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 #include <perf/event.h> 16 17 #include "map_symbol.h" 18 #include "branch.h" 19 #include "debug.h" 20 #include "dwarf-regs.h" 21 #include "env.h" 22 #include "evlist.h" 23 #include "evsel.h" 24 #include "memswap.h" 25 #include "map.h" 26 #include "symbol.h" 27 #include "session.h" 28 #include "tool.h" 29 #include "perf_regs.h" 30 #include "asm/bug.h" 31 #include "auxtrace.h" 32 #include "thread.h" 33 #include "thread-stack.h" 34 #include "sample-raw.h" 35 #include "stat.h" 36 #include "tsc.h" 37 #include "ui/progress.h" 38 #include "util.h" 39 #include "arch/common.h" 40 #include "units.h" 41 #include "annotate.h" 42 #include "perf.h" 43 #include <internal/lib.h> 44 45 static int perf_session__deliver_event(struct perf_session *session, 46 union perf_event *event, 47 const struct perf_tool *tool, 48 u64 file_offset, 49 const char *file_path); 50 51 static int perf_session__open(struct perf_session *session) 52 { 53 struct perf_data *data = session->data; 54 55 if (perf_session__read_header(session) < 0) { 56 pr_err("incompatible file format (rerun with -v to learn more)\n"); 57 return -1; 58 } 59 60 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) { 61 /* Auxiliary events may reference exited threads, hold onto dead ones. */ 62 symbol_conf.keep_exited_threads = true; 63 } 64 65 if (perf_data__is_pipe(data)) 66 return 0; 67 68 if (perf_header__has_feat(&session->header, HEADER_STAT)) 69 return 0; 70 71 if (!evlist__valid_sample_type(session->evlist)) { 72 pr_err("non matching sample_type\n"); 73 return -1; 74 } 75 76 if (!evlist__valid_sample_id_all(session->evlist)) { 77 pr_err("non matching sample_id_all\n"); 78 return -1; 79 } 80 81 if (!evlist__valid_read_format(session->evlist)) { 82 pr_err("non matching read_format\n"); 83 return -1; 84 } 85 86 return 0; 87 } 88 89 void perf_session__set_id_hdr_size(struct perf_session *session) 90 { 91 u16 id_hdr_size = evlist__id_hdr_size(session->evlist); 92 93 machines__set_id_hdr_size(&session->machines, id_hdr_size); 94 } 95 96 int perf_session__create_kernel_maps(struct perf_session *session) 97 { 98 int ret = machine__create_kernel_maps(&session->machines.host); 99 100 if (ret >= 0) 101 ret = machines__create_guest_kernel_maps(&session->machines); 102 return ret; 103 } 104 105 static void perf_session__destroy_kernel_maps(struct perf_session *session) 106 { 107 machines__destroy_kernel_maps(&session->machines); 108 } 109 110 static bool perf_session__has_comm_exec(struct perf_session *session) 111 { 112 struct evsel *evsel; 113 114 evlist__for_each_entry(session->evlist, evsel) { 115 if (evsel->core.attr.comm_exec) 116 return true; 117 } 118 119 return false; 120 } 121 122 static void perf_session__set_comm_exec(struct perf_session *session) 123 { 124 bool comm_exec = perf_session__has_comm_exec(session); 125 126 machines__set_comm_exec(&session->machines, comm_exec); 127 } 128 129 static int ordered_events__deliver_event(struct ordered_events *oe, 130 struct ordered_event *event) 131 { 132 struct perf_session *session = container_of(oe, struct perf_session, 133 ordered_events); 134 135 return perf_session__deliver_event(session, event->event, 136 session->tool, event->file_offset, 137 event->file_path); 138 } 139 140 struct perf_session *__perf_session__new(struct perf_data *data, 141 struct perf_tool *tool, 142 bool trace_event_repipe, 143 struct perf_env *host_env) 144 { 145 int ret = -ENOMEM; 146 struct perf_session *session = zalloc(sizeof(*session)); 147 148 if (!session) 149 goto out; 150 151 session->trace_event_repipe = trace_event_repipe; 152 session->tool = tool; 153 session->decomp_data.zstd_decomp = &session->zstd_data; 154 session->active_decomp = &session->decomp_data; 155 INIT_LIST_HEAD(&session->auxtrace_index); 156 machines__init(&session->machines); 157 ordered_events__init(&session->ordered_events, 158 ordered_events__deliver_event, NULL); 159 160 perf_env__init(&session->header.env); 161 if (data) { 162 ret = perf_data__open(data); 163 if (ret < 0) 164 goto out_delete; 165 166 session->data = data; 167 168 if (perf_data__is_read(data)) { 169 ret = perf_session__open(session); 170 if (ret < 0) 171 goto out_delete; 172 173 /* 174 * set session attributes that are present in perf.data 175 * but not in pipe-mode. 176 */ 177 if (!data->is_pipe) { 178 perf_session__set_id_hdr_size(session); 179 perf_session__set_comm_exec(session); 180 } 181 182 evlist__init_trace_event_sample_raw(session->evlist, &session->header.env); 183 184 /* Open the directory data. */ 185 if (data->is_dir) { 186 ret = perf_data__open_dir(data); 187 if (ret) 188 goto out_delete; 189 } 190 191 if (!symbol_conf.kallsyms_name && 192 !symbol_conf.vmlinux_name) 193 symbol_conf.kallsyms_name = perf_data__kallsyms_name(data); 194 } 195 } else { 196 assert(host_env != NULL); 197 session->machines.host.env = host_env; 198 } 199 if (session->evlist) 200 session->evlist->session = session; 201 202 session->machines.host.single_address_space = 203 perf_env__single_address_space(session->machines.host.env); 204 205 if (!data || perf_data__is_write(data)) { 206 /* 207 * In O_RDONLY mode this will be performed when reading the 208 * kernel MMAP event, in perf_event__process_mmap(). 209 */ 210 if (perf_session__create_kernel_maps(session) < 0) 211 pr_warning("Cannot read kernel map\n"); 212 } 213 214 /* 215 * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is 216 * processed, so evlist__sample_id_all is not meaningful here. 217 */ 218 if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps && 219 tool->ordered_events && !evlist__sample_id_all(session->evlist)) { 220 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 221 tool->ordered_events = false; 222 } 223 224 return session; 225 226 out_delete: 227 perf_session__delete(session); 228 out: 229 return ERR_PTR(ret); 230 } 231 232 static void perf_decomp__release_events(struct decomp *next) 233 { 234 struct decomp *decomp; 235 size_t mmap_len; 236 237 do { 238 decomp = next; 239 if (decomp == NULL) 240 break; 241 next = decomp->next; 242 mmap_len = decomp->mmap_len; 243 munmap(decomp, mmap_len); 244 } while (1); 245 } 246 247 void perf_session__delete(struct perf_session *session) 248 { 249 if (session == NULL) 250 return; 251 auxtrace__free(session); 252 auxtrace_index__free(&session->auxtrace_index); 253 debuginfo_cache__delete(); 254 perf_session__destroy_kernel_maps(session); 255 perf_decomp__release_events(session->decomp_data.decomp); 256 perf_env__exit(&session->header.env); 257 machines__exit(&session->machines); 258 if (session->data) { 259 if (perf_data__is_read(session->data)) 260 evlist__delete(session->evlist); 261 perf_data__close(session->data); 262 } 263 #ifdef HAVE_LIBTRACEEVENT 264 trace_event__cleanup(&session->tevent); 265 #endif 266 free(session); 267 } 268 269 static void swap_sample_id_all(union perf_event *event, void *data) 270 { 271 void *end = (void *) event + event->header.size; 272 int size = end - data; 273 274 BUG_ON(size % sizeof(u64)); 275 mem_bswap_64(data, size); 276 } 277 278 static void perf_event__all64_swap(union perf_event *event, 279 bool sample_id_all __maybe_unused) 280 { 281 struct perf_event_header *hdr = &event->header; 282 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 283 } 284 285 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all) 286 { 287 event->comm.pid = bswap_32(event->comm.pid); 288 event->comm.tid = bswap_32(event->comm.tid); 289 290 if (sample_id_all) { 291 void *data = &event->comm.comm; 292 293 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 294 swap_sample_id_all(event, data); 295 } 296 } 297 298 static void perf_event__mmap_swap(union perf_event *event, 299 bool sample_id_all) 300 { 301 event->mmap.pid = bswap_32(event->mmap.pid); 302 event->mmap.tid = bswap_32(event->mmap.tid); 303 event->mmap.start = bswap_64(event->mmap.start); 304 event->mmap.len = bswap_64(event->mmap.len); 305 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 306 307 if (sample_id_all) { 308 void *data = &event->mmap.filename; 309 310 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 311 swap_sample_id_all(event, data); 312 } 313 } 314 315 static void perf_event__mmap2_swap(union perf_event *event, 316 bool sample_id_all) 317 { 318 event->mmap2.pid = bswap_32(event->mmap2.pid); 319 event->mmap2.tid = bswap_32(event->mmap2.tid); 320 event->mmap2.start = bswap_64(event->mmap2.start); 321 event->mmap2.len = bswap_64(event->mmap2.len); 322 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff); 323 324 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID)) { 325 event->mmap2.maj = bswap_32(event->mmap2.maj); 326 event->mmap2.min = bswap_32(event->mmap2.min); 327 event->mmap2.ino = bswap_64(event->mmap2.ino); 328 event->mmap2.ino_generation = bswap_64(event->mmap2.ino_generation); 329 } 330 331 if (sample_id_all) { 332 void *data = &event->mmap2.filename; 333 334 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 335 swap_sample_id_all(event, data); 336 } 337 } 338 static void perf_event__task_swap(union perf_event *event, bool sample_id_all) 339 { 340 event->fork.pid = bswap_32(event->fork.pid); 341 event->fork.tid = bswap_32(event->fork.tid); 342 event->fork.ppid = bswap_32(event->fork.ppid); 343 event->fork.ptid = bswap_32(event->fork.ptid); 344 event->fork.time = bswap_64(event->fork.time); 345 346 if (sample_id_all) 347 swap_sample_id_all(event, &event->fork + 1); 348 } 349 350 static void perf_event__read_swap(union perf_event *event, bool sample_id_all) 351 { 352 event->read.pid = bswap_32(event->read.pid); 353 event->read.tid = bswap_32(event->read.tid); 354 event->read.value = bswap_64(event->read.value); 355 event->read.time_enabled = bswap_64(event->read.time_enabled); 356 event->read.time_running = bswap_64(event->read.time_running); 