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