1 #include <linux/kernel.h> 2 #include <traceevent/event-parse.h> 3 4 #include <byteswap.h> 5 #include <unistd.h> 6 #include <sys/types.h> 7 #include <sys/mman.h> 8 9 #include "evlist.h" 10 #include "evsel.h" 11 #include "session.h" 12 #include "tool.h" 13 #include "sort.h" 14 #include "util.h" 15 #include "cpumap.h" 16 #include "perf_regs.h" 17 #include "asm/bug.h" 18 #include "auxtrace.h" 19 #include "thread-stack.h" 20 21 static int perf_session__deliver_event(struct perf_session *session, 22 union perf_event *event, 23 struct perf_sample *sample, 24 struct perf_tool *tool, 25 u64 file_offset); 26 27 static int perf_session__open(struct perf_session *session) 28 { 29 struct perf_data_file *file = session->file; 30 31 if (perf_session__read_header(session) < 0) { 32 pr_err("incompatible file format (rerun with -v to learn more)"); 33 return -1; 34 } 35 36 if (perf_data_file__is_pipe(file)) 37 return 0; 38 39 if (!perf_evlist__valid_sample_type(session->evlist)) { 40 pr_err("non matching sample_type"); 41 return -1; 42 } 43 44 if (!perf_evlist__valid_sample_id_all(session->evlist)) { 45 pr_err("non matching sample_id_all"); 46 return -1; 47 } 48 49 if (!perf_evlist__valid_read_format(session->evlist)) { 50 pr_err("non matching read_format"); 51 return -1; 52 } 53 54 return 0; 55 } 56 57 void perf_session__set_id_hdr_size(struct perf_session *session) 58 { 59 u16 id_hdr_size = perf_evlist__id_hdr_size(session->evlist); 60 61 machines__set_id_hdr_size(&session->machines, id_hdr_size); 62 } 63 64 int perf_session__create_kernel_maps(struct perf_session *session) 65 { 66 int ret = machine__create_kernel_maps(&session->machines.host); 67 68 if (ret >= 0) 69 ret = machines__create_guest_kernel_maps(&session->machines); 70 return ret; 71 } 72 73 static void perf_session__destroy_kernel_maps(struct perf_session *session) 74 { 75 machines__destroy_kernel_maps(&session->machines); 76 } 77 78 static bool perf_session__has_comm_exec(struct perf_session *session) 79 { 80 struct perf_evsel *evsel; 81 82 evlist__for_each(session->evlist, evsel) { 83 if (evsel->attr.comm_exec) 84 return true; 85 } 86 87 return false; 88 } 89 90 static void perf_session__set_comm_exec(struct perf_session *session) 91 { 92 bool comm_exec = perf_session__has_comm_exec(session); 93 94 machines__set_comm_exec(&session->machines, comm_exec); 95 } 96 97 static int ordered_events__deliver_event(struct ordered_events *oe, 98 struct ordered_event *event) 99 { 100 struct perf_sample sample; 101 struct perf_session *session = container_of(oe, struct perf_session, 102 ordered_events); 103 int ret = perf_evlist__parse_sample(session->evlist, event->event, &sample); 104 105 if (ret) { 106 pr_err("Can't parse sample, err = %d\n", ret); 107 return ret; 108 } 109 110 return perf_session__deliver_event(session, event->event, &sample, 111 session->tool, event->file_offset); 112 } 113 114 struct perf_session *perf_session__new(struct perf_data_file *file, 115 bool repipe, struct perf_tool *tool) 116 { 117 struct perf_session *session = zalloc(sizeof(*session)); 118 119 if (!session) 120 goto out; 121 122 session->repipe = repipe; 123 session->tool = tool; 124 INIT_LIST_HEAD(&session->auxtrace_index); 125 machines__init(&session->machines); 126 ordered_events__init(&session->ordered_events, ordered_events__deliver_event); 127 128 if (file) { 129 if (perf_data_file__open(file)) 130 goto out_delete; 131 132 session->file = file; 133 134 if (perf_data_file__is_read(file)) { 135 if (perf_session__open(session) < 0) 136 goto out_close; 137 138 perf_session__set_id_hdr_size(session); 139 perf_session__set_comm_exec(session); 140 } 141 } 142 143 if (!file || perf_data_file__is_write(file)) { 144 /* 145 * In O_RDONLY mode this will be performed when reading the 146 * kernel MMAP event, in perf_event__process_mmap(). 147 */ 148 if (perf_session__create_kernel_maps(session) < 0) 149 pr_warning("Cannot read kernel map\n"); 150 } 151 152 if (tool && tool->ordering_requires_timestamps && 153 tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) { 154 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 155 tool->ordered_events = false; 156 } 157 158 return session; 159 160 out_close: 161 perf_data_file__close(file); 162 out_delete: 163 perf_session__delete(session); 164 out: 165 return NULL; 166 } 167 168 static void perf_session__delete_threads(struct perf_session *session) 169 { 170 machine__delete_threads(&session->machines.host); 171 } 172 173 static void perf_session_env__exit(struct perf_env *env) 174 { 175 zfree(&env->hostname); 176 zfree(&env->os_release); 177 zfree(&env->version); 178 zfree(&env->arch); 179 zfree(&env->cpu_desc); 180 zfree(&env->cpuid); 181 182 zfree(&env->cmdline); 183 zfree(&env->cmdline_argv); 184 zfree(&env->sibling_cores); 185 zfree(&env->sibling_threads); 186 zfree(&env->numa_nodes); 187 zfree(&env->pmu_mappings); 188 zfree(&env->cpu); 189 } 190 191 void perf_session__delete(struct perf_session *session) 192 { 193 auxtrace__free(session); 194 auxtrace_index__free(&session->auxtrace_index); 195 perf_session__destroy_kernel_maps(session); 196 perf_session__delete_threads(session); 197 perf_session_env__exit(&session->header.env); 198 machines__exit(&session->machines); 199 if (session->file) 200 perf_data_file__close(session->file); 201 free(session); 202 } 203 204 static int process_event_synth_tracing_data_stub(struct perf_tool *tool 205 __maybe_unused, 206 union perf_event *event 207 __maybe_unused, 208 struct perf_session *session 209 __maybe_unused) 210 { 211 dump_printf(": unhandled!\n"); 212 return 0; 213 } 214 215 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused, 216 union perf_event *event __maybe_unused, 217 struct perf_evlist **pevlist 218 __maybe_unused) 219 { 220 dump_printf(": unhandled!\n"); 221 return 0; 222 } 223 224 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused, 225 union perf_event *event __maybe_unused, 226 struct perf_sample *sample __maybe_unused, 227 struct perf_evsel *evsel __maybe_unused, 228 struct machine *machine __maybe_unused) 229 { 230 dump_printf(": unhandled!\n"); 231 return 0; 232 } 233 234 static int process_event_stub(struct perf_tool *tool __maybe_unused, 235 union perf_event *event __maybe_unused, 236 struct perf_sample *sample __maybe_unused, 237 struct machine *machine __maybe_unused) 238 { 239 dump_printf(": unhandled!\n"); 240 return 0; 241 } 242 243 static int process_build_id_stub(struct perf_tool *tool __maybe_unused, 244 union perf_event *event __maybe_unused, 245 struct perf_session *session __maybe_unused) 246 { 247 dump_printf(": unhandled!\n"); 248 return 0; 249 } 250 251 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused, 252 union perf_event *event __maybe_unused, 253 struct ordered_events *oe __maybe_unused) 254 { 255 dump_printf(": unhandled!\n"); 256 return 0; 257 } 258 259 static int process_finished_round(struct perf_tool *tool, 260 union perf_event *event, 261 struct ordered_events *oe); 262 263 static int process_id_index_stub(struct perf_tool *tool __maybe_unused, 264 union perf_event *event __maybe_unused, 265 struct perf_session *perf_session 266 __maybe_unused) 267 { 268 dump_printf(": unhandled!\n"); 269 return 0; 270 } 271 272 static int process_event_auxtrace_info_stub(struct perf_tool *tool __maybe_unused, 273 union perf_event *event __maybe_unused, 274 struct perf_session *session __maybe_unused) 275 { 276 dump_printf(": unhandled!