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 #include "vdso.h"
  21
  22 static int perf_session__open(struct perf_session *self, bool force)
  23 {
  24         struct stat input_stat;
  25
  26         if (!strcmp(self->filename, "-")) {
  27                 self->fd_pipe = true;
  28                 self->fd = STDIN_FILENO;
  29
  30                 if (perf_session__read_header(self, self->fd) < 0)
  31                         pr_err("incompatible file format (rerun with -v to learn more)");
  32
  33                 return 0;
  34         }
  35
  36         self->fd = open(self->filename, O_RDONLY);
  37         if (self->fd < 0) {
  38                 int err = errno;
  39
  40                 pr_err("failed to open %s: %s", self->filename, strerror(err));
  41                 if (err == ENOENT && !strcmp(self->filename, "perf.data"))
  42                         pr_err("  (try 'perf record' first)");
  43                 pr_err("\n");
  44                 return -errno;
  45         }
  46
  47         if (fstat(self->fd, &input_stat) < 0)
  48                 goto out_close;
  49
  50         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
  51                 pr_err("file %s not owned by current user or root\n",
  52                        self->filename);
  53                 goto out_close;
  54         }
  55
  56         if (!input_stat.st_size) {
  57                 pr_info("zero-sized file (%s), nothing to do!\n",
  58                         self->filename);
  59                 goto out_close;
  60         }
  61
  62         if (perf_session__read_header(self, self->fd) < 0) {
  63                 pr_err("incompatible file format (rerun with -v to learn more)");
  64                 goto out_close;
  65         }
  66
  67         if (!perf_evlist__valid_sample_type(self->evlist)) {
  68                 pr_err("non matching sample_type");
  69                 goto out_close;
  70         }
  71
  72         if (!perf_evlist__valid_sample_id_all(self->evlist)) {
  73                 pr_err("non matching sample_id_all");
  74                 goto out_close;
  75         }
  76
  77         self->size = input_stat.st_size;
  78         return 0;
  79
  80 out_close:
  81         close(self->fd);
  82         self->fd = -1;
  83         return -1;
  84 }
  85
  86 void perf_session__set_id_hdr_size(struct perf_session *session)
  87 {
  88         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
  89
  90         session->host_machine.id_hdr_size = id_hdr_size;
  91         machines__set_id_hdr_size(&session->machines, id_hdr_size);
  92 }
  93
  94 int perf_session__create_kernel_maps(struct perf_session *self)
  95 {
  96         int ret = machine__create_kernel_maps(&self->host_machine);
  97
  98         if (ret >= 0)
  99                 ret = machines__create_guest_kernel_maps(&self->machines);
 100         return ret;
 101 }
 102
 103 static void perf_session__destroy_kernel_maps(struct perf_session *self)
 104 {
 105         machine__destroy_kernel_maps(&self->host_machine);
 106         machines__destroy_guest_kernel_maps(&self->machines);
 107 }
 108
 109 struct perf_session *perf_session__new(const char *filename, int mode,
 110                                        bool force, bool repipe,
 111                                        struct perf_tool *tool)
 112 {
 113         struct perf_session *self;
 114         struct stat st;
 115         size_t len;
 116
 117         if (!filename || !strlen(filename)) {
 118                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 119                         filename = "-";
 120                 else
 121                         filename = "perf.data";
 122         }
 123
 124         len = strlen(filename);
 125         self = zalloc(sizeof(*self) + len);
 126
 127         if (self == NULL)
 128                 goto out;
 129
 130         memcpy(self->filename, filename, len);
 131         /*
 132          * On 64bit we can mmap the data file in one go. No need for tiny mmap
 133          * slices. On 32bit we use 32MB.
 134          */
 135 #if BITS_PER_LONG == 64
 136         self->mmap_window = ULLONG_MAX;
 137 #else
 138         self->mmap_window = 32 * 1024 * 1024ULL;
 139 #endif
 140         self->machines = RB_ROOT;
 141         self->repipe = repipe;
 142         INIT_LIST_HEAD(&self->ordered_samples.samples);
 143         INIT_LIST_HEAD(&self->ordered_samples.sample_cache);
 144         INIT_LIST_HEAD(&self->ordered_samples.to_free);
 145         machine__init(&self->host_machine, "", HOST_KERNEL_ID);
 146         hists__init(&self->hists);
 147
 148         if (mode == O_RDONLY) {
 149                 if (perf_session__open(self, force) < 0)
 150                         goto out_delete;
 151                 perf_session__set_id_hdr_size(self);
 152         } else if (mode == O_WRONLY) {
 153                 /*
 154                  * In O_RDONLY mode this will be performed when reading the
 155                  * kernel MMAP event, in perf_event__process_mmap().
 156                  */
 157                 if (perf_session__create_kernel_maps(self) < 0)
 158                         goto out_delete;
 159         }
 160
 161         if (tool && tool->ordering_requires_timestamps &&
 162             tool->ordered_samples && !perf_evlist__sample_id_all(self->evlist)) {
 163                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 164                 tool->ordered_samples = false;
 165         }
 166
 167 out:
 168         return self;
 169 out_delete:
 170         perf_session__delete(self);
 171         return NULL;
 172 }
 173
 174 static void machine__delete_dead_threads(struct machine *machine)
 175 {
 176         struct thread *n, *t;
 177
 178         list_for_each_entry_safe(t, n, &machine->dead_threads, node) {
 179                 list_del(&t->node);
 180                 thread__delete(t);
 181         }
 182 }
 183
 184 static void perf_session__delete_dead_threads(struct perf_session *session)
 185 {
 186         machine__delete_dead_threads(&session->host_machine);
 187 }
 188
 189 static void machine__delete_threads(struct machine *self)
 190 {
 191         struct rb_node *nd = rb_first(&self->threads);
 192
 193         while (nd) {
 194                 struct thread *t = rb_entry(nd, struct thread, rb_node);
 195
 196                 rb_erase(&t->rb_node, &self->threads);
 197                 nd = rb_next(nd);
 198                 thread__delete(t);
 199         }
 200 }
 201
 202 static void perf_session__delete_threads(struct perf_session *session)
 203 {
 204         machine__delete_threads(&session->host_machine);
 205 }
 206
 207 void perf_session__delete(struct perf_session *self)
 208 {
 209         perf_session__destroy_kernel_maps(self);
 210         perf_session__delete_dead_threads(self);
 211         perf_session__delete_threads(self);
 212         machine__exit(&self->host_machine);
 213         close(self->fd);
 214         free(self);
 215         vdso__exit();
 216 }
 217
 218 void machine__remove_thread(struct machine *self, struct thread *th)
 219 {
 220         self->last_match = NULL;
 221         rb_erase(&th->rb_node, &self->threads);
 222         /*
 223          * We may have references to this thread, for instance in some hist_entry
 224          * instances, so just move them to a separate list.
