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 = &sample->user_regs; 954 955 if (user_regs->regs) 956 regs__printf("user", user_regs, arch); 957 } 958 959 static void regs_intr__printf(struct perf_sample *sample, const char *arch) 960 { 961 struct regs_dump *intr_regs = &sample->intr_regs; 962 963 if (intr_regs->regs) 964 regs__printf("intr", intr_regs, arch); 965 } 966 967 static void stack_user__printf(struct stack_dump *dump) 968 { 969 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 970 dump->size, dump->offset); 971 } 972 973 static void evlist__print_tstamp(struct evlist *evlist, union perf_event *event, struct perf_sample *sample) 974 { 975 u64 sample_type = __evlist__combined_sample_type(evlist); 976 977 if (event->header.type != PERF_RECORD_SAMPLE && 978 !evlist__sample_id_all(evlist)) { 979 fputs("-1 -1 ", stdout); 980 return; 981 } 982 983 if ((sample_type & PERF_SAMPLE_CPU)) 984 printf("%u ", sample->cpu); 985 986 if (sample_type & PERF_SAMPLE_TIME) 987 printf("%" PRIu64 " ", sample->time); 988 } 989 990 static void sample_read__printf(struct perf_sample *sample, u64 read_format) 991 { 992 printf("... sample_read:\n"); 993 994 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 995 printf("...... time enabled %016" PRIx64 "\n", 996 sample->read.time_enabled); 997 998 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 999 printf("...... time running %016" PRIx64 "\n", 1000 sample->read.time_running); 1001 1002 if (read_format & PERF_FORMAT_GROUP) { 1003 struct sample_read_value *value = sample->read.group.values; 1004 1005 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr); 1006 1007 sample_read_group__for_each(value, sample->read.group.nr, read_format) { 1008 printf("..... id %016" PRIx64 1009 ", value %016" PRIx64, 1010 value->id, value->value); 1011 if (read_format & PERF_FORMAT_LOST) 1012 printf(", lost %" PRIu64, value->lost); 1013 printf("\n"); 1014 } 1015 } else { 1016 printf("..... id %016" PRIx64 ", value %016" PRIx64, 1017 sample->read.one.id, sample->read.one.value); 1018 if (read_format & PERF_FORMAT_LOST) 1019 printf(", lost %" PRIu64, sample->read.one.lost); 1020 printf("\n"); 1021 } 1022 } 1023 1024 static void dump_event(struct evlist *evlist, union perf_event *event, 1025 u64 file_offset, struct perf_sample *sample, 1026 const char *file_path) 1027 { 1028 if (!dump_trace) 1029 return; 1030 1031 printf("\n%#" PRIx64 "@%s [%#x]: event: %d\n", 1032 file_offset, file_path, event->header.size, event->header.type); 1033 1034 trace_event(event); 1035 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw) 1036 evlist->trace_event_sample_raw(evlist, event, sample); 1037 1038 if (sample) 1039 evlist__print_tstamp(evlist, event, sample); 1040 1041 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 1042 event->header.size, perf_event__name(event->header.type)); 1043 } 1044 1045 char *get_page_size_name(u64 size, char *str) 1046 { 1047 if (!size || !unit_number__scnprintf(str, PAGE_SIZE_NAME_LEN, size)) 1048 snprintf(str, PAGE_SIZE_NAME_LEN, "%s", "N/A"); 1049 1050 return str; 1051 } 1052 1053 static void dump_sample(struct evsel *evsel, union perf_event *event, 1054 struct perf_sample *sample, const char *arch) 1055 { 1056 u64 sample_type; 1057 char str[PAGE_SIZE_NAME_LEN]; 1058 1059 if (!dump_trace) 1060 return; 1061 1062 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 1063 event->header.misc, sample->pid, sample->tid, sample->ip, 1064 sample->period, sample->addr); 1065 1066 sample_type = evsel->core.attr.sample_type; 1067 1068 if (evsel__has_callchain(evsel)) 1069 callchain__printf(evsel, sample); 1070 1071 if (evsel__has_br_stack(evsel)) 1072 branch_stack__printf(sample, evsel); 1073 1074 if (sample_type & PERF_SAMPLE_REGS_USER) 1075 regs_user__printf(sample, arch); 1076 1077 if (sample_type & PERF_SAMPLE_REGS_INTR) 1078 regs_intr__printf(sample, arch); 1079 1080 if (sample_type & PERF_SAMPLE_STACK_USER) 1081 stack_user__printf(&sample->user_stack); 1082 1083 if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) { 1084 printf("... weight: %" PRIu64 "", sample->weight); 1085 if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) { 1086 printf(",0x%"PRIx16"", sample->ins_lat); 1087 printf(",0x%"PRIx16"", sample->p_stage_cyc); 1088 } 1089 printf("\n"); 1090 } 1091 1092 if (sample_type & PERF_SAMPLE_DATA_SRC) 1093 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src); 1094 1095 if (sample_type & PERF_SAMPLE_PHYS_ADDR) 1096 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr); 1097 1098 if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE) 1099 printf(" .. data page size: %s\n", get_page_size_name(sample->data_page_size, str)); 1100 1101 if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE) 1102 printf(" .. code page size: %s\n", get_page_size_name(sample->code_page_size, str)); 1103 1104 if (sample_type & PERF_SAMPLE_TRANSACTION) 1105 printf("... transaction: %" PRIx64 "\n", sample->transaction); 1106 1107 if (sample_type & PERF_SAMPLE_READ) 1108 sample_read__printf(sample, evsel->core.attr.read_format); 1109 } 1110 1111 static void dump_read(struct evsel *evsel, union perf_event *event) 1112 { 1113 struct perf_record_read *read_event = &event->read; 1114 u64 read_format; 1115 1116 if (!dump_trace) 1117 return; 1118 1119 printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid, 1120 evsel__name(evsel), event->read.value); 1121 1122 if (!evsel) 1123 return; 1124 1125 read_format = evsel->core.attr.read_format; 1126 1127 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1128 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled); 1129 1130 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1131 printf("... time running : %" PRI_lu64 "\n", read_event->time_running); 1132 1133 if (read_format & PERF_FORMAT_ID) 1134 printf("... id : %" PRI_lu64 "\n", read_event->id); 1135 1136 if (read_format & PERF_FORMAT_LOST) 1137 printf("... lost : %" PRI_lu64 "\n", read_event->lost); 1138 } 1139 1140 static struct machine *machines__find_for_cpumode(struct machines *machines, 1141 union perf_event *event, 1142 struct perf_sample *sample) 1143 { 1144 if (perf_guest && 1145 ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 1146 (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) { 1147 u32 pid; 1148 1149 if (sample->machine_pid) 1150 pid = sample->machine_pid; 1151 else if (event->header.type == PERF_RECORD_MMAP 1152 || event->header.type == PERF_RECORD_MMAP2) 1153 pid = event->mmap.pid; 1154 else 1155 pid = sample->pid; 1156 1157 /* 1158 * Guest code machine is created as needed and does not use 1159 * DEFAULT_GUEST_KERNEL_ID. 1160 */ 1161 if (symbol_conf.guest_code) 1162 return machines__findnew(machines, pid); 1163 1164 return machines__find_guest(machines, pid); 1165 } 1166 1167 return &machines->host; 1168 } 1169 1170 static int deliver_sample_value(struct evlist *evlist, 1171 const struct perf_tool *tool, 1172 union perf_event *event, 1173 struct perf_sample *sample, 1174 struct sample_read_value *v, 1175 struct machine *machine, 1176 bool per_thread) 1177 { 1178 struct perf_sample_id *sid = evlist__id2sid(evlist, v->id); 1179 struct evsel *evsel; 1180 u64 *storage = NULL; 1181 1182 if (sid) { 1183 storage = perf_sample_id__get_period_storage(sid, sample->tid, per_thread); 1184 } 1185 1186 if (storage) { 1187 sample->id = v->id; 1188 sample->period = v->value - *storage; 1189 *storage = v->value; 1190 } 1191 1192 if (!storage || sid->evsel == NULL) { 1193 ++evlist->stats.nr_unknown_id; 1194 return 0; 1195 } 1196 1197 /* 1198 * There's no reason to deliver sample 1199 * for zero period, bail out. 1200 */ 1201 if (!sample->period) 1202 return 0; 1203 1204 evsel = container_of(sid->evsel, struct evsel, core); 1205 return tool->sample(tool, event, sample, evsel, machine); 1206 } 1207 1208 static int deliver_sample_group(struct evlist *evlist, 1209 const struct perf_tool *tool, 1210 union perf_event *event, 1211 struct perf_sample *sample, 1212 struct machine *machine, 1213 u64 read_format, 1214 bool per_thread) 1215 { 1216 int ret = -EINVAL; 1217 struct sample_read_value *v = sample->read.group.values; 1218 1219 if (tool->dont_split_sample_group) 1220 return deliver_sample_value(evlist, tool, event, sample, v, machine, 1221 per_thread); 1222 1223 sample_read_group__for_each(v, sample->read.group.nr, read_format) { 1224 ret = deliver_sample_value(evlist, tool, event, sample, v, 1225 machine, per_thread); 1226 if (ret) 1227 break; 1228 } 1229 1230 return ret; 1231 } 1232 1233 static int evlist__deliver_sample(struct evlist *evlist, const struct perf_tool *tool, 1234 union perf_event *event, struct perf_sample *sample, 1235 struct evsel *evsel, struct machine *machine) 1236 { 1237 /* We know evsel != NULL. */ 1238 u64 sample_type = evsel->core.attr.sample_type; 1239 u64 read_format = evsel->core.attr.read_format; 1240 bool per_thread = perf_evsel__attr_has_per_thread_sample_period(&evsel->core); 1241 1242 /* Standard sample delivery. */ 1243 if (!(sample_type & PERF_SAMPLE_READ)) 1244 return tool->sample(tool, event, sample, evsel, machine); 1245 1246 /* For PERF_SAMPLE_READ we have either single or group mode. */ 1247 if (read_format & PERF_FORMAT_GROUP) 1248 return deliver_sample_group(evlist, tool, event, sample, 1249 machine, read_format, per_thread); 1250 else 1251 return deliver_sample_value(evlist, tool, event, sample, 1252 &sample->read.one, machine, 1253 per_thread); 1254 } 1255 1256 static int machines__deliver_event(struct machines *machines, 1257 struct evlist *evlist, 1258 union perf_event *event, 1259 struct perf_sample *sample, 1260 const struct perf_tool *tool, u64 file_offset, 1261 const char *file_path) 1262 { 1263 struct evsel *evsel; 1264 struct machine *machine; 1265 1266 dump_event(evlist, event, file_offset, sample, file_path); 1267 1268 evsel = evlist__id2evsel(evlist, sample->id); 1269 1270 machine = machines__find_for_cpumode(machines, event, sample); 1271 1272 switch (event->header.type) { 1273 case PERF_RECORD_SAMPLE: 1274 if (evsel == NULL) { 1275 ++evlist->stats.nr_unknown_id; 1276 return 0; 1277 } 1278 if (machine == NULL) { 1279 ++evlist->stats.nr_unprocessable_samples; 1280 dump_sample(evsel, event, sample, perf_env__arch(NULL)); 1281 return 0; 1282 } 1283 dump_sample(evsel, event, sample, perf_env__arch(machine->env)); 1284 return evlist__deliver_sample(evlist, tool, event, sample, evsel, machine); 1285 case PERF_RECORD_MMAP: 1286 return tool->mmap(tool, event, sample, machine); 1287 case PERF_RECORD_MMAP2: 1288 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT) 1289 ++evlist->stats.nr_proc_map_timeout; 1290 return tool->mmap2(tool, event, sample, machine); 1291 case PERF_RECORD_COMM: 1292 return tool->comm(tool, event, sample, machine); 1293 case PERF_RECORD_NAMESPACES: 1294 return tool->namespaces(tool, event, sample, machine); 1295 case PERF_RECORD_CGROUP: 1296 return tool->cgroup(tool, event, sample, machine); 1297 case PERF_RECORD_FORK: 1298 return tool->fork(tool, event, sample, machine); 1299 case PERF_RECORD_EXIT: 1300 return tool->exit(tool, event, sample, machine); 1301 case PERF_RECORD_LOST: 1302 if (tool->lost == perf_event__process_lost) 1303 evlist->stats.total_lost += event->lost.lost; 1304 return tool->lost(tool, event, sample, machine); 1305 case PERF_RECORD_LOST_SAMPLES: 1306 if (event->header.misc & PERF_RECORD_MISC_LOST_SAMPLES_BPF) 1307 evlist->stats.total_dropped_samples += event->lost_samples.lost; 1308 else if (tool->lost_samples == perf_event__process_lost_samples) 1309 evlist->stats.total_lost_samples += event->lost_samples.lost; 1310 return tool->lost_samples(tool, event, sample, machine); 1311 case PERF_RECORD_READ: 1312 dump_read(evsel, event); 1313 return tool->read(tool, event, sample, evsel, machine); 1314 case PERF_RECORD_THROTTLE: 1315 return tool->throttle(tool, event, sample, machine); 1316 case PERF_RECORD_UNTHROTTLE: 1317 return tool->unthrottle(tool, event, sample, machine); 1318 case PERF_RECORD_AUX: 1319 if (tool->aux == perf_event__process_aux) { 1320 if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) 1321 evlist->stats.total_aux_lost += 1; 1322 if (event->aux.flags & PERF_AUX_FLAG_PARTIAL) 1323 evlist->stats.total_aux_partial += 1; 1324 if (event->aux.flags & PERF_AUX_FLAG_COLLISION) 1325 evlist->stats.total_aux_collision += 1; 1326 } 1327 return tool->aux(tool, event, sample, machine); 1328 case PERF_RECORD_ITRACE_START: 1329 return tool->itrace_start(tool, event, sample, machine); 1330 case PERF_RECORD_SWITCH: 1331 case PERF_RECORD_SWITCH_CPU_WIDE: 1332 return tool->context_switch(tool, event, sample, machine); 1333 case PERF_RECORD_KSYMBOL: 1334 return