\n"); 277 return 0; 278 } 279 280 static int skipn(int fd, off_t n) 281 { 282 char buf[4096]; 283 ssize_t ret; 284 285 while (n > 0) { 286 ret = read(fd, buf, min(n, (off_t)sizeof(buf))); 287 if (ret <= 0) 288 return ret; 289 n -= ret; 290 } 291 292 return 0; 293 } 294 295 static s64 process_event_auxtrace_stub(struct perf_tool *tool __maybe_unused, 296 union perf_event *event, 297 struct perf_session *session 298 __maybe_unused) 299 { 300 dump_printf(": unhandled!\n"); 301 if (perf_data_file__is_pipe(session->file)) 302 skipn(perf_data_file__fd(session->file), event->auxtrace.size); 303 return event->auxtrace.size; 304 } 305 306 static 307 int process_event_auxtrace_error_stub(struct perf_tool *tool __maybe_unused, 308 union perf_event *event __maybe_unused, 309 struct perf_session *session __maybe_unused) 310 { 311 dump_printf(": unhandled!\n"); 312 return 0; 313 } 314 315 void perf_tool__fill_defaults(struct perf_tool *tool) 316 { 317 if (tool->sample == NULL) 318 tool->sample = process_event_sample_stub; 319 if (tool->mmap == NULL) 320 tool->mmap = process_event_stub; 321 if (tool->mmap2 == NULL) 322 tool->mmap2 = process_event_stub; 323 if (tool->comm == NULL) 324 tool->comm = process_event_stub; 325 if (tool->fork == NULL) 326 tool->fork = process_event_stub; 327 if (tool->exit == NULL) 328 tool->exit = process_event_stub; 329 if (tool->lost == NULL) 330 tool->lost = perf_event__process_lost; 331 if (tool->lost_samples == NULL) 332 tool->lost_samples = perf_event__process_lost_samples; 333 if (tool->aux == NULL) 334 tool->aux = perf_event__process_aux; 335 if (tool->itrace_start == NULL) 336 tool->itrace_start = perf_event__process_itrace_start; 337 if (tool->context_switch == NULL) 338 tool->context_switch = perf_event__process_switch; 339 if (tool->read == NULL) 340 tool->read = process_event_sample_stub; 341 if (tool->throttle == NULL) 342 tool->throttle = process_event_stub; 343 if (tool->unthrottle == NULL) 344 tool->unthrottle = process_event_stub; 345 if (tool->attr == NULL) 346 tool->attr = process_event_synth_attr_stub; 347 if (tool->tracing_data == NULL) 348 tool->tracing_data = process_event_synth_tracing_data_stub; 349 if (tool->build_id == NULL) 350 tool->build_id = process_build_id_stub; 351 if (tool->finished_round == NULL) { 352 if (tool->ordered_events) 353 tool->finished_round = process_finished_round; 354 else 355 tool->finished_round = process_finished_round_stub; 356 } 357 if (tool->id_index == NULL) 358 tool->id_index = process_id_index_stub; 359 if (tool->auxtrace_info == NULL) 360 tool->auxtrace_info = process_event_auxtrace_info_stub; 361 if (tool->auxtrace == NULL) 362 tool->auxtrace = process_event_auxtrace_stub; 363 if (tool->auxtrace_error == NULL) 364 tool->auxtrace_error = process_event_auxtrace_error_stub; 365 } 366 367 static void swap_sample_id_all(union perf_event *event, void *data) 368 { 369 void *end = (void *) event + event->header.size; 370 int size = end - data; 371 372 BUG_ON(size % sizeof(u64)); 373 mem_bswap_64(data, size); 374 } 375 376 static void perf_event__all64_swap(union perf_event *event, 377 bool sample_id_all __maybe_unused) 378 { 379 struct perf_event_header *hdr = &event->header; 380 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 381 } 382 383 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all) 384 { 385 event->comm.pid = bswap_32(event->comm.pid); 386 event->comm.tid = bswap_32(event->comm.tid); 387 388 if (sample_id_all) { 389 void *data = &event->comm.comm; 390 391 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 392 swap_sample_id_all(event, data); 393 } 394 } 395 396 static void perf_event__mmap_swap(union perf_event *event, 397 bool sample_id_all) 398 { 399 event->mmap.pid = bswap_32(event->mmap.pid); 400 event->mmap.tid = bswap_32(event->mmap.tid); 401 event->mmap.start = bswap_64(event->mmap.start); 402 event->mmap.len = bswap_64(event->mmap.len); 403 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 404 405 if (sample_id_all) { 406 void *data = &event->mmap.filename; 407 408 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 409 swap_sample_id_all(event, data); 410 } 411 } 412 413 static void perf_event__mmap2_swap(union perf_event *event, 414 bool sample_id_all) 415 { 416 event->mmap2.pid = bswap_32(event->mmap2.pid); 417 event->mmap2.tid = bswap_32(event->mmap2.tid); 418 event->mmap2.start = bswap_64(event->mmap2.start); 419 event->mmap2.len = bswap_64(event->mmap2.len); 420 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff); 421 event->mmap2.maj = bswap_32(event->mmap2.maj); 422 event->mmap2.min = bswap_32(event->mmap2.min); 423 event->mmap2.ino = bswap_64(event->mmap2.ino); 424 425 if (sample_id_all) { 426 void *data = &event->mmap2.filename; 427 428 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 429 swap_sample_id_all(event, data); 430 } 431 } 432 static void perf_event__task_swap(union perf_event *event, bool sample_id_all) 433 { 434 event->fork.pid = bswap_32(event->fork.pid); 435 event->fork.tid = bswap_32(event->fork.tid); 436 event->fork.ppid = bswap_32(event->fork.ppid); 437 event->fork.ptid = bswap_32(event->fork.ptid); 438 event->fork.time = bswap_64(event->fork.time); 439 440 if (sample_id_all) 441 swap_sample_id_all(event, &event->fork + 1); 442 } 443 444 static void perf_event__read_swap(union perf_event *event, bool sample_id_all) 445 { 446 event->read.pid = bswap_32(event->read.pid); 447 event->read.tid = bswap_32(event->read.tid); 448 event->read.value = bswap_64(event->read.value); 449 event->read.time_enabled = bswap_64(event->read.time_enabled); 450 event->read.time_running = bswap_64(event->read.time_running); 451 event->read.id = bswap_64(event->read.id); 452 453 if (sample_id_all) 454 swap_sample_id_all(event, &event->read + 1); 455 } 456 457 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all) 458 { 459 event->aux.aux_offset = bswap_64(event->aux.aux_offset); 460 event->aux.aux_size = bswap_64(event->aux.aux_size); 461 event->aux.flags = bswap_64(event->aux.flags); 462 463 if (sample_id_all) 464 swap_sample_id_all(event, &event->aux + 1); 465 } 466 467 static void perf_event__itrace_start_swap(union perf_event *event, 468 bool sample_id_all) 469 { 470 event->itrace_start.pid = bswap_32(event->itrace_start.pid); 471 event->itrace_start.tid = bswap_32(event->itrace_start.tid); 472 473 if (sample_id_all) 474 swap_sample_id_all(event, &event->itrace_start + 1); 475 } 476 477 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all) 478 { 479 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) { 480 event->context_switch.next_prev_pid = 481 bswap_32(event->context_switch.next_prev_pid); 482 event->context_switch.next_prev_tid = 483 bswap_32(event->context_switch.next_prev_tid); 484 } 485 486 if (sample_id_all) 487 swap_sample_id_all(event, &event->context_switch + 1); 488 } 489 490 static void perf_event__throttle_swap(union perf_event *event, 491 bool sample_id_all) 492 { 493 event->throttle.time = bswap_64(event->throttle.time); 494 event->throttle.id = bswap_64(event->throttle.id); 495 event->throttle.stream_id = bswap_64(event->throttle.stream_id); 496 497 if (sample_id_all) 498 swap_sample_id_all(event, &event->throttle + 1); 499 } 500 501 static u8 revbyte(u8 b) 502 { 503 int rev = (b >> 4) | ((b & 0xf) << 4); 504 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2); 505 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1); 506 return (u8) rev; 507 } 508 509 /* 510 * XXX this is hack in attempt to carry flags bitfield 511 * throught endian village. ABI says: 512 * 513 * Bit-fields are allocated from right to left (least to most significant) 514 * on little-endian implementations and from left to right (most to least 515 * significant) on big-endian implementations. 