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 "util.h" 33 #include "ui/progress.h" 34 #include "../perf.h" 35 #include "arch/common.h" 36 #include <internal/lib.h> 37 #include <linux/err.h> 38 39 #ifdef HAVE_ZSTD_SUPPORT 40 static int perf_session__process_compressed_event(struct perf_session *session, 41 union perf_event *event, u64 file_offset) 42 { 43 void *src; 44 size_t decomp_size, src_size; 45 u64 decomp_last_rem = 0; 46 size_t mmap_len, decomp_len = session->header.env.comp_mmap_len; 47 struct decomp *decomp, *decomp_last = session->decomp_last; 48 49 if (decomp_last) { 50 decomp_last_rem = decomp_last->size - decomp_last->head; 51 decomp_len += decomp_last_rem; 52 } 53 54 mmap_len = sizeof(struct decomp) + decomp_len; 55 decomp = mmap(NULL, mmap_len, PROT_READ|PROT_WRITE, 56 MAP_ANONYMOUS|MAP_PRIVATE, -1, 0); 57 if (decomp == MAP_FAILED) { 58 pr_err("Couldn't allocate memory for decompression\n"); 59 return -1; 60 } 61 62 decomp->file_pos = file_offset; 63 decomp->mmap_len = mmap_len; 64 decomp->head = 0; 65 66 if (decomp_last_rem) { 67 memcpy(decomp->data, &(decomp_last->data[decomp_last->head]), decomp_last_rem); 68 decomp->size = decomp_last_rem; 69 } 70 71 src = (void *)event + sizeof(struct perf_record_compressed); 72 src_size = event->pack.header.size - sizeof(struct perf_record_compressed); 73 74 decomp_size = zstd_decompress_stream(&(session->zstd_data), src, src_size, 75 &(decomp->data[decomp_last_rem]), decomp_len - decomp_last_rem); 76 if (!decomp_size) { 77 munmap(decomp, mmap_len); 78 pr_err("Couldn't decompress data\n"); 79 return -1; 80 } 81 82 decomp->size += decomp_size; 83 84 if (session->decomp == NULL) { 85 session->decomp = decomp; 86 session->decomp_last = decomp; 87 } else { 88 session->decomp_last->next = decomp; 89 session->decomp_last = decomp; 90 } 91 92 pr_debug("decomp (B): %ld to %ld\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) 106 { 107 struct perf_data *data = session->data; 108 109 if (perf_session__read_header(session) < 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 (!perf_evlist__valid_sample_type(session->evlist)) { 121 pr_err("non matching sample_type\n"); 122 return -1; 123 } 124 125 if (!perf_evlist__valid_sample_id_all(session->evlist)) { 126 pr_err("non matching sample_id_all\n"); 127 return -1; 128 } 129 130 if (!perf_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 = perf_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, struct perf_tool *tool) 190 { 191 int ret = -ENOMEM; 192 struct perf_session *session = zalloc(sizeof(*session)); 193 194 if (!session) 195 goto out; 196 197 session->repipe = repipe; 198 session->tool = tool; 199 INIT_LIST_HEAD(&session->auxtrace_index); 200 machines__init(&session->machines); 201 ordered_events__init(&session->ordered_events, 202 ordered_events__deliver_event, NULL); 203 204 perf_env__init(&session->header.env); 205 if (data) { 206 ret = perf_data__open(data); 207 if (ret < 0) 208 goto out_delete; 209 210 session->data = data; 211 212 if (perf_data__is_read(data)) { 213 ret = perf_session__open(session); 214 if (ret < 0) 215 goto out_delete; 216 217 /* 218 * set session attributes that are present in perf.data 219 * but not in pipe-mode. 220 */ 221 if (!data->is_pipe) { 222 perf_session__set_id_hdr_size(session); 223 perf_session__set_comm_exec(session); 224 } 225 226 perf_evlist__init_trace_event_sample_raw(session->evlist); 227 228 /* Open the directory data. */ 229 if (data->is_dir) { 230 ret = perf_data__open_dir(data); 231 if (ret) 232 goto out_delete; 233 } 234 } 235 } else { 236 session->machines.host.env = &perf_env; 237 } 238 239 session->machines.host.single_address_space = 240 perf_env__single_address_space(session->machines.host.env); 241 242 if (!data || perf_data__is_write(data)) { 243 /* 244 * In O_RDONLY mode this will be performed when reading the 245 * kernel MMAP event, in perf_event__process_mmap(). 246 */ 247 if (perf_session__create_kernel_maps(session) < 0) 248 pr_warning("Cannot read kernel map\n"); 249 } 250 251 /* 252 * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is 253 * processed, so perf_evlist__sample_id_all is not meaningful here. 254 */ 255 if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps && 256 tool->ordered_events && !perf_evlist__sample_id_all(session->evlist)) { 257 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 258 tool->ordered_events = false; 259 } 260 261 return session; 262 263 out_delete: 264 perf_session__delete(session); 265 out: 266 return ERR_PTR(ret); 267 } 268 269 static void perf_session__delete_threads(struct perf_session *session) 270 { 271 machine__delete_threads(&session->machines.host); 272 } 273 274 static void perf_session__release_decomp_events(struct perf_session *session) 275 { 276 struct decomp *next, *decomp; 277 size_t mmap_len; 278 next = session->decomp; 279 do { 280 decomp = next; 281 if (decomp == NULL) 282 break; 283 next = decomp->next; 284 mmap_len = decomp->mmap_len; 285 munmap(decomp, mmap_len); 286 } while (1); 287 } 288 289 void perf_session__delete(struct perf_session *session) 290 { 291 if (session == NULL) 292 return; 293 auxtrace__free(session); 294 auxtrace_index__free(&session->auxtrace_index); 295 perf_session__destroy_kernel_maps(session); 296 perf_session__delete_threads(session); 297 perf_session__release_decomp_events(session); 298 perf_env__exit(&session->header.env); 299 machines__exit(&session->machines); 300 if (session->data) 301 perf_data__close(session->data); 302 free(session); 303 } 304 305 static int process_event_synth_tracing_data_stub(struct perf_session *session 306 __maybe_unused, 307 union perf_event *event 308 __maybe_unused) 309 { 310 dump_printf(": unhandled!\n"); 311 return 0; 312 } 313 314 static int process_event_synth_attr_stub(struct perf_tool *tool __maybe_unused, 315 union perf_event *event __maybe_unused, 316 struct evlist **pevlist 317 __maybe_unused) 318 { 319 dump_printf(": unhandled!\n"); 320 return 0; 321 } 322 323 static int process_event_synth_event_update_stub(struct perf_tool *tool __maybe_unused, 324 union perf_event *event __maybe_unused, 325 struct evlist **pevlist 326 __maybe_unused) 327 { 328 if (dump_trace) 329 perf_event__fprintf_event_update(event, stdout); 330 331 dump_printf(": unhandled!\n"); 332 return 0; 333 } 334 335 static int process_event_sample_stub(struct perf_tool *tool __maybe_unused, 336 union perf_event *event __maybe_unused, 337 struct perf_sample *sample __maybe_unused, 338 struct evsel *evsel __maybe_unused, 339 struct machine *machine __maybe_unused) 340 { 341 dump_printf(": unhandled!\n"); 342 return 0; 343 } 344 345 static int process_event_stub(struct perf_tool *tool __maybe_unused, 346 union perf_event *event __maybe_unused, 347 struct perf_sample *sample __maybe_unused, 348 struct machine *machine __maybe_unused) 349 { 350 dump_printf(": unhandled!\n"); 351 return 0; 352 } 353 354 static int process_finished_round_stub(struct perf_tool *tool __maybe_unused, 355 union perf_event *event __maybe_unused, 356 struct ordered_events *oe __maybe_unused) 357 { 358 dump_printf(": unhandled!\n"); 359 return 0; 360 } 361 362 static int process_finished_round(struct perf_tool *tool, 363 union perf_event *event, 364 struct ordered_events *oe); 365 366 static int skipn(int fd, off_t n) 367 { 368 char buf[4096]; 369 ssize_t ret; 370 371 while (n > 0) { 372 ret = read(fd, buf, min(n, (off_t)sizeof(buf))); 373 if (ret <= 0) 374 return ret; 375 n -= ret; 376 } 377 378 return 0; 379 } 380 381 static s64 process_event_auxtrace_stub(struct perf_session *session __maybe_unused, 382 union perf_event *event) 383 { 384 dump_printf(": unhandled!\n"); 385 if (perf_data__is_pipe(session->data)) 386 skipn(perf_data__fd(session->data), event->auxtrace.size); 387 return event->auxtrace.size; 388 } 389 390 static int process_event_op2_stub(struct perf_session *session __maybe_unused, 391 union perf_event *event __maybe_unused) 392 { 393 dump_printf(": unhandled!\n"); 394 return 0; 395 } 396 397 398 static 399 int process_event_thread_map_stub(struct perf_session *session __maybe_unused, 400 union perf_event *event __maybe_unused) 401 { 402 if (dump_trace) 403 perf_event__fprintf_thread_map(event, stdout); 404 405 dump_printf(": unhandled!\n"); 406 return 0; 407 } 408 409 static 410 int process_event_cpu_map_stub(struct perf_session *session __maybe_unused, 411 union perf_event *event __maybe_unused) 412 { 413 if (dump_trace) 414 perf_event__fprintf_cpu_map(event, stdout); 415 416 dump_printf(": unhandled!\n"); 417 return 0; 418 } 419 420 static 421 int process_event_stat_config_stub(struct perf_session *session __maybe_unused, 422 union perf_event *event __maybe_unused) 423 { 424 if (dump_trace) 425 perf_event__fprintf_stat_config(event, stdout); 426 427 dump_printf(": unhandled!