tools/perf/util/session.c

   1 #include <linux/kernel.h>
   2 #include <traceevent/event-parse.h>
   3
   4 #include <byteswap.h>
   5 #include <unistd.h>
   6 #include <sys/types.h>
   7 #include <sys/mman.h>
   8
   9 #include "evlist.h"
  10 #include "evsel.h"
  11 #include "session.h"
  12 #include "tool.h"
  13 #include "sort.h"
  14 #include "util.h"
  15 #include "cpumap.h"
  16 #include "perf_regs.h"
  17 #include "asm/bug.h"
  18 #include "auxtrace.h"
  19
  20 static int perf_session__deliver_event(struct perf_session *session,
  21                                        union perf_event *event,
  22                                        struct perf_sample *sample,
  23                                        struct perf_tool *tool,
  24                                        u64 file_offset);
  25
  26 static int perf_session__open(struct perf_session *session)
  27 {
  28         struct perf_data_file *file = session->file;
  29
  30         if (perf_session__read_header(session) < 0) {
  31                 pr_err("incompatible file format (rerun with -v to learn more)");
  32                 return -1;
  33         }
  34
  35         if (perf_data_file__is_pipe(file))
  36                 return 0;
  37
  38         if (!perf_evlist__valid_sample_type(session->evlist)) {
  39                 pr_err("non matching sample_type");
  40                 return -1;
  41         }
  42
  43         if (!perf_evlist__valid_sample_id_all(session->evlist)) {
  44                 pr_err("non matching sample_id_all");
  45                 return -1;
  46         }
  47
  48         if (!perf_evlist__valid_read_format(session->evlist)) {
  49                 pr_err("non matching read_format");
  50                 return -1;
  51         }
  52
  53         return 0;
  54 }
  55
  56 void perf_session__set_id_hdr_size(struct perf_session *session)
  57 {
  58         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
  59
  60         machines__set_id_hdr_size(&session->machines, id_hdr_size);
  61 }
  62
  63 int perf_session__create_kernel_maps(struct perf_session *session)
  64 {
  65         int ret = machine__create_kernel_maps(&session->machines.host);
  66
  67         if (ret >= 0)
  68                 ret = machines__create_guest_kernel_maps(&session->machines);
  69         return ret;
  70 }
  71
  72 static void perf_session__destroy_kernel_maps(struct perf_session *session)
  73 {
  74         machines__destroy_kernel_maps(&session->machines);
  75 }
  76
  77 static bool perf_session__has_comm_exec(struct perf_session *session)
  78 {
  79         struct perf_evsel *evsel;
  80
  81         evlist__for_each(session->evlist, evsel) {
  82                 if (evsel->attr.comm_exec)
  83                         return true;
  84         }
  85
  86         return false;
  87 }
  88
  89 static void perf_session__set_comm_exec(struct perf_session *session)
  90 {
  91         bool comm_exec = perf_session__has_comm_exec(session);
  92
  93         machines__set_comm_exec(&session->machines, comm_exec);
  94 }
  95
  96 static int ordered_events__deliver_event(struct ordered_events *oe,
  97                                          struct ordered_event *event)
  98 {
  99         struct perf_sample sample;
 100         struct perf_session *session = container_of(oe, struct perf_session,
 101                                                     ordered_events);
 102         int ret = perf_evlist__parse_sample(session->evlist, event->event, &sample);
 103
 104         if (ret) {
 105                 pr_err("Can't parse sample, err = %d\n", ret);
 106                 return ret;
 107         }
 108
 109         return perf_session__deliver_event(session, event->event, &sample,
 110                                            session->tool, event->file_offset);
 111 }
 112
 113 struct perf_session *perf_session__new(struct perf_data_file *file,
 114                                        bool repipe, struct perf_tool *tool)
 115 {
 116         struct perf_session *session = zalloc(sizeof(*session));
 117
 118         if (!session)
 119                 goto out;
 120
 121         session->repipe = repipe;
 122         session->tool   = tool;
 123         machines__init(&session->machines);
 124         ordered_events__init(&session->ordered_events, ordered_events__deliver_event);
 125
 126         if (file) {
 127                 if (perf_data_file__open(file))
 128                         goto out_delete;
 129
 130                 session->file = file;
 131
 132                 if (perf_data_file__is_read(file)) {
 133                         if (perf_session__open(session) < 0)
 134                                 goto out_close;
 135
 136                         perf_session__set_id_hdr_size(session);
 137                         perf_session__set_comm_exec(session);
 138                 }
 139         }
 140
 141         if (!file || perf_data_file__is_write(file)) {
 142                 /*
 143                  * In O_RDONLY mode this will be performed when reading the
 144                  * kernel MMAP event, in perf_event__process_mmap().
 145                  */
 146                 if (perf_session__create_kernel_maps(session) < 0)
 147                         pr_warning("Cannot read kernel map\n");
 148         }
 149
 150         if (tool && tool->ordering_requires_timestamps &&
 151             tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) {
 152                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 153                 tool->ordered_events = false;
 154         }
 155
 156         return session;
 157
 158  out_close:
 159         perf_data_file__close(file);
 160  out_delete:
 161         perf_session__delete(session);
 162  out:
 163         return NULL;
 164 }
 165
 166 static void perf_session__delete_threads(struct perf_session *session)
 167 {
 168         machine__delete_threads(&session->machines.host);
 169 }
 170
 171 static void perf_session_env__delete(struct perf_session_env *env)
 172 {
 173         zfree(&env->hostname);
 174         zfree(&env->os_release);
 175         zfree(&env->version);
 176         zfree(&env->arch);
 177         zfree(&env->cpu_desc);
 178         zfree(&env->cpuid);
 179
 180         zfree(&env->cmdline);
 181         zfree(&env->sibling_cores);
 182         zfree(&env->sibling_threads);
 183         zfree(&env->numa_nodes);
 184         zfree(&env->pmu_mappings);
 185 }
 186
 187 void perf_session__delete(struct perf_session *session)
 188 {
 189         auxtrace__free(session);
 190         perf_session__destroy_kernel_maps(session);
 191         perf_session__delete_threads(session);
 192         perf_session_env__delete(&session->header.env);
 193         machines__exit(&session->machines);
 194         if (session->file)
 195                 perf_data_file__close(session->file);
 196         free(session);
 197 }
 198
 199 static int process_event_synth_tracing_data_stub(struct perf_tool *tool
 200                                                  __maybe_unused,
 201                                                  union perf_event *event
 202                                                  __maybe_unused,
 203                                                  struct perf_session *session
 204                                                 __maybe_unused)
 205 {
 206         dump_printf(": unhandled!\n");
 207         return 0;
 208 }
 209
 210 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
 211                                          union perf_event *event __maybe_unused,
 212                                          struct perf_evlist **pevlist
 213                                          __maybe_unused)
 214 {
 215         dump_printf(": unhandled!\n");
 216         return 0;
 217 }
 218
 219 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
 220                                      union perf_event *event __maybe_unused,
 221                                      struct perf_sample *sample __maybe_unused,
 222                                      struct perf_evsel *evsel __maybe_unused,
 223                                      struct machine *machine __maybe_unused)
 224 {
 225         dump_printf(": unhandled!\n");
 226         return 0;
 227 }
 228
 229 static int process_event_stub(struct perf_tool *tool __maybe_unused,
 230                               union perf_event *event __maybe_unused,
 231                               struct perf_sample *sample __maybe_unused,
 232                               struct machine *machine __maybe_unused)
 233 {
 234         dump_printf(": unhandled!\n");
 235         return 0;
 236 }
 237
 238 static int process_build_id_stub(struct perf_tool *tool __maybe_unused,
 239                                  union perf_event *event __maybe_unused,
 240                                  struct perf_session *session __maybe_unused)
 241 {
 242         dump_printf(": unhandled!\n");
 243         return 0;
 244 }
 245
 246 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
 247                                        union perf_event *event __maybe_unused,
 248                                        struct ordered_events *oe __maybe_unused)
 249 {
 250         dump_printf(": unhandled!\n");
 251         return 0;
 252 }
