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 "env.h" 19 #include "evlist.h" 20 #include "evsel.h" 21 #include "memswap.h" 22 #include "map.h" 23 #include "symbol.h" 24 #include "session.h" 25 #include "tool.h" 26 #include "perf_regs.h" 27 #include "asm/bug.h" 28 #include "auxtrace.h" 29 #include "thread.h" 30 #include "thread-stack.h" 31 #include "sample-raw.h" 32 #include "stat.h" 33 #include "tsc.h" 34 #include "ui/progress.h" 35 #include "../perf.h" 36 #include "arch/common.h" 37 #include "units.h" 38 #include <internal/lib.h> 39 40 #ifdef HAVE_ZSTD_SUPPORT 41 static int perf_session__process_compressed_event(struct perf_session *session, 42 union perf_event *event, u64 file_offset) 43 { 44 void *src; 45 size_t decomp_size, src_size; 46 u64 decomp_last_rem = 0; 47 size_t mmap_len, decomp_len = session->header.env.comp_mmap_len; 48 struct decomp *decomp, *decomp_last = session->active_decomp->decomp_last; 49 50 if (decomp_last) { 51 decomp_last_rem = decomp_last->size - decomp_last->head; 52 decomp_len += decomp_last_rem; 53 } 54 55 mmap_len = sizeof(struct decomp) + decomp_len; 56 decomp = mmap(NULL, mmap_len, PROT_READ|PROT_WRITE, 57 MAP_ANONYMOUS|MAP_PRIVATE, -1, 0); 58 if (decomp == MAP_FAILED) { 59 pr_err("Couldn't allocate memory for decompression\n"); 60 return -1; 61 } 62 63 decomp->file_pos = file_offset; 64 decomp->mmap_len = mmap_len; 65 decomp->head = 0; 66 67 if (decomp_last_rem) { 68 memcpy(decomp->data, &(decomp_last->data[decomp_last->head]), decomp_last_rem); 69 decomp->size = decomp_last_rem; 70 } 71 72 src = (void *)event + sizeof(struct perf_record_compressed); 73 src_size = event->pack.header.size - sizeof(struct perf_record_compressed); 74 75 decomp_size = zstd_decompress_stream(session->active_decomp->zstd_decomp, src, src_size, 76 &(decomp->data[decomp_last_rem]), decomp_len - decomp_last_rem); 77 if (!decomp_size) { 78 munmap(decomp, mmap_len); 79 pr_err("Couldn't decompress data\n"); 80 return -1; 81 } 82 83 decomp->size += decomp_size; 84 85 if (session->active_decomp->decomp == NULL) 86 session->active_decomp->decomp = decomp; 87 else 88 session->active_decomp->decomp_last->next = decomp; 89 90 session->active_decomp->decomp_last = decomp; 91 92 pr_debug("decomp (B): %zd to %zd\n", src_size, decomp_size); 93 94 return 0; 95 } 96 #else /* !HAVE_ZSTD_SUPPORT */ 97 #define perf_session__process_compressed_event perf_session__process_compressed_event_stub 98 #endif 99 100 static int perf_session__deliver_event(struct perf_session *session, 101 union perf_event *event, 102 struct perf_tool *tool, 103 u64 file_offset); 104 105 static int perf_session__open(struct perf_session *session, int repipe_fd) 106 { 107 struct perf_data *data = session->data; 108 109 if (perf_session__read_header(session, repipe_fd) < 0) { 110 pr_err("incompatible file format (rerun with -v to learn more)\n"); 111 return -1; 112 } 113 114 if (perf_data__is_pipe(data)) 115 return 0; 116 117 if (perf_header__has_feat(&session->header, HEADER_STAT)) 118 return 0; 119 120 if (!evlist__valid_sample_type(session->evlist)) { 121 pr_err("non matching sample_type\n"); 122 return -1; 123 } 124 125 if (!evlist__valid_sample_id_all(session->evlist)) { 126 pr_err("non matching sample_id_all\n"); 127 return -1; 128 } 129 130 if (!evlist__valid_read_format(session->evlist)) { 131 pr_err("non matching read_format\n"); 132 return -1; 133 } 134 135 return 0; 136 } 137 138 void perf_session__set_id_hdr_size(struct perf_session *session) 139 { 140 u16 id_hdr_size = evlist__id_hdr_size(session->evlist); 141 142 machines__set_id_hdr_size(&session->machines, id_hdr_size); 143 } 144 145 int perf_session__create_kernel_maps(struct perf_session *session) 146 { 147 int ret = machine__create_kernel_maps(&session->machines.host); 148 149 if (ret >= 0) 150 ret = machines__create_guest_kernel_maps(&session->machines); 151 return ret; 152 } 153 154 static void perf_session__destroy_kernel_maps(struct perf_session *session) 155 { 156 machines__destroy_kernel_maps(&session->machines); 157 } 158 159 static bool perf_session__has_comm_exec(struct perf_session *session) 160 { 161 struct evsel *evsel; 162 163 evlist__for_each_entry(session->evlist, evsel) { 164 if (evsel->core.attr.comm_exec) 165 return true; 166 } 167 168 return false; 169 } 170 171 static void perf_session__set_comm_exec(struct perf_session *session) 172 { 173 bool comm_exec = perf_session__has_comm_exec(session); 174 175 machines__set_comm_exec(&session->machines, comm_exec); 176 } 177 178 static int ordered_events__deliver_event(struct ordered_events *oe, 179 struct ordered_event *event) 180 { 181 struct perf_session *session = container_of(oe, struct perf_session, 182 ordered_events); 183 184 return perf_session__deliver_event(session, event->event, 185 session->tool, event->file_offset); 186 } 187 188 struct perf_session *__perf_session__new(struct perf_data *data, 189 bool repipe, int repipe_fd, 190 struct perf_tool *tool) 191 { 192 int ret = -ENOMEM; 193 struct perf_session *session = zalloc(sizeof(*session)); 194 195 if (!session) 196 goto out; 197 198 session->repipe = repipe; 199 session->tool = tool; 200 session->decomp_data.zstd_decomp = &session->zstd_data; 201 session->active_decomp = &session->decomp_data; 202 INIT_LIST_HEAD(&session->auxtrace_index); 203 machines__init(&session->machines); 204 ordered_events__init(&session->ordered_events, 205 ordered_events__deliver_event, NULL); 206 207 perf_env__init(&session->header.env); 208 if (data) { 209 ret = perf_data__open(data); 210 if (ret < 0) 211 goto out_delete; 212 213 session->data = data; 214 215 if (perf_data__is_read(data)) { 216 ret = perf_session__open(session, repipe_fd); 217 if (ret < 0) 218 goto out_delete; 219 220 /* 221 * set session attributes that are present in perf.data 222 * but not in pipe-mode. 223 */ 224 if (!data->is_pipe) { 225 perf_session__set_id_hdr_size(session); 226 perf_session__set_comm_exec(session); 227 } 228 229 evlist__init_trace_event_sample_raw(session->evlist); 230 231 /* Open the directory data. */ 232 if (data->is_dir) { 233 ret = perf_data__open_dir(data); 234 if (ret) 235 goto out_delete; 236 } 237 238 if (!symbol_conf.kallsyms_name && 239 !symbol_conf.vmlinux_name) 240 symbol_conf.kallsyms_name = perf_data__kallsyms_name(data); 241 } 242 } else { 243 session->machines.host.env = &perf_env; 244 } 245 246 session->machines.host.single_address_space = 247 perf_env__single_address_space(session->machines.host.env); 248 249 if (!data || perf_data__is_write(data)) { 250 /* 251 * In O_RDONLY mode this will be performed when reading the 252 * kernel MMAP event, in perf_event__process_mmap(). 253 */ 254 if (perf_session__create_kernel_maps(session) < 0) 255 pr_warning("Cannot read kernel map\n"); 256 } 257 258 /* 259 * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is 260 * processed, so evlist__sample_id_all is not meaningful here. 261 */ 262 if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps && 263 tool->ordered_events && !evlist__sample_id_all(session->evlist)) { 264 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 265 tool->ordered_events = false; 266 } 267 268 return session; 269 270 out_delete: 271 perf_session__delete(session); 272 out: 273 return ERR_PTR(ret); 274 } 275 276 static void perf_session__delete_threads(struct perf_session *session) 277 { 278 machine__delete_threads(&session->machines.host); 279 } 280 281 static void perf_decomp__release_events(struct decomp *next) 282 { 283 struct decomp *decomp; 284 size_t mmap_len; 285 286 do { 287 decomp = next; 288 if (decomp == NULL) 289 break; 290 next = decomp->next; 291 mmap_len = decomp->mmap_len; 292 munmap(decomp, mmap_len); 293 } while (1); 294 } 295 296 void perf_session__delete(struct perf_session *session) 297 { 298 if (session == NULL) 299 return; 300 auxtrace__free(session); 301 auxtrace_index__free(&session->auxtrace_index); 302 perf_session__destroy_kernel_maps(session); 303 perf_session__delete_threads(session); 304 perf_decomp__release_events(session->decomp_data.decomp); 305 perf_env__exit(&session->header.env); 306 machines__exit(&session->machines); 307 if (session->data) { 308 if (perf_data__is_read(session->data)) 309 evlist__delete(session->evlist); 310 perf_data__close(session->data); 311 } 312 trace_event__cleanup(&session->tevent); 313 free(session); 314 } 315 316 static int process_event_synth_tracing_data_stub(struct perf_session *session 317 __maybe_unused, 318 union perf_event *event 319 __maybe_unused) 320 { 321 dump_printf(": unhandled!\n"); 322 return 0; 323 } 324 325 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused, 326 union perf_event *event __maybe_unused, 327 struct evlist **pevlist 328 __maybe_unused) 329 { 330 dump_printf(": unhandled!\n"); 331 return 0; 332 } 333 334 static int process_event_synth_event_update_stub(struct perf_tool *tool __maybe_unused, 335 union perf_event *event __maybe_unused, 336 struct evlist **pevlist 337 __maybe_unused) 338 { 339 if (dump_trace) 340 perf_event__fprintf_event_update(event, stdout); 341 342 dump_printf(": unhandled!\n"); 343 return 0; 344 } 345 346 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused, 347 union perf_event *event __maybe_unused, 348 struct perf_sample *sample __maybe_unused, 349 struct evsel *evsel __maybe_unused, 350 struct machine *machine __maybe_unused) 351 { 352 dump_printf(": unhandled!\n"); 353 return 0; 354 } 355 356 static int process_event_stub(struct perf_tool *tool __maybe_unused, 357 union perf_event *event __maybe_unused, 358 struct perf_sample *sample __maybe_unused, 359 struct machine *machine __maybe_unused) 360 { 361 dump_printf(": unhandled!\n"); 362 return 0; 363 } 364 365 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused, 366 union perf_event *event __maybe_unused, 367 struct ordered_events *oe __maybe_unused) 368 { 369 dump_printf(": unhandled!