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 "tsc.h"
  33 #include "ui/progress.h"
  34 #include "../perf.h"
  35 #include "arch/common.h"
  36 #include "units.h"
  37 #include <internal/lib.h>
  38
  39 #ifdef HAVE_ZSTD_SUPPORT
  40 static int perf_session__process_compressed_event(struct perf_session *session,
  41                                                   union perf_event *event, u64 file_offset)
  42 {
  43         void *src;
  44         size_t decomp_size, src_size;
  45         u64 decomp_last_rem = 0;
  46         size_t mmap_len, decomp_len = session->header.env.comp_mmap_len;
  47         struct decomp *decomp, *decomp_last = session->decomp_last;
  48
  49         if (decomp_last) {
  50                 decomp_last_rem = decomp_last->size - decomp_last->head;
  51                 decomp_len += decomp_last_rem;
  52         }
  53
  54         mmap_len = sizeof(struct decomp) + decomp_len;
  55         decomp = mmap(NULL, mmap_len, PROT_READ|PROT_WRITE,
  56                       MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
  57         if (decomp == MAP_FAILED) {
  58                 pr_err("Couldn't allocate memory for decompression\n");
  59                 return -1;
  60         }
  61
  62         decomp->file_pos = file_offset;
  63         decomp->mmap_len = mmap_len;
  64         decomp->head = 0;
  65
  66         if (decomp_last_rem) {
  67                 memcpy(decomp->data, &(decomp_last->data[decomp_last->head]), decomp_last_rem);
  68                 decomp->size = decomp_last_rem;
  69         }
  70
  71         src = (void *)event + sizeof(struct perf_record_compressed);
  72         src_size = event->pack.header.size - sizeof(struct perf_record_compressed);
  73
  74         decomp_size = zstd_decompress_stream(&(session->zstd_data), src, src_size,
  75                                 &(decomp->data[decomp_last_rem]), decomp_len - decomp_last_rem);
  76         if (!decomp_size) {
  77                 munmap(decomp, mmap_len);
  78                 pr_err("Couldn't decompress data\n");
  79                 return -1;
  80         }
  81
  82         decomp->size += decomp_size;
  83
  84         if (session->decomp == NULL) {
  85                 session->decomp = decomp;
  86                 session->decomp_last = decomp;
  87         } else {
  88                 session->decomp_last->next = decomp;
  89                 session->decomp_last = decomp;
  90         }
  91
  92         pr_debug("decomp (B): %zd to %zd\n", src_size, decomp_size);
  93
  94         return 0;
  95 }
  96 #else /* !HAVE_ZSTD_SUPPORT */
  97 #define perf_session__process_compressed_event perf_session__process_compressed_event_stub
  98 #endif
  99
 100 static int perf_session__deliver_event(struct perf_session *session,
 101                                        union perf_event *event,
 102                                        struct perf_tool *tool,
 103                                        u64 file_offset);
 104
 105 static int perf_session__open(struct perf_session *session)
 106 {
 107         struct perf_data *data = session->data;
 108
 109         if (perf_session__read_header(session) < 0) {
 110                 pr_err("incompatible file format (rerun with -v to learn more)\n");
 111                 return -1;
 112         }
 113
 114         if (perf_data__is_pipe(data))
 115                 return 0;
 116
 117         if (perf_header__has_feat(&session->header, HEADER_STAT))
 118                 return 0;
 119
 120         if (!evlist__valid_sample_type(session->evlist)) {
 121                 pr_err("non matching sample_type\n");
 122                 return -1;
 123         }
 124
 125         if (!evlist__valid_sample_id_all(session->evlist)) {
 126                 pr_err("non matching sample_id_all\n");
 127                 return -1;
 128         }
 129
 130         if (!evlist__valid_read_format(session->evlist)) {
 131                 pr_err("non matching read_format\n");
 132                 return -1;
 133         }
 134
 135         return 0;
 136 }
 137
 138 void perf_session__set_id_hdr_size(struct perf_session *session)
 139 {
 140         u16 id_hdr_size = evlist__id_hdr_size(session->evlist);
 141
 142         machines__set_id_hdr_size(&session->machines, id_hdr_size);
 143 }
 144
 145 int perf_session__create_kernel_maps(struct perf_session *session)
 146 {
 147         int ret = machine__create_kernel_maps(&session->machines.host);
 148
 149         if (ret >= 0)
 150                 ret = machines__create_guest_kernel_maps(&session->machines);
 151         return ret;
 152 }
 153
 154 static void perf_session__destroy_kernel_maps(struct perf_session *session)
 155 {
 156         machines__destroy_kernel_maps(&session->machines);
 157 }
 158
 159 static bool perf_session__has_comm_exec(struct perf_session *session)
 160 {
 161         struct evsel *evsel;
 162
 163         evlist__for_each_entry(session->evlist, evsel) {
 164                 if (evsel->core.attr.comm_exec)
 165                         return true;
 166         }
 167
 168         return false;
 169 }
 170
 171 static void perf_session__set_comm_exec(struct perf_session *session)
 172 {
 173         bool comm_exec = perf_session__has_comm_exec(session);
 174
 175         machines__set_comm_exec(&session->machines, comm_exec);
 176 }
 177
 178 static int ordered_events__deliver_event(struct ordered_events *oe,
 179                                          struct ordered_event *event)
 180 {
 181         struct perf_session *session = container_of(oe, struct perf_session,
 182                                                     ordered_events);
 183
 184         return perf_session__deliver_event(session, event->event,
 185                                            session->tool, event->file_offset);
 186 }
 187
 188 struct perf_session *perf_session__new(struct perf_data *data,
 189                                        bool repipe, struct perf_tool *tool)
 190 {
 191         int ret = -ENOMEM;
 192         struct perf_session *session = zalloc(sizeof(*session));
 193
 194         if (!session)
 195                 goto out;
 196
 197         session->repipe = repipe;
 198         session->tool   = tool;
 199         INIT_LIST_HEAD(&session->auxtrace_index);
 200         machines__init(&session->machines);
 201         ordered_events__init(&session->ordered_events,
 202                              ordered_events__deliver_event, NULL);
 203
 204         perf_env__init(&session->header.env);
 205         if (data) {
 206                 ret = perf_data__open(data);
 207                 if (ret < 0)
 208                         goto out_delete;
 209
 210                 session->data = data;
 211
 212                 if (perf_data__is_read(data)) {
 213                         ret = perf_session__open(session);
 214                         if (ret < 0)
 215                                 goto out_delete;
 216
 217                         /*
 218                          * set session attributes that are present in perf.data
 219                          * but not in pipe-mode.
 220                          */
 221                         if (!data->is_pipe) {
 222                                 perf_session__set_id_hdr_size(session);
 223                                 perf_session__set_comm_exec(session);
 224                         }
 225
 226                         evlist__init_trace_event_sample_raw(session->evlist);
 227
 228                         /* Open the directory data. */
 229                         if (data->is_dir) {
 230                                 ret = perf_data__open_dir(data);
 231                                 if (ret)
 232                                         goto out_delete;
 233                         }
 234
 235                         if (!symbol_conf.kallsyms_name &&
 236                             !symbol_conf.vmlinux_name)
 237                                 symbol_conf.kallsyms_name = perf_data__kallsyms_name(data);
 238                 }
 239         } else  {
 240                 session->machines.host.env = &perf_env;
 241         }
 242
 243         session->machines.host.single_address_space =
 244                 perf_env__single_address_space(session->machines.host.env);
 245
 246         if (!data || perf_data__is_write(data)) {
 247                 /*
 248                  * In O_RDONLY mode this will be performed when reading the
 249                  * kernel MMAP event, in perf_event__process_mmap().
 250                  */
 251                 if (perf_session__create_kernel_maps(session) < 0)
 252                         pr_warning("Cannot read kernel map\n");
 253         }
 254
 255         /*
 256          * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is
 257          * processed, so evlist__sample_id_all is not meaningful here.
 258          */
 259         if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps &&
 260             tool->ordered_events && !evlist__sample_id_all(session->evlist)) {
 261                 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n");
 262                 tool->ordered_events = false;
 263         }
 264
 265         return session;
 266
 267  out_delete:
 268         perf_session__delete(session);
 269  out:
 270         return ERR_PTR(ret);
 271 }
 272
 273 static void perf_session__delete_threads(struct perf_session *session)
 274 {
 275         machine__delete_threads(&session->machines.host);
 276 }
 277
 278 static void perf_session__release_decomp_events(struct perf_session *session)
 279 {
 280         struct decomp *next, *decomp;
 281         size_t mmap_len;
 282         next = session->decomp;
 283         do {
 284                 decomp = next;
 285                 if (decomp == NULL)
 286                         break;
 287                 next = decomp->next;
 288                 mmap_len = decomp->mmap_len;
 289                 munmap(decomp, mmap_len);
 290         } while (1);
 291 }
 292
 293 void perf_session__delete(struct perf_session *session)
 294 {
 295         if (session == NULL)
 296                 return;
 297         auxtrace__free(session);
 298         auxtrace_index__free(&session->auxtrace_index);
 299         perf_session__destroy_kernel_maps(session);
 300         perf_session__delete_threads(session);
 301         perf_session__release_decomp_events(session);
 302         perf_env__exit(&session->header.env);
 303         machines__exit(&session->machines);
 304         if (session->data) {
 305                 if (perf_data__is_read(session->data))
 306                         evlist__delete(session->evlist);
 307                 perf_data__close(session->data);
 308         }
 309         trace_event__cleanup(&session->tevent);
 310         free(session);
 311 }
 312
 313 static int process_event_synth_tracing_data_stub(struct perf_session *session
 314                                                  __maybe_unused,
 315                                                  union perf_event *event
 316                                                  __maybe_unused)
 317 {
 318         dump_printf(": unhandled!\n");
 319         return 0;
 320 }
 321
 322 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused,
