tools/perf/util/session.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <errno.h>
   3 #include <inttypes.h>
   4 #include <linux/err.h>
   5 #include <linux/kernel.h>
   6 #include <linux/zalloc.h>
   7 #include <api/fs/fs.h>
   8
   9 #include <byteswap.h>
  10 #include <unistd.h>
  11 #include <sys/types.h>
  12 #include <sys/mman.h>
  13 #include <perf/cpumap.h>
  14
  15 #include "map_symbol.h"
  16 #include "branch.h"
  17 #include "debug.h"
  18 #include "evlist.h"
  19 #include "evsel.h"
  20 #include "memswap.h"
  21 #include "map.h"
  22 #include "symbol.h"
  23 #include "session.h"
  24 #include "tool.h"
  25 #include "perf_regs.h"
  26 #include "asm/bug.h"
  27 #include "auxtrace.h"
  28 #include "thread.h"
  29 #include "thread-stack.h"
  30 #include "sample-raw.h"
  31 #include "stat.h"
  32 #include "ui/progress.h"
  33 #include "../perf.h"
  34 #include "arch/common.h"
  35 #include <internal/lib.h>
  36
  37 #ifdef HAVE_ZSTD_SUPPORT
  38 static int perf_session__process_compressed_event(struct perf_session *session,
  39                                                   union perf_event *event, u64 file_offset)
  40 {
  41         void *src;
  42         size_t decomp_size, src_size;
  43         u64 decomp_last_rem = 0;
  44         size_t mmap_len, decomp_len = session->header.env.comp_mmap_len;
  45         struct decomp *decomp, *decomp_last = session->decomp_last;
  46
  47         if (decomp_last) {
  48                 decomp_last_rem = decomp_last->size - decomp_last->head;
  49                 decomp_len += decomp_last_rem;
  50         }
  51
  52         mmap_len = sizeof(struct decomp) + decomp_len;
  53         decomp = mmap(NULL, mmap_len, PROT_READ|PROT_WRITE,
  54                       MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
  55         if (decomp == MAP_FAILED) {
  56                 pr_err("Couldn't allocate memory for decompression\n");
  57                 return -1;
  58         }
  59
  60         decomp->file_pos = file_offset;
  61         decomp->mmap_len = mmap_len;
  62         decomp->head = 0;
  63
  64         if (decomp_last_rem) {
  65                 memcpy(decomp->data, &(decomp_last->data[decomp_last->head]), decomp_last_rem);
  66                 decomp->size = decomp_last_rem;
  67         }
  68
  69         src = (void *)event + sizeof(struct perf_record_compressed);
  70         src_size = event->pack.header.size - sizeof(struct perf_record_compressed);
  71
  72         decomp_size = zstd_decompress_stream(&(session->zstd_data), src, src_size,
  73                                 &(decomp->data[decomp_last_rem]), decomp_len - decomp_last_rem);
  74         if (!decomp_size) {
  75                 munmap(decomp, mmap_len);
  76                 pr_err("Couldn't decompress data\n");
  77                 return -1;
  78         }
  79
  80         decomp->size += decomp_size;
  81
  82         if (session->decomp == NULL) {
  83                 session->decomp = decomp;
  84                 session->decomp_last = decomp;
  85         } else {
  86                 session->decomp_last->next = decomp;
  87                 session->decomp_last = decomp;
  88         }
  89
  90         pr_debug("decomp (B): %zd to %zd\n", src_size, decomp_size);
  91
  92         return 0;
  93 }
  94 #else /* !HAVE_ZSTD_SUPPORT */
  95 #define perf_session__process_compressed_event perf_session__process_compressed_event_stub
  96 #endif
  97
  98 static int perf_session__deliver_event(struct perf_session *session,
  99                                        union perf_event *event,
 100                                        struct perf_tool *tool,
 101                                        u64 file_offset);
 102
 103 static int perf_session__open(struct perf_session *session)
 104 {
 105         struct perf_data *data = session->data;
 106
 107         if (perf_session__read_header(session) < 0) {
 108                 pr_err("incompatible file format (rerun with -v to learn more)\n");
 109                 return -1;
 110         }
 111
 112         if (perf_data__is_pipe(data))
 113                 return 0;
 114
 115         if (perf_header__has_feat(&session->header, HEADER_STAT))
 116                 return 0;
 117
 118         if (!evlist__valid_sample_type(session->evlist)) {
 119                 pr_err("non matching sample_type\n");
 120                 return -1;
 121         }
 122
 123         if (!evlist__valid_sample_id_all(session->evlist)) {
 124                 pr_err("non matching sample_id_all\n");
 125                 return -1;
 126         }
 127
 128         if (!perf_evlist__valid_read_format(session->evlist)) {
 129                 pr_err("non matching read_format\n");
 130                 return -1;
 131         }
 132
 133         return 0;
 134 }
 135
 136 void perf_session__set_id_hdr_size(struct perf_session *session)
 137 {
 138         u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist);
 139
 140         machines__set_id_hdr_size(&session->machines, id_hdr_size);
 141 }
 142
 143 int perf_session__create_kernel_maps(struct perf_session *session)
 144 {
 145         int ret = machine__create_kernel_maps(&session->machines.host);
 146
 147         if (ret >= 0)
 148                 ret = machines__create_guest_kernel_maps(&session->machines);
 149         return ret;
 150 }
 151
 152 static void perf_session__destroy_kernel_maps(struct perf_session *session)
 153 {
 154         machines__destroy_kernel_maps(&session->machines);
 155 }
 156
 157 static bool perf_session__has_comm_exec(struct perf_session *session)
 158 {
 159         struct evsel *evsel;
 160
 161         evlist__for_each_entry(session->evlist, evsel) {
 162                 if (evsel->core.attr.comm_exec)
 163                         return true;
 164         }
 165
 166         return false;
 167 }
 168
 169 static void perf_session__set_comm_exec(struct perf_session *session)
 170 {
 171         bool comm_exec = perf_session__has_comm_exec(session);
 172
 173         machines__set_comm_exec(&session->machines, comm_exec);
 174 }
 175
 176 static int ordered_events__deliver_event(struct ordered_events *oe,
 177                                          struct ordered_event *event)
 178 {
 179         struct perf_session *session = container_of(oe, struct perf_session,
 180                                                     ordered_events);
 181
 182         return perf_session__deliver_event(session, event->event,
 183                                            session->tool, event->file_offset);
 184 }
 185
 186 struct perf_session *perf_session__new(struct perf_data *data,
 187                                        bool repipe, struct perf_tool *tool)
 188 {
 189         int ret = -ENOMEM;
 190         struct perf_session *session = zalloc(sizeof(*session));
 191
 192         if (!session)
 193                 goto out;
 194
 195         session->repipe = repipe;
 196         session->tool   = tool;
 197         INIT_LIST_HEAD(&session->auxtrace_index);
 198         machines__init(&session->machines);
 199         ordered_events__init(&session->ordered_events,
 200                              ordered_events__deliver_event, NULL);
 201
 202         perf_env__init(&session->header.env);
 203         if (data) {
 204                 ret = perf_data__open(data);
 205                 if (ret < 0)
 206                         goto out_delete;
 207
 208                 session->data = data;
 209
 210                 if (perf_data__is_read(data)) {
 211                         ret = perf_session__open(session);
 212                         if (ret < 0)
 213                                 goto out_delete;
 214
 215                         /*
 216                          * set session attributes that are present in perf.data
 217                          * but not in pipe-mode.
 218                          */
 219                         if (!data->is_pipe) {
 220                                 perf_session__set_id_hdr_size(session);
 221                                 perf_session__set_comm_exec(session);
 222                         }
 223
 224                         perf_evlist__init_trace_event_sample_raw(session->evlist);
 225
 226                         /* Open the directory data. */
 227                         if (data->is_dir) {
 228                                 ret = perf_data__open_dir(data);
 229                                 if (ret)
 230                                         goto out_delete;
 231                         }
 232
 233                         if (!symbol_conf.kallsyms_name &&
 234                             !symbol_conf.vmlinux_name)
 235                                 symbol_conf.kallsyms_name = perf_data__kallsyms_name(data);
 236                 }
 237         } else  {
 238                 session->machines.host.env = &perf_env;
 239         }
 240
 241         session->machines.host.single_address_space =
 242                 perf_env__single_address_space(session->machines.host.env);
 243
 244         if (!data || perf_data__is_write(data)) {
 245                 /*
 246                  * In O_RDONLY mode this will be performed when reading the
 247                  * kernel MMAP event, in perf_event__process_mmap().
 248                  */
 249                 if (perf_session__create_kernel_maps(session) < 0)
 250                         pr_warning("Cannot read kernel map\n");
 251         }
 252
 253         /*
 254          * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is
 255          * processed, so evlist__sample_id_all is not meaningful here.