tool->ksymbol(tool, event, sample, machine); 1335 case PERF_RECORD_BPF_EVENT: 1336 return tool->bpf(tool, event, sample, machine); 1337 case PERF_RECORD_TEXT_POKE: 1338 return tool->text_poke(tool, event, sample, machine); 1339 case PERF_RECORD_AUX_OUTPUT_HW_ID: 1340 return tool->aux_output_hw_id(tool, event, sample, machine); 1341 default: 1342 ++evlist->stats.nr_unknown_events; 1343 return -1; 1344 } 1345 } 1346 1347 static int perf_session__deliver_event(struct perf_session *session, 1348 union perf_event *event, 1349 const struct perf_tool *tool, 1350 u64 file_offset, 1351 const char *file_path) 1352 { 1353 struct perf_sample sample; 1354 int ret = evlist__parse_sample(session->evlist, event, &sample); 1355 1356 if (ret) { 1357 pr_err("Can't parse sample, err = %d\n", ret); 1358 return ret; 1359 } 1360 1361 ret = auxtrace__process_event(session, event, &sample, tool); 1362 if (ret < 0) 1363 return ret; 1364 if (ret > 0) 1365 return 0; 1366 1367 ret = machines__deliver_event(&session->machines, session->evlist, 1368 event, &sample, tool, file_offset, file_path); 1369 1370 if (dump_trace && sample.aux_sample.size) 1371 auxtrace__dump_auxtrace_sample(session, &sample); 1372 1373 return ret; 1374 } 1375 1376 static s64 perf_session__process_user_event(struct perf_session *session, 1377 union perf_event *event, 1378 u64 file_offset, 1379 const char *file_path) 1380 { 1381 struct ordered_events *oe = &session->ordered_events; 1382 const struct perf_tool *tool = session->tool; 1383 struct perf_sample sample = { .time = 0, }; 1384 int fd = perf_data__fd(session->data); 1385 int err; 1386 1387 if (event->header.type != PERF_RECORD_COMPRESSED || perf_tool__compressed_is_stub(tool)) 1388 dump_event(session->evlist, event, file_offset, &sample, file_path); 1389 1390 /* These events are processed right away */ 1391 switch (event->header.type) { 1392 case PERF_RECORD_HEADER_ATTR: 1393 err = tool->attr(tool, event, &session->evlist); 1394 if (err == 0) { 1395 perf_session__set_id_hdr_size(session); 1396 perf_session__set_comm_exec(session); 1397 } 1398 return err; 1399 case PERF_RECORD_EVENT_UPDATE: 1400 return tool->event_update(tool, event, &session->evlist); 1401 case PERF_RECORD_HEADER_EVENT_TYPE: 1402 /* 1403 * Deprecated, but we need to handle it for sake 1404 * of old data files create in pipe mode. 1405 */ 1406 return 0; 1407 case PERF_RECORD_HEADER_TRACING_DATA: 1408 /* 1409 * Setup for reading amidst mmap, but only when we 1410 * are in 'file' mode. The 'pipe' fd is in proper 1411 * place already. 1412 */ 1413 if (!perf_data__is_pipe(session->data)) 1414 lseek(fd, file_offset, SEEK_SET); 1415 return tool->tracing_data(session, event); 1416 case PERF_RECORD_HEADER_BUILD_ID: 1417 return tool->build_id(session, event); 1418 case PERF_RECORD_FINISHED_ROUND: 1419 return tool->finished_round(tool, event, oe); 1420 case PERF_RECORD_ID_INDEX: 1421 return tool->id_index(session, event); 1422 case PERF_RECORD_AUXTRACE_INFO: 1423 return tool->auxtrace_info(session, event); 1424 case PERF_RECORD_AUXTRACE: 1425 /* 1426 * Setup for reading amidst mmap, but only when we 1427 * are in 'file' mode. The 'pipe' fd is in proper 1428 * place already. 1429 */ 1430 if (!perf_data__is_pipe(session->data)) 1431 lseek(fd, file_offset + event->header.size, SEEK_SET); 1432 return tool->auxtrace(session, event); 1433 case PERF_RECORD_AUXTRACE_ERROR: 1434 perf_session__auxtrace_error_inc(session, event); 1435 return tool->auxtrace_error(session, event); 1436 case PERF_RECORD_THREAD_MAP: 1437 return tool->thread_map(session, event); 1438 case PERF_RECORD_CPU_MAP: 1439 return tool->cpu_map(session, event); 1440 case PERF_RECORD_STAT_CONFIG: 1441 return tool->stat_config(session, event); 1442 case PERF_RECORD_STAT: 1443 return tool->stat(session, event); 1444 case PERF_RECORD_STAT_ROUND: 1445 return tool->stat_round(session, event); 1446 case PERF_RECORD_TIME_CONV: 1447 session->time_conv = event->time_conv; 1448 return tool->time_conv(session, event); 1449 case PERF_RECORD_HEADER_FEATURE: 1450 return tool->feature(session, event); 1451 case PERF_RECORD_COMPRESSED: 1452 err = tool->compressed(session, event, file_offset, file_path); 1453 if (err) 1454 dump_event(session->evlist, event, file_offset, &sample, file_path); 1455 return err; 1456 case PERF_RECORD_FINISHED_INIT: 1457 return tool->finished_init(session, event); 1458 default: 1459 return -EINVAL; 1460 } 1461 } 1462 1463 int perf_session__deliver_synth_event(struct perf_session *session, 1464 union perf_event *event, 1465 struct perf_sample *sample) 1466 { 1467 struct evlist *evlist = session->evlist; 1468 const struct perf_tool *tool = session->tool; 1469 1470 events_stats__inc(&evlist->stats, event->header.type); 1471 1472 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1473 return perf_session__process_user_event(session, event, 0, NULL); 1474 1475 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0, NULL); 1476 } 1477 1478 int perf_session__deliver_synth_attr_event(struct perf_session *session, 1479 const struct perf_event_attr *attr, 1480 u64 id) 1481 { 1482 union { 1483 struct { 1484 struct perf_record_header_attr attr; 1485 u64 ids[1]; 1486 } attr_id; 1487 union perf_event ev; 1488 } ev = { 1489 .attr_id.attr.header.type = PERF_RECORD_HEADER_ATTR, 1490 .attr_id.attr.header.size = sizeof(ev.attr_id), 1491 .attr_id.ids[0] = id, 1492 }; 1493 1494 if (attr->size != sizeof(ev.attr_id.attr.attr)) { 1495 pr_debug("Unexpected perf_event_attr size\n"); 1496 return -EINVAL; 1497 } 1498 ev.attr_id.attr.attr = *attr; 1499 return perf_session__deliver_synth_event(session, &ev.ev, NULL); 1500 } 1501 1502 static void event_swap(union perf_event *event, bool sample_id_all) 1503 { 1504 perf_event__swap_op swap; 1505 1506 swap = perf_event__swap_ops[event->header.type]; 1507 if (swap) 1508 swap(event, sample_id_all); 1509 } 1510 1511 int perf_session__peek_event(struct perf_session *session, off_t file_offset, 1512 void *buf, size_t buf_sz, 1513 union perf_event **event_ptr, 1514 struct perf_sample *sample) 1515 { 1516 union perf_event *event; 1517 size_t hdr_sz, rest; 1518 int fd; 1519 1520 if (session->one_mmap && !session->header.needs_swap) { 1521 event = file_offset - session->one_mmap_offset + 1522 session->one_mmap_addr; 1523 goto out_parse_sample; 1524 } 1525 1526 if (perf_data__is_pipe(session->data)) 