\n"); 428 return 0; 429 } 430 431 static int process_stat_stub(struct perf_session *perf_session __maybe_unused, 432 union perf_event *event) 433 { 434 if (dump_trace) 435 perf_event__fprintf_stat(event, stdout); 436 437 dump_printf(": unhandled!\n"); 438 return 0; 439 } 440 441 static int process_stat_round_stub(struct perf_session *perf_session __maybe_unused, 442 union perf_event *event) 443 { 444 if (dump_trace) 445 perf_event__fprintf_stat_round(event, stdout); 446 447 dump_printf(": unhandled!\n"); 448 return 0; 449 } 450 451 static int perf_session__process_compressed_event_stub(struct perf_session *session __maybe_unused, 452 union perf_event *event __maybe_unused, 453 u64 file_offset __maybe_unused) 454 { 455 dump_printf(": unhandled!\n"); 456 return 0; 457 } 458 459 void perf_tool__fill_defaults(struct perf_tool *tool) 460 { 461 if (tool->sample == NULL) 462 tool->sample = process_event_sample_stub; 463 if (tool->mmap == NULL) 464 tool->mmap = process_event_stub; 465 if (tool->mmap2 == NULL) 466 tool->mmap2 = process_event_stub; 467 if (tool->comm == NULL) 468 tool->comm = process_event_stub; 469 if (tool->namespaces == NULL) 470 tool->namespaces = process_event_stub; 471 if (tool->fork == NULL) 472 tool->fork = process_event_stub; 473 if (tool->exit == NULL) 474 tool->exit = process_event_stub; 475 if (tool->lost == NULL) 476 tool->lost = perf_event__process_lost; 477 if (tool->lost_samples == NULL) 478 tool->lost_samples = perf_event__process_lost_samples; 479 if (tool->aux == NULL) 480 tool->aux = perf_event__process_aux; 481 if (tool->itrace_start == NULL) 482 tool->itrace_start = perf_event__process_itrace_start; 483 if (tool->context_switch == NULL) 484 tool->context_switch = perf_event__process_switch; 485 if (tool->ksymbol == NULL) 486 tool->ksymbol = perf_event__process_ksymbol; 487 if (tool->bpf == NULL) 488 tool->bpf = perf_event__process_bpf; 489 if (tool->read == NULL) 490 tool->read = process_event_sample_stub; 491 if (tool->throttle == NULL) 492 tool->throttle = process_event_stub; 493 if (tool->unthrottle == NULL) 494 tool->unthrottle = process_event_stub; 495 if (tool->attr == NULL) 496 tool->attr = process_event_synth_attr_stub; 497 if (tool->event_update == NULL) 498 tool->event_update = process_event_synth_event_update_stub; 499 if (tool->tracing_data == NULL) 500 tool->tracing_data = process_event_synth_tracing_data_stub; 501 if (tool->build_id == NULL) 502 tool->build_id = process_event_op2_stub; 503 if (tool->finished_round == NULL) { 504 if (tool->ordered_events) 505 tool->finished_round = process_finished_round; 506 else 507 tool->finished_round = process_finished_round_stub; 508 } 509 if (tool->id_index == NULL) 510 tool->id_index = process_event_op2_stub; 511 if (tool->auxtrace_info == NULL) 512 tool->auxtrace_info = process_event_op2_stub; 513 if (tool->auxtrace == NULL) 514 tool->auxtrace = process_event_auxtrace_stub; 515 if (tool->auxtrace_error == NULL) 516 tool->auxtrace_error = process_event_op2_stub; 517 if (tool->thread_map == NULL) 518 tool->thread_map = process_event_thread_map_stub; 519 if (tool->cpu_map == NULL) 520 tool->cpu_map = process_event_cpu_map_stub; 521 if (tool->stat_config == NULL) 522 tool->stat_config = process_event_stat_config_stub; 523 if (tool->stat == NULL) 524 tool->stat = process_stat_stub; 525 if (tool->stat_round == NULL) 526 tool->stat_round = process_stat_round_stub; 527 if (tool->time_conv == NULL) 528 tool->time_conv = process_event_op2_stub; 529 if (tool->feature == NULL) 530 tool->feature = process_event_op2_stub; 531 if (tool->compressed == NULL) 532 tool->compressed = perf_session__process_compressed_event; 533 } 534 535 static void swap_sample_id_all(union perf_event *event, void *data) 536 { 537 void *end = (void *) event + event->header.size; 538 int size = end - data; 539 540 BUG_ON(size % sizeof(u64)); 541 mem_bswap_64(data, size); 542 } 543 544 static void perf_event__all64_swap(union perf_event *event, 545 bool sample_id_all __maybe_unused) 546 { 547 struct perf_event_header *hdr = &event->header; 548 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 549 } 550 551 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all) 552 { 553 event->comm.pid = bswap_32(event->comm.pid); 554 event->comm.tid = bswap_32(event->comm.tid); 555 556 if (sample_id_all) { 557 void *data = &event->comm.comm; 558 559 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 560 swap_sample_id_all(event, data); 561 } 562 } 563 564 static void perf_event__mmap_swap(union perf_event *event, 565 bool sample_id_all) 566 { 567 event->mmap.pid = bswap_32(event->mmap.pid); 568 event->mmap.tid = bswap_32(event->mmap.tid); 569 event->mmap.start = bswap_64(event->mmap.start); 570 event->mmap.len = bswap_64(event->mmap.len); 571 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 572 573 if (sample_id_all) { 574 void *data = &event->mmap.filename; 575 576 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 577 swap_sample_id_all(event, data); 578 } 579 } 580 581 static void perf_event__mmap2_swap(union perf_event *event, 582 bool sample_id_all) 583 { 584 event->mmap2.pid = bswap_32(event->mmap2.pid); 585 event->mmap2.tid = bswap_32(event->mmap2.tid); 586 event->mmap2.start = bswap_64(event->mmap2.start); 587 event->mmap2.len = bswap_64(event->mmap2.len); 588 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff); 589 event->mmap2.maj = bswap_32(event->mmap2.maj); 590 event->mmap2.min = bswap_32(event->mmap2.min); 591 event->mmap2.ino = bswap_64(event->mmap2.ino); 592 593 if (sample_id_all) { 594 void *data = &event->mmap2.filename; 595 596 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 597 swap_sample_id_all(event, data); 598 } 599 } 600 static void perf_event__task_swap(union perf_event *event, bool sample_id_all) 601 { 602 event->fork.pid = bswap_32(event->fork.pid); 603 event->fork.tid = bswap_32(event->fork.tid); 604 event->fork.ppid = bswap_32(event->fork.ppid); 605 event->fork.ptid = bswap_32(event->fork.ptid); 606 event->fork.time = bswap_64(event->fork.time); 607 608 if (sample_id_all) 609 swap_sample_id_all(event, &event->fork + 1); 610 } 611 612 static void perf_event__read_swap(union perf_event *event, bool sample_id_all) 613 { 614 event->read.pid = bswap_32(event->read.pid); 615 event->read.tid = bswap_32(event->read.tid); 616 event->read.value = bswap_64(event->read.value); 617 event->read.time_enabled = bswap_64(event->read.time_enabled); 618 event->read.time_running = bswap_64(event->read.time_running); 619 event->read.id = bswap_64(event->read.id); 620 621 if (sample_id_all) 622 swap_sample_id_all(event, &event->read + 1); 623 } 624 625 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all) 626 { 627 event->aux.aux_offset = bswap_64(event->aux.aux_offset); 628 event->aux.aux_size = bswap_64(event->aux.aux_size); 629 event->aux.flags = bswap_64(event->aux.flags); 630 631 if (sample_id_all) 632 swap_sample_id_all(event, &event->aux + 1); 633 } 634 635 static void perf_event__itrace_start_swap(union perf_event *event, 636 bool sample_id_all) 637 { 638 event->itrace_start.pid = bswap_32(event->itrace_start.pid); 639 event->itrace_start.tid = bswap_32(event->itrace_start.tid); 640 641 if (sample_id_all) 642 swap_sample_id_all(event, &event->itrace_start + 1); 643 } 644 645 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all) 646 { 647 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) { 648 event->context_switch.next_prev_pid = 649 bswap_32(event->context_switch.next_prev_pid); 650 event->context_switch.next_prev_tid = 651 bswap_32(event->context_switch.next_prev_tid); 652 } 653 654 if (sample_id_all) 655 swap_sample_id_all(event, &event->context_switch + 1); 656 } 657 658 static void perf_event__throttle_swap(union perf_event *event, 659 bool sample_id_all) 660 { 661 event->throttle.time = bswap_64(event->throttle.time); 662 event->throttle.id = bswap_64(event->throttle.id); 663 event->throttle.stream_id = bswap_64(event->throttle.stream_id); 664 665 if (sample_id_all) 666 swap_sample_id_all(event, &event->throttle + 1); 667 } 668 669 static void perf_event__namespaces_swap(union perf_event *event, 670 bool sample_id_all) 671 { 672 u64 i; 673 674 event->namespaces.pid = bswap_32(event->namespaces.pid); 675 event->namespaces.tid = bswap_32(event->namespaces.tid); 676 event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces); 677 678 for (i = 0; i < event->namespaces.nr_namespaces; i++) { 679 struct perf_ns_link_info *ns = &event->namespaces.link_info[i]; 680 681 ns->dev = bswap_64(ns->dev); 682 ns->ino = bswap_64(ns->ino); 683 } 684 685 if (sample_id_all) 686 swap_sample_id_all(event, &event->namespaces.link_info[i]); 687 } 688 689 static u8 revbyte(u8 b) 690 { 691 int rev = (b >> 4) | ((b & 0xf) << 4); 692 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2); 693 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1); 694 return (u8) rev; 695 } 696 697 /* 698 * XXX this is hack in attempt to carry flags bitfield 699 * through endian village. ABI says: 700 * 701 * Bit-fields are allocated from right to left (least to most significant) 702 * on little-endian implementations and from left to right (most to least 703 * significant) on big-endian implementations. 