\n"); 370 return 0; 371 } 372 373 static int process_finished_round(struct perf_tool *tool, 374 union perf_event *event, 375 struct ordered_events *oe); 376 377 static int skipn(int fd, off_t n) 378 { 379 char buf[4096]; 380 ssize_t ret; 381 382 while (n > 0) { 383 ret = read(fd, buf, min(n, (off_t)sizeof(buf))); 384 if (ret <= 0) 385 return ret; 386 n -= ret; 387 } 388 389 return 0; 390 } 391 392 static s64 process_event_auxtrace_stub(struct perf_session *session __maybe_unused, 393 union perf_event *event) 394 { 395 dump_printf(": unhandled!\n"); 396 if (perf_data__is_pipe(session->data)) 397 skipn(perf_data__fd(session->data), event->auxtrace.size); 398 return event->auxtrace.size; 399 } 400 401 static int process_event_op2_stub(struct perf_session *session __maybe_unused, 402 union perf_event *event __maybe_unused) 403 { 404 dump_printf(": unhandled!\n"); 405 return 0; 406 } 407 408 409 static 410 int process_event_thread_map_stub(struct perf_session *session __maybe_unused, 411 union perf_event *event __maybe_unused) 412 { 413 if (dump_trace) 414 perf_event__fprintf_thread_map(event, stdout); 415 416 dump_printf(": unhandled!\n"); 417 return 0; 418 } 419 420 static 421 int process_event_cpu_map_stub(struct perf_session *session __maybe_unused, 422 union perf_event *event __maybe_unused) 423 { 424 if (dump_trace) 425 perf_event__fprintf_cpu_map(event, stdout); 426 427 dump_printf(": unhandled!\n"); 428 return 0; 429 } 430 431 static 432 int process_event_stat_config_stub(struct perf_session *session __maybe_unused, 433 union perf_event *event __maybe_unused) 434 { 435 if (dump_trace) 436 perf_event__fprintf_stat_config(event, stdout); 437 438 dump_printf(": unhandled!\n"); 439 return 0; 440 } 441 442 static int process_stat_stub(struct perf_session *perf_session __maybe_unused, 443 union perf_event *event) 444 { 445 if (dump_trace) 446 perf_event__fprintf_stat(event, stdout); 447 448 dump_printf(": unhandled!\n"); 449 return 0; 450 } 451 452 static int process_stat_round_stub(struct perf_session *perf_session __maybe_unused, 453 union perf_event *event) 454 { 455 if (dump_trace) 456 perf_event__fprintf_stat_round(event, stdout); 457 458 dump_printf(": unhandled!\n"); 459 return 0; 460 } 461 462 static int process_event_time_conv_stub(struct perf_session *perf_session __maybe_unused, 463 union perf_event *event) 464 { 465 if (dump_trace) 466 perf_event__fprintf_time_conv(event, stdout); 467 468 dump_printf(": unhandled!\n"); 469 return 0; 470 } 471 472 static int perf_session__process_compressed_event_stub(struct perf_session *session __maybe_unused, 473 union perf_event *event __maybe_unused, 474 u64 file_offset __maybe_unused) 475 { 476 dump_printf(": unhandled!\n"); 477 return 0; 478 } 479 480 void perf_tool__fill_defaults(struct perf_tool *tool) 481 { 482 if (tool->sample == NULL) 483 tool->sample = process_event_sample_stub; 484 if (tool->mmap == NULL) 485 tool->mmap = process_event_stub; 486 if (tool->mmap2 == NULL) 487 tool->mmap2 = process_event_stub; 488 if (tool->comm == NULL) 489 tool->comm = process_event_stub; 490 if (tool->namespaces == NULL) 491 tool->namespaces = process_event_stub; 492 if (tool->cgroup == NULL) 493 tool->cgroup = process_event_stub; 494 if (tool->fork == NULL) 495 tool->fork = process_event_stub; 496 if (tool->exit == NULL) 497 tool->exit = process_event_stub; 498 if (tool->lost == NULL) 499 tool->lost = perf_event__process_lost; 500 if (tool->lost_samples == NULL) 501 tool->lost_samples = perf_event__process_lost_samples; 502 if (tool->aux == NULL) 503 tool->aux = perf_event__process_aux; 504 if (tool->itrace_start == NULL) 505 tool->itrace_start = perf_event__process_itrace_start; 506 if (tool->context_switch == NULL) 507 tool->context_switch = perf_event__process_switch; 508 if (tool->ksymbol == NULL) 509 tool->ksymbol = perf_event__process_ksymbol; 510 if (tool->bpf == NULL) 511 tool->bpf = perf_event__process_bpf; 512 if (tool->text_poke == NULL) 513 tool->text_poke = perf_event__process_text_poke; 514 if (tool->aux_output_hw_id == NULL) 515 tool->aux_output_hw_id = perf_event__process_aux_output_hw_id; 516 if (tool->read == NULL) 517 tool->read = process_event_sample_stub; 518 if (tool->throttle == NULL) 519 tool->throttle = process_event_stub; 520 if (tool->unthrottle == NULL) 521 tool->unthrottle = process_event_stub; 522 if (tool->attr == NULL) 523 tool->attr = process_event_synth_attr_stub; 524 if (tool->event_update == NULL) 525 tool->event_update = process_event_synth_event_update_stub; 526 if (tool->tracing_data == NULL) 527 tool->tracing_data = process_event_synth_tracing_data_stub; 528 if (tool->build_id == NULL) 529 tool->build_id = process_event_op2_stub; 530 if (tool->finished_round == NULL) { 531 if (tool->ordered_events) 532 tool->finished_round = process_finished_round; 533 else 534 tool->finished_round = process_finished_round_stub; 535 } 536 if (tool->id_index == NULL) 537 tool->id_index = process_event_op2_stub; 538 if (tool->auxtrace_info == NULL) 539 tool->auxtrace_info = process_event_op2_stub; 540 if (tool->auxtrace == NULL) 541 tool->auxtrace = process_event_auxtrace_stub; 542 if (tool->auxtrace_error == NULL) 543 tool->auxtrace_error = process_event_op2_stub; 544 if (tool->thread_map == NULL) 545 tool->thread_map = process_event_thread_map_stub; 546 if (tool->cpu_map == NULL) 547 tool->cpu_map = process_event_cpu_map_stub; 548 if (tool->stat_config == NULL) 549 tool->stat_config = process_event_stat_config_stub; 550 if (tool->stat == NULL) 551 tool->stat = process_stat_stub; 552 if (tool->stat_round == NULL) 553 tool->stat_round = process_stat_round_stub; 554 if (tool->time_conv == NULL) 555 tool->time_conv = process_event_time_conv_stub; 556 if (tool->feature == NULL) 557 tool->feature = process_event_op2_stub; 558 if (tool->compressed == NULL) 559 tool->compressed = perf_session__process_compressed_event; 560 } 561 562 static void swap_sample_id_all(union perf_event *event, void *data) 563 { 564 void *end = (void *) event + event->header.size; 565 int size = end - data; 566 567 BUG_ON(size % sizeof(u64)); 568 mem_bswap_64(data, size); 569 } 570 571 static void perf_event__all64_swap(union perf_event *event, 572 bool sample_id_all __maybe_unused) 573 { 574 struct perf_event_header *hdr = &event->header; 575 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 576 } 577 578 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all) 579 { 580 event->comm.pid = bswap_32(event->comm.pid); 581 event->comm.tid = bswap_32(event->comm.tid); 582 583 if (sample_id_all) { 584 void *data = &event->comm.comm; 585 586 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 587 swap_sample_id_all(event, data); 588 } 589 } 590 591 static void perf_event__mmap_swap(union perf_event *event, 592 bool sample_id_all) 593 { 594 event->mmap.pid = bswap_32(event->mmap.pid); 595 event->mmap.tid = bswap_32(event->mmap.tid); 596 event->mmap.start = bswap_64(event->mmap.start); 597 event->mmap.len = bswap_64(event->mmap.len); 598 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 599 600 if (sample_id_all) { 601 void *data = &event->mmap.filename; 602 603 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 604 swap_sample_id_all(event, data); 605 } 606 } 607 608 static void perf_event__mmap2_swap(union perf_event *event, 609 bool sample_id_all) 610 { 611 event->mmap2.pid = bswap_32(event->mmap2.pid); 612 event->mmap2.tid = bswap_32(event->mmap2.tid); 613 event->mmap2.start = bswap_64(event->mmap2.start); 614 event->mmap2.len = bswap_64(event->mmap2.len); 615 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff); 616 617 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID)) { 618 event->mmap2.maj = bswap_32(event->mmap2.maj); 619 event->mmap2.min = bswap_32(event->mmap2.min); 620 event->mmap2.ino = bswap_64(event->mmap2.ino); 621 event->mmap2.ino_generation = bswap_64(event->mmap2.ino_generation); 622 } 623 624 if (sample_id_all) { 625 void *data = &event->mmap2.filename; 626 627 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 628 swap_sample_id_all(event, data); 629 } 630 } 631 static void perf_event__task_swap(union perf_event *event, bool sample_id_all) 632 { 633 event->fork.pid = bswap_32(event->fork.pid); 634 event->fork.tid = bswap_32(event->fork.tid); 635 event->fork.ppid = bswap_32(event->fork.ppid); 636 event->fork.ptid = bswap_32(event->fork.ptid); 637 event->fork.time = bswap_64(event->fork.time); 638 639 if (sample_id_all) 640 swap_sample_id_all(event, &event->fork + 1); 641 } 642 643 static void perf_event__read_swap(union perf_event *event, bool sample_id_all) 644 { 645 event->read.pid = bswap_32(event->read.pid); 646 event->read.tid = bswap_32(event->read.tid); 647 event->read.value = bswap_64(event->read.value); 648 event->read.time_enabled = bswap_64(event->read.time_enabled); 649 event->read.time_running = bswap_64(event->read.time_running); 650 event->read.id = bswap_64(event->read.id); 651 652 if (sample_id_all) 653 swap_sample_id_all(event, &event->read + 1); 654 } 655 656 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all) 657 { 658 event->aux.aux_offset = bswap_64(event->aux.aux_offset); 659 event->aux.aux_size = bswap_64(event->aux.aux_size); 660 event->aux.flags = bswap_64(event->aux.flags); 661 662 if (sample_id_all) 663 swap_sample_id_all(event, &event->aux + 1); 664 } 665 666 static void perf_event__itrace_start_swap(union perf_event *event, 667 bool sample_id_all) 668 { 669 event->itrace_start.pid = bswap_32(event->itrace_start.pid); 670 event->itrace_start.tid = bswap_32(event->itrace_start.tid); 671 672 if (sample_id_all) 673 swap_sample_id_all(event, &event->itrace_start + 1); 674 } 675 676 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all) 