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 thread *thread,
 293                                struct ip_callchain *chain,
 294                                struct symbol **parent)
 295 {
 296         u8 cpumode = PERF_RECORD_MISC_USER;
 297         unsigned int i;
 298         int err;
 299
 300         callchain_cursor_reset(&callchain_cursor);
 301
 302         if (chain->nr > PERF_MAX_STACK_DEPTH) {
 303                 pr_warning("corrupted callchain. skipping...\n");
 304                 return 0;
 305         }
 306
 307         for (i = 0; i < chain->nr; i++) {
 308                 u64 ip;
 309                 struct addr_location al;
 310
 311                 if (callchain_param.order == ORDER_CALLEE)
 312                         ip = chain->ips[i];
 313                 else
 314                         ip = chain->ips[chain->nr - i - 1];
 315
 316                 if (ip >= PERF_CONTEXT_MAX) {
 317                         switch (ip) {
 318                         case PERF_CONTEXT_HV:
 319                                 cpumode = PERF_RECORD_MISC_HYPERVISOR;  break;
 320                         case PERF_CONTEXT_KERNEL:
 321                                 cpumode = PERF_RECORD_MISC_KERNEL;      break;
 322                         case PERF_CONTEXT_USER:
 323                                 cpumode = PERF_RECORD_MISC_USER;        break;
 324                         default:
 325                                 pr_debug("invalid callchain context: "
 326                                          "%"PRId64"\n", (s64) ip);
 327                                 /*
 328                                  * It seems the callchain is corrupted.
 329                                  * Discard all.
 330                                  */
 331                                 callchain_cursor_reset(&callchain_cursor);
 332                                 return 0;
 333                         }
 334                         continue;
 335                 }
 336
 337                 al.filtered = false;
 338                 thread__find_addr_location(thread, self, cpumode,
 339                                            MAP__FUNCTION, ip, &al, NULL);
 340                 if (al.sym != NULL) {
 341                         if (sort__has_parent && !*parent &&
 342                             symbol__match_parent_regex(al.sym))
 343                                 *parent = al.sym;
 344                         if (!symbol_conf.use_callchain)
 345                                 break;
 346                 }
 347
 348                 err = callchain_cursor_append(&callchain_cursor,
 349                                               ip, al.map, al.sym);
 350                 if (err)
 351                         return err;
 352         }
 353
 354         return 0;
 355 }
 356
 357 static int process_event_synth_tracing_data_stub(union perf_event *event __used,
 358                                                  struct perf_session *session __used)
 359 {
 360         dump_printf(": unhandled!\n");
 361         return 0;
 362 }
 363
 364 static int process_event_synth_attr_stub(union perf_event *event __used,
 365                                          struct perf_evlist **pevlist __used)
 366 {
 367         dump_printf(": unhandled!\n");
 368         return 0;
 369 }
 370
 371 static int process_event_sample_stub(struct perf_tool *tool __used,
 372                                      union perf_event *event __used,
 373                                      struct perf_sample *sample __used,
 374                                      struct perf_evsel *evsel __used,
 375                                      struct machine *machine __used)
 376 {
 377         dump_printf(": unhandled!\n");
 378         return 0;
 379 }
 380
 381 static int process_event_stub(struct perf_tool *tool __used,
 382                               union perf_event *event __used,
 383                               struct perf_sample *sample __used,
 384                               struct machine *machine __used)
 385 {
 386         dump_printf(": unhandled!\n");
 387         return 0;
 388 }
 389
 390 static int process_finished_round_stub(struct perf_tool *tool __used,
 391                                        union perf_event *event __used,
 392                                        struct perf_session *perf_session __used)
 393 {
 394         dump_printf(": unhandled!\n");
 395         return 0;
 396 }
 397
 398 static int process_event_type_stub(struct perf_tool *tool __used,
 399                                    union perf_event *event __used)
 400 {
 401         dump_printf(": unhandled!\n");
 402         return 0;
 403 }
 404
 405 static int process_finished_round(struct perf_tool *tool,
 406                                   union perf_event *event,
 407                                   struct perf_session *session);
 408
 409 static void perf_tool__fill_defaults(struct perf_tool *tool)
 410 {
 411         if (tool->sample == NULL)
 412                 tool->sample = process_event_sample_stub;
 413         if (tool->mmap == NULL)
 414                 tool->mmap = process_event_stub;
 415         if (tool->comm == NULL)
 416                 tool->comm = process_event_stub;
 417         if (tool->fork == NULL)
 418                 tool->fork = process_event_stub;
 419         if (tool->exit == NULL)
 420                 tool->exit = process_event_stub;
 421         if (tool->lost == NULL)
 422                 tool->lost = perf_event__process_lost;
 423         if (tool->read == NULL)
 424                 tool->read = process_event_sample_stub;
 425         if (tool->throttle == NULL)
 426                 tool->throttle = process_event_stub;
 427         if (tool->unthrottle == NULL)
 428                 tool->unthrottle = process_event_stub;
 429         if (tool->attr == NULL)
 430                 tool->attr = process_event_synth_attr_stub;
 431         if (tool->event_type == NULL)
 432                 tool->event_type = process_event_type_stub;
 433         if (tool->tracing_data == NULL)
 434                 tool->tracing_data = process_event_synth_tracing_data_stub;
 435         if (tool->build_id == NULL)
 436                 tool->build_id = process_finished_round_stub;
 437         if (tool->finished_round == NULL) {
 438                 if (tool->ordered_samples)
 439                         tool->finished_round = process_finished_round;
 440                 else
