xf) << 4);
 580         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 581         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 582         return (u8) rev;
 583 }
 584
 585 /*
 586  * XXX this is hack in attempt to carry flags bitfield
 587  * throught endian village. ABI says:
 588  *
 589  * Bit-fields are allocated from right to left (least to most significant)
 590  * on little-endian implementations and from left to right (most to least
 591  * significant) on big-endian implementations.
 592  *
 593  * The above seems to be byte specific, so we need to reverse each
 594  * byte of the bitfield. 'Internet' also says this might be implementation
 595  * specific and we probably need proper fix and carry perf_event_attr
 596  * bitfield flags in separate data file FEAT_ section. Thought this seems
 597  * to work for now.
 598  */
 599 static void swap_bitfield(u8 *p, unsigned len)
 600 {
 601         unsigned i;
 602
 603         for (i = 0; i < len; i++) {
 604                 *p = revbyte(*p);
 605                 p++;
 606         }
 607 }
 608
 609 /* exported for swapping attributes in file header */
 610 void perf_event__attr_swap(struct perf_event_attr *attr)
 611 {
 612         attr->type              = bswap_32(attr->type);
 613         attr->size              = bswap_32(attr->size);
 614         attr->config            = bswap_64(attr->config);
 615         attr->sample_period     = bswap_64(attr->sample_period);
 616         attr->sample_type       = bswap_64(attr->sample_type);
 617         attr->read_format       = bswap_64(attr->read_format);
 618         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 619         attr->bp_type           = bswap_32(attr->bp_type);
 620         attr->bp_addr           = bswap_64(attr->bp_addr);
 621         attr->bp_len            = bswap_64(attr->bp_len);
 622
 623         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 624 }
 625
 626 static void perf_event__hdr_attr_swap(union perf_event *event,
 627                                       bool sample_id_all __used)
 628 {
 629         size_t size;
 630
 631         perf_event__attr_swap(&event->attr.attr);
 632
 633         size = event->header.size;
 634         size -= (void *)&event->attr.id - (void *)event;
 635         mem_bswap_64(event->attr.id, size);
 636 }
 637
 638 static void perf_event__event_type_swap(union perf_event *event,
 639                                         bool sample_id_all __used)
 640 {
 641         event->event_type.event_type.event_id =
 642                 bswap_64(event->event_type.event_type.event_id);
 643 }
 644
 645 static void perf_event__tracing_data_swap(union perf_event *event,
 646                                           bool sample_id_all __used)
 647 {
 648         event->tracing_data.size = bswap_32(event->tracing_data.size);
 649 }
 650
 651 typedef void (*perf_event__swap_op)(union perf_event *event,
 652                                     bool sample_id_all);
 653
 654 static perf_event__swap_op perf_event__swap_ops[] = {
 655         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 656         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 657         [PERF_RECORD_FORK]                = perf_event__task_swap,
 658         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 659         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 660         [PERF_RECORD_READ]                = perf_event__read_swap,
 661         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 662         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 663         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 664         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 665         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 666         [PERF_RECORD_HEADER_MAX]          = NULL,
 667 };
 668
 669 struct sample_queue {
 670         u64                     timestamp;
 671         u64                     file_offset;
 672         union perf_event        *event;
 673         struct list_head        list;
 674 };
 675
 676 static void perf_session_free_sample_buffers(struct perf_session *session)
 677 {
 678         struct ordered_samples *os = &session->ordered_samples;
 679
 680         while (!list_empty(&os->to_free)) {
 681                 struct sample_queue *sq;
 682
 683                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 684                 list_del(&sq->list);
 685                 free(sq);
 686         }
 687 }
 688
 689 static int perf_session_deliver_event(struct perf_session *session,
 690                                       union perf_event *event,
 691                                       struct perf_sample *sample,
 692                                       struct perf_tool *tool,
 693                                       u64 file_offset);
 694
 695 static void flush_sample_queue(struct perf_session *s,
 696                                struct perf_tool *tool)
 697 {
 698         struct ordered_samples *os = &s->ordered_samples;
 699         struct list_head *head = &os->samples;
 700         struct sample_queue *tmp, *iter;
 701         struct perf_sample sample;
 702         u64 limit = os->next_flush;
 703         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 704         unsigned idx = 0, progress_next = os->nr_samples / 16;
 705         int ret;
 706
 707         if (!tool->ordered_samples || !limit)
 708                 return;
 709
 710         list_for_each_entry_safe(iter, tmp, head, list) {
 711                 if (iter->timestamp > limit)
 712                         break;
 713
 714                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample,
 715                                                 s->header.needs_swap);
 716                 if (ret)
 717                         pr_err("Can't parse sample, err = %d\n", ret);
 718                 else
 719                         perf_session_deliver_event(s, iter->event, &sample, tool,
 720                                                    iter->file_offset);
 721
 722                 os->last_flush = iter->timestamp;
 723                 list_del(&iter->list);
 724                 list_add(&iter->list, &os->sample_cache);
 725                 if (++idx >= progress_next) {
 726                         progress_next += os->nr_samples / 16;
 727                         ui_progress__update(idx, os->nr_samples,
 728                                             "Processing time ordered events...");
 729                 }
 730         }
 731
 732         if (list_empty(head)) {
 733                 os->last_sample = NULL;
 734         } else if (last_ts <= limit) {
 735                 os->last_sample =
 736                         list_entry(head->prev, struct sample_queue, list);
 737         }
 738
 739         os->nr_samples = 0;
 740 }
 741
 742 /*
 743  * When perf record finishes a pass on every buffers, it records this pseudo
 744  * event.