677 { 678 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) { 679 event->context_switch.next_prev_pid = 680 bswap_32(event->context_switch.next_prev_pid); 681 event->context_switch.next_prev_tid = 682 bswap_32(event->context_switch.next_prev_tid); 683 } 684 685 if (sample_id_all) 686 swap_sample_id_all(event, &event->context_switch + 1); 687 } 688 689 static void perf_event__text_poke_swap(union perf_event *event, bool sample_id_all) 690 { 691 event->text_poke.addr = bswap_64(event->text_poke.addr); 692 event->text_poke.old_len = bswap_16(event->text_poke.old_len); 693 event->text_poke.new_len = bswap_16(event->text_poke.new_len); 694 695 if (sample_id_all) { 696 size_t len = sizeof(event->text_poke.old_len) + 697 sizeof(event->text_poke.new_len) + 698 event->text_poke.old_len + 699 event->text_poke.new_len; 700 void *data = &event->text_poke.old_len; 701 702 data += PERF_ALIGN(len, sizeof(u64)); 703 swap_sample_id_all(event, data); 704 } 705 } 706 707 static void perf_event__throttle_swap(union perf_event *event, 708 bool sample_id_all) 709 { 710 event->throttle.time = bswap_64(event->throttle.time); 711 event->throttle.id = bswap_64(event->throttle.id); 712 event->throttle.stream_id = bswap_64(event->throttle.stream_id); 713 714 if (sample_id_all) 715 swap_sample_id_all(event, &event->throttle + 1); 716 } 717 718 static void perf_event__namespaces_swap(union perf_event *event, 719 bool sample_id_all) 720 { 721 u64 i; 722 723 event->namespaces.pid = bswap_32(event->namespaces.pid); 724 event->namespaces.tid = bswap_32(event->namespaces.tid); 725 event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces); 726 727 for (i = 0; i < event->namespaces.nr_namespaces; i++) { 728 struct perf_ns_link_info *ns = &event->namespaces.link_info[i]; 729 730 ns->dev = bswap_64(ns->dev); 731 ns->ino = bswap_64(ns->ino); 732 } 733 734 if (sample_id_all) 735 swap_sample_id_all(event, &event->namespaces.link_info[i]); 736 } 737 738 static void perf_event__cgroup_swap(union perf_event *event, bool sample_id_all) 739 { 740 event->cgroup.id = bswap_64(event->cgroup.id); 741 742 if (sample_id_all) { 743 void *data = &event->cgroup.path; 744 745 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 746 swap_sample_id_all(event, data); 747 } 748 } 749 750 static u8 revbyte(u8 b) 751 { 752 int rev = (b >> 4) | ((b & 0xf) << 4); 753 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2); 754 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1); 755 return (u8) rev; 756 } 757 758 /* 759 * XXX this is hack in attempt to carry flags bitfield 760 * through endian village. ABI says: 761 * 762 * Bit-fields are allocated from right to left (least to most significant) 763 * on little-endian implementations and from left to right (most to least 764 * significant) on big-endian implementations. 765 * 766 * The above seems to be byte specific, so we need to reverse each 767 * byte of the bitfield. 'Internet' also says this might be implementation 768 * specific and we probably need proper fix and carry perf_event_attr 769 * bitfield flags in separate data file FEAT_ section. Thought this seems 770 * to work for now. 771 */ 772 static void swap_bitfield(u8 *p, unsigned len) 773 { 774 unsigned i; 775 776 for (i = 0; i < len; i++) { 777 *p = revbyte(*p); 778 p++; 779 } 780 } 781 782 /* exported for swapping attributes in file header */ 783 void perf_event__attr_swap(struct perf_event_attr *attr) 784 { 785 attr->type = bswap_32(attr->type); 786 attr->size = bswap_32(attr->size); 787 788 #define bswap_safe(f, n) \ 789 (attr->size > (offsetof(struct perf_event_attr, f) + \ 790 sizeof(attr->f) * (n))) 791 #define bswap_field(f, sz) \ 792 do { \ 793 if (bswap_safe(f, 0)) \ 794 attr->f = bswap_##sz(attr->f); \ 795 } while(0) 796 #define bswap_field_16(f) bswap_field(f, 16) 797 #define bswap_field_32(f) bswap_field(f, 32) 798 #define bswap_field_64(f) bswap_field(f, 64) 799 800 bswap_field_64(config); 801 bswap_field_64(sample_period); 802 bswap_field_64(sample_type); 803 bswap_field_64(read_format); 804 bswap_field_32(wakeup_events); 805 bswap_field_32(bp_type); 806 bswap_field_64(bp_addr); 807 bswap_field_64(bp_len); 808 bswap_field_64(branch_sample_type); 809 bswap_field_64(sample_regs_user); 810 bswap_field_32(sample_stack_user); 811 bswap_field_32(aux_watermark); 812 bswap_field_16(sample_max_stack); 813 bswap_field_32(aux_sample_size); 814 815 /* 816 * After read_format are bitfields. Check read_format because 817 * we are unable to use offsetof on bitfield. 818 */ 819 if (bswap_safe(read_format, 1)) 820 swap_bitfield((u8 *) (&attr->read_format + 1), 821 sizeof(u64)); 822 #undef bswap_field_64 823 #undef bswap_field_32 824 #undef bswap_field 825 #undef bswap_safe 826 } 827 828 static void perf_event__hdr_attr_swap(union perf_event *event, 829 bool sample_id_all __maybe_unused) 830 { 831 size_t size; 832 833 perf_event__attr_swap(&event->attr.attr); 834 835 size = event->header.size; 836 size -= (void *)&event->attr.id - (void *)event; 837 mem_bswap_64(event->attr.id, size); 838 } 839 840 static void perf_event__event_update_swap(union perf_event *event, 841 bool sample_id_all __maybe_unused) 842 { 843 event->event_update.type = bswap_64(event->event_update.type); 844 event->event_update.id = bswap_64(event->event_update.id); 845 } 846 847 static void perf_event__event_type_swap(union perf_event *event, 848 bool sample_id_all __maybe_unused) 849 { 850 event->event_type.event_type.event_id = 851 bswap_64(event->event_type.event_type.event_id); 852 } 853 854 static void perf_event__tracing_data_swap(union perf_event *event, 855 bool sample_id_all __maybe_unused) 856 { 857 event->tracing_data.size = bswap_32(event->tracing_data.size); 858 } 859 860 static void perf_event__auxtrace_info_swap(union perf_event *event, 861 bool sample_id_all __maybe_unused) 862 { 863 size_t size; 864 865 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type); 866 867 size = event->header.size; 868 size -= (void *)&event->auxtrace_info.priv - (void *)event; 869 mem_bswap_64(event->auxtrace_info.priv, size); 870 } 871 872 static void perf_event__auxtrace_swap(union perf_event *event, 873 bool sample_id_all __maybe_unused) 874 { 875 event->auxtrace.size = bswap_64(event->auxtrace.size); 876 event->auxtrace.offset = bswap_64(event->auxtrace.offset); 877 event->auxtrace.reference = bswap_64(event->auxtrace.reference); 878 event->auxtrace.idx = bswap_32(event->auxtrace.idx); 879 event->auxtrace.tid = bswap_32(event->auxtrace.tid); 880 event->auxtrace.cpu = bswap_32(event->auxtrace.cpu); 881 } 882 883 static void perf_event__auxtrace_error_swap(union perf_event *event, 884 bool sample_id_all __maybe_unused) 885 { 886 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type); 887 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code); 888 event->auxtrace_error.cpu = bswap_32(event->auxtrace_error.cpu); 889 event->auxtrace_error.pid = bswap_32(event->auxtrace_error.pid); 890 event->auxtrace_error.tid = bswap_32(event->auxtrace_error.tid); 891 event->auxtrace_error.fmt = bswap_32(event->auxtrace_error.fmt); 892 event->auxtrace_error.ip = bswap_64(event->auxtrace_error.ip); 893 if (event->auxtrace_error.fmt) 894 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time); 895 } 896 897 static void perf_event__thread_map_swap(union perf_event *event, 898 bool sample_id_all __maybe_unused) 899 { 900 unsigned i; 901 902 event->thread_map.nr = bswap_64(event->thread_map.nr); 903 904 for (i = 0; i < event->thread_map.nr; i++) 905 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid); 906 } 907 908 static void perf_event__cpu_map_swap(union perf_event *event, 909 bool sample_id_all __maybe_unused) 910 { 911 struct perf_record_cpu_map_data *data = &event->cpu_map.data; 912 struct cpu_map_entries *cpus; 913 struct perf_record_record_cpu_map *mask; 914 unsigned i; 915 916 data->type = bswap_16(data->type); 917 918 switch (data->type) { 919 case PERF_CPU_MAP__CPUS: 920 cpus = (struct cpu_map_entries *)data->data; 921 922 cpus->nr = bswap_16(cpus->nr); 923 924 for (i = 0; i < cpus->nr; i++) 925 cpus->cpu[i] = bswap_16(cpus->cpu[i]); 926 break; 927 case PERF_CPU_MAP__MASK: 928 mask = (struct perf_record_record_cpu_map *)data->data; 929 930 mask->nr = bswap_16(mask->nr); 931 mask->long_size = bswap_16(mask->long_size); 932 933 switch (mask->long_size) { 934 case 4: mem_bswap_32(&mask->mask, mask->nr); break; 935 case 8: mem_bswap_64(&mask->mask, mask->nr); break; 936 default: 937 pr_err("cpu_map swap: unsupported long size\n"); 938 } 939 default: 940 break; 941 } 942 } 943 944 static void perf_event__stat_config_swap(union perf_event *event, 945 bool sample_id_all __maybe_unused) 946 { 947 u64 size; 948 949 size = bswap_64(event->stat_config.nr) * sizeof(event->stat_config.data[0]); 950 size += 1; /* nr item itself */ 951 mem_bswap_64(&event->stat_config.nr, size); 952 } 953 954 static void perf_event__stat_swap(union perf_event *event, 955 bool sample_id_all __maybe_unused) 956 { 957 event->stat.id = bswap_64(event->stat.id); 958 event->stat.thread = bswap_32(event->stat.thread); 959 event->stat.cpu = bswap_32(event->stat.cpu); 960 event->stat.val = bswap_64(event->stat.val); 961 event->stat.ena = bswap_64(event->stat.ena); 962 event->stat.run = bswap_64(event->stat.run); 963 } 964 965 static void perf_event__stat_round_swap(union perf_event *event, 966 bool sample_id_all __maybe_unused) 967 { 968 event->stat_round.type = bswap_64(event->stat_round.type); 969 event->stat_round.time = bswap_64(event->stat_round.time); 970 } 971 972 static void perf_event__time_conv_swap(union perf_event *event, 973 bool sample_id_all __maybe_unused) 974 { 975 event->time_conv.time_shift = bswap_64(event->time_conv.time_shift); 976 event->time_conv.time_mult = bswap_64(event->time_conv.time_mult); 977 event->time_conv.time_zero = bswap_64(event->time_conv.time_zero); 978 979 if (event_contains(event->time_conv, time_cycles)) { 980 event->time_conv.time_cycles = bswap_64(event->time_conv.time_cycles); 981 event->time_conv.time_mask = bswap_64(event->time_conv.time_mask); 982 } 983 } 984 985 typedef void (*perf_event__swap_op)(union perf_event *event, 986 bool sample_id_all); 987 988 static perf_event__swap_op perf_event__swap_ops[] = { 989 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 990 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap, 991 [PERF_RECORD_COMM] = perf_event__comm_swap, 992 [PERF_RECORD_FORK] = perf_event__task_swap, 993 [PERF_RECORD_EXIT] = perf_event__task_swap, 994 [PERF_RECORD_LOST] = perf_event__all64_swap, 995 [PERF_RECORD_READ] = perf_event__read_swap, 996 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap, 997 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap, 998 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 999 [PERF_RECORD_AUX] = perf_event__aux_swap, 1000 [PERF_RECORD_ITRACE_START] = perf_event__itrace_start_swap, 1001 [PERF_RECORD_LOST_SAMPLES] = perf_event__all64_swap, 1002 [PERF_RECORD_SWITCH] = perf_event__switch_swap, 1003 [PERF_RECORD_SWITCH_CPU_WIDE] = perf_event__switch_swap, 1004 [PERF_RECORD_NAMESPACES] = perf_event__namespaces_swap, 1005 [PERF_RECORD_CGROUP] = perf_event__cgroup_swap, 1006 [PERF_RECORD_TEXT_POKE] = perf_event__text_poke_swap, 1007 [PERF_RECORD_AUX_OUTPUT_HW_ID] = perf_event__all64_swap, 1008 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 1009 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 1010 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 1011 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 1012 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap, 1013 [PERF_RECORD_AUXTRACE_INFO] = perf_event__auxtrace_info_swap, 1014 [PERF_RECORD_AUXTRACE] = perf_event__auxtrace_swap, 1015 [PERF_RECORD_AUXTRACE_ERROR] = perf_event__auxtrace_error_swap, 1016 [PERF_RECORD_THREAD_MAP] = perf_event__thread_map_swap, 1017 [PERF_RECORD_CPU_MAP] = perf_event__cpu_map_swap, 1018 [PERF_RECORD_STAT_CONFIG] = perf_event__stat_config_swap, 1019 [PERF_RECORD_STAT] = perf_event__stat_swap, 1020 [PERF_RECORD_STAT_ROUND] = perf_event__stat_round_swap, 1021 [PERF_RECORD_EVENT_UPDATE] = perf_event__event_update_swap, 1022 [PERF_RECORD_TIME_CONV] = perf_event__time_conv_swap, 1023 [PERF_RECORD_HEADER_MAX] = NULL, 1024 }; 1025 1026 /* 1027 * When perf record finishes a pass on every buffers, it records this pseudo 1028 * event. 1029 * We record the max timestamp t found in the pass n. 1030 * Assuming these timestamps are monotonic across cpus, we know that if 1031 * a buffer still has events with timestamps below t, they will be all 1032 * available and then read in the pass n + 1. 1033 * Hence when we start to read the pass n + 2, we can safely flush every 1034 * events with timestamps below t. 1035 * 1036 * ============ PASS n ================= 1037 * CPU 0 | CPU 1 1038 * | 1039 * cnt1 timestamps | cnt2 timestamps 1040 * 1 | 2 1041 * 2 | 3 1042 * - | 4 <--- max recorded 1043 * 1044 * ============ PASS n + 1 ============== 1045 * CPU 0 | CPU 1 1046 * | 1047 * cnt1 timestamps | cnt2 timestamps 1048 * 3 | 5 1049 * 4 | 6 1050 * 5 | 7 <---- max recorded 1051 * 1052 * Flush every events below timestamp 4 1053 * 1054 * ============ PASS n + 2 ============== 1055 * CPU 0 | CPU 1 1056 * | 1057 * cnt1 timestamps | cnt2 timestamps 1058 * 6 | 8 1059 * 7 | 9 1060 * - | 10 1061 * 1062 * Flush every events below timestamp 7 1063 * etc... 1064 */ 1065 static int process_finished_round(struct perf_tool *tool __maybe_unused, 1066 union perf_event *event __maybe_unused, 1067 struct ordered_events *oe) 1068 { 1069 if (dump_trace) 1070 fprintf(stdout, "\n"); 1071 return ordered_events__flush(oe, OE_FLUSH__ROUND); 1072 } 1073 1074 int perf_session__queue_event(struct perf_session *s, union perf_event *event, 1075 u64 timestamp, u64 file_offset) 1076 { 1077 return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset); 1078 } 1079 1080 static void callchain__lbr_callstack_printf(struct perf_sample *sample) 1081 { 1082 struct ip_callchain *callchain = sample->callchain; 1083 struct branch_stack *lbr_stack = sample->branch_stack; 1084 struct branch_entry *entries = perf_sample__branch_entries(sample); 1085 u64 kernel_callchain_nr = callchain->nr; 1086 unsigned int i; 1087 1088 for (i = 0; i < kernel_callchain_nr; i++) { 1089 if (callchain->ips[i] == PERF_CONTEXT_USER) 1090 break; 1091 } 1092 1093 if ((i != kernel_callchain_nr) && lbr_stack->nr) { 1094 u64 total_nr; 1095 /* 1096 * LBR callstack can only get user call chain, 1097 * i is kernel call chain number, 1098 * 1 is PERF_CONTEXT_USER. 1099 * 1100 * The user call chain is stored in LBR registers. 1101 * LBR are pair registers. The caller is stored 1102 * in "from" register, while the callee is stored 1103 * in "to" register. 1104 * For example, there is a call stack 1105 * "A"->"B"->"C"->"D". 1106 * The LBR registers will be recorded like 1107 * "C"->"D", "B"->"C", "A"->"B". 1108 * So only the first "to" register and all "from" 1109 * registers are needed to construct the whole stack. 1110 */ 1111 total_nr = i + 1 + lbr_stack->nr + 1; 1112 kernel_callchain_nr = i + 1; 1113 1114 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr); 1115 1116 for (i = 0; i < kernel_callchain_nr; i++) 1117 printf("..... %2d: %016" PRIx64 "\n", 1118 i, callchain->ips[i]); 1119 1120 printf("..... %2d: %016" PRIx64 "\n", 1121 (int)(kernel_callchain_nr), entries[0].to); 1122 for (i = 0; i < lbr_stack->nr; i++) 1123 printf("..... %2d: %016" PRIx64 "\n", 1124 (int)(i + kernel_callchain_nr + 1), entries[i].from); 1125 } 1126 } 1127 1128 static void callchain__printf(struct evsel *evsel, 1129 struct perf_sample *sample) 1130 { 1131 unsigned int i; 1132 struct ip_callchain *callchain = sample->callchain; 1133 1134 if (evsel__has_branch_callstack(evsel)) 1135 callchain__lbr_callstack_printf(sample); 1136 1137 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr); 1138 1139 for (i = 0; i < callchain->nr; i++) 1140 printf("..... %2d: %016" PRIx64 "\n", 1141 i, callchain->ips[i]); 1142 } 1143 1144 static void branch_stack__printf(struct perf_sample *sample, bool callstack) 1145 { 1146 struct branch_entry *entries = perf_sample__branch_entries(sample); 1147 uint64_t i; 1148 1149 printf("%s: nr:%" PRIu64 "\n", 1150 !callstack ? "... branch stack" : "... branch callstack", 1151 sample->branch_stack->nr); 1152 1153 for (i = 0; i < sample->branch_stack->nr; i++) { 1154 struct branch_entry *e = &entries[i]; 1155 1156 if (!callstack) { 1157 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x\n", 1158 i, e->from, e->to, 1159 (unsigned short)e->flags.cycles, 1160 e->flags.mispred ? "M" : " ", 1161 e->flags.predicted ? "P" : " ", 1162 e->flags.abort ? "A" : " ", 1163 e->flags.in_tx ? "T" : " ", 1164 (unsigned)e->flags.reserved); 1165 } else { 1166 printf("..... %2"PRIu64": %016" PRIx64 "\n", 1167 i, i > 0 ? e->from : e->to); 1168 } 1169 } 1170 } 1171 1172 static void regs_dump__printf(u64 mask, u64 *regs, const char *arch) 1173 { 1174 unsigned rid, i = 0; 1175 1176 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) { 1177 u64 val = regs[i++]; 1178 1179 printf(".... %-5s 0x%016" PRIx64 "\n", 1180 perf_reg_name(rid, arch), val); 1181 } 1182 } 1183 1184 static const char *regs_abi[] = { 1185 [PERF_SAMPLE_REGS_ABI_NONE] = "none", 1186 [PERF_SAMPLE_REGS_ABI_32] = "32-bit", 1187 [PERF_SAMPLE_REGS_ABI_64] = "64-bit", 1188 }; 1189 1190 static inline const char *regs_dump_abi(struct regs_dump *d) 1191 { 1192 if (d->abi > PERF_SAMPLE_REGS_ABI_64) 1193 return "unknown"; 1194 1195 return regs_abi[d->abi]; 1196 } 1197 1198 static void regs__printf(const char *type, struct regs_dump *regs, const char *arch) 1199 { 1200 u64 mask = regs->mask; 1201 1202 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n", 1203 type, 1204 mask, 1205 regs_dump_abi(regs)); 1206 1207 regs_dump__printf(mask, regs->regs, arch); 1208 } 1209 1210 static void regs_user__printf(struct perf_sample *sample, const char *arch) 1211 { 1212 struct regs_dump *user_regs = &sample->user_regs; 1213 1214 if (user_regs->regs) 1215 regs__printf("user", user_regs, arch); 1216 } 1217 1218 static void regs_intr__printf(struct perf_sample *sample, const char *arch) 1219 { 1220 struct regs_dump *intr_regs = &sample->intr_regs; 1221 1222 if (intr_regs->regs) 1223 regs__printf("intr", intr_regs, arch); 1224 } 1225 1226 static void stack_user__printf(struct stack_dump *dump) 1227 { 1228 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 1229 dump->size, dump->offset); 1230 } 1231 1232 static void evlist__print_tstamp(struct evlist *evlist, union perf_event *event, struct perf_sample *sample) 1233 { 1234 u64 sample_type = __evlist__combined_sample_type(evlist); 1235 1236 if (event->header.type != PERF_RECORD_SAMPLE && 1237 !evlist__sample_id_all(evlist)) { 1238 fputs("-1 -1 ", stdout); 1239 return; 1240 } 1241 1242 if ((sample_type & PERF_SAMPLE_CPU)) 1243 printf("%u ", sample->cpu); 1244 1245 if (sample_type & PERF_SAMPLE_TIME) 1246 printf("%" PRIu64 " ", sample->time); 1247 } 1248 1249 static void sample_read__printf(struct perf_sample *sample, u64 read_format) 1250 { 1251 printf("... sample_read:\n"); 1252 1253 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1254 printf("...... time enabled %016" PRIx64 "\n", 1255 sample->read.time_enabled); 1256 1257 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1258 printf("...... time running %016" PRIx64 "\n", 1259 sample->read.time_running); 1260 1261 if (read_format & PERF_FORMAT_GROUP) { 1262 u64 i; 1263 1264 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr); 1265 