1 // SPDX-License-Identifier: GPL-2.0 2 #include <errno.h> 3 #include <signal.h> 4 #include <inttypes.h> 5 #include <linux/err.h> 6 #include <linux/kernel.h> 7 #include <linux/zalloc.h> 8 #include <api/fs/fs.h> 9 10 #include <byteswap.h> 11 #include <unistd.h> 12 #include <sys/types.h> 13 #include <sys/mman.h> 14 #include <perf/cpumap.h> 15 16 #include "map_symbol.h" 17 #include "branch.h" 18 #include "debug.h" 19 #include "env.h" 20 #include "evlist.h" 21 #include "evsel.h" 22 #include "memswap.h" 23 #include "map.h" 24 #include "symbol.h" 25 #include "session.h" 26 #include "tool.h" 27 #include "perf_regs.h" 28 #include "asm/bug.h" 29 #include "auxtrace.h" 30 #include "thread.h" 31 #include "thread-stack.h" 32 #include "sample-raw.h" 33 #include "stat.h" 34 #include "tsc.h" 35 #include "ui/progress.h" 36 #include "util.h" 37 #include "arch/common.h" 38 #include "units.h" 39 #include "annotate.h" 40 #include "perf.h" 41 #include <internal/lib.h> 42 43 static int perf_session__deliver_event(struct perf_session *session, 44 union perf_event *event, 45 const struct perf_tool *tool, 46 u64 file_offset, 47 const char *file_path); 48 49 static int perf_session__open(struct perf_session *session) 50 { 51 struct perf_data *data = session->data; 52 53 if (perf_session__read_header(session) < 0) { 54 pr_err("incompatible file format (rerun with -v to learn more)\n"); 55 return -1; 56 } 57 58 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) { 59 /* Auxiliary events may reference exited threads, hold onto dead ones. */ 60 symbol_conf.keep_exited_threads = true; 61 } 62 63 if (perf_data__is_pipe(data)) 64 return 0; 65 66 if (perf_header__has_feat(&session->header, HEADER_STAT)) 67 return 0; 68 69 if (!evlist__valid_sample_type(session->evlist)) { 70 pr_err("non matching sample_type\n"); 71 return -1; 72 } 73 74 if (!evlist__valid_sample_id_all(session->evlist)) { 75 pr_err("non matching sample_id_all\n"); 76 return -1; 77 } 78 79 if (!evlist__valid_read_format(session->evlist)) { 80 pr_err("non matching read_format\n"); 81 return -1; 82 } 83 84 return 0; 85 } 86 87 void perf_session__set_id_hdr_size(struct perf_session *session) 88 { 89 u16 id_hdr_size = evlist__id_hdr_size(session->evlist); 90 91 machines__set_id_hdr_size(&session->machines, id_hdr_size); 92 } 93 94 int perf_session__create_kernel_maps(struct perf_session *session) 95 { 96 int ret = machine__create_kernel_maps(&session->machines.host); 97 98 if (ret >= 0) 99 ret = machines__create_guest_kernel_maps(&session->machines); 100 return ret; 101 } 102 103 static void perf_session__destroy_kernel_maps(struct perf_session *session) 104 { 105 machines__destroy_kernel_maps(&session->machines); 106 } 107 108 static bool perf_session__has_comm_exec(struct perf_session *session) 109 { 110 struct evsel *evsel; 111 112 evlist__for_each_entry(session->evlist, evsel) { 113 if (evsel->core.attr.comm_exec) 114 return true; 115 } 116 117 return false; 118 } 119 120 static void perf_session__set_comm_exec(struct perf_session *session) 121 { 122 bool comm_exec = perf_session__has_comm_exec(session); 123 124 machines__set_comm_exec(&session->machines, comm_exec); 125 } 126 127 static int ordered_events__deliver_event(struct ordered_events *oe, 128 struct ordered_event *event) 129 { 130 struct perf_session *session = container_of(oe, struct perf_session, 131 ordered_events); 132 133 return perf_session__deliver_event(session, event->event, 134 session->tool, event->file_offset, 135 event->file_path); 136 } 137 138 struct perf_session *__perf_session__new(struct perf_data *data, 139 struct perf_tool *tool, 140 bool trace_event_repipe) 141 { 142 int ret = -ENOMEM; 143 struct perf_session *session = zalloc(sizeof(*session)); 144 145 if (!session) 146 goto out; 147 148 session->trace_event_repipe = trace_event_repipe; 149 session->tool = tool; 150 session->decomp_data.zstd_decomp = &session->zstd_data; 151 session->active_decomp = &session->decomp_data; 152 INIT_LIST_HEAD(&session->auxtrace_index); 153 machines__init(&session->machines); 154 ordered_events__init(&session->ordered_events, 155 ordered_events__deliver_event, NULL); 156 157 perf_env__init(&session->header.env); 158 if (data) { 159 ret = perf_data__open(data); 160 if (ret < 0) 161 goto out_delete; 162 163 session->data = data; 164 165 if (perf_data__is_read(data)) { 166 ret = perf_session__open(session); 167 if (ret < 0) 168 goto out_delete; 169 170 /* 171 * set session attributes that are present in perf.data 172 * but not in pipe-mode. 173 */ 174 if (!data->is_pipe) { 175 perf_session__set_id_hdr_size(session); 176 perf_session__set_comm_exec(session); 177 } 178 179 evlist__init_trace_event_sample_raw(session->evlist); 180 181 /* Open the directory data. */ 182 if (data->is_dir) { 183 ret = perf_data__open_dir(data); 184 if (ret) 185 goto out_delete; 186 } 187 188 if (!symbol_conf.kallsyms_name && 189 !symbol_conf.vmlinux_name) 190 symbol_conf.kallsyms_name = perf_data__kallsyms_name(data); 191 } 192 } else { 193 session->machines.host.env = &perf_env; 194 } 195 196 session->machines.host.single_address_space = 197 perf_env__single_address_space(session->machines.host.env); 198 199 if (!data || perf_data__is_write(data)) { 200 /* 201 * In O_RDONLY mode this will be performed when reading the 202 * kernel MMAP event, in perf_event__process_mmap(). 203 */ 204 if (perf_session__create_kernel_maps(session) < 0) 205 pr_warning("Cannot read kernel map\n"); 206 } 207 208 /* 209 * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is 210 * processed, so evlist__sample_id_all is not meaningful here. 211 */ 212 if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps && 213 tool->ordered_events && !evlist__sample_id_all(session->evlist)) { 214 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 215 tool->ordered_events = false; 216 } 217 218 return session; 219 220 out_delete: 221 perf_session__delete(session); 222 out: 223 return ERR_PTR(ret); 224 } 225 226 static void perf_decomp__release_events(struct decomp *next) 227 { 228 struct decomp *decomp; 229 size_t mmap_len; 230 231 do { 232 decomp = next; 233 if (decomp == NULL) 234 break; 235 next = decomp->next; 236 mmap_len = decomp->mmap_len; 237 munmap(decomp, mmap_len); 238 } while (1); 239 } 240 241 void perf_session__delete(struct perf_session *session) 242 { 243 if (session == NULL) 244 return; 245 auxtrace__free(session); 246 auxtrace_index__free(&session->auxtrace_index); 247 debuginfo_cache__delete(); 248 perf_session__destroy_kernel_maps(session); 249 perf_decomp__release_events(session->decomp_data.decomp); 250 perf_env__exit(&session->header.env); 251 machines__exit(&session->machines); 252 if (session->data) { 253 if (perf_data__is_read(session->data)) 254 evlist__delete(session->evlist); 255 perf_data__close(session->data); 256 } 257 #ifdef HAVE_LIBTRACEEVENT 258 trace_event__cleanup(&session->tevent); 259 #endif 260 free(session); 261 } 262 263 static void swap_sample_id_all(union perf_event *event, void *data) 264 { 265 void *end = (void *) event + event->header.size; 266 int size = end - data; 267 268 BUG_ON(size % sizeof(u64)); 269 mem_bswap_64(data, size); 270 } 271 272 static void perf_event__all64_swap(union perf_event *event, 273 bool sample_id_all __maybe_unused) 274 { 275 struct perf_event_header *hdr = &event->header; 276 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 277 } 278 279 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all) 280 { 281 event->comm.pid = bswap_32(event->comm.pid); 282 event->comm.tid = bswap_32(event->comm.tid); 283 284 if (sample_id_all) { 285 void *data = &event->comm.comm; 286 287 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 288 swap_sample_id_all(event, data); 289 } 290 } 291 292 static void perf_event__mmap_swap(union perf_event *event, 293 bool sample_id_all) 294 { 295 event->mmap.pid = bswap_32(event->mmap.pid); 296 event->mmap.tid = bswap_32(event->mmap.tid); 297 event->mmap.start = bswap_64(event->mmap.start); 298 event->mmap.len = bswap_64(event->mmap.len); 299 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 300 301 if (sample_id_all) { 302 void *data = &event->mmap.filename; 303 304 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 305 swap_sample_id_all(event, data); 306 } 307 } 308 309 static void perf_event__mmap2_swap(union perf_event *event, 310 bool sample_id_all) 311 { 312 event->mmap2.pid = bswap_32(event->mmap2.pid); 313 event->mmap2.tid = bswap_32(event->mmap2.tid); 314 event->mmap2.start = bswap_64(event->mmap2.start); 315 event->mmap2.len = bswap_64(event->mmap2.len); 316 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff); 317 318 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID)) { 319 event->mmap2.maj = bswap_32(event->mmap2.maj); 320 event->mmap2.min = bswap_32(event->mmap2.min); 321 event->mmap2.ino = bswap_64(event->mmap2.ino); 322 event->mmap2.ino_generation = bswap_64(event->mmap2.ino_generation); 323 } 324 325 if (sample_id_all) { 326 void *data = &event->mmap2.filename; 327 328 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 329 swap_sample_id_all(event, data); 330 } 331 } 332 static void perf_event__task_swap(union perf_event *event, bool sample_id_all) 333 { 334 event->fork.pid = bswap_32(event->fork.pid); 335 event->fork.tid = bswap_32(event->fork.tid); 336 event->fork.ppid = bswap_32(event->fork.ppid); 337 event->fork.ptid = bswap_32(event->fork.ptid); 338 event->fork.time = bswap_64(event->fork.time); 339 340 if (sample_id_all) 341 swap_sample_id_all(event, &event->fork + 1); 342 } 343 344 static void perf_event__read_swap(union perf_event *event, bool sample_id_all) 345 { 346 event->read.pid = bswap_32(event->read.pid); 347 event->read.tid = bswap_32(event->read.tid); 348 event->read.value = bswap_64(event->read.value); 349 event->read.time_enabled = bswap_64(event->read.time_enabled); 350 event->read.time_running = bswap_64(event->read.time_running); 351 event->read.id = bswap_64(event->read.id); 352 353 if (sample_id_all) 354 swap_sample_id_all(event, &event->read + 1); 355 } 356 357 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all) 358 { 359 event->aux.aux_offset = bswap_64(event->aux.aux_offset); 360 event->aux.aux_size = bswap_64(event->aux.aux_size); 361 event->aux.flags = bswap_64(event->aux.flags); 362 363 if (sample_id_all) 364 swap_sample_id_all(event, &event->aux + 1); 365 } 366 367 static void perf_event__itrace_start_swap(union perf_event *event, 368 bool sample_id_all) 369 { 370 event->itrace_start.pid = bswap_32(event->itrace_start.pid); 371 event->itrace_start.tid = bswap_32(event->itrace_start.tid); 372 373 if (sample_id_all) 374 swap_sample_id_all(event, &event->itrace_start + 1); 375 } 376 377 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all) 378 { 379 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) { 380 event->context_switch.next_prev_pid = 381 bswap_32(event->context_switch.next_prev_pid); 382 event->context_switch.next_prev_tid = 383 bswap_32(event->context_switch.next_prev_tid); 384 } 385 386 if (sample_id_all) 387 swap_sample_id_all(event, &event->context_switch + 1); 388 } 389 390 static void perf_event__text_poke_swap(union perf_event *event, bool sample_id_all) 391 { 392 event->text_poke.addr = bswap_64(event->text_poke.addr); 393 event->text_poke.old_len = bswap_16(event->text_poke.old_len); 394 event->text_poke.new_len = bswap_16(event->text_poke.new_len); 395 396 if (sample_id_all) { 397 size_t len = sizeof(event->text_poke.old_len) + 398 sizeof(event->text_poke.new_len) + 399 event->text_poke.old_len + 400 event->text_poke.new_len; 401 void *data = &event->text_poke.old_len; 402 403 data += PERF_ALIGN(len, sizeof(u64)); 404 swap_sample_id_all(event, data); 405 } 406 } 407 408 static void perf_event__throttle_swap(union perf_event *event, 409 bool sample_id_all) 410 { 411 event->throttle.time = bswap_64(event->throttle.time); 412 event->throttle.id = bswap_64(event->throttle.id); 413 event->throttle.stream_id = bswap_64(event->throttle.stream_id); 414 415 if (sample_id_all) 416 swap_sample_id_all(event, &event->throttle + 1); 417 } 418 419 static void perf_event__namespaces_swap(union perf_event *event, 420 bool sample_id_all) 421 { 422 u64 i; 423 424 event->namespaces.pid = bswap_32(event->namespaces.pid); 425 event->namespaces.tid = bswap_32(event->namespaces.tid); 426 event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces); 427 428 for (i = 0; i < event->namespaces.nr_namespaces; i++) { 429 struct perf_ns_link_info *ns = &event->namespaces.link_info[i]; 430 431 ns->dev = bswap_64(ns->dev); 432 ns->ino = bswap_64(ns->ino); 433 } 434 435 if (sample_id_all) 436 swap_sample_id_all(event, &event->namespaces.link_info[i]); 437 } 438 439 static void perf_event__cgroup_swap(union perf_event *event, bool sample_id_all) 440 { 441 event->cgroup.id = bswap_64(event->cgroup.id); 442 443 if (sample_id_all) { 444 void *data = &event->cgroup.path; 445 446 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 447 swap_sample_id_all(event, data); 448 } 449 } 450 451 static u8 revbyte(u8 b) 452 { 453 int rev = (b >> 4) | ((b & 0xf) << 4); 454 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2); 455 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1); 456 return (u8) rev; 457 } 458 459 /* 460 * XXX this is hack in attempt to carry flags bitfield 461 * through endian village. ABI says: 462 * 463 * Bit-fields are allocated from right to left (least to most significant) 464 * on little-endian implementations and from left to right (most to least 465 * significant) on big-endian implementations. 466 * 467 * The above seems to be byte specific, so we need to reverse each 468 * byte of the bitfield. 'Internet' also says this might be implementation 469 * specific and we probably need proper fix and carry perf_event_attr 470 * bitfield flags in separate data file FEAT_ section. Thought this seems 471 * to work for now. 472 */ 473 static void swap_bitfield(u8 *p, unsigned len) 474 { 475 unsigned i; 476 477 for (i = 0; i < len; i++) { 478 *p = revbyte(*p); 479 p++; 480 } 481 } 482 483 /* exported for swapping attributes in file header */ 484 void perf_event__attr_swap(struct perf_event_attr *attr) 485 { 486 attr->type = bswap_32(attr->type); 487 attr->size = bswap_32(attr->size); 488 489 #define bswap_safe(f, n) \ 490 (attr->size > (offsetof(struct perf_event_attr, f) + \ 491 sizeof(attr->f) * (n))) 492 #define bswap_field(f, sz) \ 493 do { \ 494 if (bswap_safe(f, 0)) \ 495 attr->f = bswap_##sz(attr->f); \ 496 } while(0) 497 #define bswap_field_16(f) bswap_field(f, 16) 498 #define bswap_field_32(f) bswap_field(f, 32) 499 #define bswap_field_64(f) bswap_field(f, 64) 500 501 bswap_field_64(config); 502 bswap_field_64(sample_period); 503 bswap_field_64(sample_type); 504 bswap_field_64(read_format); 505 bswap_field_32(wakeup_events); 506 bswap_field_32(bp_type); 507 bswap_field_64(bp_addr); 508 bswap_field_64(bp_len); 509 bswap_field_64(branch_sample_type); 510 bswap_field_64(sample_regs_user); 511 bswap_field_32(sample_stack_user); 512 bswap_field_32(aux_watermark); 513 bswap_field_16(sample_max_stack); 514 bswap_field_32(aux_sample_size); 515 516 /* 517 * After read_format are bitfields. Check read_format because 518 * we are unable to use offsetof on bitfield. 519 */ 520 if (bswap_safe(read_format, 1)) 521 swap_bitfield((u8 *) (&attr->read_format + 1), 522 sizeof(u64)); 523 #undef bswap_field_64 524 #undef bswap_field_32 525 #undef bswap_field 526 #undef bswap_safe 527 } 528 529 static void perf_event__hdr_attr_swap(union perf_event *event, 530 bool sample_id_all __maybe_unused) 531 { 532 size_t size; 533 534 perf_event__attr_swap(&event->attr.attr); 535 536 size = event->header.size; 537 size -= perf_record_header_attr_id(event) - (void *)event; 538 mem_bswap_64(perf_record_header_attr_id(event), size); 539 } 540 541 static void perf_event__event_update_swap(union perf_event *event, 542 bool sample_id_all __maybe_unused) 543 { 544 event->event_update.type = bswap_64(event->event_update.type); 545 event->event_update.id = bswap_64(event->event_update.id); 546 } 547 548 static void perf_event__event_type_swap(union perf_event *event, 549 bool sample_id_all __maybe_unused) 550 { 551 event->event_type.event_type.event_id = 552 bswap_64(event->event_type.event_type.event_id); 553 } 554 555 static void perf_event__tracing_data_swap(union perf_event *event, 556 bool sample_id_all __maybe_unused) 557 { 558 event->tracing_data.size = bswap_32(event->tracing_data.size); 559 } 560 561 static void perf_event__auxtrace_info_swap(union perf_event *event, 562 bool sample_id_all __maybe_unused) 563 { 564 size_t size; 565 566 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type); 567 568 size = event->header.size; 569 size -= (void *)&event->auxtrace_info.priv - (void *)event; 570 mem_bswap_64(event->auxtrace_info.priv, size); 571 } 572 573 static void perf_event__auxtrace_swap(union perf_event *event, 574 bool sample_id_all __maybe_unused) 575 { 576 event->auxtrace.size = bswap_64(event->auxtrace.size); 577 event->auxtrace.offset = bswap_64(event->auxtrace.offset); 578 event->auxtrace.reference = bswap_64(event->auxtrace.reference); 579 event->auxtrace.idx = bswap_32(event->auxtrace.idx); 580 event->auxtrace.tid = bswap_32(event->auxtrace.tid); 581 event->auxtrace.cpu = bswap_32(event->auxtrace.cpu); 582 } 583 584 static void perf_event__auxtrace_error_swap(union perf_event *event, 585 bool sample_id_all __maybe_unused) 586 { 587 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type); 588 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code); 589 event->auxtrace_error.cpu = bswap_32(event->auxtrace_error.cpu); 590 event->auxtrace_error.pid = bswap_32(event->auxtrace_error.pid); 591 event->auxtrace_error.tid = bswap_32(event->auxtrace_error.tid); 592 event->auxtrace_error.fmt = bswap_32(event->auxtrace_error.fmt); 593 event->auxtrace_error.ip = bswap_64(event->auxtrace_error.ip); 594 if (event->auxtrace_error.fmt) 595 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time); 596 if (event->auxtrace_error.fmt >= 2) { 597 event->auxtrace_error.machine_pid = bswap_32(event->auxtrace_error.machine_pid); 598 event->auxtrace_error.vcpu = bswap_32(event->auxtrace_error.vcpu); 599 } 600 } 601 602 static void perf_event__thread_map_swap(union perf_event *event, 603 bool sample_id_all __maybe_unused) 604 { 605 unsigned i; 606 607 event->thread_map.nr = bswap_64(event->thread_map.nr); 608 609 for (i = 0; i < event->thread_map.nr; i++) 610 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid); 611 } 612 613 static void perf_event__cpu_map_swap(union perf_event *event, 614 bool sample_id_all __maybe_unused) 615 { 616 struct perf_record_cpu_map_data *data = &event->cpu_map.data; 617 618 data->type = bswap_16(data->type); 619 620 switch (data->type) { 621 case PERF_CPU_MAP__CPUS: 622 data->cpus_data.nr = bswap_16(data->cpus_data.nr); 623 624 for (unsigned i = 0; i < data->cpus_data.nr; i++) 625 data->cpus_data.cpu[i] = bswap_16(data->cpus_data.cpu[i]); 626 break; 627 case PERF_CPU_MAP__MASK: 628 data->mask32_data.long_size = bswap_16(data->mask32_data.long_size); 629 630 switch (data->mask32_data.long_size) { 631 case 4: 632 data->mask32_data.nr = bswap_16(data->mask32_data.nr); 633 for (unsigned i = 0; i < data->mask32_data.nr; i++) 634 data->mask32_data.mask[i] = bswap_32(data->mask32_data.mask[i]); 635 break; 636 case 8: 637 data->mask64_data.nr = bswap_16(data->mask64_data.nr); 638 for (unsigned i = 0; i < data->mask64_data.nr; i++) 639 data->mask64_data.mask[i] = bswap_64(data->mask64_data.mask[i]); 640 break; 641 default: 642 pr_err("cpu_map swap: unsupported long size\n"); 643 } 644 break; 645 case PERF_CPU_MAP__RANGE_CPUS: 646 data->range_cpu_data.start_cpu = bswap_16(data->range_cpu_data.start_cpu); 647 data->range_cpu_data.end_cpu = bswap_16(data->range_cpu_data.end_cpu); 648 break; 649 default: 650 break; 651 } 652 } 653 654 static void perf_event__stat_config_swap(union perf_event *event, 655 bool sample_id_all __maybe_unused) 656 { 657 u64 size; 658 659 size = bswap_64(event->stat_config.nr) * sizeof(event->stat_config.data[0]); 660 size += 1; /* nr item itself */ 661 mem_bswap_64(&event->stat_config.nr, size); 662 } 663 664 static void perf_event__stat_swap(union perf_event *event, 665 bool sample_id_all __maybe_unused) 666 { 667 event->stat.id = bswap_64(event->stat.id); 668 event->stat.thread = bswap_32(event->stat.thread); 669 event->stat.cpu = bswap_32(event->stat.cpu); 670 event->stat.val = bswap_64(event->stat.val); 671 event->stat.ena = bswap_64(event->stat.ena); 672 event->stat.run = bswap_64(event->stat.run); 673 } 674 675 static void perf_event__stat_round_swap(union perf_event *event, 676 bool sample_id_all __maybe_unused) 677 { 678 event->stat_round.type = bswap_64(event->stat_round.type); 679 event->stat_round.time = bswap_64(event->stat_round.time); 680 } 681 682 static void perf_event__time_conv_swap(union perf_event *event, 683 bool sample_id_all __maybe_unused) 684 { 685 event->time_conv.time_shift = bswap_64(event->time_conv.time_shift); 686 event->time_conv.time_mult = bswap_64(event->time_conv.time_mult); 687 event->time_conv.time_zero = bswap_64(event->time_conv.time_zero); 688 689 if (event_contains(event->time_conv, time_cycles)) { 690 event->time_conv.time_cycles = bswap_64(event->time_conv.time_cycles); 691 event->time_conv.time_mask = bswap_64(event->time_conv.time_mask); 692 } 693 } 694 695 typedef void (*perf_event__swap_op)(union perf_event *event, 696 bool sample_id_all); 697 698 static perf_event__swap_op perf_event__swap_ops[] = { 699 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 700 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap, 701 [PERF_RECORD_COMM] = perf_event__comm_swap, 702 [PERF_RECORD_FORK] = perf_event__task_swap, 703 [PERF_RECORD_EXIT] = perf_event__task_swap, 704 [PERF_RECORD_LOST] = perf_event__all64_swap, 705 [PERF_RECORD_READ] = perf_event__read_swap, 706 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap, 707 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap, 708 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 709 [PERF_RECORD_AUX] = perf_event__aux_swap, 710 [PERF_RECORD_ITRACE_START] = perf_event__itrace_start_swap, 711 [PERF_RECORD_LOST_SAMPLES] = perf_event__all64_swap, 712 [PERF_RECORD_SWITCH] = perf_event__switch_swap, 713 [PERF_RECORD_SWITCH_CPU_WIDE] = perf_event__switch_swap, 714 [PERF_RECORD_NAMESPACES] = perf_event__namespaces_swap, 715 [PERF_RECORD_CGROUP] = perf_event__cgroup_swap, 716 [PERF_RECORD_TEXT_POKE] = perf_event__text_poke_swap, 717 [PERF_RECORD_AUX_OUTPUT_HW_ID] = perf_event__all64_swap, 718 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 719 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 720 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 721 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 722 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap, 723 [PERF_RECORD_AUXTRACE_INFO] = perf_event__auxtrace_info_swap, 724 [PERF_RECORD_AUXTRACE] = perf_event__auxtrace_swap, 725 [PERF_RECORD_AUXTRACE_ERROR] = perf_event__auxtrace_error_swap, 726 [PERF_RECORD_THREAD_MAP] = perf_event__thread_map_swap, 727 [PERF_RECORD_CPU_MAP] = perf_event__cpu_map_swap, 728 [PERF_RECORD_STAT_CONFIG] = perf_event__stat_config_swap, 729 [PERF_RECORD_STAT] = perf_event__stat_swap, 730 [PERF_RECORD_STAT_ROUND] = perf_event__stat_round_swap, 731 [PERF_RECORD_EVENT_UPDATE] = perf_event__event_update_swap, 732 [PERF_RECORD_TIME_CONV] = perf_event__time_conv_swap, 733 [PERF_RECORD_HEADER_MAX] = NULL, 734 }; 735 736 /* 737 * When perf record finishes a pass on every buffers, it records this pseudo 738 * event. 739 * We record the max timestamp t found in the pass n. 740 * Assuming these timestamps are monotonic across cpus, we know that if 741 * a buffer still has events with timestamps below t, they will be all 742 * available and then read in the pass n + 1. 743 * Hence when we start to read the pass n + 2, we can safely flush every 744 * events with timestamps below t. 745 * 746 * ============ PASS n ================= 747 * CPU 0 | CPU 1 748 * | 749 * cnt1 timestamps | cnt2 timestamps 750 * 1 | 2 751 * 2 | 3 752 * - | 4 <--- max recorded 753 * 754 * ============ PASS n + 1 ============== 755 * CPU 0 | CPU 1 756 * | 757 * cnt1 timestamps | cnt2 timestamps 758 * 3 | 5 759 * 4 | 6 760 * 5 | 7 <---- max recorded 761 * 762 * Flush every events below timestamp 4 763 * 764 * ============ PASS n + 2 ============== 765 * CPU 0 | CPU 1 766 * | 767 * cnt1 timestamps | cnt2 timestamps 768 * 6 | 8 769 * 7 | 9 770 * - | 10 771 * 772 * Flush every events below timestamp 7 773 * etc... 774 */ 775 int perf_event__process_finished_round(const struct perf_tool *tool __maybe_unused, 776 union perf_event *event __maybe_unused, 777 struct ordered_events *oe) 778 { 779 if (dump_trace) 780 fprintf(stdout, "\n"); 781 return ordered_events__flush(oe, OE_FLUSH__ROUND); 782 } 783 784 int perf_session__queue_event(struct perf_session *s, union perf_event *event, 785 u64 timestamp, u64 file_offset, const char *file_path) 786 { 787 return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset, file_path); 788 } 789 790 static void callchain__lbr_callstack_printf(struct perf_sample *sample) 791 { 792 struct ip_callchain *callchain = sample->callchain; 793 struct branch_stack *lbr_stack = sample->branch_stack; 794 struct branch_entry *entries = perf_sample__branch_entries(sample); 795 u64 kernel_callchain_nr = callchain->nr; 796 unsigned int i; 797 798 for (i = 0; i < kernel_callchain_nr; i++) { 799 if (callchain->ips[i] == PERF_CONTEXT_USER) 800 break; 801 } 802 803 if ((i != kernel_callchain_nr) && lbr_stack->nr) { 804 u64 total_nr; 805 /* 806 * LBR callstack can only get user call chain, 807 * i is kernel call chain number, 808 * 1 is PERF_CONTEXT_USER. 809 * 810 * The user call chain is stored in LBR registers. 811 * LBR are pair registers. The caller is stored 812 * in "from" register, while the callee is stored 813 * in "to" register. 814 * For example, there is a call stack 815 * "A"->"B"->"C"->"D". 816 * The LBR registers will be recorded like 817 * "C"->"D", "B"->"C", "A"->"B". 818 * So only the first "to" register and all "from" 819 * registers are needed to construct the whole stack. 820 */ 821 total_nr = i + 1 + lbr_stack->nr + 1; 822 kernel_callchain_nr = i + 1; 823 824 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr); 825 826 for (i = 0; i < kernel_callchain_nr; i++) 827 printf("..... %2d: %016" PRIx64 "\n", 828 i, callchain->ips[i]); 829 830 printf("..... %2d: %016" PRIx64 "\n", 831 (int)(kernel_callchain_nr), entries[0].to); 832 for (i = 0; i < lbr_stack->nr; i++) 833 printf("..... %2d: %016" PRIx64 "\n", 834 (int)(i + kernel_callchain_nr + 1), entries[i].from); 835 } 836 } 837 838 static void callchain__printf(struct evsel *evsel, 839 struct perf_sample *sample) 840 { 841 unsigned int i; 842 struct ip_callchain *callchain = sample->callchain; 843 844 if (evsel__has_branch_callstack(evsel)) 845 callchain__lbr_callstack_printf(sample); 846 847 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr); 848 849 for (i = 0; i < callchain->nr; i++) 850 printf("..... %2d: %016" PRIx64 "\n", 851 i, callchain->ips[i]); 852 } 853 854 static void branch_stack__printf(struct perf_sample *sample, 855 struct evsel *evsel) 856 { 857 struct branch_entry *entries = perf_sample__branch_entries(sample); 858 bool callstack = evsel__has_branch_callstack(evsel); 859 u64 *branch_stack_cntr = sample->branch_stack_cntr; 860 uint64_t i; 861 862 if (!callstack) { 863 printf("%s: nr:%" PRIu64 "\n", "... branch stack", sample->branch_stack->nr); 864 } else { 865 /* the reason of adding 1 to nr is because after expanding 866 * branch stack it generates nr + 1 callstack records. e.g., 867 * B()->C() 868 * A()->B() 869 * the final callstack should be: 870 * C() 871 * B() 872 * A() 873 */ 874 printf("%s: nr:%" PRIu64 "\n", "... branch callstack", sample->branch_stack->nr+1); 875 } 876 877 for (i = 0; i < sample->branch_stack->nr; i++) { 878 struct branch_entry *e = &entries[i]; 879 880 if (!callstack) { 881 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x %s %s\n", 882 i, e->from, e->to, 883 (unsigned short)e->flags.cycles, 884 e->flags.mispred ? "M" : " ", 885 e->flags.predicted ? "P" : " ", 886 e->flags.abort ? "A" : " ", 887 e->flags.in_tx ? "T" : " ", 888 (unsigned)e->flags.reserved, 889 get_branch_type(e), 890 e->flags.spec ? branch_spec_desc(e->flags.spec) : ""); 891 } else { 892 if (i == 0) { 893 printf("..... %2"PRIu64": %016" PRIx64 "\n" 894 "..... %2"PRIu64": %016" PRIx64 "\n", 895 i, e->to, i+1, e->from); 896 } else { 897 printf("..... %2"PRIu64": %016" PRIx64 "\n", i+1, e->from); 898 } 899 } 900 } 901 902 if (branch_stack_cntr) { 903 unsigned int br_cntr_width, br_cntr_nr; 904 905 perf_env__find_br_cntr_info(evsel__env(evsel), &br_cntr_nr, &br_cntr_width); 906 printf("... branch stack counters: nr:%" PRIu64 " (counter width: %u max counter nr:%u)\n", 907 sample->branch_stack->nr, br_cntr_width, br_cntr_nr); 908 for (i = 0; i < sample->branch_stack->nr; i++) 909 printf("..... %2"PRIu64": %016" PRIx64 "\n", i, branch_stack_cntr[i]); 910 } 911 } 912 913 static void regs_dump__printf(u64 mask, u64 *regs, const char *arch) 914 { 915 unsigned rid, i = 0; 916 917 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) { 918 u64 val = regs[i++]; 919 920 printf(".... %-5s 0x%016" PRIx64 "\n", 921 perf_reg_name(rid, arch), val); 922 } 923 } 924 925 static const char *regs_abi[] = { 926 [PERF_SAMPLE_REGS_ABI_NONE] = "none", 927 [PERF_SAMPLE_REGS_ABI_32] = "32-bit", 928 [PERF_SAMPLE_REGS_ABI_64] = "64-bit", 929 }; 930 931 static inline const char *regs_dump_abi(struct regs_dump *d) 932 { 933 if (d->abi > PERF_SAMPLE_REGS_ABI_64) 934 return "unknown"; 935 936 return regs_abi[d->abi]; 937 } 938 939 static void regs__printf(const char *type, struct regs_dump *regs, const char *arch) 940 { 941 u64 mask = regs->mask; 942 943 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n", 944 type, 945 mask, 946 regs_dump_abi(regs)); 947 948 regs_dump__printf(mask, regs->regs, arch); 949 } 950 951 static void regs_user__printf(struct perf_sample *sample, const char *arch) 952 { 953 struct regs_dump *user_regs; 954 955 if (!sample->user_regs) 956 return; 957 958 user_regs = perf_sample__user_regs(sample); 959 960 if (user_regs->regs) 961 regs__printf("user", user_regs, arch); 962 } 963 964 static void regs_intr__printf(struct perf_sample *sample, const char *arch) 965 { 966 struct regs_dump *intr_regs; 967 968 if (!sample->intr_regs) 969 return; 970 971 