 256          */
 257         if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps &&
 258             tool->ordered_events && !evlist__sample_id_all(session->evlist)) {
 259                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 260                 tool->ordered_events = false;
 261         }
 262
 263         return session;
 264
 265  out_delete:
 266         perf_session__delete(session);
 267  out:
 268         return ERR_PTR(ret);
 269 }
 270
 271 static void perf_session__delete_threads(struct perf_session *session)
 272 {
 273         machine__delete_threads(&session->machines.host);
 274 }
 275
 276 static void perf_session__release_decomp_events(struct perf_session *session)
 277 {
 278         struct decomp *next, *decomp;
 279         size_t mmap_len;
 280         next = session->decomp;
 281         do {
 282                 decomp = next;
 283                 if (decomp == NULL)
 284                         break;
 285                 next = decomp->next;
 286                 mmap_len = decomp->mmap_len;
 287                 munmap(decomp, mmap_len);
 288         } while (1);
 289 }
 290
 291 void perf_session__delete(struct perf_session *session)
 292 {
 293         if (session == NULL)
 294                 return;
 295         auxtrace__free(session);
 296         auxtrace_index__free(&session->auxtrace_index);
 297         perf_session__destroy_kernel_maps(session);
 298         perf_session__delete_threads(session);
 299         perf_session__release_decomp_events(session);
 300         perf_env__exit(&session->header.env);
 301         machines__exit(&session->machines);
 302         if (session->data)
 303                 perf_data__close(session->data);
 304         free(session);
 305 }
 306
 307 static int process_event_synth_tracing_data_stub(struct perf_session *session
 308                                                  __maybe_unused,
 309                                                  union perf_event *event
 310                                                  __maybe_unused)
 311 {
 312         dump_printf(": unhandled!\n");
 313         return 0;
 314 }
 315
 316 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
 317                                          union perf_event *event __maybe_unused,
 318                                          struct evlist **pevlist
 319                                          __maybe_unused)
 320 {
 321         dump_printf(": unhandled!\n");
 322         return 0;
 323 }
 324
 325 static int process_event_synth_event_update_stub(struct perf_tool *tool __maybe_unused,
 326                                                  union perf_event *event __maybe_unused,
 327                                                  struct evlist **pevlist
 328                                                  __maybe_unused)
 329 {
 330         if (dump_trace)
 331                 perf_event__fprintf_event_update(event, stdout);
 332
 333         dump_printf(": unhandled!\n");
 334         return 0;
 335 }
 336
 337 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
 338                                      union perf_event *event __maybe_unused,
 339                                      struct perf_sample *sample __maybe_unused,
 340                                      struct evsel *evsel __maybe_unused,
 341                                      struct machine *machine __maybe_unused)
 342 {
 343         dump_printf(": unhandled!\n");
 344         return 0;
 345 }
 346
 347 static int process_event_stub(struct perf_tool *tool __maybe_unused,
 348                               union perf_event *event __maybe_unused,
 349                               struct perf_sample *sample __maybe_unused,
 350                               struct machine *machine __maybe_unused)
 351 {
 352         dump_printf(": unhandled!\n");
 353         return 0;
 354 }
 355
 356 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
 357                                        union perf_event *event __maybe_unused,
 358                                        struct ordered_events *oe __maybe_unused)
 359 {
 360         dump_printf(": unhandled!\n");
 361         return 0;
 362 }
 363
 364 static int process_finished_round(struct perf_tool *tool,
 365                                   union perf_event *event,
 366                                   struct ordered_events *oe);
 367
 368 static int skipn(int fd, off_t n)
 369 {
 370         char buf[4096];
 371         ssize_t ret;
 372
 373         while (n > 0) {
 374                 ret = read(fd, buf, min(n, (off_t)sizeof(buf)));
 375                 if (ret <= 0)
 376                         return ret;
 377                 n -= ret;
 378         }
 379
 380         return 0;
 381 }
 382
 383 static s64 process_event_auxtrace_stub(struct perf_session *session __maybe_unused,
 384                                        union perf_event *event)
 385 {
 386         dump_printf(": unhandled!\n");
 387         if (perf_data__is_pipe(session->data))
 388                 skipn(perf_data__fd(session->data), event->auxtrace.size);
 389         return event->auxtrace.size;
 390 }
 391
 392 static int process_event_op2_stub(struct perf_session *session __maybe_unused,
 393                                   union perf_event *event __maybe_unused)
 394 {
 395         dump_printf(": unhandled!\n");
 396         return 0;
 397 }
 398
 399
 400 static
 401 int process_event_thread_map_stub(struct perf_session *session __maybe_unused,
 402                                   union perf_event *event __maybe_unused)
 403 {
 404         if (dump_trace)
 405                 perf_event__fprintf_thread_map(event, stdout);
 406
 407         dump_printf(": unhandled!\n");
 408         return 0;
 409 }
 410
 411 static
 412 int process_event_cpu_map_stub(struct perf_session *session __maybe_unused,
 413                                union perf_event *event __maybe_unused)
 414 {
 415         if (dump_trace)
 416                 perf_event__fprintf_cpu_map(event, stdout);
 417
 418         dump_printf(": unhandled!\n");
 419         return 0;
 420 }
 421
 422 static
 423 int process_event_stat_config_stub(struct perf_session *session __maybe_unused,
 424                                    union perf_event *event __maybe_unused)
 425 {
 426         if (dump_trace)
 427                 perf_event__fprintf_stat_config(event, stdout);
 428
 429         dump_printf(": unhandled!\n");
 430         return 0;
 431 }
 432
 433 static int process_stat_stub(struct perf_session *perf_session __maybe_unused,
 434                              union perf_event *event)
 435 {
 436         if (dump_trace)
 437                 perf_event__fprintf_stat(event, stdout);
 438
 439         dump_printf(": unhandled!\n");
 440         return 0;
 441 }
 442
 443 static int process_stat_round_stub(struct perf_session *perf_session __maybe_unused,
 444                                    union perf_event *event)
 445 {
 446         if (dump_trace)
 447                 perf_event__fprintf_stat_round(event, stdout);
 448
 449         dump_printf(": unhandled!\n");
 450         return 0;
 451 }
 452
 453 static int perf_session__process_compressed_event_stub(struct perf_session *session __maybe_unused,
 454                                                        union perf_event *event __maybe_unused,
 455                                                        u64 file_offset __maybe_unused)
 456 {
 457        dump_printf(": unhandled!\n");
 458        return 0;
 459 }
 460
 461 void perf_tool__fill_defaults(struct perf_tool *tool)
 462 {
 463         if (tool->sample == NULL)
 464                 tool->sample = process_event_sample_stub;
 465         if (tool->mmap == NULL)
 466                 tool->mmap = process_event_stub;
 467         if (tool->mmap2 == NULL)
 468                 tool->mmap2 = process_event_stub;
 469         if (tool->comm == NULL)
 470                 tool->comm = process_event_stub;
 471         if (tool->namespaces == NULL)
 472                 tool->namespaces = process_event_stub;
 473         if (tool->cgroup == NULL)
 474                 tool->cgroup = process_event_stub;
 475         if (tool->fork == NULL)
 476                 tool->fork = process_event_stub;
 477         if (tool->exit == NULL)
 478                 tool->exit = process_event_stub;
 479         if (tool->lost == NULL)
 480                 tool->lost = perf_event__process_lost;
 481         if (tool->lost_samples == NULL)
 482                 tool->lost_samples = perf_event__process_lost_samples;
 483         if (tool->aux == NULL)
 484                 tool->aux = perf_event__process_aux;
 485         if (tool->itrace_start == NULL)
 486                 tool->itrace_start = perf_event__process_itrace_start;
 487         if (tool->context_switch == NULL)
 488                 tool->context_switch = perf_event__process_switch;
 489         if (tool->ksymbol == NULL)
 490                 tool->ksymbol = perf_event__process_ksymbol;
 491         if (tool->bpf == NULL)
 492                 tool->bpf = perf_event__process_bpf;
 493         if (tool->text_poke == NULL)
 494                 tool->text_poke = perf_event__process_text_poke;
 495         if (tool->read == NULL)
 496                 tool->read = process_event_sample_stub;
 497         if (tool->throttle == NULL)
 498                 tool->throttle = process_event_stub;
 499         if (tool->unthrottle == NULL)
 500                 tool->unthrottle = process_event_stub;
 501         if (tool->attr == NULL)
 502                 tool->attr = process_event_synth_attr_stub;
 503         if (tool->event_update == NULL)
 504                 tool->event_update = process_event_synth_event_update_stub;
 505         if (tool->tracing_data == NULL)
 506                 tool->tracing_data = process_event_synth_tracing_data_stub;
 507         if (tool->build_id == NULL)
 508                 tool->build_id = process_event_op2_stub;
 509         if (tool->finished_round == NULL) {
 510                 if (tool->ordered_events)
 511                         tool->finished_round = process_finished_round;
 512                 else
 513                         tool->finished_round = process_finished_round_stub;
 514         }
 515         if (tool->id_index == NULL)
 516                 tool->id_index = process_event_op2_stub;
 517         if (tool->auxtrace_info == NULL)
 518                 tool->auxtrace_info = process_event_op2_stub;
 519         if (tool->auxtrace == NULL)
 520                 tool->auxtrace = process_event_auxtrace_stub;
 521         if (tool->auxtrace_error == NULL)
 522                 tool->auxtrace_error = process_event_op2_stub;
 523         if (tool->thread_map == NULL)
 524                 tool->thread_map = process_event_thread_map_stub;
 525         if (tool->cpu_map == NULL)
 526                 tool->cpu_map = process_event_cpu_map_stub;
 527         if (tool->stat_config == NULL)
 528                 tool->stat_config = process_event_stat_config_stub;
 529         if (tool->stat == NULL)
 530                 tool->stat = process_stat_stub;
 531         if (tool->stat_round == NULL)
 532                 tool->stat_round = process_stat_round_stub;
 533         if (tool->time_conv == NULL)
 534                 tool->time_conv = process_event_op2_stub;
 535         if (tool->feature == NULL)
 536                 tool->feature = process_event_op2_stub;
 537         if (tool->compressed == NULL)
 538                 tool->compressed = perf_session__process_compressed_event;
 539 }
 540
 541 static void swap_sample_id_all(union perf_event *event, void *data)
 542 {
 543         void *end = (void *) event + event->header.size;
 544         int size = end - data;
 545
 546         BUG_ON(size % sizeof(u64));
 547         mem_bswap_64(data, size);
 548 }
 549
 550 static void perf_event__all64_swap(union perf_event *event,
 551                                    bool sample_id_all __maybe_unused)
 552 {
 553         struct perf_event_header *hdr = &event->header;
 554         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 555 }
 556
 557 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 558 {
 559         event->comm.pid = bswap_32(event->comm.pid);
 560         event->comm.tid = bswap_32(event->comm.tid);
 561
 562         if (sample_id_all) {
 563                 void *data = &event->comm.comm;
 564
 565                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 566                 swap_sample_id_all(event, data);
 567         }
 568 }
 569
 570 static void perf_event__mmap_swap(union perf_event *event,
 571                                   bool sample_id_all)
 572 {
 573         event->mmap.pid   = bswap_32(event->mmap.pid);
 574         event->mmap.tid   = bswap_32(event->mmap.tid);
 575         event->mmap.start = bswap_64(event->mmap.start);
