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