intr_regs = perf_sample__intr_regs(sample); 972 973 if (intr_regs->regs) 974 regs__printf("intr", intr_regs, arch); 975 } 976 977 static void stack_user__printf(struct stack_dump *dump) 978 { 979 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 980 dump->size, dump->offset); 981 } 982 983 static void evlist__print_tstamp(struct evlist *evlist, union perf_event *event, struct perf_sample *sample) 984 { 985 u64 sample_type = __evlist__combined_sample_type(evlist); 986 987 if (event->header.type != PERF_RECORD_SAMPLE && 988 !evlist__sample_id_all(evlist)) { 989 fputs("-1 -1 ", stdout); 990 return; 991 } 992 993 if ((sample_type & PERF_SAMPLE_CPU)) 994 printf("%u ", sample->cpu); 995 996 if (sample_type & PERF_SAMPLE_TIME) 997 printf("%" PRIu64 " ", sample->time); 998 } 999 1000 static void sample_read__printf(struct perf_sample *sample, u64 read_format) 1001 { 1002 printf("... sample_read:\n"); 1003 1004 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1005 printf("...... time enabled %016" PRIx64 "\n", 1006 sample->read.time_enabled); 1007 1008 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1009 printf("...... time running %016" PRIx64 "\n", 1010 sample->read.time_running); 1011 1012 if (read_format & PERF_FORMAT_GROUP) { 1013 struct sample_read_value *value = sample->read.group.values; 1014 1015 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr); 1016 1017 sample_read_group__for_each(value, sample->read.group.nr, read_format) { 1018 printf("..... id %016" PRIx64 1019 ", value %016" PRIx64, 1020 value->id, value->value); 1021 if (read_format & PERF_FORMAT_LOST) 1022 printf(", lost %" PRIu64, value->lost); 1023 printf("\n"); 1024 } 1025 } else { 1026 printf("..... id %016" PRIx64 ", value %016" PRIx64, 1027 sample->read.one.id, sample->read.one.value); 1028 if (read_format & PERF_FORMAT_LOST) 1029 printf(", lost %" PRIu64, sample->read.one.lost); 1030 printf("\n"); 1031 } 1032 } 1033 1034 static void dump_event(struct evlist *evlist, union perf_event *event, 1035 u64 file_offset, struct perf_sample *sample, 1036 const char *file_path) 1037 { 1038 if (!dump_trace) 1039 return; 1040 1041 printf("\n%#" PRIx64 "@%s [%#x]: event: %d\n", 1042 file_offset, file_path, event->header.size, event->header.type); 1043 1044 trace_event(event); 1045 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw) 1046 evlist->trace_event_sample_raw(evlist, event, sample); 1047 1048 if (sample) 1049 evlist__print_tstamp(evlist, event, sample); 1050 1051 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 1052 event->header.size, perf_event__name(event->header.type)); 1053 } 1054 1055 char *get_page_size_name(u64 size, char *str) 1056 { 1057 if (!size || !unit_number__scnprintf(str, PAGE_SIZE_NAME_LEN, size)) 1058 snprintf(str, PAGE_SIZE_NAME_LEN, "%s", "N/A"); 1059 1060 return str; 1061 } 1062 1063 static void dump_sample(struct evsel *evsel, union perf_event *event, 1064 struct perf_sample *sample, const char *arch) 1065 { 1066 u64 sample_type; 1067 char str[PAGE_SIZE_NAME_LEN]; 1068 1069 if (!dump_trace) 1070 return; 1071 1072 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 1073 event->header.misc, sample->pid, sample->tid, sample->ip, 1074 sample->period, sample->addr); 1075 1076 sample_type = evsel->core.attr.sample_type; 1077 1078 if (evsel__has_callchain(evsel)) 1079 callchain__printf(evsel, sample); 1080 1081 if (evsel__has_br_stack(evsel)) 1082 branch_stack__printf(sample, evsel); 1083 1084 if (sample_type & PERF_SAMPLE_REGS_USER) 1085 regs_user__printf(sample, arch); 1086 1087 if (sample_type & PERF_SAMPLE_REGS_INTR) 1088 regs_intr__printf(sample, arch); 1089 1090 if (sample_type & PERF_SAMPLE_STACK_USER) 1091 stack_user__printf(&sample->user_stack); 1092 1093 if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) { 1094 printf("... weight: %" PRIu64 "", sample->weight); 1095 if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) { 1096 printf(",0x%"PRIx16"", sample->ins_lat); 1097 printf(",0x%"PRIx16"", sample->p_stage_cyc); 1098 } 1099 printf("\n"); 1100 } 1101 1102 if (sample_type & PERF_SAMPLE_DATA_SRC) 1103 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src); 1104 1105 if (sample_type & PERF_SAMPLE_PHYS_ADDR) 1106 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr); 1107 1108 if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE) 1109 printf(" .. data page size: %s\n", get_page_size_name(sample->data_page_size, str)); 1110 1111 if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE) 1112 printf(" .. code page size: %s\n", get_page_size_name(sample->code_page_size, str)); 1113 1114 if (sample_type & PERF_SAMPLE_TRANSACTION) 1115 printf("... transaction: %" PRIx64 "\n", sample->transaction); 1116 1117 if (sample_type & PERF_SAMPLE_READ) 1118 sample_read__printf(sample, evsel->core.attr.read_format); 1119 } 1120 1121 static void dump_read(struct evsel *evsel, union perf_event *event) 1122 { 1123 struct perf_record_read *read_event = &event->read; 1124 u64 read_format; 1125 1126 if (!dump_trace) 1127 return; 1128 1129 printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid, 1130 evsel__name(evsel), event->read.value); 1131 1132 if (!evsel) 1133 return; 1134 1135 read_format = evsel->core.attr.read_format; 1136 1137 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1138 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled); 1139 1140 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1141 printf("... time running : %" PRI_lu64 "\n", read_event->time_running); 1142 1143 if (read_format & PERF_FORMAT_ID) 1144 printf("... id : %" PRI_lu64 "\n", read_event->id); 1145 1146 if (read_format & PERF_FORMAT_LOST) 1147 printf("... lost : %" PRI_lu64 "\n", read_event->lost); 1148 } 1149 1150 static struct machine *machines__find_for_cpumode(struct machines *machines, 1151 union perf_event *event, 1152 struct perf_sample *sample) 1153 { 1154 if (perf_guest && 1155 ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 1156 (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) { 1157 u32 pid; 1158 1159 if (sample->machine_pid) 1160 pid = sample->machine_pid; 1161 else if (event->header.type == PERF_RECORD_MMAP 1162 || event->header.type == PERF_RECORD_MMAP2) 1163 pid = event->mmap.pid; 1164 else 1165 pid = sample->pid; 1166 1167 /* 1168 * Guest code machine is created as needed and does not use 1169 * DEFAULT_GUEST_KERNEL_ID. 1170 */ 1171 if (symbol_conf.guest_code) 1172 return machines__findnew(machines, pid); 1173 1174 return machines__find_guest(machines, pid); 1175 } 1176 1177 return &machines->host; 1178 } 1179 1180 static int deliver_sample_value(struct evlist *evlist, 1181 const struct perf_tool *tool, 1182 union perf_event *event, 1183 struct perf_sample *sample, 1184 struct sample_read_value *v, 1185 struct machine *machine, 1186 bool per_thread) 1187 { 1188 struct perf_sample_id *sid = evlist__id2sid(evlist, v->id); 1189 struct evsel *evsel; 1190 u64 *storage = NULL; 1191 1192 if (sid) { 1193 storage = perf_sample_id__get_period_storage(sid, sample->tid, per_thread); 1194 } 1195 1196 if (storage) { 1197 sample->id = v->id; 1198 sample->period = v->value - *storage; 1199 *storage = v->value; 1200 } 1201 1202 if (!storage || sid->evsel == NULL) { 1203 ++evlist->stats.nr_unknown_id; 1204 return 0; 1205 } 1206 1207 /* 1208 * There's no reason to deliver sample 1209 * for zero period, bail out. 1210 */ 1211 if (!sample->period) 1212 return 0; 1213 1214 evsel = container_of(sid->evsel, struct evsel, core); 1215 return tool->sample(tool, event, sample, evsel, machine); 1216 } 1217 1218 static int deliver_sample_group(struct evlist *evlist, 1219 const struct perf_tool *tool, 1220 union perf_event *event, 1221 struct perf_sample *sample, 1222 struct machine *machine, 1223 u64 read_format, 1224 bool per_thread) 1225 { 1226 int ret = -EINVAL; 1227 struct sample_read_value *v = sample->read.group.values; 1228 1229 if (tool->dont_split_sample_group) 1230 return deliver_sample_value(evlist, tool, event, sample, v, machine, 1231 per_thread); 1232 1233 sample_read_group__for_each(v, sample->read.group.nr, read_format) { 1234 ret = deliver_sample_value(evlist, tool, event, sample, v, 1235 machine, per_thread); 1236 if (ret) 1237 break; 1238 } 1239 1240 return ret; 1241 } 1242 1243 static int evlist__deliver_sample(struct evlist *evlist, const struct perf_tool *tool, 1244 union perf_event *event, struct perf_sample *sample, 1245 struct evsel *evsel, struct machine *machine) 1246 { 1247 /* We know evsel != NULL. */ 1248 u64 sample_type = evsel->core.attr.sample_type; 1249 u64 read_format = evsel->core.attr.read_format; 1250 bool per_thread = perf_evsel__attr_has_per_thread_sample_period(&evsel->core); 1251 1252 /* Standard sample delivery. */ 1253 if (!