 441                         tool->finished_round = process_finished_round_stub;
 442         }
 443 }
 444
 445 void mem_bswap_32(void *src, int byte_size)
 446 {
 447         u32 *m = src;
 448         while (byte_size > 0) {
 449                 *m = bswap_32(*m);
 450                 byte_size -= sizeof(u32);
 451                 ++m;
 452         }
 453 }
 454
 455 void mem_bswap_64(void *src, int byte_size)
 456 {
 457         u64 *m = src;
 458
 459         while (byte_size > 0) {
 460                 *m = bswap_64(*m);
 461                 byte_size -= sizeof(u64);
 462                 ++m;
 463         }
 464 }
 465
 466 static void swap_sample_id_all(union perf_event *event, void *data)
 467 {
 468         void *end = (void *) event + event->header.size;
 469         int size = end - data;
 470
 471         BUG_ON(size % sizeof(u64));
 472         mem_bswap_64(data, size);
 473 }
 474
 475 static void perf_event__all64_swap(union perf_event *event,
 476                                    bool sample_id_all __used)
 477 {
 478         struct perf_event_header *hdr = &event->header;
 479         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 480 }
 481
 482 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 483 {
 484         event->comm.pid = bswap_32(event->comm.pid);
 485         event->comm.tid = bswap_32(event->comm.tid);
 486
 487         if (sample_id_all) {
 488                 void *data = &event->comm.comm;
 489
 490                 data += ALIGN(strlen(data) + 1, sizeof(u64));
 491                 swap_sample_id_all(event, data);
 492         }
 493 }
 494
 495 static void perf_event__mmap_swap(union perf_event *event,
 496                                   bool sample_id_all)
 497 {
 498         event->mmap.pid   = bswap_32(event->mmap.pid);
 499         event->mmap.tid   = bswap_32(event->mmap.tid);
 500         event->mmap.start = bswap_64(event->mmap.start);
 501         event->mmap.len   = bswap_64(event->mmap.len);
 502         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 503
 504         if (sample_id_all) {
 505                 void *data = &event->mmap.filename;
 506
 507                 data += ALIGN(strlen(data) + 1, sizeof(u64));
 508                 swap_sample_id_all(event, data);
 509         }
 510 }
 511
 512 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 513 {
 514         event->fork.pid  = bswap_32(event->fork.pid);
 515         event->fork.tid  = bswap_32(event->fork.tid);
 516         event->fork.ppid = bswap_32(event->fork.ppid);
 517         event->fork.ptid = bswap_32(event->fork.ptid);
 518         event->fork.time = bswap_64(event->fork.time);
 519
 520         if (sample_id_all)
 521                 swap_sample_id_all(event, &event->fork + 1);
 522 }
 523
 524 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 525 {
 526         event->read.pid          = bswap_32(event->read.pid);
 527         event->read.tid          = bswap_32(event->read.tid);
 528         event->read.value        = bswap_64(event->read.value);
 529         event->read.time_enabled = bswap_64(event->read.time_enabled);
 530         event->read.time_running = bswap_64(event->read.time_running);
 531         event->read.id           = bswap_64(event->read.id);
 532
 533         if (sample_id_all)
 534                 swap_sample_id_all(event, &event->read + 1);
 535 }
 536
 537 static u8 revbyte(u8 b)
 538 {
 539         int rev = (b >> 4) | ((b & 0xf) << 4);
 540         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 541         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 542         return (u8) rev;
 543 }
 544
 545 /*
 546  * XXX this is hack in attempt to carry flags bitfield
 547  * throught endian village. ABI says:
 548  *
 549  * Bit-fields are allocated from right to left (least to most significant)
 550  * on little-endian implementations and from left to right (most to least
 551  * significant) on big-endian implementations.
 552  *
 553  * The above seems to be byte specific, so we need to reverse each
 554  * byte of the bitfield. 'Internet' also says this might be implementation
 555  * specific and we probably need proper fix and carry perf_event_attr
 556  * bitfield flags in separate data file FEAT_ section. Thought this seems
 557  * to work for now.
 558  */
 559 static void swap_bitfield(u8 *p, unsigned len)
 560 {
 561         unsigned i;
 562
 563         for (i = 0; i < len; i++) {
 564                 *p = revbyte(*p);
 565                 p++;
 566         }
 567 }
 568
 569 /* exported for swapping attributes in file header */
 570 void perf_event__attr_swap(struct perf_event_attr *attr)
 571 {
 572         attr->type              = bswap_32(attr->type);
 573         attr->size              = bswap_32(attr->size);
 574         attr->config            = bswap_64(attr->config);
 575         attr->sample_period     = bswap_64(attr->sample_period);
 576         attr->sample_type       = bswap_64(attr->sample_type);
 577         attr->read_format       = bswap_64(attr->read_format);
 578         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 579         attr->bp_type           = bswap_32(attr->bp_type);
 580         attr->bp_addr           = bswap_64(attr->bp_addr);
 581         attr->bp_len            = bswap_64(attr->bp_len);
 582
 583         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 584 }
 585
 586 static void perf_event__hdr_attr_swap(union perf_event *event,
 587                                       bool sample_id_all __used)
 588 {
 589         size_t size;
 590
 591         perf_event__attr_swap(&event->attr.attr);
 592
 593         size = event->header.size;
 594         size -= (void *)&event->attr.id - (void *)event;
 595         mem_bswap_64(event->attr.id, size);
 596 }
 597
 598 static void perf_event__event_type_swap(union perf_event *event,
 599                                         bool sample_id_all __used)
 600 {
 601         event->event_type.event_type.event_id =
 602                 bswap_64(event->event_type.event_type.event_id);
 603 }
 604
 605 static void perf_event__tracing_data_swap(union perf_event *event,
 606                                           bool sample_id_all __used)
 607 {
 608         event->tracing_data.size = bswap_32(event->tracing_data.size);
 609 }
 610
 611 typedef void (*perf_event__swap_op)(union perf_event *event,
 612                                     bool sample_id_all);
 613
 614 static perf_event__swap_op perf_event__swap_ops[] = {