357 event->read.id = bswap_64(event->read.id); 358 359 if (sample_id_all) 360 swap_sample_id_all(event, &event->read + 1); 361 } 362 363 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all) 364 { 365 event->aux.aux_offset = bswap_64(event->aux.aux_offset); 366 event->aux.aux_size = bswap_64(event->aux.aux_size); 367 event->aux.flags = bswap_64(event->aux.flags); 368 369 if (sample_id_all) 370 swap_sample_id_all(event, &event->aux + 1); 371 } 372 373 static void perf_event__itrace_start_swap(union perf_event *event, 374 bool sample_id_all) 375 { 376 event->itrace_start.pid = bswap_32(event->itrace_start.pid); 377 event->itrace_start.tid = bswap_32(event->itrace_start.tid); 378 379 if (sample_id_all) 380 swap_sample_id_all(event, &event->itrace_start + 1); 381 } 382 383 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all) 384 { 385 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) { 386 event->context_switch.next_prev_pid = 387 bswap_32(event->context_switch.next_prev_pid); 388 event->context_switch.next_prev_tid = 389 bswap_32(event->context_switch.next_prev_tid); 390 } 391 392 if (sample_id_all) 393 swap_sample_id_all(event, &event->context_switch + 1); 394 } 395 396 static void perf_event__text_poke_swap(union perf_event *event, bool sample_id_all) 397 { 398 event->text_poke.addr = bswap_64(event->text_poke.addr); 399 event->text_poke.old_len = bswap_16(event->text_poke.old_len); 400 event->text_poke.new_len = bswap_16(event->text_poke.new_len); 401 402 if (sample_id_all) { 403 size_t len = sizeof(event->text_poke.old_len) + 404 sizeof(event->text_poke.new_len) + 405 event->text_poke.old_len + 406 event->text_poke.new_len; 407 void *data = &event->text_poke.old_len; 408 409 data += PERF_ALIGN(len, sizeof(u64)); 410 swap_sample_id_all(event, data); 411 } 412 } 413 414 static void perf_event__throttle_swap(union perf_event *event, 415 bool sample_id_all) 416 { 417 event->throttle.time = bswap_64(event->throttle.time); 418 event->throttle.id = bswap_64(event->throttle.id); 419 event->throttle.stream_id = bswap_64(event->throttle.stream_id); 420 421 if (sample_id_all) 422 swap_sample_id_all(event, &event->throttle + 1); 423 } 424 425 static void perf_event__namespaces_swap(union perf_event *event, 426 bool sample_id_all) 427 { 428 u64 i; 429 430 event->namespaces.pid = bswap_32(event->namespaces.pid); 431 event->namespaces.tid = bswap_32(event->namespaces.tid); 432 event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces); 433 434 for (i = 0; i < event->namespaces.nr_namespaces; i++) { 435 struct perf_ns_link_info *ns = &event->namespaces.link_info[i]; 436 437 ns->dev = bswap_64(ns->dev); 438 ns->ino = bswap_64(ns->ino); 439 } 440 441 if (sample_id_all) 442 swap_sample_id_all(event, &event->namespaces.link_info[i]); 443 } 444 445 static void perf_event__cgroup_swap(union perf_event *event, bool sample_id_all) 446 { 447 event->cgroup.id = bswap_64(event->cgroup.id); 448 449 if (sample_id_all) { 450 void *data = &event->cgroup.path; 451 452 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 453 swap_sample_id_all(event, data); 454 } 455 } 456 457 static u8 revbyte(u8 b) 458 { 459 int rev = (b >> 4) | ((b & 0xf) << 4); 460 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2); 461 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1); 462 return (u8) rev; 463 } 464 465 /* 466 * XXX this is hack in attempt to carry flags bitfield 467 * through endian village. ABI says: 468 * 469 * Bit-fields are allocated from right to left (least to most significant) 470 * on little-endian implementations and from left to right (most to least 471 * significant) on big-endian implementations. 472 * 473 * The above seems to be byte specific, so we need to reverse each 474 * byte of the bitfield. 'Internet' also says this might be implementation 475 * specific and we probably need proper fix and carry perf_event_attr 476 * bitfield flags in separate data file FEAT_ section. Thought this seems 477 * to work for now. 478 */ 479 static void swap_bitfield(u8 *p, unsigned len) 480 { 481 unsigned i; 482 483 for (i = 0; i < len; i++) { 484 *p = revbyte(*p); 485 p++; 486 } 487 } 488 489 /* exported for swapping attributes in file header */ 490 void perf_event__attr_swap(struct perf_event_attr *attr) 491 { 492 attr->type = bswap_32(attr->type); 493 attr->size = bswap_32(attr->size); 494 495 #define bswap_safe(f, n) \ 496 (attr->size > (offsetof(struct perf_event_attr, f) + \ 497 sizeof(attr->f) * (n))) 498 #define bswap_field(f, sz) \ 499 do { \ 500 if (bswap_safe(f, 0)) \ 501 attr->f = bswap_##sz(attr->f); \ 502 } while(0) 503 #define bswap_field_16(f) bswap_field(f, 16) 504 #define bswap_field_32(f) bswap_field(f, 32) 505 #define bswap_field_64(f) bswap_field(f, 64) 506 507 bswap_field_64(config); 508 bswap_field_64(sample_period); 509 bswap_field_64(sample_type); 510 bswap_field_64(read_format); 511 bswap_field_32(wakeup_events); 512 bswap_field_32(bp_type); 513 bswap_field_64(bp_addr); 514 bswap_field_64(bp_len); 515 bswap_field_64(branch_sample_type); 516 bswap_field_64(sample_regs_user); 517 bswap_field_32(sample_stack_user); 518 bswap_field_32(aux_watermark); 519 bswap_field_16(sample_max_stack); 520 bswap_field_32(aux_sample_size); 521 522 /* 523 * After read_format are bitfields. Check read_format because 524 * we are unable to use offsetof on bitfield. 525 */ 526 if (bswap_safe(read_format, 1)) 527 swap_bitfield((u8 *) (&attr->read_format + 1), 528 sizeof(u64)); 529 #undef bswap_field_64 530 #undef bswap_field_32 531 #undef bswap_field 532 #undef bswap_safe 533 } 534 535 static void perf_event__hdr_attr_swap(union perf_event *event, 536 bool sample_id_all __maybe_unused) 537 { 538 size_t size; 539 540 perf_event__attr_swap(&event->attr.attr); 541 542 size = event->header.size; 543 size -= perf_record_header_attr_id(event) - (void *)event; 544 mem_bswap_64(perf_record_header_attr_id(event), size); 545 } 546 547 static void perf_event__event_update_swap(union perf_event *event, 548 bool sample_id_all __maybe_unused) 549 { 550 event->event_update.type = bswap_64(event->event_update.type); 551 event->event_update.id = bswap_64(event->event_update.id); 552 } 553 554 static void perf_event__event_type_swap(union perf_event *event, 555 bool sample_id_all __maybe_unused) 556 { 557 event->event_type.event_type.event_id = 558 bswap_64(event->event_type.event_type.event_id); 559 } 560 561 static void perf_event__tracing_data_swap(union perf_event *event, 562 bool sample_id_all __maybe_unused) 563 { 564 event->tracing_data.size = bswap_32(event->tracing_data.size); 565 } 566 567 static void perf_event__auxtrace_info_swap(union perf_event *event, 568 bool sample_id_all __maybe_unused) 569 { 570 size_t size; 571 572 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type); 573 574 size = event->header.size; 575 size -= (void *)&event->auxtrace_info.priv - (void *)event; 576 mem_bswap_64(event->auxtrace_info.priv, size); 577 } 578 579 static void perf_event__auxtrace_swap(union perf_event *event, 580 bool sample_id_all __maybe_unused) 581 { 582 event->auxtrace.size = bswap_64(event->auxtrace.size); 583 event->auxtrace.offset = bswap_64(event->auxtrace.offset); 584 event->auxtrace.reference = bswap_64(event->auxtrace.reference); 585 event->auxtrace.idx = bswap_32(event->auxtrace.idx); 586 event->auxtrace.tid = bswap_32(event->auxtrace.tid); 587 event->auxtrace.cpu = bswap_32(event->auxtrace.cpu); 588 } 589 590 static void perf_event__auxtrace_error_swap(union perf_event *event, 591 bool sample_id_all __maybe_unused) 592 { 593 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type); 594 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code); 595 event->auxtrace_error.cpu = bswap_32(event->auxtrace_error.cpu); 596 event->auxtrace_error.pid = bswap_32(event->auxtrace_error.pid); 597 event->auxtrace_error.tid = bswap_32(event->auxtrace_error.tid); 598 event->auxtrace_error.fmt = bswap_32(event->auxtrace_error.fmt); 599 event->auxtrace_error.ip = bswap_64(event->auxtrace_error.ip); 600 if (event->auxtrace_error.fmt) 601 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time); 602 if (event->auxtrace_error.fmt >= 2) { 603 event->auxtrace_error.machine_pid = bswap_32(event->auxtrace_error.machine_pid); 604 event->auxtrace_error.vcpu = bswap_32(event->auxtrace_error.vcpu); 605 } 606 } 607 608 static void perf_event__thread_map_swap(union perf_event *event, 609 bool sample_id_all __maybe_unused) 610 { 611 unsigned i; 612 613 event->thread_map.nr = bswap_64(event->thread_map.nr); 614 615 for (i = 0; i < event->thread_map.nr; i++) 616 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid); 617 } 618 619 static void perf_event__cpu_map_swap(union perf_event *event, 620 bool sample_id_all __maybe_unused) 621 { 622 struct perf_record_cpu_map_data *data = &event->cpu_map.data; 623 624 data->type = bswap_16(data->type); 625 626 switch (data->type) { 627 case PERF_CPU_MAP__CPUS: 628 data->cpus_data.nr = bswap_16(data->cpus_data.nr); 629 630 for (unsigned i = 0; i < data->cpus_data.nr; i++) 631 data->cpus_data.cpu[i] = bswap_16(data->cpus_data.cpu[i]); 632 break; 633 case PERF_CPU_MAP__MASK: 634 data->mask32_data.long_size = bswap_16(data->mask32_data.long_size); 635 636 switch (data->mask32_data.long_size) { 637 case 4: 638 data->mask32_data.nr = bswap_16(data->mask32_data.nr); 639 for (unsigned i = 0; i < data->mask32_data.nr; i++) 640 data->mask32_data.mask[i] = bswap_32(data->mask32_data.mask[i]); 641 break; 642 case 8: 643 data->mask64_data.nr = bswap_16(data->mask64_data.nr); 644 for (unsigned i = 0; i < data->mask64_data.nr; i++) 645 data->mask64_data.mask[i] = bswap_64(data->mask64_data.mask[i]); 646 break; 647 default: 648 pr_err("cpu_map swap: unsupported long size\n"); 649 } 650 break; 651 case PERF_CPU_MAP__RANGE_CPUS: 652 data->range_cpu_data.start_cpu = bswap_16(data->range_cpu_data.start_cpu); 653 data->range_cpu_data.end_cpu = bswap_16(data->range_cpu_data.end_cpu); 654 break; 655 default: 656 break; 657 } 658 } 659 660 static void perf_event__stat_config_swap(union perf_event *event, 661 bool sample_id_all __maybe_unused) 662 { 663 u64 size; 664 665 size = bswap_64(event->stat_config.nr) * sizeof(event->stat_config.data[0]); 666 size += 1; /* nr item itself */ 667 mem_bswap_64(&event->stat_config.nr, size); 668 } 669 670 static void perf_event__stat_swap(union perf_event *event, 671 bool sample_id_all __maybe_unused) 672 { 673 event->stat.id = bswap_64(event->stat.id); 674 event->stat.thread = bswap_32(event->stat.thread); 675 event->stat.cpu = bswap_32(event->stat.cpu); 676 event->stat.val = bswap_64(event->stat.val); 677 event->stat.ena = bswap_64(event->stat.ena); 678 event->stat.run = bswap_64(event->stat.run); 679 } 680 681 static void perf_event__stat_round_swap(union perf_event *event, 682 bool sample_id_all __maybe_unused) 683 { 684 event->stat_round.type = bswap_64(event->stat_round.type); 685 event->stat_round.time = bswap_64(event->stat_round.time); 686 } 687 688 static void perf_event__time_conv_swap(union perf_event *event, 689 bool sample_id_all __maybe_unused) 690 { 691 event->time_conv.time_shift = bswap_64(event->time_conv.time_shift); 692 event->time_conv.time_mult = bswap_64(event->time_conv.time_mult); 693 event->time_conv.time_zero = bswap_64(event->time_conv.time_zero); 694 695 if (event_contains(event->time_conv, time_cycles)) { 696 event->time_conv.time_cycles = bswap_64(event->time_conv.time_cycles); 697 event->time_conv.time_mask = bswap_64(event->time_conv.time_mask); 698 } 699 } 700 701 typedef void (*perf_event__swap_op)(union perf_event *event, 702 bool sample_id_all); 703 704 static perf_event__swap_op perf_event__swap_ops[] = { 705 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 706 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap, 707 [PERF_RECORD_COMM] = perf_event__comm_swap, 708 [PERF_RECORD_FORK] = perf_event__task_swap, 709 [PERF_RECORD_EXIT] = perf_event__task_swap, 710 [PERF_RECORD_LOST] = perf_event__all64_swap, 711 [PERF_RECORD_READ] = perf_event__read_swap, 712 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap, 713 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap, 714 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 715 [PERF_RECORD_AUX] = perf_event__aux_swap, 716 [PERF_RECORD_ITRACE_START] = perf_event__itrace_start_swap, 717 [PERF_RECORD_LOST_SAMPLES] = perf_event__all64_swap, 718 [PERF_RECORD_SWITCH] = perf_event__switch_swap, 719 [PERF_RECORD_SWITCH_CPU_WIDE] = perf_event__switch_swap, 720 [PERF_RECORD_NAMESPACES] = perf_event__namespaces_swap, 721 [PERF_RECORD_CGROUP] = perf_event__cgroup_swap, 722 [PERF_RECORD_TEXT_POKE] = perf_event__text_poke_swap, 723 [PERF_RECORD_AUX_OUTPUT_HW_ID] = perf_event__all64_swap, 724 [PERF_RECORD_CALLCHAIN_DEFERRED] = perf_event__all64_swap, 725 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 726 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 727 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 728 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 729 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap, 730 [PERF_RECORD_AUXTRACE_INFO] = perf_event__auxtrace_info_swap, 731 [PERF_RECORD_AUXTRACE] = perf_event__auxtrace_swap, 732 [PERF_RECORD_AUXTRACE_ERROR] = perf_event__auxtrace_error_swap, 733 [PERF_RECORD_THREAD_MAP] = perf_event__thread_map_swap, 734 [PERF_RECORD_CPU_MAP] = perf_event__cpu_map_swap, 735 [PERF_RECORD_STAT_CONFIG] = perf_event__stat_config_swap, 736 [PERF_RECORD_STAT] = perf_event__stat_swap, 737 [PERF_RECORD_STAT_ROUND] = perf_event__stat_round_swap, 738 [PERF_RECORD_EVENT_UPDATE] = perf_event__event_update_swap, 739 [PERF_RECORD_TIME_CONV] = perf_event__time_conv_swap, 740 [PERF_RECORD_HEADER_MAX] = NULL, 741 }; 742 743 /* 744 * When perf record finishes a pass on every buffers, it records this pseudo 745 * event. 746 * We record the max timestamp t found in the pass n. 747 * Assuming these timestamps are monotonic across cpus, we know that if 748 * a buffer still has events with timestamps below t, they will be all 749 * available and then read in the pass n + 1. 750 * Hence when we start to read the pass n + 2, we can safely flush every 751 * events with timestamps below t. 752 * 753 * ============ PASS n ================= 754 * CPU 0 | CPU 1 755 * | 756 * cnt1 timestamps | cnt2 timestamps 757 * 1 | 2 758 * 2 | 3 759 * - | 4 <--- max recorded 760 * 761 * ============ PASS n + 1 ============== 762 * CPU 0 | CPU 1 763 * | 764 * cnt1 timestamps | cnt2 timestamps 765 * 3 | 5 766 * 4 | 6 767 * 5 | 7 <---- max recorded 768 * 769 * Flush every events below timestamp 4 770 * 771 * ============ PASS n + 2 ============== 772 * CPU 0 | CPU 1 773 * | 774 * cnt1 timestamps | cnt2 timestamps 775 * 6 | 8 776 * 7 | 9 777 * - | 10 778 * 779 * Flush every events below timestamp 7 780 * etc... 781 */ 782 int perf_event__process_finished_round(const struct perf_tool *tool __maybe_unused, 783 union perf_event *event __maybe_unused, 784 struct ordered_events *oe) 785 { 786 if (dump_trace) 787 fprintf(stdout, "\n"); 788 return ordered_events__flush(oe, OE_FLUSH__ROUND); 789 } 790 791 int perf_session__queue_event(struct perf_session *s, union perf_event *event, 792 u64 timestamp, u64 file_offset, const char *file_path) 793 { 794 return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset, file_path); 795 } 796 797 static void callchain__lbr_callstack_printf(struct perf_sample *sample) 798 { 799 struct ip_callchain *callchain = sample->callchain; 800 struct branch_stack *lbr_stack = sample->branch_stack; 801 struct branch_entry *entries = perf_sample__branch_entries(sample); 802 u64 kernel_callchain_nr = callchain->nr; 803 unsigned int i; 804 805 for (i = 0; i < kernel_callchain_nr; i++) { 806 if (callchain->ips[i] == PERF_CONTEXT_USER) 807 break; 808 } 809 810 if ((i != kernel_callchain_nr) && lbr_stack->nr) { 811 u64 total_nr; 812 /* 813 * LBR callstack can only get user call chain, 814 * i is kernel call chain number, 815 * 1 is PERF_CONTEXT_USER. 816 * 817 * The user call chain is stored in LBR registers. 818 * LBR are pair registers. The caller is stored 819 * in "from" register, while the callee is stored 820 * in "to" register. 821 * For example, there is a call stack 822 * "A"->"B"->"C"->"D". 823 * The LBR registers will be recorded like 824 * "C"->"D", "B"->"C", "A"->"B". 825 * So only the first "to" register and all "from" 826 * registers are needed to construct the whole stack. 827 */ 828 total_nr = i + 1 + lbr_stack->nr + 1; 829 kernel_callchain_nr = i + 1; 830 831 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr); 832 833 for (i = 0; i < kernel_callchain_nr; i++) 834 printf("..... %2d: %016" PRIx64 "\n", 835 i, callchain->ips[i]); 836 837 printf("..... %2d: %016" PRIx64 "\n", 838 (int)(kernel_callchain_nr), entries[0].to); 839 for (i = 0; i < lbr_stack->nr; i++) 840 printf("..... %2d: %016" PRIx64 "\n", 841 (int)(i + kernel_callchain_nr + 1), entries[i].from); 842 } 843 } 844 845 static const char *callchain_context_str(u64 ip) 846 { 847 switch (ip) { 848 case PERF_CONTEXT_HV: 849 return " (PERF_CONTEXT_HV)"; 850 case PERF_CONTEXT_KERNEL: 851 return " (PERF_CONTEXT_KERNEL)"; 852 case PERF_CONTEXT_USER: 853 return " (PERF_CONTEXT_USER)"; 854 case PERF_CONTEXT_GUEST: 855 return " (PERF_CONTEXT_GUEST)"; 856 case PERF_CONTEXT_GUEST_KERNEL: 857 return " (PERF_CONTEXT_GUEST_KERNEL)"; 858 case PERF_CONTEXT_GUEST_USER: 859 return " (PERF_CONTEXT_GUEST_USER)"; 860 case PERF_CONTEXT_USER_DEFERRED: 861 return " (PERF_CONTEXT_USER_DEFERRED)"; 862 default: 863 return ""; 864 } 865 } 866 867 static void callchain__printf(struct evsel *evsel, 868 struct perf_sample *sample) 869 { 870 unsigned int i; 871 struct ip_callchain *callchain = sample->callchain; 872 873 if (evsel__has_branch_callstack(evsel)) 874 callchain__lbr_callstack_printf(sample); 875 876 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr); 877 878 for (i = 0; i < callchain->nr; i++) 879 printf("..... %2d: %016" PRIx64 "%s\n", 880 i, callchain->ips[i], 881 callchain_context_str(callchain->ips[i])); 882 883 if (sample->deferred_callchain) 884 printf("...... (deferred)\n"); 885 } 886 887 static void branch_stack__printf(struct perf_sample *sample, 888 struct evsel *evsel) 889 { 890 struct branch_entry *entries = perf_sample__branch_entries(sample); 891 bool callstack = evsel__has_branch_callstack(evsel); 892 u64 *branch_stack_cntr = sample->branch_stack_cntr; 893 uint64_t i; 894 895 if (!callstack) { 896 printf("%s: nr:%" PRIu64 "\n", "... branch stack", sample->branch_stack->nr); 897 } else { 898 /* the reason of adding 1 to nr is because after expanding 899 * branch stack it generates nr + 1 callstack records. e.g., 900 * B()->C() 901 * A()->B() 902 * the final callstack should be: 903 * C() 904 * B() 905 * A() 906 */ 907 printf("%s: nr:%" PRIu64 "\n", "... branch callstack", sample->branch_stack->nr+1); 908 } 909 910 for (i = 0; i < sample->branch_stack->nr; i++) { 911 struct branch_entry *e = &entries[i]; 912 913 if (!callstack) { 914 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x %s %s\n", 915 i, e->from, e->to, 916 (unsigned short)e->flags.cycles, 917 e->flags.mispred ? "M" : " ", 918 e->flags.predicted ? "P" : " ", 919 e->flags.abort ? "A" : " ", 920 e->flags.in_tx ? "T" : " ", 921 (unsigned)e->flags.reserved, 922 get_branch_type(e), 923 e->flags.spec ? branch_spec_desc(e->flags.spec) : ""); 924 } else { 925 if (i == 0) { 926 printf("..... %2"PRIu64": %016" PRIx64 "\n" 927 "..... %2"PRIu64": %016" PRIx64 "\n", 928 i, e->to, i+1, e->from); 929 } else { 930 printf("..... %2"PRIu64": %016" PRIx64 "\n", i+1, e->from); 931 } 932 } 933 } 934 935 if (branch_stack_cntr) { 