704 * 705 * The above seems to be byte specific, so we need to reverse each 706 * byte of the bitfield. 'Internet' also says this might be implementation 707 * specific and we probably need proper fix and carry perf_event_attr 708 * bitfield flags in separate data file FEAT_ section. Thought this seems 709 * to work for now. 710 */ 711 static void swap_bitfield(u8 *p, unsigned len) 712 { 713 unsigned i; 714 715 for (i = 0; i < len; i++) { 716 *p = revbyte(*p); 717 p++; 718 } 719 } 720 721 /* exported for swapping attributes in file header */ 722 void perf_event__attr_swap(struct perf_event_attr *attr) 723 { 724 attr->type = bswap_32(attr->type); 725 attr->size = bswap_32(attr->size); 726 727 #define bswap_safe(f, n) \ 728 (attr->size > (offsetof(struct perf_event_attr, f) + \ 729 sizeof(attr->f) * (n))) 730 #define bswap_field(f, sz) \ 731 do { \ 732 if (bswap_safe(f, 0)) \ 733 attr->f = bswap_##sz(attr->f); \ 734 } while(0) 735 #define bswap_field_16(f) bswap_field(f, 16) 736 #define bswap_field_32(f) bswap_field(f, 32) 737 #define bswap_field_64(f) bswap_field(f, 64) 738 739 bswap_field_64(config); 740 bswap_field_64(sample_period); 741 bswap_field_64(sample_type); 742 bswap_field_64(read_format); 743 bswap_field_32(wakeup_events); 744 bswap_field_32(bp_type); 745 bswap_field_64(bp_addr); 746 bswap_field_64(bp_len); 747 bswap_field_64(branch_sample_type); 748 bswap_field_64(sample_regs_user); 749 bswap_field_32(sample_stack_user); 750 bswap_field_32(aux_watermark); 751 bswap_field_16(sample_max_stack); 752 753 /* 754 * After read_format are bitfields. Check read_format because 755 * we are unable to use offsetof on bitfield. 756 */ 757 if (bswap_safe(read_format, 1)) 758 swap_bitfield((u8 *) (&attr->read_format + 1), 759 sizeof(u64)); 760 #undef bswap_field_64 761 #undef bswap_field_32 762 #undef bswap_field 763 #undef bswap_safe 764 } 765 766 static void perf_event__hdr_attr_swap(union perf_event *event, 767 bool sample_id_all __maybe_unused) 768 { 769 size_t size; 770 771 perf_event__attr_swap(&event->attr.attr); 772 773 size = event->header.size; 774 size -= (void *)&event->attr.id - (void *)event; 775 mem_bswap_64(event->attr.id, size); 776 } 777 778 static void perf_event__event_update_swap(union perf_event *event, 779 bool sample_id_all __maybe_unused) 780 { 781 event->event_update.type = bswap_64(event->event_update.type); 782 event->event_update.id = bswap_64(event->event_update.id); 783 } 784 785 static void perf_event__event_type_swap(union perf_event *event, 786 bool sample_id_all __maybe_unused) 787 { 788 event->event_type.event_type.event_id = 789 bswap_64(event->event_type.event_type.event_id); 790 } 791 792 static void perf_event__tracing_data_swap(union perf_event *event, 793 bool sample_id_all __maybe_unused) 794 { 795 event->tracing_data.size = bswap_32(event->tracing_data.size); 796 } 797 798 static void perf_event__auxtrace_info_swap(union perf_event *event, 799 bool sample_id_all __maybe_unused) 800 { 801 size_t size; 802 803 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type); 804 805 size = event->header.size; 806 size -= (void *)&event->auxtrace_info.priv - (void *)event; 807 mem_bswap_64(event->auxtrace_info.priv, size); 808 } 809 810 static void perf_event__auxtrace_swap(union perf_event *event, 811 bool sample_id_all __maybe_unused) 812 { 813 event->auxtrace.size = bswap_64(event->auxtrace.size); 814 event->auxtrace.offset = bswap_64(event->auxtrace.offset); 815 event->auxtrace.reference = bswap_64(event->auxtrace.reference); 816 event->auxtrace.idx = bswap_32(event->auxtrace.idx); 817 event->auxtrace.tid = bswap_32(event->auxtrace.tid); 818 event->auxtrace.cpu = bswap_32(event->auxtrace.cpu); 819 } 820 821 static void perf_event__auxtrace_error_swap(union perf_event *event, 822 bool sample_id_all __maybe_unused) 823 { 824 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type); 825 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code); 826 event->auxtrace_error.cpu = bswap_32(event->auxtrace_error.cpu); 827 event->auxtrace_error.pid = bswap_32(event->auxtrace_error.pid); 828 event->auxtrace_error.tid = bswap_32(event->auxtrace_error.tid); 829 event->auxtrace_error.fmt = bswap_32(event->auxtrace_error.fmt); 830 event->auxtrace_error.ip = bswap_64(event->auxtrace_error.ip); 831 if (event->auxtrace_error.fmt) 832 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time); 833 } 834 835 static void perf_event__thread_map_swap(union perf_event *event, 836 bool sample_id_all __maybe_unused) 837 { 838 unsigned i; 839 840 event->thread_map.nr = bswap_64(event->thread_map.nr); 841 842 for (i = 0; i < event->thread_map.nr; i++) 843 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid); 844 } 845 846 static void perf_event__cpu_map_swap(union perf_event *event, 847 bool sample_id_all __maybe_unused) 848 { 849 struct perf_record_cpu_map_data *data = &event->cpu_map.data; 850 struct cpu_map_entries *cpus; 851 struct perf_record_record_cpu_map *mask; 852 unsigned i; 853 854 data->type = bswap_64(data->type); 855 856 switch (data->type) { 857 case PERF_CPU_MAP__CPUS: 858 cpus = (struct cpu_map_entries *)data->data; 859 860 cpus->nr = bswap_16(cpus->nr); 861 862 for (i = 0; i < cpus->nr; i++) 863 cpus->cpu[i] = bswap_16(cpus->cpu[i]); 864 break; 865 case PERF_CPU_MAP__MASK: 866 mask = (struct perf_record_record_cpu_map *)data->data; 867 868 mask->nr = bswap_16(mask->nr); 869 mask->long_size = bswap_16(mask->long_size); 870 871 switch (mask->long_size) { 872 case 4: mem_bswap_32(&mask->mask, mask->nr); break; 873 case 8: mem_bswap_64(&mask->mask, mask->nr); break; 874 default: 875 pr_err("cpu_map swap: unsupported long size\n"); 876 } 877 default: 878 break; 879 } 880 } 881 882 static void perf_event__stat_config_swap(union perf_event *event, 883 bool sample_id_all __maybe_unused) 884 { 885 u64 size; 886 887 size = event->stat_config.nr * sizeof(event->stat_config.data[0]); 888 size += 1; /* nr item itself */ 889 mem_bswap_64(&event->stat_config.nr, size); 890 } 891 892 static void perf_event__stat_swap(union perf_event *event, 893 bool sample_id_all __maybe_unused) 894 { 895 event->stat.id = bswap_64(event->stat.id); 896 event->stat.thread = bswap_32(event->stat.thread); 897 event->stat.cpu = bswap_32(event->stat.cpu); 898 event->stat.val = bswap_64(event->stat.val); 899 event->stat.ena = bswap_64(event->stat.ena); 900 event->stat.run = bswap_64(event->stat.run); 901 } 902 903 static void perf_event__stat_round_swap(union perf_event *event, 904 bool sample_id_all __maybe_unused) 905 { 906 event->stat_round.type = bswap_64(event->stat_round.type); 907 event->stat_round.time = bswap_64(event->stat_round.time); 908 } 909 910 typedef void (*perf_event__swap_op)(union perf_event *event, 911 bool sample_id_all); 912 913 static perf_event__swap_op perf_event__swap_ops[] = { 914 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 915 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap, 916 [PERF_RECORD_COMM] = perf_event__comm_swap, 917 [PERF_RECORD_FORK] = perf_event__task_swap, 918 [PERF_RECORD_EXIT] = perf_event__task_swap, 919 [PERF_RECORD_LOST] = perf_event__all64_swap, 920 [PERF_RECORD_READ] = perf_event__read_swap, 921 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap, 922 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap, 923 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 924 [PERF_RECORD_AUX] = perf_event__aux_swap, 925 [PERF_RECORD_ITRACE_START] = perf_event__itrace_start_swap, 926 [PERF_RECORD_LOST_SAMPLES] = perf_event__all64_swap, 927 [PERF_RECORD_SWITCH] = perf_event__switch_swap, 928 [PERF_RECORD_SWITCH_CPU_WIDE] = perf_event__switch_swap, 929 [PERF_RECORD_NAMESPACES] = perf_event__namespaces_swap, 930 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 931 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 932 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 933 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 934 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap, 935 [PERF_RECORD_AUXTRACE_INFO] = perf_event__auxtrace_info_swap, 936 [PERF_RECORD_AUXTRACE] = perf_event__auxtrace_swap, 937 [PERF_RECORD_AUXTRACE_ERROR] = perf_event__auxtrace_error_swap, 938 [PERF_RECORD_THREAD_MAP] = perf_event__thread_map_swap, 939 [PERF_RECORD_CPU_MAP] = perf_event__cpu_map_swap, 940 [PERF_RECORD_STAT_CONFIG] = perf_event__stat_config_swap, 941 [PERF_RECORD_STAT] = perf_event__stat_swap, 942 [PERF_RECORD_STAT_ROUND] = perf_event__stat_round_swap, 943 [PERF_RECORD_EVENT_UPDATE] = perf_event__event_update_swap, 944 [PERF_RECORD_TIME_CONV] = perf_event__all64_swap, 945 [PERF_RECORD_HEADER_MAX] = NULL, 946 }; 947 948 /* 949 * When perf record finishes a pass on every buffers, it records this pseudo 950 * event. 