 615         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 616         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 617         [PERF_RECORD_FORK]                = perf_event__task_swap,
 618         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 619         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 620         [PERF_RECORD_READ]                = perf_event__read_swap,
 621         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 622         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 623         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 624         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 625         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 626         [PERF_RECORD_HEADER_MAX]          = NULL,
 627 };
 628
 629 struct sample_queue {
 630         u64                     timestamp;
 631         u64                     file_offset;
 632         union perf_event        *event;
 633         struct list_head        list;
 634 };
 635
 636 static void perf_session_free_sample_buffers(struct perf_session *session)
 637 {
 638         struct ordered_samples *os = &session->ordered_samples;
 639
 640         while (!list_empty(&os->to_free)) {
 641                 struct sample_queue *sq;
 642
 643                 sq = list_entry(os->to_free.next, struct sample_queue, list);
 644                 list_del(&sq->list);
 645                 free(sq);
 646         }
 647 }
 648
 649 static int perf_session_deliver_event(struct perf_session *session,
 650                                       union perf_event *event,
 651                                       struct perf_sample *sample,
 652                                       struct perf_tool *tool,
 653                                       u64 file_offset);
 654
 655 static void flush_sample_queue(struct perf_session *s,
 656                                struct perf_tool *tool)
 657 {
 658         struct ordered_samples *os = &s->ordered_samples;
 659         struct list_head *head = &os->samples;
 660         struct sample_queue *tmp, *iter;
 661         struct perf_sample sample;
 662         u64 limit = os->next_flush;
 663         u64 last_ts = os->last_sample ? os->last_sample->timestamp : 0ULL;
 664         unsigned idx = 0, progress_next = os->nr_samples / 16;
 665         int ret;
 666
 667         if (!tool->ordered_samples || !limit)
 668                 return;
 669
 670         list_for_each_entry_safe(iter, tmp, head, list) {
 671                 if (iter->timestamp > limit)
 672                         break;
 673
 674                 ret = perf_session__parse_sample(s, iter->event, &sample);
 675                 if (ret)
 676                         pr_err("Can't parse sample, err = %d\n", ret);
 677                 else
 678                         perf_session_deliver_event(s, iter->event, &sample, tool,
 679                                                    iter->file_offset);
 680
 681                 os->last_flush = iter->timestamp;
 682                 list_del(&iter->list);
 683                 list_add(&iter->list, &os->sample_cache);
 684                 if (++idx >= progress_next) {
 685                         progress_next += os->nr_samples / 16;
 686                         ui_progress__update(idx, os->nr_samples,
 687                                             "Processing time ordered events...");
 688                 }
 689         }
 690
 691         if (list_empty(head)) {
 692                 os->last_sample = NULL;
 693         } else if (last_ts <= limit) {
 694                 os->last_sample =
 695                         list_entry(head->prev, struct sample_queue, list);
 696         }
 697
 698         os->nr_samples = 0;
 699 }
 700
 701 /*
 702  * When perf record finishes a pass on every buffers, it records this pseudo
 703  * event.
 704  * We record the max timestamp t found in the pass n.
 705  * Assuming these timestamps are monotonic across cpus, we know that if
 706  * a buffer still has events with timestamps below t, they will be all
 707  * available and then read in the pass n + 1.
 708  * Hence when we start to read the pass n + 2, we can safely flush every
 709  * events with timestamps below t.
 710  *
 711  *    ============ PASS n =================
 712  *       CPU 0         |   CPU 1
 713  *                     |
 714  *    cnt1 timestamps  |   cnt2 timestamps
 715  *          1          |         2
 716  *          2          |         3
 717  *          -          |         4  <--- max recorded
 718  *
 719  *    ============ PASS n + 1 ==============
 720  *       CPU 0         |   CPU 1
 721  *                     |
 722  *    cnt1 timestamps  |   cnt2 timestamps
 723  *          3          |         5
 724  *          4          |         6
 725  *          5          |         7 <---- max recorded
 726  *
 727  *      Flush every events below timestamp 4
 728  *
 729  *    ============ PASS n + 2 ==============
 730  *       CPU 0         |   CPU 1
 731  *                     |
 732  *    cnt1 timestamps  |   cnt2 timestamps
 733  *          6          |         8
 734  *          7          |         9
 735  *          -          |         10
 736  *
 737  *      Flush every events below timestamp 7
 738  *      etc...
 739  */
 740 static int process_finished_round(struct perf_tool *tool,
 741                                   union perf_event *event __used,
 742                                   struct perf_session *session)
 743 {
 744         flush_sample_queue(session, tool);
 745         session->ordered_samples.next_flush = session->ordered_samples.max_timestamp;
 746
 747         return 0;
 748 }
 749
 750 /* The queue is ordered by time */
 751 static void __queue_event(struct sample_queue *new, struct perf_session *s)
 752 {
 753         struct ordered_samples *os = &s->ordered_samples;
 754         struct sample_queue *sample = os->last_sample;
 755         u64 timestamp = new->timestamp;
 756         struct list_head *p;
 757
 758         ++os->nr_samples;
 759         os->last_sample = new;
 760
 761         if (!sample) {
 762                 list_add(&new->list, &os->samples);
 763                 os->max_timestamp = timestamp;
 764                 return;
 765         }
 766
 767         /*
 768          * last_sample might point to some random place in the list as it's
 769          * the last queued event. We expect that the new event is close to
 770          * this.