1266 for (i = 0; i < sample->read.group.nr; i++) { 1267 struct sample_read_value *value; 1268 1269 value = &sample->read.group.values[i]; 1270 printf("..... id %016" PRIx64 1271 ", value %016" PRIx64 "\n", 1272 value->id, value->value); 1273 } 1274 } else 1275 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n", 1276 sample->read.one.id, sample->read.one.value); 1277 } 1278 1279 static void dump_event(struct evlist *evlist, union perf_event *event, 1280 u64 file_offset, struct perf_sample *sample) 1281 { 1282 if (!dump_trace) 1283 return; 1284 1285 printf("\n%#" PRIx64 " [%#x]: event: %d\n", 1286 file_offset, event->header.size, event->header.type); 1287 1288 trace_event(event); 1289 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw) 1290 evlist->trace_event_sample_raw(evlist, event, sample); 1291 1292 if (sample) 1293 evlist__print_tstamp(evlist, event, sample); 1294 1295 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 1296 event->header.size, perf_event__name(event->header.type)); 1297 } 1298 1299 char *get_page_size_name(u64 size, char *str) 1300 { 1301 if (!size || !unit_number__scnprintf(str, PAGE_SIZE_NAME_LEN, size)) 1302 snprintf(str, PAGE_SIZE_NAME_LEN, "%s", "N/A"); 1303 1304 return str; 1305 } 1306 1307 static void dump_sample(struct evsel *evsel, union perf_event *event, 1308 struct perf_sample *sample, const char *arch) 1309 { 1310 u64 sample_type; 1311 char str[PAGE_SIZE_NAME_LEN]; 1312 1313 if (!dump_trace) 1314 return; 1315 1316 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 1317 event->header.misc, sample->pid, sample->tid, sample->ip, 1318 sample->period, sample->addr); 1319 1320 sample_type = evsel->core.attr.sample_type; 1321 1322 if (evsel__has_callchain(evsel)) 1323 callchain__printf(evsel, sample); 1324 1325 if (evsel__has_br_stack(evsel)) 1326 branch_stack__printf(sample, evsel__has_branch_callstack(evsel)); 1327 1328 if (sample_type & PERF_SAMPLE_REGS_USER) 1329 regs_user__printf(sample, arch); 1330 1331 if (sample_type & PERF_SAMPLE_REGS_INTR) 1332 regs_intr__printf(sample, arch); 1333 1334 if (sample_type & PERF_SAMPLE_STACK_USER) 1335 stack_user__printf(&sample->user_stack); 1336 1337 if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) { 1338 printf("... weight: %" PRIu64 "", sample->weight); 1339 if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) { 1340 printf(",0x%"PRIx16"", sample->ins_lat); 1341 printf(",0x%"PRIx16"", sample->p_stage_cyc); 1342 } 1343 printf("\n"); 1344 } 1345 1346 if (sample_type & PERF_SAMPLE_DATA_SRC) 1347 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src); 1348 1349 if (sample_type & PERF_SAMPLE_PHYS_ADDR) 1350 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr); 1351 1352 if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE) 1353 printf(" .. data page size: %s\n", get_page_size_name(sample->data_page_size, str)); 1354 1355 if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE) 1356 printf(" .. code page size: %s\n", get_page_size_name(sample->code_page_size, str)); 1357 1358 if (sample_type & PERF_SAMPLE_TRANSACTION) 1359 printf("... transaction: %" PRIx64 "\n", sample->transaction); 1360 1361 if (sample_type & PERF_SAMPLE_READ) 1362 sample_read__printf(sample, evsel->core.attr.read_format); 1363 } 1364 1365 static void dump_read(struct evsel *evsel, union perf_event *event) 1366 { 1367 struct perf_record_read *read_event = &event->read; 1368 u64 read_format; 1369 1370 if (!dump_trace) 1371 return; 1372 1373 printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid, 1374 evsel__name(evsel), event->read.value); 1375 1376 if (!evsel) 1377 return; 1378 1379 read_format = evsel->core.attr.read_format; 1380 1381 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1382 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled); 1383 1384 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1385 printf("... time running : %" PRI_lu64 "\n", read_event->time_running); 1386 1387 if (read_format & PERF_FORMAT_ID) 1388 printf("... id : %" PRI_lu64 "\n", read_event->id); 1389 } 1390 1391 static struct machine *machines__find_for_cpumode(struct machines *machines, 1392 union perf_event *event, 1393 struct perf_sample *sample) 1394 { 1395 if (perf_guest && 1396 ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 1397 (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) { 1398 u32 pid; 1399 1400 if (event->header.type == PERF_RECORD_MMAP 1401 || event->header.type == PERF_RECORD_MMAP2) 1402 pid = event->mmap.pid; 1403 else 1404 pid = sample->pid; 1405 1406 return machines__find_guest(machines, pid); 1407 } 1408 1409 return &machines->host; 1410 } 1411 1412 static int deliver_sample_value(struct evlist *evlist, 1413 struct perf_tool *tool, 1414 union perf_event *event, 1415 struct perf_sample *sample, 1416 struct sample_read_value *v, 1417 struct machine *machine) 1418 { 1419 struct perf_sample_id *sid = evlist__id2sid(evlist, v->id); 1420 struct evsel *evsel; 1421 1422 if (sid) { 1423 sample->id = v->id; 1424 sample->period = v->value - sid->period; 1425 sid->period = v->value; 1426 } 1427 1428 if (!sid || sid->evsel == NULL) { 1429 ++evlist->stats.nr_unknown_id; 1430 return 0; 1431 } 1432 1433 /* 1434 * There's no reason to deliver sample 1435 * for zero period, bail out. 1436 */ 1437 if (!sample->period) 1438 return 0; 1439 1440 evsel = container_of(sid->evsel, struct evsel, core); 1441 return tool->sample(tool, event, sample, evsel, machine); 1442 } 1443 1444 static int deliver_sample_group(struct evlist *evlist, 1445 struct perf_tool *tool, 1446 union perf_event *event, 1447 struct perf_sample *sample, 1448 struct machine *machine) 1449 { 1450 int ret = -EINVAL; 1451 u64 i; 1452 1453 for (i = 0; i < sample->read.group.nr; i++) { 1454 ret = deliver_sample_value(evlist, tool, event, sample, 1455 &sample->read.group.values[i], 1456 machine); 1457 if (ret) 1458 break; 1459 } 1460 1461 return ret; 1462 } 1463 1464 static int evlist__deliver_sample(struct evlist *evlist, struct perf_tool *tool, 1465 union perf_event *event, struct perf_sample *sample, 1466 struct evsel *evsel, struct machine *machine) 1467 { 1468 /* We know evsel != NULL. */ 1469 u64 sample_type = evsel->core.attr.sample_type; 1470 u64 read_format = evsel->core.attr.read_format; 1471 1472 /* Standard sample delivery. */ 1473 if (!(sample_type & PERF_SAMPLE_READ)) 1474 return tool->sample(tool, event, sample, evsel, machine); 1475 1476 /* For PERF_SAMPLE_READ we have either single or group mode. */ 1477 if (read_format & PERF_FORMAT_GROUP) 1478 return deliver_sample_group(evlist, tool, event, sample, 1479 machine); 1480 else 1481 return deliver_sample_value(evlist, tool, event, sample, 1482 &sample->read.one, machine); 1483 } 1484 1485 static int machines__deliver_event(struct machines *machines, 1486 struct evlist *evlist, 1487 union perf_event *event, 1488 struct perf_sample *sample, 1489 struct perf_tool *tool, u64 file_offset) 1490 { 1491 struct evsel *evsel; 1492 struct machine *machine; 1493 1494 dump_event(evlist, event, file_offset, sample); 1495 1496 evsel = evlist__id2evsel(evlist, sample->id); 1497 1498 machine = machines__find_for_cpumode(machines, event, sample); 1499 1500 switch (event->header.type) { 1501 case PERF_RECORD_SAMPLE: 1502 if (evsel == NULL) { 1503 ++evlist->stats.nr_unknown_id; 1504 return 0; 1505 } 1506 if (machine == NULL) { 1507 ++evlist->stats.nr_unprocessable_samples; 1508 dump_sample(evsel, event, sample, perf_env__arch(NULL)); 1509 return 0; 1510 } 1511 dump_sample(evsel, event, sample, perf_env__arch(machine->env)); 1512 return evlist__deliver_sample(evlist, tool, event, sample, evsel, machine); 1513 case PERF_RECORD_MMAP: 1514 return tool->mmap(tool, event, sample, machine); 1515 case PERF_RECORD_MMAP2: 1516 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT) 1517 ++evlist->stats.nr_proc_map_timeout; 1518 return tool->mmap2(tool, event, sample, machine); 1519 case PERF_RECORD_COMM: 1520 return tool->comm(tool, event, sample, machine); 1521 case PERF_RECORD_NAMESPACES: 1522 return tool->namespaces(tool, event, sample, machine); 1523 case PERF_RECORD_CGROUP: 1524 return tool->cgroup(tool, event, sample, machine); 1525 case PERF_RECORD_FORK: 1526 return tool->fork(tool, event, sample, machine); 1527 case PERF_RECORD_EXIT: 1528 return tool->exit(tool, event, sample, machine); 1529 case PERF_RECORD_LOST: 1530 if (tool->lost == perf_event__process_lost) 1531 evlist->stats.total_lost += event->lost.lost; 1532 return tool->lost(tool, event, sample, machine); 1533 case PERF_RECORD_LOST_SAMPLES: 1534 if (tool->lost_samples == perf_event__process_lost_samples) 1535 evlist->stats.total_lost_samples += event->lost_samples.lost; 1536 return tool->lost_samples(tool, event, sample, machine); 1537 case PERF_RECORD_READ: 1538 dump_read(evsel, event); 1539 return tool->read(tool, event, sample, evsel, machine); 1540 case PERF_RECORD_THROTTLE: 1541 return tool->throttle(tool, event, sample, machine); 1542 case PERF_RECORD_UNTHROTTLE: 1543 return tool->unthrottle(tool, event, sample, machine); 1544 case PERF_RECORD_AUX: 1545 if (tool->aux == perf_event__process_aux) { 1546 if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) 1547 evlist->stats.total_aux_lost += 1; 1548 if (event->aux.flags & PERF_AUX_FLAG_PARTIAL) 1549 evlist->stats.total_aux_partial += 1; 1550 if (event->aux.flags & PERF_AUX_FLAG_COLLISION) 1551 evlist->stats.total_aux_collision += 1; 1552 } 1553 return tool->aux(tool, event, sample, machine); 1554 case PERF_RECORD_ITRACE_START: 1555 return tool->itrace_start(tool, event, sample, machine); 1556 case PERF_RECORD_SWITCH: 1557 case PERF_RECORD_SWITCH_CPU_WIDE: 1558 return tool->context_switch(tool, event, sample, machine); 1559 case PERF_RECORD_KSYMBOL: 1560 return tool->ksymbol(tool, event, sample, machine); 