 225          */
 226         list_add_tail(&th->node, &self->dead_threads);
 227 }
 228
 229 static bool symbol__match_parent_regex(struct symbol *sym)
 230 {
 231         if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
 232                 return 1;
 233
 234         return 0;
 235 }
 236
 237 static const u8 cpumodes[] = {
 238         PERF_RECORD_MISC_USER,
 239         PERF_RECORD_MISC_KERNEL,
 240         PERF_RECORD_MISC_GUEST_USER,
 241         PERF_RECORD_MISC_GUEST_KERNEL
 242 };
 243 #define NCPUMODES (sizeof(cpumodes)/sizeof(u8))
 244
 245 static void ip__resolve_ams(struct machine *self, struct thread *thread,
 246                             struct addr_map_symbol *ams,
 247                             u64 ip)
 248 {
 249         struct addr_location al;
 250         size_t i;
 251         u8 m;
 252
 253         memset(&al, 0, sizeof(al));
 254
 255         for (i = 0; i < NCPUMODES; i++) {
 256                 m = cpumodes[i];
 257                 /*
 258                  * We cannot use the header.misc hint to determine whether a
 259                  * branch stack address is user, kernel, guest, hypervisor.
 260                  * Branches may straddle the kernel/user/hypervisor boundaries.
 261                  * Thus, we have to try consecutively until we find a match
 262                  * or else, the symbol is unknown
 263                  */
 264                 thread__find_addr_location(thread, self, m, MAP__FUNCTION,
 265                                 ip, &al, NULL);
 266                 if (al.sym)
 267                         goto found;
 268         }
 269 found:
 270         ams->addr = ip;
 271         ams->al_addr = al.addr;
 272         ams->sym = al.sym;
 273         ams->map = al.map;
 274 }
 275
 276 struct branch_info *machine__resolve_bstack(struct machine *self,
 277                                             struct thread *thr,
 278                                             struct branch_stack *bs)
 279 {
 280         struct branch_info *bi;
 281         unsigned int i;
 282
 283         bi = calloc(bs->nr, sizeof(struct branch_info));
 284         if (!bi)
 285                 return NULL;
 286
 287         for (i = 0; i < bs->nr; i++) {
 288                 ip__resolve_ams(self, thr, &bi[i].to, bs->entries[i].to);
 289                 ip__resolve_ams(self, thr, &bi[i].from, bs->entries[i].from);
 290                 bi[i].flags = bs->entries[i].flags;
 291         }
 292         return bi;
 293 }
 294
 295 static int machine__resolve_callchain_sample(struct machine *machine,
 296                                              struct thread *thread,
 297                                              struct ip_callchain *chain,
 298                                              struct symbol **parent)
 299
 300 {
 301         u8 cpumode = PERF_RECORD_MISC_USER;
 302         unsigned int i;
 303         int err;
 304
 305         callchain_cursor_reset(&callchain_cursor);
 306
 307         if (chain->nr > PERF_MAX_STACK_DEPTH) {
 308                 pr_warning("corrupted callchain. skipping...\n");
 309                 return 0;
 310         }
 311
 312         for (i = 0; i < chain->nr; i++) {
 313                 u64 ip;
 314                 struct addr_location al;
 315
 316                 if (callchain_param.order == ORDER_CALLEE)
 317                         ip = chain->ips[i];
 318                 else
 319                         ip = chain->ips[chain->nr - i - 1];
 320
 321                 if (ip >= PERF_CONTEXT_MAX) {
 322                         switch (ip) {
 323                         case PERF_CONTEXT_HV:
 324                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;
 325                                 break;
 326                         case PERF_CONTEXT_KERNEL:
 327                                 cpumode = PERF_RECORD_MISC_KERNEL;
 328                                 break;
 329                         case PERF_CONTEXT_USER:
 330                                 cpumode = PERF_RECORD_MISC_USER;
 331                                 break;
 332                         default:
 333                                 pr_debug("invalid callchain context: "
 334                                          "%"PRId64"\n", (s64) ip);
 335                                 /*
 336                                  * It seems the callchain is corrupted.
 337                                  * Discard all.
 338                                  */
 339                                 callchain_cursor_reset(&callchain_cursor);
 340                                 return 0;
 341                         }
 342                         continue;
 343                 }
 344
 345                 al.filtered = false;
 346                 thread__find_addr_location(thread, machine, cpumode,
 347                                            MAP__FUNCTION, ip, &al, NULL);
 348                 if (al.sym != NULL) {
 349                         if (sort__has_parent && !*parent &&
 350                             symbol__match_parent_regex(al.sym))
 351                                 *parent = al.sym;
 352                         if (!symbol_conf.use_callchain)
 353                                 break;
 354                 }
 355
 356                 err = callchain_cursor_append(&callchain_cursor,
 357                                               ip, al.map, al.sym);
 358                 if (err)
 359                         return err;
 360         }
 361
 362         return 0;
 363 }
 364
 365 static int unwind_entry(struct unwind_entry *entry, void *arg)
 366 {
 367         struct callchain_cursor *cursor = arg;
 368         return callchain_cursor_append(cursor, entry->ip,
 369                                        entry->map, entry->sym);
 370 }
 371
 372 int machine__resolve_callchain(struct machine *machine,
 373                                struct perf_evsel *evsel,
 374                                struct thread *thread,
 375                                struct perf_sample *sample,
 376                                struct symbol **parent)
 377
 378 {
 379         int ret;
 380
 381         callchain_cursor_reset(&callchain_cursor);
 382
 383         ret = machine__resolve_callchain_sample(machine, thread,
 384                                                 sample->callchain, parent);
 385         if (ret)
 386                 return ret;
 387
 388         /* Can we do dwarf post unwind? */
 389         if (!((evsel->attr.sample_type & PERF_SAMPLE_REGS_USER) &&
 390               (evsel->attr.sample_type & PERF_SAMPLE_STACK_USER)))
 391                 return 0;
 392
 393         /* Bail out if nothing was captured. */
 394         if ((!sample->user_regs.regs) ||
 395             (!sample->user_stack.size))
 396                 return 0;
 397
 398         return unwind__get_entries(unwind_entry, &callchain_cursor, machine,
 399                                    thread, evsel->attr.sample_regs_user,
 400                                    sample);
 401
 402 }
 403
 404 static int process_event_synth_tracing_data_stub(union perf_event *event
 405                                                  __maybe_unused,
 406                                                  struct perf_session *session
 407                                                 __maybe_unused)
 408 {
 409         dump_printf(": unhandled!\n");
 410         return 0;
 411 }
 412
 413 static int process_event_synth_attr_stub(union perf_event *event __maybe_unused,
 414                                          struct perf_evlist **pevlist
 415                                          __maybe_unused)