 771          */
 772         if (sample->timestamp <= timestamp) {
 773                 while (sample->timestamp <= timestamp) {
 774                         p = sample->list.next;
 775                         if (p == &os->samples) {
 776                                 list_add_tail(&new->list, &os->samples);
 777                                 os->max_timestamp = timestamp;
 778                                 return;
 779                         }
 780                         sample = list_entry(p, struct sample_queue, list);
 781                 }
 782                 list_add_tail(&new->list, &sample->list);
 783         } else {
 784                 while (sample->timestamp > timestamp) {
 785                         p = sample->list.prev;
 786                         if (p == &os->samples) {
 787                                 list_add(&new->list, &os->samples);
 788                                 return;
 789                         }
 790                         sample = list_entry(p, struct sample_queue, list);
 791                 }
 792                 list_add(&new->list, &sample->list);
 793         }
 794 }
 795
 796 #define MAX_SAMPLE_BUFFER       (64 * 1024 / sizeof(struct sample_queue))
 797
 798 static int perf_session_queue_event(struct perf_session *s, union perf_event *event,
 799                                     struct perf_sample *sample, u64 file_offset)
 800 {
 801         struct ordered_samples *os = &s->ordered_samples;
 802         struct list_head *sc = &os->sample_cache;
 803         u64 timestamp = sample->time;
 804         struct sample_queue *new;
 805
 806         if (!timestamp || timestamp == ~0ULL)
 807                 return -ETIME;
 808
 809         if (timestamp < s->ordered_samples.last_flush) {
 810                 printf("Warning: Timestamp below last timeslice flush\n");
 811                 return -EINVAL;
 812         }
 813
 814         if (!list_empty(sc)) {
 815                 new = list_entry(sc->next, struct sample_queue, list);
 816                 list_del(&new->list);
 817         } else if (os->sample_buffer) {
 818                 new = os->sample_buffer + os->sample_buffer_idx;
 819                 if (++os->sample_buffer_idx == MAX_SAMPLE_BUFFER)
 820                         os->sample_buffer = NULL;
 821         } else {
 822                 os->sample_buffer = malloc(MAX_SAMPLE_BUFFER * sizeof(*new));
 823                 if (!os->sample_buffer)
 824                         return -ENOMEM;
 825                 list_add(&os->sample_buffer->list, &os->to_free);
 826                 os->sample_buffer_idx = 2;
 827                 new = os->sample_buffer + 1;
 828         }
 829
 830         new->timestamp = timestamp;
 831         new->file_offset = file_offset;
 832         new->event = event;
 833
 834         __queue_event(new, s);
 835
 836         return 0;
 837 }
 838
 839 static void callchain__printf(struct perf_sample *sample)
 840 {
 841         unsigned int i;
 842
 843         printf("... chain: nr:%" PRIu64 "\n", sample->callchain->nr);
 844
 845         for (i = 0; i < sample->callchain->nr; i++)
 846                 printf("..... %2d: %016" PRIx64 "\n",
 847                        i, sample->callchain->ips[i]);
 848 }
 849
 850 static void branch_stack__printf(struct perf_sample *sample)
 851 {
 852         uint64_t i;
 853
 854         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 855
 856         for (i = 0; i < sample->branch_stack->nr; i++)
 857                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 858                         i, sample->branch_stack->entries[i].from,
 859                         sample->branch_stack->entries[i].to);
 860 }
 861
 862 static void perf_session__print_tstamp(struct perf_session *session,
 863                                        union perf_event *event,
 864                                        struct perf_sample *sample)
 865 {
 866         if (event->header.type != PERF_RECORD_SAMPLE &&
 867             !session->sample_id_all) {
 868                 fputs("-1 -1 ", stdout);
 869                 return;
 870         }
 871
 872         if ((session->sample_type & PERF_SAMPLE_CPU))
 873                 printf("%u ", sample->cpu);
 874
 875         if (session->sample_type & PERF_SAMPLE_TIME)
 876                 printf("%" PRIu64 " ", sample->time);
 877 }
 878
 879 static void dump_event(struct perf_session *session, union perf_event *event,
 880                        u64 file_offset, struct perf_sample *sample)
 881 {
 882         if (!dump_trace)
 883                 return;
 884
 885         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 886                file_offset, event->header.size, event->header.type);
 887
 888         trace_event(event);
 889
 890         if (sample)
 891                 perf_session__print_tstamp(session, event, sample);
 892
 893         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 894                event->header.size, perf_event__name(event->header.type));
 895 }
 896
 897 static void dump_sample(struct perf_session *session, union perf_event *event,
 898                         struct perf_sample *sample)
 899 {
 900         if (!dump_trace)
 901                 return;
 902
 903         printf("(IP, %d): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 904                event->header.misc, sample->pid, sample->tid, sample->ip,
 905                sample->period, sample->addr);
 906
 907         if (session->sample_type & PERF_SAMPLE_CALLCHAIN)
 908                 callchain__printf(sample);
 909
 910         if (session->sample_type & PERF_SAMPLE_BRANCH_STACK)
 911                 branch_stack__printf(sample);
 912 }
 913
 914 static struct machine *
 915         perf_session__find_machine_for_cpumode(struct perf_session *session,
 916                                                union perf_event *event)
 917 {
 918         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 919
 920         if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) {
 921                 u32 pid;
 922
 923                 if (event->header.type == PERF_RECORD_MMAP)
 924                         pid = event->mmap.pid;
 925                 else
 926                         pid = event->ip.pid;
 927
 928                 return perf_session__find_machine(session, pid);
 929         }
 930
 931         return perf_session__find_host_machine(session);
 932 }
 933
 934 static int perf_session_deliver_event(struct perf_session *session,
 935                                       union perf_event *event,
 936                                       struct perf_sample *sample,
 937                                       struct perf_tool *tool,
 938                                       u64 file_offset)
 939 {
 940         struct perf_evsel *evsel;
 941         struct machine *machine;
 942
 943         dump_event(session, event, file_offset, sample);
 944
 945         evsel = perf_evlist__id2evsel(session->evlist, sample->id);
 946         if (evsel != NULL && event->header.type != PERF_RECORD_SAMPLE) {
 947                 /*
 948                  * XXX We're leaving PERF_RECORD_SAMPLE unnacounted here
 949                  * because the tools right now may apply filters, discarding
 950                  * some of the samples. For consistency, in the future we
 951                  * should have something like nr_filtered_samples and remove
 952                  * the sample->period from total_sample_period, etc, KISS for
 953                  * now tho.