 745  * We record the max timestamp t found in the pass n.
 746  * Assuming these timestamps are monotonic across cpus, we know that if
 747  * a buffer still has events with timestamps below t, they will be all
 748  * available and then read in the pass n + 1.
 749  * Hence when we start to read the pass n + 2, we can safely flush every
 750  * events with timestamps below t.
 751  *
 752  *    ============ PASS n =================
 753  *       CPU 0         |   CPU 1
 754  *                     |
 755  *    cnt1 timestamps  |   cnt2 timestamps
 756  *          1          |         2
 757  *          2          |         3
 758  *          -          |         4  <--- max recorded
 759  *
 760  *    ============ PASS n + 1 ==============
 761  *       CPU 0         |   CPU 1
 762  *                     |
 763  *    cnt1 timestamps  |   cnt2 timestamps
 764  *          3          |         5
 765  *          4          |         6
 766  *          5          |         7 <---- max recorded
 767  *
 768  *      Flush every events below timestamp 4
 769  *
 770  *    ============ PASS n + 2 ==============
 771  *       CPU 0         |   CPU 1
 772  *                     |
 773  *    cnt1 timestamps  |   cnt2 timestamps
 774  *          6          |         8
 775  *          7          |         9
 776  *          -          |         10
 777  *
 778  *      Flush every events below timestamp 7
 779  *      etc...
 780  */
 781 static int process_finished_round(struct perf_tool *tool,
 782                                   union perf_event *event __used,
 783                                   struct perf_session *session)
 784 {
 785         flush_sample_queue(session, tool);
 786         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 787
 788         return 0;
 789 }
 790
 791 /* The queue is ordered by time */
 792 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 793 {
 794         struct ordered_samples *os = &s->ordered_samples;
 795         struct sample_queue *sample = os->last_sample;
 796         u64 timestamp = new->timestamp;
 797         struct list_head *p;
 798
 799         ++os->nr_samples;
 800         os->last_sample = new;
 801
 802         if (!sample) {
 803                 list_add(&new->list, &os->samples);
 804                 os->max_timestamp = timestamp;
 805                 return;
 806         }
 807
 808         /*
 809          * last_sample might point to some random place in the list as it's
 810          * the last queued event. We expect that the new event is close to
 811          * this.
 812          */
 813         if (sample->timestamp <= timestamp) {
 814                 while (sample->timestamp <= timestamp) {
 815                         p = sample->list.next;
 816                         if (p == &os->samples) {
 817                                 list_add_tail(&new->list, &os->samples);
 818                                 os->max_timestamp = timestamp;
 819                                 return;
 820                         }
 821                         sample = list_entry(p, struct sample_queue, list);
 822                 }
 823                 list_add_tail(&new->list, &sample->list);
 824         } else {
 825                 while (sample->timestamp > timestamp) {
 826                         p = sample->list.prev;
 827                         if (p == &os->samples) {
 828                                 list_add(&new->list, &os->samples);
 829                                 return;
 830                         }
 831                         sample = list_entry(p, struct sample_queue, list);
 832                 }
 833                 list_add(&new->list, &sample->list);
 834         }
 835 }
 836
 837 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 838
 839 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 840                                     struct perf_sample *sample, u64 file_offset)
 841 {
 842         struct ordered_samples *os = &s->ordered_samples;
 843         struct list_head *sc = &os->sample_cache;
 844         u64 timestamp = sample->time;
 845         struct sample_queue *new;
 846
 847         if (!timestamp || timestamp == ~0ULL)
 848                 return -ETIME;
 849
 850         if (timestamp < s->ordered_samples.last_flush) {
 851                 printf("Warning: Timestamp below last timeslice flush\n");
 852                 return -EINVAL;
 853         }
 854
 855         if (!list_empty(sc)) {
 856                 new = list_entry(sc->next, struct sample_queue, list);
 857                 list_del(&new->list);
 858         } else if (os->sample_buffer) {
 859                 new = os->sample_buffer + os->sample_buffer_idx;
 860                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 861                         os->sample_buffer = NULL;
 862         } else {
 863                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 864                 if (!os->sample_buffer)
 865                         return -ENOMEM;
 866                 list_add(&os->sample_buffer->list, &os->to_free);
 867                 os->sample_buffer_idx = 2;
 868                 new = os->sample_buffer + 1;
 869         }
 870
 871         new->timestamp = timestamp;
 872         new->file_offset = file_offset;
 873         new->event = event;
 874
 875         __queue_event(new, s);
 876
 877         return 0;
 878 }
 879
 880 static void callchain__printf(struct perf_sample *sample)
 881 {
 882         unsigned int i;
 883
 884         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 885
 886         for (i = 0; i < sample->callchain->nr; i++)
 887                 printf("..... %2d: %016" PRIx64 "\n",