(sample_type & PERF_SAMPLE_READ)) 1254 return tool->sample(tool, event, sample, evsel, machine); 1255 1256 /* For PERF_SAMPLE_READ we have either single or group mode. */ 1257 if (read_format & PERF_FORMAT_GROUP) 1258 return deliver_sample_group(evlist, tool, event, sample, 1259 machine, read_format, per_thread); 1260 else 1261 return deliver_sample_value(evlist, tool, event, sample, 1262 &sample->read.one, machine, 1263 per_thread); 1264 } 1265 1266 static int machines__deliver_event(struct machines *machines, 1267 struct evlist *evlist, 1268 union perf_event *event, 1269 struct perf_sample *sample, 1270 const struct perf_tool *tool, u64 file_offset, 1271 const char *file_path) 1272 { 1273 struct evsel *evsel; 1274 struct machine *machine; 1275 1276 dump_event(evlist, event, file_offset, sample, file_path); 1277 1278 evsel = evlist__id2evsel(evlist, sample->id); 1279 1280 machine = machines__find_for_cpumode(machines, event, sample); 1281 1282 switch (event->header.type) { 1283 case PERF_RECORD_SAMPLE: 1284 if (evsel == NULL) { 1285 ++evlist->stats.nr_unknown_id; 1286 return 0; 1287 } 1288 if (machine == NULL) { 1289 ++evlist->stats.nr_unprocessable_samples; 1290 dump_sample(evsel, event, sample, perf_env__arch(NULL)); 1291 return 0; 1292 } 1293 dump_sample(evsel, event, sample, perf_env__arch(machine->env)); 1294 return evlist__deliver_sample(evlist, tool, event, sample, evsel, machine); 1295 case PERF_RECORD_MMAP: 1296 return tool->mmap(tool, event, sample, machine); 1297 case PERF_RECORD_MMAP2: 1298 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT) 1299 ++evlist->stats.nr_proc_map_timeout; 1300 return tool->mmap2(tool, event, sample, machine); 1301 case PERF_RECORD_COMM: 1302 return tool->comm(tool, event, sample, machine); 1303 case PERF_RECORD_NAMESPACES: 1304 return tool->namespaces(tool, event, sample, machine); 1305 case PERF_RECORD_CGROUP: 1306 return tool->cgroup(tool, event, sample, machine); 1307 case PERF_RECORD_FORK: 1308 return tool->fork(tool, event, sample, machine); 1309 case PERF_RECORD_EXIT: 1310 return tool->exit(tool, event, sample, machine); 1311 case PERF_RECORD_LOST: 1312 if (tool->lost == perf_event__process_lost) 1313 evlist->stats.total_lost += event->lost.lost; 1314 return tool->lost(tool, event, sample, machine); 1315 case PERF_RECORD_LOST_SAMPLES: 1316 if (event->header.misc & PERF_RECORD_MISC_LOST_SAMPLES_BPF) 1317 evlist->stats.total_dropped_samples += event->lost_samples.lost; 1318 else if (tool->lost_samples == perf_event__process_lost_samples) 1319 evlist->stats.total_lost_samples += event->lost_samples.lost; 1320 return tool->lost_samples(tool, event, sample, machine); 1321 case PERF_RECORD_READ: 1322 dump_read(evsel, event); 1323 return tool->read(tool, event, sample, evsel, machine); 1324 case PERF_RECORD_THROTTLE: 1325 return tool->throttle(tool, event, sample, machine); 1326 case PERF_RECORD_UNTHROTTLE: 1327 return tool->unthrottle(tool, event, sample, machine); 1328 case PERF_RECORD_AUX: 1329 if (tool->aux == perf_event__process_aux) { 1330 if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) 1331 evlist->stats.total_aux_lost += 1; 1332 if (event->aux.flags & PERF_AUX_FLAG_PARTIAL) 1333 evlist->stats.total_aux_partial += 1; 1334 if (event->aux.flags & PERF_AUX_FLAG_COLLISION) 1335 evlist->stats.total_aux_collision += 1; 1336 } 1337 return tool->aux(tool, event, sample, machine); 1338 case PERF_RECORD_ITRACE_START: 1339 return tool->itrace_start(tool, event, sample, machine); 1340 case PERF_RECORD_SWITCH: 1341 case PERF_RECORD_SWITCH_CPU_WIDE: 1342 return tool->context_switch(tool, event, sample, machine); 1343 case PERF_RECORD_KSYMBOL: 1344 return tool->ksymbol(tool, event, sample, machine); 1345 case PERF_RECORD_BPF_EVENT: 1346 return tool->bpf(tool, event, sample, machine); 1347 case PERF_RECORD_TEXT_POKE: 1348 return tool->text_poke(tool, event, sample, machine); 1349 case PERF_RECORD_AUX_OUTPUT_HW_ID: 1350 return tool->aux_output_hw_id(tool, event, sample, machine); 1351 default: 1352 ++evlist->stats.nr_unknown_events; 1353 return -1; 1354 } 1355 } 1356 1357 static int perf_session__deliver_event(struct perf_session *session, 1358 union perf_event *event, 1359 const struct perf_tool *tool, 1360 u64 file_offset, 1361 const char *file_path) 1362 { 1363 struct perf_sample sample; 1364 int ret; 1365 1366 perf_sample__init(&sample, /*all=*/false); 1367 ret = evlist__parse_sample(session->evlist, event, &sample); 1368 if (ret) { 1369 pr_err("Can't parse sample, err = %d\n", ret); 1370 goto out; 1371 } 1372 1373 ret = auxtrace__process_event(session, event, &sample, tool); 1374 if (ret < 0) 1375 goto out; 1376 if (ret > 0) { 1377 ret = 0; 1378 goto out; 1379 } 1380 1381 ret = machines__deliver_event(&session->machines, session->evlist, 1382 event, &sample, tool, file_offset, file_path); 1383 1384 if (dump_trace && sample.aux_sample.size) 1385 auxtrace__dump_auxtrace_sample(session, &sample); 1386 out: 1387 perf_sample__exit(&sample); 1388 return ret; 1389 } 1390 1391 static s64 perf_session__process_user_event(struct perf_session *session, 1392 union perf_event *event, 1393 u64 file_offset, 1394 const char *file_path) 1395 { 1396 struct ordered_events *oe = &session->ordered_events; 1397 const struct perf_tool *tool = session->tool; 1398 struct perf_sample sample; 1399 int fd = perf_data__fd(session->data); 1400 int err; 1401 1402 perf_sample__init(&sample, /*all=*/true); 1403 if (event->header.type != PERF_RECORD_COMPRESSED || perf_tool__compressed_is_stub(tool)) 1404 dump_event(session->evlist, event, file_offset, &sample, file_path); 1405 1406 /* These events are processed right away */ 1407 switch (event->header.type) { 1408 case PERF_RECORD_HEADER_ATTR: 1409 err = tool->attr(tool, event, &session->evlist); 1410 if (err == 0) { 1411 perf_session__set_id_hdr_size(session); 1412 perf_session__set_comm_exec(session); 1413 } 1414 break; 1415 case PERF_RECORD_EVENT_UPDATE: 1416 err = tool->event_update(tool, event, &session->evlist); 1417 break; 1418 case PERF_RECORD_HEADER_EVENT_TYPE: 1419 /* 1420 * Deprecated, but we need to handle it for sake 1421 * of old data files create in pipe mode. 1422 */ 1423 err = 0; 1424 break; 1425 case PERF_RECORD_HEADER_TRACING_DATA: 1426 /* 1427 * Setup for reading amidst mmap, but only when we 1428 * are in 'file' mode. The 'pipe' fd is in proper 1429 * place already. 1430 */ 1431 if (!perf_data__is_pipe(session->data)) 1432 lseek(fd, file_offset, SEEK_SET); 1433 err = tool->tracing_data(session, event); 1434 break; 1435 case PERF_RECORD_HEADER_BUILD_ID: 1436 err = tool->build_id(session, event); 1437 break; 1438 case PERF_RECORD_FINISHED_ROUND: 1439 err = tool->finished_round(tool, event, oe); 1440 break; 1441 case PERF_RECORD_ID_INDEX: 1442 err = tool->id_index(session, event); 1443 break; 1444 case PERF_RECORD_AUXTRACE_INFO: 1445 err = tool->auxtrace_info(session, event); 1446 break; 1447 case PERF_RECORD_AUXTRACE: 1448 /* 1449 * Setup for reading amidst mmap, but only when we 1450 * are in 'file' mode. The 'pipe' fd is in proper 1451 * place already. 1452 */ 1453 if (!perf_data__is_pipe(session->data)) 1454 lseek(fd, file_offset + event->header.size, SEEK_SET); 1455 err = tool->auxtrace(session, event); 1456 break; 1457 case PERF_RECORD_AUXTRACE_ERROR: 1458 perf_session__auxtrace_error_inc(session, event); 1459 err = tool->auxtrace_error(session, event); 1460 break; 1461 case PERF_RECORD_THREAD_MAP: 1462 err = tool->thread_map(session, event); 1463 break; 1464 case PERF_RECORD_CPU_MAP: 1465 err = tool->cpu_map(session, event); 1466 break; 1467 case PERF_RECORD_STAT_CONFIG: 1468 err = tool->stat_config(session, event); 1469 break; 1470 case PERF_RECORD_STAT: 1471 err = tool->stat(session, event); 1472 break; 1473 case PERF_RECORD_STAT_ROUND: 1474 err = tool->stat_round(session, event); 1475 break; 1476 case PERF_RECORD_TIME_CONV: 1477 session->time_conv = event->time_conv; 1478 err = tool->time_conv(session, event); 1479 break; 1480 case PERF_RECORD_HEADER_FEATURE: 1481 err = tool->feature(session, event); 1482 break; 1483 case PERF_RECORD_COMPRESSED: 1484 err = tool->compressed(session, event, file_offset, file_path); 1485 if (err) 1486 dump_event(session->evlist, event, file_offset, &sample, file_path); 1487 break; 1488 case PERF_RECORD_FINISHED_INIT: 1489 err = tool->finished_init(session, event); 1490 break; 1491 default: 1492 err = -EINVAL; 1493 break; 1494 } 1495 perf_sample__exit(&sample); 1496 return err; 1497 } 1498 1499 int perf_session__deliver_synth_event(struct perf_session *session, 1500 union perf_event *event, 1501 struct perf_sample *sample) 1502 { 1503 struct evlist *evlist = session->evlist; 1504 const struct perf_tool *tool = session->tool; 1505 1506 events_stats__inc(&evlist->stats, event->header.type); 1507 1508 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1509 return perf_session__process_user_event(session, event, 0, NULL); 1510 1511 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0, NULL); 1512 } 1513 1514 int perf_session__deliver_synth_attr_event(struct perf_session *session, 1515 const struct perf_event_attr *attr, 1516 u64 id) 1517 { 1518 union { 1519 struct { 1520 struct perf_record_header_attr attr; 1521 u64 ids[1]; 1522 } attr_id; 1523 union perf_event ev; 1524 } ev = { 1525 .attr_id.attr.header.type = PERF_RECORD_HEADER_ATTR, 1526 .attr_id.attr.header.size = sizeof(ev.attr_id), 1527 .attr_id.ids[0] = id, 1528 }; 1529 1530 if (attr->size != sizeof(ev.attr_id.attr.attr)) { 1531 pr_debug("Unexpected perf_event_attr size\n"); 1532 return -EINVAL; 1533 } 1534 ev.attr_id.attr.attr = *attr; 1535 return perf_session__deliver_synth_event(session, &ev.ev, NULL); 1536 } 1537 1538 static void event_swap(union perf_event *event, bool sample_id_all) 1539 { 1540 perf_event__swap_op swap; 1541 1542 swap = perf_event__swap_ops[event->header.type]; 1543 if (swap) 1544 swap(event, sample_id_all); 1545 } 1546 1547 int perf_session__peek_event(struct perf_session *session, off_t file_offset, 1548 void *buf, size_t buf_sz, 1549 union perf_event **event_ptr, 1550 struct perf_sample *sample) 1551 { 1552 union perf_event *event; 1553 size_t hdr_sz, rest; 1554 int fd; 1555 1556 if (session->one_mmap && !session->header.needs_swap) { 1557 event = file_offset - session->one_mmap_offset + 1558 session->one_mmap_addr; 1559 goto out_parse_sample; 1560 } 1561 1562 if (perf_data__is_pipe(session->data)) 1563 return -1; 1564 1565 fd = perf_data__fd(session->data); 1566 hdr_sz = sizeof(struct perf_event_header); 1567 1568 if (buf_sz < hdr_sz) 1569 return -1; 1570 1571 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 || 1572 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz) 1573 return -1; 1574 1575 event = (union perf_event *)buf; 1576 1577 if (session->header.needs_swap) 1578 perf_event_header__bswap(&event->header); 1579 1580 if (event->header.size < hdr_sz || event->header.size > buf_sz) 1581 return -1; 1582 1583 buf += hdr_sz; 1584 rest = event->header.size - hdr_sz; 1585 1586 if (readn(fd, buf, rest) != (ssize_t)rest) 1587 return -1; 1588 1589 if (session->header.needs_swap) 1590 event_swap(event, evlist__sample_id_all(session->evlist)); 1591 1592 out_parse_sample: 1593 1594 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START && 1595 evlist__parse_sample(session->evlist, event, sample)) 1596 return -1; 1597 1598 *event_ptr = event; 1599 1600 return 0; 1601 } 1602 1603 int perf_session__peek_events(struct perf_session *session, u64 offset, 1604 u64 size, peek_events_cb_t cb, void *data) 1605 { 1606 u64 max_offset = offset + size; 1607 char buf[PERF_SAMPLE_MAX_SIZE]; 1608 union perf_event *event; 1609 int err; 1610 1611 do { 1612 err = perf_session__peek_event(session, offset, buf, 1613 PERF_SAMPLE_MAX_SIZE, &event, 1614 NULL); 1615 if (err) 1616 return err; 1617 1618 err = cb(session, event, offset, data); 1619 if (err) 1620 return err; 1621 1622 offset += event->header.size; 1623 if (event->header.type == PERF_RECORD_AUXTRACE) 1624 offset += event->auxtrace.size; 1625 1626 } while (offset < max_offset); 1627 1628 return err; 1629 } 1630 1631 static s64 perf_session__process_event(struct perf_session *session, 1632 union perf_event *event, u64 file_offset, 1633 const char *file_path) 1634 { 1635 struct evlist *evlist = session->evlist; 1636 const struct perf_tool *tool = session->tool; 1637 int ret; 1638 1639 if (session->header.needs_swap) 1640 event_swap(event, evlist__sample_id_all(evlist)); 1641 1642 if (event->header.type >= PERF_RECORD_HEADER_MAX) 1643 return -EINVAL; 1644 1645 events_stats__inc(&evlist->stats, event->header.type); 1646 1647 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1648 return perf_session__process_user_event(session, event, file_offset, file_path); 1649 1650 if (tool->ordered_events) { 1651 u64 timestamp = -1ULL; 1652 1653 ret = evlist__parse_sample_timestamp(evlist, event, ×tamp); 1654 if (ret && ret != -1) 1655 return ret; 1656 1657 ret = perf_session__queue_event(session, event, timestamp, file_offset, file_path); 1658 if (ret != -ETIME) 1659 return ret; 1660 } 1661 1662 return perf_session__deliver_event(session, event, tool, file_offset, file_path); 1663 } 1664 1665 void perf_event_header__bswap(struct perf_event_header *hdr) 1666 { 1667 hdr->type = bswap_32(hdr->type); 1668 hdr->misc = bswap_16(hdr->misc); 1669 hdr->size = bswap_16(hdr->size); 1670 } 1671 1672 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1673 { 1674 return machine__findnew_thread(&session->machines.host, -1, pid); 1675 } 1676 1677 int perf_session__register_idle_thread(struct perf_session *session) 1678 { 1679 struct thread *thread = machine__idle_thread(&session->machines.host); 1680 1681 /* machine__idle_thread() got the thread, so put it */ 1682 thread__put(thread); 1683 return thread ? 0 : -1; 1684 } 1685 1686 static void 1687 perf_session__warn_order(const struct perf_session *session) 1688 { 1689 const struct ordered_events *oe = &session->ordered_events; 1690 struct evsel *evsel; 1691 bool should_warn = true; 1692 1693 evlist__for_each_entry(session->evlist, evsel) { 1694 if (evsel->core.attr.write_backward) 1695 should_warn = false; 1696 } 1697 1698 if (!should_warn) 1699 return; 1700 if (oe->nr_unordered_events != 0) 1701 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events); 1702 } 1703 1704 static void perf_session__warn_about_errors(const struct perf_session *session) 1705 { 1706 const struct events_stats *stats = &session->evlist->stats; 1707 1708 if (session->tool->lost == perf_event__process_lost && 1709 stats->nr_events[PERF_RECORD_LOST] != 0) { 1710 ui__warning("Processed %d events and lost %d chunks!\n\n" 1711 "Check IO/CPU overload!\n\n", 1712 stats->nr_events[0], 1713 stats->nr_events[PERF_RECORD_LOST]); 1714 } 1715 1716 if (session->tool->lost_samples == perf_event__process_lost_samples) { 1717 double drop_rate; 1718 1719 drop_rate = (double)stats->total_lost_samples / 1720 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples); 1721 if (drop_rate > 0.05) { 1722 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n", 1723 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples, 1724 drop_rate * 100.0); 1725 } 1726 } 1727 1728 if (session->tool->aux == perf_event__process_aux && 1729 stats->total_aux_lost != 0) { 1730 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n", 1731 stats->total_aux_lost, 1732 stats->nr_events[PERF_RECORD_AUX]); 1733 } 1734 1735 if (session->tool->aux == perf_event__process_aux && 1736 stats->total_aux_partial != 0) { 1737 bool vmm_exclusive = false; 1738 1739 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive", 1740 &vmm_exclusive); 1741 1742 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n" 1743 "Are you running a KVM guest in the background?%s\n\n", 1744 stats->total_aux_partial, 1745 stats->nr_events[PERF_RECORD_AUX], 1746 vmm_exclusive ? 1747 "\nReloading kvm_intel module with vmm_exclusive=0\n" 1748 "will reduce the gaps to only guest's timeslices." : 1749 ""); 1750 } 1751 1752 if (session->tool->aux == perf_event__process_aux && 1753 stats->total_aux_collision != 0) { 1754 ui__warning("AUX data detected collision %" PRIu64 " times out of %u!\n\n", 1755 stats->total_aux_collision, 1756 stats->nr_events[PERF_RECORD_AUX]); 1757 } 1758 1759 if (stats->nr_unknown_events != 0) { 1760 ui__warning("Found %u unknown events!\n\n" 1761 "Is this an older tool processing a perf.data " 1762 "file generated by a more recent tool?\n\n" 1763 "If that is not the case, consider " 1764 "reporting to linux-kernel@vger.kernel.org.\n\n", 1765 stats->nr_unknown_events); 1766 } 1767 1768 if (stats->nr_unknown_id != 0) { 1769 ui__warning("%u samples with id not present in the header\n", 1770 stats->nr_unknown_id); 1771 } 1772 1773 if (stats->nr_invalid_chains != 0) { 1774 ui__warning("Found invalid callchains!\n\n" 1775 "%u out of %u events were discarded for this reason.\n\n" 1776 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 1777 stats->nr_invalid_chains, 1778 stats->nr_events[PERF_RECORD_SAMPLE]); 1779 } 1780 1781 if (stats->nr_unprocessable_samples != 0) { 1782 ui__warning("%u unprocessable samples recorded.\n" 1783 "Do you have a KVM guest running and not using 'perf kvm'?\n", 1784 stats->nr_unprocessable_samples); 1785 } 1786 1787 perf_session__warn_order(session); 1788 1789 events_stats__auxtrace_error_warn(stats); 1790 1791 if (stats->nr_proc_map_timeout != 0) { 1792 ui__warning("%d map information files for pre-existing threads were\n" 1793 "not processed, if there are samples for addresses they\n" 1794 "will not be resolved, you may find out which are these\n" 1795 "threads by running with -v and redirecting the output\n" 1796 "to a file.\n" 1797 "The time limit to process proc map is too short?\n" 1798 "Increase it by --proc-map-timeout\n", 1799 stats->nr_proc_map_timeout); 1800 } 1801 } 1802 1803 static int perf_session__flush_thread_stack(struct thread *thread, 1804 void *p __maybe_unused) 1805 { 1806 return thread_stack__flush(thread); 1807 } 1808 1809 static int perf_session__flush_thread_stacks(struct perf_session *session) 1810 { 1811 return machines__for_each_thread(&session->machines, 1812 perf_session__flush_thread_stack, 1813 NULL); 1814 } 1815 1816 volatile sig_atomic_t session_done; 1817 1818 static int __perf_session__process_decomp_events(struct perf_session *session); 1819 1820 static int __perf_session__process_pipe_events(struct perf_session *session) 1821 { 1822 struct ordered_events *oe = &session->ordered_events; 1823 const struct perf_tool *tool = session->tool; 1824 struct ui_progress prog; 1825 union perf_event *event; 1826 uint32_t size, cur_size = 0; 1827 void *buf = NULL; 1828 s64 skip = 0; 1829 u64 head; 1830 ssize_t err; 1831 void *p; 1832 bool update_prog = false; 1833 1834 /* 1835 * If it's from a file saving pipe data (by redirection), it would have 1836 * a file name other than "-". Then we can get the total size and show 1837 * the progress. 1838 */ 1839 if (strcmp(session->data->path, "-") && session->data->file.size) { 1840 ui_progress__init_size(&prog, session->data->file.size, 1841 "Processing events..."); 1842 update_prog = true; 1843 } 1844 1845 head = 0; 1846 cur_size = sizeof(union perf_event); 1847 1848 buf = malloc(cur_size); 1849 if (!buf) 1850 return -errno; 1851 ordered_events__set_copy_on_queue(oe, true); 1852 more: 1853 event = buf; 1854 err = perf_data__read(session->data, event, 1855 sizeof(struct perf_event_header)); 1856 if (err <= 0) { 1857 if (err == 0) 1858 goto done; 1859 1860 pr_err("failed to read event header\n"); 1861 goto out_err; 1862 } 1863 1864 if (session->header.needs_swap) 1865 perf_event_header__bswap(&event->header); 1866 1867 size = event->header.size; 1868 if (size < sizeof(struct perf_event_header)) { 1869 pr_err("bad event header size\n"); 1870 goto out_err; 1871 } 1872 1873 if (size > cur_size) { 1874 void *new = realloc(buf, size); 1875 if (!new) { 1876 pr_err("failed to allocate memory to read event\n"); 1877 goto out_err; 1878 } 1879 buf = new; 1880 cur_size = size; 1881 event = buf; 1882 } 1883 p = event; 1884 p += sizeof(struct perf_event_header); 1885 1886 if (size - sizeof(struct perf_event_header)) { 1887 err = perf_data__read(session->data, p, 1888 size - sizeof(struct perf_event_header)); 1889 if (err <= 0) { 1890 if (err == 0) { 1891 pr_err("unexpected end of event stream\n"); 1892 goto done; 1893 } 1894 1895 pr_err("failed to read event data\n"); 1896 goto out_err; 1897 } 1898 } 1899 1900 if ((skip = perf_session__process_event(session, event, head, "pipe")) < 0) { 1901 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 1902 head, event->header.size, event->header.type); 1903 err = -EINVAL; 1904 goto out_err; 1905 } 1906 1907 head += size; 1908 1909 if (skip > 0) 1910 head += skip; 1911 1912 err = __perf_session__process_decomp_events(session); 1913 if (err) 1914 goto out_err; 1915 1916 if (update_prog) 1917 ui_progress__update(&prog, size); 1918 1919 if (!session_done()) 1920 goto more; 1921 done: 1922 /* do the final flush for ordered samples */ 1923 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 1924 if (err) 1925 goto out_err; 1926 err = auxtrace__flush_events(session, tool); 1927 if (err) 1928 goto out_err; 1929 err = perf_session__flush_thread_stacks(session); 1930 out_err: 1931 free(buf); 1932 if (update_prog) 1933 ui_progress__finish(); 1934 if (!tool->no_warn) 1935 perf_session__warn_about_errors(session); 1936 ordered_events__free(&session->ordered_events); 1937 auxtrace__free_events(session); 1938 return err; 1939 } 1940 1941 static union perf_event * 1942 prefetch_event(char *buf, u64 head, size_t mmap_size, 1943 bool needs_swap, union perf_event *error) 1944 { 1945 union perf_event *event; 1946 u16 event_size; 1947 1948 /* 1949 * Ensure we have enough space remaining to read 1950 * the size of the event in the headers. 