516 * 517 * The above seems to be byte specific, so we need to reverse each 518 * byte of the bitfield. 'Internet' also says this might be implementation 519 * specific and we probably need proper fix and carry perf_event_attr 520 * bitfield flags in separate data file FEAT_ section. Thought this seems 521 * to work for now. 522 */ 523 static void swap_bitfield(u8 *p, unsigned len) 524 { 525 unsigned i; 526 527 for (i = 0; i < len; i++) { 528 *p = revbyte(*p); 529 p++; 530 } 531 } 532 533 /* exported for swapping attributes in file header */ 534 void perf_event__attr_swap(struct perf_event_attr *attr) 535 { 536 attr->type = bswap_32(attr->type); 537 attr->size = bswap_32(attr->size); 538 539 #define bswap_safe(f, n) \ 540 (attr->size > (offsetof(struct perf_event_attr, f) + \ 541 sizeof(attr->f) * (n))) 542 #define bswap_field(f, sz) \ 543 do { \ 544 if (bswap_safe(f, 0)) \ 545 attr->f = bswap_##sz(attr->f); \ 546 } while(0) 547 #define bswap_field_32(f) bswap_field(f, 32) 548 #define bswap_field_64(f) bswap_field(f, 64) 549 550 bswap_field_64(config); 551 bswap_field_64(sample_period); 552 bswap_field_64(sample_type); 553 bswap_field_64(read_format); 554 bswap_field_32(wakeup_events); 555 bswap_field_32(bp_type); 556 bswap_field_64(bp_addr); 557 bswap_field_64(bp_len); 558 bswap_field_64(branch_sample_type); 559 bswap_field_64(sample_regs_user); 560 bswap_field_32(sample_stack_user); 561 bswap_field_32(aux_watermark); 562 563 /* 564 * After read_format are bitfields. Check read_format because 565 * we are unable to use offsetof on bitfield. 566 */ 567 if (bswap_safe(read_format, 1)) 568 swap_bitfield((u8 *) (&attr->read_format + 1), 569 sizeof(u64)); 570 #undef bswap_field_64 571 #undef bswap_field_32 572 #undef bswap_field 573 #undef bswap_safe 574 } 575 576 static void perf_event__hdr_attr_swap(union perf_event *event, 577 bool sample_id_all __maybe_unused) 578 { 579 size_t size; 580 581 perf_event__attr_swap(&event->attr.attr); 582 583 size = event->header.size; 584 size -= (void *)&event->attr.id - (void *)event; 585 mem_bswap_64(event->attr.id, size); 586 } 587 588 static void perf_event__event_type_swap(union perf_event *event, 589 bool sample_id_all __maybe_unused) 590 { 591 event->event_type.event_type.event_id = 592 bswap_64(event->event_type.event_type.event_id); 593 } 594 595 static void perf_event__tracing_data_swap(union perf_event *event, 596 bool sample_id_all __maybe_unused) 597 { 598 event->tracing_data.size = bswap_32(event->tracing_data.size); 599 } 600 601 static void perf_event__auxtrace_info_swap(union perf_event *event, 602 bool sample_id_all __maybe_unused) 603 { 604 size_t size; 605 606 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type); 607 608 size = event->header.size; 609 size -= (void *)&event->auxtrace_info.priv - (void *)event; 610 mem_bswap_64(event->auxtrace_info.priv, size); 611 } 612 613 static void perf_event__auxtrace_swap(union perf_event *event, 614 bool sample_id_all __maybe_unused) 615 { 616 event->auxtrace.size = bswap_64(event->auxtrace.size); 617 event->auxtrace.offset = bswap_64(event->auxtrace.offset); 618 event->auxtrace.reference = bswap_64(event->auxtrace.reference); 619 event->auxtrace.idx = bswap_32(event->auxtrace.idx); 620 event->auxtrace.tid = bswap_32(event->auxtrace.tid); 621 event->auxtrace.cpu = bswap_32(event->auxtrace.cpu); 622 } 623 624 static void perf_event__auxtrace_error_swap(union perf_event *event, 625 bool sample_id_all __maybe_unused) 626 { 627 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type); 628 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code); 629 event->auxtrace_error.cpu = bswap_32(event->auxtrace_error.cpu); 630 event->auxtrace_error.pid = bswap_32(event->auxtrace_error.pid); 631 event->auxtrace_error.tid = bswap_32(event->auxtrace_error.tid); 632 event->auxtrace_error.ip = bswap_64(event->auxtrace_error.ip); 633 } 634 635 typedef void (*perf_event__swap_op)(union perf_event *event, 636 bool sample_id_all); 637 638 static perf_event__swap_op perf_event__swap_ops[] = { 639 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 640 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap, 641 [PERF_RECORD_COMM] = perf_event__comm_swap, 642 [PERF_RECORD_FORK] = perf_event__task_swap, 643 [PERF_RECORD_EXIT] = perf_event__task_swap, 644 [PERF_RECORD_LOST] = perf_event__all64_swap, 645 [PERF_RECORD_READ] = perf_event__read_swap, 646 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap, 647 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap, 648 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 649 [PERF_RECORD_AUX] = perf_event__aux_swap, 650 [PERF_RECORD_ITRACE_START] = perf_event__itrace_start_swap, 651 [PERF_RECORD_LOST_SAMPLES] = perf_event__all64_swap, 652 [PERF_RECORD_SWITCH] = perf_event__switch_swap, 653 [PERF_RECORD_SWITCH_CPU_WIDE] = perf_event__switch_swap, 654 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 655 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 656 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 657 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 658 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap, 659 [PERF_RECORD_AUXTRACE_INFO] = perf_event__auxtrace_info_swap, 660 [PERF_RECORD_AUXTRACE] = perf_event__auxtrace_swap, 661 [PERF_RECORD_AUXTRACE_ERROR] = perf_event__auxtrace_error_swap, 662 [PERF_RECORD_HEADER_MAX] = NULL, 663 }; 664 665 /* 666 * When perf record finishes a pass on every buffers, it records this pseudo 667 * event. 668 * We record the max timestamp t found in the pass n. 669 * Assuming these timestamps are monotonic across cpus, we know that if 670 * a buffer still has events with timestamps below t, they will be all 671 * available and then read in the pass n + 1. 672 * Hence when we start to read the pass n + 2, we can safely flush every 673 * events with timestamps below t. 674 * 675 * ============ PASS n ================= 676 * CPU 0 | CPU 1 677 * | 678 * cnt1 timestamps | cnt2 timestamps 679 * 1 | 2 680 * 2 | 3 681 * - | 4 <--- max recorded 682 * 683 * ============ PASS n + 1 ============== 684 * CPU 0 | CPU 1 685 * | 686 * cnt1 timestamps | cnt2 timestamps 687 * 3 | 5 688 * 4 | 6 689 * 5 | 7 <---- max recorded 690 * 691 * Flush every events below timestamp 4 692 * 693 * ============ PASS n + 2 ============== 694 * CPU 0 | CPU 1 695 * | 696 * cnt1 timestamps | cnt2 timestamps 697 * 6 | 8 698 * 7 | 9 699 * - | 10 700 * 701 * Flush every events below timestamp 7 702 * etc... 703 */ 704 static int process_finished_round(struct perf_tool *tool __maybe_unused, 705 union perf_event *event __maybe_unused, 706 struct ordered_events *oe) 707 { 708 if (dump_trace) 709 fprintf(stdout, "\n"); 710 return ordered_events__flush(oe, OE_FLUSH__ROUND); 711 } 712 713 int perf_session__queue_event(struct perf_session *s, union perf_event *event, 714 struct perf_sample *sample, u64 file_offset) 715 { 716 return ordered_events__queue(&s->ordered_events, event, sample, file_offset); 717 } 718 719 static void callchain__lbr_callstack_printf(struct perf_sample *sample) 720 { 721 struct ip_callchain *callchain = sample->callchain; 722 struct branch_stack *lbr_stack = sample->branch_stack; 723 u64 kernel_callchain_nr = callchain->nr; 724 unsigned int i; 725 726 for (i = 0; i < kernel_callchain_nr; i++) { 727 if (callchain->ips[i] == PERF_CONTEXT_USER) 728 break; 729 } 730 731 if ((i != kernel_callchain_nr) && lbr_stack->nr) { 732 u64 total_nr; 733 /* 734 * LBR callstack can only get user call chain, 735 * i is kernel call chain number, 736 * 1 is PERF_CONTEXT_USER. 737 * 738 * The user call chain is stored in LBR registers. 