 888                        i, sample->callchain->ips[i]);
 889 }
 890
 891 static void branch_stack__printf(struct perf_sample *sample)
 892 {
 893         uint64_t i;
 894
 895         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 896
 897         for (i = 0; i < sample->branch_stack->nr; i++)
 898                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 899                         i, sample->branch_stack->entries[i].from,
 900                         sample->branch_stack->entries[i].to);
 901 }
 902
 903 static void regs_dump__printf(u64 mask, u64 *regs)
 904 {
 905         unsigned rid, i = 0;
 906
 907         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 908                 u64 val = regs[i++];
 909
 910                 printf(".... %-5s 0x%" PRIx64 "\n",
 911                        perf_reg_name(rid), val);
 912         }
 913 }
 914
 915 static void regs_user__printf(struct perf_sample *sample, u64 mask)
 916 {
 917         struct regs_dump *user_regs = &sample->user_regs;
 918
 919         if (user_regs->regs) {
 920                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
 921                 regs_dump__printf(mask, user_regs->regs);
 922         }
 923 }
 924
 925 static void stack_user__printf(struct stack_dump *dump)
 926 {
 927         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 928                dump->size, dump->offset);
 929 }
 930
 931 static void perf_session__print_tstamp(struct perf_session *session,
 932                                        union perf_event *event,
 933                                        struct perf_sample *sample)
 934 {
 935         u64 sample_type = perf_evlist__sample_type(session->evlist);
 936
 937         if (event->header.type != PERF_RECORD_SAMPLE &&
 938             !perf_evlist__sample_id_all(session->evlist)) {
 939                 fputs("-1 -1 ", stdout);
 940                 return;
 941         }
 942
 943         if ((sample_type & PERF_SAMPLE_CPU))
 944                 printf("%u ", sample->cpu);
 945
 946         if (sample_type & PERF_SAMPLE_TIME)
 947                 printf("%" PRIu64 " ", sample->time);
 948 }
 949
 950 static void dump_event(struct perf_session *session, union perf_event *event,
 951                        u64 file_offset, struct perf_sample *sample)
 952 {
 953         if (!dump_trace)
 954                 return;
 955
 956         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 957                file_offset, event->header.size, event->header.type);
 958
 959         trace_event(event);
 960
 961         if (sample)
 962                 perf_session__print_tstamp(session, event, sample);
 963
 964         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 965                event->header.size, perf_event__name(event->header.type));
 966 }
 967
 968 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 969                         struct perf_sample *sample)
 970 {
 971         u64 sample_type;
 972
 973         if (!dump_trace)
 974                 return;
 975
 976         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 977                event->header.misc, sample->pid, sample->tid, sample->ip,
 978                sample->period, sample->addr);
 979
 980         sample_type = evsel->attr.sample_type;
 981
 982         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 983                 callchain__printf(sample);
 984
 985         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
 986                 branch_stack__printf(sample);
 987
 988         if (sample_type & PERF_SAMPLE_REGS_USER)
 989                 regs_user__printf(sample, evsel->attr.sample_regs_user);
 990
 991         if (sample_type & PERF_SAMPLE_STACK_USER)
 992                 stack_user__printf(&sample->user_stack);
 993 }
 994
 995 static struct machine *
 996         perf_session__find_machine_for_cpumode(struct perf_session *session,
 997                                                union perf_event *event)
 998 {
 999         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1000
1001         if (perf_guest &&
1002             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
1003              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
1004                 u32 pid;
1005
1006                 if (event->header.type == PERF_RECORD_MMAP)
1007                         pid = event->mmap.pid;
1008                 else
1009                         pid = event->ip.pid;
1010
1011                 return perf_session__findnew_machine(session, pid);
1012         }
1013
1014         return perf_session__find_host_machine(session);
1015 }
1016
1017 static int perf_session_deliver_event(struct perf_session *session,
1018                                       union perf_event *event,
1019                                       struct perf_sample *sample,
1020                                       struct perf_tool *tool,
1021                                       u64 file_offset)
1022 {
1023         struct perf_evsel *evsel;
1024         struct machine *machine;
1025
1026         dump_event(session, event, file_offset, sample);
1027
1028         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
1029         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
1030                 /*
1031                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
1032                  * because the tools right now may apply filters, discarding
1033                  * some of the samples. For consistency, in the future we
1034                  * should have something like nr_filtered_samples and remove
1035                  * the sample->period from total_sample_period, etc, KISS for
1036                  * now tho.