 253
 254 static int process_finished_round(struct perf_tool *tool,
 255                                   union perf_event *event,
 256                                   struct ordered_events *oe);
 257
 258 static int process_id_index_stub(struct perf_tool *tool __maybe_unused,
 259                                  union perf_event *event __maybe_unused,
 260                                  struct perf_session *perf_session
 261                                  __maybe_unused)
 262 {
 263         dump_printf(": unhandled!\n");
 264         return 0;
 265 }
 266
 267 static int process_event_auxtrace_info_stub(struct perf_tool *tool __maybe_unused,
 268                                 union perf_event *event __maybe_unused,
 269                                 struct perf_session *session __maybe_unused)
 270 {
 271         dump_printf(": unhandled!\n");
 272         return 0;
 273 }
 274
 275 static int skipn(int fd, off_t n)
 276 {
 277         char buf[4096];
 278         ssize_t ret;
 279
 280         while (n > 0) {
 281                 ret = read(fd, buf, min(n, (off_t)sizeof(buf)));
 282                 if (ret <= 0)
 283                         return ret;
 284                 n -= ret;
 285         }
 286
 287         return 0;
 288 }
 289
 290 static s64 process_event_auxtrace_stub(struct perf_tool *tool __maybe_unused,
 291                                        union perf_event *event,
 292                                        struct perf_session *session
 293                                        __maybe_unused)
 294 {
 295         dump_printf(": unhandled!\n");
 296         if (perf_data_file__is_pipe(session->file))
 297                 skipn(perf_data_file__fd(session->file), event->auxtrace.size);
 298         return event->auxtrace.size;
 299 }
 300
 301 static
 302 int process_event_auxtrace_error_stub(struct perf_tool *tool __maybe_unused,
 303                                       union perf_event *event __maybe_unused,
 304                                       struct perf_session *session __maybe_unused)
 305 {
 306         dump_printf(": unhandled!\n");
 307         return 0;
 308 }
 309
 310 void perf_tool__fill_defaults(struct perf_tool *tool)
 311 {
 312         if (tool->sample == NULL)
 313                 tool->sample = process_event_sample_stub;
 314         if (tool->mmap == NULL)
 315                 tool->mmap = process_event_stub;
 316         if (tool->mmap2 == NULL)
 317                 tool->mmap2 = process_event_stub;
 318         if (tool->comm == NULL)
 319                 tool->comm = process_event_stub;
 320         if (tool->fork == NULL)
 321                 tool->fork = process_event_stub;
 322         if (tool->exit == NULL)
 323                 tool->exit = process_event_stub;
 324         if (tool->lost == NULL)
 325                 tool->lost = perf_event__process_lost;
 326         if (tool->read == NULL)
 327                 tool->read = process_event_sample_stub;
 328         if (tool->throttle == NULL)
 329                 tool->throttle = process_event_stub;
 330         if (tool->unthrottle == NULL)
 331                 tool->unthrottle = process_event_stub;
 332         if (tool->attr == NULL)
 333                 tool->attr = process_event_synth_attr_stub;
 334         if (tool->tracing_data == NULL)
 335                 tool->tracing_data = process_event_synth_tracing_data_stub;
 336         if (tool->build_id == NULL)
 337                 tool->build_id = process_build_id_stub;
 338         if (tool->finished_round == NULL) {
 339                 if (tool->ordered_events)
 340                         tool->finished_round = process_finished_round;
 341                 else
 342                         tool->finished_round = process_finished_round_stub;
 343         }
 344         if (tool->id_index == NULL)
 345                 tool->id_index = process_id_index_stub;
 346         if (tool->auxtrace_info == NULL)
 347                 tool->auxtrace_info = process_event_auxtrace_info_stub;
 348         if (tool->auxtrace == NULL)
 349                 tool->auxtrace = process_event_auxtrace_stub;
 350         if (tool->auxtrace_error == NULL)
 351                 tool->auxtrace_error = process_event_auxtrace_error_stub;
 352 }
 353
 354 static void swap_sample_id_all(union perf_event *event, void *data)
 355 {
 356         void *end = (void *) event + event->header.size;
 357         int size = end - data;
 358
 359         BUG_ON(size % sizeof(u64));
 360         mem_bswap_64(data, size);
 361 }
 362
 363 static void perf_event__all64_swap(union perf_event *event,
 364                                    bool sample_id_all __maybe_unused)
 365 {
 366         struct perf_event_header *hdr = &event->header;
 367         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 368 }
 369
 370 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 371 {
 372         event->comm.pid = bswap_32(event->comm.pid);
 373         event->comm.tid = bswap_32(event->comm.tid);
 374
 375         if (sample_id_all) {
 376                 void *data = &event->comm.comm;
 377
 378                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 379                 swap_sample_id_all(event, data);
 380         }
 381 }
 382
 383 static void perf_event__mmap_swap(union perf_event *event,
 384                                   bool sample_id_all)
 385 {
 386         event->mmap.pid   = bswap_32(event->mmap.pid);
 387         event->mmap.tid   = bswap_32(event->mmap.tid);
 388         event->mmap.start = bswap_64(event->mmap.start);
 389         event->mmap.len   = bswap_64(event->mmap.len);
 390         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 391
 392         if (sample_id_all) {
 393                 void *data = &event->mmap.filename;
 394
 395                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 396                 swap_sample_id_all(event, data);
 397         }
 398 }
 399
 400 static void perf_event__mmap2_swap(union perf_event *event,
 401                                   bool sample_id_all)
 402 {
 403         event->mmap2.pid   = bswap_32(event->mmap2.pid);
 404         event->mmap2.tid   = bswap_32(event->mmap2.tid);
 405         event->mmap2.start = bswap_64(event->mmap2.start);
 406         event->mmap2.len   = bswap_64(event->mmap2.len);
 407         event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
 408         event->mmap2.maj   = bswap_32(event->mmap2.maj);
 409         event->mmap2.min   = bswap_32(event->mmap2.min);
 410         event->mmap2.ino   = bswap_64(event->mmap2.ino);
 411
 412         if (sample_id_all) {
 413                 void *data = &event->mmap2.filename;
 414
 415                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 416                 swap_sample_id_all(event, data);
 417         }
 418 }
 419 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 420 {
 421         event->fork.pid  = bswap_32(event->fork.pid);
 422         event->fork.tid  = bswap_32(event->fork.tid);
 423         event->fork.ppid = bswap_32(event->fork.ppid);
 424         event->fork.ptid = bswap_32(event->fork.ptid);
 425         event->fork.time = bswap_64(event->fork.time);
 426
 427         if (sample_id_all)
 428                 swap_sample_id_all(event, &event->fork + 1);
 429 }
 430
 431 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 432 {
 433         event->read.pid          = bswap_32(event->read.pid);
 434         event->read.tid          = bswap_32(event->read.tid);
 435         event->read.value        = bswap_64(event->read.value);
 436         event->read.time_enabled = bswap_64(event->read.time_enabled);
 437         event->read.time_running = bswap_64(event->read.time_running);
 438         event->read.id           = bswap_64(event->read.id);
 439
 440         if (sample_id_all)
 441                 swap_sample_id_all(event, &event->read + 1);
 442 }
 443
 444 static void perf_event__throttle_swap(union perf_event *event,
 445                                       bool sample_id_all)
 446 {
 447         event->throttle.time      = bswap_64(event->throttle.time);
 448         event->throttle.id        = bswap_64(event->throttle.id);
 449         event->throttle.stream_id = bswap_64(event->throttle.stream_id);
 450
 451         if (sample_id_all)
 452                 swap_sample_id_all(event, &event->throttle + 1);
 453 }
 454
 455 static u8 revbyte(u8 b)
 456 {
 457         int rev = (b >> 4) | ((b & 0xf) << 4);
 458         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 459         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 460         return (u8) rev;
 461 }
 462
 463 /*
 464  * XXX this is hack in attempt to carry flags bitfield
 465  * throught endian village. ABI says:
 466  *
 467  * Bit-fields are allocated from right to left (least to most significant)
 468  * on little-endian implementations and from left to right (most to least
 469  * significant) on big-endian implementations.
 470  *
 471  * The above seems to be byte specific, so we need to reverse each
 472  * byte of the bitfield. 'Internet' also says this might be implementation
 473  * specific and we probably need proper fix and carry perf_event_attr
 474  * bitfield flags in separate data file FEAT_ section. Thought this seems
 475  * to work for now.