936 unsigned int br_cntr_width, br_cntr_nr; 937 938 perf_env__find_br_cntr_info(evsel__env(evsel), &br_cntr_nr, &br_cntr_width); 939 printf("... branch stack counters: nr:%" PRIu64 " (counter width: %u max counter nr:%u)\n", 940 sample->branch_stack->nr, br_cntr_width, br_cntr_nr); 941 for (i = 0; i < sample->branch_stack->nr; i++) 942 printf("..... %2"PRIu64": %016" PRIx64 "\n", i, branch_stack_cntr[i]); 943 } 944 } 945 946 static void regs_dump__printf(u64 mask, u64 *regs, uint16_t e_machine) 947 { 948 unsigned rid, i = 0; 949 950 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) { 951 u64 val = regs[i++]; 952 953 printf(".... %-5s 0x%016" PRIx64 "\n", 954 perf_reg_name(rid, e_machine), val); 955 } 956 } 957 958 static const char *regs_abi[] = { 959 [PERF_SAMPLE_REGS_ABI_NONE] = "none", 960 [PERF_SAMPLE_REGS_ABI_32] = "32-bit", 961 [PERF_SAMPLE_REGS_ABI_64] = "64-bit", 962 }; 963 964 static inline const char *regs_dump_abi(struct regs_dump *d) 965 { 966 if (d->abi > PERF_SAMPLE_REGS_ABI_64) 967 return "unknown"; 968 969 return regs_abi[d->abi]; 970 } 971 972 static void regs__printf(const char *type, struct regs_dump *regs, uint16_t e_machine) 973 { 974 u64 mask = regs->mask; 975 976 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n", 977 type, 978 mask, 979 regs_dump_abi(regs)); 980 981 regs_dump__printf(mask, regs->regs, e_machine); 982 } 983 984 static void regs_user__printf(struct perf_sample *sample, uint16_t e_machine) 985 { 986 struct regs_dump *user_regs; 987 988 if (!sample->user_regs) 989 return; 990 991 user_regs = perf_sample__user_regs(sample); 992 993 if (user_regs->regs) 994 regs__printf("user", user_regs, e_machine); 995 } 996 997 static void regs_intr__printf(struct perf_sample *sample, uint16_t e_machine) 998 { 999 struct regs_dump *intr_regs; 1000 1001 if (!sample->intr_regs) 1002 return; 1003 1004 intr_regs = perf_sample__intr_regs(sample); 1005 1006 if (intr_regs->regs) 1007 regs__printf("intr", intr_regs, e_machine); 1008 } 1009 1010 static void stack_user__printf(struct stack_dump *dump) 1011 { 1012 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 1013 dump->size, dump->offset); 1014 } 1015 1016 static void evlist__print_tstamp(struct evlist *evlist, union perf_event *event, struct perf_sample *sample) 1017 { 1018 u64 sample_type = __evlist__combined_sample_type(evlist); 1019 1020 if (event->header.type != PERF_RECORD_SAMPLE && 1021 !evlist__sample_id_all(evlist)) { 1022 fputs("-1 -1 ", stdout); 1023 return; 1024 } 1025 1026 if ((sample_type & PERF_SAMPLE_CPU)) 1027 printf("%u ", sample->cpu); 1028 1029 if (sample_type & PERF_SAMPLE_TIME) 1030 printf("%" PRIu64 " ", sample->time); 1031 } 1032 1033 static void sample_read__printf(struct perf_sample *sample, u64 read_format) 1034 { 1035 printf("... sample_read:\n"); 1036 1037 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1038 printf("...... time enabled %016" PRIx64 "\n", 1039 sample->read.time_enabled); 1040 1041 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1042 printf("...... time running %016" PRIx64 "\n", 1043 sample->read.time_running); 1044 1045 if (read_format & PERF_FORMAT_GROUP) { 1046 struct sample_read_value *value = sample->read.group.values; 1047 1048 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr); 1049 1050 sample_read_group__for_each(value, sample->read.group.nr, read_format) { 1051 printf("..... id %016" PRIx64 1052 ", value %016" PRIx64, 1053 value->id, value->value); 1054 if (read_format & PERF_FORMAT_LOST) 1055 printf(", lost %" PRIu64, value->lost); 1056 printf("\n"); 1057 } 1058 } else { 1059 printf("..... id %016" PRIx64 ", value %016" PRIx64, 1060 sample->read.one.id, sample->read.one.value); 1061 if (read_format & PERF_FORMAT_LOST) 1062 printf(", lost %" PRIu64, sample->read.one.lost); 1063 printf("\n"); 1064 } 1065 } 1066 1067 static void dump_event(struct evlist *evlist, union perf_event *event, 1068 u64 file_offset, struct perf_sample *sample, 1069 const char *file_path) 1070 { 1071 if (!dump_trace) 1072 return; 1073 1074 printf("\n%#" PRIx64 "@%s [%#x]: event: %d\n", 1075 file_offset, file_path, event->header.size, event->header.type); 1076 1077 trace_event(event); 1078 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw) 1079 evlist->trace_event_sample_raw(evlist, event, sample); 1080 1081 if (sample) 1082 evlist__print_tstamp(evlist, event, sample); 1083 1084 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 1085 event->header.size, perf_event__name(event->header.type)); 1086 } 1087 1088 char *get_page_size_name(u64 size, char *str) 1089 { 1090 if (!size || !unit_number__scnprintf(str, PAGE_SIZE_NAME_LEN, size)) 1091 snprintf(str, PAGE_SIZE_NAME_LEN, "%s", "N/A"); 1092 1093 return str; 1094 } 1095 1096 static void dump_sample(struct machine *machine, struct evsel *evsel, union perf_event *event, 1097 struct perf_sample *sample) 1098 { 1099 u64 sample_type; 1100 char str[PAGE_SIZE_NAME_LEN]; 1101 uint16_t e_machine = EM_NONE; 1102 1103 if (!dump_trace) 1104 return; 1105 1106 sample_type = evsel->core.attr.sample_type; 1107 1108 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR)) { 1109 struct thread *thread = machine__find_thread(machine, sample->pid, sample->pid); 1110 1111 e_machine = thread__e_machine(thread, machine); 1112 } 1113 1114 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 1115 event->header.misc, sample->pid, sample->tid, sample->ip, 1116 sample->period, sample->addr); 1117 1118 if (evsel__has_callchain(evsel)) 1119 callchain__printf(evsel, sample); 1120 1121 if (evsel__has_br_stack(evsel)) 1122 branch_stack__printf(sample, evsel); 1123 1124 if (sample_type & PERF_SAMPLE_REGS_USER) 1125 regs_user__printf(sample, e_machine); 1126 1127 if (sample_type & PERF_SAMPLE_REGS_INTR) 1128 regs_intr__printf(sample, e_machine); 1129 1130 if (sample_type & PERF_SAMPLE_STACK_USER) 1131 stack_user__printf(&sample->user_stack); 1132 1133 if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) { 1134 printf("... weight: %" PRIu64 "", sample->weight); 1135 if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) { 1136 printf(",0x%"PRIx16"", sample->ins_lat); 1137 printf(",0x%"PRIx16"", sample->weight3); 1138 } 1139 printf("\n"); 1140 } 1141 1142 if (sample_type & PERF_SAMPLE_DATA_SRC) 1143 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src); 1144 1145 if (sample_type & PERF_SAMPLE_PHYS_ADDR) 1146 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr); 1147 1148 if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE) 1149 printf(" .. data page size: %s\n", get_page_size_name(sample->data_page_size, str)); 1150 1151 if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE) 1152 printf(" .. code page size: %s\n", get_page_size_name(sample->code_page_size, str)); 1153 1154 if (sample_type & PERF_SAMPLE_TRANSACTION) 1155 printf("... transaction: %" PRIx64 "\n", sample->transaction); 1156 1157 if (sample_type & PERF_SAMPLE_READ) 1158 sample_read__printf(sample, evsel->core.attr.read_format); 1159 } 1160 1161 static void dump_deferred_callchain(struct evsel *evsel, union perf_event *event, 1162 struct perf_sample *sample) 1163 { 1164 if (!dump_trace) 1165 return; 1166 1167 printf("(IP, 0x%x): %d/%d: %#" PRIx64 "\n", 1168 event->header.misc, sample->pid, sample->tid, sample->deferred_cookie); 1169 1170 if (evsel__has_callchain(evsel)) 1171 callchain__printf(evsel, sample); 1172 } 1173 1174 static void dump_read(struct evsel *evsel, union perf_event *event) 1175 { 1176 struct perf_record_read *read_event = &event->read; 1177 u64 read_format; 1178 1179 if (!dump_trace) 1180 return; 1181 1182 printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid, 1183 evsel__name(evsel), event->read.value); 1184 1185 if (!evsel) 1186 return; 1187 1188 read_format = evsel->core.attr.read_format; 1189 1190 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1191 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled); 1192 1193 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1194 printf("... time running : %" PRI_lu64 "\n", read_event->time_running); 1195 1196 if (read_format & PERF_FORMAT_ID) 1197 printf("... id : %" PRI_lu64 "\n", read_event->id); 1198 1199 if (read_format & PERF_FORMAT_LOST) 1200 printf("... lost : %" PRI_lu64 "\n", read_event->lost); 1201 } 1202 1203 static struct machine *machines__find_for_cpumode(struct machines *machines, 1204 union perf_event *event, 1205 struct perf_sample *sample) 1206 { 1207 if (perf_guest && 1208 ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 1209 (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) { 1210 u32 pid; 1211 1212 if (sample->machine_pid) 1213 pid = sample->machine_pid; 1214 else if (event->header.type == PERF_RECORD_MMAP 1215 || event->header.type == PERF_RECORD_MMAP2) 1216 pid = event->mmap.pid; 1217 else 1218 pid = sample->pid; 1219 1220 /* 1221 * Guest code machine is created as needed and does not use 1222 * DEFAULT_GUEST_KERNEL_ID. 1223 */ 1224 if (symbol_conf.guest_code) 1225 return machines__findnew(machines, pid); 1226 1227 return machines__find_guest(machines, pid); 1228 } 1229 1230 return &machines->host; 1231 } 1232 1233 static int deliver_sample_value(struct evlist *evlist, 1234 const struct perf_tool *tool, 1235 union perf_event *event, 1236 struct perf_sample *sample, 1237 struct sample_read_value *v, 1238 struct machine *machine, 1239 bool per_thread) 1240 { 1241 struct perf_sample_id *sid = evlist__id2sid(evlist, v->id); 1242 struct evsel *evsel; 1243 u64 *storage = NULL; 1244 1245 if (sid) { 1246 storage = perf_sample_id__get_period_storage(sid, sample->tid, per_thread); 1247 } 1248 1249 if (storage) { 1250 sample->id = v->id; 1251 sample->period = v->value - *storage; 1252 *storage = v->value; 1253 } 1254 1255 if (!storage || sid->evsel == NULL) { 1256 ++evlist->stats.nr_unknown_id; 1257 return 0; 1258 } 1259 1260 /* 1261 * There's no reason to deliver sample 1262 * for zero period, bail out. 1263 */ 1264 if (!sample->period) 1265 return 0; 1266 1267 evsel = container_of(sid->evsel, struct evsel, core); 1268 return tool->sample(tool, event, sample, evsel, machine); 1269 } 1270 1271 static int deliver_sample_group(struct evlist *evlist, 1272 const struct perf_tool *tool, 1273 union perf_event *event, 1274 struct perf_sample *sample, 1275 struct machine *machine, 1276 u64 read_format, 1277 bool per_thread) 1278 { 1279 int ret = -EINVAL; 1280 struct sample_read_value *v = sample->read.group.values; 1281 1282 if (tool->dont_split_sample_group) 1283 return deliver_sample_value(evlist, tool, event, sample, v, machine, 1284 per_thread); 1285 1286 sample_read_group__for_each(v, sample->read.group.nr, read_format) { 1287 ret = deliver_sample_value(evlist, tool, event, sample, v, 1288 machine, per_thread); 1289 if (ret) 1290 break; 1291 } 1292 1293 return ret; 1294 } 1295 1296 static int evlist__deliver_sample(struct evlist *evlist, const struct perf_tool *tool, 1297 union perf_event *event, struct perf_sample *sample, 1298 struct evsel *evsel, struct machine *machine) 1299 { 1300 /* We know evsel != NULL. */ 1301 u64 sample_type = evsel->core.attr.sample_type; 1302 u64 read_format = evsel->core.attr.read_format; 1303 bool per_thread = perf_evsel__attr_has_per_thread_sample_period(&evsel->core); 1304 1305 /* Standard sample delivery. */ 1306 if (!