1561 case PERF_RECORD_BPF_EVENT: 1562 return tool->bpf(tool, event, sample, machine); 1563 case PERF_RECORD_TEXT_POKE: 1564 return tool->text_poke(tool, event, sample, machine); 1565 case PERF_RECORD_AUX_OUTPUT_HW_ID: 1566 return tool->aux_output_hw_id(tool, event, sample, machine); 1567 default: 1568 ++evlist->stats.nr_unknown_events; 1569 return -1; 1570 } 1571 } 1572 1573 static int perf_session__deliver_event(struct perf_session *session, 1574 union perf_event *event, 1575 struct perf_tool *tool, 1576 u64 file_offset) 1577 { 1578 struct perf_sample sample; 1579 int ret = evlist__parse_sample(session->evlist, event, &sample); 1580 1581 if (ret) { 1582 pr_err("Can't parse sample, err = %d\n", ret); 1583 return ret; 1584 } 1585 1586 ret = auxtrace__process_event(session, event, &sample, tool); 1587 if (ret < 0) 1588 return ret; 1589 if (ret > 0) 1590 return 0; 1591 1592 ret = machines__deliver_event(&session->machines, session->evlist, 1593 event, &sample, tool, file_offset); 1594 1595 if (dump_trace && sample.aux_sample.size) 1596 auxtrace__dump_auxtrace_sample(session, &sample); 1597 1598 return ret; 1599 } 1600 1601 static s64 perf_session__process_user_event(struct perf_session *session, 1602 union perf_event *event, 1603 u64 file_offset) 1604 { 1605 struct ordered_events *oe = &session->ordered_events; 1606 struct perf_tool *tool = session->tool; 1607 struct perf_sample sample = { .time = 0, }; 1608 int fd = perf_data__fd(session->data); 1609 int err; 1610 1611 if (event->header.type != PERF_RECORD_COMPRESSED || 1612 tool->compressed == perf_session__process_compressed_event_stub) 1613 dump_event(session->evlist, event, file_offset, &sample); 1614 1615 /* These events are processed right away */ 1616 switch (event->header.type) { 1617 case PERF_RECORD_HEADER_ATTR: 1618 err = tool->attr(tool, event, &session->evlist); 1619 if (err == 0) { 1620 perf_session__set_id_hdr_size(session); 1621 perf_session__set_comm_exec(session); 1622 } 1623 return err; 1624 case PERF_RECORD_EVENT_UPDATE: 1625 return tool->event_update(tool, event, &session->evlist); 1626 case PERF_RECORD_HEADER_EVENT_TYPE: 1627 /* 1628 * Deprecated, but we need to handle it for sake 1629 * of old data files create in pipe mode. 1630 */ 1631 return 0; 1632 case PERF_RECORD_HEADER_TRACING_DATA: 1633 /* 1634 * Setup for reading amidst mmap, but only when we 1635 * are in 'file' mode. The 'pipe' fd is in proper 1636 * place already. 1637 */ 1638 if (!perf_data__is_pipe(session->data)) 1639 lseek(fd, file_offset, SEEK_SET); 1640 return tool->tracing_data(session, event); 1641 case PERF_RECORD_HEADER_BUILD_ID: 1642 return tool->build_id(session, event); 1643 case PERF_RECORD_FINISHED_ROUND: 1644 return tool->finished_round(tool, event, oe); 1645 case PERF_RECORD_ID_INDEX: 1646 return tool->id_index(session, event); 1647 case PERF_RECORD_AUXTRACE_INFO: 1648 return tool->auxtrace_info(session, event); 1649 case PERF_RECORD_AUXTRACE: 1650 /* setup for reading amidst mmap */ 1651 lseek(fd, file_offset + event->header.size, SEEK_SET); 1652 return tool->auxtrace(session, event); 1653 case PERF_RECORD_AUXTRACE_ERROR: 1654 perf_session__auxtrace_error_inc(session, event); 1655 return tool->auxtrace_error(session, event); 1656 case PERF_RECORD_THREAD_MAP: 1657 return tool->thread_map(session, event); 1658 case PERF_RECORD_CPU_MAP: 1659 return tool->cpu_map(session, event); 1660 case PERF_RECORD_STAT_CONFIG: 1661 return tool->stat_config(session, event); 1662 case PERF_RECORD_STAT: 1663 return tool->stat(session, event); 1664 case PERF_RECORD_STAT_ROUND: 1665 return tool->stat_round(session, event); 1666 case PERF_RECORD_TIME_CONV: 1667 session->time_conv = event->time_conv; 1668 return tool->time_conv(session, event); 1669 case PERF_RECORD_HEADER_FEATURE: 1670 return tool->feature(session, event); 1671 case PERF_RECORD_COMPRESSED: 1672 err = tool->compressed(session, event, file_offset); 1673 if (err) 1674 dump_event(session->evlist, event, file_offset, &sample); 1675 return err; 1676 default: 1677 return -EINVAL; 1678 } 1679 } 1680 1681 int perf_session__deliver_synth_event(struct perf_session *session, 1682 union perf_event *event, 1683 struct perf_sample *sample) 1684 { 1685 struct evlist *evlist = session->evlist; 1686 struct perf_tool *tool = session->tool; 1687 1688 events_stats__inc(&evlist->stats, event->header.type); 1689 1690 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1691 return perf_session__process_user_event(session, event, 0); 1692 1693 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0); 1694 } 1695 1696 static void event_swap(union perf_event *event, bool sample_id_all) 1697 { 1698 perf_event__swap_op swap; 1699 1700 swap = perf_event__swap_ops[event->header.type]; 1701 if (swap) 1702 swap(event, sample_id_all); 1703 } 1704 1705 int perf_session__peek_event(struct perf_session *session, off_t file_offset, 1706 void *buf, size_t buf_sz, 1707 union perf_event **event_ptr, 1708 struct perf_sample *sample) 1709 { 1710 union perf_event *event; 1711 size_t hdr_sz, rest; 1712 int fd; 1713 1714 if (session->one_mmap && !session->header.needs_swap) { 1715 event = file_offset - session->one_mmap_offset + 1716 session->one_mmap_addr; 1717 goto out_parse_sample; 1718 } 1719 1720 if (perf_data__is_pipe(session->data)) 1721 return -1; 1722 1723 fd = perf_data__fd(session->data); 1724 hdr_sz = sizeof(struct perf_event_header); 1725 1726 if (buf_sz < hdr_sz) 1727 return -1; 1728 1729 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 || 1730 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz) 1731 return -1; 1732 1733 event = (union perf_event *)buf; 1734 1735 if (session->header.needs_swap) 1736 perf_event_header__bswap(&event->header); 1737 1738 if (event->header.size < hdr_sz || event->header.size > buf_sz) 1739 return -1; 1740 1741 buf += hdr_sz; 1742 rest = event->header.size - hdr_sz; 1743 1744 if (readn(fd, buf, rest) != (ssize_t)rest) 1745 return -1; 1746 1747 if (session->header.needs_swap) 1748 event_swap(event, evlist__sample_id_all(session->evlist)); 1749 1750 out_parse_sample: 1751 1752 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START && 1753 evlist__parse_sample(session->evlist, event, sample)) 1754 return -1; 1755 1756 *event_ptr = event; 1757 1758 return 0; 1759 } 1760 1761 int perf_session__peek_events(struct perf_session *session, u64 offset, 1762 u64 size, peek_events_cb_t cb, void *data) 1763 { 1764 u64 max_offset = offset + size; 1765 char buf[PERF_SAMPLE_MAX_SIZE]; 1766 union perf_event *event; 1767 int err; 1768 1769 do { 1770 err = perf_session__peek_event(session, offset, buf, 1771 PERF_SAMPLE_MAX_SIZE, &event, 1772 NULL); 1773 if (err) 1774 return err; 1775 1776 err = cb(session, event, offset, data); 1777 if (err) 1778 return err; 1779 1780 offset += event->header.size; 1781 if (event->header.type == PERF_RECORD_AUXTRACE) 1782 offset += event->auxtrace.size; 1783 1784 } while (offset < max_offset); 1785 1786 return err; 1787 } 1788 1789 static s64 perf_session__process_event(struct perf_session *session, 1790 union perf_event *event, u64 file_offset) 1791 { 1792 struct evlist *evlist = session->evlist; 1793 struct perf_tool *tool = session->tool; 1794 int ret; 1795 1796 if (session->header.needs_swap) 1797 event_swap(event, evlist__sample_id_all(evlist)); 1798 1799 if (event->header.type >= PERF_RECORD_HEADER_MAX) 1800 return -EINVAL; 1801 1802 events_stats__inc(&evlist->stats, event->header.type); 1803 1804 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1805 return perf_session__process_user_event(session, event, file_offset); 1806 1807 if (tool->ordered_events) { 1808 u64 timestamp = -1ULL; 1809 1810 ret = evlist__parse_sample_timestamp(evlist, event, ×tamp); 1811 if (ret && ret != -1) 1812 return ret; 1813 1814 ret = perf_session__queue_event(session, event, timestamp, file_offset); 1815 if (ret != -ETIME) 1816 return ret; 1817 } 1818 1819 return perf_session__deliver_event(session, event, tool, file_offset); 1820 } 1821 1822 void perf_event_header__bswap(struct perf_event_header *hdr) 1823 { 1824 hdr->type = bswap_32(hdr->type); 1825 hdr->misc = bswap_16(hdr->misc); 1826 hdr->size = bswap_16(hdr->size); 1827 } 1828 1829 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1830 { 1831 return machine__findnew_thread(&session->machines.host, -1, pid); 1832 } 1833 1834 int perf_session__register_idle_thread(struct perf_session *session) 1835 { 1836 struct thread *thread = machine__idle_thread(&session->machines.host); 1837 1838 /* machine__idle_thread() got the thread, so put it */ 1839 thread__put(thread); 1840 return thread ? 0 : -1; 1841 } 1842 1843 static void 1844 perf_session__warn_order(const struct perf_session *session) 1845 { 1846 const struct ordered_events *oe = &session->ordered_events; 1847 struct evsel *evsel; 1848 bool should_warn = true; 1849 1850 evlist__for_each_entry(session->evlist, evsel) { 1851 if (evsel->core.attr.write_backward) 1852 should_warn = false; 1853 } 1854 1855 if (!should_warn) 1856 return; 1857 if (oe->nr_unordered_events != 0) 1858 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events); 1859 } 1860 1861 static void perf_session__warn_about_errors(const struct perf_session *session) 1862 { 1863 const struct events_stats *stats = &session->evlist->stats; 1864 1865 if (session->tool->lost == perf_event__process_lost && 1866 stats->nr_events[PERF_RECORD_LOST] != 0) { 1867 ui__warning("Processed %d events and lost %d chunks!\n\n" 1868 "Check IO/CPU overload!