1527 return -1; 1528 1529 fd = perf_data__fd(session->data); 1530 hdr_sz = sizeof(struct perf_event_header); 1531 1532 if (buf_sz < hdr_sz) 1533 return -1; 1534 1535 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 || 1536 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz) 1537 return -1; 1538 1539 event = (union perf_event *)buf; 1540 1541 if (session->header.needs_swap) 1542 perf_event_header__bswap(&event->header); 1543 1544 if (event->header.size < hdr_sz || event->header.size > buf_sz) 1545 return -1; 1546 1547 buf += hdr_sz; 1548 rest = event->header.size - hdr_sz; 1549 1550 if (readn(fd, buf, rest) != (ssize_t)rest) 1551 return -1; 1552 1553 if (session->header.needs_swap) 1554 event_swap(event, evlist__sample_id_all(session->evlist)); 1555 1556 out_parse_sample: 1557 1558 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START && 1559 evlist__parse_sample(session->evlist, event, sample)) 1560 return -1; 1561 1562 *event_ptr = event; 1563 1564 return 0; 1565 } 1566 1567 int perf_session__peek_events(struct perf_session *session, u64 offset, 1568 u64 size, peek_events_cb_t cb, void *data) 1569 { 1570 u64 max_offset = offset + size; 1571 char buf[PERF_SAMPLE_MAX_SIZE]; 1572 union perf_event *event; 1573 int err; 1574 1575 do { 1576 err = perf_session__peek_event(session, offset, buf, 1577 PERF_SAMPLE_MAX_SIZE, &event, 1578 NULL); 1579 if (err) 1580 return err; 1581 1582 err = cb(session, event, offset, data); 1583 if (err) 1584 return err; 1585 1586 offset += event->header.size; 1587 if (event->header.type == PERF_RECORD_AUXTRACE) 1588 offset += event->auxtrace.size; 1589 1590 } while (offset < max_offset); 1591 1592 return err; 1593 } 1594 1595 static s64 perf_session__process_event(struct perf_session *session, 1596 union perf_event *event, u64 file_offset, 1597 const char *file_path) 1598 { 1599 struct evlist *evlist = session->evlist; 1600 const struct perf_tool *tool = session->tool; 1601 int ret; 1602 1603 if (session->header.needs_swap) 1604 event_swap(event, evlist__sample_id_all(evlist)); 1605 1606 if (event->header.type >= PERF_RECORD_HEADER_MAX) 1607 return -EINVAL; 1608 1609 events_stats__inc(&evlist->stats, event->header.type); 1610 1611 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1612 return perf_session__process_user_event(session, event, file_offset, file_path); 1613 1614 if (tool->ordered_events) { 1615 u64 timestamp = -1ULL; 1616 1617 ret = evlist__parse_sample_timestamp(evlist, event, ×tamp); 1618 if (ret && ret != -1) 1619 return ret; 1620 1621 ret = perf_session__queue_event(session, event, timestamp, file_offset, file_path); 1622 if (ret != -ETIME) 1623 return ret; 1624 } 1625 1626 return perf_session__deliver_event(session, event, tool, file_offset, file_path); 1627 } 1628 1629 void perf_event_header__bswap(struct perf_event_header *hdr) 1630 { 1631 hdr->type = bswap_32(hdr->type); 1632 hdr->misc = bswap_16(hdr->misc); 1633 hdr->size = bswap_16(hdr->size); 1634 } 1635 1636 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1637 { 1638 return machine__findnew_thread(&session->machines.host, -1, pid); 1639 } 1640 1641 int perf_session__register_idle_thread(struct perf_session *session) 1642 { 1643 struct thread *thread = machine__idle_thread(&session->machines.host); 1644 1645 /* machine__idle_thread() got the thread, so put it */ 1646 thread__put(thread); 1647 return thread ? 0 : -1; 1648 } 1649 1650 static void 1651 perf_session__warn_order(const struct perf_session *session) 1652 { 1653 const struct ordered_events *oe = &session->ordered_events; 1654 struct evsel *evsel; 1655 bool should_warn = true; 1656 1657 evlist__for_each_entry(session->evlist, evsel) { 1658 if (evsel->core.attr.write_backward) 1659 should_warn = false; 1660 } 1661 1662 if (!should_warn) 1663 return; 1664 if (oe->nr_unordered_events != 0) 1665 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events); 1666 } 1667 1668 static void perf_session__warn_about_errors(const struct perf_session *session) 1669 { 1670 const struct events_stats *stats = &session->evlist->stats; 1671 1672 if (session->tool->lost == perf_event__process_lost && 1673 stats->nr_events[PERF_RECORD_LOST] != 0) { 1674 ui__warning("Processed %d events and lost %d chunks!\n\n" 1675 "Check IO/CPU overload!\n\n", 1676 stats->nr_events[0], 1677 stats->nr_events[PERF_RECORD_LOST]); 1678 } 1679 1680 if (session->tool->lost_samples == perf_event__process_lost_samples) { 1681 double drop_rate; 1682 1683 drop_rate = (double)stats->total_lost_samples / 1684 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples); 1685 if (drop_rate > 0.05) { 1686 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n", 1687 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples, 1688 drop_rate * 100.0); 1689 } 1690 } 1691 1692 if (session->tool->aux == perf_event__process_aux && 1693 stats->total_aux_lost != 0) { 1694 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n", 1695 stats->total_aux_lost, 1696 stats->nr_events[PERF_RECORD_AUX]); 1697 } 1698 1699 if (session->tool->aux == perf_event__process_aux && 1700 stats->total_aux_partial != 0) { 1701 bool vmm_exclusive = false; 1702 1703 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive", 1704 &vmm_exclusive); 1705 1706 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n" 1707 "Are you running a KVM guest in the background?%s\n\n", 1708 stats->total_aux_partial, 1709 stats->nr_events[PERF_RECORD_AUX], 1710 vmm_exclusive ? 1711 "\nReloading kvm_intel module with vmm_exclusive=0\n" 1712 "will reduce the gaps to only guest's timeslices." : 1713 ""); 1714 } 1715 1716 if (session->tool->aux == perf_event__process_aux && 1717 stats->total_aux_collision != 0) { 1718 ui__warning("AUX data detected collision %" PRIu64 " times out of %u!\n\n", 1719 stats->total_aux_collision, 1720 stats->nr_events[PERF_RECORD_AUX]); 1721 } 1722 1723 if (stats->nr_unknown_events != 0) { 1724 ui__warning("Found %u unknown events!\n\n" 1725 "Is this an older tool processing a perf.data " 1726 "file generated by a more recent tool?\n\n" 1727 "If that is not the case, consider " 1728 "reporting to linux-kernel@vger.kernel.org.\n\n", 1729 stats->nr_unknown_events); 1730 } 1731 1732 if (stats->nr_unknown_id != 0) { 1733 ui__warning("%u samples with id not present in the header\n", 1734 stats->nr_unknown_id); 1735 } 1736 1737 if (stats->nr_invalid_chains != 0) { 1738 ui__warning("Found invalid callchains!