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