\n\n", 1869 stats->nr_events[0], 1870 stats->nr_events[PERF_RECORD_LOST]); 1871 } 1872 1873 if (session->tool->lost_samples == perf_event__process_lost_samples) { 1874 double drop_rate; 1875 1876 drop_rate = (double)stats->total_lost_samples / 1877 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples); 1878 if (drop_rate > 0.05) { 1879 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n", 1880 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples, 1881 drop_rate * 100.0); 1882 } 1883 } 1884 1885 if (session->tool->aux == perf_event__process_aux && 1886 stats->total_aux_lost != 0) { 1887 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n", 1888 stats->total_aux_lost, 1889 stats->nr_events[PERF_RECORD_AUX]); 1890 } 1891 1892 if (session->tool->aux == perf_event__process_aux && 1893 stats->total_aux_partial != 0) { 1894 bool vmm_exclusive = false; 1895 1896 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive", 1897 &vmm_exclusive); 1898 1899 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n" 1900 "Are you running a KVM guest in the background?%s\n\n", 1901 stats->total_aux_partial, 1902 stats->nr_events[PERF_RECORD_AUX], 1903 vmm_exclusive ? 1904 "\nReloading kvm_intel module with vmm_exclusive=0\n" 1905 "will reduce the gaps to only guest's timeslices." : 1906 ""); 1907 } 1908 1909 if (session->tool->aux == perf_event__process_aux && 1910 stats->total_aux_collision != 0) { 1911 ui__warning("AUX data detected collision %" PRIu64 " times out of %u!\n\n", 1912 stats->total_aux_collision, 1913 stats->nr_events[PERF_RECORD_AUX]); 1914 } 1915 1916 if (stats->nr_unknown_events != 0) { 1917 ui__warning("Found %u unknown events!\n\n" 1918 "Is this an older tool processing a perf.data " 1919 "file generated by a more recent tool?\n\n" 1920 "If that is not the case, consider " 1921 "reporting to linux-kernel@vger.kernel.org.\n\n", 1922 stats->nr_unknown_events); 1923 } 1924 1925 if (stats->nr_unknown_id != 0) { 1926 ui__warning("%u samples with id not present in the header\n", 1927 stats->nr_unknown_id); 1928 } 1929 1930 if (stats->nr_invalid_chains != 0) { 1931 ui__warning("Found invalid callchains!\n\n" 1932 "%u out of %u events were discarded for this reason.\n\n" 1933 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 1934 stats->nr_invalid_chains, 1935 stats->nr_events[PERF_RECORD_SAMPLE]); 1936 } 1937 1938 if (stats->nr_unprocessable_samples != 0) { 1939 ui__warning("%u unprocessable samples recorded.\n" 1940 "Do you have a KVM guest running and not using 'perf kvm'?\n", 1941 stats->nr_unprocessable_samples); 1942 } 1943 1944 perf_session__warn_order(session); 1945 1946 events_stats__auxtrace_error_warn(stats); 1947 1948 if (stats->nr_proc_map_timeout != 0) { 1949 ui__warning("%d map information files for pre-existing threads were\n" 1950 "not processed, if there are samples for addresses they\n" 1951 "will not be resolved, you may find out which are these\n" 1952 "threads by running with -v and redirecting the output\n" 1953 "to a file.\n" 1954 "The time limit to process proc map is too short?\n" 1955 "Increase it by --proc-map-timeout\n", 1956 stats->nr_proc_map_timeout); 1957 } 1958 } 1959 1960 static int perf_session__flush_thread_stack(struct thread *thread, 1961 void *p __maybe_unused) 1962 { 1963 return thread_stack__flush(thread); 1964 } 1965 1966 static int perf_session__flush_thread_stacks(struct perf_session *session) 1967 { 1968 return machines__for_each_thread(&session->machines, 1969 perf_session__flush_thread_stack, 1970 NULL); 1971 } 1972 1973 volatile int session_done; 1974 1975 static int __perf_session__process_decomp_events(struct perf_session *session); 1976 1977 static int __perf_session__process_pipe_events(struct perf_session *session) 1978 { 1979 struct ordered_events *oe = &session->ordered_events; 1980 struct perf_tool *tool = session->tool; 1981 union perf_event *event; 1982 uint32_t size, cur_size = 0; 1983 void *buf = NULL; 1984 s64 skip = 0; 1985 u64 head; 1986 ssize_t err; 1987 void *p; 1988 1989 perf_tool__fill_defaults(tool); 1990 1991 head = 0; 1992 cur_size = sizeof(union perf_event); 1993 1994 buf = malloc(cur_size); 1995 if (!buf) 1996 return -errno; 1997 ordered_events__set_copy_on_queue(oe, true); 1998 more: 1999 event = buf; 2000 err = perf_data__read(session->data, event, 2001 sizeof(struct perf_event_header)); 2002 if (err <= 0) { 2003 if (err == 0) 2004 goto done; 2005 2006 pr_err("failed to read event header\n"); 2007 goto out_err; 2008 } 2009 2010 if (session->header.needs_swap) 2011 perf_event_header__bswap(&event->header); 2012 2013 size = event->header.size; 2014 if (size < sizeof(struct perf_event_header)) { 2015 pr_err("bad event header size\n"); 2016 goto out_err; 2017 } 2018 2019 if (size > cur_size) { 2020 void *new = realloc(buf, size); 2021 if (!new) { 2022 pr_err("failed to allocate memory to read event\n"); 2023 goto out_err; 2024 } 2025 buf = new; 2026 cur_size = size; 2027 event = buf; 2028 } 2029 p = event; 2030 p += sizeof(struct perf_event_header); 2031 2032 if (size - sizeof(struct perf_event_header)) { 2033 err = perf_data__read(session->data, p, 2034 size - sizeof(struct perf_event_header)); 2035 if (err <= 0) { 2036 if (err == 0) { 2037 pr_err("unexpected end of event stream\n"); 2038 goto done; 2039 } 2040 2041 pr_err("failed to read event data\n"); 2042 goto out_err; 2043 } 2044 } 2045 2046 if ((skip = perf_session__process_event(session, event, head)) < 0) { 2047 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 2048 head, event->header.size, event->header.type); 2049 err = -EINVAL; 2050 goto out_err; 2051 } 2052 2053 head += size; 2054 2055 if (skip > 0) 2056 head += skip; 2057 2058 err = __perf_session__process_decomp_events(session); 2059 if (err) 2060 goto out_err; 2061 2062 if (!session_done()) 2063 goto more; 2064 done: 2065 /* do the final flush for ordered samples */ 2066 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2067 if (err) 2068 goto out_err; 2069 err = auxtrace__flush_events(session, tool); 2070 if (err) 2071 goto out_err; 2072 err = perf_session__flush_thread_stacks(session); 2073 out_err: 2074 free(buf); 2075 if (!tool->no_warn) 2076 perf_session__warn_about_errors(session); 2077 ordered_events__free(&session->ordered_events); 2078 auxtrace__free_events(session); 2079 return err; 2080 } 2081 2082 static union perf_event * 2083 prefetch_event(char *buf, u64 head, size_t mmap_size, 2084 bool needs_swap, union perf_event *error) 2085 { 2086 union perf_event *event; 2087 2088 /* 2089 * Ensure we have enough space remaining to read 2090 * the size of the event in the headers. 2091 */ 2092 if (head + sizeof(event->header) > mmap_size) 2093 return NULL; 2094 2095 event = (union perf_event *)(buf + head); 2096 if (needs_swap) 2097 perf_event_header__bswap(&event->header); 2098 2099 if (head + event->header.size <= mmap_size) 2100 return event; 2101 2102 /* We're not fetching the event so swap back again */ 2103 if (needs_swap) 2104 perf_event_header__bswap(&event->header); 2105 2106 pr_debug("%s: head=%#" PRIx64 " event->header_size=%#x, mmap_size=%#zx:" 2107 " fuzzed or compressed perf.data?\n",__func__, head, event->header.size, mmap_size); 2108 2109 return error; 2110 } 2111 2112 static union perf_event * 2113 fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 2114 { 2115 return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL)); 2116 } 2117 2118 static union perf_event * 2119 fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 2120 { 2121 return prefetch_event(buf, head, mmap_size, needs_swap, NULL); 2122 } 2123 2124 static int __perf_session__process_decomp_events(struct perf_session *session) 2125 { 2126 s64 skip; 2127 u64 size; 2128 struct decomp *decomp = session->active_decomp->decomp_last; 2129 2130 if (!decomp) 2131 return 0; 2132 2133 while (decomp->head < decomp->size && !session_done()) { 2134 union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data, 2135 session->header.needs_swap); 2136 2137 if (!event) 2138 break; 2139 2140 size = event->header.size; 2141 2142 if (size < sizeof(struct perf_event_header) || 2143 (skip = perf_session__process_event(session, event, decomp->file_pos)) < 0) { 2144 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 2145 decomp->file_pos + decomp->head, event->header.size, event->header.type); 2146 return -EINVAL; 2147 } 2148 2149 if (skip) 2150 size += skip; 2151 2152 decomp->head += size; 2153 } 2154 2155 return 0; 2156 } 2157 2158 /* 2159 * On 64bit we can mmap the data file in one go. No need for tiny mmap 2160 * slices. On 32bit we use 32MB. 2161 */ 2162 #if BITS_PER_LONG == 64 2163 #define MMAP_SIZE ULLONG_MAX 2164 #define NUM_MMAPS 1 2165 #else 2166 #define MMAP_SIZE (32 * 1024 * 1024ULL) 2167 #define NUM_MMAPS 128 2168 #endif 2169 2170 struct reader; 2171 2172 typedef s64 (*reader_cb_t)(struct perf_session *session, 2173 union perf_event *event, 2174 u64 file_offset); 2175 2176 struct reader { 2177 int fd; 2178 u64 data_size; 2179 u64 data_offset; 2180 reader_cb_t process; 2181 bool in_place_update; 2182 char *mmaps[NUM_MMAPS]; 2183 size_t mmap_size; 2184 int mmap_idx; 2185 char *mmap_cur; 2186 u64 file_pos; 2187 u64 file_offset; 2188 u64 head; 2189 struct zstd_data zstd_data; 2190 struct decomp_data decomp_data; 2191 }; 2192 2193 static int 2194 reader__init(struct reader *rd, bool *one_mmap) 2195 { 2196 u64 data_size = rd->data_size; 2197 char **mmaps = rd->mmaps; 2198 2199 rd->head = rd->data_offset; 2200 data_size += rd->data_offset; 2201 2202 rd->mmap_size = MMAP_SIZE; 2203 if (rd->mmap_size > data_size) { 2204 rd->mmap_size = data_size; 2205 if (one_mmap) 2206 *one_mmap = true; 2207 } 2208 2209 memset(mmaps, 0, sizeof(rd->mmaps)); 2210 2211 if (zstd_init(&rd->zstd_data, 0)) 2212 return -1; 2213 rd->decomp_data.zstd_decomp = &rd->zstd_data; 2214 2215 return 0; 2216 } 2217 2218 static void 2219 reader__release_decomp(struct reader *rd) 2220 { 2221 perf_decomp__release_events(rd->decomp_data.decomp); 2222 zstd_fini(&rd->zstd_data); 2223 } 2224 2225 static int 2226 reader__mmap(struct reader *rd, struct perf_session *session) 2227 { 2228 int mmap_prot, mmap_flags; 2229 char *buf, **mmaps = rd->mmaps; 2230 u64 page_offset; 2231 2232 mmap_prot = PROT_READ; 2233 mmap_flags = MAP_SHARED; 2234 2235 if (rd->in_place_update) { 2236 mmap_prot |= PROT_WRITE; 2237 } else if (session->header.needs_swap) { 2238 mmap_prot |= PROT_WRITE; 2239 mmap_flags = MAP_PRIVATE; 2240 } 2241 2242 if (mmaps[rd->mmap_idx]) { 2243 munmap(mmaps[rd->mmap_idx], rd->mmap_size); 2244 mmaps[rd->mmap_idx] = NULL; 2245 } 2246 2247 page_offset = page_size * (rd->head / page_size); 2248 rd->file_offset += page_offset; 2249 rd->head -= page_offset; 2250 2251 buf = mmap(NULL, rd->mmap_size, mmap_prot, mmap_flags, rd->fd, 2252 rd->file_offset); 2253 if (buf == MAP_FAILED) { 2254 pr_err("failed to mmap file\n"); 2255 return -errno; 2256 } 2257 mmaps[rd->mmap_idx] = rd->mmap_cur = buf; 2258 rd->mmap_idx = (rd->mmap_idx + 1) & (ARRAY_SIZE(rd->mmaps) - 1); 2259 rd->file_pos = rd->file_offset + rd->head; 2260 if (session->one_mmap) { 2261 session->one_mmap_addr = buf; 2262 session->one_mmap_offset = rd->file_offset; 2263 } 2264 2265 return 0; 2266 } 2267 2268 enum { 2269 READER_OK, 2270 READER_NODATA, 2271 }; 2272 2273 static int 2274 reader__read_event(struct reader *rd, struct perf_session *session, 2275 struct ui_progress *prog) 2276 { 2277 u64 size; 2278 int err = READER_OK; 2279 union perf_event *event; 2280 s64 skip; 2281 2282 event = fetch_mmaped_event(rd->head, rd->mmap_size, rd->mmap_cur, 2283 session->header.needs_swap); 2284 if (IS_ERR(event)) 2285 return PTR_ERR(event); 2286 2287 if (!event) 2288 return READER_NODATA; 2289 2290 size = event->header.size; 2291 2292 skip = -EINVAL; 2293 2294 if (size < sizeof(struct perf_event_header) || 2295 (skip = rd->process(session, event, rd->file_pos)) < 0) { 2296 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n", 2297 rd->file_offset + rd->head, event->header.size, 2298 event->header.type, strerror(-skip)); 2299 err = skip; 2300 goto out; 2301 } 2302 2303 if (skip) 2304 size += skip; 2305 2306 rd->head += size; 2307 rd->file_pos += size; 2308 2309 err = __perf_session__process_decomp_events(session); 2310 if (err) 2311 goto out; 2312 2313 ui_progress__update(prog, size); 2314 2315 out: 2316 return err; 2317 } 2318 2319 static inline bool 2320 reader__eof(struct reader *rd) 2321 { 2322 return (rd->file_pos >= rd->data_size + rd->data_offset); 2323 } 2324 2325 static int 2326 reader__process_events(struct reader *rd, struct perf_session *session, 2327 struct ui_progress *prog) 2328 { 2329 int err; 2330 2331 err = reader__init(rd, &session->one_mmap); 2332 if (err) 2333 goto out; 2334 2335 session->active_decomp = &rd->decomp_data; 2336 2337 remap: 2338 err = reader__mmap(rd, session); 2339 if (err) 2340 goto out; 2341 2342 more: 2343 err = reader__read_event(rd, session, prog); 2344 if (err < 0) 2345 goto out; 2346 else if (err == READER_NODATA) 2347 goto remap; 2348 2349 if (session_done()) 2350 goto out; 2351 2352 if (!reader__eof(rd)) 2353 goto more; 2354 2355 out: 2356 session->active_decomp = &session->decomp_data; 2357 return err; 2358 } 2359 2360 static s64 process_simple(struct perf_session *session, 2361 union perf_event *event, 2362 u64 file_offset) 2363 { 2364 return perf_session__process_event(session, event, file_offset); 2365 } 2366 2367 static int __perf_session__process_events(struct perf_session *session) 2368 { 2369 struct reader rd = { 2370 .fd = perf_data__fd(session->data), 2371 .data_size = session->header.data_size, 2372 .data_offset = session->header.data_offset, 2373 .process = process_simple, 2374 .in_place_update = session->data->in_place_update, 2375 }; 2376 struct ordered_events *oe = &session->ordered_events; 2377 struct perf_tool *tool = session->tool; 2378 struct ui_progress prog; 2379 int err; 2380 2381 perf_tool__fill_defaults(tool); 2382 2383 if (rd.data_size == 0) 2384 return -1; 2385 2386 ui_progress__init_size(&prog, rd.data_size, "Processing events..."); 2387 2388 err = reader__process_events(&rd, session, &prog); 2389 if (err) 2390 goto out_err; 2391 /* do the final flush for ordered samples */ 2392 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2393 if (err) 2394 goto out_err; 2395 err = auxtrace__flush_events(session, tool); 2396 if (err) 2397 goto out_err; 2398 err = perf_session__flush_thread_stacks(session); 2399 out_err: 2400 ui_progress__finish(); 2401 if (!tool->no_warn) 2402 perf_session__warn_about_errors(session); 2403 /* 2404 * We may switching perf.data output, make ordered_events 2405 * reusable. 2406 */ 2407 ordered_events__reinit(&session->ordered_events); 2408 auxtrace__free_events(session); 2409 reader__release_decomp(&rd); 2410 session->one_mmap = false; 2411 return err; 2412 } 2413 2414 int perf_session__process_events(struct perf_session *session) 2415 { 2416 if (perf_session__register_idle_thread(session) < 0) 2417 return -ENOMEM; 2418 2419 if (perf_data__is_pipe(session->data)) 2420 return __perf_session__process_pipe_events(session); 2421 2422 return __perf_session__process_events(session); 2423 } 2424 2425 bool perf_session__has_traces(struct perf_session *session, const char *msg) 2426 { 2427 struct evsel *evsel; 2428 2429 evlist__for_each_entry(session->evlist, evsel) { 2430 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT) 2431 return true; 2432 } 2433 2434 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 2435 return false; 2436 } 2437 2438 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr) 2439 { 2440 char *bracket; 2441 struct ref_reloc_sym *ref; 2442 struct kmap *kmap; 2443 2444 ref = zalloc(sizeof(struct ref_reloc_sym)); 2445 if (ref == NULL) 2446 return -ENOMEM; 2447 2448 ref->name = strdup(symbol_name); 2449 if (ref->name == NULL) { 2450 free(ref); 2451 return -ENOMEM; 2452 } 2453 2454 bracket = strchr(ref->name, ']'); 2455 if (bracket) 2456 *bracket = '\0'; 2457 2458 ref->addr = addr; 2459 2460 kmap = map__kmap(map); 2461 if (kmap) 2462 kmap->ref_reloc_sym = ref; 2463 2464 return 0; 2465 } 2466 2467 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp) 2468 { 2469 return machines__fprintf_dsos(&session->machines, fp); 2470 } 2471 2472 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp, 2473 bool (skip)(struct dso *dso, int parm), int parm) 2474 { 2475 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm); 2476 } 2477 2478 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp, 2479 bool skip_empty) 2480 { 2481 size_t ret; 2482 const char *msg = ""; 2483 2484 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) 2485 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)"; 2486 2487 ret = fprintf(fp, "\nAggregated stats:%s\n", msg); 2488 2489 ret += events_stats__fprintf(&session->evlist->stats, fp, skip_empty); 2490 return ret; 2491 } 2492 2493 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 2494 { 2495 /* 2496 * FIXME: Here we have to actually print all the machines in this 2497 * session, not just the host... 2498 */ 2499 return machine__fprintf(&session->machines.host, fp); 2500 } 2501 2502 struct evsel *perf_session__find_first_evtype(struct perf_session *session, 2503 unsigned int type) 2504 { 2505 struct evsel *pos; 2506 2507 evlist__for_each_entry(session->evlist, pos) { 2508 if (pos->core.attr.type == type) 2509 return pos; 2510 } 2511 return NULL; 2512 } 2513 2514 int perf_session__cpu_bitmap(struct perf_session *session, 2515 const char *cpu_list, unsigned long *cpu_bitmap) 2516 { 2517 int i, err = -1; 2518 struct perf_cpu_map *map; 2519 int nr_cpus = min(session->header.env.nr_cpus_avail, MAX_NR_CPUS); 2520 2521 for (i = 0; i < PERF_TYPE_MAX; ++i) { 2522 struct evsel *evsel; 2523 2524 evsel = perf_session__find_first_evtype(session, i); 2525 if (!evsel) 2526 continue; 2527 2528 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) { 2529 pr_err("File does not contain CPU events. " 2530 "Remove -C option to proceed.\n"); 2531 return -1; 2532 } 2533 } 2534 2535 map = perf_cpu_map__new(cpu_list); 2536 if (map == NULL) { 2537 pr_err("Invalid cpu_list\n"); 2538 return -1; 2539 } 2540 2541 for (i = 0; i < perf_cpu_map__nr(map); i++) { 2542 struct perf_cpu cpu = perf_cpu_map__cpu(map, i); 2543 2544 if (cpu.cpu >= nr_cpus) { 2545 pr_err("Requested CPU %d too large. " 2546 "Consider raising MAX_NR_CPUS\n", cpu.cpu); 2547 goto out_delete_map; 2548 } 2549 2550 set_bit(cpu.cpu, cpu_bitmap); 2551 } 2552 2553 err = 0; 2554 2555 out_delete_map: 2556 perf_cpu_map__put(map); 2557 return err; 2558 } 2559 2560 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 2561 bool full) 2562 { 2563 if (session == NULL || fp == NULL) 2564 return; 2565 2566 fprintf(fp, "# ========\n"); 2567 perf_header__fprintf_info(session, fp, full); 2568 fprintf(fp, "# ========\n#\n"); 2569 } 2570 2571 int perf_event__process_id_index(struct perf_session *session, 2572 union perf_event *event) 2573 { 2574 struct evlist *evlist = session->evlist; 2575 struct perf_record_id_index *ie = &event->id_index; 2576 size_t i, nr, max_nr; 2577 2578 max_nr = (ie->header.size - sizeof(struct perf_record_id_index)) / 2579 sizeof(struct id_index_entry); 2580 nr = ie->nr; 2581 if (nr > max_nr) 2582 return -EINVAL; 2583 2584 if (dump_trace) 2585 fprintf(stdout, " nr: %zu\n", nr); 2586 2587 for (i = 0; i < nr; i++) { 2588 struct id_index_entry *e = &ie->entries[i]; 2589 struct perf_sample_id *sid; 2590 2591 if (dump_trace) { 2592 fprintf(stdout, " ... id: %"PRI_lu64, e->id); 2593 fprintf(stdout, " idx: %"PRI_lu64, e->idx); 2594 fprintf(stdout, " cpu: %"PRI_ld64, e->cpu); 2595 fprintf(stdout, " tid: %"PRI_ld64"\n", e->tid); 2596 } 2597 2598 sid = evlist__id2sid(evlist, e->id); 2599 if (!sid) 2600 return -ENOENT; 2601 sid->idx = e->idx; 2602 sid->cpu.cpu = e->cpu; 2603 sid->tid = e->tid; 2604 } 2605 return 0; 2606 } 2607