1037                  *
1038                  * Also testing against NULL allows us to handle files without
1039                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
1040                  * future probably it'll be a good idea to restrict event
1041                  * processing via perf_session to files with both set.
1042                  */
1043                 hists__inc_nr_events(&evsel->hists, event->header.type);
1044         }
1045
1046         machine = perf_session__find_machine_for_cpumode(session, event);
1047
1048         switch (event->header.type) {
1049         case PERF_RECORD_SAMPLE:
1050                 dump_sample(evsel, event, sample);
1051                 if (evsel == NULL) {
1052                         ++session->hists.stats.nr_unknown_id;
1053                         return 0;
1054                 }
1055                 if (machine == NULL) {
1056                         ++session->hists.stats.nr_unprocessable_samples;
1057                         return 0;
1058                 }
1059                 return tool->sample(tool, event, sample, evsel, machine);
1060         case PERF_RECORD_MMAP:
1061                 return tool->mmap(tool, event, sample, machine);
1062         case PERF_RECORD_COMM:
1063                 return tool->comm(tool, event, sample, machine);
1064         case PERF_RECORD_FORK:
1065                 return tool->fork(tool, event, sample, machine);
1066         case PERF_RECORD_EXIT:
1067                 return tool->exit(tool, event, sample, machine);
1068         case PERF_RECORD_LOST:
1069                 if (tool->lost == perf_event__process_lost)
1070                         session->hists.stats.total_lost += event->lost.lost;
1071                 return tool->lost(tool, event, sample, machine);
1072         case PERF_RECORD_READ:
1073                 return tool->read(tool, event, sample, evsel, machine);
1074         case PERF_RECORD_THROTTLE:
1075                 return tool->throttle(tool, event, sample, machine);
1076         case PERF_RECORD_UNTHROTTLE:
1077                 return tool->unthrottle(tool, event, sample, machine);
1078         default:
1079                 ++session->hists.stats.nr_unknown_events;
1080                 return -1;
1081         }
1082 }
1083
1084 static int perf_session__preprocess_sample(struct perf_session *session,
1085                                            union perf_event *event, struct perf_sample *sample)
1086 {
1087         if (event->header.type != PERF_RECORD_SAMPLE ||
1088             !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN))
1089                 return 0;
1090
1091         if (!ip_callchain__valid(sample->callchain, event)) {
1092                 pr_debug("call-chain problem with event, skipping it.\n");
1093                 ++session->hists.stats.nr_invalid_chains;
1094                 session->hists.stats.total_invalid_chains += sample->period;
1095                 return -EINVAL;
1096         }
1097         return 0;
1098 }
1099
1100 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1101                                             struct perf_tool *tool, u64 file_offset)
1102 {
1103         int err;
1104
1105         dump_event(session, event, file_offset, NULL);
1106
1107         /* These events are processed right away */
1108         switch (event->header.type) {
1109         case PERF_RECORD_HEADER_ATTR:
1110                 err = tool->attr(event, &session->evlist);
1111                 if (err == 0)
1112                         perf_session__set_id_hdr_size(session);
1113                 return err;
1114         case PERF_RECORD_HEADER_EVENT_TYPE:
1115                 return tool->event_type(tool, event);
1116         case PERF_RECORD_HEADER_TRACING_DATA:
1117                 /* setup for reading amidst mmap */
1118                 lseek(session->fd, file_offset, SEEK_SET);
1119                 return tool->tracing_data(event, session);
1120         case PERF_RECORD_HEADER_BUILD_ID:
1121                 return tool->build_id(tool, event, session);
1122         case PERF_RECORD_FINISHED_ROUND:
1123                 return tool->finished_round(tool, event, session);
1124         default:
1125                 return -EINVAL;
1126         }
1127 }
1128
1129 static void event_swap(union perf_event *event, bool sample_id_all)
1130 {
1131         perf_event__swap_op swap;
1132
1133         swap = perf_event__swap_ops[event->header.type];
1134         if (swap)
1135                 swap(event, sample_id_all);
1136 }
1137
1138 static int perf_session__process_event(struct perf_session *session,
1139                                        union perf_event *event,
1140                                        struct perf_tool *tool,
1141                                        u64 file_offset)
1142 {
1143         struct perf_sample sample;
1144         int ret;
1145
1146         if (session->header.needs_swap)
1147                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1148
