1 // SPDX-License-Identifier: GPL-2.0 2 #include <errno.h> 3 #include <inttypes.h> 4 #include <linux/err.h> 5 #include <linux/kernel.h> 6 #include <linux/zalloc.h> 7 #include <api/fs/fs.h> 8 9 #include <byteswap.h> 10 #include <unistd.h> 11 #include <sys/types.h> 12 #include <sys/mman.h> 13 #include <perf/cpumap.h> 14 15 #include "map_symbol.h" 16 #include "branch.h" 17 #include "debug.h" 18 #include "evlist.h" 19 #include "evsel.h" 20 #include "memswap.h" 21 #include "map.h" 22 #include "symbol.h" 23 #include "session.h" 24 #include "tool.h" 25 #include "perf_regs.h" 26 #include "asm/bug.h" 27 #include "auxtrace.h" 28 #include "thread.h" 29 #include "thread-stack.h" 30 #include "sample-raw.h" 31 #include "stat.h" 32 #include "ui/progress.h" 33 #include "../perf.h" 34 #include "arch/common.h" 35 #include "units.h" 36 #include <internal/lib.h> 37 38 #ifdef HAVE_ZSTD_SUPPORT 39 static int perf_session__process_compressed_event(struct perf_session *session, 40 union perf_event *event, u64 file_offset) 41 { 42 void *src; 43 size_t decomp_size, src_size; 44 u64 decomp_last_rem = 0; 45 size_t mmap_len, decomp_len = session->header.env.comp_mmap_len; 46 struct decomp *decomp, *decomp_last = session->decomp_last; 47 48 if (decomp_last) { 49 decomp_last_rem = decomp_last->size - decomp_last->head; 50 decomp_len += decomp_last_rem; 51 } 52 53 mmap_len = sizeof(struct decomp) + decomp_len; 54 decomp = mmap(NULL, mmap_len, PROT_READ|PROT_WRITE, 55 MAP_ANONYMOUS|MAP_PRIVATE, -1, 0); 56 if (decomp == MAP_FAILED) { 57 pr_err("Couldn't allocate memory for decompression\n"); 58 return -1; 59 } 60 61 decomp->file_pos = file_offset; 62 decomp->mmap_len = mmap_len; 63 decomp->head = 0; 64 65 if (decomp_last_rem) { 66 memcpy(decomp->data, &(decomp_last->data[decomp_last->head]), decomp_last_rem); 67 decomp->size = decomp_last_rem; 68 } 69 70 src = (void *)event + sizeof(struct perf_record_compressed); 71 src_size = event->pack.header.size - sizeof(struct perf_record_compressed); 72 73 decomp_size = zstd_decompress_stream(&(session->zstd_data), src, src_size, 74 &(decomp->data[decomp_last_rem]), decomp_len - decomp_last_rem); 75 if (!decomp_size) { 76 munmap(decomp, mmap_len); 77 pr_err("Couldn't decompress data\n"); 78 return -1; 79 } 80 81 decomp->size += decomp_size; 82 83 if (session->decomp == NULL) { 84 session->decomp = decomp; 85 session->decomp_last = decomp; 86 } else { 87 session->decomp_last->next = decomp; 88 session->decomp_last = decomp; 89 } 90 91 pr_debug("decomp (B): %zd to %zd\n", src_size, decomp_size); 92 93 return 0; 94 } 95 #else /* !HAVE_ZSTD_SUPPORT */ 96 #define perf_session__process_compressed_event perf_session__process_compressed_event_stub 97 #endif 98 99 static int perf_session__deliver_event(struct perf_session *session, 100 union perf_event *event, 101 struct perf_tool *tool, 102 u64 file_offset); 103 104 static int perf_session__open(struct perf_session *session) 105 { 106 struct perf_data *data = session->data; 107 108 if (perf_session__read_header(session) < 0) { 109 pr_err("incompatible file format (rerun with -v to learn more)\n"); 110 return -1; 111 } 112 113 if (perf_data__is_pipe(data)) 114 return 0; 115 116 if (perf_header__has_feat(&session->header, HEADER_STAT)) 117 return 0; 118 119 if (!evlist__valid_sample_type(session->evlist)) { 120 pr_err("non matching sample_type\n"); 121 return -1; 122 } 123 124 if (!evlist__valid_sample_id_all(session->evlist)) { 125 pr_err("non matching sample_id_all\n"); 126 return -1; 127 } 128 129 if (!evlist__valid_read_format(session->evlist)) { 130 pr_err("non matching read_format\n"); 131 return -1; 132 } 133 134 return 0; 135 } 136 137 void perf_session__set_id_hdr_size(struct perf_session *session) 138 { 139 u16 id_hdr_size = evlist__id_hdr_size(session->evlist); 140 141 machines__set_id_hdr_size(&session->machines, id_hdr_size); 142 } 143 144 int perf_session__create_kernel_maps(struct perf_session *session) 145 { 146 int ret = machine__create_kernel_maps(&session->machines.host); 147 148 if (ret >= 0) 149 ret = machines__create_guest_kernel_maps(&session->machines); 150 return ret; 151 } 152 153 static void perf_session__destroy_kernel_maps(struct perf_session *session) 154 { 155 machines__destroy_kernel_maps(&session->machines); 156 } 157 158 static bool perf_session__has_comm_exec(struct perf_session *session) 159 { 160 struct evsel *evsel; 161 162 evlist__for_each_entry(session->evlist, evsel) { 163 if (evsel->core.attr.comm_exec) 164 return true; 165 } 166 167 return false; 168 } 169 170 static void perf_session__set_comm_exec(struct perf_session *session) 171 { 172 bool comm_exec = perf_session__has_comm_exec(session); 173 174 machines__set_comm_exec(&session->machines, comm_exec); 175 } 176 177 static int ordered_events__deliver_event(struct ordered_events *oe, 178 struct ordered_event *event) 179 { 180 struct perf_session *session = container_of(oe, struct perf_session, 181 ordered_events); 182 183 return perf_session__deliver_event(session, event->event, 184 session->tool, event->file_offset); 185 } 186 187 struct perf_session *perf_session__new(struct perf_data *data, 188 bool repipe, struct perf_tool *tool) 189 { 190 int ret = -ENOMEM; 191 struct perf_session *session = zalloc(sizeof(*session)); 192 193 if (!session) 194 goto out; 195 196 session->repipe = repipe; 197 session->tool = tool; 198 INIT_LIST_HEAD(&session->auxtrace_index); 199 machines__init(&session->machines); 200 ordered_events__init(&session->ordered_events, 201 ordered_events__deliver_event, NULL); 202 203 perf_env__init(&session->header.env); 204 if (data) { 205 ret = perf_data__open(data); 206 if (ret < 0) 207 goto out_delete; 208 209 session->data = data; 210 211 if (perf_data__is_read(data)) { 212 ret = perf_session__open(session); 213 if (ret < 0) 214 goto out_delete; 215 216 /* 217 * set session attributes that are present in perf.data 218 * but not in pipe-mode. 219 */ 220 if (!data->is_pipe) { 221 perf_session__set_id_hdr_size(session); 222 perf_session__set_comm_exec(session); 223 } 224 225 evlist__init_trace_event_sample_raw(session->evlist); 226 227 /* Open the directory data. */ 228 if (data->is_dir) { 229 ret = perf_data__open_dir(data); 230 if (ret) 231 goto out_delete; 232 } 233 234 if (!symbol_conf.kallsyms_name && 235 !symbol_conf.vmlinux_name) 236 symbol_conf.kallsyms_name = perf_data__kallsyms_name(data); 237 } 238 } else { 239 session->machines.host.env = &perf_env; 240 } 241 242 session->machines.host.single_address_space = 243 perf_env__single_address_space(session->machines.host.env); 244 245 if (!data || perf_data__is_write(data)) { 246 /* 247 * In O_RDONLY mode this will be performed when reading the 248 * kernel MMAP event, in perf_event__process_mmap(). 249 */ 250 if (perf_session__create_kernel_maps(session) < 0) 251 pr_warning("Cannot read kernel map\n"); 252 } 253 254 /* 255 * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is 256 * processed, so evlist__sample_id_all is not meaningful here. 257 */ 258 if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps && 259 tool->ordered_events && !evlist__sample_id_all(session->evlist)) { 260 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 261 tool->ordered_events = false; 262 } 263 264 return session; 265 266 out_delete: 267 perf_session__delete(session); 268 out: 269 return ERR_PTR(ret); 270 } 271 272 static void perf_session__delete_threads(struct perf_session *session) 273 { 274 machine__delete_threads(&session->machines.host); 275 } 276 277 static void perf_session__release_decomp_events(struct perf_session *session) 278 { 279 struct decomp *next, *decomp; 280 size_t mmap_len; 281 next = session->decomp; 282 do { 283 decomp = next; 284 if (decomp == NULL) 285 break; 286 next = decomp->next; 287 mmap_len = decomp->mmap_len; 288 munmap(decomp, mmap_len); 289 } while (1); 290 } 291 292 void perf_session__delete(struct perf_session *session) 293 { 294 if (session == NULL) 295 return; 296 auxtrace__free(session); 297 auxtrace_index__free(&session->auxtrace_index); 298 perf_session__destroy_kernel_maps(session); 299 perf_session__delete_threads(session); 300 perf_session__release_decomp_events(session); 301 perf_env__exit(&session->header.env); 302 machines__exit(&session->machines); 303 if (session->data) 304 perf_data__close(session->data); 305 free(session); 306 } 307 308 static int process_event_synth_tracing_data_stub(struct perf_session *session 309 __maybe_unused, 310 union perf_event *event 311 __maybe_unused) 312 { 313 dump_printf(": unhandled!\n"); 314 return 0; 315 } 316 317 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused, 318 union perf_event *event __maybe_unused, 319 struct evlist **pevlist 320 __maybe_unused) 321 { 322 dump_printf(": unhandled!\n"); 323 return 0; 324 } 325 326 static int process_event_synth_event_update_stub(struct perf_tool *tool __maybe_unused, 327 union perf_event *event __maybe_unused, 328 struct evlist **pevlist 329 __maybe_unused) 330 { 331 if (dump_trace) 332 perf_event__fprintf_event_update(event, stdout); 333 334 dump_printf(": unhandled!\n"); 335 return 0; 336 } 337 338 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused, 339 union perf_event *event __maybe_unused, 340 struct perf_sample *sample __maybe_unused, 341 struct evsel *evsel __maybe_unused, 342 struct machine *machine __maybe_unused) 343 { 344 dump_printf(": unhandled!