 954                  *
 955                  * Also testing against NULL allows us to handle files without
 956                  * attr.sample_id_all and/or without PERF_SAMPLE_ID. In the
 957                  * future probably it'll be a good idea to restrict event
 958                  * processing via perf_session to files with both set.
 959                  */
 960                 hists__inc_nr_events(&evsel->hists, event->header.type);
 961         }
 962
 963         machine = perf_session__find_machine_for_cpumode(session, event);
 964
 965         switch (event->header.type) {
 966         case PERF_RECORD_SAMPLE:
 967                 dump_sample(session, event, sample);
 968                 if (evsel == NULL) {
 969                         ++session->hists.stats.nr_unknown_id;
 970                         return 0;
 971                 }
 972                 if (machine == NULL) {
 973                         ++session->hists.stats.nr_unprocessable_samples;
 974                         return 0;
 975                 }
 976                 return tool->sample(tool, event, sample, evsel, machine);
 977         case PERF_RECORD_MMAP:
 978                 return tool->mmap(tool, event, sample, machine);
 979         case PERF_RECORD_COMM:
 980                 return tool->comm(tool, event, sample, machine);
 981         case PERF_RECORD_FORK:
 982                 return tool->fork(tool, event, sample, machine);
 983         case PERF_RECORD_EXIT:
 984                 return tool->exit(tool, event, sample, machine);
 985         case PERF_RECORD_LOST:
 986                 if (tool->lost == perf_event__process_lost)
 987                         session->hists.stats.total_lost += event->lost.lost;
 988                 return tool->lost(tool, event, sample, machine);
 989         case PERF_RECORD_READ:
 990                 return tool->read(tool, event, sample, evsel, machine);
 991         case PERF_RECORD_THROTTLE:
 992                 return tool->throttle(tool, event, sample, machine);
 993         case PERF_RECORD_UNTHROTTLE:
 994                 return tool->unthrottle(tool, event, sample, machine);
 995         default:
 996                 ++session->hists.stats.nr_unknown_events;
 997                 return -1;
 998         }
 999 }
1000
1001 static int perf_session__preprocess_sample(struct perf_session *session,
1002                                            union perf_event *event, struct perf_sample *sample)
1003 {
1004         if (event->header.type != PERF_RECORD_SAMPLE ||
1005             !(session->sample_type & PERF_SAMPLE_CALLCHAIN))
1006                 return 0;
1007
1008         if (!ip_callchain__valid(sample->callchain, event)) {
1009                 pr_debug("call-chain problem with event, skipping it.\n");
1010                 ++session->hists.stats.nr_invalid_chains;
1011                 session->hists.stats.total_invalid_chains += sample->period;
1012                 return -EINVAL;
1013         }
1014         return 0;
1015 }
1016
1017 static int perf_session__process_user_event(struct perf_session *session, union perf_event *event,
1018                                             struct perf_tool *tool, u64 file_offset)
1019 {
1020         int err;
1021
1022         dump_event(session, event, file_offset, NULL);
1023
1024         /* These events are processed right away */
1025         switch (event->header.type) {
1026         case PERF_RECORD_HEADER_ATTR:
1027                 err = tool->attr(event, &session->evlist);
1028                 if (err == 0)
1029                         perf_session__update_sample_type(session);
1030                 return err;
1031         case PERF_RECORD_HEADER_EVENT_TYPE:
1032                 return tool->event_type(tool, event);
1033         case PERF_RECORD_HEADER_TRACING_DATA:
1034                 /* setup for reading amidst mmap */
1035                 lseek(session->fd, file_offset, SEEK_SET);
1036                 return tool->tracing_data(event, session);
1037         case PERF_RECORD_HEADER_BUILD_ID:
1038                 return tool->build_id(tool, event, session);
1039         case PERF_RECORD_FINISHED_ROUND:
1040                 return tool->finished_round(tool, event, session);
1041         default:
1042                 return -EINVAL;
1043         }
1044 }
1045
1046 static void event_swap(union perf_event *event, bool sample_id_all)
1047 {
1048         perf_event__swap_op swap;
1049
1050         swap = perf_event__swap_ops[event->header.type];
1051         if (swap)
1052                 swap(event, sample_id_all);
1053 }
1054
1055 static int perf_session__process_event(struct perf_session *session,
1056                                        union perf_event *event,
1057                                        struct perf_tool *tool,
1058                                        u64 file_offset)
1059 {
1060         struct perf_sample sample;
1061         int ret;
1062
1063         if (session->header.needs_swap)
1064                 event_swap(event, session->sample_id_all);
1065
1066         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1067                 return -EINVAL;
1068
1069         hists__inc_nr_events(&session->hists, event->header.type);
1070
1071         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1072                 return perf_session__process_user_event(session, event, tool, file_offset);
1073
1074         /*
1075          * For all kernel events we get the sample data
1076          */
1077         ret = perf_session__parse_sample(session, event, &sample);
1078         if (ret)
1079                 return ret;
1080
1081         /* Preprocess sample records - precheck callchains */
1082         if (perf_session__preprocess_sample(session, event, &sample))
1083                 return 0;
1084
1085         if (tool->ordered_samples) {
1086                 ret = perf_session_queue_event(session, event, &sample,
1087                                                file_offset);
1088                 if (ret != -ETIME)
1089                         return ret;
1090         }
1091
1092         return perf_session_deliver_event(session, event, &sample, tool,
1093                                           file_offset);
1094 }
1095