 476  */
 477 static void swap_bitfield(u8 *p, unsigned len)
 478 {
 479         unsigned i;
 480
 481         for (i = 0; i < len; i++) {
 482                 *p = revbyte(*p);
 483                 p++;
 484         }
 485 }
 486
 487 /* exported for swapping attributes in file header */
 488 void perf_event__attr_swap(struct perf_event_attr *attr)
 489 {
 490         attr->type              = bswap_32(attr->type);
 491         attr->size              = bswap_32(attr->size);
 492         attr->config            = bswap_64(attr->config);
 493         attr->sample_period     = bswap_64(attr->sample_period);
 494         attr->sample_type       = bswap_64(attr->sample_type);
 495         attr->read_format       = bswap_64(attr->read_format);
 496         attr->wakeup_events     = bswap_32(attr->wakeup_events);
 497         attr->bp_type           = bswap_32(attr->bp_type);
 498         attr->bp_addr           = bswap_64(attr->bp_addr);
 499         attr->bp_len            = bswap_64(attr->bp_len);
 500         attr->branch_sample_type = bswap_64(attr->branch_sample_type);
 501         attr->sample_regs_user   = bswap_64(attr->sample_regs_user);
 502         attr->sample_stack_user  = bswap_32(attr->sample_stack_user);
 503
 504         swap_bitfield((u8 *) (&attr->read_format + 1), sizeof(u64));
 505 }
 506
 507 static void perf_event__hdr_attr_swap(union perf_event *event,
 508                                       bool sample_id_all __maybe_unused)
 509 {
 510         size_t size;
 511
 512         perf_event__attr_swap(&event->attr.attr);
 513
 514         size = event->header.size;
 515         size -= (void *)&event->attr.id - (void *)event;
 516         mem_bswap_64(event->attr.id, size);
 517 }
 518
 519 static void perf_event__event_type_swap(union perf_event *event,
 520                                         bool sample_id_all __maybe_unused)
 521 {
 522         event->event_type.event_type.event_id =
 523                 bswap_64(event->event_type.event_type.event_id);
 524 }
 525
 526 static void perf_event__tracing_data_swap(union perf_event *event,
 527                                           bool sample_id_all __maybe_unused)
 528 {
 529         event->tracing_data.size = bswap_32(event->tracing_data.size);
 530 }
 531
 532 static void perf_event__auxtrace_info_swap(union perf_event *event,
 533                                            bool sample_id_all __maybe_unused)
 534 {
 535         size_t size;
 536
 537         event->auxtrace_info.type = bswap_32(event->auxtrace_info.type);
 538
 539         size = event->header.size;
 540         size -= (void *)&event->auxtrace_info.priv - (void *)event;
 541         mem_bswap_64(event->auxtrace_info.priv, size);
 542 }
 543
 544 static void perf_event__auxtrace_swap(union perf_event *event,
 545                                       bool sample_id_all __maybe_unused)
 546 {
 547         event->auxtrace.size      = bswap_64(event->auxtrace.size);
 548         event->auxtrace.offset    = bswap_64(event->auxtrace.offset);
 549         event->auxtrace.reference = bswap_64(event->auxtrace.reference);
 550         event->auxtrace.idx       = bswap_32(event->auxtrace.idx);
 551         event->auxtrace.tid       = bswap_32(event->auxtrace.tid);
 552         event->auxtrace.cpu       = bswap_32(event->auxtrace.cpu);
 553 }
 554
 555 static void perf_event__auxtrace_error_swap(union perf_event *event,
 556                                             bool sample_id_all __maybe_unused)
 557 {
 558         event->auxtrace_error.type = bswap_32(event->auxtrace_error.type);
 559         event->auxtrace_error.code = bswap_32(event->auxtrace_error.code);
 560         event->auxtrace_error.cpu  = bswap_32(event->auxtrace_error.cpu);
 561         event->auxtrace_error.pid  = bswap_32(event->auxtrace_error.pid);
 562         event->auxtrace_error.tid  = bswap_32(event->auxtrace_error.tid);
 563         event->auxtrace_error.ip   = bswap_64(event->auxtrace_error.ip);
 564 }
 565
 566 typedef void (*perf_event__swap_op)(union perf_event *event,
 567                                     bool sample_id_all);
 568
 569 static perf_event__swap_op perf_event__swap_ops[] = {
 570         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 571         [PERF_RECORD_MMAP2]               = perf_event__mmap2_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_THROTTLE]            = perf_event__throttle_swap,
 578         [PERF_RECORD_UNTHROTTLE]          = perf_event__throttle_swap,
 579         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 580         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 581         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 582         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 583         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 584         [PERF_RECORD_ID_INDEX]            = perf_event__all64_swap,
 585         [PERF_RECORD_AUXTRACE_INFO]       = perf_event__auxtrace_info_swap,
 586         [PERF_RECORD_AUXTRACE]            = perf_event__auxtrace_swap,
 587         [PERF_RECORD_AUXTRACE_ERROR]      = perf_event__auxtrace_error_swap,
 588         [PERF_RECORD_HEADER_MAX]          = NULL,
 589 };
 590
 591 /*
 592  * When perf record finishes a pass on every buffers, it records this pseudo
 593  * event.
 594  * We record the max timestamp t found in the pass n.
 595  * Assuming these timestamps are monotonic across cpus, we know that if
 596  * a buffer still has events with timestamps below t, they will be all
 597  * available and then read in the pass n + 1.
 598  * Hence when we start to read the pass n + 2, we can safely flush every
 599  * events with timestamps below t.
 600  *
 601  *    ============ PASS n =================
 602  *       CPU 0         |   CPU 1
 603  *                     |
 604  *    cnt1 timestamps  |   cnt2 timestamps
 605  *          1          |         2
 606  *          2          |         3
 607  *          -          |         4  <--- max recorded
 608  *
 609  *    ============ PASS n + 1 ==============
 610  *       CPU 0         |   CPU 1
 611  *                     |
 612  *    cnt1 timestamps  |   cnt2 timestamps
 613  *          3          |         5
 614  *          4          |         6
 615  *          5          |         7 <---- max recorded
 616  *
 617  *      Flush every events below timestamp 4
 618  *
 619  *    ============ PASS n + 2 ==============
 620  *       CPU 0         |   CPU 1
 621  *                     |
 622  *    cnt1 timestamps  |   cnt2 timestamps
 623  *          6          |         8
 624  *          7          |         9
 625  *          -          |         10
 626  *
 627  *      Flush every events below timestamp 7
 628  *      etc...
 629  */
 630 static int process_finished_round(struct perf_tool *tool __maybe_unused,
 631                                   union perf_event *event __maybe_unused,
 632                                   struct ordered_events *oe)
 633 {
 634         return ordered_events__flush(oe, OE_FLUSH__ROUND);
 635 }
 636
 637 int perf_session__queue_event(struct perf_session *s, union perf_event *event,
 638                               struct perf_sample *sample, u64 file_offset)
 639 {
 640         return ordered_events__queue(&s->ordered_events, event, sample, file_offset);
 641 }
 642
 643 static void callchain__lbr_callstack_printf(struct perf_sample *sample)
 644 {
 645         struct ip_callchain *callchain = sample->callchain;
 646         struct branch_stack *lbr_stack = sample->branch_stack;
 647         u64 kernel_callchain_nr = callchain->nr;
 648         unsigned int i;
 649
 650         for (i = 0; i < kernel_callchain_nr; i++) {
 651                 if (callchain->ips[i] == PERF_CONTEXT_USER)
 652                         break;
 653         }
 654
 655         if ((i != kernel_callchain_nr) && lbr_stack->nr) {
 656                 u64 total_nr;
 657                 /*
 658                  * LBR callstack can only get user call chain,
 659                  * i is kernel call chain number,
 660                  * 1 is PERF_CONTEXT_USER.
 661                  *
 662                  * The user call chain is stored in LBR registers.
 663                  * LBR are pair registers. The caller is stored
 664                  * in "from" register, while the callee is stored
 665                  * in "to" register.
 666                  * For example, there is a call stack
 667                  * "A"->"B"->"C"->"D".
 668                  * The LBR registers will recorde like
 669                  * "C"->"D", "B"->"C", "A"->"B".
 670                  * So only the first "to" register and all "from"
 671                  * registers are needed to construct the whole stack.