 323                                          union perf_event *event __maybe_unused,
 324                                          struct evlist **pevlist
 325                                          __maybe_unused)
 326 {
 327         dump_printf(": unhandled!\n");
 328         return 0;
 329 }
 330
 331 static int process_event_synth_event_update_stub(struct perf_tool *tool __maybe_unused,
 332                                                  union perf_event *event __maybe_unused,
 333                                                  struct evlist **pevlist
 334                                                  __maybe_unused)
 335 {
 336         if (dump_trace)
 337                 perf_event__fprintf_event_update(event, stdout);
 338
 339         dump_printf(": unhandled!\n");
 340         return 0;
 341 }
 342
 343 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused,
 344                                      union perf_event *event __maybe_unused,
 345                                      struct perf_sample *sample __maybe_unused,
 346                                      struct evsel *evsel __maybe_unused,
 347                                      struct machine *machine __maybe_unused)
 348 {
 349         dump_printf(": unhandled!\n");
 350         return 0;
 351 }
 352
 353 static int process_event_stub(struct perf_tool *tool __maybe_unused,
 354                               union perf_event *event __maybe_unused,
 355                               struct perf_sample *sample __maybe_unused,
 356                               struct machine *machine __maybe_unused)
 357 {
 358         dump_printf(": unhandled!\n");
 359         return 0;
 360 }
 361
 362 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused,
 363                                        union perf_event *event __maybe_unused,
 364                                        struct ordered_events *oe __maybe_unused)
 365 {
 366         dump_printf(": unhandled!\n");
 367         return 0;
 368 }
 369
 370 static int process_finished_round(struct perf_tool *tool,
 371                                   union perf_event *event,
 372                                   struct ordered_events *oe);
 373
 374 static int skipn(int fd, off_t n)
 375 {
 376         char buf[4096];
 377         ssize_t ret;
 378
 379         while (n > 0) {
 380                 ret = read(fd, buf, min(n, (off_t)sizeof(buf)));
 381                 if (ret <= 0)
 382                         return ret;
 383                 n -= ret;
 384         }
 385
 386         return 0;
 387 }
 388
 389 static s64 process_event_auxtrace_stub(struct perf_session *session __maybe_unused,
 390                                        union perf_event *event)
 391 {
 392         dump_printf(": unhandled!\n");
 393         if (perf_data__is_pipe(session->data))
 394                 skipn(perf_data__fd(session->data), event->auxtrace.size);
 395         return event->auxtrace.size;
 396 }
 397
 398 static int process_event_op2_stub(struct perf_session *session __maybe_unused,
 399                                   union perf_event *event __maybe_unused)
 400 {
 401         dump_printf(": unhandled!\n");
 402         return 0;
 403 }
 404
 405
 406 static
 407 int process_event_thread_map_stub(struct perf_session *session __maybe_unused,
 408                                   union perf_event *event __maybe_unused)
 409 {
 410         if (dump_trace)
 411                 perf_event__fprintf_thread_map(event, stdout);
 412
 413         dump_printf(": unhandled!\n");
 414         return 0;
 415 }
 416
 417 static
 418 int process_event_cpu_map_stub(struct perf_session *session __maybe_unused,
 419                                union perf_event *event __maybe_unused)
 420 {
 421         if (dump_trace)
 422                 perf_event__fprintf_cpu_map(event, stdout);
 423
 424         dump_printf(": unhandled!\n");
 425         return 0;
 426 }
 427
 428 static
 429 int process_event_stat_config_stub(struct perf_session *session __maybe_unused,
 430                                    union perf_event *event __maybe_unused)
 431 {
 432         if (dump_trace)
 433                 perf_event__fprintf_stat_config(event, stdout);
 434
 435         dump_printf(": unhandled!\n");
 436         return 0;
 437 }
 438
 439 static int process_stat_stub(struct perf_session *perf_session __maybe_unused,
 440                              union perf_event *event)
 441 {
 442         if (dump_trace)
 443                 perf_event__fprintf_stat(event, stdout);
 444
 445         dump_printf(": unhandled!\n");
 446         return 0;
 447 }
 448
 449 static int process_stat_round_stub(struct perf_session *perf_session __maybe_unused,
 450                                    union perf_event *event)
 451 {
 452         if (dump_trace)
 453                 perf_event__fprintf_stat_round(event, stdout);
 454
 455         dump_printf(": unhandled!\n");
 456         return 0;
 457 }
 458
 459 static int process_event_time_conv_stub(struct perf_session *perf_session __maybe_unused,
 460                                         union perf_event *event)
 461 {
 462         if (dump_trace)
 463                 perf_event__fprintf_time_conv(event, stdout);
 464
 465         dump_printf(": unhandled!\n");
 466         return 0;
 467 }
 468
 469 static int perf_session__process_compressed_event_stub(struct perf_session *session __maybe_unused,
 470                                                        union perf_event *event __maybe_unused,
 471                                                        u64 file_offset __maybe_unused)
 472 {
 473        dump_printf(": unhandled!\n");
 474        return 0;
 475 }
 476
 477 void perf_tool__fill_defaults(struct perf_tool *tool)
 478 {
 479         if (tool->sample == NULL)
 480                 tool->sample = process_event_sample_stub;
 481         if (tool->mmap == NULL)
 482                 tool->mmap = process_event_stub;
 483         if (tool->mmap2 == NULL)
 484                 tool->mmap2 = process_event_stub;
 485         if (tool->comm == NULL)
 486                 tool->comm = process_event_stub;
 487         if (tool->namespaces == NULL)
 488                 tool->namespaces = process_event_stub;
 489         if (tool->cgroup == NULL)
 490                 tool->cgroup = process_event_stub;
 491         if (tool->fork == NULL)
 492                 tool->fork = process_event_stub;
 493         if (tool->exit == NULL)
 494                 tool->exit = process_event_stub;
 495         if (tool->lost == NULL)
 496                 tool->lost = perf_event__process_lost;
 497         if (tool->lost_samples == NULL)
 498                 tool->lost_samples = perf_event__process_lost_samples;
 499         if (tool->aux == NULL)
 500                 tool->aux = perf_event__process_aux;
 501         if (tool->itrace_start == NULL)
 502                 tool->itrace_start = perf_event__process_itrace_start;
 503         if (tool->context_switch == NULL)
 504                 tool->context_switch = perf_event__process_switch;
 505         if (tool->ksymbol == NULL)
 506                 tool->ksymbol = perf_event__process_ksymbol;
 507         if (tool->bpf == NULL)
 508                 tool->bpf = perf_event__process_bpf;
 509         if (tool->text_poke == NULL)
 510                 tool->text_poke = perf_event__process_text_poke;
 511         if (tool->read == NULL)
 512                 tool->read = process_event_sample_stub;
 513         if (tool->throttle == NULL)
 514                 tool->throttle = process_event_stub;
 515         if (tool->unthrottle == NULL)
 516                 tool->unthrottle = process_event_stub;
 517         if (tool->attr == NULL)
 518                 tool->attr = process_event_synth_attr_stub;
 519         if (tool->event_update == NULL)
 520                 tool->event_update = process_event_synth_event_update_stub;
 521         if (tool->tracing_data == NULL)
 522                 tool->tracing_data = process_event_synth_tracing_data_stub;
 523         if (tool->build_id == NULL)
 524                 tool->build_id = process_event_op2_stub;
 525         if (tool->finished_round == NULL) {
 526                 if (tool->ordered_events)
 527                         tool->finished_round = process_finished_round;
 528                 else
 529                         tool->finished_round = process_finished_round_stub;
 530         }
 531         if (tool->id_index == NULL)
 532                 tool->id_index = process_event_op2_stub;
 533         if (tool->auxtrace_info == NULL)
 534                 tool->auxtrace_info = process_event_op2_stub;
 535         if (tool->auxtrace == NULL)
 536                 tool->auxtrace = process_event_auxtrace_stub;
 537         if (tool->auxtrace_error == NULL)
 538                 tool->auxtrace_error = process_event_op2_stub;
 539         if (tool->thread_map == NULL)
 540                 tool->thread_map = process_event_thread_map_stub;
 541         if (tool->cpu_map == NULL)
 542                 tool->cpu_map = process_event_cpu_map_stub;
 543         if (tool->stat_config == NULL)
 544                 tool->stat_config = process_event_stat_config_stub;
 545         if (tool->stat == NULL)
 546                 tool->stat = process_stat_stub;
 547         if (tool->stat_round == NULL)
 548                 tool->stat_round = process_stat_round_stub;
 549         if (tool->time_conv == NULL)
 550                 tool->time_conv = process_event_time_conv_stub;
 551         if (tool->feature == NULL)
 552                 tool->feature = process_event_op2_stub;
 553         if (tool->compressed == NULL)
 554                 tool->compressed = perf_session__process_compressed_event;
 555 }
 556
 557 static void swap_sample_id_all(union perf_event *event, void *data)
 558 {
 559         void *end = (void *) event + event->header.size;
 560         int size = end - data;
 561
 562         BUG_ON(size % sizeof(u64));
 563         mem_bswap_64(data, size);
 564 }
 565
 566 static void perf_event__all64_swap(union perf_event *event,
 567                                    bool sample_id_all __maybe_unused)
 568 {
 569         struct perf_event_header *hdr = &event->header;
 570         mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr));
 571 }
 572
 573 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all)
 574 {
 575         event->comm.pid = bswap_32(event->comm.pid);
 576         event->comm.tid = bswap_32(event->comm.tid);
 577
 578         if (sample_id_all) {
 579                 void *data = &event->comm.comm;
 580
 581                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 582                 swap_sample_id_all(event, data);
 583         }
 584 }
 585
 586 static void perf_event__mmap_swap(union perf_event *event,
 587                                   bool sample_id_all)
 588 {
 589         event->mmap.pid   = bswap_32(event->mmap.pid);
 590         event->mmap.tid   = bswap_32(event->mmap.tid);
 591         event->mmap.start = bswap_64(event->mmap.start);
 592         event->mmap.len   = bswap_64(event->mmap.len);
 593         event->mmap.pgoff = bswap_64(event->mmap.pgoff);
 594
 595         if (sample_id_all) {
 596                 void *data = &event->mmap.filename;
 597