739 * LBR are pair registers. The caller is stored 740 * in "from" register, while the callee is stored 741 * in "to" register. 742 * For example, there is a call stack 743 * "A"->"B"->"C"->"D". 744 * The LBR registers will recorde like 745 * "C"->"D", "B"->"C", "A"->"B". 746 * So only the first "to" register and all "from" 747 * registers are needed to construct the whole stack. 748 */ 749 total_nr = i + 1 + lbr_stack->nr + 1; 750 kernel_callchain_nr = i + 1; 751 752 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr); 753 754 for (i = 0; i < kernel_callchain_nr; i++) 755 printf("..... %2d: %016" PRIx64 "\n", 756 i, callchain->ips[i]); 757 758 printf("..... %2d: %016" PRIx64 "\n", 759 (int)(kernel_callchain_nr), lbr_stack->entries[0].to); 760 for (i = 0; i < lbr_stack->nr; i++) 761 printf("..... %2d: %016" PRIx64 "\n", 762 (int)(i + kernel_callchain_nr + 1), lbr_stack->entries[i].from); 763 } 764 } 765 766 static void callchain__printf(struct perf_evsel *evsel, 767 struct perf_sample *sample) 768 { 769 unsigned int i; 770 struct ip_callchain *callchain = sample->callchain; 771 772 if (has_branch_callstack(evsel)) 773 callchain__lbr_callstack_printf(sample); 774 775 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr); 776 777 for (i = 0; i < callchain->nr; i++) 778 printf("..... %2d: %016" PRIx64 "\n", 779 i, callchain->ips[i]); 780 } 781 782 static void branch_stack__printf(struct perf_sample *sample) 783 { 784 uint64_t i; 785 786 printf("... branch stack: nr:%" PRIu64 "\n", sample->branch_stack->nr); 787 788 for (i = 0; i < sample->branch_stack->nr; i++) { 789 struct branch_entry *e = &sample->branch_stack->entries[i]; 790 791 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x\n", 792 i, e->from, e->to, 793 e->flags.cycles, 794 e->flags.mispred ? "M" : " ", 795 e->flags.predicted ? "P" : " ", 796 e->flags.abort ? "A" : " ", 797 e->flags.in_tx ? "T" : " ", 798 (unsigned)e->flags.reserved); 799 } 800 } 801 802 static void regs_dump__printf(u64 mask, u64 *regs) 803 { 804 unsigned rid, i = 0; 805 806 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) { 807 u64 val = regs[i++]; 808 809 printf(".... %-5s 0x%" PRIx64 "\n", 810 perf_reg_name(rid), val); 811 } 812 } 813 814 static const char *regs_abi[] = { 815 [PERF_SAMPLE_REGS_ABI_NONE] = "none", 816 [PERF_SAMPLE_REGS_ABI_32] = "32-bit", 817 [PERF_SAMPLE_REGS_ABI_64] = "64-bit", 818 }; 819 820 static inline const char *regs_dump_abi(struct regs_dump *d) 821 { 822 if (d->abi > PERF_SAMPLE_REGS_ABI_64) 823 return "unknown"; 824 825 return regs_abi[d->abi]; 826 } 827 828 static void regs__printf(const char *type, struct regs_dump *regs) 829 { 830 u64 mask = regs->mask; 831 832 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n", 833 type, 834 mask, 835 regs_dump_abi(regs)); 836 837 regs_dump__printf(mask, regs->regs); 838 } 839 840 static void regs_user__printf(struct perf_sample *sample) 841 { 842 struct regs_dump *user_regs = &sample->user_regs; 843 844 if (user_regs->regs) 845 regs__printf("user", user_regs); 846 } 847 848 static void regs_intr__printf(struct perf_sample *sample) 849 { 850 struct regs_dump *intr_regs = &sample->intr_regs; 851 852 if (intr_regs->regs) 853 regs__printf("intr", intr_regs); 854 } 855 856 static void stack_user__printf(struct stack_dump *dump) 857 { 858 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 859 dump->size, dump->offset); 860 } 861 862 static void perf_evlist__print_tstamp(struct perf_evlist *evlist, 863 union perf_event *event, 864 struct perf_sample *sample) 865 { 866 u64 sample_type = __perf_evlist__combined_sample_type(evlist); 867 868 if (event->header.type != PERF_RECORD_SAMPLE && 869 !perf_evlist__sample_id_all(evlist)) { 870 fputs("-1 -1 ", stdout); 871 return; 872 } 873 874 if ((sample_type & PERF_SAMPLE_CPU)) 875 printf("%u ", sample->cpu); 876 877 if (sample_type & PERF_SAMPLE_TIME) 878 printf("%" PRIu64 " ", sample->time); 879 } 880 881 static void sample_read__printf(struct perf_sample *sample, u64 read_format) 882 { 883 printf("... sample_read:\n"); 884 885 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 886 printf("...... time enabled %016" PRIx64 "\n", 887 sample->read.time_enabled); 888 889 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 890 printf("...... time running %016" PRIx64 "\n", 891 sample->read.time_running); 892 893 if (read_format & PERF_FORMAT_GROUP) { 894 u64 i; 895 896 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr); 897 898 for (i = 0; i < sample->read.group.nr; i++) { 899 struct sample_read_value *value; 900 901 value = &sample->read.group.values[i]; 902 printf("..... id %016" PRIx64 903 ", value %016" PRIx64 "\n", 904 value->id, value->value); 905 } 906 } else 907 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n", 908 sample->read.one.id, sample->read.one.value); 909 } 910 911 static void dump_event(struct perf_evlist *evlist, union perf_event *event, 912 u64 file_offset, struct perf_sample *sample) 913 { 914 if (!dump_trace) 915 return; 916 917 printf("\n%#" PRIx64 " [%#x]: event: %d\n", 918 file_offset, event->header.size, event->header.type); 919 920 trace_event(event); 921 922 if (sample) 923 perf_evlist__print_tstamp(evlist, event, sample); 924 925 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 926 event->header.size, perf_event__name(event->header.type)); 927 } 928 929 static void dump_sample(struct perf_evsel *evsel, union perf_event *event, 930 struct perf_sample *sample) 931 { 932 u64 sample_type; 933 934 if (!dump_trace) 935 return; 936 937 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 938 event->header.misc, sample->pid, sample->tid, sample->ip, 939 sample->period, sample->addr); 940 941 sample_type = evsel->attr.sample_type; 942 943 if (sample_type & PERF_SAMPLE_CALLCHAIN) 944 callchain__printf(evsel, sample); 945 946 if ((sample_type & PERF_SAMPLE_BRANCH_STACK) && !has_branch_callstack(evsel)) 947 branch_stack__printf(sample); 948 949 if (sample_type & PERF_SAMPLE_REGS_USER) 950 regs_user__printf(sample); 951 952 if (sample_type & PERF_SAMPLE_REGS_INTR) 953 regs_intr__printf(sample); 954 955 if (sample_type & PERF_SAMPLE_STACK_USER) 956 stack_user__printf(&sample->user_stack); 957 958 if (sample_type & PERF_SAMPLE_WEIGHT) 959 printf("... weight: %" PRIu64 "\n", sample->weight); 960 961 if (sample_type & PERF_SAMPLE_DATA_SRC) 962 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src); 963 964 if (sample_type & PERF_SAMPLE_TRANSACTION) 965 printf("... transaction: %" PRIx64 "\n", sample->transaction); 966 967 if (sample_type & PERF_SAMPLE_READ) 968 sample_read__printf(sample, evsel->attr.read_format); 969 } 970 971 static struct machine *machines__find_for_cpumode(struct machines *machines, 972 union perf_event *event, 973 struct perf_sample *sample) 974 { 975 const u8 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK; 976 struct machine *machine; 977 978 if (perf_guest && 979 ((cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 980 (cpumode == PERF_RECORD_MISC_GUEST_USER))) { 981 u32 pid; 982 983 if (event->header.type == PERF_RECORD_MMAP 984 || event->header.type == PERF_RECORD_MMAP2) 985 pid = event->mmap.pid; 986 else 987 pid = sample->pid; 988 989 machine = machines__find(machines, pid); 990 if (!machine) 991 machine = machines__find(machines, DEFAULT_GUEST_KERNEL_ID); 992 return machine; 993 } 994 995 return &machines->host; 996 } 997 998 static int deliver_sample_value(struct perf_evlist *evlist, 999 struct perf_tool *tool, 1000 union perf_event *event, 1001 struct perf_sample *sample, 1002 struct