1149         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1150                 return -EINVAL;
1151
1152         hists__inc_nr_events(&session->hists, event->header.type);
1153
1154         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1155                 return perf_session__process_user_event(session, event, tool, file_offset);
1156
1157         /*
1158          * For all kernel events we get the sample data
1159          */
1160         ret = perf_evlist__parse_sample(session->evlist, event, &sample,
1161                                         session->header.needs_swap);
1162         if (ret)
1163                 return ret;
1164
1165         /* Preprocess sample records - precheck callchains */
1166         if (perf_session__preprocess_sample(session, event, &sample))
1167                 return 0;
1168
1169         if (tool->ordered_samples) {
1170                 ret = perf_session_queue_event(session, event, &sample,
1171                                                file_offset);
1172                 if (ret != -ETIME)
1173                         return ret;
1174         }
1175
1176         return perf_session_deliver_event(session, event, &sample, tool,
1177                                           file_offset);
1178 }
1179
1180 void perf_event_header__bswap(struct perf_event_header *self)
1181 {
1182         self->type = bswap_32(self->type);
1183         self->misc = bswap_16(self->misc);
1184         self->size = bswap_16(self->size);
1185 }
1186
1187 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1188 {
1189         return machine__findnew_thread(&session->host_machine, pid);
1190 }
1191
1192 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1193 {
1194         struct thread *thread = perf_session__findnew(self, 0);
1195
1196         if (thread == NULL || thread__set_comm(thread, "swapper")) {
1197                 pr_err("problem inserting idle task.\n");
1198                 thread = NULL;
1199         }
1200
1201         return thread;
1202 }
1203
1204 static void perf_session__warn_about_errors(const struct perf_session *session,
1205                                             const struct perf_tool *tool)
1206 {
1207         if (tool->lost == perf_event__process_lost &&
1208             session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1209                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1210                             "Check IO/CPU overload!\n\n",
1211                             session->hists.stats.nr_events[0],
1212                             session->hists.stats.nr_events[PERF_RECORD_LOST]);
1213         }
1214
1215         if (session->hists.stats.nr_unknown_events != 0) {
1216                 ui__warning("Found %u unknown events!\n\n"
1217                             "Is this an older tool processing a perf.data "
1218                             "file generated by a more recent tool?\n\n"
1219                             "If that is not the case, consider "
1220                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1221                             session->hists.stats.nr_unknown_events);
1222         }
1223
1224         if (session->hists.stats.nr_unknown_id != 0) {
1225                 ui__warning("%u samples with id not present in the header\n",
1226                             session->hists.stats.nr_unknown_id);
1227         }
1228
1229         if (session->hists.stats.nr_invalid_chains != 0) {
1230                 ui__warning("Found invalid callchains!\n\n"
1231                             "%u out of %u events were discarded for this reason.\n\n"
1232                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1233                             session->hists.stats.nr_invalid_chains,
1234                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1235         }
1236
1237         if (session->hists.stats.nr_unprocessable_samples != 0) {
1238                 ui__warning("%u unprocessable samples recorded.\n"
1239                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1240                             session->hists.stats.nr_unprocessable_samples);
1241         }
1242 }
1243
1244 #define session_done()  (*(volatile int *)(&session_done))
1245 volatile int session_done;
1246
1247 static int __perf_session__process_pipe_events(struct perf_session *self,
1248                                                struct perf_tool *tool)
1249 {
1250         union perf_event *event;
1251         uint32_t size, cur_size = 0;
1252         void *buf = NULL;
1253         int skip = 0;
1254         u64 head;
1255         int err;
1256         void *p;
1257
1258         perf_tool__fill_defaults(tool);
1259
1260         head = 0;
1261         cur_size = sizeof(union perf_event);
1262
1263         buf = malloc(cur_size);
1264         if (!buf)
1265                 return -errno;
1266 more:
1267         event = buf;
1268         err = readn(self->fd, event, sizeof(struct perf_event_header));
1269         if (err <= 0) {
1270                 if (err == 0)