 576         event->mmap.len   = bswap_64(event->mmap.len);
 577         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 578
 579         if (sample_id_all) {
 580                 void *data = &event->mmap.filename;
 581
 582                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 583                 swap_sample_id_all(event, data);
 584         }
 585 }
 586
 587 static void perf_event__mmap2_swap(union perf_event *event,
 588                                   bool sample_id_all)
 589 {
 590         event->mmap2.pid   = bswap_32(event->mmap2.pid);
 591         event->mmap2.tid   = bswap_32(event->mmap2.tid);
 592         event->mmap2.start = bswap_64(event->mmap2.start);
 593         event->mmap2.len   = bswap_64(event->mmap2.len);
 594         event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
 595         event->mmap2.maj   = bswap_32(event->mmap2.maj);
 596         event->mmap2.min   = bswap_32(event->mmap2.min);
 597         event->mmap2.ino   = bswap_64(event->mmap2.ino);
 598         event->mmap2.ino_generation = bswap_64(event->mmap2.ino_generation);
 599
 600         if (sample_id_all) {
 601                 void *data = &event->mmap2.filename;
 602
 603                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 604                 swap_sample_id_all(event, data);
 605         }
 606 }
 607 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 608 {
 609         event->fork.pid  = bswap_32(event->fork.pid);
 610         event->fork.tid  = bswap_32(event->fork.tid);
 611         event->fork.ppid = bswap_32(event->fork.ppid);
 612         event->fork.ptid = bswap_32(event->fork.ptid);
 613         event->fork.time = bswap_64(event->fork.time);
 614
 615         if (sample_id_all)
 616                 swap_sample_id_all(event, &event->fork + 1);
 617 }
 618
 619 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 620 {
 621         event->read.pid          = bswap_32(event->read.pid);
 622         event->read.tid          = bswap_32(event->read.tid);
 623         event->read.value        = bswap_64(event->read.value);
 624         event->read.time_enabled = bswap_64(event->read.time_enabled);
 625         event->read.time_running = bswap_64(event->read.time_running);
 626         event->read.id           = bswap_64(event->read.id);
 627
 628         if (sample_id_all)
 629                 swap_sample_id_all(event, &event->read + 1);
 630 }
 631
 632 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all)
 633 {
 634         event->aux.aux_offset = bswap_64(event->aux.aux_offset);
 635         event->aux.aux_size   = bswap_64(event->aux.aux_size);
 636         event->aux.flags      = bswap_64(event->aux.flags);
 637
 638         if (sample_id_all)
 639                 swap_sample_id_all(event, &event->aux + 1);
 640 }
 641
 642 static void perf_event__itrace_start_swap(union perf_event *event,
 643                                           bool sample_id_all)
 644 {
 645         event->itrace_start.pid  = bswap_32(event->itrace_start.pid);
 646         event->itrace_start.tid  = bswap_32(event->itrace_start.tid);
 647
 648         if (sample_id_all)
 649                 swap_sample_id_all(event, &event->itrace_start + 1);
 650 }
 651
 652 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all)
 653 {
 654         if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) {
 655                 event->context_switch.next_prev_pid =
 656                                 bswap_32(event->context_switch.next_prev_pid);
 657                 event->context_switch.next_prev_tid =
 658                                 bswap_32(event->context_switch.next_prev_tid);
 659         }
 660
 661         if (sample_id_all)
 662                 swap_sample_id_all(event, &event->context_switch + 1);
 663 }
 664
 665 static void perf_event__text_poke_swap(union perf_event *event, bool sample_id_all)
 666 {
 667         event->text_poke.addr    = bswap_64(event->text_poke.addr);
 668         event->text_poke.old_len = bswap_16(event->text_poke.old_len);
 669         event->text_poke.new_len = bswap_16(event->text_poke.new_len);
 670
 671         if (sample_id_all) {
 672                 size_t len = sizeof(event->text_poke.old_len) +
 673                              sizeof(event->text_poke.new_len) +
 674                              event->text_poke.old_len +
 675                              event->text_poke.new_len;
 676                 void *data = &event->text_poke.old_len;
 677
 678                 data += PERF_ALIGN(len, sizeof(u64));
 679                 swap_sample_id_all(event, data);
 680         }
 681 }
 682
 683 static void perf_event__throttle_swap(union perf_event *event,
 684                                       bool sample_id_all)
 685 {
 686         event->throttle.time      = bswap_64(event->throttle.time);
 687         event->throttle.id        = bswap_64(event->throttle.id);
 688         event->throttle.stream_id = bswap_64(event->throttle.stream_id);
 689
 690         if (sample_id_all)
 691                 swap_sample_id_all(event, &event->throttle + 1);
 692 }
 693
 694 static void perf_event__namespaces_swap(union perf_event *event,
 695                                         bool sample_id_all)
 696 {
 697         u64 i;
 698
 699         event->namespaces.pid           = bswap_32(event->namespaces.pid);
 700         event->namespaces.tid           = bswap_32(event->namespaces.tid);
 701         event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces);
 702
 703         for (i = 0; i < event->namespaces.nr_namespaces; i++) {
 704                 struct perf_ns_link_info *ns = &event->namespaces.link_info[i];
 705
 706                 ns->dev = bswap_64(ns->dev);
 707                 ns->ino = bswap_64(ns->ino);
 708         }
 709
 710         if (sample_id_all)
 711                 swap_sample_id_all(event, &event->namespaces.link_info[i]);
 712 }
 713
 714 static void perf_event__cgroup_swap(union perf_event *event, bool sample_id_all)
 715 {
 716         event->cgroup.id = bswap_64(event->cgroup.id);
 717
 718         if (sample_id_all) {
 719                 void *data = &event->cgroup.path;
 720
 721                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 722                 swap_sample_id_all(event, data);
 723         }
 724 }
 725
 726 static u8 revbyte(u8 b)
 727 {
 728         int rev = (b >> 4) | ((b & 0xf) << 4);
 729         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 730         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 731         return (u8) rev;
 732 }
 733
 734 /*
 735  * XXX this is hack in attempt to carry flags bitfield
 736  * through endian village. ABI says:
 737  *
 738  * Bit-fields are allocated from right to left (least to most significant)
 739  * on little-endian implementations and from left to right (most to least
 740  * significant) on big-endian implementations.
 741  *
 742  * The above seems to be byte specific, so we need to reverse each
 743  * byte of the bitfield. 'Internet' also says this might be implementation
 744  * specific and we probably need proper fix and carry perf_event_attr
 745  * bitfield flags in separate data file FEAT_ section. Thought this seems
 746  * to work for now.
 747  */
 748 static void swap_bitfield(u8 *p, unsigned len)
 749 {
 750         unsigned i;
 751
 752         for (i = 0; i < len; i++) {
 753                 *p = revbyte(*p);
 754                 p++;
 755         }
 756 }
 757
 758 /* exported for swapping attributes in file header */
 759 void perf_event__attr_swap(struct perf_event_attr *attr)
 760 {
 761         attr->type              = bswap_32(attr->type);
 762         attr->size              = bswap_32(attr->size);
 763
 764 #define bswap_safe(f, n)                                        \
 765         (attr->size > (offsetof(struct perf_event_attr, f) +    \
 766                        sizeof(attr->f) * (n)))
 767 #define bswap_field(f, sz)                      \
 768 do {                                            \
 769         if (bswap_safe(f, 0))                   \
 770                 attr->f = bswap_##sz(attr->f);  \
 771 } while(0)
 772 #define bswap_field_16(f) bswap_field(f, 16)
 773 #define bswap_field_32(f) bswap_field(f, 32)
 774 #define bswap_field_64(f) bswap_field(f, 64)
 775
 776         bswap_field_64(config);
 777         bswap_field_64(sample_period);
 778         bswap_field_64(sample_type);
 779         bswap_field_64(read_format);
 780         bswap_field_32(wakeup_events);
 781         bswap_field_32(bp_type);
 782         bswap_field_64(bp_addr);
 783         bswap_field_64(bp_len);
 784         bswap_field_64(branch_sample_type);
 785         bswap_field_64(sample_regs_user);
 786         bswap_field_32(sample_stack_user);
 787         bswap_field_32(aux_watermark);
 788         bswap_field_16(sample_max_stack);
 789         bswap_field_32(aux_sample_size);
 790
 791         /*
 792          * After read_format are bitfields. Check read_format because
 793          * we are unable to use offsetof on bitfield.
 794          */
 795         if (bswap_safe(read_format, 1))
 796                 swap_bitfield((u8 *) (&attr->read_format + 1),
 797                               sizeof(u64));
 798 #undef bswap_field_64
 799 #undef bswap_field_32
 800 #undef bswap_field
 801 #undef bswap_safe
 802 }
 803
 804 static void perf_event__hdr_attr_swap(union perf_event *event,
 805                                       bool sample_id_all __maybe_unused)
 806 {
 807         size_t size;
 808
 809         perf_event__attr_swap(&event->attr.attr);
 810
 811         size = event->header.size;
 812         size -= (void *)&event->attr.id - (void *)event;
 813         mem_bswap_64(event->attr.id, size);
 814 }
 815
 816 static void perf_event__event_update_swap(union perf_event *event,
 817                                           bool sample_id_all __maybe_unused)
 818 {
 819         event->event_update.type = bswap_64(event->event_update.type);
 820         event->event_update.id   = bswap_64(event->event_update.id);
 821 }
 822
 823 static void perf_event__event_type_swap(union perf_event *event,
 824                                         bool sample_id_all __maybe_unused)
 825 {
 826         event->event_type.event_type.event_id =
 827                 bswap_64(event->event_type.event_type.event_id);
 828 }
 829
 830 static void perf_event__tracing_data_swap(union perf_event *event,
 831                                           bool sample_id_all __maybe_unused)
 832 {
 833         event->tracing_data.size = bswap_32(event->tracing_data.size);
 834 }
 835
 836 static void perf_event__auxtrace_info_swap(union perf_event *event,
 837                                            bool sample_id_all __maybe_unused)
 838 {
 839         size_t size;
 840
 841         event->auxtrace_info.type = bswap_32(event->auxtrace_info.type);
 842
 843         size = event->header.size;
 844         size -= (void *)&event->auxtrace_info.priv - (void *)event;
 845         mem_bswap_64(event->auxtrace_info.priv, size);
 846 }
 847
 848 static void perf_event__auxtrace_swap(union perf_event *event,
 849                                       bool sample_id_all __maybe_unused)
 850 {
 851         event->auxtrace.size      = bswap_64(event->auxtrace.size);
 852         event->auxtrace.offset    = bswap_64(event->auxtrace.offset);
 853         event->auxtrace.reference = bswap_64(event->auxtrace.reference);
 854         event->auxtrace.idx       = bswap_32(event->auxtrace.idx);
 855         event->auxtrace.tid       = bswap_32(event->auxtrace.tid);
 856         event->auxtrace.cpu       = bswap_32(event->auxtrace.cpu);
 857 }
 858
 859 static void perf_event__auxtrace_error_swap(union perf_event *event,
 860                                             bool sample_id_all __maybe_unused)
 861 {
 862         event->auxtrace_error.type = bswap_32(event->auxtrace_error.type);
 863         event->auxtrace_error.code = bswap_32(event->auxtrace_error.code);
 864         event->auxtrace_error.cpu  = bswap_32(event->auxtrace_error.cpu);
 865         event->auxtrace_error.pid  = bswap_32(event->auxtrace_error.pid);
 866         event->auxtrace_error.tid  = bswap_32(event->auxtrace_error.tid);
 867         event->auxtrace_error.fmt  = bswap_32(event->auxtrace_error.fmt);
 868         event->auxtrace_error.ip   = bswap_64(event->auxtrace_error.ip);
 869         if (event->auxtrace_error.fmt)
 870                 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time);
 871 }
 872
 873 static void perf_event__thread_map_swap(union perf_event *event,
 874                                         bool sample_id_all __maybe_unused)
 875 {
 876         unsigned i;
 877
 878         event->thread_map.nr = bswap_64(event->thread_map.nr);
 879
 880         for (i = 0; i < event->thread_map.nr; i++)