\n"); 345 return 0; 346 } 347 348 static int process_event_stub(struct perf_tool *tool __maybe_unused, 349 union perf_event *event __maybe_unused, 350 struct perf_sample *sample __maybe_unused, 351 struct machine *machine __maybe_unused) 352 { 353 dump_printf(": unhandled!\n"); 354 return 0; 355 } 356 357 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused, 358 union perf_event *event __maybe_unused, 359 struct ordered_events *oe __maybe_unused) 360 { 361 dump_printf(": unhandled!\n"); 362 return 0; 363 } 364 365 static int process_finished_round(struct perf_tool *tool, 366 union perf_event *event, 367 struct ordered_events *oe); 368 369 static int skipn(int fd, off_t n) 370 { 371 char buf[4096]; 372 ssize_t ret; 373 374 while (n > 0) { 375 ret = read(fd, buf, min(n, (off_t)sizeof(buf))); 376 if (ret <= 0) 377 return ret; 378 n -= ret; 379 } 380 381 return 0; 382 } 383 384 static s64 process_event_auxtrace_stub(struct perf_session *session __maybe_unused, 385 union perf_event *event) 386 { 387 dump_printf(": unhandled!\n"); 388 if (perf_data__is_pipe(session->data)) 389 skipn(perf_data__fd(session->data), event->auxtrace.size); 390 return event->auxtrace.size; 391 } 392 393 static int process_event_op2_stub(struct perf_session *session __maybe_unused, 394 union perf_event *event __maybe_unused) 395 { 396 dump_printf(": unhandled!\n"); 397 return 0; 398 } 399 400 401 static 402 int process_event_thread_map_stub(struct perf_session *session __maybe_unused, 403 union perf_event *event __maybe_unused) 404 { 405 if (dump_trace) 406 perf_event__fprintf_thread_map(event, stdout); 407 408 dump_printf(": unhandled!\n"); 409 return 0; 410 } 411 412 static 413 int process_event_cpu_map_stub(struct perf_session *session __maybe_unused, 414 union perf_event *event __maybe_unused) 415 { 416 if (dump_trace) 417 perf_event__fprintf_cpu_map(event, stdout); 418 419 dump_printf(": unhandled!\n"); 420 return 0; 421 } 422 423 static 424 int process_event_stat_config_stub(struct perf_session *session __maybe_unused, 425 union perf_event *event __maybe_unused) 426 { 427 if (dump_trace) 428 perf_event__fprintf_stat_config(event, stdout); 429 430 dump_printf(": unhandled!\n"); 431 return 0; 432 } 433 434 static int process_stat_stub(struct perf_session *perf_session __maybe_unused, 435 union perf_event *event) 436 { 437 if (dump_trace) 438 perf_event__fprintf_stat(event, stdout); 439 440 dump_printf(": unhandled!\n"); 441 return 0; 442 } 443 444 static int process_stat_round_stub(struct perf_session *perf_session __maybe_unused, 445 union perf_event *event) 446 { 447 if (dump_trace) 448 perf_event__fprintf_stat_round(event, stdout); 449 450 dump_printf(": unhandled!\n"); 451 return 0; 452 } 453 454 static int perf_session__process_compressed_event_stub(struct perf_session *session __maybe_unused, 455 union perf_event *event __maybe_unused, 456 u64 file_offset __maybe_unused) 457 { 458 dump_printf(": unhandled!\n"); 459 return 0; 460 } 461 462 void perf_tool__fill_defaults(struct perf_tool *tool) 463 { 464 if (tool->sample == NULL) 465 tool->sample = process_event_sample_stub; 466 if (tool->mmap == NULL) 467 tool->mmap = process_event_stub; 468 if (tool->mmap2 == NULL) 469 tool->mmap2 = process_event_stub; 470 if (tool->comm == NULL) 471 tool->comm = process_event_stub; 472 if (tool->namespaces == NULL) 473 tool->namespaces = process_event_stub; 474 if (tool->cgroup == NULL) 475 tool->cgroup = process_event_stub; 476 if (tool->fork == NULL) 477 tool->fork = process_event_stub; 478 if (tool->exit == NULL) 479 tool->exit = process_event_stub; 480 if (tool->lost == NULL) 481 tool->lost = perf_event__process_lost; 482 if (tool->lost_samples == NULL) 483 tool->lost_samples = perf_event__process_lost_samples; 484 if (tool->aux == NULL) 485 tool->aux = perf_event__process_aux; 486 if (tool->itrace_start == NULL) 487 tool->itrace_start = perf_event__process_itrace_start; 488 if (tool->context_switch == NULL) 489 tool->context_switch = perf_event__process_switch; 490 if (tool->ksymbol == NULL) 491 tool->ksymbol = perf_event__process_ksymbol; 492 if (tool->bpf == NULL) 493 tool->bpf = perf_event__process_bpf; 494 if (tool->text_poke == NULL) 495 tool->text_poke = perf_event__process_text_poke; 496 if (tool->read == NULL) 497 tool->read = process_event_sample_stub; 498 if (tool->throttle == NULL) 499 tool->throttle = process_event_stub; 500 if (tool->unthrottle == NULL) 501 tool->unthrottle = process_event_stub; 502 if (tool->attr == NULL) 503 tool->attr = process_event_synth_attr_stub; 504 if (tool->event_update == NULL) 505 tool->event_update = process_event_synth_event_update_stub; 506 if (tool->tracing_data == NULL) 507 tool->tracing_data = process_event_synth_tracing_data_stub; 508 if (tool->build_id == NULL) 509 tool->build_id = process_event_op2_stub; 510 if (tool->finished_round == NULL) { 511 if (tool->ordered_events) 512 tool->finished_round = process_finished_round; 513 else 514 tool->finished_round = process_finished_round_stub; 515 } 516 if (tool->id_index == NULL) 517 tool->id_index = process_event_op2_stub; 518 if (tool->auxtrace_info == NULL) 519 tool->auxtrace_info = process_event_op2_stub; 520 if (tool->auxtrace == NULL) 521 tool->auxtrace = process_event_auxtrace_stub; 522 if (tool->auxtrace_error == NULL) 523 tool->auxtrace_error = process_event_op2_stub; 524 if (tool->thread_map == NULL) 525 tool->thread_map = process_event_thread_map_stub; 526 if (tool->cpu_map == NULL) 527 tool->cpu_map = process_event_cpu_map_stub; 528 if (tool->stat_config == NULL) 529 tool->stat_config = process_event_stat_config_stub; 530 if (tool->stat == NULL) 531 tool->stat = process_stat_stub; 532 if (tool->stat_round == NULL) 533 tool->stat_round = process_stat_round_stub; 534 if (tool->time_conv == NULL) 535 tool->time_conv = process_event_op2_stub; 536 if (tool->feature == NULL) 537 tool->feature = process_event_op2_stub; 538 if (tool->compressed == NULL) 539 tool->compressed = perf_session__process_compressed_event; 540 } 541 542 static void swap_sample_id_all(union perf_event *event, void *data) 543 { 544 void *end = (void *) event + event->header.size; 545 int size = end - data; 546 547 BUG_ON(size % sizeof(u64)); 548 mem_bswap_64(data, size); 549 } 550 551 static void perf_event__all64_swap(union perf_event *event, 552 bool sample_id_all __maybe_unused) 553 { 554 struct perf_event_header *hdr = &event->header; 555 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 556 } 557 558 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all) 559 { 560 event->comm.pid = bswap_32(event->comm.pid); 561 event->comm.tid = bswap_32(event->comm.tid); 562 563 if (sample_id_all) { 564 void *data = &event->comm.comm; 565 566 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 567 swap_sample_id_all(event, data); 568 } 569 } 570 571 static void perf_event__mmap_swap(union perf_event *event, 572 bool sample_id_all) 573 { 574 event->mmap.pid = bswap_32(event->mmap.pid); 575 event->mmap.tid = bswap_32(event->mmap.tid); 576 event->mmap.start = bswap_64(event->mmap.start); 577 event->mmap.len = bswap_64(event->mmap.len); 578 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 579 580 if (sample_id_all) { 581 void *data = &event->mmap.filename; 582 583 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 584 swap_sample_id_all(event, data); 585 } 586 } 587 588 static void perf_event__mmap2_swap(union perf_event *event, 589 bool sample_id_all) 590 { 591 event->mmap2.pid = bswap_32(event->mmap2.pid); 592 event->mmap2.tid = bswap_32(event->mmap2.tid); 593 event->mmap2.start = bswap_64(event->mmap2.start); 594 event->mmap2.len = bswap_64(event->mmap2.len); 595 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff); 596 597 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID)) { 598 event->mmap2.maj = bswap_32(event->mmap2.maj); 599 event->mmap2.min = bswap_32(event->mmap2.min); 600 event->mmap2.ino = bswap_64(event->mmap2.ino); 601 event->mmap2.ino_generation = bswap_64(event->mmap2.ino_generation); 602 } 603 604 if (sample_id_all) { 605 void *data = &event->mmap2.filename; 606 607 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 608 swap_sample_id_all(event, data); 609 } 610 } 611 static void perf_event__task_swap(union perf_event *event, bool sample_id_all) 612 { 613 event->fork.pid = bswap_32(event->fork.pid); 614 event->fork.tid = bswap_32(event->fork.tid); 615 event->fork.ppid = bswap_32(event->fork.ppid); 616 event->fork.ptid = bswap_32(event->fork.ptid); 617 event->fork.time = bswap_64(event->fork.time); 618 619 if (sample_id_all) 620 swap_sample_id_all(event, &event->fork + 1); 621 } 622 623 static void perf_event__read_swap(union perf_event *event, bool sample_id_all) 624 { 625 event->read.pid = bswap_32(event->read.pid); 626 event->read.tid = bswap_32(event->read.tid); 627 event->read.value = bswap_64(event->read.value); 628 event->read.time_enabled = bswap_64(event->read.time_enabled); 629 event->read.time_running = bswap_64(event->read.time_running); 630 event->read.id = bswap_64(event->read.id); 631 632 if (sample_id_all) 633 swap_sample_id_all(event, &event->read + 1); 634 } 635 636 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all) 637 { 638 event->aux.aux_offset = bswap_64(event->aux.aux_offset); 639 event->aux.aux_size = bswap_64(event->aux.aux_size); 640 event->aux.flags = bswap_64(event->aux.flags); 641 642 if (sample_id_all) 643 swap_sample_id_all(event, &event->aux + 1); 644 } 645 646 static void perf_event__itrace_start_swap(union perf_event *event, 647 bool sample_id_all) 648 { 649 event->itrace_start.pid = bswap_32(event->itrace_start.pid); 650 event->itrace_start.tid = bswap_32(event->itrace_start.tid); 651 652 if (sample_id_all) 653 swap_sample_id_all(event, &event->itrace_start + 1); 654 } 655 656 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all) 657 { 658 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) { 659 event->context_switch.next_prev_pid = 660 bswap_32(event->context_switch.next_prev_pid); 661 event->context_switch.next_prev_tid = 662 bswap_32(event->context_switch.next_prev_tid); 663 } 664 665 if (sample_id_all) 666 swap_sample_id_all(event, &event->context_switch + 1); 667 } 668 669 static void perf_event__text_poke_swap(union perf_event *event, bool sample_id_all) 670 { 671 event->text_poke.addr = bswap_64(event->text_poke.addr); 672 event->text_poke.old_len = bswap_16(event->text_poke.old_len); 673 event->text_poke.new_len = bswap_16(event->text_poke.new_len); 674 675 if (sample_id_all) { 676 size_t len = sizeof(event->text_poke.old_len) + 677 sizeof(event->text_poke.new_len) + 678 event->text_poke.old_len + 679 event->text_poke.new_len; 680 void *data = &event->text_poke.old_len; 681 682 data += PERF_ALIGN(len, sizeof(u64)); 683 swap_sample_id_all(event, data); 684 } 685 } 686 687 static void perf_event__throttle_swap(union perf_event *event, 688 bool sample_id_all) 689 { 690 event->throttle.time = bswap_64(event->throttle.time); 691 event->throttle.id = bswap_64(event->throttle.id); 692 event->throttle.stream_id = bswap_64(event->throttle.stream_id); 693 694 if (sample_id_all) 695 swap_sample_id_all(event, &event->throttle + 1); 696 } 697 698 static void perf_event__namespaces_swap(union perf_event *event, 699 bool sample_id_all) 700 { 701 u64 i; 702 703 event->namespaces.pid = bswap_32(event->namespaces.pid); 704 event->namespaces.tid = bswap_32(event->namespaces.tid); 705 event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces); 706 707 for (i = 0; i < event->namespaces.nr_namespaces; i++) { 708 struct perf_ns_link_info *ns = &event->namespaces.link_info[i]; 709 710 ns->dev = bswap_64(ns->dev); 711 ns->ino = bswap_64(ns->ino); 712 } 713 714 if (sample_id_all) 715 swap_sample_id_all(event, &event->namespaces.link_info[i]); 716 } 717 718 static void perf_event__cgroup_swap(union perf_event *event, bool sample_id_all) 719 { 720 event->cgroup.id = bswap_64(event->cgroup.id); 721 722 if (sample_id_all) { 723 void *data = &event->cgroup.path; 724 725 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 726 swap_sample_id_all(event, data); 727 } 728 } 729 730 static u8 revbyte(u8 b) 731 { 732 int rev = (b >> 4) | ((b & 0xf) << 4); 733 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2); 734 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1); 735 return (u8) rev; 736 } 737 738 /* 739 * XXX this is hack in attempt to carry flags bitfield 740 * through endian village. ABI says: 741 * 742 * Bit-fields are allocated from right to left (least to most significant) 743 * on little-endian implementations and from left to right (most to least 744 * significant) on big-endian implementations. 745 * 746 * The above seems to be byte specific, so we need to reverse each 747 * byte of the bitfield. 'Internet' also says this might be implementation 748 * specific and we probably need proper fix and carry perf_event_attr 749 * bitfield flags in separate data file FEAT_ section. Thought this seems 750 * to work for now. 751 */ 752 static void swap_bitfield(u8 *p, unsigned len) 753 { 754 unsigned i; 755 756 for (i = 0; i < len; i++) { 757 *p = revbyte(*p); 758 p++; 759 } 760 } 761 762 /* exported for swapping attributes in file header */ 763 void perf_event__attr_swap(struct perf_event_attr *attr) 764 { 765 attr->type = bswap_32(attr->type); 766 attr->size = bswap_32(attr->size); 767 768 #define bswap_safe(f, n) \ 769 (attr->size > (offsetof(struct perf_event_attr, f) + \ 770 sizeof(attr->f) * (n))) 771 #define bswap_field(f, sz) \ 772 do { \ 773 if (bswap_safe(f, 0)) \ 774 attr->f = bswap_##sz(attr->f); \ 775 } while(0) 776 #define bswap_field_16(f) bswap_field(f, 16) 777 #define bswap_field_32(f) bswap_field(f, 32) 778 #define bswap_field_64(f) bswap_field(f, 64) 779 780 bswap_field_64(config); 781 bswap_field_64(sample_period); 782 bswap_field_64(sample_type); 783 bswap_field_64(read_format); 784 bswap_field_32(wakeup_events); 785 bswap_field_32(bp_type); 786 bswap_field_64(bp_addr); 787 bswap_field_64(bp_len); 788 bswap_field_64(branch_sample_type); 789 bswap_field_64(sample_regs_user); 790 bswap_field_32(sample_stack_user); 791 bswap_field_32(aux_watermark); 792 bswap_field_16(sample_max_stack); 793 bswap_field_32(aux_sample_size); 794 795 /* 796 * After read_format are bitfields. Check read_format because 797 * we are unable to use offsetof on bitfield. 798 */ 799 if (bswap_safe(read_format, 1)) 800 swap_bitfield((u8 *) (&attr->read_format + 1), 801 sizeof(u64)); 802 #undef bswap_field_64 803 #undef bswap_field_32 804 #undef bswap_field 805 #undef bswap_safe 806 } 807 808 static void perf_event__hdr_attr_swap(union perf_event *event, 809 bool sample_id_all __maybe_unused) 810 { 811 size_t size; 812 813 perf_event__attr_swap(&event->attr.attr); 814 815 size = event->header.size; 816 size -= (void *)&event->attr.id - (void *)event; 817 mem_bswap_64(event->attr.id, size); 818 } 819 820 static void perf_event__event_update_swap(union perf_event *event, 821 bool sample_id_all __maybe_unused) 822 { 823 event->event_update.type = bswap_64(event->event_update.type); 824 event->event_update.id = bswap_64(event->event_update.id); 825 } 826 827 static void perf_event__event_type_swap(union perf_event *event, 828 bool sample_id_all __maybe_unused) 829 { 830 event->event_type.event_type.event_id = 831 bswap_64(event->event_type.event_type.event_id); 832 } 833 834 static void perf_event__tracing_data_swap(union perf_event *event, 835 bool sample_id_all __maybe_unused) 836 { 837 event->tracing_data.size = bswap_32(event->tracing_data.size); 838 } 839 840 static void perf_event__auxtrace_info_swap(union perf_event *event, 841 bool sample_id_all __maybe_unused) 842 { 843 size_t size; 844 845 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type); 846 847 size = event->header.size; 848 size -= (void *)&event->auxtrace_info.priv - (void *)event; 849 mem_bswap_64(event->auxtrace_info.priv, size); 850 } 851 852 static void perf_event__auxtrace_swap(union perf_event *event, 853 bool sample_id_all __maybe_unused) 854 { 855 event->auxtrace.size = bswap_64(event->auxtrace.size); 856 event->auxtrace.offset = bswap_64(event->auxtrace.offset); 857 event->auxtrace.reference = bswap_64(event->auxtrace.reference); 858 event->auxtrace.idx = bswap_32(event->auxtrace.idx); 859 event->auxtrace.tid = bswap_32(event->auxtrace.tid); 860 event->auxtrace.cpu = bswap_32(event->auxtrace.cpu); 861 } 862 863 static void perf_event__auxtrace_error_swap(union perf_event *event, 864 bool sample_id_all __maybe_unused) 865 { 866 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type); 867 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code); 868 event->auxtrace_error.cpu = bswap_32(event->auxtrace_error.cpu); 869 event->auxtrace_error.pid = bswap_32(event->auxtrace_error.pid); 870 event->auxtrace_error.tid = bswap_32(event->auxtrace_error.tid); 871 event->auxtrace_error.fmt = bswap_32(event->auxtrace_error.fmt); 872 event->auxtrace_error.ip = bswap_64(event->auxtrace_error.ip); 873 if (event->auxtrace_error.fmt) 874 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time); 875 } 876 877 static void perf_event__thread_map_swap(union perf_event *event, 878 bool sample_id_all __maybe_unused) 879 { 880 unsigned i; 881 882 event->thread_map.nr = bswap_64(event->thread_map.nr); 883 884 for (i = 0; i < event->thread_map.nr; i++) 885 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid); 886 } 887 888 static void perf_event__cpu_map_swap(union perf_event *event, 889 bool sample_id_all __maybe_unused) 890 { 891 struct perf_record_cpu_map_data *data = &event->cpu_map.data; 892 struct cpu_map_entries *cpus; 893 struct perf_record_record_cpu_map *mask; 894 unsigned i; 895 896 data->type = bswap_64(data->type); 897 898 switch (data->type) { 899 case PERF_CPU_MAP__CPUS: 900 cpus = (struct cpu_map_entries *)data->data; 901 902 cpus->nr = bswap_16(cpus->nr); 903 904 for (i = 0; i < cpus->nr; i++) 905 cpus->cpu[i] = bswap_16(cpus->cpu[i]); 906 break; 907 case PERF_CPU_MAP__MASK: 908 mask = (struct perf_record_record_cpu_map *)data->data; 909 910 mask->nr = bswap_16(mask->nr); 911 mask->long_size = bswap_16(mask->long_size); 912 913 switch (mask->long_size) { 914 case 4: mem_bswap_32(&mask->mask, mask->nr); break; 915 case 8: mem_bswap_64(&mask->mask, mask->nr); break; 916 default: 917 pr_err("cpu_map swap: unsupported long size\n"); 918 } 919 default: 920 break; 921 } 922 } 923 924 static void perf_event__stat_config_swap(union perf_event *event, 925 bool sample_id_all __maybe_unused) 926 { 927 u64 size; 928 929 size = event->stat_config.nr * sizeof(event->stat_config.data[0]); 930 size += 1; /* nr item itself */ 931 mem_bswap_64(&event->stat_config.nr, size); 932 } 933 934 static void perf_event__stat_swap(union perf_event *event, 935 bool sample_id_all __maybe_unused) 936 { 937 event->stat.id = bswap_64(event->stat.id); 938 event->stat.thread = bswap_32(event->stat.thread); 939 event->stat.cpu = bswap_32(event->stat.cpu); 