1096 void perf_event_header__bswap(struct perf_event_header *self)
1097 {
1098         self->type = bswap_32(self->type);
1099         self->misc = bswap_16(self->misc);
1100         self->size = bswap_16(self->size);
1101 }
1102
1103 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1104 {
1105         return machine__findnew_thread(&session->host_machine, pid);
1106 }
1107
1108 static struct thread *perf_session__register_idle_thread(struct perf_session *self)
1109 {
1110         struct thread *thread = perf_session__findnew(self, 0);
1111
1112         if (thread == NULL || thread__set_comm(thread, "swapper")) {
1113                 pr_err("problem inserting idle task.\n");
1114                 thread = NULL;
1115         }
1116
1117         return thread;
1118 }
1119
1120 static void perf_session__warn_about_errors(const struct perf_session *session,
1121                                             const struct perf_tool *tool)
1122 {
1123         if (tool->lost == perf_event__process_lost &&
1124             session->hists.stats.nr_events[PERF_RECORD_LOST] != 0) {
1125                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1126                             "Check IO/CPU overload!\n\n",
1127                             session->hists.stats.nr_events[0],
1128                             session->hists.stats.nr_events[PERF_RECORD_LOST]);
1129         }
1130
1131         if (session->hists.stats.nr_unknown_events != 0) {
1132                 ui__warning("Found %u unknown events!\n\n"
1133                             "Is this an older tool processing a perf.data "
1134                             "file generated by a more recent tool?\n\n"
1135                             "If that is not the case, consider "
1136                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1137                             session->hists.stats.nr_unknown_events);
1138         }
1139
1140         if (session->hists.stats.nr_unknown_id != 0) {
1141                 ui__warning("%u samples with id not present in the header\n",
1142                             session->hists.stats.nr_unknown_id);
1143         }
1144
1145         if (session->hists.stats.nr_invalid_chains != 0) {
1146                 ui__warning("Found invalid callchains!\n\n"
1147                             "%u out of %u events were discarded for this reason.\n\n"
1148                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1149                             session->hists.stats.nr_invalid_chains,
1150                             session->hists.stats.nr_events[PERF_RECORD_SAMPLE]);
1151         }
1152
1153         if (session->hists.stats.nr_unprocessable_samples != 0) {
1154                 ui__warning("%u unprocessable samples recorded.\n"
1155                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1156                             session->hists.stats.nr_unprocessable_samples);
1157         }
1158 }
1159
1160 #define session_done()  (*(volatile int *)(&session_done))
1161 volatile int session_done;
1162
1163 static int __perf_session__process_pipe_events(struct perf_session *self,
1164                                                struct perf_tool *tool)
1165 {
1166         union perf_event *event;
1167         uint32_t size, cur_size = 0;
1168         void *buf = NULL;
1169         int skip = 0;
1170         u64 head;
1171         int err;
1172         void *p;
1173
1174         perf_tool__fill_defaults(tool);
1175
1176         head = 0;
1177         cur_size = sizeof(union perf_event);
1178
1179         buf = malloc(cur_size);
1180         if (!buf)
1181                 return -errno;
1182 more:
1183         event = buf;
1184         err = readn(self->fd, event, sizeof(struct perf_event_header));
1185         if (err <= 0) {
1186                 if (err == 0)
1187                         goto done;
1188
1189                 pr_err("failed to read event header\n");
1190                 goto out_err;
1191         }
1192
1193         if (self->header.needs_swap)
1194                 perf_event_header__bswap(&event->header);
1195
1196         size = event->header.size;
1197         if (size == 0)
1198                 size = 8;
1199
1200         if (size > cur_size) {
1201                 void *new = realloc(buf, size);
1202                 if (!new) {
1203                         pr_err("failed to allocate memory to read event\n");
1204                         goto out_err;
1205                 }
1206                 buf = new;
1207                 cur_size = size;
1208                 event = buf;
1209         }
1210         p = event;
1211         p += sizeof(struct perf_event_header);
1212
1213         if (size - sizeof(struct perf_event_header)) {
1214                 err = readn(self->fd, p, size - sizeof(struct perf_event_header));
1215                 if (err <= 0) {
1216                         if (err == 0) {
1217                                 pr_err("unexpected end of event stream\n");
1218                                 goto done;
1219                         }
1220
1221                         pr_err("failed to read event data\n");
1222                         goto out_err;
1223                 }
1224         }
1225
1226         if ((skip = perf_session__process_event(self, event, tool, head)) < 0) {
1227                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1228                        head, event->header.size, event->header.type);
1229                 err = -EINVAL;
1230                 goto out_err;
1231         }
1232
1233         head += size;
1234
1235         if (skip > 0)
1236                 head += skip;
1237
1238         if (!session_done())
1239                 goto more;
1240 done:
1241         err = 0;
1242 out_err:
1243         free(buf);
1244         perf_session__warn_about_errors(self, tool);
1245         perf_session_free_sample_buffers(self);
1246         return err;
1247 }
1248
1249 static union perf_event *
1250 fetch_mmaped_event(struct perf_session *session,
1251                    u64 head, size_t mmap_size, char *buf)
1252 {
1253         union perf_event *event;
1254
1255         /*
1256          * Ensure we have enough space remaining to read
1257          * the size of the event in the headers.