 598                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 599                 swap_sample_id_all(event, data);
 600         }
 601 }
 602
 603 static void perf_event__mmap2_swap(union perf_event *event,
 604                                   bool sample_id_all)
 605 {
 606         event->mmap2.pid   = bswap_32(event->mmap2.pid);
 607         event->mmap2.tid   = bswap_32(event->mmap2.tid);
 608         event->mmap2.start = bswap_64(event->mmap2.start);
 609         event->mmap2.len   = bswap_64(event->mmap2.len);
 610         event->mmap2.pgoff = bswap_64(event->mmap2.pgoff);
 611
 612         if (!(event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID)) {
 613                 event->mmap2.maj   = bswap_32(event->mmap2.maj);
 614                 event->mmap2.min   = bswap_32(event->mmap2.min);
 615                 event->mmap2.ino   = bswap_64(event->mmap2.ino);
 616                 event->mmap2.ino_generation = bswap_64(event->mmap2.ino_generation);
 617         }
 618
 619         if (sample_id_all) {
 620                 void *data = &event->mmap2.filename;
 621
 622                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 623                 swap_sample_id_all(event, data);
 624         }
 625 }
 626 static void perf_event__task_swap(union perf_event *event, bool sample_id_all)
 627 {
 628         event->fork.pid  = bswap_32(event->fork.pid);
 629         event->fork.tid  = bswap_32(event->fork.tid);
 630         event->fork.ppid = bswap_32(event->fork.ppid);
 631         event->fork.ptid = bswap_32(event->fork.ptid);
 632         event->fork.time = bswap_64(event->fork.time);
 633
 634         if (sample_id_all)
 635                 swap_sample_id_all(event, &event->fork + 1);
 636 }
 637
 638 static void perf_event__read_swap(union perf_event *event, bool sample_id_all)
 639 {
 640         event->read.pid          = bswap_32(event->read.pid);
 641         event->read.tid          = bswap_32(event->read.tid);
 642         event->read.value        = bswap_64(event->read.value);
 643         event->read.time_enabled = bswap_64(event->read.time_enabled);
 644         event->read.time_running = bswap_64(event->read.time_running);
 645         event->read.id           = bswap_64(event->read.id);
 646
 647         if (sample_id_all)
 648                 swap_sample_id_all(event, &event->read + 1);
 649 }
 650
 651 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all)
 652 {
 653         event->aux.aux_offset = bswap_64(event->aux.aux_offset);
 654         event->aux.aux_size   = bswap_64(event->aux.aux_size);
 655         event->aux.flags      = bswap_64(event->aux.flags);
 656
 657         if (sample_id_all)
 658                 swap_sample_id_all(event, &event->aux + 1);
 659 }
 660
 661 static void perf_event__itrace_start_swap(union perf_event *event,
 662                                           bool sample_id_all)
 663 {
 664         event->itrace_start.pid  = bswap_32(event->itrace_start.pid);
 665         event->itrace_start.tid  = bswap_32(event->itrace_start.tid);
 666
 667         if (sample_id_all)
 668                 swap_sample_id_all(event, &event->itrace_start + 1);
 669 }
 670
 671 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all)
 672 {
 673         if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) {
 674                 event->context_switch.next_prev_pid =
 675                                 bswap_32(event->context_switch.next_prev_pid);
 676                 event->context_switch.next_prev_tid =
 677                                 bswap_32(event->context_switch.next_prev_tid);
 678         }
 679
 680         if (sample_id_all)
 681                 swap_sample_id_all(event, &event->context_switch + 1);
 682 }
 683
 684 static void perf_event__text_poke_swap(union perf_event *event, bool sample_id_all)
 685 {
 686         event->text_poke.addr    = bswap_64(event->text_poke.addr);
 687         event->text_poke.old_len = bswap_16(event->text_poke.old_len);
 688         event->text_poke.new_len = bswap_16(event->text_poke.new_len);
 689
 690         if (sample_id_all) {
 691                 size_t len = sizeof(event->text_poke.old_len) +
 692                              sizeof(event->text_poke.new_len) +
 693                              event->text_poke.old_len +
 694                              event->text_poke.new_len;
 695                 void *data = &event->text_poke.old_len;
 696
 697                 data += PERF_ALIGN(len, sizeof(u64));
 698                 swap_sample_id_all(event, data);
 699         }
 700 }
 701
 702 static void perf_event__throttle_swap(union perf_event *event,
 703                                       bool sample_id_all)
 704 {
 705         event->throttle.time      = bswap_64(event->throttle.time);
 706         event->throttle.id        = bswap_64(event->throttle.id);
 707         event->throttle.stream_id = bswap_64(event->throttle.stream_id);
 708
 709         if (sample_id_all)
 710                 swap_sample_id_all(event, &event->throttle + 1);
 711 }
 712
 713 static void perf_event__namespaces_swap(union perf_event *event,
 714                                         bool sample_id_all)
 715 {
 716         u64 i;
 717
 718         event->namespaces.pid           = bswap_32(event->namespaces.pid);
 719         event->namespaces.tid           = bswap_32(event->namespaces.tid);
 720         event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces);
 721
 722         for (i = 0; i < event->namespaces.nr_namespaces; i++) {
 723                 struct perf_ns_link_info *ns = &event->namespaces.link_info[i];
 724
 725                 ns->dev = bswap_64(ns->dev);
 726                 ns->ino = bswap_64(ns->ino);
 727         }
 728
 729         if (sample_id_all)
 730                 swap_sample_id_all(event, &event->namespaces.link_info[i]);
 731 }
 732
 733 static void perf_event__cgroup_swap(union perf_event *event, bool sample_id_all)
 734 {
 735         event->cgroup.id = bswap_64(event->cgroup.id);
 736
 737         if (sample_id_all) {
 738                 void *data = &event->cgroup.path;
 739
 740                 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64));
 741                 swap_sample_id_all(event, data);
 742         }
 743 }
 744
 745 static u8 revbyte(u8 b)
 746 {
 747         int rev = (b >> 4) | ((b & 0xf) << 4);
 748         rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2);
 749         rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1);
 750         return (u8) rev;
 751 }
 752
 753 /*
 754  * XXX this is hack in attempt to carry flags bitfield
 755  * through endian village. ABI says:
 756  *
 757  * Bit-fields are allocated from right to left (least to most significant)
 758  * on little-endian implementations and from left to right (most to least
 759  * significant) on big-endian implementations.
 760  *
 761  * The above seems to be byte specific, so we need to reverse each
 762  * byte of the bitfield. 'Internet' also says this might be implementation
 763  * specific and we probably need proper fix and carry perf_event_attr
 764  * bitfield flags in separate data file FEAT_ section. Thought this seems
 765  * to work for now.
 766  */
 767 static void swap_bitfield(u8 *p, unsigned len)
 768 {
 769         unsigned i;
 770
 771         for (i = 0; i < len; i++) {
 772                 *p = revbyte(*p);
 773                 p++;
 774         }
 775 }
 776
 777 /* exported for swapping attributes in file header */
 778 void perf_event__attr_swap(struct perf_event_attr *attr)
 779 {
 780         attr->type              = bswap_32(attr->type);
 781         attr->size              = bswap_32(attr->size);
 782
 783 #define bswap_safe(f, n)                                        \
 784         (attr->size > (offsetof(struct perf_event_attr, f) +    \
 785                        sizeof(attr->f) * (n)))
 786 #define bswap_field(f, sz)                      \
 787 do {                                            \
 788         if (bswap_safe(f, 0))                   \
 789                 attr->f = bswap_##sz(attr->f);  \
 790 } while(0)
 791 #define bswap_field_16(f) bswap_field(f, 16)
 792 #define bswap_field_32(f) bswap_field(f, 32)
 793 #define bswap_field_64(f) bswap_field(f, 64)
 794
 795         bswap_field_64(config);
 796         bswap_field_64(sample_period);
 797         bswap_field_64(sample_type);
 798         bswap_field_64(read_format);
 799         bswap_field_32(wakeup_events);
 800         bswap_field_32(bp_type);
 801         bswap_field_64(bp_addr);
 802         bswap_field_64(bp_len);
 803         bswap_field_64(branch_sample_type);
 804         bswap_field_64(sample_regs_user);
 805         bswap_field_32(sample_stack_user);
 806         bswap_field_32(aux_watermark);
 807         bswap_field_16(sample_max_stack);
 808         bswap_field_32(aux_sample_size);
 809
 810         /*
 811          * After read_format are bitfields. Check read_format because
 812          * we are unable to use offsetof on bitfield.
 813          */
 814         if (bswap_safe(read_format, 1))
 815                 swap_bitfield((u8 *) (&attr->read_format + 1),
 816                               sizeof(u64));
 817 #undef bswap_field_64
 818 #undef bswap_field_32
 819 #undef bswap_field
 820 #undef bswap_safe
 821 }
 822
 823 static void perf_event__hdr_attr_swap(union perf_event *event,
 824                                       bool sample_id_all __maybe_unused)
 825 {
 826         size_t size;
 827
 828         perf_event__attr_swap(&event->attr.attr);
 829
 830         size = event->header.size;
 831         size -= (void *)&event->attr.id - (void *)event;
 832         mem_bswap_64(event->attr.id, size);
 833 }
 834
 835 static void perf_event__event_update_swap(union perf_event *event,
 836                                           bool sample_id_all __maybe_unused)
 837 {
 838         event->event_update.type = bswap_64(event->event_update.type);
 839         event->event_update.id   = bswap_64(event->event_update.id);
 840 }
 841
 842 static void perf_event__event_type_swap(union perf_event *event,
 843                                         bool sample_id_all __maybe_unused)
 844 {
 845         event->event_type.event_type.event_id =
 846                 bswap_64(event->event_type.event_type.event_id);
 847 }
 848
 849 static void perf_event__tracing_data_swap(union perf_event *event,
 850                                           bool sample_id_all __maybe_unused)
 851 {
 852         event->tracing_data.size = bswap_32(event->tracing_data.size);
 853 }
 854
 855 static void perf_event__auxtrace_info_swap(union perf_event *event,
 856                                            bool sample_id_all __maybe_unused)
 857 {
 858         size_t size;
 859
 860         event->auxtrace_info.type = bswap_32(event->auxtrace_info.type);
 861
 862         size = event->header.size;
 863         size -= (void *)&event->auxtrace_info.priv - (void *)event;
 864         mem_bswap_64(event->auxtrace_info.priv, size);
 865 }
 866
 867 static void perf_event__auxtrace_swap(union perf_event *event,
 868                                       bool sample_id_all __maybe_unused)
 869 {
 870         event->auxtrace.size      = bswap_64(event->auxtrace.size);