(sample_type & PERF_SAMPLE_READ)) 1307 return tool->sample(tool, event, sample, evsel, machine); 1308 1309 /* For PERF_SAMPLE_READ we have either single or group mode. */ 1310 if (read_format & PERF_FORMAT_GROUP) 1311 return deliver_sample_group(evlist, tool, event, sample, 1312 machine, read_format, per_thread); 1313 else 1314 return deliver_sample_value(evlist, tool, event, sample, 1315 &sample->read.one, machine, 1316 per_thread); 1317 } 1318 1319 /* 1320 * Samples with deferred callchains should wait for the next matching 1321 * PERF_RECORD_CALLCHAIN_RECORD entries. Keep the events in a list and 1322 * deliver them once it finds the callchains. 1323 */ 1324 struct deferred_event { 1325 struct list_head list; 1326 union perf_event *event; 1327 }; 1328 1329 /* 1330 * This is called when a deferred callchain record comes up. Find all matching 1331 * samples, merge the callchains and process them. 1332 */ 1333 static int evlist__deliver_deferred_callchain(struct evlist *evlist, 1334 const struct perf_tool *tool, 1335 union perf_event *event, 1336 struct perf_sample *sample, 1337 struct machine *machine) 1338 { 1339 struct deferred_event *de, *tmp; 1340 struct evsel *evsel; 1341 int ret = 0; 1342 1343 if (!tool->merge_deferred_callchains) { 1344 evsel = evlist__id2evsel(evlist, sample->id); 1345 return tool->callchain_deferred(tool, event, sample, 1346 evsel, machine); 1347 } 1348 1349 list_for_each_entry_safe(de, tmp, &evlist->deferred_samples, list) { 1350 struct perf_sample orig_sample; 1351 1352 ret = evlist__parse_sample(evlist, de->event, &orig_sample); 1353 if (ret < 0) { 1354 pr_err("failed to parse original sample\n"); 1355 break; 1356 } 1357 1358 if (sample->tid != orig_sample.tid) 1359 continue; 1360 1361 if (event->callchain_deferred.cookie == orig_sample.deferred_cookie) 1362 sample__merge_deferred_callchain(&orig_sample, sample); 1363 else 1364 orig_sample.deferred_callchain = false; 1365 1366 evsel = evlist__id2evsel(evlist, orig_sample.id); 1367 ret = evlist__deliver_sample(evlist, tool, de->event, 1368 &orig_sample, evsel, machine); 1369 1370 if (orig_sample.deferred_callchain) 1371 free(orig_sample.callchain); 1372 1373 list_del(&de->list); 1374 free(de->event); 1375 free(de); 1376 1377 if (ret) 1378 break; 1379 } 1380 return ret; 1381 } 1382 1383 /* 1384 * This is called at the end of the data processing for the session. Flush the 1385 * remaining samples as there's no hope for matching deferred callchains. 1386 */ 1387 static int session__flush_deferred_samples(struct perf_session *session, 1388 const struct perf_tool *tool) 1389 { 1390 struct evlist *evlist = session->evlist; 1391 struct machine *machine = &session->machines.host; 1392 struct deferred_event *de, *tmp; 1393 struct evsel *evsel; 1394 int ret = 0; 1395 1396 list_for_each_entry_safe(de, tmp, &evlist->deferred_samples, list) { 1397 struct perf_sample sample; 1398 1399 ret = evlist__parse_sample(evlist, de->event, &sample); 1400 if (ret < 0) { 1401 pr_err("failed to parse original sample\n"); 1402 break; 1403 } 1404 1405 evsel = evlist__id2evsel(evlist, sample.id); 1406 ret = evlist__deliver_sample(evlist, tool, de->event, 1407 &sample, evsel, machine); 1408 1409 list_del(&de->list); 1410 free(de->event); 1411 free(de); 1412 1413 if (ret) 1414 break; 1415 } 1416 return ret; 1417 } 1418 1419 static int machines__deliver_event(struct machines *machines, 1420 struct evlist *evlist, 1421 union perf_event *event, 1422 struct perf_sample *sample, 1423 const struct perf_tool *tool, u64 file_offset, 1424 const char *file_path) 1425 { 1426 struct evsel *evsel; 1427 struct machine *machine; 1428 1429 dump_event(evlist, event, file_offset, sample, file_path); 1430 1431 evsel = evlist__id2evsel(evlist, sample->id); 1432 1433 machine = machines__find_for_cpumode(machines, event, sample); 1434 1435 switch (event->header.type) { 1436 case PERF_RECORD_SAMPLE: 1437 if (evsel == NULL) { 1438 ++evlist->stats.nr_unknown_id; 1439 return 0; 1440 } 1441 if (machine == NULL) { 1442 ++evlist->stats.nr_unprocessable_samples; 1443 dump_sample(machine, evsel, event, sample); 1444 return 0; 1445 } 1446 dump_sample(machine, evsel, event, sample); 1447 if (sample->deferred_callchain && tool->merge_deferred_callchains) { 1448 struct deferred_event *de = malloc(sizeof(*de)); 1449 size_t sz = event->header.size; 1450 1451 if (de == NULL) 1452 return -ENOMEM; 1453 1454 de->event = malloc(sz); 1455 if (de->event == NULL) { 1456 free(de); 1457 return -ENOMEM; 1458 } 1459 memcpy(de->event, event, sz); 1460 list_add_tail(&de->list, &evlist->deferred_samples); 1461 return 0; 1462 } 1463 return evlist__deliver_sample(evlist, tool, event, sample, evsel, machine); 1464 case PERF_RECORD_MMAP: 1465 return tool->mmap(tool, event, sample, machine); 1466 case PERF_RECORD_MMAP2: 1467 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT) 1468 ++evlist->stats.nr_proc_map_timeout; 1469 return tool->mmap2(tool, event, sample, machine); 1470 case PERF_RECORD_COMM: 1471 return tool->comm(tool, event, sample, machine); 1472 case PERF_RECORD_NAMESPACES: 1473 return tool->namespaces(tool, event, sample, machine); 1474 case PERF_RECORD_CGROUP: 1475 return tool->cgroup(tool, event, sample, machine); 1476 case PERF_RECORD_FORK: 1477 return tool->fork(tool, event, sample, machine); 1478 case PERF_RECORD_EXIT: 1479 return tool->exit(tool, event, sample, machine); 1480 case PERF_RECORD_LOST: 1481 if (tool->lost == perf_event__process_lost) 1482 evlist->stats.total_lost += event->lost.lost; 1483 return tool->lost(tool, event, sample, machine); 1484 case PERF_RECORD_LOST_SAMPLES: 1485 if (event->header.misc & PERF_RECORD_MISC_LOST_SAMPLES_BPF) 1486 evlist->stats.total_dropped_samples += event->lost_samples.lost; 1487 else if (tool->lost_samples == perf_event__process_lost_samples) 1488 evlist->stats.total_lost_samples += event->lost_samples.lost; 1489 return tool->lost_samples(tool, event, sample, machine); 1490 case PERF_RECORD_READ: 1491 dump_read(evsel, event); 1492 return tool->read(tool, event, sample, evsel, machine); 1493 case PERF_RECORD_THROTTLE: 1494 return tool->throttle(tool, event, sample, machine); 1495 case PERF_RECORD_UNTHROTTLE: 1496 return tool->unthrottle(tool, event, sample, machine); 1497 case PERF_RECORD_AUX: 1498 if (tool->aux == perf_event__process_aux) { 1499 if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) 1500 evlist->stats.total_aux_lost += 1; 1501 if (event->aux.flags & PERF_AUX_FLAG_PARTIAL) 1502 evlist->stats.total_aux_partial += 1; 1503 if (event->aux.flags & PERF_AUX_FLAG_COLLISION) 1504 evlist->stats.total_aux_collision += 1; 1505 } 1506 return tool->aux(tool, event, sample, machine); 1507 case PERF_RECORD_ITRACE_START: 1508 return tool->itrace_start(tool, event, sample, machine); 1509 case PERF_RECORD_SWITCH: 1510 case PERF_RECORD_SWITCH_CPU_WIDE: 1511 return tool->context_switch(tool, event, sample, machine); 1512 case PERF_RECORD_KSYMBOL: 1513 return tool->ksymbol(tool, event, sample, machine); 1514 case PERF_RECORD_BPF_EVENT: 1515 return tool->bpf(tool, event, sample, machine); 1516 case PERF_RECORD_TEXT_POKE: 1517 return tool->text_poke(tool, event, sample, machine); 1518 case PERF_RECORD_AUX_OUTPUT_HW_ID: 1519 return tool->aux_output_hw_id(tool, event, sample, machine); 1520 case PERF_RECORD_CALLCHAIN_DEFERRED: 1521 dump_deferred_callchain(evsel, event, sample); 1522 return evlist__deliver_deferred_callchain(evlist, tool, event, 1523 sample, machine); 1524 default: 1525 ++evlist->stats.nr_unknown_events; 1526 return -1; 1527 } 1528 } 1529 1530 static int perf_session__deliver_event(struct perf_session *session, 1531 union perf_event *event, 1532 const struct perf_tool *tool, 1533 u64 file_offset, 1534 const char *file_path) 1535 { 1536 struct perf_sample sample; 1537 int ret; 1538 1539 perf_sample__init(&sample, /*all=*/false); 1540 ret = evlist__parse_sample(session->evlist, event, &sample); 1541 if (ret) { 1542 pr_err("Can't parse sample, err = %d\n", ret); 1543 goto out; 1544 } 1545 1546 ret = auxtrace__process_event(session, event, &sample, tool); 1547 if (ret < 0) 1548 goto out; 1549 if (ret > 0) { 1550 ret = 0; 1551 goto out; 1552 } 1553 1554 ret = machines__deliver_event(&session->machines, session->evlist, 1555 event, &sample, tool, file_offset, file_path); 1556 1557 if (dump_trace && sample.aux_sample.size) 1558 auxtrace__dump_auxtrace_sample(session, &sample); 