sample_read_value *v, 1003 struct machine *machine) 1004 { 1005 struct perf_sample_id *sid = perf_evlist__id2sid(evlist, v->id); 1006 1007 if (sid) { 1008 sample->id = v->id; 1009 sample->period = v->value - sid->period; 1010 sid->period = v->value; 1011 } 1012 1013 if (!sid || sid->evsel == NULL) { 1014 ++evlist->stats.nr_unknown_id; 1015 return 0; 1016 } 1017 1018 return tool->sample(tool, event, sample, sid->evsel, machine); 1019 } 1020 1021 static int deliver_sample_group(struct perf_evlist *evlist, 1022 struct perf_tool *tool, 1023 union perf_event *event, 1024 struct perf_sample *sample, 1025 struct machine *machine) 1026 { 1027 int ret = -EINVAL; 1028 u64 i; 1029 1030 for (i = 0; i < sample->read.group.nr; i++) { 1031 ret = deliver_sample_value(evlist, tool, event, sample, 1032 &sample->read.group.values[i], 1033 machine); 1034 if (ret) 1035 break; 1036 } 1037 1038 return ret; 1039 } 1040 1041 static int 1042 perf_evlist__deliver_sample(struct perf_evlist *evlist, 1043 struct perf_tool *tool, 1044 union perf_event *event, 1045 struct perf_sample *sample, 1046 struct perf_evsel *evsel, 1047 struct machine *machine) 1048 { 1049 /* We know evsel != NULL. */ 1050 u64 sample_type = evsel->attr.sample_type; 1051 u64 read_format = evsel->attr.read_format; 1052 1053 /* Standard sample delievery. */ 1054 if (!(sample_type & PERF_SAMPLE_READ)) 1055 return tool->sample(tool, event, sample, evsel, machine); 1056 1057 /* For PERF_SAMPLE_READ we have either single or group mode. */ 1058 if (read_format & PERF_FORMAT_GROUP) 1059 return deliver_sample_group(evlist, tool, event, sample, 1060 machine); 1061 else 1062 return deliver_sample_value(evlist, tool, event, sample, 1063 &sample->read.one, machine); 1064 } 1065 1066 static int machines__deliver_event(struct machines *machines, 1067 struct perf_evlist *evlist, 1068 union perf_event *event, 1069 struct perf_sample *sample, 1070 struct perf_tool *tool, u64 file_offset) 1071 { 1072 struct perf_evsel *evsel; 1073 struct machine *machine; 1074 1075 dump_event(evlist, event, file_offset, sample); 1076 1077 evsel = perf_evlist__id2evsel(evlist, sample->id); 1078 1079 machine = machines__find_for_cpumode(machines, event, sample); 1080 1081 switch (event->header.type) { 1082 case PERF_RECORD_SAMPLE: 1083 if (evsel == NULL) { 1084 ++evlist->stats.nr_unknown_id; 1085 return 0; 1086 } 1087 dump_sample(evsel, event, sample); 1088 if (machine == NULL) { 1089 ++evlist->stats.nr_unprocessable_samples; 1090 return 0; 1091 } 1092 return perf_evlist__deliver_sample(evlist, tool, event, sample, evsel, machine); 1093 case PERF_RECORD_MMAP: 1094 return tool->mmap(tool, event, sample, machine); 1095 case PERF_RECORD_MMAP2: 1096 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT) 1097 ++evlist->stats.nr_proc_map_timeout; 1098 return tool->mmap2(tool, event, sample, machine); 1099 case PERF_RECORD_COMM: 1100 return tool->comm(tool, event, sample, machine); 1101 case PERF_RECORD_FORK: 1102 return tool->fork(tool, event, sample, machine); 1103 case PERF_RECORD_EXIT: 1104 return tool->exit(tool, event, sample, machine); 1105 case PERF_RECORD_LOST: 1106 if (tool->lost == perf_event__process_lost) 1107 evlist->stats.total_lost += event->lost.lost; 1108 return tool->lost(tool, event, sample, machine); 1109 case PERF_RECORD_LOST_SAMPLES: 1110 if (tool->lost_samples == perf_event__process_lost_samples) 1111 evlist->stats.total_lost_samples += event->lost_samples.lost; 1112 return tool->lost_samples(tool, event, sample, machine); 1113 case PERF_RECORD_READ: 1114 return tool->read(tool, event, sample, evsel, machine); 1115 case PERF_RECORD_THROTTLE: 1116 return tool->throttle(tool, event, sample, machine); 1117 case PERF_RECORD_UNTHROTTLE: 1118 return tool->unthrottle(tool, event, sample, machine); 1119 case PERF_RECORD_AUX: 1120 return tool->aux(tool, event, sample, machine); 1121 case PERF_RECORD_ITRACE_START: 1122 return tool->itrace_start(tool, event, sample, machine); 1123 case PERF_RECORD_SWITCH: 1124 case PERF_RECORD_SWITCH_CPU_WIDE: 1125 return tool->context_switch(tool, event, sample, machine); 1126 default: 1127 ++evlist->stats.nr_unknown_events; 1128 return -1; 1129 } 1130 } 1131 1132 static int perf_session__deliver_event(struct perf_session *session, 1133 union perf_event *event, 1134 struct perf_sample *sample, 1135 struct perf_tool *tool, 1136 u64 file_offset) 1137 { 1138 int ret; 1139 1140 ret = auxtrace__process_event(session, event, sample, tool); 1141 if (ret < 0) 1142 return ret; 1143 if (ret > 0) 1144 return 0; 1145 1146 return machines__deliver_event(&session->machines, session->evlist, 1147 event, sample, tool, file_offset); 1148 } 1149 1150 static s64 perf_session__process_user_event(struct perf_session *session, 1151 union perf_event *event, 1152 u64 file_offset) 1153 { 1154 struct ordered_events *oe = &session->ordered_events; 1155 struct perf_tool *tool = session->tool; 1156 int fd = perf_data_file__fd(session->file); 1157 int err; 1158 1159 dump_event(session->evlist, event, file_offset, NULL); 1160 1161 /* These events are processed right away */ 1162 switch (event->header.type) { 1163 case PERF_RECORD_HEADER_ATTR: 1164 err = tool->attr(tool, event, &session->evlist); 1165 if (err == 0) { 1166 perf_session__set_id_hdr_size(session); 1167 perf_session__set_comm_exec(session); 1168 } 1169 return err; 1170 case PERF_RECORD_HEADER_EVENT_TYPE: 1171 /* 1172 * Depreceated, but we need to handle it for sake 1173 * of old data files create in pipe mode. 1174 */ 1175 return 0; 1176 case PERF_RECORD_HEADER_TRACING_DATA: 1177 /* setup for reading amidst mmap */ 1178 lseek(fd, file_offset, SEEK_SET); 1179 return tool->tracing_data(tool, event, session); 1180 case PERF_RECORD_HEADER_BUILD_ID: 1181 return tool->build_id(tool, event, session); 1182 case PERF_RECORD_FINISHED_ROUND: 1183 return tool->finished_round(tool, event, oe); 1184 case PERF_RECORD_ID_INDEX: 1185 return tool->id_index(tool, event, session); 1186 case PERF_RECORD_AUXTRACE_INFO: 1187 return tool->auxtrace_info(tool, event, session); 1188 case PERF_RECORD_AUXTRACE: 1189 /* setup for reading amidst mmap */ 1190 lseek(fd, file_offset + event->header.size, SEEK_SET); 1191 return tool->auxtrace(tool, event, session); 1192 case PERF_RECORD_AUXTRACE_ERROR: 1193 perf_session__auxtrace_error_inc(session, event); 1194 return tool->auxtrace_error(tool, event, session); 1195 default: 1196 return -EINVAL; 1197 } 1198 } 1199 1200 int perf_session__deliver_synth_event(struct perf_session *session, 1201 union perf_event *event, 1202 struct perf_sample *sample) 1203 { 1204 struct perf_evlist *evlist = session->evlist; 1205 struct perf_tool *tool = session->tool; 1206 1207 events_stats__inc(&evlist->stats, event->header.type); 1208 1209 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1210 return perf_session__process_user_event(session, event, 0); 1211 1212 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0); 1213 } 1214 1215 static void event_swap(union perf_event *event, bool sample_id_all) 1216 { 1217 perf_event__swap_op swap; 1218 1219 swap = perf_event__swap_ops[event->header.type]; 1220 if (swap) 1221 swap(event, sample_id_all); 1222 } 1223 1224 int perf_session__peek_event(struct perf_session *session, off_t file_offset, 1225 void *buf, size_t buf_sz, 1226 union perf_event **event_ptr, 1227 struct perf_sample *sample) 1228 { 1229 union perf_event *event; 1230 size_t hdr_sz, rest; 1231 int fd; 1232 1233 if (session->one_mmap && !session->header.needs_swap) { 1234 event = file_offset - session->one_mmap_offset + 1235 session->one_mmap_addr; 1236 goto out_parse_sample; 