 871         event->auxtrace.offset    = bswap_64(event->auxtrace.offset);
 872         event->auxtrace.reference = bswap_64(event->auxtrace.reference);
 873         event->auxtrace.idx       = bswap_32(event->auxtrace.idx);
 874         event->auxtrace.tid       = bswap_32(event->auxtrace.tid);
 875         event->auxtrace.cpu       = bswap_32(event->auxtrace.cpu);
 876 }
 877
 878 static void perf_event__auxtrace_error_swap(union perf_event *event,
 879                                             bool sample_id_all __maybe_unused)
 880 {
 881         event->auxtrace_error.type = bswap_32(event->auxtrace_error.type);
 882         event->auxtrace_error.code = bswap_32(event->auxtrace_error.code);
 883         event->auxtrace_error.cpu  = bswap_32(event->auxtrace_error.cpu);
 884         event->auxtrace_error.pid  = bswap_32(event->auxtrace_error.pid);
 885         event->auxtrace_error.tid  = bswap_32(event->auxtrace_error.tid);
 886         event->auxtrace_error.fmt  = bswap_32(event->auxtrace_error.fmt);
 887         event->auxtrace_error.ip   = bswap_64(event->auxtrace_error.ip);
 888         if (event->auxtrace_error.fmt)
 889                 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time);
 890 }
 891
 892 static void perf_event__thread_map_swap(union perf_event *event,
 893                                         bool sample_id_all __maybe_unused)
 894 {
 895         unsigned i;
 896
 897         event->thread_map.nr = bswap_64(event->thread_map.nr);
 898
 899         for (i = 0; i < event->thread_map.nr; i++)
 900                 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid);
 901 }
 902
 903 static void perf_event__cpu_map_swap(union perf_event *event,
 904                                      bool sample_id_all __maybe_unused)
 905 {
 906         struct perf_record_cpu_map_data *data = &event->cpu_map.data;
 907         struct cpu_map_entries *cpus;
 908         struct perf_record_record_cpu_map *mask;
 909         unsigned i;
 910
 911         data->type = bswap_16(data->type);
 912
 913         switch (data->type) {
 914         case PERF_CPU_MAP__CPUS:
 915                 cpus = (struct cpu_map_entries *)data->data;
 916
 917                 cpus->nr = bswap_16(cpus->nr);
 918
 919                 for (i = 0; i < cpus->nr; i++)
 920                         cpus->cpu[i] = bswap_16(cpus->cpu[i]);
 921                 break;
 922         case PERF_CPU_MAP__MASK:
 923                 mask = (struct perf_record_record_cpu_map *)data->data;
 924
 925                 mask->nr = bswap_16(mask->nr);
 926                 mask->long_size = bswap_16(mask->long_size);
 927
 928                 switch (mask->long_size) {
 929                 case 4: mem_bswap_32(&mask->mask, mask->nr); break;
 930                 case 8: mem_bswap_64(&mask->mask, mask->nr); break;
 931                 default:
 932                         pr_err("cpu_map swap: unsupported long size\n");
 933                 }
 934         default:
 935                 break;
 936         }
 937 }
 938
 939 static void perf_event__stat_config_swap(union perf_event *event,
 940                                          bool sample_id_all __maybe_unused)
 941 {
 942         u64 size;
 943
 944         size  = bswap_64(event->stat_config.nr) * sizeof(event->stat_config.data[0]);
 945         size += 1; /* nr item itself */
 946         mem_bswap_64(&event->stat_config.nr, size);
 947 }
 948
 949 static void perf_event__stat_swap(union perf_event *event,
 950                                   bool sample_id_all __maybe_unused)
 951 {
 952         event->stat.id     = bswap_64(event->stat.id);
 953         event->stat.thread = bswap_32(event->stat.thread);
 954         event->stat.cpu    = bswap_32(event->stat.cpu);
 955         event->stat.val    = bswap_64(event->stat.val);
 956         event->stat.ena    = bswap_64(event->stat.ena);
 957         event->stat.run    = bswap_64(event->stat.run);
 958 }
 959
 960 static void perf_event__stat_round_swap(union perf_event *event,
 961                                         bool sample_id_all __maybe_unused)
 962 {
 963         event->stat_round.type = bswap_64(event->stat_round.type);
 964         event->stat_round.time = bswap_64(event->stat_round.time);
 965 }
 966
 967 static void perf_event__time_conv_swap(union perf_event *event,
 968                                        bool sample_id_all __maybe_unused)
 969 {
 970         event->time_conv.time_shift = bswap_64(event->time_conv.time_shift);
 971         event->time_conv.time_mult  = bswap_64(event->time_conv.time_mult);
 972         event->time_conv.time_zero  = bswap_64(event->time_conv.time_zero);
 973
 974         if (event_contains(event->time_conv, time_cycles)) {
 975                 event->time_conv.time_cycles = bswap_64(event->time_conv.time_cycles);
 976                 event->time_conv.time_mask = bswap_64(event->time_conv.time_mask);
 977         }
 978 }
 979
 980 typedef void (*perf_event__swap_op)(union perf_event *event,
 981                                     bool sample_id_all);
 982
 983 static perf_event__swap_op perf_event__swap_ops[] = {
 984         [PERF_RECORD_MMAP]                = perf_event__mmap_swap,
 985         [PERF_RECORD_MMAP2]               = perf_event__mmap2_swap,
 986         [PERF_RECORD_COMM]                = perf_event__comm_swap,
 987         [PERF_RECORD_FORK]                = perf_event__task_swap,
 988         [PERF_RECORD_EXIT]                = perf_event__task_swap,
 989         [PERF_RECORD_LOST]                = perf_event__all64_swap,
 990         [PERF_RECORD_READ]                = perf_event__read_swap,
 991         [PERF_RECORD_THROTTLE]            = perf_event__throttle_swap,
 992         [PERF_RECORD_UNTHROTTLE]          = perf_event__throttle_swap,
 993         [PERF_RECORD_SAMPLE]              = perf_event__all64_swap,
 994         [PERF_RECORD_AUX]                 = perf_event__aux_swap,
 995         [PERF_RECORD_ITRACE_START]        = perf_event__itrace_start_swap,
 996         [PERF_RECORD_LOST_SAMPLES]        = perf_event__all64_swap,
 997         [PERF_RECORD_SWITCH]              = perf_event__switch_swap,
 998         [PERF_RECORD_SWITCH_CPU_WIDE]     = perf_event__switch_swap,
 999         [PERF_RECORD_NAMESPACES]          = perf_event__namespaces_swap,
1000         [PERF_RECORD_CGROUP]              = perf_event__cgroup_swap,
1001         [PERF_RECORD_TEXT_POKE]           = perf_event__text_poke_swap,
1002         [PERF_RECORD_HEADER_ATTR]         = perf_event__hdr_attr_swap,
1003         [PERF_RECORD_HEADER_EVENT_TYPE]   = perf_event__event_type_swap,
1004         [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap,
1005         [PERF_RECORD_HEADER_BUILD_ID]     = NULL,
1006         [PERF_RECORD_ID_INDEX]            = perf_event__all64_swap,
1007         [PERF_RECORD_AUXTRACE_INFO]       = perf_event__auxtrace_info_swap,
1008         [PERF_RECORD_AUXTRACE]            = perf_event__auxtrace_swap,
1009         [PERF_RECORD_AUXTRACE_ERROR]      = perf_event__auxtrace_error_swap,
1010         [PERF_RECORD_THREAD_MAP]          = perf_event__thread_map_swap,
1011         [PERF_RECORD_CPU_MAP]             = perf_event__cpu_map_swap,
1012         [PERF_RECORD_STAT_CONFIG]         = perf_event__stat_config_swap,
1013         [PERF_RECORD_STAT]                = perf_event__stat_swap,
1014         [PERF_RECORD_STAT_ROUND]          = perf_event__stat_round_swap,
1015         [PERF_RECORD_EVENT_UPDATE]        = perf_event__event_update_swap,
1016         [PERF_RECORD_TIME_CONV]           = perf_event__time_conv_swap,
1017         [PERF_RECORD_HEADER_MAX]          = NULL,
1018 };
1019
1020 /*
1021  * When perf record finishes a pass on every buffers, it records this pseudo
1022  * event.
1023  * We record the max timestamp t found in the pass n.
1024  * Assuming these timestamps are monotonic across cpus, we know that if
1025  * a buffer still has events with timestamps below t, they will be all
1026  * available and then read in the pass n + 1.
1027  * Hence when we start to read the pass n + 2, we can safely flush every
1028  * events with timestamps below t.
1029  *
1030  *    ============ PASS n =================
1031  *       CPU 0         |   CPU 1
1032  *                     |
1033  *    cnt1 timestamps  |   cnt2 timestamps
1034  *          1          |         2
1035  *          2          |         3
1036  *          -          |         4  <--- max recorded
1037  *
1038  *    ============ PASS n + 1 ==============
1039  *       CPU 0         |   CPU 1
1040  *                     |
1041  *    cnt1 timestamps  |   cnt2 timestamps
1042  *          3          |         5
1043  *          4          |         6
1044  *          5          |         7 <---- max recorded
1045  *
1046  *      Flush every events below timestamp 4
1047  *
1048  *    ============ PASS n + 2 ==============
1049  *       CPU 0         |   CPU 1
1050  *                     |
1051  *    cnt1 timestamps  |   cnt2 timestamps
1052  *          6          |         8
1053  *          7          |         9
1054  *          -          |         10
1055  *
1056  *      Flush every events below timestamp 7
1057  *      etc...
1058  */
1059 static int process_finished_round(struct perf_tool *tool __maybe_unused,
1060                                   union perf_event *event __maybe_unused,
1061                                   struct ordered_events *oe)
1062 {
1063         if (dump_trace)
1064                 fprintf(stdout, "\n");
1065         return ordered_events__flush(oe, OE_FLUSH__ROUND);
1066 }
1067
1068 int perf_session__queue_event(struct perf_session *s, union perf_event *event,
1069                               u64 timestamp, u64 file_offset)
1070 {
1071         return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset);
1072 }
1073
1074 static void callchain__lbr_callstack_printf(struct perf_sample *sample)
1075 {
1076         struct ip_callchain *callchain = sample->callchain;
1077         struct branch_stack *lbr_stack = sample->branch_stack;
1078         struct branch_entry *entries = perf_sample__branch_entries(sample);
1079         u64 kernel_callchain_nr = callchain->nr;
1080         unsigned int i;
1081
1082         for (i = 0; i < kernel_callchain_nr; i++) {
1083                 if (callchain->ips[i] == PERF_CONTEXT_USER)
1084                         break;
1085         }
1086
1087         if ((i != kernel_callchain_nr) && lbr_stack->nr) {
1088                 u64 total_nr;
1089                 /*
1090                  * LBR callstack can only get user call chain,
1091                  * i is kernel call chain number,
1092                  * 1 is PERF_CONTEXT_USER.
1093                  *
1094                  * The user call chain is stored in LBR registers.
1095                  * LBR are pair registers. The caller is stored
1096                  * in "from" register, while the callee is stored
1097                  * in "to" register.
1098                  * For example, there is a call stack
1099                  * "A"->"B"->"C"->"D".
1100                  * The LBR registers will be recorded like
1101                  * "C"->"D", "B"->"C", "A"->"B".
1102                  * So only the first "to" register and all "from"
1103                  * registers are needed to construct the whole stack.