1559 out: 1560 perf_sample__exit(&sample); 1561 return ret; 1562 } 1563 1564 static s64 perf_session__process_user_event(struct perf_session *session, 1565 union perf_event *event, 1566 u64 file_offset, 1567 const char *file_path) 1568 { 1569 struct ordered_events *oe = &session->ordered_events; 1570 const struct perf_tool *tool = session->tool; 1571 struct perf_sample sample; 1572 int fd = perf_data__fd(session->data); 1573 s64 err; 1574 1575 perf_sample__init(&sample, /*all=*/true); 1576 if ((event->header.type != PERF_RECORD_COMPRESSED && 1577 event->header.type != PERF_RECORD_COMPRESSED2) || 1578 perf_tool__compressed_is_stub(tool)) 1579 dump_event(session->evlist, event, file_offset, &sample, file_path); 1580 1581 /* These events are processed right away */ 1582 switch (event->header.type) { 1583 case PERF_RECORD_HEADER_ATTR: 1584 err = tool->attr(tool, event, &session->evlist); 1585 if (err == 0) { 1586 perf_session__set_id_hdr_size(session); 1587 perf_session__set_comm_exec(session); 1588 } 1589 break; 1590 case PERF_RECORD_EVENT_UPDATE: 1591 err = tool->event_update(tool, event, &session->evlist); 1592 break; 1593 case PERF_RECORD_HEADER_EVENT_TYPE: 1594 /* 1595 * Deprecated, but we need to handle it for sake 1596 * of old data files create in pipe mode. 1597 */ 1598 err = 0; 1599 break; 1600 case PERF_RECORD_HEADER_TRACING_DATA: 1601 /* 1602 * Setup for reading amidst mmap, but only when we 1603 * are in 'file' mode. The 'pipe' fd is in proper 1604 * place already. 1605 */ 1606 if (!perf_data__is_pipe(session->data)) 1607 lseek(fd, file_offset, SEEK_SET); 1608 err = tool->tracing_data(tool, session, event); 1609 break; 1610 case PERF_RECORD_HEADER_BUILD_ID: 1611 err = tool->build_id(tool, session, event); 1612 break; 1613 case PERF_RECORD_FINISHED_ROUND: 1614 err = tool->finished_round(tool, event, oe); 1615 break; 1616 case PERF_RECORD_ID_INDEX: 1617 err = tool->id_index(tool, session, event); 1618 break; 1619 case PERF_RECORD_AUXTRACE_INFO: 1620 err = tool->auxtrace_info(tool, session, event); 1621 break; 1622 case PERF_RECORD_AUXTRACE: 1623 /* 1624 * Setup for reading amidst mmap, but only when we 1625 * are in 'file' mode. The 'pipe' fd is in proper 1626 * place already. 1627 */ 1628 if (!perf_data__is_pipe(session->data)) 1629 lseek(fd, file_offset + event->header.size, SEEK_SET); 1630 err = tool->auxtrace(tool, session, event); 1631 break; 1632 case PERF_RECORD_AUXTRACE_ERROR: 1633 perf_session__auxtrace_error_inc(session, event); 1634 err = tool->auxtrace_error(tool, session, event); 1635 break; 1636 case PERF_RECORD_THREAD_MAP: 1637 err = tool->thread_map(tool, session, event); 1638 break; 1639 case PERF_RECORD_CPU_MAP: 1640 err = tool->cpu_map(tool, session, event); 1641 break; 1642 case PERF_RECORD_STAT_CONFIG: 1643 err = tool->stat_config(tool, session, event); 1644 break; 1645 case PERF_RECORD_STAT: 1646 err = tool->stat(tool, session, event); 1647 break; 1648 case PERF_RECORD_STAT_ROUND: 1649 err = tool->stat_round(tool, session, event); 1650 break; 1651 case PERF_RECORD_TIME_CONV: 1652 session->time_conv = event->time_conv; 1653 err = tool->time_conv(tool, session, event); 1654 break; 1655 case PERF_RECORD_HEADER_FEATURE: 1656 err = tool->feature(tool, session, event); 1657 break; 1658 case PERF_RECORD_COMPRESSED: 1659 case PERF_RECORD_COMPRESSED2: 1660 err = tool->compressed(tool, session, event, file_offset, file_path); 1661 if (err) 1662 dump_event(session->evlist, event, file_offset, &sample, file_path); 1663 break; 1664 case PERF_RECORD_FINISHED_INIT: 1665 err = tool->finished_init(tool, session, event); 1666 break; 1667 case PERF_RECORD_BPF_METADATA: 1668 err = tool->bpf_metadata(tool, session, event); 1669 break; 1670 default: 1671 err = -EINVAL; 1672 break; 1673 } 1674 perf_sample__exit(&sample); 1675 return err; 1676 } 1677 1678 int perf_session__deliver_synth_event(struct perf_session *session, 1679 union perf_event *event, 1680 struct perf_sample *sample) 1681 { 1682 struct evlist *evlist = session->evlist; 1683 const struct perf_tool *tool = session->tool; 1684 1685 events_stats__inc(&evlist->stats, event->header.type); 1686 1687 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1688 return perf_session__process_user_event(session, event, 0, NULL); 1689 1690 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0, NULL); 1691 } 1692 1693 int perf_session__deliver_synth_attr_event(struct perf_session *session, 1694 const struct perf_event_attr *attr, 1695 u64 id) 1696 { 1697 union { 1698 struct { 1699 struct perf_record_header_attr attr; 1700 u64 ids[1]; 1701 } attr_id; 1702 union perf_event ev; 1703 } ev = { 1704 .attr_id.attr.header.type = PERF_RECORD_HEADER_ATTR, 1705 .attr_id.attr.header.size = sizeof(ev.attr_id), 1706 .attr_id.ids[0] = id, 1707 }; 1708 1709 if (attr->size != sizeof(ev.attr_id.attr.attr)) { 1710 pr_debug("Unexpected perf_event_attr size\n"); 1711 return -EINVAL; 1712 } 1713 ev.attr_id.attr.attr = *attr; 1714 return perf_session__deliver_synth_event(session, &ev.ev, NULL); 1715 } 1716 1717 static void event_swap(union perf_event *event, bool sample_id_all) 1718 { 1719 perf_event__swap_op swap; 1720 1721 swap = perf_event__swap_ops[event->header.type]; 1722 if (swap) 1723 swap(event, sample_id_all); 1724 } 1725 1726 int perf_session__peek_event(struct perf_session *session, off_t file_offset, 1727 void *buf, size_t buf_sz, 1728 union perf_event **event_ptr, 1729 struct perf_sample *sample) 1730 { 1731 union perf_event *event; 1732 size_t hdr_sz, rest; 1733 int fd; 1734 1735 if (session->one_mmap && !session->header.needs_swap) { 1736 event = file_offset - session->one_mmap_offset + 1737 session->one_mmap_addr; 1738 goto out_parse_sample; 1739 } 1740 1741 if (perf_data__is_pipe(session->data)) 1742 return -1; 1743 1744 fd = perf_data__fd(session->data); 1745 hdr_sz = sizeof(struct perf_event_header); 1746 1747 if (buf_sz < hdr_sz) 1748 return -1; 1749 1750 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 || 1751 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz) 1752 return -1; 1753 1754 event = (union perf_event *)buf; 1755 1756 if (session->header.needs_swap) 1757 perf_event_header__bswap(&event->header); 1758 1759 if (event->header.size < hdr_sz || event->header.size > buf_sz) 1760 return -1; 1761 1762 buf += hdr_sz; 1763 rest = event->header.size - hdr_sz; 1764 1765 if (readn(fd, buf, rest) != (ssize_t)rest) 1766 return -1; 1767 1768 if (session->header.needs_swap) 1769 event_swap(event, evlist__sample_id_all(session->evlist)); 1770 1771 out_parse_sample: 1772 1773 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START && 1774 evlist__parse_sample(session->evlist, event, sample)) 1775 return -1; 1776 1777 *event_ptr = event; 1778 1779 return 0; 1780 } 1781 1782 int perf_session__peek_events(struct perf_session *session, u64 offset, 1783 u64 size, peek_events_cb_t cb, void *data) 1784 { 1785 u64 max_offset = offset + size; 1786 char buf[PERF_SAMPLE_MAX_SIZE]; 1787 union perf_event *event; 1788 int err; 1789 1790 do { 1791 err = perf_session__peek_event(session, offset, buf, 1792 PERF_SAMPLE_MAX_SIZE, &event, 1793 NULL); 1794 if (err) 1795 return err; 1796 1797 err = cb(session, event, offset, data); 1798 if (err) 1799 return err; 1800 1801 offset += event->header.size; 1802 if (event->header.type == PERF_RECORD_AUXTRACE) 1803 offset += event->auxtrace.size; 1804 1805 } while (offset < max_offset); 1806 1807 return err; 1808 } 1809 1810 static s64 perf_session__process_event(struct perf_session *session, 1811 union perf_event *event, u64 file_offset, 1812 const char *file_path) 1813 { 1814 struct evlist *evlist = session->evlist; 1815 const struct perf_tool *tool = session->tool; 1816 int ret; 1817 1818 if (session->header.needs_swap) 1819 event_swap(event, evlist__sample_id_all(evlist)); 1820 1821 if (event->header.type >= PERF_RECORD_HEADER_MAX) { 1822 /* perf should not support unaligned event, stop here. */ 1823 if (event->header.size % sizeof(u64)) 1824 return -EINVAL; 1825 1826 /* This perf is outdated and does not support the latest event type. */ 1827 ui__warning("Unsupported header type %u, please consider updating perf.\n", 1828 event->header.type); 1829 /* Skip unsupported event by returning its size. */ 1830 return event->header.size; 1831 } 1832 1833 events_stats__inc(&evlist->stats, event->header.type); 1834 1835 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1836 return perf_session__process_user_event(session, event, file_offset, file_path); 1837 1838 if (tool->ordered_events) { 1839 u64 timestamp = -1ULL; 1840 1841 ret = evlist__parse_sample_timestamp(evlist, event, ×tamp); 1842 if (ret && ret != -1) 1843 return ret; 1844 1845 ret = perf_session__queue_event(session, event, timestamp, file_offset, file_path); 1846 if (ret != -ETIME) 1847 return ret; 1848 } 1849 1850 return perf_session__deliver_event(session, event, tool, file_offset, file_path); 1851 } 1852 1853 void perf_event_header__bswap(struct perf_event_header *hdr) 1854 { 1855 hdr->type = bswap_32(hdr->type); 1856 hdr->misc = bswap_16(hdr->misc); 1857 hdr->size = bswap_16(hdr->size); 1858 } 1859 1860 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1861 { 1862 return machine__findnew_thread(&session->machines.host, -1, pid); 1863 } 1864 1865 int perf_session__register_idle_thread(struct perf_session *session) 1866 { 1867 struct thread *thread = machine__idle_thread(&session->machines.host); 1868 1869 /* machine__idle_thread() got the thread, so put it */ 1870 thread__put(thread); 1871 return thread ? 