 881                 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid);
 882 }
 883
 884 static void perf_event__cpu_map_swap(union perf_event *event,
 885                                      bool sample_id_all __maybe_unused)
 886 {
 887         struct perf_record_cpu_map_data *data = &event->cpu_map.data;
 888         struct cpu_map_entries *cpus;
 889         struct perf_record_record_cpu_map *mask;
 890         unsigned i;
 891
 892         data->type = bswap_64(data->type);
 893
 894         switch (data->type) {
 895         case PERF_CPU_MAP__CPUS:
 896                 cpus = (struct cpu_map_entries *)data->data;
 897
 898                 cpus->nr = bswap_16(cpus->nr);
 899
 900                 for (i = 0; i < cpus->nr; i++)
 901                         cpus->cpu[i] = bswap_16(cpus->cpu[i]);
 902                 break;
 903         case PERF_CPU_MAP__MASK:
 904                 mask = (struct perf_record_record_cpu_map *)data->data;
 905
 906                 mask->nr = bswap_16(mask->nr);
 907                 mask->long_size = bswap_16(mask->long_size);
 908
 909                 switch (mask->long_size) {
 910                 case 4: mem_bswap_32(&mask->mask, mask->nr); break;
 911                 case 8: mem_bswap_64(&mask->mask, mask->nr); break;
 912                 default:
 913                         pr_err("cpu_map swap: unsupported long size\n");
 914                 }
 915         default:
 916                 break;
 917         }
 918 }
 919
 920 static void perf_event__stat_config_swap(union perf_event *event,
 921                                          bool sample_id_all __maybe_unused)
 922 {
 923         u64 size;
 924
 925         size  = event->stat_config.nr * sizeof(event->stat_config.data[0]);
 926         size += 1; /* nr item itself */
 927         mem_bswap_64(&event->stat_config.nr, size);
 928 }
 929
 930 static void perf_event__stat_swap(union perf_event *event,
 931                                   bool sample_id_all __maybe_unused)
 932 {
 933         event->stat.id     = bswap_64(event->stat.id);
 934         event->stat.thread = bswap_32(event->stat.thread);
 935         event->stat.cpu    = bswap_32(event->stat.cpu);
 936         event->stat.val    = bswap_64(event->stat.val);
 937         event->stat.ena    = bswap_64(event->stat.ena);
 938         event->stat.run    = bswap_64(event->stat.run);
 939 }
 940
 941 static void perf_event__stat_round_swap(union perf_event *event,
 942                                         bool sample_id_all __maybe_unused)
 943 {
 944         event->stat_round.type = bswap_64(event->stat_round.type);
 945         event->stat_round.time = bswap_64(event->stat_round.time);
 946 }
 947
 948 typedef void (*perf_event__swap_op)(union perf_event *event,
 949                                     bool sample_id_all);
 950
 951 static perf_event__swap_op perf_event__swap_ops[] = {
 952         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 953         [PERF_RECORD_MMAP2]               = perf_event__mmap2_swap,
 954         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 955         [PERF_RECORD_FORK]                = perf_event__task_swap,
 956         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 957         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 958         [PERF_RECORD_READ]                = perf_event__read_swap,
 959         [PERF_RECORD_THROTTLE]            = perf_event__throttle_swap,
 960         [PERF_RECORD_UNTHROTTLE]          = perf_event__throttle_swap,
 961         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 962         [PERF_RECORD_AUX]                 = perf_event__aux_swap,
 963         [PERF_RECORD_ITRACE_START]        = perf_event__itrace_start_swap,
 964         [PERF_RECORD_LOST_SAMPLES]        = perf_event__all64_swap,
 965         [PERF_RECORD_SWITCH]              = perf_event__switch_swap,
 966         [PERF_RECORD_SWITCH_CPU_WIDE]     = perf_event__switch_swap,
 967         [PERF_RECORD_NAMESPACES]          = perf_event__namespaces_swap,
 968         [PERF_RECORD_CGROUP]              = perf_event__cgroup_swap,
 969         [PERF_RECORD_TEXT_POKE]           = perf_event__text_poke_swap,
 970         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
 971         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
 972         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
 973         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
 974         [PERF_RECORD_ID_INDEX]            = perf_event__all64_swap,
 975         [PERF_RECORD_AUXTRACE_INFO]       = perf_event__auxtrace_info_swap,
 976         [PERF_RECORD_AUXTRACE]            = perf_event__auxtrace_swap,
 977         [PERF_RECORD_AUXTRACE_ERROR]      = perf_event__auxtrace_error_swap,
 978         [PERF_RECORD_THREAD_MAP]          = perf_event__thread_map_swap,
 979         [PERF_RECORD_CPU_MAP]             = perf_event__cpu_map_swap,
 980         [PERF_RECORD_STAT_CONFIG]         = perf_event__stat_config_swap,
 981         [PERF_RECORD_STAT]                = perf_event__stat_swap,
 982         [PERF_RECORD_STAT_ROUND]          = perf_event__stat_round_swap,
 983         [PERF_RECORD_EVENT_UPDATE]        = perf_event__event_update_swap,
 984         [PERF_RECORD_TIME_CONV]           = perf_event__all64_swap,
 985         [PERF_RECORD_HEADER_MAX]          = NULL,
 986 };
 987
 988 /*
 989  * When perf record finishes a pass on every buffers, it records this pseudo
 990  * event.
 991  * We record the max timestamp t found in the pass n.
 992  * Assuming these timestamps are monotonic across cpus, we know that if
 993  * a buffer still has events with timestamps below t, they will be all
 994  * available and then read in the pass n + 1.
 995  * Hence when we start to read the pass n + 2, we can safely flush every
 996  * events with timestamps below t.
 997  *
 998  *    ============ PASS n =================
 999  *       CPU 0         |   CPU 1
1000  *                     |
1001  *    cnt1 timestamps  |   cnt2 timestamps
1002  *          1          |         2
1003  *          2          |         3
1004  *          -          |         4  <--- max recorded
1005  *
1006  *    ============ PASS n + 1 ==============
1007  *       CPU 0         |   CPU 1
1008  *                     |
1009  *    cnt1 timestamps  |   cnt2 timestamps
1010  *          3          |         5
1011  *          4          |         6
1012  *          5          |         7 <---- max recorded
1013  *
1014  *      Flush every events below timestamp 4
1015  *
1016  *    ============ PASS n + 2 ==============
1017  *       CPU 0         |   CPU 1
1018  *                     |
1019  *    cnt1 timestamps  |   cnt2 timestamps
1020  *          6          |         8
1021  *          7          |         9
1022  *          -          |         10
1023  *
1024  *      Flush every events below timestamp 7
1025  *      etc...
1026  */
1027 static int process_finished_round(struct perf_tool *tool __maybe_unused,
1028                                   union perf_event *event __maybe_unused,
1029                                   struct ordered_events *oe)
1030 {
1031         if (dump_trace)
1032                 fprintf(stdout, "\n");
1033         return ordered_events__flush(oe, OE_FLUSH__ROUND);
1034 }
1035
1036 int perf_session__queue_event(struct perf_session *s, union perf_event *event,
1037                               u64 timestamp, u64 file_offset)
1038 {
1039         return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset);
1040 }
1041
1042 static void callchain__lbr_callstack_printf(struct perf_sample *sample)
1043 {
1044         struct ip_callchain *callchain = sample->callchain;
1045         struct branch_stack *lbr_stack = sample->branch_stack;
1046         struct branch_entry *entries = perf_sample__branch_entries(sample);
1047         u64 kernel_callchain_nr = callchain->nr;
1048         unsigned int i;
1049
1050         for (i = 0; i < kernel_callchain_nr; i++) {
1051                 if (callchain->ips[i] == PERF_CONTEXT_USER)
1052                         break;
1053         }
1054
1055         if ((i != kernel_callchain_nr) && lbr_stack->nr) {
1056                 u64 total_nr;
1057                 /*
1058                  * LBR callstack can only get user call chain,
1059                  * i is kernel call chain number,
1060                  * 1 is PERF_CONTEXT_USER.
1061                  *
1062                  * The user call chain is stored in LBR registers.
1063                  * LBR are pair registers. The caller is stored
1064                  * in "from" register, while the callee is stored
1065                  * in "to" register.
1066                  * For example, there is a call stack
1067                  * "A"->"B"->"C"->"D".
1068                  * The LBR registers will recorde like
1069                  * "C"->"D", "B"->"C", "A"->"B".
1070                  * So only the first "to" register and all "from"
1071                  * registers are needed to construct the whole stack.
1072                  */
1073                 total_nr = i + 1 + lbr_stack->nr + 1;
1074                 kernel_callchain_nr = i + 1;
1075
1076                 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr);
1077
1078                 for (i = 0; i < kernel_callchain_nr; i++)
1079                         printf("..... %2d: %016" PRIx64 "\n",
1080                                i, callchain->ips[i]);
1081
1082                 printf("..... %2d: %016" PRIx64 "\n",
1083                        (int)(kernel_callchain_nr), entries[0].to);
1084                 for (i = 0; i < lbr_stack->nr; i++)
1085                         printf("..... %2d: %016" PRIx64 "\n",
1086                                (int)(i + kernel_callchain_nr + 1), entries[i].from);
1087         }
1088 }
1089
1090 static void callchain__printf(struct evsel *evsel,
1091                               struct perf_sample *sample)
1092 {
1093         unsigned int i;
1094         struct ip_callchain *callchain = sample->callchain;
1095
1096         if (evsel__has_branch_callstack(evsel))
1097                 callchain__lbr_callstack_printf(sample);
1098
1099         printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr);
1100
1101         for (i = 0; i < callchain->nr; i++)
1102                 printf("..... %2d: %016" PRIx64 "\n",
1103                        i, callchain->ips[i]);
1104 }
1105
1106 static void branch_stack__printf(struct perf_sample *sample, bool callstack)
1107 {
1108         struct branch_entry *entries = perf_sample__branch_entries(sample);
1109         uint64_t i;
1110
1111         printf("%s: nr:%" PRIu64 "\n",
1112                 !callstack ? "... branch stack" : "... branch callstack",
1113                 sample->branch_stack->nr);
1114
1115         for (i = 0; i < sample->branch_stack->nr; i++) {
1116                 struct branch_entry *e = &entries[i];
1117
1118                 if (!callstack) {
1119                         printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x\n",
1120                                 i, e->from, e->to,
1121                                 (unsigned short)e->flags.cycles,
1122                                 e->flags.mispred ? "M" : " ",
1123                                 e->flags.predicted ? "P" : " ",
1124                                 e->flags.abort ? "A" : " ",
1125                                 e->flags.in_tx ? "T" : " ",
1126                                 (unsigned)e->flags.reserved);
1127                 } else {
1128                         printf("..... %2"PRIu64": %016" PRIx64 "\n",
1129                                 i, i > 0 ? e->from : e->to);
1130                 }
1131         }
1132 }
1133
1134 static void regs_dump__printf(u64 mask, u64 *regs)
1135 {
1136         unsigned rid, i = 0;
1137
1138         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
1139                 u64 val = regs[i++];
1140
1141                 printf(".... %-5s 0x%016" PRIx64 "\n",
1142                        perf_reg_name(rid), val);
1143         }
1144 }
1145
1146 static const char *regs_abi[] = {
1147         [PERF_SAMPLE_REGS_ABI_NONE] = "none",
1148         [PERF_SAMPLE_REGS_ABI_32] = "32-bit",
1149         [PERF_SAMPLE_REGS_ABI_64] = "64-bit",
1150 };
1151
1152 static inline const char *regs_dump_abi(struct regs_dump *d)
1153 {
1154         if (d->abi > PERF_SAMPLE_REGS_ABI_64)
1155                 return "unknown";
1156
1157         return regs_abi[d->abi];
1158 }
1159
1160 static void regs__printf(const char *type, struct regs_dump *regs)
1161 {
1162         u64 mask = regs->mask;
1163
1164         printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n",
1165                type,
1166                mask,
1167                regs_dump_abi(regs));
1168
1169         regs_dump__printf(mask, regs->regs);
1170 }