1104                  */
1105                 total_nr = i + 1 + lbr_stack->nr + 1;
1106                 kernel_callchain_nr = i + 1;
1107
1108                 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr);
1109
1110                 for (i = 0; i < kernel_callchain_nr; i++)
1111                         printf("..... %2d: %016" PRIx64 "\n",
1112                                i, callchain->ips[i]);
1113
1114                 printf("..... %2d: %016" PRIx64 "\n",
1115                        (int)(kernel_callchain_nr), entries[0].to);
1116                 for (i = 0; i < lbr_stack->nr; i++)
1117                         printf("..... %2d: %016" PRIx64 "\n",
1118                                (int)(i + kernel_callchain_nr + 1), entries[i].from);
1119         }
1120 }
1121
1122 static void callchain__printf(struct evsel *evsel,
1123                               struct perf_sample *sample)
1124 {
1125         unsigned int i;
1126         struct ip_callchain *callchain = sample->callchain;
1127
1128         if (evsel__has_branch_callstack(evsel))
1129                 callchain__lbr_callstack_printf(sample);
1130
1131         printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr);
1132
1133         for (i = 0; i < callchain->nr; i++)
1134                 printf("..... %2d: %016" PRIx64 "\n",
1135                        i, callchain->ips[i]);
1136 }
1137
1138 static void branch_stack__printf(struct perf_sample *sample, bool callstack)
1139 {
1140         struct branch_entry *entries = perf_sample__branch_entries(sample);
1141         uint64_t i;
1142
1143         printf("%s: nr:%" PRIu64 "\n",
1144                 !callstack ? "... branch stack" : "... branch callstack",
1145                 sample->branch_stack->nr);
1146
1147         for (i = 0; i < sample->branch_stack->nr; i++) {
1148                 struct branch_entry *e = &entries[i];
1149
1150                 if (!callstack) {
1151                         printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x\n",
1152                                 i, e->from, e->to,
1153                                 (unsigned short)e->flags.cycles,
1154                                 e->flags.mispred ? "M" : " ",
1155                                 e->flags.predicted ? "P" : " ",
1156                                 e->flags.abort ? "A" : " ",
1157                                 e->flags.in_tx ? "T" : " ",
1158                                 (unsigned)e->flags.reserved);
1159                 } else {
1160                         printf("..... %2"PRIu64": %016" PRIx64 "\n",
1161                                 i, i > 0 ? e->from : e->to);
1162                 }
1163         }
1164 }
1165
1166 static void regs_dump__printf(u64 mask, u64 *regs)
1167 {
1168         unsigned rid, i = 0;
1169
1170         for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) {
1171                 u64 val = regs[i++];
1172
1173                 printf(".... %-5s 0x%016" PRIx64 "\n",
1174                        perf_reg_name(rid), val);
1175         }
1176 }
1177
1178 static const char *regs_abi[] = {
1179         [PERF_SAMPLE_REGS_ABI_NONE] = "none",
1180         [PERF_SAMPLE_REGS_ABI_32] = "32-bit",
1181         [PERF_SAMPLE_REGS_ABI_64] = "64-bit",
1182 };
1183
1184 static inline const char *regs_dump_abi(struct regs_dump *d)
1185 {
1186         if (d->abi > PERF_SAMPLE_REGS_ABI_64)
1187                 return "unknown";
1188
1189         return regs_abi[d->abi];
1190 }
1191
1192 static void regs__printf(const char *type, struct regs_dump *regs)
1193 {
1194         u64 mask = regs->mask;
1195
1196         printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n",
1197                type,
1198                mask,
1199                regs_dump_abi(regs));
1200
1201         regs_dump__printf(mask, regs->regs);
1202 }
1203
1204 static void regs_user__printf(struct perf_sample *sample)
1205 {
1206         struct regs_dump *user_regs = &sample->user_regs;
1207
1208         if (user_regs->regs)
1209                 regs__printf("user", user_regs);
1210 }
1211
1212 static void regs_intr__printf(struct perf_sample *sample)
1213 {
1214         struct regs_dump *intr_regs = &sample->intr_regs;
1215
1216         if (intr_regs->regs)
1217                 regs__printf("intr", intr_regs);
1218 }
1219
1220 static void stack_user__printf(struct stack_dump *dump)
1221 {
1222         printf("... ustack: size %" PRIu64 ", offset 0x%x\n",
1223                dump->size, dump->offset);
1224 }
1225
1226 static void evlist__print_tstamp(struct evlist *evlist, union perf_event *event, struct perf_sample *sample)
1227 {
1228         u64 sample_type = __evlist__combined_sample_type(evlist);
1229
1230         if (event->header.type != PERF_RECORD_SAMPLE &&
1231             !evlist__sample_id_all(evlist)) {
1232                 fputs("-1 -1 ", stdout);
1233                 return;
1234         }
1235
1236         if ((sample_type & PERF_SAMPLE_CPU))
1237                 printf("%u ", sample->cpu);
1238
1239         if (sample_type & PERF_SAMPLE_TIME)
1240                 printf("%" PRIu64 " ", sample->time);
1241 }
1242
1243 static void sample_read__printf(struct perf_sample *sample, u64 read_format)
1244 {
1245         printf("... sample_read:\n");
1246
1247         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1248                 printf("...... time enabled %016" PRIx64 "\n",
1249                        sample->read.time_enabled);
1250
1251         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1252                 printf("...... time running %016" PRIx64 "\n",
1253                        sample->read.time_running);
1254
1255         if (read_format & PERF_FORMAT_GROUP) {
1256                 u64 i;
1257
1258                 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr);
1259
1260                 for (i = 0; i < sample->read.group.nr; i++) {
1261                         struct sample_read_value *value;
1262
1263                         value = &sample->read.group.values[i];
1264                         printf("..... id %016" PRIx64
1265                                ", value %016" PRIx64 "\n",
1266                                value->id, value->value);
1267                 }
1268         } else
1269                 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n",
1270                         sample->read.one.id, sample->read.one.value);
1271 }
1272
1273 static void dump_event(struct evlist *evlist, union perf_event *event,
1274                        u64 file_offset, struct perf_sample *sample)
1275 {
1276         if (!dump_trace)
1277                 return;
1278
1279         printf("\n%#" PRIx64 " [%#x]: event: %d\n",
1280                file_offset, event->header.size, event->header.type);
1281
1282         trace_event(event);
1283         if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw)
1284                 evlist->trace_event_sample_raw(evlist, event, sample);
1285
1286         if (sample)
1287                 evlist__print_tstamp(evlist, event, sample);
1288
1289         printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset,
1290                event->header.size, perf_event__name(event->header.type));
1291 }
1292
1293 char *get_page_size_name(u64 size, char *str)
1294 {
1295         if (!size || !unit_number__scnprintf(str, PAGE_SIZE_NAME_LEN, size))
1296                 snprintf(str, PAGE_SIZE_NAME_LEN, "%s", "N/A");
1297
1298         return str;
1299 }
1300
1301 static void dump_sample(struct evsel *evsel, union perf_event *event,
1302                         struct perf_sample *sample)
1303 {
1304         u64 sample_type;
1305         char str[PAGE_SIZE_NAME_LEN];
1306
1307         if (!dump_trace)
1308                 return;
1309
1310         printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n",
1311                event->header.misc, sample->pid, sample->tid, sample->ip,
1312                sample->period, sample->addr);
1313
1314         sample_type = evsel->core.attr.sample_type;
1315
1316         if (evsel__has_callchain(evsel))
1317                 callchain__printf(evsel, sample);
1318
1319         if (evsel__has_br_stack(evsel))
1320                 branch_stack__printf(sample, evsel__has_branch_callstack(evsel));
1321
1322         if (sample_type & PERF_SAMPLE_REGS_USER)
1323                 regs_user__printf(sample);
1324
1325         if (sample_type & PERF_SAMPLE_REGS_INTR)
1326                 regs_intr__printf(sample);
1327
1328         if (sample_type & PERF_SAMPLE_STACK_USER)
1329                 stack_user__printf(&sample->user_stack);
1330
1331         if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) {
1332                 printf("... weight: %" PRIu64 "", sample->weight);
1333                         if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) {
1334                                 printf(",0x%"PRIx16"", sample->ins_lat);
1335                                 printf(",0x%"PRIx16"", sample->p_stage_cyc);
1336                         }
1337                 printf("\n");
1338         }
1339
1340         if (sample_type & PERF_SAMPLE_DATA_SRC)
1341                 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src);
1342
1343         if (sample_type & PERF_SAMPLE_PHYS_ADDR)
1344                 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr);
1345
1346         if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE)
1347                 printf(" .. data page size: %s\n", get_page_size_name(sample->data_page_size, str));
1348
1349         if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE)
1350                 printf(" .. code page size: %s\n", get_page_size_name(sample->code_page_size, str));
1351
1352         if (sample_type & PERF_SAMPLE_TRANSACTION)
1353                 printf("... transaction: %" PRIx64 "\n", sample->transaction);
1354
1355         if (sample_type & PERF_SAMPLE_READ)
1356                 sample_read__printf(sample, evsel->core.attr.read_format);
1357 }
1358
1359 static void dump_read(struct evsel *evsel, union perf_event *event)
1360 {
1361         struct perf_record_read *read_event = &event->read;
1362         u64 read_format;
1363
1364         if (!dump_trace)
1365                 return;
1366
1367         printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid,
1368                evsel__name(evsel), event->read.value);
1369
1370         if (!evsel)
1371                 return;
1372
1373         read_format = evsel->core.attr.read_format;
1374
1375         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1376                 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled);
1377
1378         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1379                 printf("... time running : %" PRI_lu64 "\n", read_event->time_running);
1380
1381         if (read_format & PERF_FORMAT_ID)
1382                 printf("... id           : %" PRI_lu64 "\n", read_event->id);
1383 }
1384
1385 static struct machine *machines__find_for_cpumode(struct machines *machines,
1386                                                union perf_event *event,
1387                                                struct perf_sample *sample)
1388 {
1389         if (perf_guest &&
1390             ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) ||
1391              (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) {
1392                 u32 pid;
1393
1394                 if (event->header.type == PERF_RECORD_MMAP
1395                     || event->header.type == PERF_RECORD_MMAP2)
1396                         pid = event->mmap.pid;
1397                 else
1398                         pid = sample->pid;
1399
1400                 return machines__find_guest(machines, pid);
1401         }
1402
1403         return &machines->host;
1404 }
1405
1406 static int deliver_sample_value(struct evlist *evlist,
1407                                 struct perf_tool *tool,
1408                                 union perf_event *event,
1409                                 struct perf_sample *sample,
1410                                 struct sample_read_value *v,
1411                                 struct machine *machine)
1412 {
1413         struct perf_sample_id *sid = evlist__id2sid(evlist, v->id);
1414         struct evsel *evsel;
1415
1416         if (sid) {
1417                 sample->id     = v->id;
1418                 sample->period = v->value - sid->period;
1419                 sid->period    = v->value;
1420         }
1421
1422         if (!sid || sid->evsel == NULL) {
1423                 ++evlist->stats.nr_unknown_id;
1424                 return 0;
1425         }
1426
1427         /*
1428          * There's no reason to deliver sample
1429          * for zero period, bail out.
1430          */
1431         if (!sample->period)
1432                 return 0;
1433
1434         evsel = container_of(sid->evsel, struct evsel, core);
1435         return tool->sample(tool, event, sample, evsel, machine);
1436 }
1437
1438 static int deliver_sample_group(struct evlist *evlist,
1439                                 struct perf_tool *tool,
1440                                 union  perf_event *event,
1441                                 struct perf_sample *sample,
1442                                 struct machine *machine)
1443 {
1444         int ret = -EINVAL;
1445         u64 i;
1446
1447         for (i = 0; i < sample->read.group.nr; i++) {
1448                 ret = deliver_sample_value(evlist, tool, event, sample,
1449                                            &sample->read.group.values[i],
1450                                            machine);
1451                 if (ret)
1452                         break;
1453         }
1454
1455         return ret;
1456 }
1457
1458 static int evlist__deliver_sample(struct evlist *evlist, struct perf_tool *tool,
1459                                   union  perf_event *event, struct perf_sample *sample,
1460                                   struct evsel *evsel, struct machine *machine)
1461 {
1462         /* We know evsel != NULL. */
1463         u64 sample_type = evsel->core.attr.sample_type;
1464         u64 read_format = evsel->core.attr.read_format;
1465
1466         /* Standard sample delivery. */
1467         if (!(sample_type & PERF_SAMPLE_READ))
1468                 return tool->sample(tool, event, sample, evsel, machine);
1469
1470         /* For PERF_SAMPLE_READ we have either single or group mode. */
1471         if (read_format & PERF_FORMAT_GROUP)
1472                 return deliver_sample_group(evlist, tool, event, sample,
1473                                             machine);
1474         else
1475                 return deliver_sample_value(evlist, tool, event, sample,
1476                                             &sample->read.one, machine);
1477 }
1478
1479 static int machines__deliver_event(struct machines *machines,
1480                                    struct evlist *evlist,
1481                                    union perf_event *event,
1482                                    struct perf_sample *sample,
1483                                    struct perf_tool *tool, u64 file_offset)
1484 {
1485         struct evsel *evsel;
1486         struct machine *machine;
1487
1488         dump_event(evlist, event, file_offset, sample);
1489
1490         evsel = evlist__id2evsel(evlist, sample->id);
1491
1492         machine = machines__find_for_cpumode(machines, event, sample);
1493
1494         switch (event->header.type) {
1495         case PERF_RECORD_SAMPLE:
1496                 if (evsel == NULL) {
1497                         ++evlist->stats.nr_unknown_id;
1498                         return 0;
1499                 }
1500                 dump_sample(evsel, event, sample);
1501                 if (machine == NULL) {
1502                         ++evlist->stats.nr_unprocessable_samples;
1503                         return 0;
1504                 }
1505                 return evlist__deliver_sample(evlist, tool, event, sample, evsel, machine);
1506         case PERF_RECORD_MMAP:
1507                 return tool->mmap(tool, event, sample, machine);
1508         case PERF_RECORD_MMAP2:
1509                 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT)
1510                         ++evlist->stats.nr_proc_map_timeout;
1511                 return tool->mmap2(tool, event, sample, machine);
1512         case PERF_RECORD_COMM:
1513                 return tool->comm(tool, event, sample, machine);
1514         case PERF_RECORD_NAMESPACES:
1515                 return tool->namespaces(tool, event, sample, machine);
1516         case PERF_RECORD_CGROUP:
1517                 return tool->cgroup(tool, event, sample, machine);
1518         case PERF_RECORD_FORK:
1519                 return tool->fork(tool, event, sample, machine);
1520         case PERF_RECORD_EXIT:
1521                 return tool->exit(tool, event, sample, machine);
1522         case PERF_RECORD_LOST:
1523                 if (tool->lost == perf_event__process_lost)
1524                         evlist->stats.total_lost += event->lost.lost;
1525                 return tool->lost(tool, event, sample, machine);
1526         case PERF_RECORD_LOST_SAMPLES:
1527                 if (tool->lost_samples == perf_event__process_lost_samples)
1528                         evlist->stats.total_lost_samples += event->lost_samples.lost;
1529                 return tool->lost_samples(tool, event, sample, machine);
1530         case PERF_RECORD_READ:
1531                 dump_read(evsel, event);
1532                 return tool->read(tool, event, sample, evsel, machine);
1533         case PERF_RECORD_THROTTLE:
1534                 return tool->throttle(tool, event, sample, machine);
1535         case PERF_RECORD_UNTHROTTLE:
1536                 return tool->unthrottle(tool, event, sample, machine);
1537         case PERF_RECORD_AUX:
1538                 if (tool->aux == perf_event__process_aux) {
1539                         if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED)
1540                                 evlist->stats.total_aux_lost += 1;
1541                         if (event->aux.flags & PERF_AUX_FLAG_PARTIAL)
1542                                 evlist->stats.total_aux_partial += 1;
1543                 }
1544                 return tool->aux(tool, event, sample, machine);
1545         case PERF_RECORD_ITRACE_START:
1546                 return tool->itrace_start(tool, event, sample, machine);
1547         case PERF_RECORD_SWITCH:
1548         case PERF_RECORD_SWITCH_CPU_WIDE:
1549                 return tool->context_switch(tool, event, sample, machine);
1550         case PERF_RECORD_KSYMBOL:
1551                 return tool->ksymbol(tool, event, sample, machine);
1552         case PERF_RECORD_BPF_EVENT:
1553                 return tool->bpf(tool, event, sample, machine);
1554         case PERF_RECORD_TEXT_POKE:
1555                 return tool->text_poke(tool, event, sample, machine);
1556         default:
1557                 ++evlist->stats.nr_unknown_events;
1558                 return -1;
1559         }
1560 }
1561
1562 static int perf_session__deliver_event(struct perf_session *session,
1563                                        union perf_event *event,
1564                                        struct perf_tool *tool,
1565                                        u64 file_offset)
1566 {
1567         struct perf_sample sample;
1568         int ret = evlist__parse_sample(session->evlist, event, &sample);
1569
1570         if (ret) {
1571                 pr_err("Can't parse sample, err = %d\n", ret);
1572                 return ret;
1573         }
1574
1575         ret = auxtrace__process_event(session, event, &sample, tool);
1576         if (ret < 0)
1577                 return ret;
1578         if (ret > 0)
1579                 return 0;
1580
1581         ret = machines__deliver_event(&session->machines, session->evlist,
1582                                       event, &sample, tool, file_offset);
1583
1584         if (dump_trace && sample.aux_sample.size)
1585                 auxtrace__dump_auxtrace_sample(session, &sample);
1586
1587         return ret;
1588 }
1589
1590 static s64 perf_session__process_user_event(struct perf_session *session,
1591                                             union perf_event *event,
1592                                             u64 file_offset)
1593 {
1594         struct ordered_events *oe = &session->ordered_events;
1595         struct perf_tool *tool = session->tool;
1596         struct perf_sample sample = { .time = 0, };
1597         int fd = perf_data__fd(session->data);
1598         int err;
1599
1600         if (event->header.type != PERF_RECORD_COMPRESSED ||
1601             tool->compressed == perf_session__process_compressed_event_stub)
1602                 dump_event(session->evlist, event, file_offset, &sample);
1603
1604         /* These events are processed right away */
1605         switch (event->header.type) {
1606         case PERF_RECORD_HEADER_ATTR:
1607                 err = tool->attr(tool, event, &session->evlist);
1608                 if (err == 0) {
1609                         perf_session__set_id_hdr_size(session);
1610                         perf_session__set_comm_exec(session);
1611                 }
1612                 return err;
1613         case PERF_RECORD_EVENT_UPDATE:
1614                 return tool->event_update(tool, event, &session->evlist);
1615         case PERF_RECORD_HEADER_EVENT_TYPE:
1616                 /*
1617                  * Deprecated, but we need to handle it for sake
1618                  * of old data files create in pipe mode.
1619                  */
1620                 return 0;
1621         case PERF_RECORD_HEADER_TRACING_DATA:
1622                 /*
1623                  * Setup for reading amidst mmap, but only when we
1624                  * are in 'file' mode. The 'pipe' fd is in proper
1625                  * place already.
1626                  */
1627                 if (!perf_data__is_pipe(session->data))
1628                         lseek(fd, file_offset, SEEK_SET);
1629                 return tool->tracing_data(session, event);
1630         case PERF_RECORD_HEADER_BUILD_ID:
1631                 return tool->build_id(session, event);
1632         case PERF_RECORD_FINISHED_ROUND:
1633                 return tool->finished_round(tool, event, oe);
1634         case PERF_RECORD_ID_INDEX:
1635                 return tool->id_index(session, event);
1636         case PERF_RECORD_AUXTRACE_INFO:
1637                 return tool->auxtrace_info(session, event);
1638         case PERF_RECORD_AUXTRACE:
1639                 /* setup for reading amidst mmap */
1640                 lseek(fd, file_offset + event->header.size, SEEK_SET);
1641                 return tool->auxtrace(session, event);
1642         case PERF_RECORD_AUXTRACE_ERROR:
1643                 perf_session__auxtrace_error_inc(session, event);
1644                 return tool->auxtrace_error(session, event);
1645         case PERF_RECORD_THREAD_MAP:
1646                 return tool->thread_map(session, event);
1647         case PERF_RECORD_CPU_MAP:
1648                 return tool->cpu_map(session, event);
1649         case PERF_RECORD_STAT_CONFIG:
1650                 return tool->stat_config(session, event);
1651         case PERF_RECORD_STAT:
1652                 return tool->stat(session, event);
1653         case PERF_RECORD_STAT_ROUND:
1654                 return tool->stat_round(session, event);
1655         case PERF_RECORD_TIME_CONV:
1656                 session->time_conv = event->time_conv;
1657                 return tool->time_conv(session, event);
1658         case PERF_RECORD_HEADER_FEATURE:
1659                 return tool->feature(session, event);
1660         case PERF_RECORD_COMPRESSED:
1661                 err = tool->compressed(session, event, file_offset);
1662                 if (err)
1663                         dump_event(session->evlist, event, file_offset, &sample);
1664                 return err;
1665         default:
1666                 return -EINVAL;
1667         }
1668 }
1669
1670 int perf_session__deliver_synth_event(struct perf_session *session,
1671                                       union perf_event *event,
1672                                       struct perf_sample *sample)
1673 {
1674         struct evlist *evlist = session->evlist;
1675         struct perf_tool *tool = session->tool;
1676
1677         events_stats__inc(&evlist->stats, event->header.type);
1678
1679         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1680                 return perf_session__process_user_event(session, event, 0);
1681
1682         return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0);
1683 }
1684
1685 static void event_swap(union perf_event *event, bool sample_id_all)
1686 {
1687         perf_event__swap_op swap;
1688
1689         swap = perf_event__swap_ops[event->header.type];
1690         if (swap)
1691                 swap(event, sample_id_all);
1692 }
1693
1694 int perf_session__peek_event(struct perf_session *session, off_t file_offset,
1695                              void *buf, size_t buf_sz,
1696                              union perf_event **event_ptr,
1697                              struct perf_sample *sample)
1698 {
1699         union perf_event *event;
1700         size_t hdr_sz, rest;
1701         int fd;
1702
1703         if (session->one_mmap && !session->header.needs_swap) {
1704                 event = file_offset - session->one_mmap_offset +
1705                         session->one_mmap_addr;
1706                 goto out_parse_sample;
1707         }
1708
1709         if (perf_data__is_pipe(session->data))
1710                 return -1;
1711
1712         fd = perf_data__fd(session->data);
1713         hdr_sz = sizeof(struct perf_event_header);
1714
1715         if (buf_sz < hdr_sz)
1716                 return -1;
1717
1718         if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 ||
1719             readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz)
1720                 return -1;
1721
1722         event = (union perf_event *)buf;
1723
1724         if (session->header.needs_swap)
1725                 perf_event_header__bswap(&event->header);
1726
1727         if (event->header.size < hdr_sz || event->header.size > buf_sz)
1728                 return -1;
1729
1730         buf += hdr_sz;
1731         rest = event->header.size - hdr_sz;
1732
1733         if (readn(fd, buf, rest) != (ssize_t)rest)
1734                 return -1;
1735
1736         if (session->header.needs_swap)
1737                 event_swap(event, evlist__sample_id_all(session->evlist));
1738
1739 out_parse_sample:
1740
1741         if (sample && event->header.type < PERF_RECORD_USER_TYPE_START &&
1742             evlist__parse_sample(session->evlist, event, sample))
1743                 return -1;
1744
1745         *event_ptr = event;
1746
1747         return 0;
1748 }
1749
1750 int perf_session__peek_events(struct perf_session *session, u64 offset,
1751                               u64 size, peek_events_cb_t cb, void *data)
1752 {
1753         u64 max_offset = offset + size;
1754         char buf[PERF_SAMPLE_MAX_SIZE];
1755         union perf_event *event;
1756         int err;
1757
1758         do {
1759                 err = perf_session__peek_event(session, offset, buf,
1760                                                PERF_SAMPLE_MAX_SIZE, &event,
1761                                                NULL);
1762                 if (err)
1763                         return err;
1764
1765                 err = cb(session, event, offset, data);
1766                 if (err)
1767                         return err;
1768
1769                 offset += event->header.size;
1770                 if (event->header.type == PERF_RECORD_AUXTRACE)
1771                         offset += event->auxtrace.size;
1772
1773         } while (offset < max_offset);
1774
1775         return err;
1776 }
1777
1778 static s64 perf_session__process_event(struct perf_session *session,
1779                                        union perf_event *event, u64 file_offset)
1780 {
1781         struct evlist *evlist = session->evlist;
1782         struct perf_tool *tool = session->tool;
1783         int ret;
1784
1785         if (session->header.needs_swap)
1786                 event_swap(event, evlist__sample_id_all(evlist));
1787
1788         if (event->header.type >= PERF_RECORD_HEADER_MAX)
1789                 return -EINVAL;
1790
1791         events_stats__inc(&evlist->stats, event->header.type);
1792
1793         if (event->header.type >= PERF_RECORD_USER_TYPE_START)
1794                 return perf_session__process_user_event(session, event, file_offset);
1795
1796         if (tool->ordered_events) {
1797                 u64 timestamp = -1ULL;
1798
1799                 ret = evlist__parse_sample_timestamp(evlist, event, &timestamp);
1800                 if (ret && ret != -1)
1801                         return ret;
1802
1803                 ret = perf_session__queue_event(session, event, timestamp, file_offset);
1804                 if (ret != -ETIME)
1805                         return ret;
1806         }
1807
1808         return perf_session__deliver_event(session, event, tool, file_offset);
1809 }
1810
1811 void perf_event_header__bswap(struct perf_event_header *hdr)
1812 {
1813         hdr->type = bswap_32(hdr->type);
1814         hdr->misc = bswap_16(hdr->misc);
1815         hdr->size = bswap_16(hdr->size);
1816 }
1817
1818 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid)
1819 {
1820         return machine__findnew_thread(&session->machines.host, -1, pid);
1821 }
1822
1823 int perf_session__register_idle_thread(struct perf_session *session)
1824 {
1825         struct thread *thread = machine__idle_thread(&session->machines.host);
1826
1827         /* machine__idle_thread() got the thread, so put it */
1828         thread__put(thread);
1829         return thread ? 0 : -1;
1830 }
1831
1832 static void
1833 perf_session__warn_order(const struct perf_session *session)
1834 {
1835         const struct ordered_events *oe = &session->ordered_events;
1836         struct evsel *evsel;
1837         bool should_warn = true;
1838
1839         evlist__for_each_entry(session->evlist, evsel) {
1840                 if (evsel->core.attr.write_backward)
1841                         should_warn = false;
1842         }
1843
1844         if (!should_warn)
1845                 return;
1846         if (oe->nr_unordered_events != 0)
1847                 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events);
1848 }
1849
1850 static void perf_session__warn_about_errors(const struct perf_session *session)
1851 {
1852         const struct events_stats *stats = &session->evlist->stats;
1853
1854         if (session->tool->lost == perf_event__process_lost &&
1855             stats->nr_events[PERF_RECORD_LOST] != 0) {
1856                 ui__warning("Processed %d events and lost %d chunks!\n\n"
1857                             "Check IO/CPU overload!\n\n",
1858                             stats->nr_events[0],
1859                             stats->nr_events[PERF_RECORD_LOST]);
1860         }
1861
1862         if (session->tool->lost_samples == perf_event__process_lost_samples) {
1863                 double drop_rate;
1864
1865                 drop_rate = (double)stats->total_lost_samples /
1866                             (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples);
1867                 if (drop_rate > 0.05) {
1868                         ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n",
1869                                     stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples,
1870                                     drop_rate * 100.0);
1871                 }
1872         }
1873
1874         if (session->tool->aux == perf_event__process_aux &&
1875             stats->total_aux_lost != 0) {
1876                 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n",
1877                             stats->total_aux_lost,
1878                             stats->nr_events[PERF_RECORD_AUX]);
1879         }
1880
1881         if (session->tool->aux == perf_event__process_aux &&
1882             stats->total_aux_partial != 0) {
1883                 bool vmm_exclusive = false;
1884
1885                 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive",
1886                                        &vmm_exclusive);
1887
1888                 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n"
1889                             "Are you running a KVM guest in the background?%s\n\n",
1890                             stats->total_aux_partial,
1891                             stats->nr_events[PERF_RECORD_AUX],
1892                             vmm_exclusive ?