951 * We record the max timestamp t found in the pass n. 952 * Assuming these timestamps are monotonic across cpus, we know that if 953 * a buffer still has events with timestamps below t, they will be all 954 * available and then read in the pass n + 1. 955 * Hence when we start to read the pass n + 2, we can safely flush every 956 * events with timestamps below t. 957 * 958 * ============ PASS n ================= 959 * CPU 0 | CPU 1 960 * | 961 * cnt1 timestamps | cnt2 timestamps 962 * 1 | 2 963 * 2 | 3 964 * - | 4 <--- max recorded 965 * 966 * ============ PASS n + 1 ============== 967 * CPU 0 | CPU 1 968 * | 969 * cnt1 timestamps | cnt2 timestamps 970 * 3 | 5 971 * 4 | 6 972 * 5 | 7 <---- max recorded 973 * 974 * Flush every events below timestamp 4 975 * 976 * ============ PASS n + 2 ============== 977 * CPU 0 | CPU 1 978 * | 979 * cnt1 timestamps | cnt2 timestamps 980 * 6 | 8 981 * 7 | 9 982 * - | 10 983 * 984 * Flush every events below timestamp 7 985 * etc... 986 */ 987 static int process_finished_round(struct perf_tool *tool __maybe_unused, 988 union perf_event *event __maybe_unused, 989 struct ordered_events *oe) 990 { 991 if (dump_trace) 992 fprintf(stdout, "\n"); 993 return ordered_events__flush(oe, OE_FLUSH__ROUND); 994 } 995 996 int perf_session__queue_event(struct perf_session *s, union perf_event *event, 997 u64 timestamp, u64 file_offset) 998 { 999 return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset); 1000 } 1001 1002 static void callchain__lbr_callstack_printf(struct perf_sample *sample) 1003 { 1004 struct ip_callchain *callchain = sample->callchain; 1005 struct branch_stack *lbr_stack = sample->branch_stack; 1006 u64 kernel_callchain_nr = callchain->nr; 1007 unsigned int i; 1008 1009 for (i = 0; i < kernel_callchain_nr; i++) { 1010 if (callchain->ips[i] == PERF_CONTEXT_USER) 1011 break; 1012 } 1013 1014 if ((i != kernel_callchain_nr) && lbr_stack->nr) { 1015 u64 total_nr; 1016 /* 1017 * LBR callstack can only get user call chain, 1018 * i is kernel call chain number, 1019 * 1 is PERF_CONTEXT_USER. 1020 * 1021 * The user call chain is stored in LBR registers. 1022 * LBR are pair registers. The caller is stored 1023 * in "from" register, while the callee is stored 1024 * in "to" register. 1025 * For example, there is a call stack 1026 * "A"->"B"->"C"->"D". 1027 * The LBR registers will recorde like 1028 * "C"->"D", "B"->"C", "A"->"B". 1029 * So only the first "to" register and all "from" 1030 * registers are needed to construct the whole stack. 1031 */ 1032 total_nr = i + 1 + lbr_stack->nr + 1; 1033 kernel_callchain_nr = i + 1; 1034 1035 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr); 1036 1037 for (i = 0; i < kernel_callchain_nr; i++) 1038 printf("..... %2d: %016" PRIx64 "\n", 1039 i, callchain->ips[i]); 1040 1041 printf("..... %2d: %016" PRIx64 "\n", 1042 (int)(kernel_callchain_nr), lbr_stack->entries[0].to); 1043 for (i = 0; i < lbr_stack->nr; i++) 1044 printf("..... %2d: %016" PRIx64 "\n", 1045 (int)(i + kernel_callchain_nr + 1), lbr_stack->entries[i].from); 1046 } 1047 } 1048 1049 static void callchain__printf(struct evsel *evsel, 1050 struct perf_sample *sample) 1051 { 1052 unsigned int i; 1053 struct ip_callchain *callchain = sample->callchain; 1054 1055 if (perf_evsel__has_branch_callstack(evsel)) 1056 callchain__lbr_callstack_printf(sample); 1057 1058 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr); 1059 1060 for (i = 0; i < callchain->nr; i++) 1061 printf("..... %2d: %016" PRIx64 "\n", 1062 i, callchain->ips[i]); 1063 } 1064 1065 static void branch_stack__printf(struct perf_sample *sample, bool callstack) 1066 { 1067 uint64_t i; 1068 1069 printf("%s: nr:%" PRIu64 "\n", 1070 !callstack ? "... branch stack" : "... branch callstack", 1071 sample->branch_stack->nr); 1072 1073 for (i = 0; i < sample->branch_stack->nr; i++) { 1074 struct branch_entry *e = &sample->branch_stack->entries[i]; 1075 1076 if (!callstack) { 1077 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x\n", 1078 i, e->from, e->to, 1079 (unsigned short)e->flags.cycles, 1080 e->flags.mispred ? "M" : " ", 1081 e->flags.predicted ? "P" : " ", 1082 e->flags.abort ? "A" : " ", 1083 e->flags.in_tx ? "T" : " ", 1084 (unsigned)e->flags.reserved); 1085 } else { 1086 printf("..... %2"PRIu64": %016" PRIx64 "\n", 1087 i, i > 0 ? e->from : e->to); 1088 } 1089 } 1090 } 1091 1092 static void regs_dump__printf(u64 mask, u64 *regs) 1093 { 1094 unsigned rid, i = 0; 1095 1096 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) { 1097 u64 val = regs[i++]; 1098 1099 printf(".... %-5s 0x%" PRIx64 "\n", 1100 perf_reg_name(rid), val); 1101 } 1102 } 1103 1104 static const char *regs_abi[] = { 1105 [PERF_SAMPLE_REGS_ABI_NONE] = "none", 1106 [PERF_SAMPLE_REGS_ABI_32] = "32-bit", 1107 [PERF_SAMPLE_REGS_ABI_64] = "64-bit", 1108 }; 1109 1110 static inline const char *regs_dump_abi(struct regs_dump *d) 1111 { 1112 if (d->abi > PERF_SAMPLE_REGS_ABI_64) 1113 return "unknown"; 1114 1115 return regs_abi[d->abi]; 1116 } 1117 1118 static void regs__printf(const char *type, struct regs_dump *regs) 1119 { 1120 u64 mask = regs->mask; 1121 1122 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n", 1123 type, 1124 mask, 1125 regs_dump_abi(regs)); 1126 1127 regs_dump__printf(mask, regs->regs); 1128 } 1129 1130 static void regs_user__printf(struct perf_sample *sample) 1131 { 1132 struct regs_dump *user_regs = &sample->user_regs; 1133 1134 if (user_regs->regs) 1135 regs__printf("user", user_regs); 1136 } 1137 1138 static void regs_intr__printf(struct perf_sample *sample) 1139 { 1140 struct regs_dump *intr_regs = &sample->intr_regs; 1141 1142 if (intr_regs->regs) 1143 regs__printf("intr", intr_regs); 1144 } 1145 1146 static void stack_user__printf(struct stack_dump *dump) 1147 { 1148 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 1149 dump->size, dump->offset); 1150 } 1151 1152 static void perf_evlist__print_tstamp(struct evlist *evlist, 1153 union perf_event *event, 1154 struct perf_sample *sample) 1155 { 1156 u64 sample_type = __perf_evlist__combined_sample_type(evlist); 1157 1158 if (event->header.type != PERF_RECORD_SAMPLE && 1159 !perf_evlist__sample_id_all(evlist)) { 1160 fputs("-1 -1 ", stdout); 1161 return; 1162 } 1163 1164 if ((sample_type & PERF_SAMPLE_CPU)) 1165 printf("%u ", sample->cpu); 1166 1167 if (sample_type & PERF_SAMPLE_TIME) 1168 printf("%" PRIu64 " ", sample->time); 1169 } 1170 1171 static void sample_read__printf(struct perf_sample *sample, u64 read_format) 1172 { 1173 printf("... sample_read:\n"); 1174 1175 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1176 printf("...... time enabled %016" PRIx64 "\n", 1177 sample->read.time_enabled); 1178 1179 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1180 printf("...... time running %016" PRIx64 "\n", 1181 sample->read.time_running); 1182 1183 if (read_format & PERF_FORMAT_GROUP) { 1184 u64 i; 1185 1186 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr); 1187 1188 for (i = 0; i < sample->read.group.nr; i++) { 1189 struct sample_read_value *value; 1190 1191 value = &sample->read.group.values[i]; 1192 printf("..... id %016" PRIx64 1193 ", value %016" PRIx64 "\n", 1194 value->id, value->value); 1195 } 1196 } else 1197 printf("..... id %016" PRIx64 ", value %016" PRIx64 "\n", 1198 sample->read.one.id, sample->read.one.value); 1199 } 1200 1201 static void dump_event(struct evlist *evlist, union perf_event *event, 1202 u64 file_offset, struct perf_sample *sample) 1203 { 1204 if (!dump_trace) 1205 return; 1206 1207 printf("\n%#" PRIx64 " [%#x]: event: %d\n", 1208 file_offset, event->header.size, event->header.type); 1209 1210 trace_event(event); 1211 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw) 1212 evlist->trace_event_sample_raw(evlist, event, sample); 1213 1214 if (sample) 1215 perf_evlist__print_tstamp(evlist, event, sample); 1216 1217 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 1218 event->header.size, perf_event__name(event->header.type)); 1219 } 1220 1221 static void dump_sample(struct evsel *evsel, union perf_event *event, 1222 struct perf_sample *sample) 1223 { 1224 u64 sample_type; 1225 1226 if (!dump_trace) 1227 return; 1228 1229 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 1230 event->header.misc, sample->pid, sample->tid, sample->ip, 1231 sample->period, sample->addr); 1232 1233 sample_type = evsel->core.attr.sample_type; 1234 1235 if (evsel__has_callchain(evsel)) 1236 callchain__printf(evsel, sample); 1237 1238 if (sample_type & PERF_SAMPLE_BRANCH_STACK) 1239 branch_stack__printf(sample, perf_evsel__has_branch_callstack(evsel)); 1240 1241 if (sample_type & PERF_SAMPLE_REGS_USER) 1242 regs_user__printf(sample); 1243 1244 if (sample_type & PERF_SAMPLE_REGS_INTR) 1245 regs_intr__printf(sample); 1246 1247 if (sample_type & PERF_SAMPLE_STACK_USER) 1248 stack_user__printf(&sample->user_stack); 1249 1250 if (sample_type & PERF_SAMPLE_WEIGHT) 1251 printf("... weight: %" PRIu64 "\n", sample->weight); 1252 1253 if (sample_type & PERF_SAMPLE_DATA_SRC) 1254 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src); 1255 1256 if (sample_type & PERF_SAMPLE_PHYS_ADDR) 1257 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr); 1258 1259 if (sample_type & PERF_SAMPLE_TRANSACTION) 1260 printf("... transaction: %" PRIx64 "\n", sample->transaction); 1261 1262 if (sample_type & PERF_SAMPLE_READ) 1263 sample_read__printf(sample, evsel->core.attr.read_format); 1264 } 1265 1266 static void dump_read(struct evsel *evsel, union perf_event *event) 1267 { 1268 struct perf_record_read *read_event = &event->read; 1269 u64 read_format; 1270 1271 if (!dump_trace) 1272 return; 1273 1274 printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid, 1275 perf_evsel__name(evsel), 1276 event->read.value); 1277 1278 if (!evsel) 1279 return; 1280 1281 read_format = evsel->core.attr.read_format; 1282 1283 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1284 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled); 1285 1286 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1287 printf("... time running : %" PRI_lu64 "\n", read_event->time_running); 1288 1289 if (read_format & PERF_FORMAT_ID) 1290 printf("... id : %" PRI_lu64 "\n", read_event->id); 1291 } 1292 1293 static struct machine *machines__find_for_cpumode(struct machines *machines, 1294 union perf_event *event, 1295 struct perf_sample *sample) 1296 { 1297 struct machine *machine; 1298 1299 if (perf_guest && 1300 ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 1301 (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) { 1302 u32 pid; 1303 1304 if (event->header.type == PERF_RECORD_MMAP 1305 || event->header.type == PERF_RECORD_MMAP2) 1306 pid = event->mmap.pid; 1307 else 1308 pid = sample->pid; 1309 1310 machine = machines__find(machines, pid); 1311 if (!machine) 1312 machine = machines__findnew(machines, DEFAULT_GUEST_KERNEL_ID); 1313 return machine; 1314 } 1315 1316 return &machines->host; 1317 } 1318 1319 static int deliver_sample_value(struct evlist *evlist, 1320 struct perf_tool *tool, 1321 union perf_event *event, 1322 struct perf_sample *sample, 1323 struct sample_read_value *v, 1324 struct machine *machine) 1325 { 1326 struct perf_sample_id *sid = perf_evlist__id2sid(evlist, v->id); 1327 1328 if (sid) { 1329 sample->id = v->id; 1330 sample->period = v->value - sid->period; 1331 sid->period = v->value; 1332 } 1333 1334 if (!sid || sid->evsel == NULL) { 1335 ++evlist->stats.nr_unknown_id; 1336 return 0; 1337 } 1338 1339 /* 1340 * There's no reason to deliver sample 1341 * for zero period, bail out. 1342 */ 1343 if (!sample->period) 1344 return 0; 1345 1346 return tool->sample(tool, event, sample, sid->evsel, machine); 1347 } 1348 1349 static int deliver_sample_group(struct evlist *evlist, 1350 struct perf_tool *tool, 1351 union perf_event *event, 1352 struct perf_sample *sample, 1353 struct machine *machine) 1354 { 1355 int ret = -EINVAL; 1356 u64 i; 1357 1358 for (i = 0; i < sample->read.group.nr; i++) { 1359 ret = deliver_sample_value(evlist, tool, event, sample, 1360 &sample->read.group.values[i], 1361 machine); 1362 if (ret) 1363 break; 1364 } 1365 1366 return ret; 1367 } 1368 1369 static int 1370 perf_evlist__deliver_sample(struct evlist *evlist, 1371 struct perf_tool *tool, 1372 union perf_event *event, 1373 struct perf_sample *sample, 1374 struct evsel *evsel, 1375 struct machine *machine) 1376 { 1377 /* We know evsel != NULL. */ 1378 u64 sample_type = evsel->core.attr.sample_type; 1379 u64 read_format = evsel->core.attr.read_format; 1380 1381 /* Standard sample delivery. */ 1382 if (!(sample_type & PERF_SAMPLE_READ)) 1383 return tool->sample(tool, event, sample, evsel, machine); 1384 1385 /* For PERF_SAMPLE_READ we have either single or group mode. */ 1386 if (read_format & PERF_FORMAT_GROUP) 1387 return deliver_sample_group(evlist, tool, event, sample, 1388 machine); 1389 else 1390 return deliver_sample_value(evlist, tool, event, sample, 1391 &sample->read.one, machine); 1392 } 1393 1394 static int machines__deliver_event(struct machines *machines, 1395 struct evlist *evlist, 1396 union perf_event *event, 1397 struct perf_sample *sample, 1398 struct perf_tool *tool, u64 file_offset) 1399 { 1400 struct evsel *evsel; 1401 struct machine *machine; 1402 1403 dump_event(evlist, event, file_offset, sample); 1404 1405 evsel = perf_evlist__id2evsel(evlist, sample->id); 1406 1407 machine = machines__find_for_cpumode(machines, event, sample); 1408 1409 switch (event->header.type) { 1410 case PERF_RECORD_SAMPLE: 1411 if (evsel == NULL) { 1412 ++evlist->stats.nr_unknown_id; 1413 return 0; 1414 } 1415 dump_sample(evsel, event, sample); 1416 if (machine == NULL) { 1417 ++evlist->stats.nr_unprocessable_samples; 1418 return 0; 1419 } 1420 return perf_evlist__deliver_sample(evlist, tool, event, sample, evsel, machine); 1421 case PERF_RECORD_MMAP: 1422 return tool->mmap(tool, event, sample, machine); 1423 case PERF_RECORD_MMAP2: 1424 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT) 1425 ++evlist->stats.nr_proc_map_timeout; 1426 return tool->mmap2(tool, event, sample, machine); 1427 case PERF_RECORD_COMM: 1428 return tool->comm(tool, event, sample, machine); 1429 case PERF_RECORD_NAMESPACES: 1430 return tool->namespaces(tool, event, sample, machine); 1431 case PERF_RECORD_FORK: 1432 return tool->fork(tool, event, sample, machine); 1433 case PERF_RECORD_EXIT: 1434 return tool->exit(tool, event, sample, machine); 1435 case PERF_RECORD_LOST: 1436 if (tool->lost == perf_event__process_lost) 1437 evlist->stats.total_lost += event->lost.lost; 1438 return tool->lost(tool, event, sample, machine); 1439 case PERF_RECORD_LOST_SAMPLES: 1440 if (tool->lost_samples == perf_event__process_lost_samples) 1441 evlist->stats.total_lost_samples += event->lost_samples.lost; 1442 return tool->lost_samples(tool, event, sample, machine); 1443 case PERF_RECORD_READ: 1444 dump_read(evsel, event); 1445 return tool->read(tool, event, sample, evsel, machine); 1446 case PERF_RECORD_THROTTLE: 1447 return tool->throttle(tool, event, sample, machine); 1448 case PERF_RECORD_UNTHROTTLE: 1449 return tool->unthrottle(tool, event, sample, machine); 1450 case PERF_RECORD_AUX: 1451 if (tool->aux == perf_event__process_aux) { 1452 if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) 1453 evlist->stats.total_aux_lost += 1; 1454 if (event->aux.flags & PERF_AUX_FLAG_PARTIAL) 1455 evlist->stats.total_aux_partial += 1; 1456 } 1457 return tool->aux(tool, event, sample, machine); 1458 case PERF_RECORD_ITRACE_START: 1459 return tool->itrace_start(tool, event, sample, machine); 1460 case PERF_RECORD_SWITCH: 1461 case PERF_RECORD_SWITCH_CPU_WIDE: 1462 return tool->context_switch(tool, event, sample, machine); 1463 case PERF_RECORD_KSYMBOL: 1464 return tool->ksymbol(tool, event, sample, machine); 1465 case PERF_RECORD_BPF_EVENT: 1466 return tool->bpf(tool, event, sample, machine); 1467 default: 1468 ++evlist->stats.nr_unknown_events; 1469 return -1; 1470 } 1471 } 1472 1473 static int perf_session__deliver_event(struct perf_session *session, 1474 union perf_event *event, 1475 struct perf_tool *tool, 1476 u64 file_offset) 1477 { 1478 struct perf_sample sample; 1479 int ret; 1480 1481 ret = perf_evlist__parse_sample(session->evlist, event, &sample); 1482 if (ret) { 1483 pr_err("Can't parse sample, err = %d\n", ret); 1484 return ret; 1485 } 1486 1487 ret = auxtrace__process_event(session, event, &sample, tool); 1488 if (ret < 0) 1489 return ret; 1490 if (ret > 0) 1491 return 0; 1492 1493 return machines__deliver_event(&session->machines, session->evlist, 1494 event, &sample, tool, file_offset); 1495 } 1496 1497 static s64 perf_session__process_user_event(struct perf_session *session, 1498 union perf_event *event, 1499 u64 file_offset) 1500 { 1501 struct ordered_events *oe = &session->ordered_events; 1502 struct perf_tool *tool = session->tool; 1503 struct