1271                         goto done;
1272
1273                 pr_err("failed to read event header\n");
1274                 goto out_err;
1275         }
1276
1277         if (self->header.needs_swap)
1278                 perf_event_header__bswap(&event->header);
1279
1280         size = event->header.size;
1281         if (size == 0)
1282                 size = 8;
1283
1284         if (size > cur_size) {
1285                 void *new = realloc(buf, size);
1286                 if (!new) {
1287                         pr_err("failed to allocate memory to read event\n");
1288                         goto out_err;
1289                 }
1290                 buf = new;
1291                 cur_size = size;
1292                 event = buf;
1293         }
1294         p = event;
1295         p += sizeof(struct perf_event_header);
1296
1297         if (size - sizeof(struct perf_event_header)) {
1298                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1299                 if (err <= 0) {
1300                         if (err == 0) {
1301                                 pr_err("unexpected end of event stream\n");
1302                                 goto done;
1303                         }
1304
1305                         pr_err("failed to read event data\n");
1306                         goto out_err;
1307                 }
1308         }
1309
1310         if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1311                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1312                        head, event->header.size, event->header.type);
1313                 err = -EINVAL;
1314                 goto out_err;
1315         }
1316
1317         head += size;
1318
1319         if (skip > 0)
1320                 head += skip;
1321
1322         if (!session_done())
1323                 goto more;
1324 done:
1325         err = 0;
1326 out_err:
1327         free(buf);
1328         perf_session__warn_about_errors(self, tool);
1329         perf_session_free_sample_buffers(self);
1330         return err;
1331 }
1332
1333 static union perf_event *
1334 fetch_mmaped_event(struct perf_session *session,
1335                    u64 head, size_t mmap_size, char *buf)
1336 {
1337         union perf_event *event;
1338
1339         /*
1340          * Ensure we have enough space remaining to read
1341          * the size of the event in the headers.
1342          */
1343         if (head + sizeof(event->header) > mmap_size)
1344                 return NULL;
1345
1346         event = (union perf_event *)(buf + head);
1347
1348         if (session->header.needs_swap)
1349                 perf_event_header__bswap(&event->header);
1350
1351         if (head + event->header.size > mmap_size)
1352                 return NULL;
1353
1354         return event;
1355 }
1356
1357 int __perf_session__process_events(struct perf_session *session,
1358                                    u64 data_offset, u64 data_size,
1359                                    u64 file_size, struct perf_tool *tool)
1360 {
1361         u64 head, page_offset, file_offset, file_pos, progress_next;
1362         int err, mmap_prot, mmap_flags, map_idx = 0;
1363         size_t  page_size, mmap_size;
1364         char *buf, *mmaps[8];
1365         union perf_event *event;
1366         uint32_t size;
1367
1368         perf_tool__fill_defaults(tool);
1369
1370         page_size = sysconf(_SC_PAGESIZE);
1371
1372         page_offset = page_size * (data_offset / page_size);
1373         file_offset = page_offset;
1374         head = data_offset - page_offset;
1375
1376         if (data_offset + data_size < file_size)
1377                 file_size = data_offset + data_size;
1378
1379         progress_next = file_size / 16;
1380
1381         mmap_size = session->mmap_window;
1382         if (mmap_size > file_size)
1383                 mmap_size = file_size;
1384
1385         memset(mmaps, 0, sizeof(mmaps));
1386
1387         mmap_prot  = PROT_READ;
1388         mmap_flags = MAP_SHARED;
1389
1390         if (session->header.needs_swap) {
1391                 mmap_prot  |= PROT_WRITE;
1392                 mmap_flags = MAP_PRIVATE;
1393         }
1394 remap:
1395         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1396                    file_offset);
1397         if (buf == MAP_FAILED) {
1398                 pr_err("failed to mmap file\n");
1399                 err = -errno;
1400                 goto out_err;
1401         }
1402         mmaps[map_idx] = buf;
1403         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1404         file_pos = file_offset + head;
1405
1406 more:
1407         event = fetch_mmaped_event(session, head, mmap_size, buf);
1408         if (!event) {
1409                 if (mmaps[map_idx]) {
1410                         munmap(mmaps[map_idx], mmap_size);
1411                         mmaps[map_idx] = NULL;
1412                 }
1413
1414                 page_offset = page_size * (head / page_size);
1415                 file_offset += page_offset;