 672                  */
 673                 total_nr = i + 1 + lbr_stack->nr + 1;
 674                 kernel_callchain_nr = i + 1;
 675
 676                 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr);
 677
 678                 for (i = 0; i < kernel_callchain_nr; i++)
 679                         printf("..... %2d: %016" PRIx64 "\n",
 680                                i, callchain->ips[i]);
 681
 682                 printf("..... %2d: %016" PRIx64 "\n",
 683                        (int)(kernel_callchain_nr), lbr_stack->entries[0].to);
 684                 for (i = 0; i < lbr_stack->nr; i++)
 685                         printf("..... %2d: %016" PRIx64 "\n",
 686                                (int)(i + kernel_callchain_nr + 1), lbr_stack->entries[i].from);
 687         }
 688 }
 689
 690 static void callchain__printf(struct perf_evsel *evsel,
 691                               struct perf_sample *sample)
 692 {
 693         unsigned int i;
 694         struct ip_callchain *callchain = sample->callchain;
 695
 696         if (has_branch_callstack(evsel))
 697                 callchain__lbr_callstack_printf(sample);
 698
 699         printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr);
 700
 701         for (i = 0; i < callchain->nr; i++)
 702                 printf("..... %2d: %016" PRIx64 "\n",
 703                        i, callchain->ips[i]);
 704 }
 705
 706 static void branch_stack__printf(struct perf_sample *sample)
 707 {
 708         uint64_t i;
 709
 710         printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr);
 711
 712         for (i = 0; i < sample->branch_stack->nr; i++)
 713                 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 "\n",
 714                         i, sample->branch_stack->entries[i].from,
 715                         sample->branch_stack->entries[i].to);
 716 }
 717
 718 static void regs_dump__printf(u64 mask, u64 *regs)
 719 {
 720         unsigned rid, i = 0;
 721
 722         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
 723                 u64 val = regs[i++];
 724
 725                 printf(".... %-5s 0x%" PRIx64 "\n",
 726                        perf_reg_name(rid), val);
 727         }
 728 }
 729
 730 static const char *regs_abi[] = {
 731         [PERF_SAMPLE_REGS_ABI_NONE] = "none",
 732         [PERF_SAMPLE_REGS_ABI_32] = "32-bit",
 733         [PERF_SAMPLE_REGS_ABI_64] = "64-bit",
 734 };
 735
 736 static inline const char *regs_dump_abi(struct regs_dump *d)
 737 {
 738         if (d->abi > PERF_SAMPLE_REGS_ABI_64)
 739                 return "unknown";
 740
 741         return regs_abi[d->abi];
 742 }
 743
 744 static void regs__printf(const char *type, struct regs_dump *regs)
 745 {
 746         u64 mask = regs->mask;
 747
 748         printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n",
 749                type,
 750                mask,
 751                regs_dump_abi(regs));
 752
 753         regs_dump__printf(mask, regs->regs);
 754 }
 755
 756 static void regs_user__printf(struct perf_sample *sample)
 757 {
 758         struct regs_dump *user_regs = &sample->user_regs;
 759
 760         if (user_regs->regs)
 761                 regs__printf("user", user_regs);
 762 }
 763
 764 static void regs_intr__printf(struct perf_sample *sample)
 765 {
 766         struct regs_dump *intr_regs = &sample->intr_regs;
 767
 768         if (intr_regs->regs)
 769                 regs__printf("intr", intr_regs);
 770 }
 771
 772 static void stack_user__printf(struct stack_dump *dump)
 773 {
 774         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
 775                dump->size, dump->offset);
 776 }
 777
 778 static void perf_evlist__print_tstamp(struct perf_evlist *evlist,
 779                                        union perf_event *event,
 780                                        struct perf_sample *sample)
 781 {
 782         u64 sample_type = __perf_evlist__combined_sample_type(evlist);
 783
 784         if (event->header.type != PERF_RECORD_SAMPLE &&
 785             !perf_evlist__sample_id_all(evlist)) {
 786                 fputs("-1 -1 ", stdout);
 787                 return;
 788         }
 789
 790         if ((sample_type & PERF_SAMPLE_CPU))
 791                 printf("%u ", sample->cpu);
 792
 793         if (sample_type & PERF_SAMPLE_TIME)
 794                 printf("%" PRIu64 " ", sample->time);
 795 }
 796
 797 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
 798 {
 799         printf("... sample_read:\n");
 800
 801         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
 802                 printf("...... time enabled %016" PRIx64 "\n",
 803                        sample->read.time_enabled);
 804
 805         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
 806                 printf("...... time running %016" PRIx64 "\n",
 807                        sample->read.time_running);
 808
 809         if (read_format & PERF_FORMAT_GROUP) {
 810                 u64 i;
 811
 812                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
 813
 814                 for (i = 0; i < sample->read.group.nr; i++) {
 815                         struct sample_read_value *value;
 816
 817                         value = &sample->read.group.values[i];
 818                         printf("..... id %016" PRIx64
 819                                ", value %016" PRIx64 "\n",
 820                                value->id, value->value);
 821                 }
 822         } else
 823                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
 824                         sample->read.one.id, sample->read.one.value);
 825 }
 826
 827 static void dump_event(struct perf_evlist *evlist, union perf_event *event,
 828                        u64 file_offset, struct perf_sample *sample)
 829 {
 830         if (!dump_trace)
 831                 return;
 832
 833         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
 834                file_offset, event->header.size, event->header.type);
 835
 836         trace_event(event);
 837
 838         if (sample)
 839                 perf_evlist__print_tstamp(evlist, event, sample);
 840
 841         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
 842                event->header.size, perf_event__name(event->header.type));
 843 }
 844
 845 static void dump_sample(struct perf_evsel *evsel, union perf_event *event,
 846                         struct perf_sample *sample)
 847 {
 848         u64 sample_type;
 849
 850         if (!dump_trace)
 851                 return;
 852
 853         printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
 854                event->header.misc, sample->pid, sample->tid, sample->ip,
 855                sample->period, sample->addr);
 856
 857         sample_type = evsel->attr.sample_type;
 858
 859         if (sample_type & PERF_SAMPLE_CALLCHAIN)
 860                 callchain__printf(evsel, sample);
 861
 862         if ((sample_type & PERF_SAMPLE_BRANCH_STACK) && !has_branch_callstack(evsel))
 863                 branch_stack__printf(sample);
 864
 865         if (sample_type & PERF_SAMPLE_REGS_USER)
 866                 regs_user__printf(sample);
 867
 868         if (sample_type & PERF_SAMPLE_REGS_INTR)
 869                 regs_intr__printf(sample);
 870
 871         if (sample_type & PERF_SAMPLE_STACK_USER)
 872                 stack_user__printf(&sample->user_stack);
 873
 874         if (sample_type & PERF_SAMPLE_WEIGHT)
 875                 printf("... weight: %" PRIu64 "\n", sample->weight);
 876
 877         if (sample_type & PERF_SAMPLE_DATA_SRC)
 878                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
 879
 880         if (sample_type & PERF_SAMPLE_TRANSACTION)
 881                 printf("... transaction: %" PRIx64 "\n", sample->transaction);
 882
 883         if (sample_type & PERF_SAMPLE_READ)
 884                 sample_read__printf(sample, evsel->attr.read_format);
 885 }
 886
 887 static struct machine *machines__find_for_cpumode(struct machines *machines,
 888                                                union perf_event *event,
 889                                                struct perf_sample *sample)
 890 {
 891         const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
 892         struct machine *machine;
 893
 894         if (perf_guest &&
 895             ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
 896              (cpumode == PERF_RECORD_MISC_GUEST_USER))) {
 897                 u32 pid;
 898
 899                 if (event->header.type == PERF_RECORD_MMAP
 900                     || event->header.type == PERF_RECORD_MMAP2)
 901                         pid = event->mmap.pid;
 902                 else
 903                         pid = sample->pid;
 904
 905                 machine = machines__find(machines, pid);
 906                 if (!machine)
 907                         machine = machines__find(machines, DEFAULT_GUEST_KERNEL_ID);
 908                 return machine;
 909         }
 910
 911         return &machines->host;
 912 }
 913
 914 static int deliver_sample_value(struct perf_evlist *evlist,
 915                                 struct perf_tool *tool,
 916                                 union perf_event *event,
 917                                 struct perf_sample *sample,
 918                                 struct sample_read_value *v,
 919                                 struct machine *machine)
 920 {
 921         struct perf_sample_id *sid = perf_evlist__id2sid(evlist, v->id);
 922
 923         if (sid) {
 924                 sample->id     = v->id;
 925                 sample->period = v->value - sid->period;
 926                 sid->period    = v->value;
 927         }
 928
 929         if (!sid || sid->evsel == NULL) {