perf_sample sample = { .time = 0, }; 1504 int fd = perf_data__fd(session->data); 1505 int err; 1506 1507 if (event->header.type != PERF_RECORD_COMPRESSED || 1508 tool->compressed == perf_session__process_compressed_event_stub) 1509 dump_event(session->evlist, event, file_offset, &sample); 1510 1511 /* These events are processed right away */ 1512 switch (event->header.type) { 1513 case PERF_RECORD_HEADER_ATTR: 1514 err = tool->attr(tool, event, &session->evlist); 1515 if (err == 0) { 1516 perf_session__set_id_hdr_size(session); 1517 perf_session__set_comm_exec(session); 1518 } 1519 return err; 1520 case PERF_RECORD_EVENT_UPDATE: 1521 return tool->event_update(tool, event, &session->evlist); 1522 case PERF_RECORD_HEADER_EVENT_TYPE: 1523 /* 1524 * Depreceated, but we need to handle it for sake 1525 * of old data files create in pipe mode. 1526 */ 1527 return 0; 1528 case PERF_RECORD_HEADER_TRACING_DATA: 1529 /* setup for reading amidst mmap */ 1530 lseek(fd, file_offset, SEEK_SET); 1531 return tool->tracing_data(session, event); 1532 case PERF_RECORD_HEADER_BUILD_ID: 1533 return tool->build_id(session, event); 1534 case PERF_RECORD_FINISHED_ROUND: 1535 return tool->finished_round(tool, event, oe); 1536 case PERF_RECORD_ID_INDEX: 1537 return tool->id_index(session, event); 1538 case PERF_RECORD_AUXTRACE_INFO: 1539 return tool->auxtrace_info(session, event); 1540 case PERF_RECORD_AUXTRACE: 1541 /* setup for reading amidst mmap */ 1542 lseek(fd, file_offset + event->header.size, SEEK_SET); 1543 return tool->auxtrace(session, event); 1544 case PERF_RECORD_AUXTRACE_ERROR: 1545 perf_session__auxtrace_error_inc(session, event); 1546 return tool->auxtrace_error(session, event); 1547 case PERF_RECORD_THREAD_MAP: 1548 return tool->thread_map(session, event); 1549 case PERF_RECORD_CPU_MAP: 1550 return tool->cpu_map(session, event); 1551 case PERF_RECORD_STAT_CONFIG: 1552 return tool->stat_config(session, event); 1553 case PERF_RECORD_STAT: 1554 return tool->stat(session, event); 1555 case PERF_RECORD_STAT_ROUND: 1556 return tool->stat_round(session, event); 1557 case PERF_RECORD_TIME_CONV: 1558 session->time_conv = event->time_conv; 1559 return tool->time_conv(session, event); 1560 case PERF_RECORD_HEADER_FEATURE: 1561 return tool->feature(session, event); 1562 case PERF_RECORD_COMPRESSED: 1563 err = tool->compressed(session, event, file_offset); 1564 if (err) 1565 dump_event(session->evlist, event, file_offset, &sample); 1566 return err; 1567 default: 1568 return -EINVAL; 1569 } 1570 } 1571 1572 int perf_session__deliver_synth_event(struct perf_session *session, 1573 union perf_event *event, 1574 struct perf_sample *sample) 1575 { 1576 struct evlist *evlist = session->evlist; 1577 struct perf_tool *tool = session->tool; 1578 1579 events_stats__inc(&evlist->stats, event->header.type); 1580 1581 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1582 return perf_session__process_user_event(session, event, 0); 1583 1584 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0); 1585 } 1586 1587 static void event_swap(union perf_event *event, bool sample_id_all) 1588 { 1589 perf_event__swap_op swap; 1590 1591 swap = perf_event__swap_ops[event->header.type]; 1592 if (swap) 1593 swap(event, sample_id_all); 1594 } 1595 1596 int perf_session__peek_event(struct perf_session *session, off_t file_offset, 1597 void *buf, size_t buf_sz, 1598 union perf_event **event_ptr, 1599 struct perf_sample *sample) 1600 { 1601 union perf_event *event; 1602 size_t hdr_sz, rest; 1603 int fd; 1604 1605 if (session->one_mmap && !session->header.needs_swap) { 1606 event = file_offset - session->one_mmap_offset + 1607 session->one_mmap_addr; 1608 goto out_parse_sample; 1609 } 1610 1611 if (perf_data__is_pipe(session->data)) 1612 return -1; 1613 1614 fd = perf_data__fd(session->data); 1615 hdr_sz = sizeof(struct perf_event_header); 1616 1617 if (buf_sz < hdr_sz) 1618 return -1; 1619 1620 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 || 1621 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz) 1622 return -1; 1623 1624 event = (union perf_event *)buf; 1625 1626 if (session->header.needs_swap) 1627 perf_event_header__bswap(&event->header); 1628 1629 if (event->header.size < hdr_sz || event->header.size > buf_sz) 1630 return -1; 1631 1632 rest = event->header.size - hdr_sz; 1633 1634 if (readn(fd, buf, rest) != (ssize_t)rest) 1635 return -1; 1636 1637 if (session->header.needs_swap) 1638 event_swap(event, perf_evlist__sample_id_all(session->evlist)); 1639 1640 out_parse_sample: 1641 1642 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START && 1643 perf_evlist__parse_sample(session->evlist, event, sample)) 1644 return -1; 1645 1646 *event_ptr = event; 1647 1648 return 0; 1649 } 1650 1651 static s64 perf_session__process_event(struct perf_session *session, 1652 union perf_event *event, u64 file_offset) 1653 { 1654 struct evlist *evlist = session->evlist; 1655 struct perf_tool *tool = session->tool; 1656 int ret; 1657 1658 if (session->header.needs_swap) 1659 event_swap(event, perf_evlist__sample_id_all(evlist)); 1660 1661 if (event->header.type >= PERF_RECORD_HEADER_MAX) 1662 return -EINVAL; 1663 1664 events_stats__inc(&evlist->stats, event->header.type); 1665 1666 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1667 return perf_session__process_user_event(session, event, file_offset); 1668 1669 if (tool->ordered_events) { 1670 u64 timestamp = -1ULL; 1671 1672 ret = perf_evlist__parse_sample_timestamp(evlist, event, ×tamp); 1673 if (ret && ret != -1) 1674 return ret; 1675 1676 ret = perf_session__queue_event(session, event, timestamp, file_offset); 1677 if (ret != -ETIME) 1678 return ret; 1679 } 1680 1681 return perf_session__deliver_event(session, event, tool, file_offset); 1682 } 1683 1684 void perf_event_header__bswap(struct perf_event_header *hdr) 1685 { 1686 hdr->type = bswap_32(hdr->type); 1687 hdr->misc = bswap_16(hdr->misc); 1688 hdr->size = bswap_16(hdr->size); 1689 } 1690 1691 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1692 { 1693 return machine__findnew_thread(&session->machines.host, -1, pid); 1694 } 1695 1696 /* 1697 * Threads are identified by pid and tid, and the idle task has pid == tid == 0. 1698 * So here a single thread is created for that, but actually there is a separate 1699 * idle task per cpu, so there should be one 'struct thread' per cpu, but there 1700 * is only 1. That causes problems for some tools, requiring workarounds. For 1701 * example get_idle_thread() in builtin-sched.c, or thread_stack__per_cpu(). 1702 */ 1703 int perf_session__register_idle_thread(struct perf_session *session) 1704 { 1705 struct thread *thread; 1706 int err = 0; 1707 1708 thread = machine__findnew_thread(&session->machines.host, 0, 0); 1709 if (thread == NULL || thread__set_comm(thread, "swapper", 0)) { 1710 pr_err("problem inserting idle task.\n"); 1711 err = -1; 1712 } 1713 1714 if (thread == NULL || thread__set_namespaces(thread, 0, NULL)) { 1715 pr_err("problem inserting idle task.\n"); 1716 err = -1; 1717 } 1718 1719 /* machine__findnew_thread() got the thread, so put it */ 1720 thread__put(thread); 1721 return err; 1722 } 1723 1724 static void 1725 perf_session__warn_order(const struct perf_session *session) 1726 { 1727 const struct ordered_events *oe = &session->ordered_events; 1728 struct evsel *evsel; 1729 bool should_warn = true; 1730 1731 evlist__for_each_entry(session->evlist, evsel) { 1732 if (evsel->core.attr.write_backward) 1733 should_warn = false; 1734 } 1735 1736 if (!should_warn) 1737 return; 1738 if (oe->nr_unordered_events != 0) 1739 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events); 1740 } 1741 1742 static void perf_session__warn_about_errors(const struct perf_session *session) 1743 { 1744 const struct events_stats *stats = &session->evlist->stats; 1745 1746 if (session->tool->lost == perf_event__process_lost && 1747 stats->nr_events[PERF_RECORD_LOST] != 0) { 1748 ui__warning("Processed %d events and lost %d chunks!\n\n" 1749 "Check IO/CPU overload!\n\n", 1750 stats->nr_events[0], 1751 stats->nr_events[PERF_RECORD_LOST]); 1752 } 1753 1754 if (session->tool->lost_samples == perf_event__process_lost_samples) { 1755 double drop_rate; 1756 1757 drop_rate = (double)stats->total_lost_samples / 1758 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples); 1759 if (drop_rate > 0.05) { 1760 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n", 1761 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples, 1762 drop_rate * 100.0); 1763 } 1764 } 1765 1766 if (session->tool->aux == perf_event__process_aux && 1767 stats->total_aux_lost != 0) { 1768 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n", 1769 stats->total_aux_lost, 1770 stats->nr_events[PERF_RECORD_AUX]); 1771 } 1772 1773 if (session->tool->aux == perf_event__process_aux && 1774 stats->total_aux_partial != 0) { 1775 bool vmm_exclusive = false; 1776 1777 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive", 1778 &vmm_exclusive); 1779 1780 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n" 1781 "Are you running a KVM guest in the background?