 416 {
 417         dump_printf(": unhandled!\n");
 418         return 0;
 419 }
 420
 421 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
 422                                      union perf_event *event __maybe_unused,
 423                                      struct perf_sample *sample __maybe_unused,
 424                                      struct perf_evsel *evsel __maybe_unused,
 425                                      struct machine *machine __maybe_unused)
 426 {
 427         dump_printf(": unhandled!\n");
 428         return 0;
 429 }
 430
 431 static int process_event_stub(struct perf_tool *tool __maybe_unused,
 432                               union perf_event *event __maybe_unused,
 433                               struct perf_sample *sample __maybe_unused,
 434                               struct machine *machine __maybe_unused)
 435 {
 436         dump_printf(": unhandled!\n");
 437         return 0;
 438 }
 439
 440 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
 441                                        union perf_event *event __maybe_unused,
 442                                        struct perf_session *perf_session
 443                                        __maybe_unused)
 444 {
 445         dump_printf(": unhandled!\n");
 446         return 0;
 447 }
 448
 449 static int process_event_type_stub(struct perf_tool *tool __maybe_unused,
 450                                    union perf_event *event __maybe_unused)
 451 {
 452         dump_printf(": unhandled!\n");
 453         return 0;
 454 }
 455
 456 static int process_finished_round(struct perf_tool *tool,
 457                                   union perf_event *event,
 458                                   struct perf_session *session);
 459
 460 static void perf_tool__fill_defaults(struct perf_tool *tool)
 461 {
 462         if (tool->sample == NULL)
 463                 tool->sample = process_event_sample_stub;
 464         if (tool->mmap == NULL)
 465                 tool->mmap = process_event_stub;
 466         if (tool->comm == NULL)
 467                 tool->comm = process_event_stub;
 468         if (tool->fork == NULL)
 469                 tool->fork = process_event_stub;
 470         if (tool->exit == NULL)
 471                 tool->exit = process_event_stub;
 472         if (tool->lost == NULL)
 473                 tool->lost = perf_event__process_lost;
 474         if (tool->read == NULL)
 475                 tool->read = process_event_sample_stub;
 476         if (tool->throttle == NULL)
 477                 tool->throttle = process_event_stub;
 478         if (tool->unthrottle == NULL)
 479                 tool->unthrottle = process_event_stub;
 480         if (tool->attr == NULL)
 481                 tool->attr = process_event_synth_attr_stub;
 482         if (tool->event_type == NULL)
 483                 tool->event_type = process_event_type_stub;
 484         if (tool->tracing_data == NULL)
 485                 tool->tracing_data = process_event_synth_tracing_data_stub;
 486         if (tool->build_id == NULL)
 487                 tool->build_id = process_finished_round_stub;
 488         if (tool->finished_round == NULL) {
 489                 if (tool->ordered_samples)
 490                         tool->finished_round = process_finished_round;
 491                 else
 492                         tool->finished_round = process_finished_round_stub;
 493         }
 494 }
 495
 496 void mem_bswap_32(void *src, int byte_size)
 497 {
 498         u32 *m = src;
 499         while (byte_size > 0) {
 500                 *m = bswap_32(*m);
 501                 byte_size -= sizeof(u32);
 502                 ++m;
 503         }
 504 }
 505
 506 void mem_bswap_64(void *src, int byte_size)
 507 {
 508         u64 *m = src;
 509
 510         while (byte_size > 0) {
 511                 *m = bswap_64(*m);
 512                 byte_size -= sizeof(u64);
 513                 ++m;
 514         }
 515 }
 516
 517 static void swap_sample_id_all(union perf_event *event, void *data)
 518 {
 519         void *end = (void *) event + event->header.size;
 520         int size = end - data;
 521
 522         BUG_ON(size % sizeof(u64));
 523         mem_bswap_64(data, size);
 524 }
 525
 526 static void perf_event__all64_swap(union perf_event *event,
 527                                    bool sample_id_all __maybe_unused)
 528 {
 529         struct perf_event_header *hdr = &event->header;
 530         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 531 }
 532
 533 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 534 {
 535         event->comm.pid = bswap_32(event->comm.pid);
 536         event->comm.tid = bswap_32(event->comm.tid);
 537
 538         if (sample_id_all) {
 539                 void *data = &event->comm.comm;
 540
 541                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 542                 swap_sample_id_all(event, data);
 543         }
 544 }
 545
 546 static void perf_event__mmap_swap(union perf_event *event,
 547                                   bool sample_id_all)
 548 {
 549         event->mmap.pid   = bswap_32(event->mmap.pid);
 550         event->mmap.tid   = bswap_32(event->mmap.tid);
 551         event->mmap.start = bswap_64(event->mmap.start);
 552         event->mmap.len   = bswap_64(event->mmap.len);
 553         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 554
 555         if (sample_id_all) {
 556                 void *data = &event->mmap.filename;
 557
 558                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 559                 swap_sample_id_all(event, data);
 560         }
 561 }
 562
 563 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 564 {
 565         event->fork.pid  = bswap_32(event->fork.pid);
 566         event->fork.tid  = bswap_32(event->fork.tid);
 567         event->fork.ppid = bswap_32(event->fork.ppid);
 568         event->fork.ptid = bswap_32(event->fork.ptid);
 569         event->fork.time = bswap_64(event->fork.time);
 570
 571         if (sample_id_all)
 572                 swap_sample_id_all(event, &event->fork + 1);
 573 }
 574
 575 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 576 {
 577         event->read.pid          = bswap_32(event->read.pid);
 578         event->read.tid          = bswap_32(event->read.tid);
 579         event->read.value        = bswap_64(event->read.value);
 580         event->read.time_enabled = bswap_64(event->read.time_enabled);
 581         event->read.time_running = bswap_64(event->read.time_running);
 582         event->read.id           = bswap_64(event->read.id);
 583
 584         if (sample_id_all)
 585                 swap_sample_id_all(event, &event->read + 1);
 586 }
 587
 588 static u8 revbyte(u8 b)
 589 {
 590         int rev = (b >> 4) | ((b & 0xf) << 4);
 591         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 592         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 593         return (u8) rev;
 594 }
 595
 596 /*
 597  * XXX this is hack in attempt to carry flags bitfield
 598  * throught endian village. ABI says:
 599  *
 600  * Bit-fields are allocated from right to left (least to most significant)
 601  * on little-endian implementations and from left to right (most to least
 602  * significant) on big-endian implementations.
 603  *
 604  * The above seems to be byte specific, so we need to reverse each
 605  * byte of the bitfield. 'Internet' also says this might be implementation
 606  * specific and we probably need proper fix and carry perf_event_attr
 607  * bitfield flags in separate data file FEAT_ section. Thought this seems
 608  * to work for now.
