tools/perf/util/session.c

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