1258          */
1259         if (head + sizeof(event->header) > mmap_size)
1260                 return NULL;
1261
1262         event = (union perf_event *)(buf + head);
1263
1264         if (session->header.needs_swap)
1265                 perf_event_header__bswap(&event->header);
1266
1267         if (head + event->header.size > mmap_size)
1268                 return NULL;
1269
1270         return event;
1271 }
1272
1273 int __perf_session__process_events(struct perf_session *session,
1274                                    u64 data_offset, u64 data_size,
1275                                    u64 file_size, struct perf_tool *tool)
1276 {
1277         u64 head, page_offset, file_offset, file_pos, progress_next;
1278         int err, mmap_prot, mmap_flags, map_idx = 0;
1279         size_t  page_size, mmap_size;
1280         char *buf, *mmaps[8];
1281         union perf_event *event;
1282         uint32_t size;
1283
1284         perf_tool__fill_defaults(tool);
1285
1286         page_size = sysconf(_SC_PAGESIZE);
1287
1288         page_offset = page_size * (data_offset / page_size);
1289         file_offset = page_offset;
1290         head = data_offset - page_offset;
1291
1292         if (data_offset + data_size < file_size)
1293                 file_size = data_offset + data_size;
1294
1295         progress_next = file_size / 16;
1296
1297         mmap_size = session->mmap_window;
1298         if (mmap_size > file_size)
1299                 mmap_size = file_size;
1300
1301         memset(mmaps, 0, sizeof(mmaps));
1302
1303         mmap_prot  = PROT_READ;
1304         mmap_flags = MAP_SHARED;
1305
1306         if (session->header.needs_swap) {
1307                 mmap_prot  |= PROT_WRITE;
1308                 mmap_flags = MAP_PRIVATE;
1309         }
1310 remap:
1311         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, session->fd,
1312                    file_offset);
1313         if (buf == MAP_FAILED) {
1314                 pr_err("failed to mmap file\n");
1315                 err = -errno;
1316                 goto out_err;
1317         }
1318         mmaps[map_idx] = buf;
1319         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1320         file_pos = file_offset + head;
1321
1322 more:
1323         event = fetch_mmaped_event(session, head, mmap_size, buf);
1324         if (!event) {
1325                 if (mmaps[map_idx]) {
1326                         munmap(mmaps[map_idx], mmap_size);
1327                         mmaps[map_idx] = NULL;
1328                 }
1329
1330                 page_offset = page_size * (head / page_size);
1331                 file_offset += page_offset;
1332                 head -= page_offset;
1333                 goto remap;
1334         }
1335
1336         size = event->header.size;
1337
1338         if (size == 0 ||
1339             perf_session__process_event(session, event, tool, file_pos) < 0) {
1340                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1341                        file_offset + head, event->header.size,
1342                        event->header.type);
1343                 err = -EINVAL;
1344                 goto out_err;
1345         }
1346
1347         head += size;
1348         file_pos += size;
1349
1350         if (file_pos >= progress_next) {
1351                 progress_next += file_size / 16;
1352                 ui_progress__update(file_pos, file_size,
1353                                     "Processing events...");
1354         }
1355
1356         if (file_pos < file_size)
1357                 goto more;
1358
1359         err = 0;
1360         /* do the final flush for ordered samples */
1361         session->ordered_samples.next_flush = ULLONG_MAX;
1362         flush_sample_queue(session, tool);
1363 out_err:
1364         perf_session__warn_about_errors(session, tool);
1365         perf_session_free_sample_buffers(session);
1366         return err;
1367 }
1368
1369 int perf_session__process_events(struct perf_session *self,
1370                                  struct perf_tool *tool)
1371 {
1372         int err;
1373
1374         if (perf_session__register_idle_thread(self) == NULL)
1375                 return -ENOMEM;
1376
1377         if (!self->fd_pipe)
1378                 err = __perf_session__process_events(self,
1379                                                      self->header.data_offset,
1380                                                      self->header.data_size,
1381                                                      self->size, tool);
1382         else
1383                 err = __perf_session__process_pipe_events(self, tool);
1384
1385         return err;
1386 }
1387
1388 bool perf_session__has_traces(struct perf_session *self, const char *msg)
1389 {
1390         if (!(self->sample_type & PERF_SAMPLE_RAW)) {
1391                 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1392                 return false;
1393         }
1394
1395         return true;
1396 }
1397
1398 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1399                                      const char *symbol_name, u64 addr)
1400 {
1401         char *bracket;
1402         enum map_type i;
1403         struct ref_reloc_sym *ref;
1404
1405         ref = zalloc(sizeof(struct ref_reloc_sym));
1406         if (ref == NULL)
1407                 return -ENOMEM;
1408
1409         ref->name = strdup(symbol_name);
1410         if (ref->name == NULL) {
1411                 free(ref);
1412                 return -ENOMEM;
1413         }
1414
1415         bracket = strchr(ref->name, ']');
1416         if (bracket)
1417                 *bracket = '\0';
1418
1419         ref->addr = addr;
1420
1421         for (i = 0; i < MAP__NR_TYPES; ++i) {
1422                 struct kmap *kmap = map__kmap(maps[i]);
1423                 kmap->ref_reloc_sym = ref;
1424         }
1425
1426         return 0;
1427 }
1428
1429 size_t perf_session__fprintf_dsos(struct perf_session *self, FILE *fp)
1430 {
1431         return __dsos__fprintf(&self->host_machine.kernel_dsos, fp) +
1432                __dsos__fprintf(&self->host_machine.user_dsos, fp) +
1433                machines__fprintf_dsos(&self->machines, fp);
1434 }
1435
1436 size_t perf_session__fprintf_dsos_buildid(struct perf_session *self, FILE *fp,
1437                                           bool with_hits)
1438 {
1439         size_t ret = machine__fprintf_dsos_buildid(&self->host_machine, fp, with_hits);
1440         return ret + machines__fprintf_dsos_buildid(&self->machines, fp, with_hits);
1441 }
1442
1443 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1444 {
1445         struct perf_evsel *pos;
1446         size_t ret = fprintf(fp, "Aggregated stats:\n");
1447
1448         ret += hists__fprintf_nr_events(&session->hists, fp);
1449
1450         list_for_each_entry(pos, &session->evlist->entries, node) {
1451                 ret += fprintf(fp, "%s stats:\n", perf_evsel__name(pos));
1452                 ret += hists__fprintf_nr_events(&pos->hists, fp);
1453         }
1454
1455         return ret;
1456 }
1457
1458 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1459 {
1460         /*
1461          * FIXME: Here we have to actually print all the machines in this
1462          * session, not just the host...