 609  */
 610 static void swap_bitfield(u8 *p, unsigned len)
 611 {
 612         unsigned i;
 613
 614         for (i = 0; i < len; i++) {
 615                 *p = revbyte(*p);
 616                 p++;
 617         }
 618 }
 619
 620 /* exported for swapping attributes in file header */
 621 void perf_event__attr_swap(struct perf_event_attr *attr)
 622 {
 623         attr->type              = bswap_32(attr->type);
 624         attr->size              = bswap_32(attr->size);
 625         attr->config            = bswap_64(attr->config);
 626         attr->sample_period     = bswap_64(attr->sample_period);
 627         attr->sample_type       = bswap_64(attr->sample_type);
 628         attr->read_format       = bswap_64(attr->read_format);
 629         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 630         attr->bp_type           = bswap_32(attr->bp_type);
 631         attr->bp_addr           = bswap_64(attr->bp_addr);
 632         attr->bp_len            = bswap_64(attr->bp_len);
 633
 634         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 635 }
 636
 637 static void perf_event__hdr_attr_swap(union perf_event *event,
 638                                       bool sample_id_all __maybe_unused)
 639 {
 640         size_t size;
 641
 642         perf_event__attr_swap(&event->attr.attr);
 643
 644         size = event->header.size;
 645         size -= (void *)&event->attr.id - (void *)event;
 646         mem_bswap_64(event->attr.id, size);
 647 }
 648
 649 static void perf_event__event_type_swap(union perf_event *event,
 650                                         bool sample_id_all __maybe_unused)
 651 {
 652         event->event_type.event_type.event_id =
 653                 bswap_64(event->event_type.event_type.event_id);
 654 }
 655
 656 static void perf_event__tracing_data_swap(union perf_event *event,
 657                                           bool sample_id_all __maybe_unused)
 658 {
 659         event->tracing_data.size = bswap_32(event->tracing_data.size);
 660 }
 661
 662 typedef void (*perf_event__swap_op)(union perf_event *event,
 663                                     bool sample_id_all);
 664
 665 static perf_event__swap_op perf_event__swap_ops[] = {
 666         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 667         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 668         [PERF_RECORD_FORK]                = perf_event__task_swap,
 669         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 670         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 671         [PERF_RECORD_READ]                = perf_event__read_swap,
 672         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 673         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 674         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 675         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 676         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 677         [PERF_RECORD_HEADER_MAX]          = NULL,
 678 };
 679
 680 struct sample_queue {
 681         u64                     timestamp;
 682         u64                     file_offset;
 683         union perf_event        *event;
 684         struct list_head        list;
 685 };
 686
 687 static void perf_session_free_sample_buffers(struct perf_session *session)
 688 {
 689         struct ordered_samples *os = &session->ordered_samples;
 690
 691         while (!list_empty(&os->to_free)) {
 692                 struct sample_queue *sq;
 693
 694                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 695                 list_del(&sq->list);
 696                 free(sq);
 697         }
 698 }
 699
 700 static int perf_session_deliver_event(struct perf_session *session,
 701                                       union perf_event *event,
 702                                       struct perf_sample *sample,
 703                                       struct perf_tool *tool,
 704                                       u64 file_offset);
 705
 706 static int flush_sample_queue(struct perf_session *s,
 707                                struct perf_tool *tool)
 708 {
 709         struct ordered_samples *os = &s->ordered_samples;
 710         struct list_head *head = &os->samples;
 711         struct sample_queue *tmp, *iter;
 712         struct perf_sample sample;
 713         u64 limit = os->next_flush;
 714         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 715         unsigned idx = 0, progress_next = os->nr_samples / 16;
 716         int ret;
 717
 718         if (!tool->ordered_samples || !limit)
 719                 return 0;
 720
 721         list_for_each_entry_safe(iter, tmp, head, list) {
 722                 if (iter->timestamp > limit)
 723                         break;
 724
 725                 ret = perf_evlist__parse_sample(s->evlist, iter->event, &sample);
 726                 if (ret)
 727                         pr_err("Can't parse sample, err = %d\n", ret);
 728                 else {
 729                         ret = perf_session_deliver_event(s, iter->event, &sample, tool,
 730                                                          iter->file_offset);
 731                         if (ret)
 732                                 return ret;
 733                 }
 734
 735                 os->last_flush = iter->timestamp;
 736                 list_del(&iter->list);
 737                 list_add(&iter->list, &os->sample_cache);
 738                 if (++idx >= progress_next) {
 739                         progress_next += os->nr_samples / 16;
 740                         ui_progress__update(idx, os->nr_samples,
 741                                             "Processing time ordered events...");
 742                 }
 743         }
 744
 745         if (list_empty(head)) {
 746                 os->last_sample = NULL;
 747         } else if (last_ts <= limit) {
 748                 os->last_sample =
 749                         list_entry(head->prev, struct sample_queue, list);
 750         }
 751
 752         os->nr_samples = 0;
 753
 754         return 0;
 755 }
 756
 757 /*
 758  * When perf record finishes a pass on every buffers, it records this pseudo
 759  * event.
 760  * We record the max timestamp t found in the pass n.
 761  * Assuming these timestamps are monotonic across cpus, we know that if
 762  * a buffer still has events with timestamps below t, they will be all
 763  * available and then read in the pass n + 1.
 764  * Hence when we start to read the pass n + 2, we can safely flush every
 765  * events with timestamps below t.
 766  *
 767  *    ============ PASS n =================
 768  *       CPU 0         |   CPU 1
 769  *                     |
 770  *    cnt1 timestamps  |   cnt2 timestamps
 771  *          1          |         2
 772  *          2          |         3
 773  *          -          |         4  <--- max recorded
 774  *
 775  *    ============ PASS n + 1 ==============
 776  *       CPU 0         |   CPU 1
 777  *                     |
 778  *    cnt1 timestamps  |   cnt2 timestamps
 779  *          3          |         5
 780  *          4          |         6
 781  *          5          |         7 <---- max recorded
 782  *
 783  *      Flush every events below timestamp 4
 784  *
 785  *    ============ PASS n + 2 ==============
 786  *       CPU 0         |   CPU 1
 787  *                     |
 788  *    cnt1 timestamps  |   cnt2 timestamps
 789  *          6          |         8
 790  *          7          |         9
 791  *          -          |         10
 792  *
 793  *      Flush every events below timestamp 7
 794  *      etc...
 795  */
 796 static int process_finished_round(struct perf_tool *tool,
 797                                   union perf_event *event __maybe_unused,
 798                                   struct perf_session *session)
 799 {
 800         int ret = flush_sample_queue(session, tool);
 801         if (!ret)
 802                 session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 803
 804         return ret;
 805 }
 806
 807 /* The queue is ordered by time */
 808 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 809 {
 810         struct ordered_samples *os = &s->ordered_samples;
 811         struct sample_queue *sample = os->last_sample;
 812         u64 timestamp = new->timestamp;
 813         struct list_head *p;
 814
 815         ++os->nr_samples;
 816         os->last_sample = new;
 817
 818         if (!sample) {
 819                 list_add(&new->list, &os->samples);
 820                 os->max_timestamp = timestamp;
 821                 return;
 822         }
 823
 824         /*
 825          * last_sample might point to some random place in the list as it's
 826          * the last queued event. We expect that the new event is close to
 827          * this.