940 event->stat.val = bswap_64(event->stat.val); 941 event->stat.ena = bswap_64(event->stat.ena); 942 event->stat.run = bswap_64(event->stat.run); 943 } 944 945 static void perf_event__stat_round_swap(union perf_event *event, 946 bool sample_id_all __maybe_unused) 947 { 948 event->stat_round.type = bswap_64(event->stat_round.type); 949 event->stat_round.time = bswap_64(event->stat_round.time); 950 } 951 952 typedef void (*perf_event__swap_op)(union perf_event *event, 953 bool sample_id_all); 954 955 static perf_event__swap_op perf_event__swap_ops[] = { 956 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 957 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap, 958 [PERF_RECORD_COMM] = perf_event__comm_swap, 959 [PERF_RECORD_FORK] = perf_event__task_swap, 960 [PERF_RECORD_EXIT] = perf_event__task_swap, 961 [PERF_RECORD_LOST] = perf_event__all64_swap, 962 [PERF_RECORD_READ] = perf_event__read_swap, 963 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap, 964 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap, 965 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 966 [PERF_RECORD_AUX] = perf_event__aux_swap, 967 [PERF_RECORD_ITRACE_START] = perf_event__itrace_start_swap, 968 [PERF_RECORD_LOST_SAMPLES] = perf_event__all64_swap, 969 [PERF_RECORD_SWITCH] = perf_event__switch_swap, 970 [PERF_RECORD_SWITCH_CPU_WIDE] = perf_event__switch_swap, 971 [PERF_RECORD_NAMESPACES] = perf_event__namespaces_swap, 972 [PERF_RECORD_CGROUP] = perf_event__cgroup_swap, 973 [PERF_RECORD_TEXT_POKE] = perf_event__text_poke_swap, 974 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 975 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 976 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 977 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 978 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap, 979 [PERF_RECORD_AUXTRACE_INFO] = perf_event__auxtrace_info_swap, 980 [PERF_RECORD_AUXTRACE] = perf_event__auxtrace_swap, 981 [PERF_RECORD_AUXTRACE_ERROR] = perf_event__auxtrace_error_swap, 982 [PERF_RECORD_THREAD_MAP] = perf_event__thread_map_swap, 983 [PERF_RECORD_CPU_MAP] = perf_event__cpu_map_swap, 984 [PERF_RECORD_STAT_CONFIG] = perf_event__stat_config_swap, 985 [PERF_RECORD_STAT] = perf_event__stat_swap, 986 [PERF_RECORD_STAT_ROUND] = perf_event__stat_round_swap, 987 [PERF_RECORD_EVENT_UPDATE] = perf_event__event_update_swap, 988 [PERF_RECORD_TIME_CONV] = perf_event__all64_swap, 989 [PERF_RECORD_HEADER_MAX] = NULL, 990 }; 991 992 /* 993 * When perf record finishes a pass on every buffers, it records this pseudo 994 * event. 995 * We record the max timestamp t found in the pass n. 996 * Assuming these timestamps are monotonic across cpus, we know that if 997 * a buffer still has events with timestamps below t, they will be all 998 * available and then read in the pass n + 1. 999 * Hence when we start to read the pass n + 2, we can safely flush every 1000 * events with timestamps below t. 1001 * 1002 * ============ PASS n ================= 1003 * CPU 0 | CPU 1 1004 * | 1005 * cnt1 timestamps | cnt2 timestamps 1006 * 1 | 2 1007 * 2 | 3 1008 * - | 4 <--- max recorded 1009 * 1010 * ============ PASS n + 1 ============== 1011 * CPU 0 | CPU 1 1012 * | 1013 * cnt1 timestamps | cnt2 timestamps 1014 * 3 | 5 1015 * 4 | 6 1016 * 5 | 7 <---- max recorded 1017 * 1018 * Flush every events below timestamp 4 1019 * 1020 * ============ PASS n + 2 ============== 1021 * CPU 0 | CPU 1 1022 * | 1023 * cnt1 timestamps | cnt2 timestamps 1024 * 6 | 8 1025 * 7 | 9 1026 * - | 10 1027 * 1028 * Flush every events below timestamp 7 1029 * etc... 1030 */ 1031 static int process_finished_round(struct perf_tool *tool __maybe_unused, 1032 union perf_event *event __maybe_unused, 1033 struct ordered_events *oe) 1034 { 1035 if (dump_trace) 1036 fprintf(stdout, "\n"); 1037 return ordered_events__flush(oe, OE_FLUSH__ROUND); 1038 } 1039 1040 int perf_session__queue_event(struct perf_session *s, union perf_event *event, 1041 u64 timestamp, u64 file_offset) 1042 { 1043 return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset); 1044 } 1045 1046 static void callchain__lbr_callstack_printf(struct perf_sample *sample) 1047 { 1048 struct ip_callchain *callchain = sample->callchain; 1049 struct branch_stack *lbr_stack = sample->branch_stack; 1050 struct branch_entry *entries = perf_sample__branch_entries(sample); 1051 u64 kernel_callchain_nr = callchain->nr; 1052 unsigned int i; 1053 1054 for (i = 0; i < kernel_callchain_nr; i++) { 1055 if (callchain->ips[i] == PERF_CONTEXT_USER) 1056 break; 1057 } 1058 1059 if ((i != kernel_callchain_nr) && lbr_stack->nr) { 1060 u64 total_nr; 1061 /* 1062 * LBR callstack can only get user call chain, 1063 * i is kernel call chain number, 1064 * 1 is PERF_CONTEXT_USER. 1065 * 1066 * The user call chain is stored in LBR registers. 1067 * LBR are pair registers. The caller is stored 1068 * in "from" register, while the callee is stored 1069 * in "to" register. 1070 * For example, there is a call stack 1071 * "A"->"B"->"C"->"D". 1072 * The LBR registers will recorde like 1073 * "C"->"D", "B"->"C", "A"->"B". 1074 * So only the first "to" register and all "from" 1075 * registers are needed to construct the whole stack. 1076 */ 1077 total_nr = i + 1 + lbr_stack->nr + 1; 1078 kernel_callchain_nr = i + 1; 1079 1080 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr); 1081 1082 for (i = 0; i < kernel_callchain_nr; i++) 1083 printf("..... %2d: %016" PRIx64 "\n", 1084 i, callchain->ips[i]); 1085 1086 printf("..... %2d: %016" PRIx64 "\n", 1087 (int)(kernel_callchain_nr), entries[0].to); 1088 for (i = 0; i < lbr_stack->nr; i++) 1089 printf("..... %2d: %016" PRIx64 "\n", 1090 (int)(i + kernel_callchain_nr + 1), entries[i].from); 1091 } 1092 } 1093 1094 static void callchain__printf(struct evsel *evsel, 1095 struct perf_sample *sample) 1096 { 1097 unsigned int i; 1098 struct ip_callchain *callchain = sample->callchain; 1099 1100 if (evsel__has_branch_callstack(evsel)) 1101 callchain__lbr_callstack_printf(sample); 1102 1103 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr); 1104 1105 for (i = 0; i < callchain->nr; i++) 1106 printf("..... %2d: %016" PRIx64 "\n", 1107 i, callchain->ips[i]); 1108 } 1109 1110 static void branch_stack__printf(struct perf_sample *sample, bool callstack) 1111 { 1112 struct branch_entry *entries = perf_sample__branch_entries(sample); 1113 uint64_t i; 1114 1115 printf("%s: nr:%" PRIu64 "\n", 1116 !callstack ? "... branch stack" : "... branch callstack", 1117 sample->branch_stack->nr); 1118 1119 for (i = 0; i < sample->branch_stack->nr; i++) { 1120 struct branch_entry *e = &entries[i]; 1121 1122 if (!callstack) { 1123 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x\n", 1124 i, e->from, e->to, 1125 (unsigned short)e->flags.cycles, 1126 e->flags.mispred ? "M" : " ", 1127 e->flags.predicted ? "P" : " ", 1128 e->flags.abort ? "A" : " ", 1129 e->flags.in_tx ? "T" : " ", 1130 (unsigned)e->flags.reserved); 1131 } else { 1132 printf("..... %2"PRIu64": %016" PRIx64 "\n", 1133 i, i > 0 ? e->from : e->to); 1134 } 1135 } 1136 } 1137 1138 static void regs_dump__printf(u64 mask, u64 *regs) 1139 { 1140 unsigned rid, i = 0; 1141 1142 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) { 1143 u64 val = regs[i++]; 1144 1145 printf(".... %-5s 0x%016" PRIx64 "\n", 1146 perf_reg_name(rid), val); 1147 } 1148 } 1149 1150 static const char *regs_abi[] = { 1151 [PERF_SAMPLE_REGS_ABI_NONE] = "none", 1152 [PERF_SAMPLE_REGS_ABI_32] = "32-bit", 1153 [PERF_SAMPLE_REGS_ABI_64] = "64-bit", 1154 }; 1155 1156 static inline const char *regs_dump_abi(struct regs_dump *d) 1157 { 1158 if (d->abi > PERF_SAMPLE_REGS_ABI_64) 1159 return "unknown"; 1160 1161 return regs_abi[d->abi]; 1162 } 1163 1164 static void regs__printf(const char *type, struct regs_dump *regs) 1165 { 1166 u64 mask = regs->mask; 1167 1168 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n", 1169 type, 1170 mask, 1171 regs_dump_abi(regs)); 1172 1173 regs_dump__printf(mask, regs->regs); 1174 } 1175 1176 static void regs_user__printf(struct perf_sample *sample) 1177 { 1178 struct regs_dump *user_regs = &sample->user_regs; 1179 1180 if (user_regs->regs) 1181 regs__printf("user", user_regs); 1182 } 1183 1184 static void regs_intr__printf(struct perf_sample *sample) 1185 { 1186 struct regs_dump *intr_regs = &sample->intr_regs; 1187 1188 if (intr_regs->regs) 1189 regs__printf("intr", intr_regs); 1190 } 1191 1192 static void stack_user__printf(struct stack_dump *dump) 1193 { 1194 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 1195 dump->size, dump->offset); 1196 } 1197 1198 static void evlist__print_tstamp(struct evlist *evlist, union perf_event *event, struct perf_sample *sample) 1199 { 1200 u64 sample_type = __evlist__combined_sample_type(evlist); 1201 1202 if (event->header.type != PERF_RECORD_SAMPLE && 1203 !evlist__sample_id_all(evlist)) { 1204 fputs("-1 -1 ", stdout); 1205 return; 1206 } 1207 1208 if ((sample_type & PERF_SAMPLE_CPU)) 1209 printf("%u ", sample->cpu); 1210 1211 if (sample_type & PERF_SAMPLE_TIME) 1212 printf("%" PRIu64 " ", sample->time); 1213 } 1214 1215 static void sample_read__printf(struct perf_sample *sample, u64 read_format) 1216 { 1217 printf("... sample_read:\n"); 1218 1219 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1220 printf("...... time enabled %016" PRIx64 "\n", 1221 sample->read.time_enabled); 1222 1223 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1224 