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