 828          */
 829         if (sample->timestamp <= timestamp) {
 830                 while (sample->timestamp <= timestamp) {
 831                         p = sample->list.next;
 832                         if (p == &os->samples) {
 833                                 list_add_tail(&new->list, &os->samples);
 834                                 os->max_timestamp = timestamp;
 835                                 return;
 836                         }
 837                         sample = list_entry(p, struct sample_queue, list);
 838                 }
 839                 list_add_tail(&new->list, &sample->list);
 840         } else {
 841                 while (sample->timestamp > timestamp) {
 842                         p = sample->list.prev;
 843                         if (p == &os->samples) {
 844                                 list_add(&new->list, &os->samples);
 845                                 return;
 846                         }
 847                         sample = list_entry(p, struct sample_queue, list);
 848                 }
 849                 list_add(&new->list, &sample->list);
 850         }
 851 }
 852
 853 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 854
 855 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 856                                     struct perf_sample *sample, u64 file_offset)
 857 {
 858         struct ordered_samples *os = &s->ordered_samples;
 859         struct list_head *sc = &os->sample_cache;
 860         u64 timestamp = sample->time;
 861         struct sample_queue *new;
 862
 863         if (!timestamp || timestamp == ~0ULL)
 864                 return -ETIME;
 865
 866         if (timestamp < s->ordered_samples.last_flush) {
 867                 printf("Warning: Timestamp below last timeslice flush\n");
 868                 return -EINVAL;
 869         }
 870
 871         if (!list_empty(sc)) {
 872                 new = list_entry(sc->next, struct sample_queue, list);
 873                 list_del(&new->list);
 874         } else if (os->sample_buffer) {
 875                 new = os->sample_buffer + os->sample_buffer_idx;
 876                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 877                         os->sample_buffer = NULL;
 878         } else {
 879                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 880                 if (!os->sample_buffer)
 881                         return -ENOMEM;
 882                 list_add(&os->sample_buffer->list, &os->to_free);
 883                 os->sample_buffer_idx = 2;
 884                 new = os->sample_buffer + 1;
 885         }
 886
 887         new->timestamp = timestamp;
 888         new->file_offset = file_offset;
 889         new->event = event;
 890
 891         __queue_event(new, s);
 892
 893         return 0;
 894 }
 895
 896 static void callchain__printf(struct perf_sample *sample)
 897 {
 898         unsigned int i;
 899
 900         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 901
 902         for (i = 0; i < sample->callchain->nr; i++)
 903                 printf("..... %2d: %016" PRIx64 "\n",
 904                        i, sample->callchain->ips[i]);
 905 }
 906
 907 static void branch_stack__printf(struct perf_sample *sample)
 908 {
 909         uint64_t i;
 910
 911         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 912
 913         for (i = 0; i < sample->branch_stack->nr; i++)
 914                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 915                         i, sample->branch_stack->entries[i].from,
 916                         sample->branch_stack->entries[i].to);
 917 }
 918
 919 static void regs_dump__printf(u64 mask, u64 *regs)
 920 {
 921         unsigned rid, i = 0;
 922
 923         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 924                 u64 val = regs[i++];
 925
 926                 printf(".... %-5s 0x%" PRIx64 "\n",
 927                        perf_reg_name(rid), val);
 928         }
 929 }
 930
 931 static void regs_user__printf(struct perf_sample *sample, u64 mask)
 932 {
 933         struct regs_dump *user_regs = &sample->user_regs;
 934
 935         if (user_regs->regs) {
 936                 printf("... user regs: mask 0x%" PRIx64 "\n", mask);
 937                 regs_dump__printf(mask, user_regs->regs);
 938         }
 939 }
 940
 941 static void stack_user__printf(struct stack_dump *dump)
 942 {
 943         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 944                dump->size, dump->offset);
 945 }
 946
 947 static void perf_session__print_tstamp(struct perf_session *session,
 948                                        union perf_event *event,
 949                                        struct perf_sample *sample)
 950 {
 951         u64 sample_type = perf_evlist__sample_type(session->evlist);
 952
 953         if (event->header.type != PERF_RECORD_SAMPLE &&
 954             !perf_evlist__sample_id_all(session->evlist)) {
 955                 fputs("-1 -1 ", stdout);
 956                 return;
 957         }
 958
 959         if ((sample_type & PERF_SAMPLE_CPU))
 960                 printf("%u ", sample->cpu);
 961
 962         if (sample_type & PERF_SAMPLE_TIME)
 963                 printf("%" PRIu64 " ", sample->time);
 964 }
 965
 966 static void dump_event(struct perf_session *session, union perf_event *event,
 967                        u64 file_offset, struct perf_sample *sample)
 968 {
 969         if (!dump_trace)
 970                 return;
 971
 972         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 973                file_offset, event->header.size, event->header.type);
 974
 975         trace_event(event);
 976
 977         if (sample)
 978                 perf_session__print_tstamp(session, event, sample);
 979
 980         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 981                event->header.size, perf_event__name(event->header.type));
 982 }
 983
 984 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 985                         struct perf_sample *sample)
 986 {
 987         u64 sample_type;
 988
 989         if (!dump_trace)
 990                 return;
 991
 992         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 993                event->header.misc, sample->pid, sample->tid, sample->ip,
 994                sample->period, sample->addr);
 995
 996         sample_type = evsel->attr.sample_type;
 997
 998         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 999                 callchain__printf(sample);
1000
1001         if (sample_type & PERF_SAMPLE_BRANCH_STACK)
1002                 branch_stack__printf(sample);
1003
1004         if (sample_type & PERF_SAMPLE_REGS_USER)
1005                 regs_user__printf(sample, evsel->attr.sample_regs_user);
1006
1007         if (sample_type & PERF_SAMPLE_STACK_USER)
1008                 stack_user__printf(&sample->user_stack);
1009 }
1010
1011 static struct machine *
1012         perf_session__find_machine_for_cpumode(struct perf_session *session,
1013                                                union perf_event *event)
1014 {
1015         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1016
1017         if (perf_guest &&
1018             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
1019              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
1020                 u32 pid;
1021
1022                 if (event->header.type == PERF_RECORD_MMAP)
1023                         pid = event->mmap.pid;
1024                 else
1025                         pid = event->ip.pid;
1026
1027                 return perf_session__findnew_machine(session, pid);
1028         }
1029
1030         return perf_session__find_host_machine(session);
1031 }
1032
1033 static int perf_session_deliver_event(struct perf_session *session,
1034                                       union perf_event *event,
1035                                       struct perf_sample *sample,
1036                                       struct perf_tool *tool,
1037                                       u64 file_offset)
1038 {
1039         struct perf_evsel *evsel;
1040         struct machine *machine;
1041
1042         dump_event(session, event, file_offset, sample);
1043
1044         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
1045         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
1046                 /*
1047                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
1048                  * because the tools right now may apply filters, discarding
1049                  * some of the samples. For consistency, in the future we
1050                  * should have something like nr_filtered_samples and remove
1051                  * the sample->period from total_sample_period, etc, KISS for
1052                  * now tho.
1053                  *
1054                  * Also testing against NULL allows us to handle files without
1055                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
1056                  * future probably it'll be a good idea to restrict event
1057                  * processing via perf_session to files with both set.