 930                 ++evlist->stats.nr_unknown_id;
 931                 return 0;
 932         }
 933
 934         return tool->sample(tool, event, sample, sid->evsel, machine);
 935 }
 936
 937 static int deliver_sample_group(struct perf_evlist *evlist,
 938                                 struct perf_tool *tool,
 939                                 union  perf_event *event,
 940                                 struct perf_sample *sample,
 941                                 struct machine *machine)
 942 {
 943         int ret = -EINVAL;
 944         u64 i;
 945
 946         for (i = 0; i < sample->read.group.nr; i++) {
 947                 ret = deliver_sample_value(evlist, tool, event, sample,
 948                                            &sample->read.group.values[i],
 949                                            machine);
 950                 if (ret)
 951                         break;
 952         }
 953
 954         return ret;
 955 }
 956
 957 static int
 958  perf_evlist__deliver_sample(struct perf_evlist *evlist,
 959                              struct perf_tool *tool,
 960                              union  perf_event *event,
 961                              struct perf_sample *sample,
 962                              struct perf_evsel *evsel,
 963                              struct machine *machine)
 964 {
 965         /* We know evsel != NULL. */
 966         u64 sample_type = evsel->attr.sample_type;
 967         u64 read_format = evsel->attr.read_format;
 968
 969         /* Standard sample delievery. */
 970         if (!(sample_type & PERF_SAMPLE_READ))
 971                 return tool->sample(tool, event, sample, evsel, machine);
 972
 973         /* For PERF_SAMPLE_READ we have either single or group mode. */
 974         if (read_format & PERF_FORMAT_GROUP)
 975                 return deliver_sample_group(evlist, tool, event, sample,
 976                                             machine);
 977         else
 978                 return deliver_sample_value(evlist, tool, event, sample,
 979                                             &sample->read.one, machine);
 980 }
 981
 982 static int machines__deliver_event(struct machines *machines,
 983                                    struct perf_evlist *evlist,
 984                                    union perf_event *event,
 985                                    struct perf_sample *sample,
 986                                    struct perf_tool *tool, u64 file_offset)
 987 {
 988         struct perf_evsel *evsel;
 989         struct machine *machine;
 990
 991         dump_event(evlist, event, file_offset, sample);
 992
 993         evsel = perf_evlist__id2evsel(evlist, sample->id);
 994
 995         machine = machines__find_for_cpumode(machines, event, sample);
 996
 997         switch (event->header.type) {
 998         case PERF_RECORD_SAMPLE:
 999                 dump_sample(evsel, event, sample);
1000                 if (evsel == NULL) {
1001                         ++evlist->stats.nr_unknown_id;
1002                         return 0;
1003                 }
1004                 if (machine == NULL) {
1005                         ++evlist->stats.nr_unprocessable_samples;
1006                         return 0;
1007                 }
1008                 return perf_evlist__deliver_sample(evlist, tool, event, sample, evsel, machine);
1009         case PERF_RECORD_MMAP:
1010                 return tool->mmap(tool, event, sample, machine);
1011         case PERF_RECORD_MMAP2:
1012                 return tool->mmap2(tool, event, sample, machine);
1013         case PERF_RECORD_COMM:
1014                 return tool->comm(tool, event, sample, machine);
1015         case PERF_RECORD_FORK:
1016                 return tool->fork(tool, event, sample, machine);
1017         case PERF_RECORD_EXIT:
1018                 return tool->exit(tool, event, sample, machine);
1019         case PERF_RECORD_LOST:
1020                 if (tool->lost == perf_event__process_lost)
1021                         evlist->stats.total_lost += event->lost.lost;
1022                 return tool->lost(tool, event, sample, machine);
1023         case PERF_RECORD_READ:
1024                 return tool->read(tool, event, sample, evsel, machine);
1025         case PERF_RECORD_THROTTLE:
1026                 return tool->throttle(tool, event, sample, machine);
1027         case PERF_RECORD_UNTHROTTLE:
1028                 return tool->unthrottle(tool, event, sample, machine);
1029         default:
1030                 ++evlist->stats.nr_unknown_events;
1031                 return -1;
1032         }
1033 }
1034
1035 static int perf_session__deliver_event(struct perf_session *session,
1036                                        union perf_event *event,
1037                                        struct perf_sample *sample,
1038                                        struct perf_tool *tool,
1039                                        u64 file_offset)
1040 {
1041         int ret;
1042
1043         ret = auxtrace__process_event(session, event, sample, tool);
1044         if (ret < 0)
1045                 return ret;
1046         if (ret > 0)
1047                 return 0;
1048
1049         return machines__deliver_event(&session->machines, session->evlist,
1050                                        event, sample, tool, file_offset);
1051 }
1052
1053 static s64 perf_session__process_user_event(struct perf_session *session,
1054                                             union perf_event *event,
1055                                             u64 file_offset)
1056 {
1057         struct ordered_events *oe = &session->ordered_events;
1058         struct perf_tool *tool = session->tool;
1059         int fd = perf_data_file__fd(session->file);
1060         int err;
1061
1062         dump_event(session->evlist, event, file_offset, NULL);
1063
1064         /* These events are processed right away */
1065         switch (event->header.type) {
1066         case PERF_RECORD_HEADER_ATTR:
1067                 err = tool->attr(tool, event, &session->evlist);
1068                 if (err == 0) {
1069                         perf_session__set_id_hdr_size(session);
1070                         perf_session__set_comm_exec(session);
1071                 }
1072                 return err;
1073         case PERF_RECORD_HEADER_EVENT_TYPE:
1074                 /*
1075                  * Depreceated, but we need to handle it for sake
1076                  * of old data files create in pipe mode.
1077                  */
1078                 return 0;
1079         case PERF_RECORD_HEADER_TRACING_DATA:
1080                 /* setup for reading amidst mmap */
1081                 lseek(fd, file_offset, SEEK_SET);
1082                 return tool->tracing_data(tool, event, session);
1083         case PERF_RECORD_HEADER_BUILD_ID:
1084                 return tool->build_id(tool, event, session);
1085         case PERF_RECORD_FINISHED_ROUND:
1086                 return tool->finished_round(tool, event, oe);
1087         case PERF_RECORD_ID_INDEX:
1088                 return tool->id_index(tool, event, session);
1089         case PERF_RECORD_AUXTRACE_INFO:
1090                 return tool->auxtrace_info(tool, event, session);
1091         case PERF_RECORD_AUXTRACE:
1092                 /* setup for reading amidst mmap */
1093                 lseek(fd, file_offset + event->header.size, SEEK_SET);
1094                 return tool->auxtrace(tool, event, session);
1095         case PERF_RECORD_AUXTRACE_ERROR:
1096                 perf_session__auxtrace_error_inc(session, event);
1097                 return tool->auxtrace_error(tool, event, session);
1098         default:
1099                 return -EINVAL;
1100         }
1101 }
1102
1103 int perf_session__deliver_synth_event(struct perf_session *session,
1104                                       union perf_event *event,
1105                                       struct perf_sample *sample)
1106 {
1107         struct perf_evlist *evlist = session->evlist;
1108         struct perf_tool *tool = session->tool;
1109
1110         events_stats__inc(&evlist->stats, event->header.type);
1111
1112         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1113                 return perf_session__process_user_event(session, event, 0);
1114
1115         return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0);
1116 }
1117
1118 static void event_swap(union perf_event *event, bool sample_id_all)
1119 {
1120         perf_event__swap_op swap;
1121
1122         swap = perf_event__swap_ops[event->header.type];
1123         if (swap)
1124                 swap(event, sample_id_all);
1125 }
1126
1127 int perf_session__peek_event(struct perf_session *session, off_t file_offset,
1128                              void *buf, size_t buf_sz,
1129                              union perf_event **event_ptr,
1130                              struct perf_sample *sample)
1131 {
1132         union perf_event *event;
1133         size_t hdr_sz, rest;
1134         int fd;
1135
1136         if (session->one_mmap && !session->header.needs_swap) {
1137                 event = file_offset - session->one_mmap_offset +
1138                         session->one_mmap_addr;
1139                 goto out_parse_sample;
1140         }
1141
1142         if (perf_data_file__is_pipe(session->file))
1143                 return -1;
1144
1145         fd = perf_data_file__fd(session->file);
1146         hdr_sz = sizeof(struct perf_event_header);
1147
1148         if (buf_sz < hdr_sz)
1149                 return -1;
1150
1151         if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 ||
1152             readn(fd, &buf, hdr_sz) != (ssize_t)hdr_sz)
1153                 return -1;
1154