1893                             "\nReloading kvm_intel module with vmm_exclusive=0\n"
1894                             "will reduce the gaps to only guest's timeslices." :
1895                             "");
1896         }
1897
1898         if (stats->nr_unknown_events != 0) {
1899                 ui__warning("Found %u unknown events!\n\n"
1900                             "Is this an older tool processing a perf.data "
1901                             "file generated by a more recent tool?\n\n"
1902                             "If that is not the case, consider "
1903                             "reporting to linux-kernel@vger.kernel.org.\n\n",
1904                             stats->nr_unknown_events);
1905         }
1906
1907         if (stats->nr_unknown_id != 0) {
1908                 ui__warning("%u samples with id not present in the header\n",
1909                             stats->nr_unknown_id);
1910         }
1911
1912         if (stats->nr_invalid_chains != 0) {
1913                 ui__warning("Found invalid callchains!\n\n"
1914                             "%u out of %u events were discarded for this reason.\n\n"
1915                             "Consider reporting to linux-kernel@vger.kernel.org.\n\n",
1916                             stats->nr_invalid_chains,
1917                             stats->nr_events[PERF_RECORD_SAMPLE]);
1918         }
1919
1920         if (stats->nr_unprocessable_samples != 0) {
1921                 ui__warning("%u unprocessable samples recorded.\n"
1922                             "Do you have a KVM guest running and not using 'perf kvm'?\n",
1923                             stats->nr_unprocessable_samples);
1924         }
1925
1926         perf_session__warn_order(session);
1927
1928         events_stats__auxtrace_error_warn(stats);
1929
1930         if (stats->nr_proc_map_timeout != 0) {
1931                 ui__warning("%d map information files for pre-existing threads were\n"
1932                             "not processed, if there are samples for addresses they\n"
1933                             "will not be resolved, you may find out which are these\n"
1934                             "threads by running with -v and redirecting the output\n"
1935                             "to a file.\n"
1936                             "The time limit to process proc map is too short?\n"
1937                             "Increase it by --proc-map-timeout\n",
1938                             stats->nr_proc_map_timeout);
1939         }
1940 }
1941
1942 static int perf_session__flush_thread_stack(struct thread *thread,
1943                                             void *p __maybe_unused)
1944 {
1945         return thread_stack__flush(thread);
1946 }
1947
1948 static int perf_session__flush_thread_stacks(struct perf_session *session)
1949 {
1950         return machines__for_each_thread(&session->machines,
1951                                          perf_session__flush_thread_stack,
1952                                          NULL);
1953 }
1954
1955 volatile int session_done;
1956
1957 static int __perf_session__process_decomp_events(struct perf_session *session);
1958
1959 static int __perf_session__process_pipe_events(struct perf_session *session)
1960 {
1961         struct ordered_events *oe = &session->ordered_events;
1962         struct perf_tool *tool = session->tool;
1963         union perf_event *event;
1964         uint32_t size, cur_size = 0;
1965         void *buf = NULL;
1966         s64 skip = 0;
1967         u64 head;
1968         ssize_t err;
1969         void *p;
1970
1971         perf_tool__fill_defaults(tool);
1972
1973         head = 0;
1974         cur_size = sizeof(union perf_event);
1975
1976         buf = malloc(cur_size);
1977         if (!buf)
1978                 return -errno;
1979         ordered_events__set_copy_on_queue(oe, true);
1980 more:
1981         event = buf;
1982         err = perf_data__read(session->data, event,
1983                               sizeof(struct perf_event_header));
1984         if (err <= 0) {
1985                 if (err == 0)
1986                         goto done;
1987
1988                 pr_err("failed to read event header\n");
1989                 goto out_err;
1990         }
1991
1992         if (session->header.needs_swap)
1993                 perf_event_header__bswap(&event->header);
1994
1995         size = event->header.size;
1996         if (size < sizeof(struct perf_event_header)) {
1997                 pr_err("bad event header size\n");
1998                 goto out_err;
1999         }
2000
2001         if (size > cur_size) {
2002                 void *new = realloc(buf, size);
2003                 if (!new) {
2004                         pr_err("failed to allocate memory to read event\n");
2005                         goto out_err;
2006                 }
2007                 buf = new;
2008                 cur_size = size;
2009                 event = buf;
2010         }
2011         p = event;
2012         p += sizeof(struct perf_event_header);
2013
2014         if (size - sizeof(struct perf_event_header)) {
2015                 err = perf_data__read(session->data, p,
2016                                       size - sizeof(struct perf_event_header));
2017                 if (err <= 0) {
2018                         if (err == 0) {
2019                                 pr_err("unexpected end of event stream\n");
2020                                 goto done;
2021                         }
2022
2023                         pr_err("failed to read event data\n");
2024                         goto out_err;
2025                 }
2026         }
2027
2028         if ((skip = perf_session__process_event(session, event, head)) < 0) {
2029                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
2030                        head, event->header.size, event->header.type);
2031                 err = -EINVAL;
2032                 goto out_err;
2033         }
2034
2035         head += size;
2036
2037         if (skip > 0)
2038                 head += skip;
2039
2040         err = __perf_session__process_decomp_events(session);
2041         if (err)
2042                 goto out_err;
2043
2044         if (!session_done())
2045                 goto more;
2046 done:
2047         /* do the final flush for ordered samples */
2048         err = ordered_events__flush(oe, OE_FLUSH__FINAL);
2049         if (err)
2050                 goto out_err;
2051         err = auxtrace__flush_events(session, tool);
2052         if (err)
2053                 goto out_err;
2054         err = perf_session__flush_thread_stacks(session);
2055 out_err:
2056         free(buf);
2057         if (!tool->no_warn)
2058                 perf_session__warn_about_errors(session);
2059         ordered_events__free(&session->ordered_events);
2060         auxtrace__free_events(session);
2061         return err;
2062 }
2063
2064 static union perf_event *
2065 prefetch_event(char *buf, u64 head, size_t mmap_size,
2066                bool needs_swap, union perf_event *error)
2067 {
2068         union perf_event *event;
2069
2070         /*
2071          * Ensure we have enough space remaining to read
2072          * the size of the event in the headers.
2073          */
2074         if (head + sizeof(event->header) > mmap_size)
2075                 return NULL;
2076
2077         event = (union perf_event *)(buf + head);
2078         if (needs_swap)
2079                 perf_event_header__bswap(&event->header);
2080
2081         if (head + event->header.size <= mmap_size)
2082                 return event;
2083
2084         /* We're not fetching the event so swap back again */
2085         if (needs_swap)
2086                 perf_event_header__bswap(&event->header);
2087
2088         pr_debug("%s: head=%#" PRIx64 " event->header_size=%#x, mmap_size=%#zx:"
2089                  " fuzzed or compressed perf.data?\n",__func__, head, event->header.size, mmap_size);
2090
2091         return error;
2092 }
2093
2094 static union perf_event *
2095 fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap)
2096 {
2097         return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL));
2098 }
2099
2100 static union perf_event *
2101 fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap)
2102 {
2103         return prefetch_event(buf, head, mmap_size, needs_swap, NULL);
2104 }
2105
2106 static int __perf_session__process_decomp_events(struct perf_session *session)
2107 {
2108         s64 skip;
2109         u64 size, file_pos = 0;
2110         struct decomp *decomp = session->decomp_last;
2111
2112         if (!decomp)
2113                 return 0;
2114
2115         while (decomp->head < decomp->size && !session_done()) {
2116                 union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data,
2117                                                              session->header.needs_swap);
2118
2119                 if (!event)
2120                         break;
2121
2122                 size = event->header.size;
2123
2124                 if (size < sizeof(struct perf_event_header) ||
2125                     (skip = perf_session__process_event(session, event, file_pos)) < 0) {
2126                         pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n",
2127                                 decomp->file_pos + decomp->head, event->header.size, event->header.type);
2128                         return -EINVAL;
2129                 }
2130
2131                 if (skip)
2132                         size += skip;
2133
2134                 decomp->head += size;
2135         }
2136
2137         return 0;
2138 }
2139
2140 /*
2141  * On 64bit we can mmap the data file in one go. No need for tiny mmap
2142  * slices. On 32bit we use 32MB.