0 : -1; 1872 } 1873 1874 static void 1875 perf_session__warn_order(const struct perf_session *session) 1876 { 1877 const struct ordered_events *oe = &session->ordered_events; 1878 struct evsel *evsel; 1879 bool should_warn = true; 1880 1881 evlist__for_each_entry(session->evlist, evsel) { 1882 if (evsel->core.attr.write_backward) 1883 should_warn = false; 1884 } 1885 1886 if (!should_warn) 1887 return; 1888 if (oe->nr_unordered_events != 0) 1889 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events); 1890 } 1891 1892 static void perf_session__warn_about_errors(const struct perf_session *session) 1893 { 1894 const struct events_stats *stats = &session->evlist->stats; 1895 1896 if (session->tool->lost == perf_event__process_lost && 1897 stats->nr_events[PERF_RECORD_LOST] != 0) { 1898 ui__warning("Processed %d events and lost %d chunks!\n\n" 1899 "Check IO/CPU overload!\n\n", 1900 stats->nr_events[0], 1901 stats->nr_events[PERF_RECORD_LOST]); 1902 } 1903 1904 if (session->tool->lost_samples == perf_event__process_lost_samples) { 1905 double drop_rate; 1906 1907 drop_rate = (double)stats->total_lost_samples / 1908 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples); 1909 if (drop_rate > 0.05) { 1910 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n", 1911 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples, 1912 drop_rate * 100.0); 1913 } 1914 } 1915 1916 if (session->tool->aux == perf_event__process_aux && 1917 stats->total_aux_lost != 0) { 1918 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n", 1919 stats->total_aux_lost, 1920 stats->nr_events[PERF_RECORD_AUX]); 1921 } 1922 1923 if (session->tool->aux == perf_event__process_aux && 1924 stats->total_aux_partial != 0) { 1925 bool vmm_exclusive = false; 1926 1927 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive", 1928 &vmm_exclusive); 1929 1930 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n" 1931 "Are you running a KVM guest in the background?%s\n\n", 1932 stats->total_aux_partial, 1933 stats->nr_events[PERF_RECORD_AUX], 1934 vmm_exclusive ? 1935 "\nReloading kvm_intel module with vmm_exclusive=0\n" 1936 "will reduce the gaps to only guest's timeslices." : 1937 ""); 1938 } 1939 1940 if (session->tool->aux == perf_event__process_aux && 1941 stats->total_aux_collision != 0) { 1942 ui__warning("AUX data detected collision %" PRIu64 " times out of %u!\n\n", 1943 stats->total_aux_collision, 1944 stats->nr_events[PERF_RECORD_AUX]); 1945 } 1946 1947 if (stats->nr_unknown_events != 0) { 1948 ui__warning("Found %u unknown events!\n\n" 1949 "Is this an older tool processing a perf.data " 1950 "file generated by a more recent tool?\n\n" 1951 "If that is not the case, consider " 1952 "reporting to linux-kernel@vger.kernel.org.\n\n", 1953 stats->nr_unknown_events); 1954 } 1955 1956 if (stats->nr_unknown_id != 0) { 1957 ui__warning("%u samples with id not present in the header\n", 1958 stats->nr_unknown_id); 1959 } 1960 1961 if (stats->nr_invalid_chains != 0) { 1962 ui__warning("Found invalid callchains!\n\n" 1963 "%u out of %u events were discarded for this reason.\n\n" 1964 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 1965 stats->nr_invalid_chains, 1966 stats->nr_events[PERF_RECORD_SAMPLE]); 1967 } 1968 1969 if (stats->nr_unprocessable_samples != 0) { 1970 ui__warning("%u unprocessable samples recorded.\n" 1971 "Do you have a KVM guest running and not using 'perf kvm'?\n", 1972 stats->nr_unprocessable_samples); 1973 } 1974 1975 perf_session__warn_order(session); 1976 1977 events_stats__auxtrace_error_warn(stats); 1978 1979 if (stats->nr_proc_map_timeout != 0) { 1980 ui__warning("%d map information files for pre-existing threads were\n" 1981 "not processed, if there are samples for addresses they\n" 1982 "will not be resolved, you may find out which are these\n" 1983 "threads by running with -v and redirecting the output\n" 1984 "to a file.\n" 1985 "The time limit to process proc map is too short?\n" 1986 "Increase it by --proc-map-timeout\n", 1987 stats->nr_proc_map_timeout); 1988 } 1989 } 1990 1991 static int perf_session__flush_thread_stack(struct thread *thread, 1992 void *p __maybe_unused) 1993 { 1994 return thread_stack__flush(thread); 1995 } 1996 1997 static int perf_session__flush_thread_stacks(struct perf_session *session) 1998 { 1999 return machines__for_each_thread(&session->machines, 2000 perf_session__flush_thread_stack, 2001 NULL); 2002 } 2003 2004 volatile sig_atomic_t session_done; 2005 2006 static int __perf_session__process_decomp_events(struct perf_session *session); 2007 2008 static int __perf_session__process_pipe_events(struct perf_session *session) 2009 { 2010 struct ordered_events *oe = &session->ordered_events; 2011 const struct perf_tool *tool = session->tool; 2012 struct ui_progress prog; 2013 union perf_event *event; 2014 uint32_t size, cur_size = 0; 2015 void *buf = NULL; 2016 s64 skip = 0; 2017 u64 head; 2018 ssize_t err; 2019 void *p; 2020 bool update_prog = false; 2021 2022 /* 2023 * If it's from a file saving pipe data (by redirection), it would have 2024 * a file name other than "-". Then we can get the total size and show 2025 * the progress. 2026 */ 2027 if (strcmp(session->data->path, "-") && session->data->file.size) { 2028 ui_progress__init_size(&prog, session->data->file.size, 2029 "Processing events..."); 2030 update_prog = true; 2031 } 2032 2033 head = 0; 2034 cur_size = sizeof(union perf_event); 2035 2036 buf = malloc(cur_size); 2037 if (!buf) 2038 return -errno; 2039 ordered_events__set_copy_on_queue(oe, true); 2040 more: 2041 event = buf; 2042 err = perf_data__read(session->data, event, 2043 sizeof(struct perf_event_header)); 2044 if (err <= 0) { 2045 if (err == 0) 2046 goto done; 2047 2048 pr_err("failed to read event header\n"); 2049 goto out_err; 2050 } 2051 2052 if (session->header.needs_swap) 2053 perf_event_header__bswap(&event->header); 2054 2055 size = event->header.size; 2056 if (size < sizeof(struct perf_event_header)) { 2057 pr_err("bad event header size\n"); 2058 goto out_err; 2059 } 2060 2061 if (size > cur_size) { 2062 void *new = realloc(buf, size); 2063 if (!new) { 2064 pr_err("failed to allocate memory to read event\n"); 2065 goto out_err; 2066 } 2067 buf = new; 2068 cur_size = size; 2069 event = buf; 2070 } 2071 p = event; 2072 p += sizeof(struct perf_event_header); 2073 2074 if (size - sizeof(struct perf_event_header)) { 2075 err = perf_data__read(session->data, p, 2076 size - sizeof(struct perf_event_header)); 2077 if (err <= 0) { 2078 if (err == 0) { 2079 pr_err("unexpected end of event stream\n"); 2080 goto done; 2081 } 2082 2083 pr_err("failed to read event data\n"); 2084 goto out_err; 2085 } 2086 } 2087 2088 if ((skip = perf_session__process_event(session, event, head, "pipe")) < 0) { 2089 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 2090 head, event->header.size, event->header.type); 2091 err = -EINVAL; 2092 goto out_err; 2093 } 2094 2095 head += size; 2096 2097 if (skip > 0) 2098 head += skip; 2099 2100 err = __perf_session__process_decomp_events(session); 2101 if (err) 2102 goto out_err; 2103 2104 if (update_prog) 2105 ui_progress__update(&prog, size); 2106 2107 if (!session_done()) 2108 goto more; 2109 done: 2110 /* do the final flush for ordered samples */ 2111 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2112 if (err) 2113 goto out_err; 2114 err = session__flush_deferred_samples(session, tool); 2115 if (err) 2116 goto out_err; 2117 err = auxtrace__flush_events(session, tool); 2118 if (err) 2119 goto out_err; 2120 err = perf_session__flush_thread_stacks(session); 2121 out_err: 2122 free(buf); 2123 if (update_prog) 2124 ui_progress__finish(); 2125 if (!tool->no_warn) 2126 perf_session__warn_about_errors(session); 2127 ordered_events__free(&session->ordered_events); 2128 auxtrace__free_events(session); 2129 return err; 2130 } 2131 2132 static union perf_event * 2133 prefetch_event(char *buf, u64 head, size_t mmap_size, 2134 bool needs_swap, union perf_event *error) 2135 { 2136 union perf_event *event; 2137 u16 event_size; 2138 2139 /* 2140 * Ensure we have enough space remaining to read 2141 * the size of the event in the headers. 2142 */ 2143 if (head + sizeof(event->header) > mmap_size) 2144 return NULL; 2145 2146 event = (union perf_event *)(buf + head); 2147 if (needs_swap) 2148 perf_event_header__bswap(&event->header); 2149 2150 event_size = event->header.size; 2151 if (head + event_size <= mmap_size) 2152 return event; 2153 2154 /* We're not fetching the event so swap back again */ 2155 if (needs_swap) 2156 perf_event_header__bswap(&event->header); 2157 2158 /* Check if the event fits into the next mmapped buf. */ 2159 if (event_size <= mmap_size - head % page_size) { 2160 /* Remap buf and fetch again. */ 2161 return NULL; 2162 } 2163 2164 /* Invalid input. Event size should never exceed mmap_size. */ 2165 pr_debug("%s: head=%#" PRIx64 " event->header.size=%#x, mmap_size=%#zx:" 2166 " fuzzed or compressed perf.data?