1951 */ 1952 if (head + sizeof(event->header) > mmap_size) 1953 return NULL; 1954 1955 event = (union perf_event *)(buf + head); 1956 if (needs_swap) 1957 perf_event_header__bswap(&event->header); 1958 1959 event_size = event->header.size; 1960 if (head + event_size <= mmap_size) 1961 return event; 1962 1963 /* We're not fetching the event so swap back again */ 1964 if (needs_swap) 1965 perf_event_header__bswap(&event->header); 1966 1967 /* Check if the event fits into the next mmapped buf. */ 1968 if (event_size <= mmap_size - head % page_size) { 1969 /* Remap buf and fetch again. */ 1970 return NULL; 1971 } 1972 1973 /* Invalid input. Event size should never exceed mmap_size. */ 1974 pr_debug("%s: head=%#" PRIx64 " event->header.size=%#x, mmap_size=%#zx:" 1975 " fuzzed or compressed perf.data?\n", __func__, head, event_size, mmap_size); 1976 1977 return error; 1978 } 1979 1980 static union perf_event * 1981 fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 1982 { 1983 return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL)); 1984 } 1985 1986 static union perf_event * 1987 fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 1988 { 1989 return prefetch_event(buf, head, mmap_size, needs_swap, NULL); 1990 } 1991 1992 static int __perf_session__process_decomp_events(struct perf_session *session) 1993 { 1994 s64 skip; 1995 u64 size; 1996 struct decomp *decomp = session->active_decomp->decomp_last; 1997 1998 if (!decomp) 1999 return 0; 2000 2001 while (decomp->head < decomp->size && !session_done()) { 2002 union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data, 2003 session->header.needs_swap); 2004 2005 if (!event) 2006 break; 2007 2008 size = event->header.size; 2009 2010 if (size < sizeof(struct perf_event_header) || 2011 (skip = perf_session__process_event(session, event, decomp->file_pos, 2012 decomp->file_path)) < 0) { 2013 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 2014 decomp->file_pos + decomp->head, event->header.size, event->header.type); 2015 return -EINVAL; 2016 } 2017 2018 if (skip) 2019 size += skip; 2020 2021 decomp->head += size; 2022 } 2023 2024 return 0; 2025 } 2026 2027 /* 2028 * On 64bit we can mmap the data file in one go. No need for tiny mmap 2029 * slices. On 32bit we use 32MB. 2030 */ 2031 #if BITS_PER_LONG == 64 2032 #define MMAP_SIZE ULLONG_MAX 2033 #define NUM_MMAPS 1 2034 #else 2035 #define MMAP_SIZE (32 * 1024 * 1024ULL) 2036 #define NUM_MMAPS 128 2037 #endif 2038 2039 struct reader; 2040 2041 typedef s64 (*reader_cb_t)(struct perf_session *session, 2042 union perf_event *event, 2043 u64 file_offset, 2044 const char *file_path); 2045 2046 struct reader { 2047 int fd; 2048 const char *path; 2049 u64 data_size; 2050 u64 data_offset; 2051 reader_cb_t process; 2052 bool in_place_update; 2053 char *mmaps[NUM_MMAPS]; 2054 size_t mmap_size; 2055 int mmap_idx; 2056 char *mmap_cur; 2057 u64 file_pos; 2058 u64 file_offset; 2059 u64 head; 2060 u64 size; 2061 bool done; 2062 struct zstd_data zstd_data; 2063 struct decomp_data decomp_data; 2064 }; 2065 2066 static int 2067 reader__init(struct reader *rd, bool *one_mmap) 2068 { 2069 u64 data_size = rd->data_size; 2070 char **mmaps = rd->mmaps; 2071 2072 rd->head = rd->data_offset; 2073 data_size += rd->data_offset; 2074 2075 rd->mmap_size = MMAP_SIZE; 2076 if (rd->mmap_size > data_size) { 2077 rd->mmap_size = data_size; 2078 if (one_mmap) 2079 *one_mmap = true; 2080 } 2081 2082 memset(mmaps, 0, sizeof(rd->mmaps)); 2083 2084 if (zstd_init(&rd->zstd_data, 0)) 2085 return -1; 2086 rd->decomp_data.zstd_decomp = &rd->zstd_data; 2087 2088 return 0; 2089 } 2090 2091 static void 2092 reader__release_decomp(struct reader *rd) 2093 { 2094 perf_decomp__release_events(rd->decomp_data.decomp); 2095 zstd_fini(&rd->zstd_data); 2096 } 2097 2098 static int 2099 reader__mmap(struct reader *rd, struct perf_session *session) 2100 { 2101 int mmap_prot, mmap_flags; 2102 char *buf, **mmaps = rd->mmaps; 2103 u64 page_offset; 2104 2105 mmap_prot = PROT_READ; 2106 mmap_flags = MAP_SHARED; 2107 2108 if (rd->in_place_update) { 2109 mmap_prot |= PROT_WRITE; 2110 } else if (session->header.needs_swap) { 2111 mmap_prot |= PROT_WRITE; 2112 mmap_flags = MAP_PRIVATE; 2113 } 2114 2115 if (mmaps[rd->mmap_idx]) { 2116 munmap(mmaps[rd->mmap_idx], rd->mmap_size); 2117 mmaps[rd->mmap_idx] = NULL; 2118 } 2119 2120 page_offset = page_size * (rd->head / page_size); 2121 rd->file_offset += page_offset; 2122 rd->head -= page_offset; 2123 2124 buf = mmap(NULL, rd->mmap_size, mmap_prot, mmap_flags, rd->fd, 2125 rd->file_offset); 2126 if (buf == MAP_FAILED) { 2127 pr_err("failed to mmap file\n"); 2128 return -errno; 2129 } 2130 mmaps[rd->mmap_idx] = rd->mmap_cur = buf; 2131 rd->mmap_idx = (rd->mmap_idx + 1) & (ARRAY_SIZE(rd->mmaps) - 1); 2132 rd->file_pos = rd->file_offset + rd->head; 2133 if (session->one_mmap) { 2134 session->one_mmap_addr = buf; 2135 session->one_mmap_offset = rd->file_offset; 2136 } 2137 2138 return 0; 2139 } 2140 2141 enum { 2142 READER_OK, 2143 READER_NODATA, 2144 }; 2145 2146 static int 2147 reader__read_event(struct reader *rd, struct perf_session *session, 2148 struct ui_progress *prog) 2149 { 2150 u64 size; 2151 int err = READER_OK; 2152 union perf_event *event; 2153 s64 skip; 2154 2155 event = fetch_mmaped_event(rd->head, rd->mmap_size, rd->mmap_cur, 2156 session->header.needs_swap); 2157 if (IS_ERR(event)) 2158 return PTR_ERR(event); 2159 2160 if (!event) 2161 return READER_NODATA; 2162 2163 size = event->header.size; 2164 2165 skip = -EINVAL; 2166 2167 if (size < sizeof(struct perf_event_header) || 2168 (skip = rd->process(session, event, rd->file_pos, rd->path)) < 0) { 2169 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n", 2170 rd->file_offset + rd->head, event->header.size, 2171 event->header.type, strerror(-skip)); 2172 err = skip; 2173 goto out; 2174 } 2175 2176 if (skip) 2177 size += skip; 2178 2179 rd->size += size; 2180 rd->head += size; 2181 rd->file_pos += size; 2182 2183 err = __perf_session__process_decomp_events(session); 2184 if (err) 2185 goto out; 2186 2187 ui_progress__update(prog, size); 2188 2189 out: 2190 return err; 2191 } 2192 2193 static inline bool 2194 reader__eof(struct reader *rd) 2195 { 2196 return (rd->file_pos >= rd->data_size + rd->data_offset); 2197 } 2198 2199 static int 2200 reader__process_events(struct reader *rd, struct perf_session *session, 2201 struct ui_progress *prog) 2202 { 2203 int err; 2204 2205 err = reader__init(rd, &session->one_mmap); 2206 if (err) 2207 goto out; 2208 2209 session->active_decomp = &rd->decomp_data; 2210 2211 remap: 2212 err = reader__mmap(rd, session); 2213 if (err) 2214 goto out; 2215 2216 more: 2217 err = reader__read_event(rd, session, prog); 2218 if (err < 0) 2219 goto out; 2220 else if (err == READER_NODATA) 2221 goto remap; 2222 2223 if (session_done()) 2224 goto out; 2225 2226 if (!reader__eof(rd)) 2227 goto more; 2228 2229 out: 2230 session->active_decomp = &session->decomp_data; 2231 return err; 2232 } 2233 2234 static s64 process_simple(struct perf_session *session, 2235 union perf_event *event, 2236 u64 file_offset, 2237 const char *file_path) 2238 { 2239 return perf_session__process_event(session, event, file_offset, file_path); 2240 } 2241 2242 static int __perf_session__process_events(struct perf_session *session) 2243 { 2244 struct reader rd = { 2245 .fd = perf_data__fd(session->data), 2246 .path = session->data->file.path, 2247 .data_size = session->header.data_size, 2248 .data_offset = session->header.data_offset, 2249 .process = process_simple, 2250 .in_place_update = session->data->in_place_update, 2251 }; 2252 struct ordered_events *oe = &session->ordered_events; 2253 const struct perf_tool *tool = session->tool; 2254 struct ui_progress prog; 2255 int err; 2256 2257 if (rd.data_size == 0) 2258 return -1; 2259 2260 ui_progress__init_size(&prog, rd.data_size, "Processing events..."); 2261 2262 err = reader__process_events(&rd, session, &prog); 2263 if (err) 2264 goto out_err; 2265 /* do the final flush for ordered samples */ 2266 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2267 if (err) 2268 goto out_err; 2269 err = auxtrace__flush_events(session, tool); 2270 if (err) 2271 goto out_err; 2272 err = perf_session__flush_thread_stacks(session); 2273 out_err: 2274 ui_progress__finish(); 2275 if (!tool->no_warn) 2276 perf_session__warn_about_errors(session); 2277 /* 2278 * We may switching perf.data output, make ordered_events 2279 * reusable. 2280 */ 2281 ordered_events__reinit(&session->ordered_events); 2282 auxtrace__free_events(session); 2283 reader__release_decomp(&rd); 2284 session->one_mmap = false; 2285 return err; 2286 } 2287 2288 /* 2289 * Processing 2 MB of data from each reader in sequence, 2290 * because that's the way the ordered events sorting works 2291 * most efficiently. 2292 */ 2293 #define READER_MAX_SIZE (2 * 1024 * 1024) 2294 2295 /* 2296 * This function reads, merge and process directory data. 2297 * It assumens the version 1 of directory data, where each 2298 * data file holds per-cpu data, already sorted by kernel. 