1171
1172 static void regs_user__printf(struct perf_sample *sample)
1173 {
1174         struct regs_dump *user_regs = &sample->user_regs;
1175
1176         if (user_regs->regs)
1177                 regs__printf("user", user_regs);
1178 }
1179
1180 static void regs_intr__printf(struct perf_sample *sample)
1181 {
1182         struct regs_dump *intr_regs = &sample->intr_regs;
1183
1184         if (intr_regs->regs)
1185                 regs__printf("intr", intr_regs);
1186 }
1187
1188 static void stack_user__printf(struct stack_dump *dump)
1189 {
1190         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
1191                dump->size, dump->offset);
1192 }
1193
1194 static void perf_evlist__print_tstamp(struct evlist *evlist,
1195                                        union perf_event *event,
1196                                        struct perf_sample *sample)
1197 {
1198         u64 sample_type = __evlist__combined_sample_type(evlist);
1199
1200         if (event->header.type != PERF_RECORD_SAMPLE &&
1201             !evlist__sample_id_all(evlist)) {
1202                 fputs("-1 -1 ", stdout);
1203                 return;
1204         }
1205
1206         if ((sample_type & PERF_SAMPLE_CPU))
1207                 printf("%u ", sample->cpu);
1208
1209         if (sample_type & PERF_SAMPLE_TIME)
1210                 printf("%" PRIu64 " ", sample->time);
1211 }
1212
1213 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
1214 {
1215         printf("... sample_read:\n");
1216
1217         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1218                 printf("...... time enabled %016" PRIx64 "\n",
1219                        sample->read.time_enabled);
1220
1221         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1222                 printf("...... time running %016" PRIx64 "\n",
1223                        sample->read.time_running);
1224
1225         if (read_format & PERF_FORMAT_GROUP) {
1226                 u64 i;
1227
1228                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
1229
1230                 for (i = 0; i < sample->read.group.nr; i++) {
1231                         struct sample_read_value *value;
1232
1233                         value = &sample->read.group.values[i];
1234                         printf("..... id %016" PRIx64
1235                                ", value %016" PRIx64 "\n",
1236                                value->id, value->value);
1237                 }
1238         } else
1239                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
1240                         sample->read.one.id, sample->read.one.value);
1241 }
1242
1243 static void dump_event(struct evlist *evlist, union perf_event *event,
1244                        u64 file_offset, struct perf_sample *sample)
1245 {
1246         if (!dump_trace)
1247                 return;
1248
1249         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
1250                file_offset, event->header.size, event->header.type);
1251
1252         trace_event(event);
1253         if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw)
1254                 evlist->trace_event_sample_raw(evlist, event, sample);
1255
1256         if (sample)
1257                 perf_evlist__print_tstamp(evlist, event, sample);
1258
1259         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
1260                event->header.size, perf_event__name(event->header.type));
1261 }
1262
1263 static void dump_sample(struct evsel *evsel, union perf_event *event,
1264                         struct perf_sample *sample)
1265 {
1266         u64 sample_type;
1267
1268         if (!dump_trace)
1269                 return;
1270
1271         printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
1272                event->header.misc, sample->pid, sample->tid, sample->ip,
1273                sample->period, sample->addr);
1274
1275         sample_type = evsel->core.attr.sample_type;
1276
1277         if (evsel__has_callchain(evsel))
1278                 callchain__printf(evsel, sample);
1279
1280         if (evsel__has_br_stack(evsel))
1281                 branch_stack__printf(sample, evsel__has_branch_callstack(evsel));
1282
1283         if (sample_type & PERF_SAMPLE_REGS_USER)
1284                 regs_user__printf(sample);
1285
1286         if (sample_type & PERF_SAMPLE_REGS_INTR)
1287                 regs_intr__printf(sample);
1288
1289         if (sample_type & PERF_SAMPLE_STACK_USER)
1290                 stack_user__printf(&sample->user_stack);
1291
1292         if (sample_type & PERF_SAMPLE_WEIGHT)
1293                 printf("... weight: %" PRIu64 "\n", sample->weight);
1294
1295         if (sample_type & PERF_SAMPLE_DATA_SRC)
1296                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
1297
1298         if (sample_type & PERF_SAMPLE_PHYS_ADDR)
1299                 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr);
1300
1301         if (sample_type & PERF_SAMPLE_TRANSACTION)
1302                 printf("... transaction: %" PRIx64 "\n", sample->transaction);
1303
1304         if (sample_type & PERF_SAMPLE_READ)
1305                 sample_read__printf(sample, evsel->core.attr.read_format);
1306 }
1307
1308 static void dump_read(struct evsel *evsel, union perf_event *event)
1309 {
1310         struct perf_record_read *read_event = &event->read;
1311         u64 read_format;
1312
1313         if (!dump_trace)
1314                 return;
1315
1316         printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid,
1317                evsel__name(evsel), event->read.value);
1318
1319         if (!evsel)
1320                 return;
1321
1322         read_format = evsel->core.attr.read_format;
1323
1324         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1325                 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled);
1326
1327         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1328                 printf("... time running : %" PRI_lu64 "\n", read_event->time_running);
1329
1330         if (read_format & PERF_FORMAT_ID)
1331                 printf("... id           : %" PRI_lu64 "\n", read_event->id);
1332 }
1333
1334 static struct machine *machines__find_for_cpumode(struct machines *machines,
1335                                                union perf_event *event,
1336                                                struct perf_sample *sample)
1337 {
1338         struct machine *machine;
1339
1340         if (perf_guest &&
1341             ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
1342              (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) {
1343                 u32 pid;
1344
1345                 if (event->header.type == PERF_RECORD_MMAP
1346                     || event->header.type == PERF_RECORD_MMAP2)
1347                         pid = event->mmap.pid;
1348                 else
1349                         pid = sample->pid;
1350
1351                 machine = machines__find(machines, pid);
1352                 if (!machine)
1353                         machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID);
1354                 return machine;
1355         }
1356
1357         return &machines->host;
1358 }
1359
1360 static int deliver_sample_value(struct evlist *evlist,
1361                                 struct perf_tool *tool,
1362                                 union perf_event *event,
1363                                 struct perf_sample *sample,
1364                                 struct sample_read_value *v,
1365                                 struct machine *machine)
1366 {
1367         struct perf_sample_id *sid = perf_evlist__id2sid(evlist, v->id);
1368         struct evsel *evsel;
1369
1370         if (sid) {
1371                 sample->id     = v->id;
1372                 sample->period = v->value - sid->period;
1373                 sid->period    = v->value;
1374         }
1375
1376         if (!sid || sid->evsel == NULL) {
1377                 ++evlist->stats.nr_unknown_id;
1378                 return 0;
1379         }
1380
1381         /*
1382          * There's no reason to deliver sample
1383          * for zero period, bail out.
1384          */
1385         if (!sample->period)
1386                 return 0;
1387
1388         evsel = container_of(sid->evsel, struct evsel, core);
1389         return tool->sample(tool, event, sample, evsel, machine);
1390 }
1391
1392 static int deliver_sample_group(struct evlist *evlist,
1393                                 struct perf_tool *tool,
1394                                 union  perf_event *event,
1395                                 struct perf_sample *sample,
1396                                 struct machine *machine)
1397 {
1398         int ret = -EINVAL;
1399         u64 i;
1400
1401         for (i = 0; i < sample->read.group.nr; i++) {
1402                 ret = deliver_sample_value(evlist, tool, event, sample,
1403                                            &sample->read.group.values[i],
1404                                            machine);
1405                 if (ret)
1406                         break;
1407         }
1408
1409         return ret;
1410 }
1411
1412 static int
1413  perf_evlist__deliver_sample(struct evlist *evlist,
1414                              struct perf_tool *tool,
1415                              union  perf_event *event,
1416                              struct perf_sample *sample,
1417                              struct evsel *evsel,
1418                              struct machine *machine)
1419 {
1420         /* We know evsel != NULL. */
1421         u64 sample_type = evsel->core.attr.sample_type;
1422         u64 read_format = evsel->core.attr.read_format;
1423
1424         /* Standard sample delivery. */
1425         if (!(sample_type & PERF_SAMPLE_READ))
1426                 return tool->sample(tool, event, sample, evsel, machine);
1427
1428         /* For PERF_SAMPLE_READ we have either single or group mode. */
1429         if (read_format & PERF_FORMAT_GROUP)
1430                 return deliver_sample_group(evlist, tool, event, sample,
1431                                             machine);
1432         else
1433                 return deliver_sample_value(evlist, tool, event, sample,
1434                                             &sample->read.one, machine);
1435 }
1436
1437 static int machines__deliver_event(struct machines *machines,
1438                                    struct evlist *evlist,
1439                                    union perf_event *event,
1440                                    struct perf_sample *sample,
1441                                    struct perf_tool *tool, u64 file_offset)
1442 {
1443         struct evsel *evsel;
1444         struct machine *machine;
1445
1446         dump_event(evlist, event, file_offset, sample);
1447
1448         evsel = perf_evlist__id2evsel(evlist, sample->id);
1449
1450         machine = machines__find_for_cpumode(machines, event, sample);
1451
1452         switch (event->header.type) {
1453         case PERF_RECORD_SAMPLE:
1454                 if (evsel == NULL) {
1455                         ++evlist->stats.nr_unknown_id;
1456                         return 0;
1457                 }
1458                 dump_sample(evsel, event, sample);
1459                 if (machine == NULL) {
1460                         ++evlist->stats.nr_unprocessable_samples;
1461                         return 0;
1462                 }
1463                 return perf_evlist__deliver_sample(evlist, tool, event, sample, evsel, machine);
1464         case PERF_RECORD_MMAP:
1465                 return tool->mmap(tool, event, sample, machine);
1466         case PERF_RECORD_MMAP2:
1467                 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT)
1468                         ++evlist->stats.nr_proc_map_timeout;
1469                 return tool->mmap2(tool, event, sample, machine);
1470         case PERF_RECORD_COMM:
1471                 return tool->comm(tool, event, sample, machine);
1472         case PERF_RECORD_NAMESPACES:
1473                 return tool->namespaces(tool, event, sample, machine);
1474         case PERF_RECORD_CGROUP:
1475                 return tool->cgroup(tool, event, sample, machine);
1476         case PERF_RECORD_FORK:
1477                 return tool->fork(tool, event, sample, machine);
1478         case PERF_RECORD_EXIT:
1479                 return tool->exit(tool, event, sample, machine);
1480         case PERF_RECORD_LOST:
1481                 if (tool->lost == perf_event__process_lost)
1482                         evlist->stats.total_lost += event->lost.lost;
1483                 return tool->lost(tool, event, sample, machine);
1484         case PERF_RECORD_LOST_SAMPLES:
1485                 if (tool->lost_samples == perf_event__process_lost_samples)