\n\n" 1739 "%u out of %u events were discarded for this reason.\n\n" 1740 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 1741 stats->nr_invalid_chains, 1742 stats->nr_events[PERF_RECORD_SAMPLE]); 1743 } 1744 1745 if (stats->nr_unprocessable_samples != 0) { 1746 ui__warning("%u unprocessable samples recorded.\n" 1747 "Do you have a KVM guest running and not using 'perf kvm'?\n", 1748 stats->nr_unprocessable_samples); 1749 } 1750 1751 perf_session__warn_order(session); 1752 1753 events_stats__auxtrace_error_warn(stats); 1754 1755 if (stats->nr_proc_map_timeout != 0) { 1756 ui__warning("%d map information files for pre-existing threads were\n" 1757 "not processed, if there are samples for addresses they\n" 1758 "will not be resolved, you may find out which are these\n" 1759 "threads by running with -v and redirecting the output\n" 1760 "to a file.\n" 1761 "The time limit to process proc map is too short?\n" 1762 "Increase it by --proc-map-timeout\n", 1763 stats->nr_proc_map_timeout); 1764 } 1765 } 1766 1767 static int perf_session__flush_thread_stack(struct thread *thread, 1768 void *p __maybe_unused) 1769 { 1770 return thread_stack__flush(thread); 1771 } 1772 1773 static int perf_session__flush_thread_stacks(struct perf_session *session) 1774 { 1775 return machines__for_each_thread(&session->machines, 1776 perf_session__flush_thread_stack, 1777 NULL); 1778 } 1779 1780 volatile sig_atomic_t session_done; 1781 1782 static int __perf_session__process_decomp_events(struct perf_session *session); 1783 1784 static int __perf_session__process_pipe_events(struct perf_session *session) 1785 { 1786 struct ordered_events *oe = &session->ordered_events; 1787 const struct perf_tool *tool = session->tool; 1788 struct ui_progress prog; 1789 union perf_event *event; 1790 uint32_t size, cur_size = 0; 1791 void *buf = NULL; 1792 s64 skip = 0; 1793 u64 head; 1794 ssize_t err; 1795 void *p; 1796 bool update_prog = false; 1797 1798 /* 1799 * If it's from a file saving pipe data (by redirection), it would have 1800 * a file name other than "-". Then we can get the total size and show 1801 * the progress. 1802 */ 1803 if (strcmp(session->data->path, "-") && session->data->file.size) { 1804 ui_progress__init_size(&prog, session->data->file.size, 1805 "Processing events..."); 1806 update_prog = true; 1807 } 1808 1809 head = 0; 1810 cur_size = sizeof(union perf_event); 1811 1812 buf = malloc(cur_size); 1813 if (!buf) 1814 return -errno; 1815 ordered_events__set_copy_on_queue(oe, true); 1816 more: 1817 event = buf; 1818 err = perf_data__read(session->data, event, 1819 sizeof(struct perf_event_header)); 1820 if (err <= 0) { 1821 if (err == 0) 1822 goto done; 1823 1824 pr_err("failed to read event header\n"); 1825 goto out_err; 1826 } 1827 1828 if (session->header.needs_swap) 1829 perf_event_header__bswap(&event->header); 1830 1831 size = event->header.size; 1832 if (size < sizeof(struct perf_event_header)) { 1833 pr_err("bad event header size\n"); 1834 goto out_err; 1835 } 1836 1837 if (size > cur_size) { 1838 void *new = realloc(buf, size); 1839 if (!new) { 1840 pr_err("failed to allocate memory to read event\n"); 1841 goto out_err; 1842 } 1843 buf = new; 1844 cur_size = size; 1845 event = buf; 1846 } 1847 p = event; 1848 p += sizeof(struct perf_event_header); 1849 1850 if (size - sizeof(struct perf_event_header)) { 1851 err = perf_data__read(session->data, p, 1852 size - sizeof(struct perf_event_header)); 1853 if (err <= 0) { 1854 if (err == 0) { 1855 pr_err("unexpected end of event stream\n"); 1856 goto done; 1857 } 1858 1859 pr_err("failed to read event data\n"); 1860 goto out_err; 1861 } 1862 } 1863 1864 if ((skip = perf_session__process_event(session, event, head, "pipe")) < 0) { 1865 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 1866 head, event->header.size, event->header.type); 1867 err = -EINVAL; 1868 goto out_err; 1869 } 1870 1871 head += size; 1872 1873 if (skip > 0) 1874 head += skip; 1875 1876 err = __perf_session__process_decomp_events(session); 1877 if (err) 1878 goto out_err; 1879 1880 if (update_prog) 1881 ui_progress__update(&prog, size); 1882 1883 if (!session_done()) 1884 goto more; 1885 done: 1886 /* do the final flush for ordered samples */ 1887 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 1888 if (err) 1889 goto out_err; 1890 err = auxtrace__flush_events(session, tool); 1891 if (err) 1892 goto out_err; 1893 err = perf_session__flush_thread_stacks(session); 1894 out_err: 1895 free(buf); 1896 if (update_prog) 1897 ui_progress__finish(); 1898 if (!tool->no_warn) 1899 perf_session__warn_about_errors(session); 1900 ordered_events__free(&session->ordered_events); 1901 auxtrace__free_events(session); 1902 return err; 1903 } 1904 1905 static union perf_event * 1906 prefetch_event(char *buf, u64 head, size_t mmap_size, 1907 bool needs_swap, union perf_event *error) 1908 { 1909 union perf_event *event; 1910 u16 event_size; 1911 1912 /* 1913 * Ensure we have enough space remaining to read 1914 * the size of the event in the headers. 1915 */ 1916 if (head + sizeof(event->header) > mmap_size) 1917 return NULL; 1918 1919 event = (union perf_event *)(buf + head); 1920 if (needs_swap) 1921 perf_event_header__bswap(&event->header); 1922 1923 event_size = event->header.size; 1924 if (head + event_size <= mmap_size) 1925 return event; 1926 1927 /* We're not fetching the event so swap back again */ 1928 if (needs_swap) 1929 perf_event_header__bswap(&event->header); 1930 1931 /* Check if the event fits into the next mmapped buf. */ 1932 if (event_size <= mmap_size - head % page_size) { 1933 /* Remap buf and fetch again. */ 1934 return NULL; 1935 } 1936 1937 /* Invalid input. Event size should never exceed mmap_size. */ 1938 pr_debug("%s: head=%#" PRIx64 " event->header.size=%#x, mmap_size=%#zx:" 1939 " fuzzed or compressed perf.data?