1463          */
1464         return machine__fprintf(&session->host_machine, fp);
1465 }
1466
1467 void perf_session__remove_thread(struct perf_session *session,
1468                                  struct thread *th)
1469 {
1470         /*
1471          * FIXME: This one makes no sense, we need to remove the thread from
1472          * the machine it belongs to, perf_session can have many machines, so
1473          * doing it always on ->host_machine is wrong.  Fix when auditing all
1474          * the 'perf kvm' code.
1475          */
1476         machine__remove_thread(&session->host_machine, th);
1477 }
1478
1479 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1480                                               unsigned int type)
1481 {
1482         struct perf_evsel *pos;
1483
1484         list_for_each_entry(pos, &session->evlist->entries, node) {
1485                 if (pos->attr.type == type)
1486                         return pos;
1487         }
1488         return NULL;
1489 }
1490
1491 void perf_event__print_ip(union perf_event *event, struct perf_sample *sample,
1492                           struct machine *machine, int print_sym,
1493                           int print_dso, int print_symoffset)
1494 {
1495         struct addr_location al;
1496         struct callchain_cursor_node *node;
1497
1498         if (perf_event__preprocess_sample(event, machine, &al, sample,
1499                                           NULL) < 0) {
1500                 error("problem processing %d event, skipping it.\n",
1501                         event->header.type);
1502                 return;
1503         }
1504
1505         if (symbol_conf.use_callchain && sample->callchain) {
1506
1507                 if (machine__resolve_callchain(machine, al.thread,
1508                                                 sample->callchain, NULL) != 0) {
1509                         if (verbose)
1510                                 error("Failed to resolve callchain. Skipping\n");
1511                         return;
1512                 }
1513                 callchain_cursor_commit(&callchain_cursor);
1514
1515                 while (1) {
1516                         node = callchain_cursor_current(&callchain_cursor);
1517                         if (!node)
1518                                 break;
1519
1520                         printf("\t%16" PRIx64, node->ip);
1521                         if (print_sym) {
1522                                 printf(" ");
1523                                 symbol__fprintf_symname(node->sym, stdout);
1524                         }
1525                         if (print_dso) {
1526                                 printf(" (");
1527                                 map__fprintf_dsoname(node->map, stdout);
1528                                 printf(")");
1529                         }
1530                         printf("\n");
1531
1532                         callchain_cursor_advance(&callchain_cursor);
1533                 }
1534
1535         } else {
1536                 printf("%16" PRIx64, sample->ip);
1537                 if (print_sym) {
1538                         printf(" ");
1539                         if (print_symoffset)
1540                                 symbol__fprintf_symname_offs(al.sym, &al,
1541                                                              stdout);
1542                         else
1543                                 symbol__fprintf_symname(al.sym, stdout);
1544                 }
1545
1546                 if (print_dso) {
1547                         printf(" (");
1548                         map__fprintf_dsoname(al.map, stdout);
1549                         printf(")");
1550                 }
1551         }
1552 }
1553
1554 int perf_session__cpu_bitmap(struct perf_session *session,
1555                              const char *cpu_list, unsigned long *cpu_bitmap)
1556 {
1557         int i;
1558         struct cpu_map *map;
1559
1560         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1561                 struct perf_evsel *evsel;
1562
1563                 evsel = perf_session__find_first_evtype(session, i);
1564                 if (!evsel)
1565                         continue;
1566
1567                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1568                         pr_err("File does not contain CPU events. "
1569                                "Remove -c option to proceed.\n");
1570                         return -1;
1571                 }
1572         }
1573
1574         map = cpu_map__new(cpu_list);
1575         if (map == NULL) {
1576                 pr_err("Invalid cpu_list\n");
1577                 return -1;
1578         }
1579
1580         for (i = 0; i < map->nr; i++) {
1581                 int cpu = map->map[i];
1582
1583                 if (cpu >= MAX_NR_CPUS) {
1584                         pr_err("Requested CPU %d too large. "
1585                                "Consider raising MAX_NR_CPUS\n", cpu);
1586                         return -1;
1587                 }
1588
1589                 set_bit(cpu, cpu_bitmap);
1590         }
1591
1592         return 0;
1593 }
1594
1595 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1596                                 bool full)
1597 {
1598         struct stat st;
1599         int ret;
1600
1601         if (session == NULL || fp == NULL)
1602                 return;
1603
1604         ret = fstat(session->fd, &st);
1605         if (ret == -1)
1606                 return;
1607
1608         fprintf(fp, "# ========\n");
1609         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1610         perf_header__fprintf_info(session, fp, full);
1611         fprintf(fp, "# ========\n#\n");
1612 }