1486                         evlist->stats.total_lost_samples += event->lost_samples.lost;
1487                 return tool->lost_samples(tool, event, sample, machine);
1488         case PERF_RECORD_READ:
1489                 dump_read(evsel, event);
1490                 return tool->read(tool, event, sample, evsel, machine);
1491         case PERF_RECORD_THROTTLE:
1492                 return tool->throttle(tool, event, sample, machine);
1493         case PERF_RECORD_UNTHROTTLE:
1494                 return tool->unthrottle(tool, event, sample, machine);
1495         case PERF_RECORD_AUX:
1496                 if (tool->aux == perf_event__process_aux) {
1497                         if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED)
1498                                 evlist->stats.total_aux_lost += 1;
1499                         if (event->aux.flags & PERF_AUX_FLAG_PARTIAL)
1500                                 evlist->stats.total_aux_partial += 1;
1501                 }
1502                 return tool->aux(tool, event, sample, machine);
1503         case PERF_RECORD_ITRACE_START:
1504                 return tool->itrace_start(tool, event, sample, machine);
1505         case PERF_RECORD_SWITCH:
1506         case PERF_RECORD_SWITCH_CPU_WIDE:
1507                 return tool->context_switch(tool, event, sample, machine);
1508         case PERF_RECORD_KSYMBOL:
1509                 return tool->ksymbol(tool, event, sample, machine);
1510         case PERF_RECORD_BPF_EVENT:
1511                 return tool->bpf(tool, event, sample, machine);
1512         case PERF_RECORD_TEXT_POKE:
1513                 return tool->text_poke(tool, event, sample, machine);
1514         default:
1515                 ++evlist->stats.nr_unknown_events;
1516                 return -1;
1517         }
1518 }
1519
1520 static int perf_session__deliver_event(struct perf_session *session,
1521                                        union perf_event *event,
1522                                        struct perf_tool *tool,
1523                                        u64 file_offset)
1524 {
1525         struct perf_sample sample;
1526         int ret;
1527
1528         ret = perf_evlist__parse_sample(session->evlist, event, &sample);
1529         if (ret) {
1530                 pr_err("Can't parse sample, err = %d\n", ret);
1531                 return ret;
1532         }
1533
1534         ret = auxtrace__process_event(session, event, &sample, tool);
1535         if (ret < 0)
1536                 return ret;
1537         if (ret > 0)
1538                 return 0;
1539
1540         ret = machines__deliver_event(&session->machines, session->evlist,
1541                                       event, &sample, tool, file_offset);
1542
1543         if (dump_trace && sample.aux_sample.size)
1544                 auxtrace__dump_auxtrace_sample(session, &sample);
1545
1546         return ret;
1547 }
1548
1549 static s64 perf_session__process_user_event(struct perf_session *session,
1550                                             union perf_event *event,
1551                                             u64 file_offset)
1552 {
1553         struct ordered_events *oe = &session->ordered_events;
1554         struct perf_tool *tool = session->tool;
1555         struct perf_sample sample = { .time = 0, };
1556         int fd = perf_data__fd(session->data);
1557         int err;
1558
1559         if (event->header.type != PERF_RECORD_COMPRESSED ||
1560             tool->compressed == perf_session__process_compressed_event_stub)
1561                 dump_event(session->evlist, event, file_offset, &sample);
1562
1563         /* These events are processed right away */
1564         switch (event->header.type) {
1565         case PERF_RECORD_HEADER_ATTR:
1566                 err = tool->attr(tool, event, &session->evlist);
1567                 if (err == 0) {
1568                         perf_session__set_id_hdr_size(session);
1569                         perf_session__set_comm_exec(session);
1570                 }
1571                 return err;
1572         case PERF_RECORD_EVENT_UPDATE:
1573                 return tool->event_update(tool, event, &session->evlist);
1574         case PERF_RECORD_HEADER_EVENT_TYPE:
1575                 /*
1576                  * Depreceated, but we need to handle it for sake
1577                  * of old data files create in pipe mode.
1578                  */
1579                 return 0;
1580         case PERF_RECORD_HEADER_TRACING_DATA:
1581                 /*
1582                  * Setup for reading amidst mmap, but only when we
1583                  * are in 'file' mode. The 'pipe' fd is in proper
1584                  * place already.
1585                  */
1586                 if (!perf_data__is_pipe(session->data))
1587                         lseek(fd, file_offset, SEEK_SET);
1588                 return tool->tracing_data(session, event);
1589         case PERF_RECORD_HEADER_BUILD_ID:
1590                 return tool->build_id(session, event);
1591         case PERF_RECORD_FINISHED_ROUND:
1592                 return tool->finished_round(tool, event, oe);
1593         case PERF_RECORD_ID_INDEX:
1594                 return tool->id_index(session, event);
1595         case PERF_RECORD_AUXTRACE_INFO:
1596                 return tool->auxtrace_info(session, event);
1597         case PERF_RECORD_AUXTRACE:
1598                 /* setup for reading amidst mmap */
1599                 lseek(fd, file_offset + event->header.size, SEEK_SET);
1600                 return tool->auxtrace(session, event);
1601         case PERF_RECORD_AUXTRACE_ERROR:
1602                 perf_session__auxtrace_error_inc(session, event);
1603                 return tool->auxtrace_error(session, event);
1604         case PERF_RECORD_THREAD_MAP:
1605                 return tool->thread_map(session, event);
1606         case PERF_RECORD_CPU_MAP:
1607                 return tool->cpu_map(session, event);
1608         case PERF_RECORD_STAT_CONFIG:
1609                 return tool->stat_config(session, event);
1610         case PERF_RECORD_STAT:
1611                 return tool->stat(session, event);
1612         case PERF_RECORD_STAT_ROUND:
1613                 return tool->stat_round(session, event);
1614         case PERF_RECORD_TIME_CONV:
1615                 session->time_conv = event->time_conv;
1616                 return tool->time_conv(session, event);
1617         case PERF_RECORD_HEADER_FEATURE:
1618                 return tool->feature(session, event);
1619         case PERF_RECORD_COMPRESSED:
1620                 err = tool->compressed(session, event, file_offset);
1621                 if (err)
1622                         dump_event(session->evlist, event, file_offset, &sample);
1623                 return err;
1624         default:
1625                 return -EINVAL;
1626         }
1627 }
1628
1629 int perf_session__deliver_synth_event(struct perf_session *session,
1630                                       union perf_event *event,
1631                                       struct perf_sample *sample)
1632 {
1633         struct evlist *evlist = session->evlist;
1634         struct perf_tool *tool = session->tool;
1635
1636         events_stats__inc(&evlist->stats, event->header.type);
1637
1638         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1639                 return perf_session__process_user_event(session, event, 0);
1640
1641         return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0);
1642 }
1643
1644 static void event_swap(union perf_event *event, bool sample_id_all)
1645 {
1646         perf_event__swap_op swap;
1647
1648         swap = perf_event__swap_ops[event->header.type];
1649         if (swap)
1650                 swap(event, sample_id_all);
1651 }
1652
1653 int perf_session__peek_event(struct perf_session *session, off_t file_offset,
1654                              void *buf, size_t buf_sz,
1655                              union perf_event **event_ptr,
1656                              struct perf_sample *sample)
1657 {
1658         union perf_event *event;
1659         size_t hdr_sz, rest;
1660         int fd;
1661
1662         if (session->one_mmap && !session->header.needs_swap) {
1663                 event = file_offset - session->one_mmap_offset +
1664                         session->one_mmap_addr;
1665                 goto out_parse_sample;
1666         }
1667
1668         if (perf_data__is_pipe(session->data))
1669                 return -1;
1670
1671         fd = perf_data__fd(session->data);
1672         hdr_sz = sizeof(struct perf_event_header);
1673
1674         if (buf_sz < hdr_sz)
1675                 return -1;
1676
1677         if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 ||
1678             readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz)
1679                 return -1;
1680
1681         event = (union perf_event *)buf;
1682
1683         if (session->header.needs_swap)
1684                 perf_event_header__bswap(&event->header);
1685
1686         if (event->header.size < hdr_sz || event->header.size > buf_sz)
1687                 return -1;
1688
1689         rest = event->header.size - hdr_sz;
1690
1691         if (readn(fd, buf, rest) != (ssize_t)rest)
1692                 return -1;
1693
1694         if (session->header.needs_swap)
1695                 event_swap(event, evlist__sample_id_all(session->evlist));
1696
1697 out_parse_sample:
1698
1699         if (sample && event->header.type < PERF_RECORD_USER_TYPE_START &&
1700             perf_evlist__parse_sample(session->evlist, event, sample))
1701                 return -1;
1702
1703         *event_ptr = event;
1704
1705         return 0;
1706 }
1707
1708 int perf_session__peek_events(struct perf_session *session, u64 offset,
1709                               u64 size, peek_events_cb_t cb, void *data)
1710 {
1711         u64 max_offset = offset + size;
1712         char buf[PERF_SAMPLE_MAX_SIZE];
1713         union perf_event *event;
1714         int err;
1715
1716         do {
1717                 err = perf_session__peek_event(session, offset, buf,
1718                                                PERF_SAMPLE_MAX_SIZE, &event,
1719                                                NULL);
1720                 if (err)
1721                         return err;
1722
1723                 err = cb(session, event, offset, data);
1724                 if (err)
1725                         return err;
1726
1727                 offset += event->header.size;
1728                 if (event->header.type == PERF_RECORD_AUXTRACE)
1729                         offset += event->auxtrace.size;
1730
1731         } while (offset < max_offset);
1732
1733         return err;
1734 }
1735
1736 static s64 perf_session__process_event(struct perf_session *session,
1737                                        union perf_event *event, u64 file_offset)
1738 {
1739         struct evlist *evlist = session->evlist;
1740         struct perf_tool *tool = session->tool;
1741         int ret;
1742
1743         if (session->header.needs_swap)
1744                 event_swap(event, evlist__sample_id_all(evlist));
1745
1746         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1747                 return -EINVAL;
1748
1749         events_stats__inc(&evlist->stats, event->header.type);
1750
1751         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1752                 return perf_session__process_user_event(session, event, file_offset);
1753
1754         if (tool->ordered_events) {
1755                 u64 timestamp = -1ULL;
1756
1757                 ret = perf_evlist__parse_sample_timestamp(evlist, event, &timestamp);
1758                 if (ret && ret != -1)
1759                         return ret;
1760
1761                 ret = perf_session__queue_event(session, event, timestamp, file_offset);
1762                 if (ret != -ETIME)
1763                         return ret;
1764         }
1765
1766         return perf_session__deliver_event(session, event, tool, file_offset);
1767 }
1768
1769 void perf_event_header__bswap(struct perf_event_header *hdr)
1770 {
1771         hdr->type = bswap_32(hdr->type);
1772         hdr->misc = bswap_16(hdr->misc);
1773         hdr->size = bswap_16(hdr->size);
1774 }
1775
1776 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1777 {
1778         return machine__findnew_thread(&session->machines.host, -1, pid);
1779 }
1780
1781 /*
1782  * Threads are identified by pid and tid, and the idle task has pid == tid == 0.
1783  * So here a single thread is created for that, but actually there is a separate
1784  * idle task per cpu, so there should be one 'struct thread' per cpu, but there
1785  * is only 1. That causes problems for some tools, requiring workarounds. For
1786  * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu().