1237 } 1238 1239 if (perf_data_file__is_pipe(session->file)) 1240 return -1; 1241 1242 fd = perf_data_file__fd(session->file); 1243 hdr_sz = sizeof(struct perf_event_header); 1244 1245 if (buf_sz < hdr_sz) 1246 return -1; 1247 1248 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 || 1249 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz) 1250 return -1; 1251 1252 event = (union perf_event *)buf; 1253 1254 if (session->header.needs_swap) 1255 perf_event_header__bswap(&event->header); 1256 1257 if (event->header.size < hdr_sz || event->header.size > buf_sz) 1258 return -1; 1259 1260 rest = event->header.size - hdr_sz; 1261 1262 if (readn(fd, buf, rest) != (ssize_t)rest) 1263 return -1; 1264 1265 if (session->header.needs_swap) 1266 event_swap(event, perf_evlist__sample_id_all(session->evlist)); 1267 1268 out_parse_sample: 1269 1270 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START && 1271 perf_evlist__parse_sample(session->evlist, event, sample)) 1272 return -1; 1273 1274 *event_ptr = event; 1275 1276 return 0; 1277 } 1278 1279 static s64 perf_session__process_event(struct perf_session *session, 1280 union perf_event *event, u64 file_offset) 1281 { 1282 struct perf_evlist *evlist = session->evlist; 1283 struct perf_tool *tool = session->tool; 1284 struct perf_sample sample; 1285 int ret; 1286 1287 if (session->header.needs_swap) 1288 event_swap(event, perf_evlist__sample_id_all(evlist)); 1289 1290 if (event->header.type >= PERF_RECORD_HEADER_MAX) 1291 return -EINVAL; 1292 1293 events_stats__inc(&evlist->stats, event->header.type); 1294 1295 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1296 return perf_session__process_user_event(session, event, file_offset); 1297 1298 /* 1299 * For all kernel events we get the sample data 1300 */ 1301 ret = perf_evlist__parse_sample(evlist, event, &sample); 1302 if (ret) 1303 return ret; 1304 1305 if (tool->ordered_events) { 1306 ret = perf_session__queue_event(session, event, &sample, file_offset); 1307 if (ret != -ETIME) 1308 return ret; 1309 } 1310 1311 return perf_session__deliver_event(session, event, &sample, tool, 1312 file_offset); 1313 } 1314 1315 void perf_event_header__bswap(struct perf_event_header *hdr) 1316 { 1317 hdr->type = bswap_32(hdr->type); 1318 hdr->misc = bswap_16(hdr->misc); 1319 hdr->size = bswap_16(hdr->size); 1320 } 1321 1322 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1323 { 1324 return machine__findnew_thread(&session->machines.host, -1, pid); 1325 } 1326 1327 static struct thread *perf_session__register_idle_thread(struct perf_session *session) 1328 { 1329 struct thread *thread; 1330 1331 thread = machine__findnew_thread(&session->machines.host, 0, 0); 1332 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) { 1333 pr_err("problem inserting idle task.\n"); 1334 thread = NULL; 1335 } 1336 1337 return thread; 1338 } 1339 1340 static void perf_session__warn_about_errors(const struct perf_session *session) 1341 { 1342 const struct events_stats *stats = &session->evlist->stats; 1343 const struct ordered_events *oe = &session->ordered_events; 1344 1345 if (session->tool->lost == perf_event__process_lost && 1346 stats->nr_events[PERF_RECORD_LOST] != 0) { 1347 ui__warning("Processed %d events and lost %d chunks!\n\n" 1348 "Check IO/CPU overload!\n\n", 1349 stats->nr_events[0], 1350 stats->nr_events[PERF_RECORD_LOST]); 1351 } 1352 1353 if (session->tool->lost_samples == perf_event__process_lost_samples) { 1354 double drop_rate; 1355 1356 drop_rate = (double)stats->total_lost_samples / 1357 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples); 1358 if (drop_rate > 0.05) { 1359 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%% samples!\n\n", 1360 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples, 1361 drop_rate * 100.0); 1362 } 1363 } 1364 1365 if (stats->nr_unknown_events != 0) { 1366 ui__warning("Found %u unknown events!\n\n" 1367 "Is this an older tool processing a perf.data " 1368 "file generated by a more recent tool?\n\n" 1369 "If that is not the case, consider " 1370 "reporting to linux-kernel@vger.kernel.org.\n\n", 1371 stats->nr_unknown_events); 1372 } 1373 1374 if (stats->nr_unknown_id != 0) { 1375 ui__warning("%u samples with id not present in the header\n", 1376 stats->nr_unknown_id); 1377 } 1378 1379 if (stats->nr_invalid_chains != 0) { 1380 ui__warning("Found invalid callchains!\n\n" 1381 "%u out of %u events were discarded for this reason.\n\n" 1382 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 1383 stats->nr_invalid_chains, 1384 stats->nr_events[PERF_RECORD_SAMPLE]); 1385 } 1386 1387 if (stats->nr_unprocessable_samples != 0) { 1388 ui__warning("%u unprocessable samples recorded.\n" 1389 "Do you have a KVM guest running and not using 'perf kvm'?\n", 1390 stats->nr_unprocessable_samples); 1391 } 1392 1393 if (oe->nr_unordered_events != 0) 1394 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events); 1395 1396 events_stats__auxtrace_error_warn(stats); 1397 1398 if (stats->nr_proc_map_timeout != 0) { 1399 ui__warning("%d map information files for pre-existing threads were\n" 1400 "not processed, if there are samples for addresses they\n" 1401 "will not be resolved, you may find out which are these\n" 1402 "threads by running with -v and redirecting the output\n" 1403 "to a file.\n" 1404 "The time limit to process proc map is too short?\n" 1405 "Increase it by --proc-map-timeout\n", 1406 stats->nr_proc_map_timeout); 1407 } 1408 } 1409 1410 static int perf_session__flush_thread_stack(struct thread *thread, 1411 void *p __maybe_unused) 1412 { 1413 return thread_stack__flush(thread); 1414 } 1415 1416 static int perf_session__flush_thread_stacks(struct perf_session *session) 1417 { 1418 return machines__for_each_thread(&session->machines, 1419 perf_session__flush_thread_stack, 1420 NULL); 1421 } 1422 1423 volatile int session_done; 1424 1425 static int __perf_session__process_pipe_events(struct perf_session *session) 1426 { 1427 struct ordered_events *oe = &session->ordered_events; 1428 struct perf_tool *tool = session->tool; 1429 int fd = perf_data_file__fd(session->file); 1430 union perf_event *event; 1431 uint32_t size, cur_size = 0; 1432 void *buf = NULL; 1433 s64 skip = 0; 1434 u64 head; 1435 ssize_t err; 1436 void *p; 1437 1438 perf_tool__fill_defaults(tool); 1439 1440 head = 0; 1441 cur_size = sizeof(union perf_event); 1442 1443 buf = malloc(cur_size); 1444 if (!buf) 1445 return -errno; 1446 more: 1447 event = buf; 1448 err = readn(fd, event, sizeof(struct perf_event_header)); 1449 if (err <= 0) { 1450 if (err == 0) 1451 goto done; 1452 1453 pr_err("failed to read event header\n"); 1454 goto out_err; 1455 } 1456 1457 if (session->header.needs_swap) 1458 perf_event_header__bswap(&event->header); 1459 1460 size = event->header.size; 1461 if (size < sizeof(struct perf_event_header)) { 1462 pr_err("bad event header size\n"); 1463 goto out_err; 1464 } 1465 1466 if (size > cur_size) { 1467 void *new = realloc(buf, size); 1468 if (!new) { 1469 pr_err("failed to allocate memory to read event\n"); 1470 goto out_err; 1471 } 1472 buf = new; 1473 cur_size = size; 1474 event = buf; 1475 } 1476 p = event; 1477 p += sizeof(struct perf_event_header); 1478 1479 if (size - sizeof(struct perf_event_header)) { 1480 err = readn(fd, p, size - sizeof(struct perf_event_header)); 1481 if (err <= 0) { 1482 if (err == 0) { 1483 pr_err("unexpected end of event stream\n"); 1484 goto done; 1485 } 1486 1487 pr_err("failed to read event data\n"); 1488 goto out_err; 1489 } 1490 } 1491 1492 if ((skip = perf_session__process_event(session, event, head)) < 0) { 1493 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 1494 head, event->header.size, event->header.type); 1495 err = -EINVAL; 1496 goto out_err; 1497 } 1498 1499 head += size; 1500 1501 if (skip > 0) 1502 head += skip; 1503 1504 if (!session_done()) 1505 goto more; 1506 done: 1507 /* do the final flush for ordered samples */ 1508 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 1509 if (err) 1510 goto out_err; 1511 err = auxtrace__flush_events(session, tool); 1512 if (err) 1513 goto out_err; 1514 err = perf_session__flush_thread_stacks(session); 1515 out_err: 1516 free(buf); 1517 perf_session__warn_about_errors(session); 1518 ordered_events__free(&session->ordered_events); 1519 auxtrace__free_events(session); 1520 return err; 1521 } 1522 1523 static union perf_event * 1524 fetch_mmaped_event(struct perf_session *session, 1525 u64 head, size_t mmap_size, char *buf) 1526 { 1527 union perf_event *event; 1528 1529 /* 1530 * Ensure we have enough space remaining to read 1531 * the size of the event in the headers. 