\n", __func__, head, event_size, mmap_size); 1940 1941 return error; 1942 } 1943 1944 static union perf_event * 1945 fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 1946 { 1947 return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL)); 1948 } 1949 1950 static union perf_event * 1951 fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 1952 { 1953 return prefetch_event(buf, head, mmap_size, needs_swap, NULL); 1954 } 1955 1956 static int __perf_session__process_decomp_events(struct perf_session *session) 1957 { 1958 s64 skip; 1959 u64 size; 1960 struct decomp *decomp = session->active_decomp->decomp_last; 1961 1962 if (!decomp) 1963 return 0; 1964 1965 while (decomp->head < decomp->size && !session_done()) { 1966 union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data, 1967 session->header.needs_swap); 1968 1969 if (!event) 1970 break; 1971 1972 size = event->header.size; 1973 1974 if (size < sizeof(struct perf_event_header) || 1975 (skip = perf_session__process_event(session, event, decomp->file_pos, 1976 decomp->file_path)) < 0) { 1977 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 1978 decomp->file_pos + decomp->head, event->header.size, event->header.type); 1979 return -EINVAL; 1980 } 1981 1982 if (skip) 1983 size += skip; 1984 1985 decomp->head += size; 1986 } 1987 1988 return 0; 1989 } 1990 1991 /* 1992 * On 64bit we can mmap the data file in one go. No need for tiny mmap 1993 * slices. On 32bit we use 32MB. 1994 */ 1995 #if BITS_PER_LONG == 64 1996 #define MMAP_SIZE ULLONG_MAX 1997 #define NUM_MMAPS 1 1998 #else 1999 #define MMAP_SIZE (32 * 1024 * 1024ULL) 2000 #define NUM_MMAPS 128 2001 #endif 2002 2003 struct reader; 2004 2005 typedef s64 (*reader_cb_t)(struct perf_session *session, 2006 union perf_event *event, 2007 u64 file_offset, 2008 const char *file_path); 2009 2010 struct reader { 2011 int fd; 2012 const char *path; 2013 u64 data_size; 2014 u64 data_offset; 2015 reader_cb_t process; 2016 bool in_place_update; 2017 char *mmaps[NUM_MMAPS]; 2018 size_t mmap_size; 2019 int mmap_idx; 2020 char *mmap_cur; 2021 u64 file_pos; 2022 u64 file_offset; 2023 u64 head; 2024 u64 size; 2025 bool done; 2026 struct zstd_data zstd_data; 2027 struct decomp_data decomp_data; 2028 }; 2029 2030 static int 2031 reader__init(struct reader *rd, bool *one_mmap) 2032 { 2033 u64 data_size = rd->data_size; 2034 char **mmaps = rd->mmaps; 2035 2036 rd->head = rd->data_offset; 2037 data_size += rd->data_offset; 2038 2039 rd->mmap_size = MMAP_SIZE; 2040 if (rd->mmap_size > data_size) { 2041 rd->mmap_size = data_size; 2042 if (one_mmap) 2043 *one_mmap = true; 2044 } 2045 2046 memset(mmaps, 0, sizeof(rd->mmaps)); 2047 2048 if (zstd_init(&rd->zstd_data, 0)) 2049 return -1; 2050 rd->decomp_data.zstd_decomp = &rd->zstd_data; 2051 2052 return 0; 2053 } 2054 2055 static void 2056 reader__release_decomp(struct reader *rd) 2057 { 2058 perf_decomp__release_events(rd->decomp_data.decomp); 2059 zstd_fini(&rd->zstd_data); 2060 } 2061 2062 static int 2063 reader__mmap(struct reader *rd, struct perf_session *session) 2064 { 2065 int mmap_prot, mmap_flags; 2066 char *buf, **mmaps = rd->mmaps; 2067 u64 page_offset; 2068 2069 mmap_prot = PROT_READ; 2070 mmap_flags = MAP_SHARED; 2071 2072 if (rd->in_place_update) { 2073 mmap_prot |= PROT_WRITE; 2074 } else if (session->header.needs_swap) { 2075 mmap_prot |= PROT_WRITE; 2076 mmap_flags = MAP_PRIVATE; 2077 } 2078 2079 if (mmaps[rd->mmap_idx]) { 2080 munmap(mmaps[rd->mmap_idx], rd->mmap_size); 2081 mmaps[rd->mmap_idx] = NULL; 2082 } 2083 2084 page_offset = page_size * (rd->head / page_size); 2085 rd->file_offset += page_offset; 2086 rd->head -= page_offset; 2087 2088 buf = mmap(NULL, rd->mmap_size, mmap_prot, mmap_flags, rd->fd, 2089 rd->file_offset); 2090 if (buf == MAP_FAILED) { 2091 pr_err("failed to mmap file\n"); 2092 return -errno; 2093 } 2094 mmaps[rd->mmap_idx] = rd->mmap_cur = buf; 2095 rd->mmap_idx = (rd->mmap_idx + 1) & (ARRAY_SIZE(rd->mmaps) - 1); 2096 rd->file_pos = rd->file_offset + rd->head; 2097 if (session->one_mmap) { 2098 session->one_mmap_addr = buf; 2099 session->one_mmap_offset = rd->file_offset; 2100 } 2101 2102 return 0; 2103 } 2104 2105 enum { 2106 READER_OK, 2107 READER_NODATA, 2108 }; 2109 2110 static int 2111 reader__read_event(struct reader *rd, struct perf_session *session, 2112 struct ui_progress *prog) 2113 { 2114 u64 size; 2115 int err = READER_OK; 2116 union perf_event *event; 2117 s64 skip; 2118 2119 event = fetch_mmaped_event(rd->head, rd->mmap_size, rd->mmap_cur, 2120 session->header.needs_swap); 2121 if (IS_ERR(event)) 2122 return PTR_ERR(event); 2123 2124 if (!event) 2125 return READER_NODATA; 2126 2127 size = event->header.size; 2128 2129 skip = -EINVAL; 2130 2131 if (size < sizeof(struct perf_event_header) || 2132 (skip = rd->process(session, event, rd->file_pos, rd->path)) < 0) { 2133 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%s]\n", 2134 rd->file_offset + rd->head, event->header.size, 2135 event->header.type, strerror(-skip)); 2136 err = skip; 2137 goto out; 2138 } 2139 2140 if (skip) 2141 size += skip; 2142 2143 rd->size += size; 2144 rd->head += size; 2145 rd->file_pos += size; 2146 2147 err = __perf_session__process_decomp_events(session); 2148 if (err) 2149 goto out; 2150 2151 ui_progress__update(prog, size); 2152 2153 out: 2154 return err; 2155 } 2156 2157 static inline bool 2158 reader__eof(struct reader *rd) 2159 { 2160 return (rd->file_pos >= rd->data_size + rd->data_offset); 2161 } 2162 2163 static int 2164 reader__process_events(struct reader *rd, struct perf_session *session, 2165 struct ui_progress *prog) 2166 { 2167 int err; 2168 2169 err = reader__init(rd, &session->one_mmap); 2170 if (err) 2171 goto out; 2172 2173 session->active_decomp = &rd->decomp_data; 2174 2175 remap: 2176 err = reader__mmap(rd, session); 2177 if (err) 2178 goto out; 2179 2180 more: 2181 err = reader__read_event(rd, session, prog); 2182 if (err < 0) 2183 goto out; 2184 else if (err == READER_NODATA) 2185 goto remap; 2186 2187 if (session_done()) 2188 goto out; 2189 2190 if (!reader__eof(rd)) 2191 goto more; 2192 2193 out: 2194 session->active_decomp = &session->decomp_data; 2195 return err; 2196 } 2197 2198 static s64 process_simple(struct perf_session *session, 2199 union perf_event *event, 2200 u64 file_offset, 2201 const char *file_path) 2202 { 2203 return perf_session__process_event(session, event, file_offset, file_path); 2204 } 2205 2206 static int __perf_session__process_events(struct perf_session *session) 2207 { 2208 struct reader rd = { 2209 .fd = perf_data__fd(session->data), 2210 .path = session->data->file.path, 2211 .data_size = session->header.data_size, 2212 .data_offset = session->header.data_offset, 