1058                  */
1059                 hists__inc_nr_events(&evsel->hists, event->header.type);
1060         }
1061
1062         machine = perf_session__find_machine_for_cpumode(session, event);
1063
1064         switch (event->header.type) {
1065         case PERF_RECORD_SAMPLE:
1066                 dump_sample(evsel, event, sample);
1067                 if (evsel == NULL) {
1068                         ++session->hists.stats.nr_unknown_id;
1069                         return 0;
1070                 }
1071                 if (machine == NULL) {
1072                         ++session->hists.stats.nr_unprocessable_samples;
1073                         return 0;
1074                 }
1075                 return tool->sample(tool, event, sample, evsel, machine);
1076         case PERF_RECORD_MMAP:
1077                 return tool->mmap(tool, event, sample, machine);
1078         case PERF_RECORD_COMM:
1079                 return tool->comm(tool, event, sample, machine);
1080         case PERF_RECORD_FORK:
1081                 return tool->fork(tool, event, sample, machine);
1082         case PERF_RECORD_EXIT:
1083                 return tool->exit(tool, event, sample, machine);
1084         case PERF_RECORD_LOST:
1085                 if (tool->lost == perf_event__process_lost)
1086                         session->hists.stats.total_lost += event->lost.lost;
1087                 return tool->lost(tool, event, sample, machine);
1088         case PERF_RECORD_READ:
1089                 return tool->read(tool, event, sample, evsel, machine);
1090         case PERF_RECORD_THROTTLE:
1091                 return tool->throttle(tool, event, sample, machine);
1092         case PERF_RECORD_UNTHROTTLE:
1093                 return tool->unthrottle(tool, event, sample, machine);
1094         default:
1095                 ++session->hists.stats.nr_unknown_events;
1096                 return -1;
1097         }
1098 }
1099
1100 static int perf_session__preprocess_sample(struct perf_session *session,
1101                                            union perf_event *event, struct perf_sample *sample)
1102 {
1103         if (event->header.type != PERF_RECORD_SAMPLE ||
1104             !(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_CALLCHAIN))
1105                 return 0;
1106
1107         if (!ip_callchain__valid(sample->callchain, event)) {
1108                 pr_debug("call-chain problem with event, skipping it.\n");
1109                 ++session->hists.stats.nr_invalid_chains;
1110                 session->hists.stats.total_invalid_chains += sample->period;
1111                 return -EINVAL;
1112         }
1113         return 0;
1114 }
1115
1116 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1117                                             struct perf_tool *tool, u64 file_offset)
1118 {
1119         int err;
1120
1121         dump_event(session, event, file_offset, NULL);
1122
1123         /* These events are processed right away */
1124         switch (event->header.type) {
1125         case PERF_RECORD_HEADER_ATTR:
1126                 err = tool->attr(event, &session->evlist);
1127                 if (err == 0)
1128                         perf_session__set_id_hdr_size(session);
1129                 return err;
1130         case PERF_RECORD_HEADER_EVENT_TYPE:
1131                 return tool->event_type(tool, event);
1132         case PERF_RECORD_HEADER_TRACING_DATA:
1133                 /* setup for reading amidst mmap */
1134                 lseek(session->fd, file_offset, SEEK_SET);
1135                 return tool->tracing_data(event, session);
1136         case PERF_RECORD_HEADER_BUILD_ID:
1137                 return tool->build_id(tool, event, session);
1138         case PERF_RECORD_FINISHED_ROUND:
1139                 return tool->finished_round(tool, event, session);
1140         default:
1141                 return -EINVAL;
1142         }
1143 }
1144
1145 static void event_swap(union perf_event *event, bool sample_id_all)
1146 {
1147         perf_event__swap_op swap;
1148
1149         swap = perf_event__swap_ops[event->header.type];
1150         if (swap)
1151                 swap(event, sample_id_all);
1152 }
1153
1154 static int perf_session__process_event(struct perf_session *session,
1155                                        union perf_event *event,
1156                                        struct perf_tool *tool,
1157                                        u64 file_offset)
1158 {
1159         struct perf_sample sample;
1160         int ret;
1161
1162         if (session->header.needs_swap)
1163                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1164
1165         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1166                 return -EINVAL;
1167
1168         hists__inc_nr_events(&session->hists, event->header.type);
1169
1170         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1171                 return perf_session__process_user_event(session, event, tool, file_offset);
1172
1173         /*
1174          * For all kernel events we get the sample data
1175          */
1176         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1177         if (ret)
1178                 return ret;
1179
1180         /* Preprocess sample records - precheck callchains */
1181         if (perf_session__preprocess_sample(session, event, &sample))
1182                 return 0;
1183
1184         if (tool->ordered_samples) {
1185                 ret = perf_session_queue_event(session, event, &sample,
1186                                                file_offset);
1187                 if (ret != -ETIME)
1188                         return ret;
1189         }
1190
1191         return perf_session_deliver_event(session, event, &sample, tool,
1192                                           file_offset);
1193 }
1194
1195 void perf_event_header__bswap(struct perf_event_header *self)
1196 {
1197         self->type = bswap_32(self->type);
1198         self->misc = bswap_16(self->misc);
1199         self->size = bswap_16(self->size);
1200 }
1201
1202 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1203 {
1204         return machine__findnew_thread(&session->host_machine, pid);
1205 }
1206
1207 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1208 {
1209         struct thread *thread = perf_session__findnew(self, 0);
1210
1211         if (thread == NULL || thread__set_comm(thread, "swapper")) {
1212                 pr_err("problem inserting idle task.\n");
1213                 thread = NULL;
1214         }
1215
1216         return thread;
1217 }
1218
1219 static void perf_session__warn_about_errors(const struct perf_session *session,
1220                                             const struct perf_tool *tool)
1221 {
1222         if (tool->lost == perf_event__process_lost &&
1223             session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1224                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1225                             "Check IO/CPU overload!\n\n",
1226                             session->hists.stats.nr_events[0],
1227                             session->hists.stats.nr_events[PERF_RECORD_LOST]);
1228         }
1229
1230         if (session->hists.stats.nr_unknown_events != 0) {
1231                 ui__warning("Found %u unknown events!\n\n"
1232                             "Is this an older tool processing a perf.data "
1233                             "file generated by a more recent tool?\n\n"
1234                             "If that is not the case, consider "
1235                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1236                             session->hists.stats.nr_unknown_events);
1237         }
1238
1239         if (session->hists.stats.nr_unknown_id != 0) {
1240                 ui__warning("%u samples with id not present in the header\n",
1241                             session->hists.stats.nr_unknown_id);
1242         }
1243
1244         if (session->hists.stats.nr_invalid_chains != 0) {
1245                 ui__warning("Found invalid callchains!\n\n"
1246                             "%u out of %u events were discarded for this reason.\n\n"
1247                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1248                             session->hists.stats.nr_invalid_chains,
1249                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1250         }
1251
1252         if (session->hists.stats.nr_unprocessable_samples != 0) {
1253                 ui__warning("%u unprocessable samples recorded.\n"
1254                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1255                             session->hists.stats.nr_unprocessable_samples);
1256         }
1257 }
1258
1259 #define session_done()  (*(volatile int *)(&session_done))
1260 volatile int session_done;
1261
1262 static int __perf_session__process_pipe_events(struct perf_session *self,
1263                                                struct perf_tool *tool)
1264 {
1265         union perf_event *event;
1266         uint32_t size, cur_size = 0;
1267         void *buf = NULL;
1268         int skip = 0;
1269         u64 head;
1270         int err;
1271         void *p;
1272
1273         perf_tool__fill_defaults(tool);
1274
1275         head = 0;
1276         cur_size = sizeof(union perf_event);
1277
1278         buf = malloc(cur_size);
1279         if (!buf)
1280                 return -errno;
1281 more:
1282         event = buf;
1283         err = readn(self->fd, event, sizeof(struct perf_event_header));
1284         if (err <= 0) {
1285                 if (err == 0)
1286                         goto done;
1287
1288                 pr_err("failed to read event header\n");
1289                 goto out_err;
1290         }
1291
1292         if (self->header.needs_swap)
1293                 perf_event_header__bswap(&event->header);
1294
1295         size = event->header.size;
1296         if (size == 0)
1297                 size = 8;
1298
1299         if (size > cur_size) {
1300                 void *new = realloc(buf, size);
1301                 if (!new) {
1302                         pr_err("failed to allocate memory to read event\n");
1303                         goto out_err;
1304                 }
1305                 buf = new;
1306                 cur_size = size;
1307                 event = buf;
1308         }
1309         p = event;
1310         p += sizeof(struct perf_event_header);
1311
1312         if (size - sizeof(struct perf_event_header)) {
1313                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1314                 if (err <= 0) {
1315                         if (err == 0) {
1316                                 pr_err("unexpected end of event stream\n");
1317                                 goto done;
1318                         }
1319
1320                         pr_err("failed to read event data\n");
1321                         goto out_err;
1322                 }
1323         }
1324
1325         if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1326                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1327                        head, event->header.size, event->header.type);
1328                 err = -EINVAL;
1329                 goto out_err;
1330         }
1331
1332         head += size;
1333
1334         if (skip > 0)
1335                 head += skip;
1336
1337         if (!session_done())
1338                 goto more;
1339 done:
1340         err = 0;
1341 out_err:
1342         free(buf);
1343         perf_session__warn_about_errors(self, tool);
1344         perf_session_free_sample_buffers(self);
1345         return err;
1346 }
1347
1348 static union perf_event *
1349 fetch_mmaped_event(struct perf_session *session,
1350                    u64 head, size_t mmap_size, char *buf)
1351 {
1352         union perf_event *event;
1353
1354         /*
1355          * Ensure we have enough space remaining to read
1356          * the size of the event in the headers.