1532 */ 1533 if (head + sizeof(event->header) > mmap_size) 1534 return NULL; 1535 1536 event = (union perf_event *)(buf + head); 1537 1538 if (session->header.needs_swap) 1539 perf_event_header__bswap(&event->header); 1540 1541 if (head + event->header.size > mmap_size) { 1542 /* We're not fetching the event so swap back again */ 1543 if (session->header.needs_swap) 1544 perf_event_header__bswap(&event->header); 1545 return NULL; 1546 } 1547 1548 return event; 1549 } 1550 1551 /* 1552 * On 64bit we can mmap the data file in one go. No need for tiny mmap 1553 * slices. On 32bit we use 32MB. 1554 */ 1555 #if BITS_PER_LONG == 64 1556 #define MMAP_SIZE ULLONG_MAX 1557 #define NUM_MMAPS 1 1558 #else 1559 #define MMAP_SIZE (32 * 1024 * 1024ULL) 1560 #define NUM_MMAPS 128 1561 #endif 1562 1563 static int __perf_session__process_events(struct perf_session *session, 1564 u64 data_offset, u64 data_size, 1565 u64 file_size) 1566 { 1567 struct ordered_events *oe = &session->ordered_events; 1568 struct perf_tool *tool = session->tool; 1569 int fd = perf_data_file__fd(session->file); 1570 u64 head, page_offset, file_offset, file_pos, size; 1571 int err, mmap_prot, mmap_flags, map_idx = 0; 1572 size_t mmap_size; 1573 char *buf, *mmaps[NUM_MMAPS]; 1574 union perf_event *event; 1575 struct ui_progress prog; 1576 s64 skip; 1577 1578 perf_tool__fill_defaults(tool); 1579 1580 page_offset = page_size * (data_offset / page_size); 1581 file_offset = page_offset; 1582 head = data_offset - page_offset; 1583 1584 if (data_size && (data_offset + data_size < file_size)) 1585 file_size = data_offset + data_size; 1586 1587 ui_progress__init(&prog, file_size, "Processing events..."); 1588 1589 mmap_size = MMAP_SIZE; 1590 if (mmap_size > file_size) { 1591 mmap_size = file_size; 1592 session->one_mmap = true; 1593 } 1594 1595 memset(mmaps, 0, sizeof(mmaps)); 1596 1597 mmap_prot = PROT_READ; 1598 mmap_flags = MAP_SHARED; 1599 1600 if (session->header.needs_swap) { 1601 mmap_prot |= PROT_WRITE; 1602 mmap_flags = MAP_PRIVATE; 1603 } 1604 remap: 1605 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, fd, 1606 file_offset); 1607 if (buf == MAP_FAILED) { 1608 pr_err("failed to mmap file\n"); 1609 err = -errno; 1610 goto out_err; 1611 } 1612 mmaps[map_idx] = buf; 1613 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1); 1614 file_pos = file_offset + head; 1615 if (session->one_mmap) { 1616 session->one_mmap_addr = buf; 1617 session->one_mmap_offset = file_offset; 1618 } 1619 1620 more: 1621 event = fetch_mmaped_event(session, head, mmap_size, buf); 1622 if (!event) { 1623 if (mmaps[map_idx]) { 1624 munmap(mmaps[map_idx], mmap_size); 1625 mmaps[map_idx] = NULL; 1626 } 1627 1628 page_offset = page_size * (head / page_size); 1629 file_offset += page_offset; 1630 head -= page_offset; 1631 goto remap; 1632 } 1633 1634 size = event->header.size; 1635 1636 if (size < sizeof(struct perf_event_header) || 1637 (skip = perf_session__process_event(session, event, file_pos)) < 0) { 1638 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 1639 file_offset + head, event->header.size, 1640 event->header.type); 1641 err = -EINVAL; 1642 goto out_err; 1643 } 1644 1645 if (skip) 1646 size += skip; 1647 1648 head += size; 1649 file_pos += size; 1650 1651 ui_progress__update(&prog, size); 1652 1653 if (session_done()) 1654 goto out; 1655 1656 if (file_pos < file_size) 1657 goto more; 1658 1659 out: 1660 /* do the final flush for ordered samples */ 1661 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 1662 if (err) 1663 goto out_err; 1664 err = auxtrace__flush_events(session, tool); 1665 if (err) 1666 goto out_err; 1667 err = perf_session__flush_thread_stacks(session); 1668 out_err: 1669 ui_progress__finish(); 1670 perf_session__warn_about_errors(session); 1671 ordered_events__free(&session->ordered_events); 1672 auxtrace__free_events(session); 1673 session->one_mmap = false; 1674 return err; 1675 } 1676 1677 int perf_session__process_events(struct perf_session *session) 1678 { 1679 u64 size = perf_data_file__size(session->file); 1680 int err; 1681 1682 if (perf_session__register_idle_thread(session) == NULL) 1683 return -ENOMEM; 1684 1685 if (!perf_data_file__is_pipe(session->file)) 1686 err = __perf_session__process_events(session, 1687 session->header.data_offset, 1688 session->header.data_size, size); 1689 else 1690 err = __perf_session__process_pipe_events(session); 1691 1692 return err; 1693 } 1694 1695 bool perf_session__has_traces(struct perf_session *session, const char *msg) 1696 { 1697 struct perf_evsel *evsel; 1698 1699 evlist__for_each(session->evlist, evsel) { 1700 if (evsel->attr.type == PERF_TYPE_TRACEPOINT) 1701 return true; 1702 } 1703 1704 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 1705 return false; 1706 } 1707 1708 int maps__set_kallsyms_ref_reloc_sym(struct map **maps, 1709 const char *symbol_name, u64 addr) 1710 { 1711 char *bracket; 1712 enum map_type i; 1713 struct ref_reloc_sym *ref; 1714 1715 ref = zalloc(sizeof(struct ref_reloc_sym)); 1716 if (ref == NULL) 1717 return -ENOMEM; 1718 1719 ref->name = strdup(symbol_name); 1720 if (ref->name == NULL) { 1721 free(ref); 1722 return -ENOMEM; 1723 } 1724 1725 bracket = strchr(ref->name, ']'); 1726 if (bracket) 1727 *bracket = '\0'; 1728 1729 ref->addr = addr; 1730 1731 for (i = 0; i < MAP__NR_TYPES; ++i) { 1732 struct kmap *kmap = map__kmap(maps[i]); 1733 1734 if (!kmap) 1735 continue; 1736 kmap->ref_reloc_sym = ref; 1737 } 1738 1739 return 0; 1740 } 1741 1742 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp) 1743 { 1744 return machines__fprintf_dsos(&session->machines, fp); 1745 } 1746 1747 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp, 1748 bool (skip)(struct dso *dso, int parm), int parm) 1749 { 1750 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm); 1751 } 1752 1753 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 1754 { 1755 size_t ret; 1756 const char *msg = ""; 1757 1758 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) 1759 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)"; 1760 1761 ret = fprintf(fp, "\nAggregated stats:%s\n", msg); 1762 1763 ret += events_stats__fprintf(&session->evlist->stats, fp); 1764 return ret; 1765 } 1766 1767 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 1768 { 1769 /* 1770 * FIXME: Here we have to actually print all the machines in this 1771 * session, not just the host... 1772 */ 1773 return machine__fprintf(&session->machines.host, fp); 1774 } 1775 1776 struct perf_evsel *perf_session__find_first_evtype(struct