1416                 head -= page_offset;
1417                 goto remap;
1418         }
1419
1420         size = event->header.size;
1421
1422         if (size == 0 ||
1423             perf_session__process_event(session, event, tool, file_pos) < 0) {
1424                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1425                        file_offset + head, event->header.size,
1426                        event->header.type);
1427                 err = -EINVAL;
1428                 goto out_err;
1429         }
1430
1431         head += size;
1432         file_pos += size;
1433
1434         if (file_pos >= progress_next) {
1435                 progress_next += file_size / 16;
1436                 ui_progress__update(file_pos, file_size,
1437                                     "Processing events...");
1438         }
1439
1440         if (file_pos < file_size)
1441                 goto more;
1442
1443         err = 0;
1444         /* do the final flush for ordered samples */
1445         session->ordered_samples.next_flush = ULLONG_MAX;
1446         flush_sample_queue(session, tool);
1447 out_err:
1448         perf_session__warn_about_errors(session, tool);
1449         perf_session_free_sample_buffers(session);
1450         return err;
1451 }
1452
1453 int perf_session__process_events(struct perf_session *self,
1454                                  struct perf_tool *tool)
1455 {
1456         int err;
1457
1458         if (perf_session__register_idle_thread(self) == NULL)
1459                 return -ENOMEM;
1460
1461         if (!self->fd_pipe)
1462                 err = __perf_session__process_events(self,
1463                                                      self->header.data_offset,
1464                                                      self->header.data_size,
1465                                                      self->size, tool);
1466         else
1467                 err = __perf_session__process_pipe_events(self, tool);
1468
1469         return err;
1470 }
1471
1472 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1473 {
1474         if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) {
1475                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1476                 return false;
1477         }
1478
1479         return true;
1480 }
1481
1482 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1483                                      const char *symbol_name, u64 addr)
1484 {
1485         char *bracket;
1486         enum map_type i;
1487         struct ref_reloc_sym *ref;
1488
1489         ref = zalloc(sizeof(struct ref_reloc_sym));
1490         if (ref == NULL)
1491                 return -ENOMEM;
1492
1493         ref->name = strdup(symbol_name);
1494         if (ref->name == NULL) {
1495                 free(ref);
1496                 return -ENOMEM;
1497         }
1498
1499         bracket = strchr(ref->name, ']');
1500         if (bracket)
1501                 *bracket = '\0';
1502
1503         ref->addr = addr;
1504
1505         for (i = 0; i < MAP__NR_TYPES; ++i) {
1506                 struct kmap *kmap = map__kmap(maps[i]);
1507                 kmap->ref_reloc_sym = ref;
1508         }
1509
1510         return 0;
1511 }
1512
1513 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1514 {
1515         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1516                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1517                machines__fprintf_dsos(&self->machines, fp);
1518 }
1519
1520 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1521                                           bool with_hits)
1522 {
1523         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1524         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1525 }
1526
1527 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1528 {
1529         struct perf_evsel *pos;
1530         size_t ret = fprintf(fp, "Aggregated stats:\n");
1531
1532         ret += hists__fprintf_nr_events(&session->hists, fp);
1533
1534         list_for_each_entry(pos, &session->evlist->entries, node) {
1535                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1536                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1537         }
1538
1539         return ret;
1540 }
1541
1542 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1543 {
1544         /*
1545          * FIXME: Here we have to actually print all the machines in this
1546          * session, not just the host...
1547          */
1548         return machine__fprintf(&session->host_machine, fp);
1549 }
1550
1551 void perf_session__remove_thread(struct perf_session *session,
1552                                  struct thread *th)
1553 {
1554         /*
1555          * FIXME: This one makes no sense, we need to remove the thread from
1556          * the machine it belongs to, perf_session can have many machines, so
1557          * doing it always on ->host_machine is wrong.  Fix when auditing all
1558          * the 'perf kvm' code.