printf("...... time running %016" PRIx64 "\n", 1225 sample->read.time_running); 1226 1227 if (read_format & PERF_FORMAT_GROUP) { 1228 u64 i; 1229 1230 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr); 1231 1232 for (i = 0; i < sample->read.group.nr; i++) { 1233 struct sample_read_value *value; 1234 1235 value = &sample->read.group.values[i]; 1236 printf("..... id %016" PRIx64 1237 ", value %016" PRIx64 "\n", 1238 value->id, value->value); 1239 } 1240 } else 1241 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n", 1242 sample->read.one.id, sample->read.one.value); 1243 } 1244 1245 static void dump_event(struct evlist *evlist, union perf_event *event, 1246 u64 file_offset, struct perf_sample *sample) 1247 { 1248 if (!dump_trace) 1249 return; 1250 1251 printf("\n%#" PRIx64 " [%#x]: event: %d\n", 1252 file_offset, event->header.size, event->header.type); 1253 1254 trace_event(event); 1255 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw) 1256 evlist->trace_event_sample_raw(evlist, event, sample); 1257 1258 if (sample) 1259 evlist__print_tstamp(evlist, event, sample); 1260 1261 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 1262 event->header.size, perf_event__name(event->header.type)); 1263 } 1264 1265 char *get_page_size_name(u64 size, char *str) 1266 { 1267 if (!size || !unit_number__scnprintf(str, PAGE_SIZE_NAME_LEN, size)) 1268 snprintf(str, PAGE_SIZE_NAME_LEN, "%s", "N/A"); 1269 1270 return str; 1271 } 1272 1273 static void dump_sample(struct evsel *evsel, union perf_event *event, 1274 struct perf_sample *sample) 1275 { 1276 u64 sample_type; 1277 char str[PAGE_SIZE_NAME_LEN]; 1278 1279 if (!dump_trace) 1280 return; 1281 1282 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 1283 event->header.misc, sample->pid, sample->tid, sample->ip, 1284 sample->period, sample->addr); 1285 1286 sample_type = evsel->core.attr.sample_type; 1287 1288 if (evsel__has_callchain(evsel)) 1289 callchain__printf(evsel, sample); 1290 1291 if (evsel__has_br_stack(evsel)) 1292 branch_stack__printf(sample, evsel__has_branch_callstack(evsel)); 1293 1294 if (sample_type & PERF_SAMPLE_REGS_USER) 1295 regs_user__printf(sample); 1296 1297 if (sample_type & PERF_SAMPLE_REGS_INTR) 1298 regs_intr__printf(sample); 1299 1300 if (sample_type & PERF_SAMPLE_STACK_USER) 1301 stack_user__printf(&sample->user_stack); 1302 1303 if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) { 1304 printf("... weight: %" PRIu64 "", sample->weight); 1305 if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) 1306 printf(",0x%"PRIx16"", sample->ins_lat); 1307 printf("\n"); 1308 } 1309 1310 if (sample_type & PERF_SAMPLE_DATA_SRC) 1311 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src); 1312 1313 if (sample_type & PERF_SAMPLE_PHYS_ADDR) 1314 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr); 1315 1316 if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE) 1317 printf(" .. data page size: %s\n", get_page_size_name(sample->data_page_size, str)); 1318 1319 if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE) 1320 printf(" .. code page size: %s\n", get_page_size_name(sample->code_page_size, str)); 1321 1322 if (sample_type & PERF_SAMPLE_TRANSACTION) 1323 printf("... transaction: %" PRIx64 "\n", sample->transaction); 1324 1325 if (sample_type & PERF_SAMPLE_READ) 1326 sample_read__printf(sample, evsel->core.attr.read_format); 1327 } 1328 1329 static void dump_read(struct evsel *evsel, union perf_event *event) 1330 { 1331 struct perf_record_read *read_event = &event->read; 1332 u64 read_format; 1333 1334 if (!dump_trace) 1335 return; 1336 1337 printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid, 1338 evsel__name(evsel), event->read.value); 1339 1340 if (!evsel) 1341 return; 1342 1343 read_format = evsel->core.attr.read_format; 1344 1345 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1346 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled); 1347 1348 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1349 printf("... time running : %" PRI_lu64 "\n", read_event->time_running); 1350 1351 if (read_format & PERF_FORMAT_ID) 1352 printf("... id : %" PRI_lu64 "\n", read_event->id); 1353 } 1354 1355 static struct machine *machines__find_for_cpumode(struct machines *machines, 1356 union perf_event *event, 1357 struct perf_sample *sample) 1358 { 1359 struct machine *machine; 1360 1361 if (perf_guest && 1362 ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 1363 (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) { 1364 u32 pid; 1365 1366 if (event->header.type == PERF_RECORD_MMAP 1367 || event->header.type == PERF_RECORD_MMAP2) 1368 pid = event->mmap.pid; 1369 else 1370 pid = sample->pid; 1371 1372 machine = machines__find(machines, pid); 1373 if (!machine) 1374 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID); 1375 return machine; 1376 } 1377 1378 return &machines->host; 1379 } 1380 1381 static int deliver_sample_value(struct evlist *evlist, 1382 struct perf_tool *tool, 1383 union perf_event *event, 1384 struct perf_sample *sample, 1385 struct sample_read_value *v, 1386 struct machine *machine) 1387 { 1388 struct perf_sample_id *sid = evlist__id2sid(evlist, v->id); 1389 struct evsel *evsel; 1390 1391 if (sid) { 1392 sample->id = v->id; 1393 sample->period = v->value - sid->period; 1394 sid->period = v->value; 1395 } 1396 1397 if (!sid || sid->evsel == NULL) { 1398 ++evlist->stats.nr_unknown_id; 1399 return 0; 1400 } 1401 1402 /* 1403 * There's no reason to deliver sample 1404 * for zero period, bail out. 1405 */ 1406 if (!sample->period) 1407 return 0; 1408 1409 evsel = container_of(sid->evsel, struct evsel, core); 1410 return tool->sample(tool, event, sample, evsel, machine); 1411 } 1412 1413 static int deliver_sample_group(struct evlist *evlist, 1414 struct perf_tool *tool, 1415 union perf_event *event, 1416 struct perf_sample *sample, 1417 struct machine *machine) 1418 { 1419 int ret = -EINVAL; 1420 u64 i; 1421 1422 for (i = 0; i < sample->read.group.nr; i++) { 1423 ret = deliver_sample_value(evlist, tool, event, sample, 1424 &sample->read.group.values[i], 1425 machine); 1426 if (ret) 1427 break; 1428 } 1429 1430 return ret; 1431 } 1432 1433 static int evlist__deliver_sample(struct evlist *evlist, struct perf_tool *tool, 1434 union perf_event *event, struct perf_sample *sample, 1435 struct evsel *evsel, struct machine *machine) 1436 { 1437 /* We know evsel != NULL. */ 1438 u64 sample_type = evsel->core.attr.sample_type; 1439 u64 read_format = evsel->core.attr.read_format; 1440 1441 /* Standard sample delivery. */ 1442 if (!(sample_type & PERF_SAMPLE_READ)) 1443 return tool->sample(tool, event, sample, evsel, machine); 1444 1445 /* For PERF_SAMPLE_READ we have either single or group mode. */ 1446 if (read_format & PERF_FORMAT_GROUP) 1447 return deliver_sample_group(evlist, tool, event, sample, 1448 machine); 1449 else 1450 return deliver_sample_value(evlist, tool, event, sample, 1451 &sample->read.one, machine); 1452 } 1453 1454 static int machines__deliver_event(struct machines *machines, 1455 struct evlist *evlist, 1456 union perf_event *event, 1457 struct perf_sample *sample, 1458 struct perf_tool *tool, u64 file_offset) 1459 { 1460 struct evsel *evsel; 1461 struct machine *machine; 1462 1463 dump_event(evlist, event, file_offset, sample); 1464 1465 evsel = evlist__id2evsel(evlist, sample->id); 1466 1467 machine = machines__find_for_cpumode(machines, event, sample); 1468 1469 switch (event->header.type) { 1470 case PERF_RECORD_SAMPLE: 1471 if (evsel == NULL) { 1472 ++evlist->stats.nr_unknown_id; 1473 return 0; 1474 } 1475 dump_sample(evsel, event, sample); 1476 if (machine == NULL) { 1477 ++evlist->stats.nr_unprocessable_samples; 1478 return 0; 1479 } 1480 return evlist__deliver_sample(evlist, tool, event, sample, evsel, machine); 1481 case PERF_RECORD_MMAP: 1482 return tool->mmap(tool, event, sample, machine); 1483 case PERF_RECORD_MMAP2: 1484 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT) 1485 ++evlist->stats.nr_proc_map_timeout; 1486 return tool->mmap2(tool, event, sample, machine); 1487 case PERF_RECORD_COMM: 1488 return tool->comm(tool, event, sample, machine); 1489 case PERF_RECORD_NAMESPACES: 1490 return tool->namespaces(tool, event, sample, machine); 1491 case PERF_RECORD_CGROUP: 1492 return tool->cgroup(tool, event, sample, machine); 1493 case PERF_RECORD_FORK: 1494 return tool->fork(tool, event, sample, machine); 1495 case PERF_RECORD_EXIT: 1496 return tool->exit(tool, event, sample, machine); 1497 case PERF_RECORD_LOST: 1498 if (tool->lost == perf_event__process_lost) 1499 evlist->stats.total_lost += event->lost.lost; 1500 return tool->lost(tool, event, sample, machine); 1501 case PERF_RECORD_LOST_SAMPLES: 1502 if (tool->lost_samples == perf_event__process_lost_samples) 1503 evlist->stats.total_lost_samples += event->lost_samples.lost; 1504 return tool->lost_samples(tool, event, sample, machine); 1505 case PERF_RECORD_READ: 1506 dump_read(evsel, event); 1507 return tool->read(tool, event, sample, evsel, machine); 1508 case PERF_RECORD_THROTTLE: 1509 return tool->throttle(tool, event, sample, machine); 1510 case PERF_RECORD_UNTHROTTLE: 1511 return tool->unthrottle(tool, event, sample, machine); 1512 case PERF_RECORD_AUX: 1513 if (tool->aux == perf_event__process_aux) { 1514 if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) 1515 evlist->stats.total_aux_lost += 1; 1516 if (event->aux.flags & PERF_AUX_FLAG_PARTIAL) 1517 