1155         event = (union perf_event *)buf;
1156
1157         if (session->header.needs_swap)
1158                 perf_event_header__bswap(&event->header);
1159
1160         if (event->header.size < hdr_sz)
1161                 return -1;
1162
1163         rest = event->header.size - hdr_sz;
1164
1165         if (readn(fd, &buf, rest) != (ssize_t)rest)
1166                 return -1;
1167
1168         if (session->header.needs_swap)
1169                 event_swap(event, perf_evlist__sample_id_all(session->evlist));
1170
1171 out_parse_sample:
1172
1173         if (sample && event->header.type < PERF_RECORD_USER_TYPE_START &&
1174             perf_evlist__parse_sample(session->evlist, event, sample))
1175                 return -1;
1176
1177         *event_ptr = event;
1178
1179         return 0;
1180 }
1181
1182 static s64 perf_session__process_event(struct perf_session *session,
1183                                        union perf_event *event, u64 file_offset)
1184 {
1185         struct perf_evlist *evlist = session->evlist;
1186         struct perf_tool *tool = session->tool;
1187         struct perf_sample sample;
1188         int ret;
1189
1190         if (session->header.needs_swap)
1191                 event_swap(event, perf_evlist__sample_id_all(evlist));
1192
1193         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1194                 return -EINVAL;
1195
1196         events_stats__inc(&evlist->stats, event->header.type);
1197
1198         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1199                 return perf_session__process_user_event(session, event, file_offset);
1200
1201         /*
1202          * For all kernel events we get the sample data
1203          */
1204         ret = perf_evlist__parse_sample(evlist, event, &sample);
1205         if (ret)
1206                 return ret;
1207
1208         if (tool->ordered_events) {
1209                 ret = perf_session__queue_event(session, event, &sample, file_offset);
1210                 if (ret != -ETIME)
1211                         return ret;
1212         }
1213
1214         return perf_session__deliver_event(session, event, &sample, tool,
1215                                            file_offset);
1216 }
1217
1218 void perf_event_header__bswap(struct perf_event_header *hdr)
1219 {
1220         hdr->type = bswap_32(hdr->type);
1221         hdr->misc = bswap_16(hdr->misc);
1222         hdr->size = bswap_16(hdr->size);
1223 }
1224
1225 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1226 {
1227         return machine__findnew_thread(&session->machines.host, -1, pid);
1228 }
1229
1230 static struct thread *perf_session__register_idle_thread(struct perf_session *session)
1231 {
1232         struct thread *thread;
1233
1234         thread = machine__findnew_thread(&session->machines.host, 0, 0);
1235         if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1236                 pr_err("problem inserting idle task.\n");
1237                 thread = NULL;
1238         }
1239
1240         return thread;
1241 }
1242
1243 static void perf_session__warn_about_errors(const struct perf_session *session)
1244 {
1245         const struct events_stats *stats = &session->evlist->stats;
1246         const struct ordered_events *oe = &session->ordered_events;
1247
1248         if (session->tool->lost == perf_event__process_lost &&
1249             stats->nr_events[PERF_RECORD_LOST] != 0) {
1250                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1251                             "Check IO/CPU overload!\n\n",
1252                             stats->nr_events[0],
1253                             stats->nr_events[PERF_RECORD_LOST]);
1254         }
1255
1256         if (stats->nr_unknown_events != 0) {
1257                 ui__warning("Found %u unknown events!\n\n"
1258                             "Is this an older tool processing a perf.data "
1259                             "file generated by a more recent tool?\n\n"
1260                             "If that is not the case, consider "
1261                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1262                             stats->nr_unknown_events);
1263         }
1264
1265         if (stats->nr_unknown_id != 0) {
1266                 ui__warning("%u samples with id not present in the header\n",
1267                             stats->nr_unknown_id);
1268         }
1269
1270         if (stats->nr_invalid_chains != 0) {
1271                 ui__warning("Found invalid callchains!\n\n"
1272                             "%u out of %u events were discarded for this reason.\n\n"
1273                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1274                             stats->nr_invalid_chains,
1275                             stats->nr_events[PERF_RECORD_SAMPLE]);
1276         }
1277
1278         if (stats->nr_unprocessable_samples != 0) {
1279                 ui__warning("%u unprocessable samples recorded.\n"
1280                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1281                             stats->nr_unprocessable_samples);
1282         }
1283
1284         if (oe->nr_unordered_events != 0)
1285                 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events);
1286
1287         events_stats__auxtrace_error_warn(stats);
1288 }
1289
1290 volatile int session_done;
1291
1292 static int __perf_session__process_pipe_events(struct perf_session *session)
1293 {
1294         struct ordered_events *oe = &session->ordered_events;
1295         struct perf_tool *tool = session->tool;
1296         int fd = perf_data_file__fd(session->file);
1297         union perf_event *event;
1298         uint32_t size, cur_size = 0;
1299         void *buf = NULL;
1300         s64 skip = 0;
1301         u64 head;
1302         ssize_t err;
1303         void *p;
1304
1305         perf_tool__fill_defaults(tool);
1306
1307         head = 0;
1308         cur_size = sizeof(union perf_event);
1309
1310         buf = malloc(cur_size);
1311         if (!buf)
1312                 return -errno;
1313 more:
1314         event = buf;
1315         err = readn(fd, event, sizeof(struct perf_event_header));
1316         if (err <= 0) {
1317                 if (err == 0)
1318                         goto done;
1319
1320                 pr_err("failed to read event header\n");
1321                 goto out_err;
1322         }
1323
1324         if (session->header.needs_swap)
1325                 perf_event_header__bswap(&event->header);
1326
1327         size = event->header.size;
1328         if (size < sizeof(struct perf_event_header)) {
1329                 pr_err("bad event header size\n");
1330                 goto out_err;
1331         }
1332
1333         if (size > cur_size) {
1334                 void *new = realloc(buf, size);
1335                 if (!new) {
1336                         pr_err("failed to allocate memory to read event\n");
1337                         goto out_err;
1338                 }
1339                 buf = new;
1340                 cur_size = size;
1341                 event = buf;
1342         }
1343         p = event;
1344         p += sizeof(struct perf_event_header);
1345
1346         if (size - sizeof(struct perf_event_header)) {
1347                 err = readn(fd, p, size - sizeof(struct perf_event_header));
1348                 if (err <= 0) {
1349                         if (err == 0) {
1350                                 pr_err("unexpected end of event stream\n");
1351                                 goto done;
1352                         }
1353
1354                         pr_err("failed to read event data\n");
1355                         goto out_err;
1356                 }
1357         }
1358
1359         if ((skip = perf_session__process_event(session, event, head)) < 0) {
1360                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1361                        head, event->header.size, event->header.type);
1362                 err = -EINVAL;
1363                 goto out_err;
1364         }
1365
1366         head += size;
1367
1368         if (skip > 0)
1369                 head += skip;
1370
1371         if (!session_done())
1372                 goto more;
1373 done:
1374         /* do the final flush for ordered samples */
1375         err = ordered_events__flush(oe, OE_FLUSH__FINAL);
1376         if (err)
1377                 goto out_err;
1378         err = auxtrace__flush_events(session, tool);
1379 out_err:
1380         free(buf);
1381         perf_session__warn_about_errors(session);
1382         ordered_events__free(&session->ordered_events);
1383         auxtrace__free_events(session);
1384         return err;
1385 }
1386
1387 static union perf_event *
1388 fetch_mmaped_event(struct perf_session *session,
1389                    u64 head, size_t mmap_size, char *buf)
1390 {
1391         union perf_event *event;
1392
1393         /*
1394          * Ensure we have enough space remaining to read
1395          * the size of the event in the headers.
1396          */
1397         if (head + sizeof(event->header) > mmap_size)
1398                 return NULL;
1399
1400         event = (union perf_event *)(buf + head);
1401
1402         if (session->header.needs_swap)
1403                 perf_event_header__bswap(&event->header);
1404
1405         if (head + event->header.size > mmap_size) {
1406                 /* We're not fetching the event so swap back again */
1407                 if (session->header.needs_swap)
1408                         perf_event_header__bswap(&event->header);
1409                 return NULL;
1410         }
1411
1412         return event;
1413 }
1414
1415 /*
1416  * On 64bit we can mmap the data file in one go. No need for tiny mmap
1417  * slices. On 32bit we use 32MB.