1357          */
1358         if (head + sizeof(event->header) > mmap_size)
1359                 return NULL;
1360
1361         event = (union perf_event *)(buf + head);
1362
1363         if (session->header.needs_swap)
1364                 perf_event_header__bswap(&event->header);
1365
1366         if (head + event->header.size > mmap_size)
1367                 return NULL;
1368
1369         return event;
1370 }
1371
1372 int __perf_session__process_events(struct perf_session *session,
1373                                    u64 data_offset, u64 data_size,
1374                                    u64 file_size, struct perf_tool *tool)
1375 {
1376         u64 head, page_offset, file_offset, file_pos, progress_next;
1377         int err, mmap_prot, mmap_flags, map_idx = 0;
1378         size_t  page_size, mmap_size;
1379         char *buf, *mmaps[8];
1380         union perf_event *event;
1381         uint32_t size;
1382
1383         perf_tool__fill_defaults(tool);
1384
1385         page_size = sysconf(_SC_PAGESIZE);
1386
1387         page_offset = page_size * (data_offset / page_size);
1388         file_offset = page_offset;
1389         head = data_offset - page_offset;
1390
1391         if (data_offset + data_size < file_size)
1392                 file_size = data_offset + data_size;
1393
1394         progress_next = file_size / 16;
1395
1396         mmap_size = session->mmap_window;
1397         if (mmap_size > file_size)
1398                 mmap_size = file_size;
1399
1400         memset(mmaps, 0, sizeof(mmaps));
1401
1402         mmap_prot  = PROT_READ;
1403         mmap_flags = MAP_SHARED;
1404
1405         if (session->header.needs_swap) {
1406                 mmap_prot  |= PROT_WRITE;
1407                 mmap_flags = MAP_PRIVATE;
1408         }
1409 remap:
1410         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1411                    file_offset);
1412         if (buf == MAP_FAILED) {
1413                 pr_err("failed to mmap file\n");
1414                 err = -errno;
1415                 goto out_err;
1416         }
1417         mmaps[map_idx] = buf;
1418         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1419         file_pos = file_offset + head;
1420
1421 more:
1422         event = fetch_mmaped_event(session, head, mmap_size, buf);
1423         if (!event) {
1424                 if (mmaps[map_idx]) {
1425                         munmap(mmaps[map_idx], mmap_size);
1426                         mmaps[map_idx] = NULL;
1427                 }
1428
1429                 page_offset = page_size * (head / page_size);
1430                 file_offset += page_offset;
1431                 head -= page_offset;
1432                 goto remap;
1433         }
1434
1435         size = event->header.size;
1436
1437         if (size == 0 ||
1438             perf_session__process_event(session, event, tool, file_pos) < 0) {
1439                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1440                        file_offset + head, event->header.size,
1441                        event->header.type);
1442                 err = -EINVAL;
1443                 goto out_err;
1444         }
1445
1446         head += size;
1447         file_pos += size;
1448
1449         if (file_pos >= progress_next) {
1450                 progress_next += file_size / 16;
1451                 ui_progress__update(file_pos, file_size,
1452                                     "Processing events...");
1453         }
1454
1455         if (file_pos < file_size)
1456                 goto more;
1457
1458         err = 0;
1459         /* do the final flush for ordered samples */
1460         session->ordered_samples.next_flush = ULLONG_MAX;
1461         err = flush_sample_queue(session, tool);
1462 out_err:
1463         perf_session__warn_about_errors(session, tool);
1464         perf_session_free_sample_buffers(session);
1465         return err;
1466 }
1467
1468 int perf_session__process_events(struct perf_session *self,
1469                                  struct perf_tool *tool)
1470 {
1471         int err;
1472
1473         if (perf_session__register_idle_thread(self) == NULL)
1474                 return -ENOMEM;
1475
1476         if (!self->fd_pipe)
1477                 err = __perf_session__process_events(self,
1478                                                      self->header.data_offset,
1479                                                      self->header.data_size,
1480                                                      self->size, tool);
1481         else
1482                 err = __perf_session__process_pipe_events(self, tool);
1483
1484         return err;
1485 }
1486
1487 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1488 {
1489         if (!(perf_evlist__sample_type(session->evlist) & PERF_SAMPLE_RAW)) {
1490                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1491                 return false;
1492         }
1493
1494         return true;
1495 }
1496
1497 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1498                                      const char *symbol_name, u64 addr)
1499 {
1500         char *bracket;
1501         enum map_type i;
1502         struct ref_reloc_sym *ref;
1503
1504         ref = zalloc(sizeof(struct ref_reloc_sym));
1505         if (ref == NULL)
1506                 return -ENOMEM;
1507
1508         ref->name = strdup(symbol_name);
1509         if (ref->name == NULL) {
1510                 free(ref);
1511                 return -ENOMEM;
1512         }
1513
1514         bracket = strchr(ref->name, ']');
1515         if (bracket)
1516                 *bracket = '\0';
1517
1518         ref->addr = addr;
1519
1520         for (i = 0; i < MAP__NR_TYPES; ++i) {
1521                 struct kmap *kmap = map__kmap(maps[i]);
1522                 kmap->ref_reloc_sym = ref;
1523         }
1524
1525         return 0;
1526 }
1527
1528 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1529 {
1530         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1531                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1532                machines__fprintf_dsos(&self->machines, fp);
1533 }
1534
1535 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1536                                           bool with_hits)
1537 {
1538         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1539         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1540 }
1541
1542 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1543 {
1544         struct perf_evsel *pos;
1545         size_t ret = fprintf(fp, "Aggregated stats:\n");
1546
1547         ret += hists__fprintf_nr_events(&session->hists, fp);
1548
1549         list_for_each_entry(pos, &session->evlist->entries, node) {
1550                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1551                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1552         }
1553
1554         return ret;
1555 }
1556
1557 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1558 {
1559         /*
1560          * FIXME: Here we have to actually print all the machines in this
1561          * session, not just the host...