evlist->stats.total_aux_partial += 1; 1518 } 1519 return tool->aux(tool, event, sample, machine); 1520 case PERF_RECORD_ITRACE_START: 1521 return tool->itrace_start(tool, event, sample, machine); 1522 case PERF_RECORD_SWITCH: 1523 case PERF_RECORD_SWITCH_CPU_WIDE: 1524 return tool->context_switch(tool, event, sample, machine); 1525 case PERF_RECORD_KSYMBOL: 1526 return tool->ksymbol(tool, event, sample, machine); 1527 case PERF_RECORD_BPF_EVENT: 1528 return tool->bpf(tool, event, sample, machine); 1529 case PERF_RECORD_TEXT_POKE: 1530 return tool->text_poke(tool, event, sample, machine); 1531 default: 1532 ++evlist->stats.nr_unknown_events; 1533 return -1; 1534 } 1535 } 1536 1537 static int perf_session__deliver_event(struct perf_session *session, 1538 union perf_event *event, 1539 struct perf_tool *tool, 1540 u64 file_offset) 1541 { 1542 struct perf_sample sample; 1543 int ret = evlist__parse_sample(session->evlist, event, &sample); 1544 1545 if (ret) { 1546 pr_err("Can't parse sample, err = %d\n", ret); 1547 return ret; 1548 } 1549 1550 ret = auxtrace__process_event(session, event, &sample, tool); 1551 if (ret < 0) 1552 return ret; 1553 if (ret > 0) 1554 return 0; 1555 1556 ret = machines__deliver_event(&session->machines, session->evlist, 1557 event, &sample, tool, file_offset); 1558 1559 if (dump_trace && sample.aux_sample.size) 1560 auxtrace__dump_auxtrace_sample(session, &sample); 1561 1562 return ret; 1563 } 1564 1565 static s64 perf_session__process_user_event(struct perf_session *session, 1566 union perf_event *event, 1567 u64 file_offset) 1568 { 1569 struct ordered_events *oe = &session->ordered_events; 1570 struct perf_tool *tool = session->tool; 1571 struct perf_sample sample = { .time = 0, }; 1572 int fd = perf_data__fd(session->data); 1573 int err; 1574 1575 if (event->header.type != PERF_RECORD_COMPRESSED || 1576 tool->compressed == perf_session__process_compressed_event_stub) 1577 dump_event(session->evlist, event, file_offset, &sample); 1578 1579 /* These events are processed right away */ 1580 switch (event->header.type) { 1581 case PERF_RECORD_HEADER_ATTR: 1582 err = tool->attr(tool, event, &session->evlist); 1583 if (err == 0) { 1584 perf_session__set_id_hdr_size(session); 1585 perf_session__set_comm_exec(session); 1586 } 1587 return err; 1588 case PERF_RECORD_EVENT_UPDATE: 1589 return tool->event_update(tool, event, &session->evlist); 1590 case PERF_RECORD_HEADER_EVENT_TYPE: 1591 /* 1592 * Depreceated, but we need to handle it for sake 1593 * of old data files create in pipe mode. 1594 */ 1595 return 0; 1596 case PERF_RECORD_HEADER_TRACING_DATA: 1597 /* 1598 * Setup for reading amidst mmap, but only when we 1599 * are in 'file' mode. The 'pipe' fd is in proper 1600 * place already. 1601 */ 1602 if (!perf_data__is_pipe(session->data)) 1603 lseek(fd, file_offset, SEEK_SET); 1604 return tool->tracing_data(session, event); 1605 case PERF_RECORD_HEADER_BUILD_ID: 1606 return tool->build_id(session, event); 1607 case PERF_RECORD_FINISHED_ROUND: 1608 return tool->finished_round(tool, event, oe); 1609 case PERF_RECORD_ID_INDEX: 1610 return tool->id_index(session, event); 1611 case PERF_RECORD_AUXTRACE_INFO: 1612 return tool->auxtrace_info(session, event); 1613 case PERF_RECORD_AUXTRACE: 1614 /* setup for reading amidst mmap */ 1615 lseek(fd, file_offset + event->header.size, SEEK_SET); 1616 return tool->auxtrace(session, event); 1617 case PERF_RECORD_AUXTRACE_ERROR: 1618 perf_session__auxtrace_error_inc(session, event); 1619 return tool->auxtrace_error(session, event); 1620 case PERF_RECORD_THREAD_MAP: 1621 return tool->thread_map(session, event); 1622 case PERF_RECORD_CPU_MAP: 1623 return tool->cpu_map(session, event); 1624 case PERF_RECORD_STAT_CONFIG: 1625 return tool->stat_config(session, event); 1626 case PERF_RECORD_STAT: 1627 return tool->stat(session, event); 1628 case PERF_RECORD_STAT_ROUND: 1629 return tool->stat_round(session, event); 1630 case PERF_RECORD_TIME_CONV: 1631 session->time_conv = event->time_conv; 1632 return tool->time_conv(session, event); 1633 case PERF_RECORD_HEADER_FEATURE: 1634 return tool->feature(session, event); 1635 case PERF_RECORD_COMPRESSED: 1636 err = tool->compressed(session, event, file_offset); 1637 if (err) 1638 dump_event(session->evlist, event, file_offset, &sample); 1639 return err; 1640 default: 1641 return -EINVAL; 1642 } 1643 } 1644 1645 int perf_session__deliver_synth_event(struct perf_session *session, 1646 union perf_event *event, 1647 struct perf_sample *sample) 1648 { 1649 struct evlist *evlist = session->evlist; 1650 struct perf_tool *tool = session->tool; 1651 1652 events_stats__inc(&evlist->stats, event->header.type); 1653 1654 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1655 return perf_session__process_user_event(session, event, 0); 1656 1657 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0); 1658 } 1659 1660 static void event_swap(union perf_event *event, bool sample_id_all) 1661 { 1662 perf_event__swap_op swap; 1663 1664 swap = perf_event__swap_ops[event->header.type]; 1665 if (swap) 1666 swap(event, sample_id_all); 1667 } 1668 1669 int perf_session__peek_event(struct perf_session *session, off_t file_offset, 1670 void *buf, size_t buf_sz, 1671 union perf_event **event_ptr, 1672 struct perf_sample *sample) 1673 { 1674 union perf_event *event; 1675 size_t hdr_sz, rest; 1676 int fd; 1677 1678 if (session->one_mmap && !session->header.needs_swap) { 1679 event = file_offset - session->one_mmap_offset + 1680 session->one_mmap_addr; 1681 goto out_parse_sample; 1682 } 1683 1684 if (perf_data__is_pipe(session->data)) 1685 return -1; 1686 1687 fd = perf_data__fd(session->data); 1688 hdr_sz = sizeof(struct perf_event_header); 1689 1690 if (buf_sz < hdr_sz) 1691 return -1; 1692 1693 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 || 1694 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz) 1695 return -1; 1696 1697 event = (union perf_event *)buf; 1698 1699 if (session->header.needs_swap) 1700 perf_event_header__bswap(&event->header); 1701 1702 if (event->header.size < hdr_sz || event->header.size > buf_sz) 1703 return -1; 1704 1705 rest = event->header.size - hdr_sz; 1706 1707 if (readn(fd, buf, rest) != (ssize_t)rest) 1708 return -1; 1709 1710 if (session->header.needs_swap) 1711 event_swap(event, evlist__sample_id_all(session->evlist)); 1712 1713 out_parse_sample: 1714 1715 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START && 1716 evlist__parse_sample(session->evlist, event, sample)) 1717 return -1; 1718 1719 *event_ptr = event; 1720 1721 return 0; 1722 } 1723 1724 int perf_session__peek_events(struct perf_session *session, u64 offset, 1725 u64 size, peek_events_cb_t cb, void *data) 1726 { 1727 u64 max_offset = offset + size; 1728 char buf[PERF_SAMPLE_MAX_SIZE]; 1729 union perf_event *event; 1730 int err; 1731 1732 do { 1733 err = perf_session__peek_event(session, offset, buf, 1734 PERF_SAMPLE_MAX_SIZE, &event, 1735 NULL); 1736 if (err) 1737 return err; 1738 1739 err = cb(session, event, offset, data); 1740 if (err) 1741 return err; 1742 1743 offset += event->header.size; 1744 if (event->header.type == PERF_RECORD_AUXTRACE) 1745 offset += event->auxtrace.size; 1746 1747 } while (offset < max_offset); 1748 1749 return err; 1750 } 1751 1752 static s64 perf_session__process_event(struct perf_session *session, 1753 union perf_event *event, u64 file_offset) 1754 { 1755 struct evlist *evlist = session->evlist; 1756 struct perf_tool *tool = session->tool; 1757 int ret; 1758 1759 if (session->header.needs_swap) 1760 event_swap(event, evlist__sample_id_all(evlist)); 1761 1762 if (event->header.type >= PERF_RECORD_HEADER_MAX) 1763 return -EINVAL; 1764 1765 events_stats__inc(&evlist->stats, event->header.type); 1766 1767 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1768 return perf_session__process_user_event(session, event, file_offset); 1769 1770 if (tool->ordered_events) { 1771 u64 timestamp = -1ULL; 1772 1773 ret = evlist__parse_sample_timestamp(evlist, event, ×tamp); 1774 if (ret && ret != -1) 1775 return ret; 1776 1777 ret = perf_session__queue_event(session, event, timestamp, file_offset); 1778 if (ret != -ETIME) 1779 return ret; 1780 } 1781 1782 return perf_session__deliver_event(session, event, tool, file_offset); 1783 } 1784 1785 void perf_event_header__bswap(struct perf_event_header *hdr) 1786 { 1787 hdr->type = bswap_32(hdr->type); 1788 hdr->misc = bswap_16(hdr->misc); 1789 hdr->size = bswap_16(hdr->size); 1790 } 1791 1792 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1793 { 1794 return machine__findnew_thread(&session->machines.host, -1, pid); 1795 } 1796 1797 /* 1798 * Threads are identified by pid and tid, and the idle task has pid == tid == 0. 1799 * So here a single thread is created for that, but actually there is a separate 1800 * idle task per cpu, so there should be one 'struct thread' per cpu, but there 1801 * is only 1. That causes problems for some tools, requiring workarounds. For 1802 * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu(). 1803 */ 1804 int perf_session__register_idle_thread(struct perf_session *session) 1805 { 1806 struct thread *thread; 1807 int err = 0; 1808 1809 thread = machine__findnew_thread(&session->machines.host, 0, 0); 1810 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) { 1811 pr_err("problem inserting idle task.\n"); 1812 err = -1; 1813 } 1814 1815 if (thread == NULL || thread__set_namespaces(thread, 0, NULL)) { 1816 pr_err("problem inserting idle task.