2143  */
2144 #if BITS_PER_LONG == 64
2145 #define MMAP_SIZE ULLONG_MAX
2146 #define NUM_MMAPS 1
2147 #else
2148 #define MMAP_SIZE (32 * 1024 * 1024ULL)
2149 #define NUM_MMAPS 128
2150 #endif
2151
2152 struct reader;
2153
2154 typedef s64 (*reader_cb_t)(struct perf_session *session,
2155                            union perf_event *event,
2156                            u64 file_offset);
2157
2158 struct reader {
2159         int              fd;
2160         u64              data_size;
2161         u64              data_offset;
2162         reader_cb_t      process;
2163         bool             in_place_update;
2164 };
2165
2166 static int
2167 reader__process_events(struct reader *rd, struct perf_session *session,
2168                        struct ui_progress *prog)
2169 {
2170         u64 data_size = rd->data_size;
2171         u64 head, page_offset, file_offset, file_pos, size;
2172         int err = 0, mmap_prot, mmap_flags, map_idx = 0;
2173         size_t  mmap_size;
2174         char *buf, *mmaps[NUM_MMAPS];
2175         union perf_event *event;
2176         s64 skip;
2177
2178         page_offset = page_size * (rd->data_offset / page_size);
2179         file_offset = page_offset;
2180         head = rd->data_offset - page_offset;
2181
2182         ui_progress__init_size(prog, data_size, "Processing events...");
2183
2184         data_size += rd->data_offset;
2185
2186         mmap_size = MMAP_SIZE;
2187         if (mmap_size > data_size) {
2188                 mmap_size = data_size;
2189                 session->one_mmap = true;
2190         }
2191
2192         memset(mmaps, 0, sizeof(mmaps));
2193
2194         mmap_prot  = PROT_READ;
2195         mmap_flags = MAP_SHARED;
2196
2197         if (rd->in_place_update) {
2198                 mmap_prot  |= PROT_WRITE;
2199         } else if (session->header.needs_swap) {
2200                 mmap_prot  |= PROT_WRITE;
2201                 mmap_flags = MAP_PRIVATE;
2202         }
2203 remap:
2204         buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, rd->fd,
2205                    file_offset);
2206         if (buf == MAP_FAILED) {
2207                 pr_err("failed to mmap file\n");
2208                 err = -errno;
2209                 goto out;
2210         }
2211         mmaps[map_idx] = buf;
2212         map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1);
2213         file_pos = file_offset + head;
2214         if (session->one_mmap) {
2215                 session->one_mmap_addr = buf;
2216                 session->one_mmap_offset = file_offset;
2217         }
2218
2219 more:
2220         event = fetch_mmaped_event(head, mmap_size, buf, session->header.needs_swap);
2221         if (IS_ERR(event))
2222                 return PTR_ERR(event);
2223
2224         if (!event) {
2225                 if (mmaps[map_idx]) {
2226                         munmap(mmaps[map_idx], mmap_size);
2227                         mmaps[map_idx] = NULL;
2228                 }
2229
2230                 page_offset = page_size * (head / page_size);
2231                 file_offset += page_offset;
2232                 head -= page_offset;
2233                 goto remap;
2234         }
2235
2236         size = event->header.size;
2237
2238         skip = -EINVAL;
2239
2240         if (size < sizeof(struct perf_event_header) ||
2241             (skip = rd->process(session, event, file_pos)) < 0) {
2242                 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n",
2243                        file_offset + head, event->header.size,
2244                        event->header.type, strerror(-skip));
2245                 err = skip;
2246                 goto out;
2247         }
2248
2249         if (skip)
2250                 size += skip;
2251
2252         head += size;
2253         file_pos += size;
2254
2255         err = __perf_session__process_decomp_events(session);
2256         if (err)
2257                 goto out;
2258
2259         ui_progress__update(prog, size);
2260
2261         if (session_done())
2262                 goto out;
2263
2264         if (file_pos < data_size)
2265                 goto more;
2266
2267 out:
2268         return err;
2269 }
2270
2271 static s64 process_simple(struct perf_session *session,
2272                           union perf_event *event,
2273                           u64 file_offset)
2274 {
2275         return perf_session__process_event(session, event, file_offset);
2276 }
2277
2278 static int __perf_session__process_events(struct perf_session *session)
2279 {
2280         struct reader rd = {
2281                 .fd             = perf_data__fd(session->data),
2282                 .data_size      = session->header.data_size,
2283                 .data_offset    = session->header.data_offset,
2284                 .process        = process_simple,
2285                 .in_place_update = session->data->in_place_update,
2286         };
2287         struct ordered_events *oe = &session->ordered_events;
2288         struct perf_tool *tool = session->tool;
2289         struct ui_progress prog;
2290         int err;
2291
2292         perf_tool__fill_defaults(tool);
2293
2294         if (rd.data_size == 0)
2295                 return -1;
2296
2297         ui_progress__init_size(&prog, rd.data_size, "Processing events...");
2298
2299         err = reader__process_events(&rd, session, &prog);
2300         if (err)
2301                 goto out_err;
2302         /* do the final flush for ordered samples */
2303         err = ordered_events__flush(oe, OE_FLUSH__FINAL);
2304         if (err)
2305                 goto out_err;
2306         err = auxtrace__flush_events(session, tool);
2307         if (err)
2308                 goto out_err;
2309         err = perf_session__flush_thread_stacks(session);
2310 out_err:
2311         ui_progress__finish();
2312         if (!tool->no_warn)
2313                 perf_session__warn_about_errors(session);
2314         /*
2315          * We may switching perf.data output, make ordered_events
2316          * reusable.
2317          */
2318         ordered_events__reinit(&session->ordered_events);
2319         auxtrace__free_events(session);
2320         session->one_mmap = false;
2321         return err;
2322 }
2323
2324 int perf_session__process_events(struct perf_session *session)
2325 {
2326         if (perf_session__register_idle_thread(session) < 0)
2327                 return -ENOMEM;
2328
2329         if (perf_data__is_pipe(session->data))
2330                 return __perf_session__process_pipe_events(session);
2331
2332         return __perf_session__process_events(session);
2333 }
2334
2335 bool perf_session__has_traces(struct perf_session *session, const char *msg)
2336 {
2337         struct evsel *evsel;
2338
2339         evlist__for_each_entry(session->evlist, evsel) {
2340                 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT)
2341                         return true;
2342         }
2343
2344         pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg);
2345         return false;
2346 }
2347
2348 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr)
2349 {
2350         char *bracket;
2351         struct ref_reloc_sym *ref;
2352         struct kmap *kmap;
2353
2354         ref = zalloc(sizeof(struct ref_reloc_sym));
2355         if (ref == NULL)
2356                 return -ENOMEM;
2357
2358         ref->name = strdup(symbol_name);
2359         if (ref->name == NULL) {
2360                 free(ref);
2361                 return -ENOMEM;
2362         }
2363
2364         bracket = strchr(ref->name, ']');
2365         if (bracket)
2366                 *bracket = '\0';
2367
2368         ref->addr = addr;
2369
2370         kmap = map__kmap(map);
2371         if (kmap)
2372                 kmap->ref_reloc_sym = ref;
2373
2374         return 0;
2375 }
2376
2377 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp)
2378 {
2379         return machines__fprintf_dsos(&session->machines, fp);
2380 }
2381
2382 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp,
2383                                           bool (skip)(struct dso *dso, int parm), int parm)
2384 {
2385         return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm);
2386 }
2387
2388 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp,
2389                                        bool skip_empty)
2390 {
2391         size_t ret;
2392         const char *msg = "";
2393
2394         if (perf_header__has_feat(&session->header, HEADER_AUXTRACE))
2395                 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)";
2396
2397         ret = fprintf(fp, "\nAggregated stats:%s\n", msg);
2398
2399         ret += events_stats__fprintf(&session->evlist->stats, fp, skip_empty);
2400         return ret;
2401 }
2402
2403 size_t perf_session__fprintf(struct perf_session *session, FILE *fp)
2404 {
2405         /*
2406          * FIXME: Here we have to actually print all the machines in this
2407          * session, not just the host...
2408          */
2409         return machine__fprintf(&session->machines.host, fp);
2410 }
2411
2412 struct evsel *perf_session__find_first_evtype(struct perf_session *session,
2413                                               unsigned int type)
2414 {
2415         struct evsel *pos;
2416
2417         evlist__for_each_entry(session->evlist, pos) {
2418                 if (pos->core.attr.type == type)
2419                         return pos;
2420         }
2421         return NULL;
2422 }
2423
2424 int perf_session__cpu_bitmap(struct perf_session *session,
2425                              const char *cpu_list, unsigned long *cpu_bitmap)
2426 {
2427         int i, err = -1;
2428         struct perf_cpu_map *map;
2429         int nr_cpus = min(session->header.env.nr_cpus_avail, MAX_NR_CPUS);
2430
2431         for (i = 0; i < PERF_TYPE_MAX; ++i) {
2432                 struct evsel *evsel;
2433
2434                 evsel = perf_session__find_first_evtype(session, i);
2435                 if (!evsel)
2436                         continue;
2437
2438                 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) {
2439                         pr_err("File does not contain CPU events. "
2440                                "Remove -C option to proceed.\n");
2441                         return -1;
2442                 }
2443         }
2444
2445         map = perf_cpu_map__new(cpu_list);
2446         if (map == NULL) {
2447                 pr_err("Invalid cpu_list\n");
2448                 return -1;
2449         }
2450
2451         for (i = 0; i < map->nr; i++) {
2452                 int cpu = map->map[i];
2453
2454                 if (cpu >= nr_cpus) {
2455                         pr_err("Requested CPU %d too large. "
2456                                "Consider raising MAX_NR_CPUS\n", cpu);
2457                         goto out_delete_map;
2458                 }
2459
2460                 set_bit(cpu, cpu_bitmap);
2461         }
2462
2463         err = 0;
2464
2465 out_delete_map:
2466         perf_cpu_map__put(map);
2467         return err;
2468 }
2469
2470 void perf_session__fprintf_info(struct perf_session *session, FILE *fp,
2471                                 bool full)
2472 {
2473         if (session == NULL || fp == NULL)
2474                 return;
2475
2476         fprintf(fp, "# ========\n");
2477         perf_header__fprintf_info(session, fp, full);
2478         fprintf(fp, "# ========\n#\n");
2479 }
2480
2481 int perf_event__process_id_index(struct perf_session *session,
2482                                  union perf_event *event)
2483 {
2484         struct evlist *evlist = session->evlist;
2485         struct perf_record_id_index *ie = &event->id_index;
2486         size_t i, nr, max_nr;
2487
2488         max_nr = (ie->header.size - sizeof(struct perf_record_id_index)) /
2489                  sizeof(struct id_index_entry);
2490         nr = ie->nr;
2491         if (nr > max_nr)
2492                 return -EINVAL;
2493
2494         if (dump_trace)
2495                 fprintf(stdout, " nr: %zu\n", nr);
2496
2497         for (i = 0; i < nr; i++) {
2498                 struct id_index_entry *e = &ie->entries[i];
2499                 struct perf_sample_id *sid;
2500
2501                 if (dump_trace) {
2502                         fprintf(stdout, " ... id: %"PRI_lu64, e->id);
2503                         fprintf(stdout, "  idx: %"PRI_lu64, e->idx);
2504                         fprintf(stdout, "  cpu: %"PRI_ld64, e->cpu);
2505                         fprintf(stdout, "  tid: %"PRI_ld64"\n", e->tid);
2506                 }
2507
2508                 sid = evlist__id2sid(evlist, e->id);
2509                 if (!sid)
2510                         return -ENOENT;
2511                 sid->idx = e->idx;
2512                 sid->cpu = e->cpu;
2513                 sid->tid = e->tid;
2514         }
2515         return 0;
2516 }