2213 .process = process_simple, 2214 .in_place_update = session->data->in_place_update, 2215 }; 2216 struct ordered_events *oe = &session->ordered_events; 2217 const struct perf_tool *tool = session->tool; 2218 struct ui_progress prog; 2219 int err; 2220 2221 if (rd.data_size == 0) 2222 return -1; 2223 2224 ui_progress__init_size(&prog, rd.data_size, "Processing events..."); 2225 2226 err = reader__process_events(&rd, session, &prog); 2227 if (err) 2228 goto out_err; 2229 /* do the final flush for ordered samples */ 2230 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2231 if (err) 2232 goto out_err; 2233 err = auxtrace__flush_events(session, tool); 2234 if (err) 2235 goto out_err; 2236 err = perf_session__flush_thread_stacks(session); 2237 out_err: 2238 ui_progress__finish(); 2239 if (!tool->no_warn) 2240 perf_session__warn_about_errors(session); 2241 /* 2242 * We may switching perf.data output, make ordered_events 2243 * reusable. 2244 */ 2245 ordered_events__reinit(&session->ordered_events); 2246 auxtrace__free_events(session); 2247 reader__release_decomp(&rd); 2248 session->one_mmap = false; 2249 return err; 2250 } 2251 2252 /* 2253 * Processing 2 MB of data from each reader in sequence, 2254 * because that's the way the ordered events sorting works 2255 * most efficiently. 2256 */ 2257 #define READER_MAX_SIZE (2 * 1024 * 1024) 2258 2259 /* 2260 * This function reads, merge and process directory data. 2261 * It assumens the version 1 of directory data, where each 2262 * data file holds per-cpu data, already sorted by kernel. 2263 */ 2264 static int __perf_session__process_dir_events(struct perf_session *session) 2265 { 2266 struct perf_data *data = session->data; 2267 const struct perf_tool *tool = session->tool; 2268 int i, ret, readers, nr_readers; 2269 struct ui_progress prog; 2270 u64 total_size = perf_data__size(session->data); 2271 struct reader *rd; 2272 2273 ui_progress__init_size(&prog, total_size, "Processing events..."); 2274 2275 nr_readers = 1; 2276 for (i = 0; i < data->dir.nr; i++) { 2277 if (data->dir.files[i].size) 2278 nr_readers++; 2279 } 2280 2281 rd = zalloc(nr_readers * sizeof(struct reader)); 2282 if (!rd) 2283 return -ENOMEM; 2284 2285 rd[0] = (struct reader) { 2286 .fd = perf_data__fd(session->data), 2287 .path = session->data->file.path, 2288 .data_size = session->header.data_size, 2289 .data_offset = session->header.data_offset, 2290 .process = process_simple, 2291 .in_place_update = session->data->in_place_update, 2292 }; 2293 ret = reader__init(&rd[0], NULL); 2294 if (ret) 2295 goto out_err; 2296 ret = reader__mmap(&rd[0], session); 2297 if (ret) 2298 goto out_err; 2299 readers = 1; 2300 2301 for (i = 0; i < data->dir.nr; i++) { 2302 if (!data->dir.files[i].size) 2303 continue; 2304 rd[readers] = (struct reader) { 2305 .fd = data->dir.files[i].fd, 2306 .path = data->dir.files[i].path, 2307 .data_size = data->dir.files[i].size, 2308 .data_offset = 0, 2309 .process = process_simple, 2310 .in_place_update = session->data->in_place_update, 2311 }; 2312 ret = reader__init(&rd[readers], NULL); 2313 if (ret) 2314 goto out_err; 2315 ret = reader__mmap(&rd[readers], session); 2316 if (ret) 2317 goto out_err; 2318 readers++; 2319 } 2320 2321 i = 0; 2322 while (readers) { 2323 if (session_done()) 2324 break; 2325 2326 if (rd[i].done) { 2327 i = (i + 1) % nr_readers; 2328 continue; 2329 } 2330 if (reader__eof(&rd[i])) { 2331 rd[i].done = true; 2332 readers--; 2333 continue; 2334 } 2335 2336 session->active_decomp = &rd[i].decomp_data; 2337 ret = reader__read_event(&rd[i], session, &prog); 2338 if (ret < 0) { 2339 goto out_err; 2340 } else if (ret == READER_NODATA) { 2341 ret = reader__mmap(&rd[i], session); 2342 if (ret) 2343 goto out_err; 2344 } 2345 2346 if (rd[i].size >= READER_MAX_SIZE) { 2347 rd[i].size = 0; 2348 i = (i + 1) % nr_readers; 2349 } 2350 } 2351 2352 ret = ordered_events__flush(&session->ordered_events, OE_FLUSH__FINAL); 2353 if (ret) 2354 goto out_err; 2355 2356 ret = perf_session__flush_thread_stacks(session); 2357 out_err: 2358 ui_progress__finish(); 2359 2360 if (!tool->no_warn) 2361 perf_session__warn_about_errors(session); 2362 2363 /* 2364 * We may switching perf.data output, make ordered_events 2365 * reusable. 2366 */ 2367 ordered_events__reinit(&session->ordered_events); 2368 2369 session->one_mmap = false; 2370 2371 session->active_decomp = &session->decomp_data; 2372 for (i = 0; i < nr_readers; i++) 2373 reader__release_decomp(&rd[i]); 2374 zfree(&rd); 2375 2376 return ret; 2377 } 2378 2379 int perf_session__process_events(struct perf_session *session) 2380 { 2381 if (perf_session__register_idle_thread(session) < 0) 2382 return -ENOMEM; 2383 2384 if (perf_data__is_pipe(session->data)) 2385 return __perf_session__process_pipe_events(session); 2386 2387 if (perf_data__is_dir(session->data) && session->data->dir.nr) 2388 return __perf_session__process_dir_events(session); 2389 2390 return __perf_session__process_events(session); 2391 } 2392 2393 bool perf_session__has_traces(struct perf_session *session, const char *msg) 2394 { 2395 struct evsel *evsel; 2396 2397 evlist__for_each_entry(session->evlist, evsel) { 2398 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT) 2399 return true; 2400 } 2401 2402 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 2403 return false; 2404 } 2405 2406 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr) 2407 { 2408 char *bracket; 2409 struct ref_reloc_sym *ref; 2410 struct kmap *kmap; 2411 2412 ref = zalloc(sizeof(struct ref_reloc_sym)); 2413 if (ref == NULL) 2414 return -ENOMEM; 2415 2416 ref->name = strdup(symbol_name); 2417 if (ref->name == NULL) { 2418 free(ref); 2419 return -ENOMEM; 2420 } 2421 2422 bracket = strchr(ref->name, ']'); 2423 if (bracket) 2424 *bracket = '\0'; 2425 2426 ref->addr = addr; 2427 2428 kmap = map__kmap(map); 2429 if (kmap) 2430 kmap->ref_reloc_sym = ref; 2431 2432 return 0; 2433 } 2434 2435 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp) 2436 { 2437 return machines__fprintf_dsos(&session->machines, fp); 2438 } 2439 2440 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp, 2441 bool (skip)(struct dso *dso, int parm), int parm) 2442 { 2443 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm); 2444 } 2445 2446 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 