\n"); 1817 err = -1; 1818 } 1819 1820 /* machine__findnew_thread() got the thread, so put it */ 1821 thread__put(thread); 1822 return err; 1823 } 1824 1825 static void 1826 perf_session__warn_order(const struct perf_session *session) 1827 { 1828 const struct ordered_events *oe = &session->ordered_events; 1829 struct evsel *evsel; 1830 bool should_warn = true; 1831 1832 evlist__for_each_entry(session->evlist, evsel) { 1833 if (evsel->core.attr.write_backward) 1834 should_warn = false; 1835 } 1836 1837 if (!should_warn) 1838 return; 1839 if (oe->nr_unordered_events != 0) 1840 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events); 1841 } 1842 1843 static void perf_session__warn_about_errors(const struct perf_session *session) 1844 { 1845 const struct events_stats *stats = &session->evlist->stats; 1846 1847 if (session->tool->lost == perf_event__process_lost && 1848 stats->nr_events[PERF_RECORD_LOST] != 0) { 1849 ui__warning("Processed %d events and lost %d chunks!\n\n" 1850 "Check IO/CPU overload!\n\n", 1851 stats->nr_events[0], 1852 stats->nr_events[PERF_RECORD_LOST]); 1853 } 1854 1855 if (session->tool->lost_samples == perf_event__process_lost_samples) { 1856 double drop_rate; 1857 1858 drop_rate = (double)stats->total_lost_samples / 1859 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples); 1860 if (drop_rate > 0.05) { 1861 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n", 1862 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples, 1863 drop_rate * 100.0); 1864 } 1865 } 1866 1867 if (session->tool->aux == perf_event__process_aux && 1868 stats->total_aux_lost != 0) { 1869 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n", 1870 stats->total_aux_lost, 1871 stats->nr_events[PERF_RECORD_AUX]); 1872 } 1873 1874 if (session->tool->aux == perf_event__process_aux && 1875 stats->total_aux_partial != 0) { 1876 bool vmm_exclusive = false; 1877 1878 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive", 1879 &vmm_exclusive); 1880 1881 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n" 1882 "Are you running a KVM guest in the background?%s\n\n", 1883 stats->total_aux_partial, 1884 stats->nr_events[PERF_RECORD_AUX], 1885 vmm_exclusive ? 1886 "\nReloading kvm_intel module with vmm_exclusive=0\n" 1887 "will reduce the gaps to only guest's timeslices." : 1888 ""); 1889 } 1890 1891 if (stats->nr_unknown_events != 0) { 1892 ui__warning("Found %u unknown events!\n\n" 1893 "Is this an older tool processing a perf.data " 1894 "file generated by a more recent tool?\n\n" 1895 "If that is not the case, consider " 1896 "reporting to linux-kernel@vger.kernel.org.\n\n", 1897 stats->nr_unknown_events); 1898 } 1899 1900 if (stats->nr_unknown_id != 0) { 1901 ui__warning("%u samples with id not present in the header\n", 1902 stats->nr_unknown_id); 1903 } 1904 1905 if (stats->nr_invalid_chains != 0) { 1906 ui__warning("Found invalid callchains!\n\n" 1907 "%u out of %u events were discarded for this reason.\n\n" 1908 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 1909 stats->nr_invalid_chains, 1910 stats->nr_events[PERF_RECORD_SAMPLE]); 1911 } 1912 1913 if (stats->nr_unprocessable_samples != 0) { 1914 ui__warning("%u unprocessable samples recorded.\n" 1915 "Do you have a KVM guest running and not using 'perf kvm'?\n", 1916 stats->nr_unprocessable_samples); 1917 } 1918 1919 perf_session__warn_order(session); 1920 1921 events_stats__auxtrace_error_warn(stats); 1922 1923 if (stats->nr_proc_map_timeout != 0) { 1924 ui__warning("%d map information files for pre-existing threads were\n" 1925 "not processed, if there are samples for addresses they\n" 1926 "will not be resolved, you may find out which are these\n" 1927 "threads by running with -v and redirecting the output\n" 1928 "to a file.\n" 1929 "The time limit to process proc map is too short?\n" 1930 "Increase it by --proc-map-timeout\n", 1931 stats->nr_proc_map_timeout); 1932 } 1933 } 1934 1935 static int perf_session__flush_thread_stack(struct thread *thread, 1936 void *p __maybe_unused) 1937 { 1938 return thread_stack__flush(thread); 1939 } 1940 1941 static int perf_session__flush_thread_stacks(struct perf_session *session) 1942 { 1943 return machines__for_each_thread(&session->machines, 1944 perf_session__flush_thread_stack, 1945 NULL); 1946 } 1947 1948 volatile int session_done; 1949 1950 static int __perf_session__process_decomp_events(struct perf_session *session); 1951 1952 static int __perf_session__process_pipe_events(struct perf_session *session) 1953 { 1954 struct ordered_events *oe = &session->ordered_events; 1955 struct perf_tool *tool = session->tool; 1956 union perf_event *event; 1957 uint32_t size, cur_size = 0; 1958 void *buf = NULL; 1959 s64 skip = 0; 1960 u64 head; 1961 ssize_t err; 1962 void *p; 1963 1964 perf_tool__fill_defaults(tool); 1965 1966 head = 0; 1967 cur_size = sizeof(union perf_event); 1968 1969 buf = malloc(cur_size); 1970 if (!buf) 1971 return -errno; 1972 ordered_events__set_copy_on_queue(oe, true); 1973 more: 1974 event = buf; 1975 err = perf_data__read(session->data, event, 1976 sizeof(struct perf_event_header)); 1977 if (err <= 0) { 1978 if (err == 0) 1979 goto done; 1980 1981 pr_err("failed to read event header\n"); 1982 goto out_err; 1983 } 1984 1985 if (session->header.needs_swap) 1986 perf_event_header__bswap(&event->header); 1987 1988 size = event->header.size; 1989 if (size < sizeof(struct perf_event_header)) { 1990 pr_err("bad event header size\n"); 1991 goto out_err; 1992 } 1993 1994 if (size > cur_size) { 1995 void *new = realloc(buf, size); 1996 if (!new) { 1997 pr_err("failed to allocate memory to read event\n"); 1998 goto out_err; 1999 } 2000 buf = new; 2001 cur_size = size; 2002 event = buf; 2003 } 2004 p = event; 2005 p += sizeof(struct perf_event_header); 2006 2007 if (size - sizeof(struct perf_event_header)) { 2008 err = perf_data__read(session->data, p, 2009 size - sizeof(struct perf_event_header)); 2010 if (err <= 0) { 2011 if (err == 0) { 2012 pr_err("unexpected end of event stream\n"); 2013 goto done; 2014 } 2015 2016 pr_err("failed to read event data\n"); 2017 goto out_err; 2018 } 2019 } 2020 2021 if ((skip = perf_session__process_event(session, event, head)) < 0) { 2022 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 2023 head, event->header.size, event->header.type); 2024 err = -EINVAL; 2025 goto out_err; 2026 } 2027 2028 head += size; 2029 2030 if (skip > 0) 2031 head += skip; 2032 2033 err = __perf_session__process_decomp_events(session); 2034 if (err) 2035 goto out_err; 2036 2037 if (!session_done()) 2038 goto more; 2039 done: 2040 /* do the final flush for ordered samples */ 2041 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2042 if (err) 2043 goto out_err; 2044 err = auxtrace__flush_events(session, tool); 2045 if (err) 2046 goto out_err; 2047 err = perf_session__flush_thread_stacks(session); 2048 out_err: 2049 free(buf); 2050 if (!tool->no_warn) 2051 perf_session__warn_about_errors(session); 2052 ordered_events__free(&session->ordered_events); 2053 auxtrace__free_events(session); 2054 return err; 2055 } 2056 2057 static union perf_event * 2058 prefetch_event(char *buf, u64 head, size_t mmap_size, 2059 bool needs_swap, union perf_event *error) 2060 { 2061 union perf_event *event; 2062 2063 /* 2064 * Ensure we have enough space remaining to read 2065 * the size of the event in the headers. 2066 */ 2067 if (head + sizeof(event->header) > mmap_size) 2068 return NULL; 2069 2070 event = (union perf_event *)(buf + head); 2071 if (needs_swap) 2072 perf_event_header__bswap(&event->header); 2073 2074 if (head + event->header.size <= mmap_size) 2075 return event; 2076 2077 /* We're not fetching the event so swap back again */ 2078 if (needs_swap) 2079 perf_event_header__bswap(&event->header); 2080 2081 pr_debug("%s: head=%#" PRIx64 " event->header_size=%#x, mmap_size=%#zx:" 2082 " fuzzed or compressed perf.data?\n",__func__, head, event->header.size, mmap_size); 2083 2084 return error; 2085 } 2086 2087 static union perf_event * 2088 fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 2089 { 2090 return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL)); 2091 } 2092 2093 static union perf_event * 2094 fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 2095 { 2096 return prefetch_event(buf, head, mmap_size, needs_swap, NULL); 2097 } 2098 2099 static int __perf_session__process_decomp_events(struct perf_session *session) 2100 { 2101 s64 skip; 2102 u64 size, file_pos = 0; 2103 struct decomp *decomp = session->decomp_last; 2104 2105 if (!decomp) 2106 return 0; 2107 2108 while (decomp->head < decomp->size && !session_done()) { 2109 union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data, 2110 session->header.needs_swap); 2111 2112 if (!event) 2113 break; 2114 2115 size = event->header.size; 2116 2117 if (size < sizeof(struct perf_event_header) || 2118 (skip = perf_session__process_event(session, event, file_pos)) < 0) { 2119 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 2120 decomp->file_pos + decomp->head, event->header.size, event->header.type); 2121 return -EINVAL; 2122 } 2123 2124 if (skip) 2125 size += skip; 2126 2127 decomp->head += size; 2128 } 2129 2130 return 0; 2131 } 2132 2133 /* 2134 * On 64bit we can mmap the data file in one go. No need for tiny mmap 2135 * slices. On 32bit we use 32MB. 2136 */ 2137 #if BITS_PER_LONG == 64 2138 #define MMAP_SIZE ULLONG_MAX 