1418  */
1419 #if BITS_PER_LONG == 64
1420 #define MMAP_SIZE ULLONG_MAX
1421 #define NUM_MMAPS 1
1422 #else
1423 #define MMAP_SIZE (32 * 1024 * 1024ULL)
1424 #define NUM_MMAPS 128
1425 #endif
1426
1427 static int __perf_session__process_events(struct perf_session *session,
1428                                           u64 data_offset, u64 data_size,
1429                                           u64 file_size)
1430 {
1431         struct ordered_events *oe = &session->ordered_events;
1432         struct perf_tool *tool = session->tool;
1433         int fd = perf_data_file__fd(session->file);
1434         u64 head, page_offset, file_offset, file_pos, size;
1435         int err, mmap_prot, mmap_flags, map_idx = 0;
1436         size_t  mmap_size;
1437         char *buf, *mmaps[NUM_MMAPS];
1438         union perf_event *event;
1439         struct ui_progress prog;
1440         s64 skip;
1441
1442         perf_tool__fill_defaults(tool);
1443
1444         page_offset = page_size * (data_offset / page_size);
1445         file_offset = page_offset;
1446         head = data_offset - page_offset;
1447
1448         if (data_size && (data_offset + data_size < file_size))
1449                 file_size = data_offset + data_size;
1450
1451         ui_progress__init(&prog, file_size, "Processing events...");
1452
1453         mmap_size = MMAP_SIZE;
1454         if (mmap_size > file_size) {
1455                 mmap_size = file_size;
1456                 session->one_mmap = true;
1457         }
1458
1459         memset(mmaps, 0, sizeof(mmaps));
1460
1461         mmap_prot  = PROT_READ;
1462         mmap_flags = MAP_SHARED;
1463
1464         if (session->header.needs_swap) {
1465                 mmap_prot  |= PROT_WRITE;
1466                 mmap_flags = MAP_PRIVATE;
1467         }
1468 remap:
1469         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd,
1470                    file_offset);
1471         if (buf == MAP_FAILED) {
1472                 pr_err("failed to mmap file\n");
1473                 err = -errno;
1474                 goto out_err;
1475         }
1476         mmaps[map_idx] = buf;
1477         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
1478         file_pos = file_offset + head;
1479         if (session->one_mmap) {
1480                 session->one_mmap_addr = buf;
1481                 session->one_mmap_offset = file_offset;
1482         }
1483
1484 more:
1485         event = fetch_mmaped_event(session, head, mmap_size, buf);
1486         if (!event) {
1487                 if (mmaps[map_idx]) {
1488                         munmap(mmaps[map_idx], mmap_size);
1489                         mmaps[map_idx] = NULL;
1490                 }
1491
1492                 page_offset = page_size * (head / page_size);
1493                 file_offset += page_offset;
1494                 head -= page_offset;
1495                 goto remap;
1496         }
1497
1498         size = event->header.size;
1499
1500         if (size < sizeof(struct perf_event_header) ||
1501             (skip = perf_session__process_event(session, event, file_pos)) < 0) {
1502                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
1503                        file_offset + head, event->header.size,
1504                        event->header.type);
1505                 err = -EINVAL;
1506                 goto out_err;
1507         }
1508
1509         if (skip)
1510                 size += skip;
1511
1512         head += size;
1513         file_pos += size;
1514
1515         ui_progress__update(&prog, size);
1516
1517         if (session_done())
1518                 goto out;
1519
1520         if (file_pos < file_size)
1521                 goto more;
1522
1523 out:
1524         /* do the final flush for ordered samples */
1525         err = ordered_events__flush(oe, OE_FLUSH__FINAL);
1526         if (err)
1527                 goto out_err;
1528         err = auxtrace__flush_events(session, tool);
1529 out_err:
1530         ui_progress__finish();
1531         perf_session__warn_about_errors(session);
1532         ordered_events__free(&session->ordered_events);
1533         auxtrace__free_events(session);
1534         session->one_mmap = false;
1535         return err;
1536 }
1537
1538 int perf_session__process_events(struct perf_session *session)
1539 {
1540         u64 size = perf_data_file__size(session->file);
1541         int err;
1542
1543         if (perf_session__register_idle_thread(session) == NULL)
1544                 return -ENOMEM;
1545
1546         if (!perf_data_file__is_pipe(session->file))
1547                 err = __perf_session__process_events(session,
1548                                                      session->header.data_offset,
1549                                                      session->header.data_size, size);
1550         else
1551                 err = __perf_session__process_pipe_events(session);
1552
1553         return err;
1554 }
1555
1556 bool perf_session__has_traces(struct perf_session *session, const char *msg)
1557 {
1558         struct perf_evsel *evsel;
1559
1560         evlist__for_each(session->evlist, evsel) {
1561                 if (evsel->attr.type == PERF_TYPE_TRACEPOINT)
1562                         return true;
1563         }
1564
1565         pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
1566         return false;
1567 }
1568
1569 int maps__set_kallsyms_ref_reloc_sym(struct map **maps,
1570                                      const char *symbol_name, u64 addr)
1571 {
1572         char *bracket;
1573         enum map_type i;
1574         struct ref_reloc_sym *ref;
1575
1576         ref = zalloc(sizeof(struct ref_reloc_sym));
1577         if (ref == NULL)
1578                 return -ENOMEM;
1579
1580         ref->name = strdup(symbol_name);
1581         if (ref->name == NULL) {
1582                 free(ref);
1583                 return -ENOMEM;
1584         }
1585
1586         bracket = strchr(ref->name, ']');
1587         if (bracket)
1588                 *bracket = '\0';
1589
1590         ref->addr = addr;
1591
1592         for (i = 0; i < MAP__NR_TYPES; ++i) {
1593                 struct kmap *kmap = map__kmap(maps[i]);
1594
1595                 if (!kmap)
1596                         continue;
1597                 kmap->ref_reloc_sym = ref;
1598         }
1599
1600         return 0;
1601 }
1602
1603 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
1604 {
1605         return machines__fprintf_dsos(&session->machines, fp);
1606 }
1607
1608 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
1609                                           bool (skip)(struct dso *dso, int parm), int parm)
1610 {
1611         return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
1612 }
1613
1614 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
1615 {
1616         size_t ret;
1617         const char *msg = "";
1618
1619         if (perf_header__has_feat(&session->header, HEADER_AUXTRACE))
1620                 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)";
1621
1622         ret = fprintf(fp, "Aggregated stats:%s\n", msg);
1623
1624         ret += events_stats__fprintf(&session->evlist->stats, fp);
1625         return ret;
1626 }
1627
1628 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
1629 {
1630         /*
1631          * FIXME: Here we have to actually print all the machines in this
1632          * session, not just the host...