1787  */
1788 int perf_session__register_idle_thread(struct perf_session *session)
1789 {
1790         struct thread *thread;
1791         int err = 0;
1792
1793         thread = machine__findnew_thread(&session->machines.host, 0, 0);
1794         if (thread == NULL || thread__set_comm(thread, "swapper", 0)) {
1795                 pr_err("problem inserting idle task.\n");
1796                 err = -1;
1797         }
1798
1799         if (thread == NULL || thread__set_namespaces(thread, 0, NULL)) {
1800                 pr_err("problem inserting idle task.\n");
1801                 err = -1;
1802         }
1803
1804         /* machine__findnew_thread() got the thread, so put it */
1805         thread__put(thread);
1806         return err;
1807 }
1808
1809 static void
1810 perf_session__warn_order(const struct perf_session *session)
1811 {
1812         const struct ordered_events *oe = &session->ordered_events;
1813         struct evsel *evsel;
1814         bool should_warn = true;
1815
1816         evlist__for_each_entry(session->evlist, evsel) {
1817                 if (evsel->core.attr.write_backward)
1818                         should_warn = false;
1819         }
1820
1821         if (!should_warn)
1822                 return;
1823         if (oe->nr_unordered_events != 0)
1824                 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events);
1825 }
1826
1827 static void perf_session__warn_about_errors(const struct perf_session *session)
1828 {
1829         const struct events_stats *stats = &session->evlist->stats;
1830
1831         if (session->tool->lost == perf_event__process_lost &&
1832             stats->nr_events[PERF_RECORD_LOST] != 0) {
1833                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1834                             "Check IO/CPU overload!\n\n",
1835                             stats->nr_events[0],
1836                             stats->nr_events[PERF_RECORD_LOST]);
1837         }
1838
1839         if (session->tool->lost_samples == perf_event__process_lost_samples) {
1840                 double drop_rate;
1841
1842                 drop_rate = (double)stats->total_lost_samples /
1843                             (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples);
1844                 if (drop_rate > 0.05) {
1845                         ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n",
1846                                     stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples,
1847                                     drop_rate * 100.0);
1848                 }
1849         }
1850
1851         if (session->tool->aux == perf_event__process_aux &&
1852             stats->total_aux_lost != 0) {
1853                 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n",
1854                             stats->total_aux_lost,
1855                             stats->nr_events[PERF_RECORD_AUX]);
1856         }
1857
1858         if (session->tool->aux == perf_event__process_aux &&
1859             stats->total_aux_partial != 0) {
1860                 bool vmm_exclusive = false;
1861
1862                 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive",
1863                                        &vmm_exclusive);
1864
1865                 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n"
1866                             "Are you running a KVM guest in the background?%s\n\n",
1867                             stats->total_aux_partial,
1868                             stats->nr_events[PERF_RECORD_AUX],
1869                             vmm_exclusive ?
1870                             "\nReloading kvm_intel module with vmm_exclusive=0\n"
1871                             "will reduce the gaps to only guest's timeslices." :
1872                             "");
1873         }
1874
1875         if (stats->nr_unknown_events != 0) {
1876                 ui__warning("Found %u unknown events!\n\n"
1877                             "Is this an older tool processing a perf.data "
1878                             "file generated by a more recent tool?\n\n"
1879                             "If that is not the case, consider "
1880                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1881                             stats->nr_unknown_events);
1882         }
1883
1884         if (stats->nr_unknown_id != 0) {
1885                 ui__warning("%u samples with id not present in the header\n",
1886                             stats->nr_unknown_id);
1887         }
1888
1889         if (stats->nr_invalid_chains != 0) {
1890                 ui__warning("Found invalid callchains!\n\n"
1891                             "%u out of %u events were discarded for this reason.\n\n"
1892                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1893                             stats->nr_invalid_chains,
1894                             stats->nr_events[PERF_RECORD_SAMPLE]);
1895         }
1896
1897         if (stats->nr_unprocessable_samples != 0) {
1898                 ui__warning("%u unprocessable samples recorded.\n"
1899                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1900                             stats->nr_unprocessable_samples);
1901         }
1902
1903         perf_session__warn_order(session);
1904
1905         events_stats__auxtrace_error_warn(stats);
1906
1907         if (stats->nr_proc_map_timeout != 0) {
1908                 ui__warning("%d map information files for pre-existing threads were\n"
1909                             "not processed, if there are samples for addresses they\n"
1910                             "will not be resolved, you may find out which are these\n"
1911                             "threads by running with -v and redirecting the output\n"
1912                             "to a file.\n"
1913                             "The time limit to process proc map is too short?\n"
1914                             "Increase it by --proc-map-timeout\n",
1915                             stats->nr_proc_map_timeout);
1916         }
1917 }
1918
1919 static int perf_session__flush_thread_stack(struct thread *thread,
1920                                             void *p __maybe_unused)
1921 {
1922         return thread_stack__flush(thread);
1923 }
1924
1925 static int perf_session__flush_thread_stacks(struct perf_session *session)
1926 {
1927         return machines__for_each_thread(&session->machines,
1928                                          perf_session__flush_thread_stack,
1929                                          NULL);
1930 }
1931
1932 volatile int session_done;
1933
1934 static int __perf_session__process_decomp_events(struct perf_session *session);
1935
1936 static int __perf_session__process_pipe_events(struct perf_session *session)
1937 {
1938         struct ordered_events *oe = &session->ordered_events;
1939         struct perf_tool *tool = session->tool;
1940         union perf_event *event;
1941         uint32_t size, cur_size = 0;
1942         void *buf = NULL;
1943         s64 skip = 0;
1944         u64 head;
1945         ssize_t err;
1946         void *p;
1947
1948         perf_tool__fill_defaults(tool);
1949
1950         head = 0;
1951         cur_size = sizeof(union perf_event);
1952
1953         buf = malloc(cur_size);
1954         if (!buf)
1955                 return -errno;
1956         ordered_events__set_copy_on_queue(oe, true);
1957 more:
1958         event = buf;
1959         err = perf_data__read(session->data, event,
1960                               sizeof(struct perf_event_header));
1961         if (err <= 0) {
1962                 if (err == 0)
1963                         goto done;
1964
1965                 pr_err("failed to read event header\n");
1966                 goto out_err;
1967         }
1968
1969         if (session->header.needs_swap)
1970                 perf_event_header__bswap(&event->header);
1971
1972         size = event->header.size;
1973         if (size < sizeof(struct perf_event_header)) {
1974                 pr_err("bad event header size\n");
1975                 goto out_err;
1976         }
1977
1978         if (size > cur_size) {
1979                 void *new = realloc(buf, size);
1980                 if (!new) {
1981                         pr_err("failed to allocate memory to read event\n");
1982                         goto out_err;
1983                 }
1984                 buf = new;
1985                 cur_size = size;
1986                 event = buf;
1987         }
1988         p = event;
1989         p += sizeof(struct perf_event_header);
1990
1991         if (size - sizeof(struct perf_event_header)) {
1992                 err = perf_data__read(session->data, p,
1993                                       size - sizeof(struct perf_event_header));
1994                 if (err <= 0) {
1995                         if (err == 0) {
1996                                 pr_err("unexpected end of event stream\n");
1997                                 goto done;
1998                         }
1999
2000                         pr_err("failed to read event data\n");
2001                         goto out_err;
2002                 }
2003         }
2004
2005         if ((skip = perf_session__process_event(session, event, head)) < 0) {
2006                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
2007                        head, event->header.size, event->header.type);
2008                 err = -EINVAL;
2009                 goto out_err;
2010         }
2011
2012         head += size;
2013
2014         if (skip > 0)
2015                 head += skip;
2016
2017         err = __perf_session__process_decomp_events(session);
2018         if (err)
2019                 goto out_err;
2020
2021         if (!session_done())
2022                 goto more;
2023 done:
2024         /* do the final flush for ordered samples */
2025         err = ordered_events__flush(oe, OE_FLUSH__FINAL);
2026         if (err)
2027                 goto out_err;
2028         err = auxtrace__flush_events(session, tool);
2029         if (err)
2030                 goto out_err;
2031         err = perf_session__flush_thread_stacks(session);
2032 out_err:
2033         free(buf);
2034         if (!tool->no_warn)
2035                 perf_session__warn_about_errors(session);
2036         ordered_events__free(&session->ordered_events);
2037         auxtrace__free_events(session);
2038         return err;
2039 }
2040
2041 static union perf_event *
2042 prefetch_event(char *buf, u64 head, size_t mmap_size,
2043                bool needs_swap, union perf_event *error)
2044 {
2045         union perf_event *event;
2046
2047         /*
2048          * Ensure we have enough space remaining to read
2049          * the size of the event in the headers.
2050          */
2051         if (head + sizeof(event->header) > mmap_size)
2052                 return NULL;
2053
2054         event = (union perf_event *)(buf + head);
2055         if (needs_swap)
2056                 perf_event_header__bswap(&event->header);
2057
2058         if (head + event->header.size <= mmap_size)
2059                 return event;
2060
2061         /* We're not fetching the event so swap back again */
2062         if (needs_swap)
2063                 perf_event_header__bswap(&event->header);
2064
2065         pr_debug("%s: head=%#" PRIx64 " event->header_size=%#x, mmap_size=%#zx:"
2066                  " fuzzed or compressed perf.data?\n",__func__, head, event->header.size, mmap_size);
2067
2068         return error;
2069 }
2070
2071 static union perf_event *
2072 fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap)
2073 {
2074         return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL));
2075 }
2076
2077 static union perf_event *
2078 fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap)
2079 {
2080         return prefetch_event(buf, head, mmap_size, needs_swap, NULL);
2081 }
2082
2083 static int __perf_session__process_decomp_events(struct perf_session *session)
2084 {
2085         s64 skip;
2086         u64 size, file_pos = 0;
2087         struct decomp *decomp = session->decomp_last;
2088
2089         if (!decomp)
2090                 return 0;
2091
2092         while (decomp->head < decomp->size && !session_done()) {
2093                 union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data,
2094                                                              session->header.needs_swap);
2095
2096                 if (!event)
2097                         break;
2098
2099                 size = event->header.size;
2100
2101                 if (size < sizeof(struct perf_event_header) ||
2102                     (skip = perf_session__process_event(session, event, file_pos)) < 0) {
2103                         pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
2104                                 decomp->file_pos + decomp->head, event->header.size, event->header.type);
2105                         return -EINVAL;
2106                 }
2107
2108                 if (skip)
2109                         size += skip;
2110
2111                 decomp->head += size;
2112         }
2113
2114         return 0;
2115 }
2116
2117 /*
2118  * On 64bit we can mmap the data file in one go. No need for tiny mmap
2119  * slices. On 32bit we use 32MB.