2447 { 2448 size_t ret; 2449 const char *msg = ""; 2450 2451 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) 2452 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)"; 2453 2454 ret = fprintf(fp, "\nAggregated stats:%s\n", msg); 2455 2456 ret += events_stats__fprintf(&session->evlist->stats, fp); 2457 return ret; 2458 } 2459 2460 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 2461 { 2462 /* 2463 * FIXME: Here we have to actually print all the machines in this 2464 * session, not just the host... 2465 */ 2466 return machine__fprintf(&session->machines.host, fp); 2467 } 2468 2469 void perf_session__dump_kmaps(struct perf_session *session) 2470 { 2471 int save_verbose = verbose; 2472 2473 fflush(stdout); 2474 fprintf(stderr, "Kernel and module maps:\n"); 2475 verbose = 0; /* Suppress verbose to print a summary only */ 2476 maps__fprintf(machine__kernel_maps(&session->machines.host), stderr); 2477 verbose = save_verbose; 2478 } 2479 2480 struct evsel *perf_session__find_first_evtype(struct perf_session *session, 2481 unsigned int type) 2482 { 2483 struct evsel *pos; 2484 2485 evlist__for_each_entry(session->evlist, pos) { 2486 if (pos->core.attr.type == type) 2487 return pos; 2488 } 2489 return NULL; 2490 } 2491 2492 int perf_session__cpu_bitmap(struct perf_session *session, 2493 const char *cpu_list, unsigned long *cpu_bitmap) 2494 { 2495 int i, err = -1; 2496 struct perf_cpu_map *map; 2497 int nr_cpus = min(session->header.env.nr_cpus_avail, MAX_NR_CPUS); 2498 struct perf_cpu cpu; 2499 2500 for (i = 0; i < PERF_TYPE_MAX; ++i) { 2501 struct evsel *evsel; 2502 2503 evsel = perf_session__find_first_evtype(session, i); 2504 if (!evsel) 2505 continue; 2506 2507 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) { 2508 pr_err("File does not contain CPU events. " 2509 "Remove -C option to proceed.\n"); 2510 return -1; 2511 } 2512 } 2513 2514 map = perf_cpu_map__new(cpu_list); 2515 if (map == NULL) { 2516 pr_err("Invalid cpu_list\n"); 2517 return -1; 2518 } 2519 2520 perf_cpu_map__for_each_cpu(cpu, i, map) { 2521 if (cpu.cpu >= nr_cpus) { 2522 pr_err("Requested CPU %d too large. " 2523 "Consider raising MAX_NR_CPUS\n", cpu.cpu); 2524 goto out_delete_map; 2525 } 2526 2527 __set_bit(cpu.cpu, cpu_bitmap); 2528 } 2529 2530 err = 0; 2531 2532 out_delete_map: 2533 perf_cpu_map__put(map); 2534 return err; 2535 } 2536 2537 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 2538 bool full) 2539 { 2540 if (session == NULL || fp == NULL) 2541 return; 2542 2543 fprintf(fp, "# ========\n"); 2544 perf_header__fprintf_info(session, fp, full); 2545 fprintf(fp, "# ========\n#\n"); 2546 } 2547 2548 static int perf_session__register_guest(struct perf_session *session, pid_t machine_pid) 2549 { 2550 struct machine *machine = machines__findnew(&session->machines, machine_pid); 2551 struct thread *thread; 2552 2553 if (!machine) 2554 return -ENOMEM; 2555 2556 machine->single_address_space = session->machines.host.single_address_space; 2557 2558 thread = machine__idle_thread(machine); 2559 if (!thread) 2560 return -ENOMEM; 2561 thread__put(thread); 2562 2563 machine->kallsyms_filename = perf_data__guest_kallsyms_name(session->data, machine_pid); 2564 2565 return 0; 2566 } 2567 2568 static int perf_session__set_guest_cpu(struct perf_session *session, pid_t pid, 2569 pid_t tid, int guest_cpu) 2570 { 2571 struct machine *machine = &session->machines.host; 2572 struct thread *thread = machine__findnew_thread(machine, pid, tid); 2573 2574 if (!thread) 2575 return -ENOMEM; 2576 thread__set_guest_cpu(thread, guest_cpu); 2577 thread__put(thread); 2578 2579 return 0; 2580 } 2581 2582 int perf_event__process_id_index(struct perf_session *session, 2583 union perf_event *event) 2584 { 2585 struct evlist *evlist = session->evlist; 2586 struct perf_record_id_index *ie = &event->id_index; 2587 size_t sz = ie->header.size - sizeof(*ie); 2588 size_t i, nr, max_nr; 2589 size_t e1_sz = sizeof(struct id_index_entry); 2590 size_t e2_sz = sizeof(struct id_index_entry_2); 2591 size_t etot_sz = e1_sz + e2_sz; 2592 struct id_index_entry_2 *e2; 2593 pid_t last_pid = 0; 2594 2595 max_nr = sz / e1_sz; 2596 nr = ie->nr; 2597 if (nr > max_nr) { 2598 printf("Too big: nr %zu max_nr %zu\n", nr, max_nr); 2599 return -EINVAL; 2600 } 2601 2602 if (sz >= nr * etot_sz) { 2603 max_nr = sz / etot_sz; 2604 if (nr > max_nr) { 2605 printf("Too big2: nr %zu max_nr %zu\n", nr, max_nr); 2606 return -EINVAL; 2607 } 2608 e2 = (void *)ie + sizeof(*ie) + nr * e1_sz; 2609 } else { 2610 e2 = NULL; 2611 } 2612 2613 if (dump_trace) 2614 fprintf(stdout, " nr: %zu\n", nr); 2615 2616 for (i = 0; i < nr; i++, (e2 ? e2++ : 0)) { 2617 struct id_index_entry *e = &ie->entries[i]; 2618 struct perf_sample_id *sid; 2619 int ret; 2620 2621 if (dump_trace) { 2622 fprintf(stdout, " ... id: %"PRI_lu64, e->id); 2623 fprintf(stdout, " idx: %"PRI_lu64, e->idx); 2624 fprintf(stdout, " cpu: %"PRI_ld64, e->cpu); 2625 fprintf(stdout, " tid: %"PRI_ld64, e->tid); 2626 if (e2) { 2627 fprintf(stdout, " machine_pid: %"PRI_ld64, e2->machine_pid); 2628 fprintf(stdout, " vcpu: %"PRI_lu64"\n", e2->vcpu); 2629 } else { 2630 fprintf(stdout, "\n"); 2631 } 2632 } 2633 2634 sid = evlist__id2sid(evlist, e->id); 2635 if (!sid) 2636 return -ENOENT; 2637 2638 sid->idx = e->idx; 2639 sid->cpu.cpu = e->cpu; 2640 sid->tid = e->tid; 2641 2642 if (!e2) 2643 continue; 2644 2645 sid->machine_pid = e2->machine_pid; 2646 sid->vcpu.cpu = e2->vcpu; 2647 2648 if (!sid->machine_pid) 2649 continue; 2650 2651 if (sid->machine_pid != last_pid) { 2652 ret = perf_session__register_guest(session, sid->machine_pid); 2653 if (ret) 2654 return ret; 2655 last_pid = sid->machine_pid; 2656 perf_guest = true; 2657 } 2658 2659 ret = perf_session__set_guest_cpu(session, sid->machine_pid, e->tid, e2->vcpu); 2660 if (ret) 2661 return ret; 2662 } 2663 return 0; 2664 } 2665 2666 int perf_session__dsos_hit_all(struct perf_session *session) 2667 { 2668 struct rb_node *nd; 2669 int err; 2670 2671 err = machine__hit_all_dsos(&session->machines.host); 2672 if (err) 2673 return err; 2674 2675 for (nd = rb_first_cached(&session->machines.guests); nd; 2676 nd = rb_next(nd)) { 2677 struct machine *pos = rb_entry(nd, struct machine, rb_node); 2678 2679 err = machine__hit_all_dsos(pos); 2680 if (err) 2681 return err; 2682 } 2683 2684 return 0; 2685 } 2686