2139 #define NUM_MMAPS 1 2140 #else 2141 #define MMAP_SIZE (32 * 1024 * 1024ULL) 2142 #define NUM_MMAPS 128 2143 #endif 2144 2145 struct reader; 2146 2147 typedef s64 (*reader_cb_t)(struct perf_session *session, 2148 union perf_event *event, 2149 u64 file_offset); 2150 2151 struct reader { 2152 int fd; 2153 u64 data_size; 2154 u64 data_offset; 2155 reader_cb_t process; 2156 }; 2157 2158 static int 2159 reader__process_events(struct reader *rd, struct perf_session *session, 2160 struct ui_progress *prog) 2161 { 2162 u64 data_size = rd->data_size; 2163 u64 head, page_offset, file_offset, file_pos, size; 2164 int err = 0, mmap_prot, mmap_flags, map_idx = 0; 2165 size_t mmap_size; 2166 char *buf, *mmaps[NUM_MMAPS]; 2167 union perf_event *event; 2168 s64 skip; 2169 2170 page_offset = page_size * (rd->data_offset / page_size); 2171 file_offset = page_offset; 2172 head = rd->data_offset - page_offset; 2173 2174 ui_progress__init_size(prog, data_size, "Processing events..."); 2175 2176 data_size += rd->data_offset; 2177 2178 mmap_size = MMAP_SIZE; 2179 if (mmap_size > data_size) { 2180 mmap_size = data_size; 2181 session->one_mmap = true; 2182 } 2183 2184 memset(mmaps, 0, sizeof(mmaps)); 2185 2186 mmap_prot = PROT_READ; 2187 mmap_flags = MAP_SHARED; 2188 2189 if (session->header.needs_swap) { 2190 mmap_prot |= PROT_WRITE; 2191 mmap_flags = MAP_PRIVATE; 2192 } 2193 remap: 2194 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, rd->fd, 2195 file_offset); 2196 if (buf == MAP_FAILED) { 2197 pr_err("failed to mmap file\n"); 2198 err = -errno; 2199 goto out; 2200 } 2201 mmaps[map_idx] = buf; 2202 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1); 2203 file_pos = file_offset + head; 2204 if (session->one_mmap) { 2205 session->one_mmap_addr = buf; 2206 session->one_mmap_offset = file_offset; 2207 } 2208 2209 more: 2210 event = fetch_mmaped_event(head, mmap_size, buf, session->header.needs_swap); 2211 if (IS_ERR(event)) 2212 return PTR_ERR(event); 2213 2214 if (!event) { 2215 if (mmaps[map_idx]) { 2216 munmap(mmaps[map_idx], mmap_size); 2217 mmaps[map_idx] = NULL; 2218 } 2219 2220 page_offset = page_size * (head / page_size); 2221 file_offset += page_offset; 2222 head -= page_offset; 2223 goto remap; 2224 } 2225 2226 size = event->header.size; 2227 2228 skip = -EINVAL; 2229 2230 if (size < sizeof(struct perf_event_header) || 2231 (skip = rd->process(session, event, file_pos)) < 0) { 2232 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n", 2233 file_offset + head, event->header.size, 2234 event->header.type, strerror(-skip)); 2235 err = skip; 2236 goto out; 2237 } 2238 2239 if (skip) 2240 size += skip; 2241 2242 head += size; 2243 file_pos += size; 2244 2245 err = __perf_session__process_decomp_events(session); 2246 if (err) 2247 goto out; 2248 2249 ui_progress__update(prog, size); 2250 2251 if (session_done()) 2252 goto out; 2253 2254 if (file_pos < data_size) 2255 goto more; 2256 2257 out: 2258 return err; 2259 } 2260 2261 static s64 process_simple(struct perf_session *session, 2262 union perf_event *event, 2263 u64 file_offset) 2264 { 2265 return perf_session__process_event(session, event, file_offset); 2266 } 2267 2268 static int __perf_session__process_events(struct perf_session *session) 2269 { 2270 struct reader rd = { 2271 .fd = perf_data__fd(session->data), 2272 .data_size = session->header.data_size, 2273 .data_offset = session->header.data_offset, 2274 .process = process_simple, 2275 }; 2276 struct ordered_events *oe = &session->ordered_events; 2277 struct perf_tool *tool = session->tool; 2278 struct ui_progress prog; 2279 int err; 2280 2281 perf_tool__fill_defaults(tool); 2282 2283 if (rd.data_size == 0) 2284 return -1; 2285 2286 ui_progress__init_size(&prog, rd.data_size, "Processing events..."); 2287 2288 err = reader__process_events(&rd, session, &prog); 2289 if (err) 2290 goto out_err; 2291 /* do the final flush for ordered samples */ 2292 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2293 if (err) 2294 goto out_err; 2295 err = auxtrace__flush_events(session, tool); 2296 if (err) 2297 goto out_err; 2298 err = perf_session__flush_thread_stacks(session); 2299 out_err: 2300 ui_progress__finish(); 2301 if (!tool->no_warn) 2302 perf_session__warn_about_errors(session); 2303 /* 2304 * We may switching perf.data output, make ordered_events 2305 * reusable. 2306 */ 2307 ordered_events__reinit(&session->ordered_events); 2308 auxtrace__free_events(session); 2309 session->one_mmap = false; 2310 return err; 2311 } 2312 2313 int perf_session__process_events(struct perf_session *session) 2314 { 2315 if (perf_session__register_idle_thread(session) < 0) 2316 return -ENOMEM; 2317 2318 if (perf_data__is_pipe(session->data)) 2319 return __perf_session__process_pipe_events(session); 2320 2321 return __perf_session__process_events(session); 2322 } 2323 2324 bool perf_session__has_traces(struct perf_session *session, const char *msg) 2325 { 2326 struct evsel *evsel; 2327 2328 evlist__for_each_entry(session->evlist, evsel) { 2329 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT) 2330 return true; 2331 } 2332 2333 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 2334 return false; 2335 } 2336 2337 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr) 2338 { 2339 char *bracket; 2340 struct ref_reloc_sym *ref; 2341 struct kmap *kmap; 2342 2343 ref = zalloc(sizeof(struct ref_reloc_sym)); 2344 if (ref == NULL) 2345 return -ENOMEM; 2346 2347 ref->name = strdup(symbol_name); 2348 if (ref->name == NULL) { 2349 free(ref); 2350 return -ENOMEM; 2351 } 2352 2353 bracket = strchr(ref->name, ']'); 2354 if (bracket) 2355 *bracket = '\0'; 2356 2357 ref->addr = addr; 2358 2359 kmap = map__kmap(map); 2360 if (kmap) 2361 kmap->ref_reloc_sym = ref; 2362 2363 return 0; 2364 } 2365 2366 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp) 2367 { 2368 return machines__fprintf_dsos(&session->machines, fp); 2369 } 2370 2371 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp, 2372 bool (skip)(struct dso *dso, int parm), int parm) 2373 { 2374 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm); 2375 } 2376 2377 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 2378 { 2379 size_t ret; 2380 const char *msg = ""; 2381 2382 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) 2383 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)"; 2384 2385 ret = fprintf(fp, "\nAggregated stats:%s\n", msg); 2386 2387 ret += events_stats__fprintf(&session->evlist->stats, fp); 2388 return ret; 2389 } 2390 2391 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 2392 { 2393 /* 2394 * FIXME: Here we have to actually print all the machines in this 2395 * session, not just the host... 2396 */ 2397 return machine__fprintf(&session->machines.host, fp); 2398 } 2399 2400 struct evsel *perf_session__find_first_evtype(struct perf_session *session, 2401 unsigned int type) 2402 { 2403 struct evsel *pos; 2404 2405 evlist__for_each_entry(session->evlist, pos) { 2406 if (pos->core.attr.type == type) 2407 return pos; 2408 } 2409 return NULL; 2410 } 2411 2412 int perf_session__cpu_bitmap(struct perf_session *session, 2413 const char *cpu_list, unsigned long *cpu_bitmap) 2414 { 2415 int i, err = -1; 2416 struct perf_cpu_map *map; 2417 int nr_cpus = min(session->header.env.nr_cpus_avail, MAX_NR_CPUS); 2418 2419 for (i = 0; i < PERF_TYPE_MAX; ++i) { 2420 struct evsel *evsel; 2421 2422 evsel = perf_session__find_first_evtype(session, i); 2423 if (!evsel) 2424 continue; 2425 2426 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) { 2427 pr_err("File does not contain CPU events. " 2428 "Remove -C option to proceed.\n"); 2429 return -1; 2430 } 2431 } 2432 2433 map = perf_cpu_map__new(cpu_list); 2434 if (map == NULL) { 2435 pr_err("Invalid cpu_list\n"); 2436 return -1; 2437 } 2438 2439 for (i = 0; i < map->nr; i++) { 2440 int cpu = map->map[i]; 2441 2442 if (cpu >= nr_cpus) { 2443 pr_err("Requested CPU %d too large. " 2444 "Consider raising MAX_NR_CPUS\n", cpu); 2445 goto out_delete_map; 2446 } 2447 2448 set_bit(cpu, cpu_bitmap); 2449 } 2450 2451 err = 0; 2452 2453 out_delete_map: 2454 perf_cpu_map__put(map); 2455 return err; 2456 } 2457 2458 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 2459 bool full) 2460 { 2461 if (session == NULL || fp == NULL) 2462 return; 2463 2464 fprintf(fp, "# ========\n"); 2465 perf_header__fprintf_info(session, fp, full); 2466 fprintf(fp, "# ========\n#\n"); 2467 } 2468 2469 int perf_event__process_id_index(struct perf_session *session, 2470 union perf_event *event) 2471 { 2472 struct evlist *evlist = session->evlist; 2473 struct perf_record_id_index *ie = &event->id_index; 2474 size_t i, nr, max_nr; 2475 2476 max_nr = (ie->header.size - sizeof(struct perf_record_id_index)) / 2477 sizeof(struct id_index_entry); 2478 nr = ie->nr; 2479 if (nr > max_nr) 2480 return -EINVAL; 2481 2482 if (dump_trace) 2483 fprintf(stdout, " nr: %zu\n", nr); 2484 2485 for (i = 0; i < nr; i++) { 2486 struct id_index_entry *e = &ie->entries[i]; 2487 struct perf_sample_id *sid; 2488 2489 if (dump_trace) { 2490 fprintf(stdout, " ... id: %"PRI_lu64, e->id); 2491 fprintf(stdout, " idx: %"PRI_lu64, e->idx); 2492 fprintf(stdout, " cpu: %"PRI_ld64, e->cpu); 2493 fprintf(stdout, " tid: %"PRI_ld64"\n", e->tid); 2494 } 2495 2496 sid = evlist__id2sid(evlist, e->id); 2497 if (!sid) 2498 return -ENOENT; 2499 sid->idx = e->idx; 2500 sid->cpu = e->cpu; 2501 sid->tid = e->tid; 2502 } 2503 return 0; 2504 } 2505