%s\n\n", 1782 stats->total_aux_partial, 1783 stats->nr_events[PERF_RECORD_AUX], 1784 vmm_exclusive ? 1785 "\nReloading kvm_intel module with vmm_exclusive=0\n" 1786 "will reduce the gaps to only guest's timeslices." : 1787 ""); 1788 } 1789 1790 if (stats->nr_unknown_events != 0) { 1791 ui__warning("Found %u unknown events!\n\n" 1792 "Is this an older tool processing a perf.data " 1793 "file generated by a more recent tool?\n\n" 1794 "If that is not the case, consider " 1795 "reporting to linux-kernel@vger.kernel.org.\n\n", 1796 stats->nr_unknown_events); 1797 } 1798 1799 if (stats->nr_unknown_id != 0) { 1800 ui__warning("%u samples with id not present in the header\n", 1801 stats->nr_unknown_id); 1802 } 1803 1804 if (stats->nr_invalid_chains != 0) { 1805 ui__warning("Found invalid callchains!\n\n" 1806 "%u out of %u events were discarded for this reason.\n\n" 1807 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 1808 stats->nr_invalid_chains, 1809 stats->nr_events[PERF_RECORD_SAMPLE]); 1810 } 1811 1812 if (stats->nr_unprocessable_samples != 0) { 1813 ui__warning("%u unprocessable samples recorded.\n" 1814 "Do you have a KVM guest running and not using 'perf kvm'?\n", 1815 stats->nr_unprocessable_samples); 1816 } 1817 1818 perf_session__warn_order(session); 1819 1820 events_stats__auxtrace_error_warn(stats); 1821 1822 if (stats->nr_proc_map_timeout != 0) { 1823 ui__warning("%d map information files for pre-existing threads were\n" 1824 "not processed, if there are samples for addresses they\n" 1825 "will not be resolved, you may find out which are these\n" 1826 "threads by running with -v and redirecting the output\n" 1827 "to a file.\n" 1828 "The time limit to process proc map is too short?\n" 1829 "Increase it by --proc-map-timeout\n", 1830 stats->nr_proc_map_timeout); 1831 } 1832 } 1833 1834 static int perf_session__flush_thread_stack(struct thread *thread, 1835 void *p __maybe_unused) 1836 { 1837 return thread_stack__flush(thread); 1838 } 1839 1840 static int perf_session__flush_thread_stacks(struct perf_session *session) 1841 { 1842 return machines__for_each_thread(&session->machines, 1843 perf_session__flush_thread_stack, 1844 NULL); 1845 } 1846 1847 volatile int session_done; 1848 1849 static int __perf_session__process_decomp_events(struct perf_session *session); 1850 1851 static int __perf_session__process_pipe_events(struct perf_session *session) 1852 { 1853 struct ordered_events *oe = &session->ordered_events; 1854 struct perf_tool *tool = session->tool; 1855 int fd = perf_data__fd(session->data); 1856 union perf_event *event; 1857 uint32_t size, cur_size = 0; 1858 void *buf = NULL; 1859 s64 skip = 0; 1860 u64 head; 1861 ssize_t err; 1862 void *p; 1863 1864 perf_tool__fill_defaults(tool); 1865 1866 head = 0; 1867 cur_size = sizeof(union perf_event); 1868 1869 buf = malloc(cur_size); 1870 if (!buf) 1871 return -errno; 1872 ordered_events__set_copy_on_queue(oe, true); 1873 more: 1874 event = buf; 1875 err = readn(fd, event, sizeof(struct perf_event_header)); 1876 if (err <= 0) { 1877 if (err == 0) 1878 goto done; 1879 1880 pr_err("failed to read event header\n"); 1881 goto out_err; 1882 } 1883 1884 if (session->header.needs_swap) 1885 perf_event_header__bswap(&event->header); 1886 1887 size = event->header.size; 1888 if (size < sizeof(struct perf_event_header)) { 1889 pr_err("bad event header size\n"); 1890 goto out_err; 1891 } 1892 1893 if (size > cur_size) { 1894 void *new = realloc(buf, size); 1895 if (!new) { 1896 pr_err("failed to allocate memory to read event\n"); 1897 goto out_err; 1898 } 1899 buf = new; 1900 cur_size = size; 1901 event = buf; 1902 } 1903 p = event; 1904 p += sizeof(struct perf_event_header); 1905 1906 if (size - sizeof(struct perf_event_header)) { 1907 err = readn(fd, p, size - sizeof(struct perf_event_header)); 1908 if (err <= 0) { 1909 if (err == 0) { 1910 pr_err("unexpected end of event stream\n"); 1911 goto done; 1912 } 1913 1914 pr_err("failed to read event data\n"); 1915 goto out_err; 1916 } 1917 } 1918 1919 if ((skip = perf_session__process_event(session, event, head)) < 0) { 1920 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 1921 head, event->header.size, event->header.type); 1922 err = -EINVAL; 1923 goto out_err; 1924 } 1925 1926 head += size; 1927 1928 if (skip > 0) 1929 head += skip; 1930 1931 err = __perf_session__process_decomp_events(session); 1932 if (err) 1933 goto out_err; 1934 1935 if (!session_done()) 1936 goto more; 1937 done: 1938 /* do the final flush for ordered samples */ 1939 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 1940 if (err) 1941 goto out_err; 1942 err = auxtrace__flush_events(session, tool); 1943 if (err) 1944 goto out_err; 1945 err = perf_session__flush_thread_stacks(session); 1946 out_err: 1947 free(buf); 1948 if (!tool->no_warn) 1949 perf_session__warn_about_errors(session); 1950 ordered_events__free(&session->ordered_events); 1951 auxtrace__free_events(session); 1952 return err; 1953 } 1954 1955 static union perf_event * 1956 fetch_mmaped_event(struct perf_session *session, 1957 u64 head, size_t mmap_size, char *buf) 1958 { 1959 union perf_event *event; 1960 1961 /* 1962 * Ensure we have enough space remaining to read 1963 * the size of the event in the headers. 1964 */ 1965 if (head + sizeof(event->header) > mmap_size) 1966 return NULL; 1967 1968 event = (union perf_event *)(buf + head); 1969 1970 if (session->header.needs_swap) 1971 perf_event_header__bswap(&event->header); 1972 1973 if (head + event->header.size > mmap_size) { 1974 /* We're not fetching the event so swap back again */ 1975 if (session->header.needs_swap) 1976 perf_event_header__bswap(&event->header); 1977 pr_debug("%s: head=%#" PRIx64 " event->header_size=%#x, mmap_size=%#zx: fuzzed perf.data?\n", 1978 __func__, head, event->header.size, mmap_size); 1979 return ERR_PTR(-EINVAL); 1980 } 1981 1982 return event; 1983 } 1984 1985 static int __perf_session__process_decomp_events(struct perf_session *session) 1986 { 1987 s64 skip; 1988 u64 size, file_pos = 0; 1989 struct decomp *decomp = session->decomp_last; 1990 1991 if (!decomp) 1992 return 0; 1993 1994 while (decomp->head < decomp->size && !session_done()) { 1995 union perf_event *event = fetch_mmaped_event(session, decomp->head, decomp->size, decomp->data); 1996 1997 if (IS_ERR(event)) 1998 return PTR_ERR(event); 1999 2000 if (!event) 2001 break; 2002 2003 size = event->header.size; 2004 2005 if (size < sizeof(struct perf_event_header) || 2006 (skip = perf_session__process_event(session, event, file_pos)) < 0) { 2007 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 2008 decomp->file_pos + decomp->head, event->header.size, event->header.type); 2009 return -EINVAL; 2010 } 2011 2012 if (skip) 2013 size += skip; 2014 2015 decomp->head += size; 2016 } 2017 2018 return 0; 2019 } 2020 2021 /* 2022 * On 64bit we can mmap the data file in one go. No need for tiny mmap 2023 * slices. On 32bit we use 32MB. 2024 */ 2025 #if BITS_PER_LONG == 64 2026 #define MMAP_SIZE ULLONG_MAX 2027 #define NUM_MMAPS 1 2028 #else 2029 #define MMAP_SIZE (32 * 1024 * 1024ULL) 2030 #define NUM_MMAPS 128 2031 #endif 2032 2033 struct reader; 2034 2035 typedef s64 (*reader_cb_t)(struct perf_session *session, 2036 union perf_event *event, 2037 u64 file_offset); 2038 2039 struct reader { 2040 int fd; 2041 u64 data_size; 2042 u64 data_offset; 2043 reader_cb_t process; 2044 }; 2045 2046 static int 2047 reader__process_events(struct reader *rd, struct perf_session *session, 2048 struct ui_progress *prog) 2049 { 2050 u64 data_size = rd->data_size; 2051 u64 head, page_offset, file_offset, file_pos, size; 2052 int err = 0, mmap_prot, mmap_flags, map_idx = 0; 2053 size_t mmap_size; 2054 char *buf, *mmaps[NUM_MMAPS]; 2055 union perf_event *event; 2056 s64 skip; 2057 2058 page_offset = page_size * (rd->data_offset / page_size); 2059 file_offset = page_offset; 2060 head = rd->data_offset - page_offset; 2061 2062 ui_progress__init_size(prog, data_size, "Processing events..."); 2063 2064 data_size += rd->data_offset; 2065 2066 mmap_size = MMAP_SIZE; 2067 if (mmap_size > data_size) { 2068 mmap_size = data_size; 2069 session->one_mmap = true; 2070 } 2071 2072 memset(mmaps, 0, sizeof(mmaps)); 2073 2074 mmap_prot = PROT_READ; 2075 mmap_flags = MAP_SHARED; 2076 2077 if (session->header.needs_swap) { 2078 mmap_prot |= PROT_WRITE; 2079 mmap_flags = MAP_PRIVATE; 2080 } 2081 remap: 2082 buf = mmap(NULL, mmap_size, mmap_prot, mmap_flags, rd->fd, 2083 file_offset); 2084 if (buf == MAP_FAILED) { 2085 pr_err("failed to mmap file\n"); 2086 err = -errno; 2087 goto out; 2088 } 2089 mmaps[map_idx] = buf; 2090 map_idx = (map_idx + 1) & (ARRAY_SIZE(mmaps) - 1); 2091 file_pos = file_offset + head; 2092 if (session->one_mmap) { 2093 session->one_mmap_addr = buf; 2094 session->one_mmap_offset = file_offset; 2095 } 2096 2097 more: 2098 event = fetch_mmaped_event(session, head, mmap_size, buf); 2099 if (IS_ERR(event)) 2100 return PTR_ERR(event); 2101 2102 if (!event) { 2103 if (mmaps[map_idx]) { 2104 munmap(mmaps[map_idx], mmap_size); 2105 mmaps[map_idx] = NULL; 2106 } 2107 2108 page_offset = page_size * (head / page_size); 2109 file_offset += page_offset; 2110 head -= page_offset; 2111 goto remap; 2112 } 2113 2114 size = event->header.size; 2115 2116 skip = -EINVAL; 2117 2118 if (size < sizeof(struct perf_event_header) || 2119 (skip = rd->process(session, event, file_pos)) < 0) { 2120 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n", 2121 file_offset + head, event->header.size, 2122 event->header.type, strerror(-skip)); 2123 err = skip; 2124 goto out; 2125 } 2126 2127 if (skip) 2128 size += skip; 2129 2130 head += size; 2131 file_pos += size; 2132 2133 err = __perf_session__process_decomp_events(session); 2134 if (err) 2135 goto out; 2136 2137 ui_progress__update(prog, size); 2138 2139 if (session_done()) 2140 goto out; 2141 2142 if (file_pos < data_size) 2143 goto more; 2144 2145 out: 2146 return err; 2147 } 2148 2149 static s64 process_simple(struct perf_session *session, 2150 union perf_event *event, 2151 u64 file_offset) 2152 { 2153 return perf_session__process_event(session, event, file_offset); 2154 } 2155 2156 static int __perf_session__process_events(struct perf_session *session) 2157 { 2158 struct reader rd = { 2159 .fd = perf_data__fd(session->data), 2160 .data_size = session->header.data_size, 2161 .data_offset = session->header.data_offset, 2162 .process = process_simple, 2163 }; 2164 struct ordered_events *oe = &session->ordered_events; 2165 struct perf_tool *tool = session->tool; 2166 struct ui_progress prog; 2167 int err; 2168 2169 perf_tool__fill_defaults(tool); 2170 2171 if (rd.data_size == 0) 2172 return -1; 2173 2174 ui_progress__init_size(&prog, rd.data_size, "Processing events..."); 2175 2176 err = reader__process_events(&rd, session, &prog); 2177 if (err) 2178 goto out_err; 2179 /* do the final flush for ordered samples */ 2180 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2181 if (err) 2182 goto out_err; 2183 err = auxtrace__flush_events(session, tool); 2184 if (err) 2185 goto out_err; 2186 err = perf_session__flush_thread_stacks(session); 2187 out_err: 2188 ui_progress__finish(); 2189 if (!tool->no_warn) 2190 perf_session__warn_about_errors(session); 2191 /* 2192 * We may switching perf.data output, make ordered_events 2193 * reusable. 2194 */ 2195 ordered_events__reinit(&session->ordered_events); 2196 auxtrace__free_events(session); 2197 session->one_mmap = false; 2198 return err; 2199 } 2200 2201 int perf_session__process_events(struct perf_session *session) 2202 { 2203 if (perf_session__register_idle_thread(session) < 0) 2204 return -ENOMEM; 2205 2206 if (perf_data__is_pipe(session->data)) 2207 return __perf_session__process_pipe_events(session); 2208 2209 return __perf_session__process_events(session); 2210 } 2211 2212 bool perf_session__has_traces(struct perf_session *session, const char *msg) 2213 { 2214 struct evsel *evsel; 2215 2216 evlist__for_each_entry(session->evlist, evsel) { 2217 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT) 2218 return true; 2219 } 2220 2221 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 2222 return false; 2223 } 2224 2225 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr) 2226 { 2227 char *bracket; 2228 struct ref_reloc_sym *ref; 2229 struct kmap *kmap; 2230 2231 ref = zalloc(sizeof(struct ref_reloc_sym)); 2232 if (ref == NULL) 2233 return -ENOMEM; 2234 2235 ref->name = strdup(symbol_name); 2236 if (ref->name == NULL) { 2237 free(ref); 2238 return -ENOMEM; 2239 } 2240 2241 bracket = strchr(ref->name, ']'); 2242 if (bracket) 2243 *bracket = '\0'; 2244 2245 ref->addr = addr; 2246 2247 kmap = map__kmap(map); 2248 if (kmap) 2249 kmap->ref_reloc_sym = ref; 2250 2251 return 0; 2252 } 2253 2254 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp) 2255 { 2256 return machines__fprintf_dsos(&session->machines, fp); 2257 } 2258 2259 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp, 2260 bool (skip)(struct dso *dso, int parm), int parm) 2261 { 2262 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm); 2263 } 2264 2265 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 2266 { 2267 size_t ret; 2268 const char *msg = ""; 2269 2270 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) 2271 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)"; 2272 2273 ret = fprintf(fp, "\nAggregated stats:%s\n", msg); 2274 2275 ret += events_stats__fprintf(&session->evlist->stats, fp); 2276 return ret; 2277 } 2278 2279 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 2280 { 2281 /* 2282 * FIXME: Here we have to actually print all the machines in this 2283 * session, not just the host... 2284 */ 2285 return machine__fprintf(&session->machines.host, fp); 2286 } 2287 2288 struct evsel *perf_session__find_first_evtype(struct perf_session *session, 2289 unsigned int type) 2290 { 2291 struct evsel *pos; 2292 2293 evlist__for_each_entry(session->evlist, pos) { 2294 if (pos->core.attr.type == type) 2295 return pos; 2296 } 2297 return NULL; 2298 } 2299 2300 int perf_session__cpu_bitmap(struct perf_session *session, 2301 const char *cpu_list, unsigned long *cpu_bitmap) 2302 { 2303 int i, err = -1; 2304 struct perf_cpu_map *map; 2305 int nr_cpus = min(session->header.env.nr_cpus_online, MAX_NR_CPUS); 2306 2307 for (i = 0; i < PERF_TYPE_MAX; ++i) { 2308 struct evsel *evsel; 2309 2310 evsel = perf_session__find_first_evtype(session, i); 2311 if (!evsel) 2312 continue; 2313 2314 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) { 2315 pr_err("File does not contain CPU events. " 2316 "Remove -C option to proceed.\n"); 2317 return -1; 2318 } 2319 } 2320 2321 map = perf_cpu_map__new(cpu_list); 2322 if (map == NULL) { 2323 pr_err("Invalid cpu_list\n"); 2324 return -1; 2325 } 2326 2327 for (i = 0; i < map->nr; i++) { 2328 int cpu = map->map[i]; 2329 2330 if (cpu >= nr_cpus) { 2331 pr_err("Requested CPU %d too large. " 2332 "Consider raising MAX_NR_CPUS\n", cpu); 2333 goto out_delete_map; 2334 } 2335 2336 set_bit(cpu, cpu_bitmap); 2337 } 2338 2339 err = 0; 2340 2341 out_delete_map: 2342 perf_cpu_map__put(map); 2343 return err; 2344 } 2345 2346 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 2347 bool full) 2348 { 2349 if (session == NULL || fp == NULL) 2350 return; 2351 2352 fprintf(fp, "# ========\n"); 2353 perf_header__fprintf_info(session, fp, full); 2354 fprintf(fp, "# ========\n#\n"); 2355 } 2356 2357 2358 int __perf_session__set_tracepoints_handlers(struct perf_session *session, 2359 const struct evsel_str_handler *assocs, 2360 size_t nr_assocs) 2361 { 2362 struct evsel *evsel; 2363 size_t i; 2364 int err; 2365 2366 for (i = 0; i < nr_assocs; i++) { 2367 /* 2368 * Adding a handler for an event not in the session, 2369 * just ignore it. 2370 */ 2371 evsel = perf_evlist__find_tracepoint_by_name(session->evlist, assocs[i].name); 2372 if (evsel == NULL) 2373 continue; 2374 2375 err = -EEXIST; 2376 if (evsel->handler != NULL) 2377 goto out; 2378 evsel->handler = assocs[i].handler; 2379 } 2380 2381 err = 0; 2382 out: 2383 return err; 2384 } 2385 2386 int perf_event__process_id_index(struct perf_session *session, 2387 union perf_event *event) 2388 { 2389 struct evlist *evlist = session->evlist; 2390 struct perf_record_id_index *ie = &event->id_index; 2391 size_t i, nr, max_nr; 2392 2393 max_nr = (ie->header.size - sizeof(struct perf_record_id_index)) / 2394 sizeof(struct id_index_entry); 2395 nr = ie->nr; 2396 if (nr > max_nr) 2397 return -EINVAL; 2398 2399 if (dump_trace) 2400 fprintf(stdout, " nr: %zu\n", nr); 2401 2402 for (i = 0; i < nr; i++) { 2403 struct id_index_entry *e = &ie->entries[i]; 2404 struct perf_sample_id *sid; 2405 2406 if (dump_trace) { 2407 fprintf(stdout, " ... id: %"PRI_lu64, e->id); 2408 fprintf(stdout, " idx: %"PRI_lu64, e->idx); 2409 fprintf(stdout, " cpu: %"PRI_ld64, e->cpu); 2410 fprintf(stdout, " tid: %"PRI_ld64"\n", e->tid); 2411 } 2412 2413 sid = perf_evlist__id2sid(evlist, e->id); 2414 if (!sid) 2415 return -ENOENT; 2416 sid->idx = e->idx; 2417 sid->cpu = e->cpu; 2418 sid->tid = e->tid; 2419 } 2420 return 0; 2421 } 2422