perf_session *session, 1777 unsigned int type) 1778 { 1779 struct perf_evsel *pos; 1780 1781 evlist__for_each(session->evlist, pos) { 1782 if (pos->attr.type == type) 1783 return pos; 1784 } 1785 return NULL; 1786 } 1787 1788 void perf_evsel__print_ip(struct perf_evsel *evsel, struct perf_sample *sample, 1789 struct addr_location *al, 1790 unsigned int print_opts, unsigned int stack_depth) 1791 { 1792 struct callchain_cursor_node *node; 1793 int print_ip = print_opts & PRINT_IP_OPT_IP; 1794 int print_sym = print_opts & PRINT_IP_OPT_SYM; 1795 int print_dso = print_opts & PRINT_IP_OPT_DSO; 1796 int print_symoffset = print_opts & PRINT_IP_OPT_SYMOFFSET; 1797 int print_oneline = print_opts & PRINT_IP_OPT_ONELINE; 1798 int print_srcline = print_opts & PRINT_IP_OPT_SRCLINE; 1799 char s = print_oneline ? ' ' : '\t'; 1800 1801 if (symbol_conf.use_callchain && sample->callchain) { 1802 struct addr_location node_al; 1803 1804 if (thread__resolve_callchain(al->thread, evsel, 1805 sample, NULL, NULL, 1806 PERF_MAX_STACK_DEPTH) != 0) { 1807 if (verbose) 1808 error("Failed to resolve callchain. Skipping\n"); 1809 return; 1810 } 1811 callchain_cursor_commit(&callchain_cursor); 1812 1813 if (print_symoffset) 1814 node_al = *al; 1815 1816 while (stack_depth) { 1817 u64 addr = 0; 1818 1819 node = callchain_cursor_current(&callchain_cursor); 1820 if (!node) 1821 break; 1822 1823 if (node->sym && node->sym->ignore) 1824 goto next; 1825 1826 if (print_ip) 1827 printf("%c%16" PRIx64, s, node->ip); 1828 1829 if (node->map) 1830 addr = node->map->map_ip(node->map, node->ip); 1831 1832 if (print_sym) { 1833 printf(" "); 1834 if (print_symoffset) { 1835 node_al.addr = addr; 1836 node_al.map = node->map; 1837 symbol__fprintf_symname_offs(node->sym, &node_al, stdout); 1838 } else 1839 symbol__fprintf_symname(node->sym, stdout); 1840 } 1841 1842 if (print_dso) { 1843 printf(" ("); 1844 map__fprintf_dsoname(node->map, stdout); 1845 printf(")"); 1846 } 1847 1848 if (print_srcline) 1849 map__fprintf_srcline(node->map, addr, "\n ", 1850 stdout); 1851 1852 if (!print_oneline) 1853 printf("\n"); 1854 1855 stack_depth--; 1856 next: 1857 callchain_cursor_advance(&callchain_cursor); 1858 } 1859 1860 } else { 1861 if (al->sym && al->sym->ignore) 1862 return; 1863 1864 if (print_ip) 1865 printf("%16" PRIx64, sample->ip); 1866 1867 if (print_sym) { 1868 printf(" "); 1869 if (print_symoffset) 1870 symbol__fprintf_symname_offs(al->sym, al, 1871 stdout); 1872 else 1873 symbol__fprintf_symname(al->sym, stdout); 1874 } 1875 1876 if (print_dso) { 1877 printf(" ("); 1878 map__fprintf_dsoname(al->map, stdout); 1879 printf(")"); 1880 } 1881 1882 if (print_srcline) 1883 map__fprintf_srcline(al->map, al->addr, "\n ", stdout); 1884 } 1885 } 1886 1887 int perf_session__cpu_bitmap(struct perf_session *session, 1888 const char *cpu_list, unsigned long *cpu_bitmap) 1889 { 1890 int i, err = -1; 1891 struct cpu_map *map; 1892 1893 for (i = 0; i < PERF_TYPE_MAX; ++i) { 1894 struct perf_evsel *evsel; 1895 1896 evsel = perf_session__find_first_evtype(session, i); 1897 if (!evsel) 1898 continue; 1899 1900 if (!(evsel->attr.sample_type & PERF_SAMPLE_CPU)) { 1901 pr_err("File does not contain CPU events. " 1902 "Remove -c option to proceed.\n"); 1903 return -1; 1904 } 1905 } 1906 1907 map = cpu_map__new(cpu_list); 1908 if (map == NULL) { 1909 pr_err("Invalid cpu_list\n"); 1910 return -1; 1911 } 1912 1913 for (i = 0; i < map->nr; i++) { 1914 int cpu = map->map[i]; 1915 1916 if (cpu >= MAX_NR_CPUS) { 1917 pr_err("Requested CPU %d too large. " 1918 "Consider raising MAX_NR_CPUS\n", cpu); 1919 goto out_delete_map; 1920 } 1921 1922 set_bit(cpu, cpu_bitmap); 1923 } 1924 1925 err = 0; 1926 1927 out_delete_map: 1928 cpu_map__put(map); 1929 return err; 1930 } 1931 1932 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 1933 bool full) 1934 { 1935 struct stat st; 1936 int fd, ret; 1937 1938 if (session == NULL || fp == NULL) 1939 return; 1940 1941 fd = perf_data_file__fd(session->file); 1942 1943 ret = fstat(fd, &st); 1944 if (ret == -1) 1945 return; 1946 1947 fprintf(fp, "# ========\n"); 1948 fprintf(fp, "# captured on: %s", ctime(&st.st_ctime)); 1949 perf_header__fprintf_info(session, fp, full); 1950 fprintf(fp, "# ========\n#\n"); 1951 } 1952 1953 1954 int __perf_session__set_tracepoints_handlers(struct perf_session *session, 1955 const struct perf_evsel_str_handler *assocs, 1956 size_t nr_assocs) 1957 { 1958 struct perf_evsel *evsel; 1959 size_t i; 1960 int err; 1961 1962 for (i = 0; i < nr_assocs; i++) { 1963 /* 1964 * Adding a handler for an event not in the session, 1965 * just ignore it. 1966 */ 1967 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name); 1968 if (evsel == NULL) 1969 continue; 1970 1971 err = -EEXIST; 1972 if (evsel->handler != NULL) 1973 goto out; 1974 evsel->handler = assocs[i].handler; 1975 } 1976 1977 err = 0; 1978 out: 1979 return err; 1980 } 1981 1982 int perf_event__process_id_index(struct perf_tool *tool __maybe_unused, 1983 union perf_event *event, 1984 struct perf_session *session) 1985 { 1986 struct perf_evlist *evlist = session->evlist; 1987 struct id_index_event *ie = &event->id_index; 1988 size_t i, nr, max_nr; 1989 1990 max_nr = (ie->header.size - sizeof(struct id_index_event)) / 1991 sizeof(struct id_index_entry); 1992 nr = ie->nr; 1993 if (nr > max_nr) 1994 return -EINVAL; 1995 1996 if (dump_trace) 1997 fprintf(stdout, " nr: %zu\n", nr); 1998 1999 for (i = 0; i < nr; i++) { 2000 struct id_index_entry *e = &ie->entries[i]; 2001 struct perf_sample_id *sid; 2002 2003 if (dump_trace) { 2004 fprintf(stdout, " ... id: %"PRIu64, e->id); 2005 fprintf(stdout, " idx: %"PRIu64, e->idx); 2006 fprintf(stdout, " cpu: %"PRId64, e->cpu); 2007 fprintf(stdout, " tid: %"PRId64"\n", e->tid); 2008 } 2009 2010 sid = perf_evlist__id2sid(evlist, e->id); 2011 if (!sid) 2012 return -ENOENT; 2013 sid->idx = e->idx; 2014 sid->cpu = e->cpu; 2015 sid->tid = e->tid; 2016 } 2017 return 0; 2018 } 2019 2020 int perf_event__synthesize_id_index(struct perf_tool *tool, 2021 perf_event__handler_t process, 2022 struct perf_evlist *evlist, 2023 struct machine *machine) 2024 { 2025 union perf_event *ev; 2026 struct perf_evsel *evsel; 2027 size_t nr = 0, i = 0, sz, max_nr, n; 2028 int err; 2029 2030 pr_debug2("Synthesizing id index\n"); 2031 2032 max_nr = (UINT16_MAX - sizeof(struct id_index_event)) / 2033 sizeof(struct id_index_entry); 2034 2035 evlist__for_each(evlist, evsel) 2036 nr += evsel->ids; 2037 2038 n = nr > max_nr ? max_nr : nr; 2039 sz = sizeof(struct id_index_event) + n * sizeof(struct id_index_entry); 2040 ev = zalloc(sz); 2041 if (!ev) 2042 return -ENOMEM; 2043 2044 ev->id_index.header.type = PERF_RECORD_ID_INDEX; 2045 ev->id_index.header.size = sz; 2046 ev->id_index.nr = n; 2047 2048 evlist__for_each(evlist, evsel) { 2049 u32 j; 2050 2051 for (j = 0; j < evsel->ids; j++) { 2052 struct id_index_entry *e; 2053 struct perf_sample_id *sid; 2054 2055 if (i >= n) { 2056 err = process(tool, ev, NULL, machine); 2057 if (err) 2058 goto out_err; 2059 nr -= n; 2060 i = 0; 2061 } 2062 2063 e = &ev->id_index.entries[i++]; 2064 2065 e->id = evsel->id[j]; 2066 2067 sid = perf_evlist__id2sid(evlist, e->id); 2068 if (!sid) { 2069 free(ev); 2070 return -ENOENT; 2071 } 2072 2073 e->idx = sid->idx; 2074 e->cpu = sid->cpu; 2075 e->tid = sid->tid; 2076 } 2077 } 2078 2079 sz = sizeof(struct id_index_event) + nr * sizeof(struct id_index_entry); 2080 ev->id_index.header.size = sz; 2081 ev->id_index.nr = nr; 2082 2083 err = process(tool, ev, NULL, machine); 2084 out_err: 2085 free(ev); 2086 2087 return err; 2088 } 2089