1559          */
1560         machine__remove_thread(&session->host_machine, th);
1561 }
1562
1563 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1564                                               unsigned int type)
1565 {
1566         struct perf_evsel *pos;
1567
1568         list_for_each_entry(pos, &session->evlist->entries, node) {
1569                 if (pos->attr.type == type)
1570                         return pos;
1571         }
1572         return NULL;
1573 }
1574
1575 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1576                           struct perf_sample *sample, struct machine *machine,
1577                           int print_sym, int print_dso, int print_symoffset)
1578 {
1579         struct addr_location al;
1580         struct callchain_cursor_node *node;
1581
1582         if (perf_event__preprocess_sample(event, machine, &al, sample,
1583                                           NULL) < 0) {
1584                 error("problem processing %d event, skipping it.\n",
1585                         event->header.type);
1586                 return;
1587         }
1588
1589         if (symbol_conf.use_callchain && sample->callchain) {
1590
1591
1592                 if (machine__resolve_callchain(machine, evsel, al.thread,
1593                                                sample, NULL) != 0) {
1594                         if (verbose)
1595                                 error("Failed to resolve callchain. Skipping\n");
1596                         return;
1597                 }
1598                 callchain_cursor_commit(&callchain_cursor);
1599
1600                 while (1) {
1601                         node = callchain_cursor_current(&callchain_cursor);
1602                         if (!node)
1603                                 break;
1604
1605                         printf("\t%16" PRIx64, node->ip);
1606                         if (print_sym) {
1607                                 printf(" ");
1608                                 symbol__fprintf_symname(node->sym, stdout);
1609                         }
1610                         if (print_dso) {
1611                                 printf(" (");
1612                                 map__fprintf_dsoname(node->map, stdout);
1613                                 printf(")");
1614                         }
1615                         printf("\n");
1616
1617                         callchain_cursor_advance(&callchain_cursor);
1618                 }
1619
1620         } else {
1621                 printf("%16" PRIx64, sample->ip);
1622                 if (print_sym) {
1623                         printf(" ");
1624                         if (print_symoffset)
1625                                 symbol__fprintf_symname_offs(al.sym, &al,
1626                                                              stdout);
1627                         else
1628                                 symbol__fprintf_symname(al.sym, stdout);
1629                 }
1630
1631                 if (print_dso) {
1632                         printf(" (");
1633                         map__fprintf_dsoname(al.map, stdout);
1634                         printf(")");
1635                 }
1636         }
1637 }
1638
1639 int perf_session__cpu_bitmap(struct perf_session *session,
1640                              const char *cpu_list, unsigned long *cpu_bitmap)
1641 {
1642         int i;
1643         struct cpu_map *map;
1644
1645         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1646                 struct perf_evsel *evsel;
1647
1648                 evsel = perf_session__find_first_evtype(session, i);
1649                 if (!evsel)
1650                         continue;
1651
1652                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1653                         pr_err("File does not contain CPU events. "
1654                                "Remove -c option to proceed.\n");
1655                         return -1;
1656                 }
1657         }
1658
1659         map = cpu_map__new(cpu_list);
1660         if (map == NULL) {
1661                 pr_err("Invalid cpu_list\n");
1662                 return -1;
1663         }
1664
1665         for (i = 0; i < map->nr; i++) {
1666                 int cpu = map->map[i];
1667
1668                 if (cpu >= MAX_NR_CPUS) {
1669                         pr_err("Requested CPU %d too large. "
1670                                "Consider raising MAX_NR_CPUS\n", cpu);
1671                         return -1;
1672                 }
1673
1674                 set_bit(cpu, cpu_bitmap);
1675         }
1676
1677         return 0;
1678 }
1679
1680 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1681                                 bool full)
1682 {
1683         struct stat st;
1684         int ret;
1685
1686         if (session == NULL || fp == NULL)
1687                 return;
1688
1689         ret = fstat(session->fd, &st);
1690         if (ret == -1)
1691                 return;
1692
1693         fprintf(fp, "# ========\n");
1694         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1695         perf_header__fprintf_info(session, fp, full);
1696         fprintf(fp, "# ========\n#\n");
1697 }
1698
1699
1700 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1701                                              const struct perf_evsel_str_handler *assocs,
1702                                              size_t nr_assocs)
1703 {
1704         struct perf_evlist *evlist = session->evlist;
1705         struct event_format *format;
1706         struct perf_evsel *evsel;
1707         char *tracepoint, *name;
1708         size_t i;
1709         int err;
1710
1711         for (i = 0; i < nr_assocs; i++) {
1712                 err = -ENOMEM;
1713                 tracepoint = strdup(assocs[i].name);
1714                 if (tracepoint == NULL)
1715                         goto out;
1716
1717                 err = -ENOENT;
1718                 name = strchr(tracepoint, ':');
1719                 if (name == NULL)
1720                         goto out_free;
1721
1722                 *name++ = '\0';
1723                 format = pevent_find_event_by_name(session->pevent,
1724                                                    tracepoint, name);
1725                 if (format == NULL) {
1726                         /*
1727                          * Adding a handler for an event not in the session,
1728                          * just ignore it.
1729                          */
1730                         goto next;
1731                 }
1732
1733                 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1734                 if (evsel == NULL)
1735                         goto next;
1736
1737                 err = -EEXIST;
1738                 if (evsel->handler.func != NULL)
1739                         goto out_free;
1740                 evsel->handler.func = assocs[i].handler;
1741 next:
1742                 free(tracepoint);
1743         }
1744
1745         err = 0;
1746 out:
1747         return err;
1748
1749 out_free:
1750         free(tracepoint);
1751         goto out;
1752 }