1633          */
1634         return machine__fprintf(&session->machines.host, fp);
1635 }
1636
1637 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session,
1638                                               unsigned int type)
1639 {
1640         struct perf_evsel *pos;
1641
1642         evlist__for_each(session->evlist, pos) {
1643                 if (pos->attr.type == type)
1644                         return pos;
1645         }
1646         return NULL;
1647 }
1648
1649 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample,
1650                           struct addr_location *al,
1651                           unsigned int print_opts, unsigned int stack_depth)
1652 {
1653         struct callchain_cursor_node *node;
1654         int print_ip = print_opts & PRINT_IP_OPT_IP;
1655         int print_sym = print_opts & PRINT_IP_OPT_SYM;
1656         int print_dso = print_opts & PRINT_IP_OPT_DSO;
1657         int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET;
1658         int print_oneline = print_opts & PRINT_IP_OPT_ONELINE;
1659         int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE;
1660         char s = print_oneline ? ' ' : '\t';
1661
1662         if (symbol_conf.use_callchain && sample->callchain) {
1663                 struct addr_location node_al;
1664
1665                 if (thread__resolve_callchain(al->thread, evsel,
1666                                               sample, NULL, NULL,
1667                                               PERF_MAX_STACK_DEPTH) != 0) {
1668                         if (verbose)
1669                                 error("Failed to resolve callchain. Skipping\n");
1670                         return;
1671                 }
1672                 callchain_cursor_commit(&callchain_cursor);
1673
1674                 if (print_symoffset)
1675                         node_al = *al;
1676
1677                 while (stack_depth) {
1678                         u64 addr = 0;
1679
1680                         node = callchain_cursor_current(&callchain_cursor);
1681                         if (!node)
1682                                 break;
1683
1684                         if (node->sym && node->sym->ignore)
1685                                 goto next;
1686
1687                         if (print_ip)
1688                                 printf("%c%16" PRIx64, s, node->ip);
1689
1690                         if (node->map)
1691                                 addr = node->map->map_ip(node->map, node->ip);
1692
1693                         if (print_sym) {
1694                                 printf(" ");
1695                                 if (print_symoffset) {
1696                                         node_al.addr = addr;
1697                                         node_al.map  = node->map;
1698                                         symbol__fprintf_symname_offs(node->sym, &node_al, stdout);
1699                                 } else
1700                                         symbol__fprintf_symname(node->sym, stdout);
1701                         }
1702
1703                         if (print_dso) {
1704                                 printf(" (");
1705                                 map__fprintf_dsoname(node->map, stdout);
1706                                 printf(")");
1707                         }
1708
1709                         if (print_srcline)
1710                                 map__fprintf_srcline(node->map, addr, "\n  ",
1711                                                      stdout);
1712
1713                         if (!print_oneline)
1714                                 printf("\n");
1715
1716                         stack_depth--;
1717 next:
1718                         callchain_cursor_advance(&callchain_cursor);
1719                 }
1720
1721         } else {
1722                 if (al->sym && al->sym->ignore)
1723                         return;
1724
1725                 if (print_ip)
1726                         printf("%16" PRIx64, sample->ip);
1727
1728                 if (print_sym) {
1729                         printf(" ");
1730                         if (print_symoffset)
1731                                 symbol__fprintf_symname_offs(al->sym, al,
1732                                                              stdout);
1733                         else
1734                                 symbol__fprintf_symname(al->sym, stdout);
1735                 }
1736
1737                 if (print_dso) {
1738                         printf(" (");
1739                         map__fprintf_dsoname(al->map, stdout);
1740                         printf(")");
1741                 }
1742
1743                 if (print_srcline)
1744                         map__fprintf_srcline(al->map, al->addr, "\n  ", stdout);
1745         }
1746 }
1747
1748 int perf_session__cpu_bitmap(struct perf_session *session,
1749                              const char *cpu_list, unsigned long *cpu_bitmap)
1750 {
1751         int i, err = -1;
1752         struct cpu_map *map;
1753
1754         for (i = 0; i < PERF_TYPE_MAX; ++i) {
1755                 struct perf_evsel *evsel;
1756
1757                 evsel = perf_session__find_first_evtype(session, i);
1758                 if (!evsel)
1759                         continue;
1760
1761                 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) {
1762                         pr_err("File does not contain CPU events. "
1763                                "Remove -c option to proceed.\n");
1764                         return -1;
1765                 }
1766         }
1767
1768         map = cpu_map__new(cpu_list);
1769         if (map == NULL) {
1770                 pr_err("Invalid cpu_list\n");
1771                 return -1;
1772         }
1773
1774         for (i = 0; i < map->nr; i++) {
1775                 int cpu = map->map[i];
1776
1777                 if (cpu >= MAX_NR_CPUS) {
1778                         pr_err("Requested CPU %d too large. "
1779                                "Consider raising MAX_NR_CPUS\n", cpu);
1780                         goto out_delete_map;
1781                 }
1782
1783                 set_bit(cpu, cpu_bitmap);
1784         }
1785
1786         err = 0;
1787
1788 out_delete_map:
1789         cpu_map__delete(map);
1790         return err;
1791 }
1792
1793 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
1794                                 bool full)
1795 {
1796         struct stat st;
1797         int fd, ret;
1798
1799         if (session == NULL || fp == NULL)
1800                 return;
1801
1802         fd = perf_data_file__fd(session->file);
1803
1804         ret = fstat(fd, &st);
1805         if (ret == -1)
1806                 return;
1807
1808         fprintf(fp, "# ========\n");
1809         fprintf(fp, "# captured on: %s", ctime(&st.st_ctime));
1810         perf_header__fprintf_info(session, fp, full);
1811         fprintf(fp, "# ========\n#\n");
1812 }
1813
1814
1815 int __perf_session__set_tracepoints_handlers(struct perf_session *session,
1816                                              const struct perf_evsel_str_handler *assocs,
1817                                              size_t nr_assocs)
1818 {
1819         struct perf_evsel *evsel;
1820         size_t i;
1821         int err;
1822
1823         for (i = 0; i < nr_assocs; i++) {
1824                 /*
1825                  * Adding a handler for an event not in the session,
1826                  * just ignore it.
1827                  */
1828                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name);
1829                 if (evsel == NULL)
1830                         continue;
1831
1832                 err = -EEXIST;
1833                 if (evsel->handler != NULL)
1834                         goto out;
1835                 evsel->handler = assocs[i].handler;
1836         }
1837
1838         err = 0;
1839 out:
1840         return err;
1841 }
1842
1843 int perf_event__process_id_index(struct perf_tool *tool __maybe_unused,
1844                                  union perf_event *event,
1845                                  struct perf_session *session)
1846 {
1847         struct perf_evlist *evlist = session->evlist;
1848         struct id_index_event *ie = &event->id_index;
1849         size_t i, nr, max_nr;
1850
1851         max_nr = (ie->header.size - sizeof(struct id_index_event)) /
1852                  sizeof(struct id_index_entry);
1853         nr = ie->nr;
1854         if (nr > max_nr)
1855                 return -EINVAL;
1856
1857         if (dump_trace)
1858                 fprintf(stdout, " nr: %zu\n", nr);
1859
1860         for (i = 0; i < nr; i++) {
1861                 struct id_index_entry *e = &ie->entries[i];
1862                 struct perf_sample_id *sid;
1863
1864                 if (dump_trace) {
1865                         fprintf(stdout, " ... id: %"PRIu64, e->id);
1866                         fprintf(stdout, "  idx: %"PRIu64, e->idx);
1867                         fprintf(stdout, "  cpu: %"PRId64, e->cpu);
1868                         fprintf(stdout, "  tid: %"PRId64"\n", e->tid);
1869                 }
1870
1871                 sid = perf_evlist__id2sid(evlist, e->id);
1872                 if (!sid)
1873                         return -ENOENT;
1874                 sid->idx = e->idx;
1875                 sid->cpu = e->cpu;
1876                 sid->tid = e->tid;
1877         }
1878         return 0;
1879 }
1880
1881 int perf_event__synthesize_id_index(struct perf_tool *tool,
1882                                     perf_event__handler_t process,
1883                                     struct perf_evlist *evlist,
1884                                     struct machine *machine)
1885 {
1886         union perf_event *ev;
1887         struct perf_evsel *evsel;
1888         size_t nr = 0, i = 0, sz, max_nr, n;
1889         int err;
1890
1891         pr_debug2("Synthesizing id index\n");
1892
1893         max_nr = (UINT16_MAX - sizeof(struct id_index_event)) /
1894                  sizeof(struct id_index_entry);
1895
1896         evlist__for_each(evlist, evsel)
1897                 nr += evsel->ids;
1898
1899         n = nr > max_nr ? max_nr : nr;
1900         sz = sizeof(struct id_index_event) + n * sizeof(struct id_index_entry);
1901         ev = zalloc(sz);
1902         if (!ev)
1903                 return -ENOMEM;
1904
1905         ev->id_index.header.type = PERF_RECORD_ID_INDEX;
1906         ev->id_index.header.size = sz;
1907         ev->id_index.nr = n;
1908
1909         evlist__for_each(evlist, evsel) {
1910                 u32 j;
1911
1912                 for (j = 0; j < evsel->ids; j++) {
1913                         struct id_index_entry *e;
1914                         struct perf_sample_id *sid;
1915
1916                         if (i >= n) {
1917                                 err = process(tool, ev, NULL, machine);
1918                                 if (err)
1919                                         goto out_err;
1920                                 nr -= n;
1921                                 i = 0;
1922                         }
1923
1924                         e = &ev->id_index.entries[i++];
1925
1926                         e->id = evsel->id[j];
1927
1928                         sid = perf_evlist__id2sid(evlist, e->id);
1929                         if (!sid) {
1930                                 free(ev);
1931                                 return -ENOENT;
1932                         }
1933
1934                         e->idx = sid->idx;
1935                         e->cpu = sid->cpu;
1936                         e->tid = sid->tid;
1937                 }
1938         }
1939
1940         sz = sizeof(struct id_index_event) + nr * sizeof(struct id_index_entry);
1941         ev->id_index.header.size = sz;
1942         ev->id_index.nr = nr;
1943
1944         err = process(tool, ev, NULL, machine);
1945 out_err:
1946         free(ev);
1947
1948         return err;
1949 }