2120  */
2121 #if BITS_PER_LONG == 64
2122 #define MMAP_SIZE ULLONG_MAX
2123 #define NUM_MMAPS 1
2124 #else
2125 #define MMAP_SIZE (32 * 1024 * 1024ULL)
2126 #define NUM_MMAPS 128
2127 #endif
2128
2129 struct reader;
2130
2131 typedef s64 (*reader_cb_t)(struct perf_session *session,
2132                            union perf_event *event,
2133                            u64 file_offset);
2134
2135 struct reader {
2136         int              fd;
2137         u64              data_size;
2138         u64              data_offset;
2139         reader_cb_t      process;
2140 };
2141
2142 static int
2143 reader__process_events(struct reader *rd, struct perf_session *session,
2144                        struct ui_progress *prog)
2145 {
2146         u64 data_size = rd->data_size;
2147         u64 head, page_offset, file_offset, file_pos, size;
2148         int err = 0, mmap_prot, mmap_flags, map_idx = 0;
2149         size_t  mmap_size;
2150         char *buf, *mmaps[NUM_MMAPS];
2151         union perf_event *event;
2152         s64 skip;
2153
2154         page_offset = page_size * (rd->data_offset / page_size);
2155         file_offset = page_offset;
2156         head = rd->data_offset - page_offset;
2157
2158         ui_progress__init_size(prog, data_size, "Processing events...");
2159
2160         data_size += rd->data_offset;
2161
2162         mmap_size = MMAP_SIZE;
2163         if (mmap_size > data_size) {
2164                 mmap_size = data_size;
2165                 session->one_mmap = true;
2166         }
2167
2168         memset(mmaps, 0, sizeof(mmaps));
2169
2170         mmap_prot  = PROT_READ;
2171         mmap_flags = MAP_SHARED;
2172
2173         if (session->header.needs_swap) {
2174                 mmap_prot  |= PROT_WRITE;
2175                 mmap_flags = MAP_PRIVATE;
2176         }
2177 remap:
2178         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, rd->fd,
2179                    file_offset);
2180         if (buf == MAP_FAILED) {
2181                 pr_err("failed to mmap file\n");
2182                 err = -errno;
2183                 goto out;
2184         }
2185         mmaps[map_idx] = buf;
2186         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
2187         file_pos = file_offset + head;
2188         if (session->one_mmap) {
2189                 session->one_mmap_addr = buf;
2190                 session->one_mmap_offset = file_offset;
2191         }
2192
2193 more:
2194         event = fetch_mmaped_event(head, mmap_size, buf, session->header.needs_swap);
2195         if (IS_ERR(event))
2196                 return PTR_ERR(event);
2197
2198         if (!event) {
2199                 if (mmaps[map_idx]) {
2200                         munmap(mmaps[map_idx], mmap_size);
2201                         mmaps[map_idx] = NULL;
2202                 }
2203
2204                 page_offset = page_size * (head / page_size);
2205                 file_offset += page_offset;
2206                 head -= page_offset;
2207                 goto remap;
2208         }
2209
2210         size = event->header.size;
2211
2212         skip = -EINVAL;
2213
2214         if (size < sizeof(struct perf_event_header) ||
2215             (skip = rd->process(session, event, file_pos)) < 0) {
2216                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n",
2217                        file_offset + head, event->header.size,
2218                        event->header.type, strerror(-skip));
2219                 err = skip;
2220                 goto out;
2221         }
2222
2223         if (skip)
2224                 size += skip;
2225
2226         head += size;
2227         file_pos += size;
2228
2229         err = __perf_session__process_decomp_events(session);
2230         if (err)
2231                 goto out;
2232
2233         ui_progress__update(prog, size);
2234
2235         if (session_done())
2236                 goto out;
2237
2238         if (file_pos < data_size)
2239                 goto more;
2240
2241 out:
2242         return err;
2243 }
2244
2245 static s64 process_simple(struct perf_session *session,
2246                           union perf_event *event,
2247                           u64 file_offset)
2248 {
2249         return perf_session__process_event(session, event, file_offset);
2250 }
2251
2252 static int __perf_session__process_events(struct perf_session *session)
2253 {
2254         struct reader rd = {
2255                 .fd             = perf_data__fd(session->data),
2256                 .data_size      = session->header.data_size,
2257                 .data_offset    = session->header.data_offset,
2258                 .process        = process_simple,
2259         };
2260         struct ordered_events *oe = &session->ordered_events;
2261         struct perf_tool *tool = session->tool;
2262         struct ui_progress prog;
2263         int err;
2264
2265         perf_tool__fill_defaults(tool);
2266
2267         if (rd.data_size == 0)
2268                 return -1;
2269
2270         ui_progress__init_size(&prog, rd.data_size, "Processing events...");
2271
2272         err = reader__process_events(&rd, session, &prog);
2273         if (err)
2274                 goto out_err;
2275         /* do the final flush for ordered samples */
2276         err = ordered_events__flush(oe, OE_FLUSH__FINAL);
2277         if (err)
2278                 goto out_err;
2279         err = auxtrace__flush_events(session, tool);
2280         if (err)
2281                 goto out_err;
2282         err = perf_session__flush_thread_stacks(session);
2283 out_err:
2284         ui_progress__finish();
2285         if (!tool->no_warn)
2286                 perf_session__warn_about_errors(session);
2287         /*
2288          * We may switching perf.data output, make ordered_events
2289          * reusable.
2290          */
2291         ordered_events__reinit(&session->ordered_events);
2292         auxtrace__free_events(session);
2293         session->one_mmap = false;
2294         return err;
2295 }
2296
2297 int perf_session__process_events(struct perf_session *session)
2298 {
2299         if (perf_session__register_idle_thread(session) < 0)
2300                 return -ENOMEM;
2301
2302         if (perf_data__is_pipe(session->data))
2303                 return __perf_session__process_pipe_events(session);
2304
2305         return __perf_session__process_events(session);
2306 }
2307
2308 bool perf_session__has_traces(struct perf_session *session, const char *msg)
2309 {
2310         struct evsel *evsel;
2311
2312         evlist__for_each_entry(session->evlist, evsel) {
2313                 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT)
2314                         return true;
2315         }
2316
2317         pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
2318         return false;
2319 }
2320
2321 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr)
2322 {
2323         char *bracket;
2324         struct ref_reloc_sym *ref;
2325         struct kmap *kmap;
2326
2327         ref = zalloc(sizeof(struct ref_reloc_sym));
2328         if (ref == NULL)
2329                 return -ENOMEM;
2330
2331         ref->name = strdup(symbol_name);
2332         if (ref->name == NULL) {
2333                 free(ref);
2334                 return -ENOMEM;
2335         }
2336
2337         bracket = strchr(ref->name, ']');
2338         if (bracket)
2339                 *bracket = '\0';
2340
2341         ref->addr = addr;
2342
2343         kmap = map__kmap(map);
2344         if (kmap)
2345                 kmap->ref_reloc_sym = ref;
2346
2347         return 0;
2348 }
2349
2350 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
2351 {
2352         return machines__fprintf_dsos(&session->machines, fp);
2353 }
2354
2355 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
2356                                           bool (skip)(struct dso *dso, int parm), int parm)
2357 {
2358         return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
2359 }
2360
2361 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp)
2362 {
2363         size_t ret;
2364         const char *msg = "";
2365
2366         if (perf_header__has_feat(&session->header, HEADER_AUXTRACE))
2367                 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)";
2368
2369         ret = fprintf(fp, "\nAggregated stats:%s\n", msg);
2370
2371         ret += events_stats__fprintf(&session->evlist->stats, fp);
2372         return ret;
2373 }
2374
2375 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
2376 {
2377         /*
2378          * FIXME: Here we have to actually print all the machines in this
2379          * session, not just the host...
2380          */
2381         return machine__fprintf(&session->machines.host, fp);
2382 }
2383
2384 struct evsel *perf_session__find_first_evtype(struct perf_session *session,
2385                                               unsigned int type)
2386 {
2387         struct evsel *pos;
2388
2389         evlist__for_each_entry(session->evlist, pos) {
2390                 if (pos->core.attr.type == type)
2391                         return pos;
2392         }
2393         return NULL;
2394 }
2395
2396 int perf_session__cpu_bitmap(struct perf_session *session,
2397                              const char *cpu_list, unsigned long *cpu_bitmap)
2398 {
2399         int i, err = -1;
2400         struct perf_cpu_map *map;
2401         int nr_cpus = min(session->header.env.nr_cpus_online, MAX_NR_CPUS);
2402
2403         for (i = 0; i < PERF_TYPE_MAX; ++i) {
2404                 struct evsel *evsel;
2405
2406                 evsel = perf_session__find_first_evtype(session, i);
2407                 if (!evsel)
2408                         continue;
2409
2410                 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) {
2411                         pr_err("File does not contain CPU events. "
2412                                "Remove -C option to proceed.\n");
2413                         return -1;
2414                 }
2415         }
2416
2417         map = perf_cpu_map__new(cpu_list);
2418         if (map == NULL) {
2419                 pr_err("Invalid cpu_list\n");
2420                 return -1;
2421         }
2422
2423         for (i = 0; i < map->nr; i++) {
2424                 int cpu = map->map[i];
2425
2426                 if (cpu >= nr_cpus) {
2427                         pr_err("Requested CPU %d too large. "
2428                                "Consider raising MAX_NR_CPUS\n", cpu);
2429                         goto out_delete_map;
2430                 }
2431
2432                 set_bit(cpu, cpu_bitmap);
2433         }
2434
2435         err = 0;
2436
2437 out_delete_map:
2438         perf_cpu_map__put(map);
2439         return err;
2440 }
2441
2442 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
2443                                 bool full)
2444 {
2445         if (session == NULL || fp == NULL)
2446                 return;
2447
2448         fprintf(fp, "# ========\n");
2449         perf_header__fprintf_info(session, fp, full);
2450         fprintf(fp, "# ========\n#\n");
2451 }
2452
2453 int perf_event__process_id_index(struct perf_session *session,
2454                                  union perf_event *event)
2455 {
2456         struct evlist *evlist = session->evlist;
2457         struct perf_record_id_index *ie = &event->id_index;
2458         size_t i, nr, max_nr;
2459
2460         max_nr = (ie->header.size - sizeof(struct perf_record_id_index)) /
2461                  sizeof(struct id_index_entry);
2462         nr = ie->nr;
2463         if (nr > max_nr)
2464                 return -EINVAL;
2465
2466         if (dump_trace)
2467                 fprintf(stdout, " nr: %zu\n", nr);
2468
2469         for (i = 0; i < nr; i++) {
2470                 struct id_index_entry *e = &ie->entries[i];
2471                 struct perf_sample_id *sid;
2472
2473                 if (dump_trace) {
2474                         fprintf(stdout, " ... id: %"PRI_lu64, e->id);
2475                         fprintf(stdout, "  idx: %"PRI_lu64, e->idx);
2476                         fprintf(stdout, "  cpu: %"PRI_ld64, e->cpu);
2477                         fprintf(stdout, "  tid: %"PRI_ld64"\n", e->tid);
2478                 }
2479
2480                 sid = perf_evlist__id2sid(evlist, e->id);
2481                 if (!sid)
2482                         return -ENOENT;
2483                 sid->idx = e->idx;
2484                 sid->cpu = e->cpu;
2485                 sid->tid = e->tid;
2486         }
2487         return 0;
2488 }