1562          */
1563         return machine__fprintf(&session->host_machine, fp);
1564 }
1565
1566 void perf_session__remove_thread(struct perf_session *session,
1567                                  struct thread *th)
1568 {
1569         /*
1570          * FIXME: This one makes no sense, we need to remove the thread from
1571          * the machine it belongs to, perf_session can have many machines, so
1572          * doing it always on ->host_machine is wrong.  Fix when auditing all
1573          * the 'perf kvm' code.
1574          */
1575         machine__remove_thread(&session->host_machine, th);
1576 }
1577
1578 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1579                                               unsigned int type)
1580 {
1581         struct perf_evsel *pos;
1582
1583         list_for_each_entry(pos, &session->evlist->entries, node) {
1584                 if (pos->attr.type == type)
1585                         return pos;
1586         }
1587         return NULL;
1588 }
1589
1590 void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
1591                           struct perf_sample *sample, struct machine *machine,
1592                           int print_sym, int print_dso, int print_symoffset)
1593 {
1594         struct addr_location al;
1595         struct callchain_cursor_node *node;
1596
1597         if (perf_event__preprocess_sample(event, machine, &al, sample,
1598                                           NULL) < 0) {
1599                 error("problem processing %d event, skipping it.\n",
1600                         event->header.type);
1601                 return;
1602         }
1603
1604         if (symbol_conf.use_callchain && sample->callchain) {
1605
1606
1607                 if (machine__resolve_callchain(machine, evsel, al.thread,
1608                                                sample, NULL) != 0) {
1609                         if (verbose)
1610                                 error("Failed to resolve callchain. Skipping\n");
1611                         return;
1612                 }
1613                 callchain_cursor_commit(&callchain_cursor);
1614
1615                 while (1) {
1616                         node = callchain_cursor_current(&callchain_cursor);
1617                         if (!node)
1618                                 break;
1619
1620                         printf("\t%16" PRIx64, node->ip);
1621                         if (print_sym) {
1622                                 printf(" ");
1623                                 symbol__fprintf_symname(node->sym, stdout);
1624                         }
1625                         if (print_dso) {
1626                                 printf(" (");
1627                                 map__fprintf_dsoname(node->map, stdout);
1628                                 printf(")");
1629                         }
1630                         printf("\n");
1631
1632                         callchain_cursor_advance(&callchain_cursor);
1633                 }
1634
1635         } else {
1636                 printf("%16" PRIx64, sample->ip);
1637                 if (print_sym) {
1638                         printf(" ");
1639                         if (print_symoffset)
1640                                 symbol__fprintf_symname_offs(al.sym, &al,
1641                                                              stdout);
1642                         else
1643                                 symbol__fprintf_symname(al.sym, stdout);
1644                 }
1645
1646                 if (print_dso) {
1647                         printf(" (");
1648                         map__fprintf_dsoname(al.map, stdout);
1649                         printf(")");
1650                 }
1651         }
1652 }
1653
1654 int perf_session__cpu_bitmap(struct perf_session *session,
1655                              const char *cpu_list, unsigned long *cpu_bitmap)
1656 {
1657         int i;
1658         struct cpu_map *map;
1659
1660         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1661                 struct perf_evsel *evsel;
1662
1663                 evsel = perf_session__find_first_evtype(session, i);
1664                 if (!evsel)
1665                         continue;
1666
1667                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1668                         pr_err("File does not contain CPU events. "
1669                                "Remove -c option to proceed.\n");
1670                         return -1;
1671                 }
1672         }
1673
1674         map = cpu_map__new(cpu_list);
1675         if (map == NULL) {
1676                 pr_err("Invalid cpu_list\n");
1677                 return -1;
1678         }
1679
1680         for (i = 0; i < map->nr; i++) {
1681                 int cpu = map->map[i];
1682
1683                 if (cpu >= MAX_NR_CPUS) {
1684                         pr_err("Requested CPU %d too large. "
1685                                "Consider raising MAX_NR_CPUS\n", cpu);
1686                         return -1;
1687                 }
1688
1689                 set_bit(cpu, cpu_bitmap);
1690         }
1691
1692         return 0;
1693 }
1694
1695 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1696                                 bool full)
1697 {
1698         struct stat st;
1699         int ret;
1700
1701         if (session == NULL || fp == NULL)
1702                 return;
1703
1704         ret = fstat(session->fd, &st);
1705         if (ret == -1)
1706                 return;
1707
1708         fprintf(fp, "# ========\n");
1709         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1710         perf_header__fprintf_info(session, fp, full);
1711         fprintf(fp, "# ========\n#\n");
1712 }
1713
1714
1715 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1716                                              const struct perf_evsel_str_handler *assocs,
1717                                              size_t nr_assocs)
1718 {
1719         struct perf_evlist *evlist = session->evlist;
1720         struct event_format *format;
1721         struct perf_evsel *evsel;
1722         char *tracepoint, *name;
1723         size_t i;
1724         int err;
1725
1726         for (i = 0; i < nr_assocs; i++) {
1727                 err = -ENOMEM;
1728                 tracepoint = strdup(assocs[i].name);
1729                 if (tracepoint == NULL)
1730                         goto out;
1731
1732                 err = -ENOENT;
1733                 name = strchr(tracepoint, ':');
1734                 if (name == NULL)
1735                         goto out_free;
1736
1737                 *name++ = '\0';
1738                 format = pevent_find_event_by_name(session->pevent,
1739                                                    tracepoint, name);
1740                 if (format == NULL) {
1741                         /*
1742                          * Adding a handler for an event not in the session,
1743                          * just ignore it.
1744                          */
1745                         goto next;
1746                 }
1747
1748                 evsel = perf_evlist__find_tracepoint_by_id(evlist, format->id);
1749                 if (evsel == NULL)
1750                         goto next;
1751
1752                 err = -EEXIST;
1753                 if (evsel->handler.func != NULL)
1754                         goto out_free;
1755                 evsel->handler.func = assocs[i].handler;
1756 next:
1757                 free(tracepoint);
1758         }
1759
1760         err = 0;
1761 out:
1762         return err;
1763
1764 out_free:
1765         free(tracepoint);
1766         goto out;
1767 }