2299 */ 2300 static int __perf_session__process_dir_events(struct perf_session *session) 2301 { 2302 struct perf_data *data = session->data; 2303 const struct perf_tool *tool = session->tool; 2304 int i, ret, readers, nr_readers; 2305 struct ui_progress prog; 2306 u64 total_size = perf_data__size(session->data); 2307 struct reader *rd; 2308 2309 ui_progress__init_size(&prog, total_size, "Processing events..."); 2310 2311 nr_readers = 1; 2312 for (i = 0; i < data->dir.nr; i++) { 2313 if (data->dir.files[i].size) 2314 nr_readers++; 2315 } 2316 2317 rd = zalloc(nr_readers * sizeof(struct reader)); 2318 if (!rd) 2319 return -ENOMEM; 2320 2321 rd[0] = (struct reader) { 2322 .fd = perf_data__fd(session->data), 2323 .path = session->data->file.path, 2324 .data_size = session->header.data_size, 2325 .data_offset = session->header.data_offset, 2326 .process = process_simple, 2327 .in_place_update = session->data->in_place_update, 2328 }; 2329 ret = reader__init(&rd[0], NULL); 2330 if (ret) 2331 goto out_err; 2332 ret = reader__mmap(&rd[0], session); 2333 if (ret) 2334 goto out_err; 2335 readers = 1; 2336 2337 for (i = 0; i < data->dir.nr; i++) { 2338 if (!data->dir.files[i].size) 2339 continue; 2340 rd[readers] = (struct reader) { 2341 .fd = data->dir.files[i].fd, 2342 .path = data->dir.files[i].path, 2343 .data_size = data->dir.files[i].size, 2344 .data_offset = 0, 2345 .process = process_simple, 2346 .in_place_update = session->data->in_place_update, 2347 }; 2348 ret = reader__init(&rd[readers], NULL); 2349 if (ret) 2350 goto out_err; 2351 ret = reader__mmap(&rd[readers], session); 2352 if (ret) 2353 goto out_err; 2354 readers++; 2355 } 2356 2357 i = 0; 2358 while (readers) { 2359 if (session_done()) 2360 break; 2361 2362 if (rd[i].done) { 2363 i = (i + 1) % nr_readers; 2364 continue; 2365 } 2366 if (reader__eof(&rd[i])) { 2367 rd[i].done = true; 2368 readers--; 2369 continue; 2370 } 2371 2372 session->active_decomp = &rd[i].decomp_data; 2373 ret = reader__read_event(&rd[i], session, &prog); 2374 if (ret < 0) { 2375 goto out_err; 2376 } else if (ret == READER_NODATA) { 2377 ret = reader__mmap(&rd[i], session); 2378 if (ret) 2379 goto out_err; 2380 } 2381 2382 if (rd[i].size >= READER_MAX_SIZE) { 2383 rd[i].size = 0; 2384 i = (i + 1) % nr_readers; 2385 } 2386 } 2387 2388 ret = ordered_events__flush(&session->ordered_events, OE_FLUSH__FINAL); 2389 if (ret) 2390 goto out_err; 2391 2392 ret = perf_session__flush_thread_stacks(session); 2393 out_err: 2394 ui_progress__finish(); 2395 2396 if (!tool->no_warn) 2397 perf_session__warn_about_errors(session); 2398 2399 /* 2400 * We may switching perf.data output, make ordered_events 2401 * reusable. 2402 */ 2403 ordered_events__reinit(&session->ordered_events); 2404 2405 session->one_mmap = false; 2406 2407 session->active_decomp = &session->decomp_data; 2408 for (i = 0; i < nr_readers; i++) 2409 reader__release_decomp(&rd[i]); 2410 zfree(&rd); 2411 2412 return ret; 2413 } 2414 2415 int perf_session__process_events(struct perf_session *session) 2416 { 2417 if (perf_session__register_idle_thread(session) < 0) 2418 return -ENOMEM; 2419 2420 if (perf_data__is_pipe(session->data)) 2421 return __perf_session__process_pipe_events(session); 2422 2423 if (perf_data__is_dir(session->data) && session->data->dir.nr) 2424 return __perf_session__process_dir_events(session); 2425 2426 return __perf_session__process_events(session); 2427 } 2428 2429 bool perf_session__has_traces(struct perf_session *session, const char *msg) 2430 { 2431 struct evsel *evsel; 2432 2433 evlist__for_each_entry(session->evlist, evsel) { 2434 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT) 2435 return true; 2436 } 2437 2438 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 2439 return false; 2440 } 2441 2442 bool perf_session__has_switch_events(struct perf_session *session) 2443 { 2444 struct evsel *evsel; 2445 2446 evlist__for_each_entry(session->evlist, evsel) { 2447 if (evsel->core.attr.context_switch) 2448 return true; 2449 } 2450 2451 return false; 2452 } 2453 2454 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr) 2455 { 2456 char *bracket; 2457 struct ref_reloc_sym *ref; 2458 struct kmap *kmap; 2459 2460 ref = zalloc(sizeof(struct ref_reloc_sym)); 2461 if (ref == NULL) 2462 return -ENOMEM; 2463 2464 ref->name = strdup(symbol_name); 2465 if (ref->name == NULL) { 2466 free(ref); 2467 return -ENOMEM; 2468 } 2469 2470 bracket = strchr(ref->name, ']'); 2471 if (bracket) 2472 *bracket = '\0'; 2473 2474 ref->addr = addr; 2475 2476 kmap = map__kmap(map); 2477 if (kmap) 2478 kmap->ref_reloc_sym = ref; 2479 2480 return 0; 2481 } 2482 2483 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp) 2484 { 2485 return machines__fprintf_dsos(&session->machines, fp); 2486 } 2487 2488 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp, 2489 bool (skip)(struct dso *dso, int parm), int parm) 2490 { 2491 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm); 2492 } 2493 2494 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 2495 { 2496 size_t ret; 2497 const char *msg = ""; 2498 2499 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) 2500 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)"; 2501 2502 ret = fprintf(fp, "\nAggregated stats:%s\n", msg); 2503 2504 ret += events_stats__fprintf(&session->evlist->stats, fp); 2505 return ret; 2506 } 2507 2508 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 2509 { 2510 /* 2511 * FIXME: Here we have to actually print all the machines in this 2512 * session, not just the host... 2513 */ 2514 return machine__fprintf(&session->machines.host, fp); 2515 } 2516 2517 void perf_session__dump_kmaps(struct perf_session *session) 2518 { 2519 int save_verbose = verbose; 2520 2521 fflush(stdout); 2522 fprintf(stderr, "Kernel and module maps:\n"); 2523 verbose = 0; /* Suppress verbose to print a summary only */ 2524 maps__fprintf(machine__kernel_maps(&session->machines.host), stderr); 2525 verbose = save_verbose; 2526 } 2527 2528 struct evsel *perf_session__find_first_evtype(struct perf_session *session, 2529 unsigned int type) 2530 { 2531 struct evsel *pos; 2532 2533 evlist__for_each_entry(session->evlist, pos) { 2534 if (pos->core.attr.type == type) 2535 return pos; 2536 } 2537 return NULL; 2538 } 2539 2540 int perf_session__cpu_bitmap(struct perf_session *session, 2541 const char *cpu_list, unsigned long *cpu_bitmap) 2542 { 2543 int i, err = -1; 2544 struct perf_cpu_map *map; 2545 int nr_cpus = min(session->header.env.nr_cpus_avail, MAX_NR_CPUS); 2546 struct perf_cpu cpu; 2547 2548 for (i = 0; i < PERF_TYPE_MAX; ++i) { 2549 struct evsel *evsel; 2550 2551 evsel = perf_session__find_first_evtype(session, i); 2552 if (!evsel) 2553 continue; 2554 2555 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) { 2556 pr_err("File does not contain CPU events. " 2557 "Remove -C option to proceed.\n"); 2558 return -1; 2559 } 2560 } 2561 2562 map = perf_cpu_map__new(cpu_list); 2563 if (map == NULL) { 2564 pr_err("Invalid cpu_list\n"); 2565 return -1; 2566 } 2567 2568 perf_cpu_map__for_each_cpu(cpu, i, map) { 2569 if (cpu.cpu >= nr_cpus) { 2570 pr_err("Requested CPU %d too large. " 2571 "Consider raising MAX_NR_CPUS\n", cpu.cpu); 2572 goto out_delete_map; 2573 } 2574 2575 __set_bit(cpu.cpu, cpu_bitmap); 2576 } 2577 2578 err = 0; 2579 2580 out_delete_map: 2581 perf_cpu_map__put(map); 2582 return err; 2583 } 2584 2585 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 2586 bool full) 2587 { 2588 if (session == NULL || fp == NULL) 2589 return; 2590 2591 fprintf(fp, "# ========\n"); 2592 perf_header__fprintf_info(session, fp, full); 2593 fprintf(fp, "# ========\n#\n"); 2594 } 2595 2596 static int perf_session__register_guest(struct perf_session *session, pid_t machine_pid) 2597 { 2598 struct machine *machine = machines__findnew(&session->machines, machine_pid); 2599 struct thread *thread; 2600 2601 if (!machine) 2602 return -ENOMEM; 2603 2604 machine->single_address_space = session->machines.host.single_address_space; 2605 2606 thread = machine__idle_thread(machine); 2607 if (!thread) 2608 return -ENOMEM; 2609 thread__put(thread); 2610 2611 machine->kallsyms_filename = perf_data__guest_kallsyms_name(session->data, machine_pid); 2612 2613 return 0; 2614 } 2615 2616 static int perf_session__set_guest_cpu(struct perf_session *session, pid_t pid, 2617 pid_t tid, int guest_cpu) 2618 { 2619 struct machine *machine = &session->machines.host; 2620 struct thread *thread = machine__findnew_thread(machine, pid, tid); 2621 2622 if (!thread) 2623 return -ENOMEM; 2624 thread__set_guest_cpu(thread, guest_cpu); 2625 thread__put(thread); 2626 2627 return 0; 2628 } 2629 2630 int perf_event__process_id_index(struct perf_session *session, 2631 union perf_event *event) 2632 { 2633 struct evlist *evlist = session->evlist; 2634 struct perf_record_id_index *ie = &event->id_index; 2635 size_t sz = ie->header.size - sizeof(*ie); 2636 size_t i, nr, max_nr; 2637 size_t e1_sz = sizeof(struct id_index_entry); 2638 size_t e2_sz = sizeof(struct id_index_entry_2); 2639 size_t etot_sz = e1_sz + e2_sz; 2640 struct id_index_entry_2 *e2; 2641 pid_t last_pid = 0; 2642 2643 max_nr = sz / e1_sz; 2644 nr = ie->nr; 2645 if (nr > max_nr) { 2646 printf("Too big: nr %zu max_nr %zu\n", nr, max_nr); 2647 return -EINVAL; 2648 } 2649 2650 if (sz >= nr * etot_sz) { 2651 max_nr = sz / etot_sz; 2652 if (nr > max_nr) { 2653 printf("Too big2: nr %zu max_nr %zu\n", nr, max_nr); 2654 return -EINVAL; 2655 } 2656 e2 = (void *)ie + sizeof(*ie) + nr * e1_sz; 2657 } else { 2658 e2 = NULL; 2659 } 2660 2661 if (dump_trace) 2662 fprintf(stdout, " nr: %zu\n", nr); 2663 2664 for (i = 0; i < nr; i++, (e2 ? e2++ : 0)) { 2665 struct id_index_entry *e = &ie->entries[i]; 2666 struct perf_sample_id *sid; 2667 int ret; 2668 2669 if (dump_trace) { 2670 fprintf(stdout, " ... id: %"PRI_lu64, e->id); 2671 fprintf(stdout, " idx: %"PRI_lu64, e->idx); 2672 fprintf(stdout, " cpu: %"PRI_ld64, e->cpu); 2673 fprintf(stdout, " tid: %"PRI_ld64, e->tid); 2674 if (e2) { 2675 fprintf(stdout, " machine_pid: %"PRI_ld64, e2->machine_pid); 2676 fprintf(stdout, " vcpu: %"PRI_lu64"\n", e2->vcpu); 2677 } else { 2678 fprintf(stdout, "\n"); 2679 } 2680 } 2681 2682 sid = evlist__id2sid(evlist, e->id); 2683 if (!sid) 2684 return -ENOENT; 2685 2686 sid->idx = e->idx; 2687 sid->cpu.cpu = e->cpu; 2688 sid->tid = e->tid; 2689 2690 if (!e2) 2691 continue; 2692 2693 sid->machine_pid = e2->machine_pid; 2694 sid->vcpu.cpu = e2->vcpu; 2695 2696 if (!sid->machine_pid) 2697 continue; 2698 2699 if (sid->machine_pid != last_pid) { 2700 ret = perf_session__register_guest(session, sid->machine_pid); 2701 if (ret) 2702 return ret; 2703 last_pid = sid->machine_pid; 2704 perf_guest = true; 2705 } 2706 2707 ret = perf_session__set_guest_cpu(session, sid->machine_pid, e->tid, e2->vcpu); 2708 if (ret) 2709 return ret; 2710 } 2711 return 0; 2712 } 2713 2714 int perf_session__dsos_hit_all(struct perf_session *session) 2715 { 2716 struct rb_node *nd; 2717 int err; 2718 2719 err = machine__hit_all_dsos(&session->machines.host); 2720 if (err) 2721 return err; 2722 2723 for (nd = rb_first_cached(&session->machines.guests); nd; 2724 nd = rb_next(nd)) { 2725 struct machine *pos = rb_entry(nd, struct machine, rb_node); 2726 2727 err = machine__hit_all_dsos(pos); 2728 if (err) 2729 return err; 2730 } 2731 2732 return 0; 2733 } 2734