\n", __func__, head, event_size, mmap_size); 2167 2168 return error; 2169 } 2170 2171 static union perf_event * 2172 fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 2173 { 2174 return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL)); 2175 } 2176 2177 static union perf_event * 2178 fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 2179 { 2180 return prefetch_event(buf, head, mmap_size, needs_swap, NULL); 2181 } 2182 2183 static int __perf_session__process_decomp_events(struct perf_session *session) 2184 { 2185 s64 skip; 2186 u64 size; 2187 struct decomp *decomp = session->active_decomp->decomp_last; 2188 2189 if (!decomp) 2190 return 0; 2191 2192 while (decomp->head < decomp->size && !session_done()) { 2193 union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data, 2194 session->header.needs_swap); 2195 2196 if (!event) 2197 break; 2198 2199 size = event->header.size; 2200 2201 if (size < sizeof(struct perf_event_header) || 2202 (skip = perf_session__process_event(session, event, decomp->file_pos, 2203 decomp->file_path)) < 0) { 2204 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d\n", 2205 decomp->file_pos + decomp->head, event->header.size, event->header.type); 2206 return -EINVAL; 2207 } 2208 2209 if (skip) 2210 size += skip; 2211 2212 decomp->head += size; 2213 } 2214 2215 return 0; 2216 } 2217 2218 /* 2219 * On 64bit we can mmap the data file in one go. No need for tiny mmap 2220 * slices. On 32bit we use 32MB. 2221 */ 2222 #if BITS_PER_LONG == 64 2223 #define MMAP_SIZE ULLONG_MAX 2224 #define NUM_MMAPS 1 2225 #else 2226 #define MMAP_SIZE (32 * 1024 * 1024ULL) 2227 #define NUM_MMAPS 128 2228 #endif 2229 2230 struct reader; 2231 2232 typedef s64 (*reader_cb_t)(struct perf_session *session, 2233 union perf_event *event, 2234 u64 file_offset, 2235 const char *file_path); 2236 2237 struct reader { 2238 int fd; 2239 const char *path; 2240 u64 data_size; 2241 u64 data_offset; 2242 reader_cb_t process; 2243 bool in_place_update; 2244 char *mmaps[NUM_MMAPS]; 2245 size_t mmap_size; 2246 int mmap_idx; 2247 char *mmap_cur; 2248 u64 file_pos; 2249 u64 file_offset; 2250 u64 head; 2251 u64 size; 2252 bool done; 2253 struct zstd_data zstd_data; 2254 struct decomp_data decomp_data; 2255 }; 2256 2257 static int 2258 reader__init(struct reader *rd, bool *one_mmap) 2259 { 2260 u64 data_size = rd->data_size; 2261 char **mmaps = rd->mmaps; 2262 2263 rd->head = rd->data_offset; 2264 data_size += rd->data_offset; 2265 2266 rd->mmap_size = MMAP_SIZE; 2267 if (rd->mmap_size > data_size) { 2268 rd->mmap_size = data_size; 2269 if (one_mmap) 2270 *one_mmap = true; 2271 } 2272 2273 memset(mmaps, 0, sizeof(rd->mmaps)); 2274 2275 if (zstd_init(&rd->zstd_data, 0)) 2276 return -1; 2277 rd->decomp_data.zstd_decomp = &rd->zstd_data; 2278 2279 return 0; 2280 } 2281 2282 static void 2283 reader__release_decomp(struct reader *rd) 2284 { 2285 perf_decomp__release_events(rd->decomp_data.decomp); 2286 zstd_fini(&rd->zstd_data); 2287 } 2288 2289 static int 2290 reader__mmap(struct reader *rd, struct perf_session *session) 2291 { 2292 int mmap_prot, mmap_flags; 2293 char *buf, **mmaps = rd->mmaps; 2294 u64 page_offset; 2295 2296 mmap_prot = PROT_READ; 2297 mmap_flags = MAP_SHARED; 2298 2299 if (rd->in_place_update) { 2300 mmap_prot |= PROT_WRITE; 2301 } else if (session->header.needs_swap) { 2302 mmap_prot |= PROT_WRITE; 2303 mmap_flags = MAP_PRIVATE; 2304 } 2305 2306 if (mmaps[rd->mmap_idx]) { 2307 munmap(mmaps[rd->mmap_idx], rd->mmap_size); 2308 mmaps[rd->mmap_idx] = NULL; 2309 } 2310 2311 page_offset = page_size * (rd->head / page_size); 2312 rd->file_offset += page_offset; 2313 rd->head -= page_offset; 2314 2315 buf = mmap(NULL, rd->mmap_size, mmap_prot, mmap_flags, rd->fd, 2316 rd->file_offset); 2317 if (buf == MAP_FAILED) { 2318 pr_err("failed to mmap file\n"); 2319 return -errno; 2320 } 2321 mmaps[rd->mmap_idx] = rd->mmap_cur = buf; 2322 rd->mmap_idx = (rd->mmap_idx + 1) & (ARRAY_SIZE(rd->mmaps) - 1); 2323 rd->file_pos = rd->file_offset + rd->head; 2324 if (session->one_mmap) { 2325 session->one_mmap_addr = buf; 2326 session->one_mmap_offset = rd->file_offset; 2327 } 2328 2329 return 0; 2330 } 2331 2332 enum { 2333 READER_OK, 2334 READER_NODATA, 2335 }; 2336 2337 static int 2338 reader__read_event(struct reader *rd, struct perf_session *session, 2339 struct ui_progress *prog) 2340 { 2341 u64 size; 2342 int err = READER_OK; 2343 union perf_event *event; 2344 s64 skip; 2345 2346 event = fetch_mmaped_event(rd->head, rd->mmap_size, rd->mmap_cur, 2347 session->header.needs_swap); 2348 if (IS_ERR(event)) 2349 return PTR_ERR(event); 2350 2351 if (!event) 2352 return READER_NODATA; 2353 2354 size = event->header.size; 2355 2356 skip = -EINVAL; 2357 2358 if (size < sizeof(struct perf_event_header) || 2359 (skip = rd->process(session, event, rd->file_pos, rd->path)) < 0) { 2360 errno = -skip; 2361 pr_err("%#" PRIx64 " [%#x]: failed to process type: %d [%m]\n", 2362 rd->file_offset + rd->head, event->header.size, 2363 event->header.type); 2364 err = skip; 2365 goto out; 2366 } 2367 2368 if (skip) 2369 size += skip; 2370 2371 rd->size += size; 2372 rd->head += size; 2373 rd->file_pos += size; 2374 2375 err = __perf_session__process_decomp_events(session); 2376 if (err) 2377 goto out; 2378 2379 ui_progress__update(prog, size); 2380 2381 out: 2382 return err; 2383 } 2384 2385 static inline bool 2386 reader__eof(struct reader *rd) 2387 { 2388 return (rd->file_pos >= rd->data_size + rd->data_offset); 2389 } 2390 2391 static int 2392 reader__process_events(struct reader *rd, struct perf_session *session, 2393 struct ui_progress *prog) 2394 { 2395 int err; 2396 2397 err = reader__init(rd, &session->one_mmap); 2398 if (err) 2399 goto out; 2400 2401 session->active_decomp = &rd->decomp_data; 2402 2403 remap: 2404 err = reader__mmap(rd, session); 2405 if (err) 2406 goto out; 2407 2408 more: 2409 err = reader__read_event(rd, session, prog); 2410 if (err < 0) 2411 goto out; 2412 else if (err == READER_NODATA) 2413 goto remap; 2414 2415 if (session_done()) 2416 goto out; 2417 2418 if (!reader__eof(rd)) 2419 goto more; 2420 2421 out: 2422 session->active_decomp = &session->decomp_data; 2423 return err; 2424 } 2425 2426 static s64 process_simple(struct perf_session *session, 2427 union perf_event *event, 2428 u64 file_offset, 2429 const char *file_path) 2430 { 2431 return perf_session__process_event(session, event, file_offset, file_path); 2432 } 2433 2434 static int __perf_session__process_events(struct perf_session *session) 2435 { 2436 struct reader rd = { 2437 .fd = perf_data__fd(session->data), 2438 .path = session->data->file.path, 2439 .data_size = session->header.data_size, 2440 .data_offset = session->header.data_offset, 2441 .process = process_simple, 2442 .in_place_update = session->data->in_place_update, 2443 }; 2444 struct ordered_events *oe = &session->ordered_events; 2445 const struct perf_tool *tool = session->tool; 2446 struct ui_progress prog; 2447 int err; 2448 2449 if (rd.data_size == 0) 2450 return -1; 2451 2452 ui_progress__init_size(&prog, rd.data_size, "Processing events..."); 2453 2454 err = reader__process_events(&rd, session, &prog); 2455 if (err) 2456 goto out_err; 2457 /* do the final flush for ordered samples */ 2458 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2459 if (err) 2460 goto out_err; 2461 err = auxtrace__flush_events(session, tool); 2462 if (err) 2463 goto out_err; 2464 err = session__flush_deferred_samples(session, tool); 2465 if (err) 2466 goto out_err; 2467 err = perf_session__flush_thread_stacks(session); 2468 out_err: 2469 ui_progress__finish(); 2470 if (!tool->no_warn) 2471 perf_session__warn_about_errors(session); 2472 /* 2473 * We may switching perf.data output, make ordered_events 2474 * reusable. 2475 */ 2476 ordered_events__reinit(&session->ordered_events); 2477 auxtrace__free_events(session); 2478 reader__release_decomp(&rd); 2479 session->one_mmap = false; 2480 return err; 2481 } 2482 2483 /* 2484 * Processing 2 MB of data from each reader in sequence, 2485 * because that's the way the ordered events sorting works 2486 * most efficiently. 2487 */ 2488 #define READER_MAX_SIZE (2 * 1024 * 1024) 2489 2490 /* 2491 * This function reads, merge and process directory data. 2492 * It assumens the version 1 of directory data, where each 2493 * data file holds per-cpu data, already sorted by kernel. 2494 */ 2495 static int __perf_session__process_dir_events(struct perf_session *session) 2496 { 2497 struct perf_data *data = session->data; 2498 const struct perf_tool *tool = session->tool; 2499 int i, ret, readers, nr_readers; 2500 struct ui_progress prog; 2501 u64 total_size = perf_data__size(session->data); 2502 struct reader *rd; 2503 2504 ui_progress__init_size(&prog, total_size, "Processing events..."); 2505 2506 nr_readers = 1; 2507 for (i = 0; i < data->dir.nr; i++) { 2508 if (data->dir.files[i].size) 2509 nr_readers++; 2510 } 2511 2512 rd = zalloc(nr_readers * sizeof(struct reader)); 2513 if (!rd) 2514 return -ENOMEM; 2515 2516 rd[0] = (struct reader) { 2517 .fd = perf_data__fd(session->data), 2518 .path = session->data->file.path, 2519 .data_size = session->header.data_size, 2520 .data_offset = session->header.data_offset, 2521 .process = process_simple, 2522 .in_place_update = session->data->in_place_update, 2523 }; 2524 ret = reader__init(&rd[0], NULL); 2525 if (ret) 2526 goto out_err; 2527 ret = reader__mmap(&rd[0], session); 2528 if (ret) 2529 goto out_err; 2530 readers = 1; 2531 2532 for (i = 0; i < data->dir.nr; i++) { 2533 if (!data->dir.files[i].size) 2534 continue; 2535 rd[readers] = (struct reader) { 2536 .fd = data->dir.files[i].fd, 2537 .path = data->dir.files[i].path, 2538 .data_size = data->dir.files[i].size, 2539 .data_offset = 0, 2540 .process = process_simple, 2541 .in_place_update = session->data->in_place_update, 2542 }; 2543 ret = reader__init(&rd[readers], NULL); 2544 if (ret) 2545 goto out_err; 2546 ret = reader__mmap(&rd[readers], session); 2547 if (ret) 2548 goto out_err; 2549 readers++; 2550 } 2551 2552 i = 0; 2553 while (readers) { 2554 if (session_done()) 2555 break; 2556 2557 if (rd[i].done) { 2558 i = (i + 1) % nr_readers; 2559 continue; 2560 } 2561 if (reader__eof(&rd[i])) { 2562 rd[i].done = true; 2563 readers--; 2564 continue; 2565 } 2566 2567 session->active_decomp = &rd[i].decomp_data; 2568 ret = reader__read_event(&rd[i], session, &prog); 2569 if (ret < 0) { 2570 goto out_err; 2571 } else if (ret == READER_NODATA) { 2572 ret = reader__mmap(&rd[i], session); 2573 if (ret) 2574 goto out_err; 2575 } 2576 2577 if (rd[i].size >= READER_MAX_SIZE) { 2578 rd[i].size = 0; 2579 i = (i + 1) % nr_readers; 2580 } 2581 } 2582 2583 ret = ordered_events__flush(&session->ordered_events, OE_FLUSH__FINAL); 2584 if (ret) 2585 goto out_err; 2586 2587 ret = session__flush_deferred_samples(session, tool); 2588 if (ret) 2589 goto out_err; 2590 2591 ret = perf_session__flush_thread_stacks(session); 2592 out_err: 2593 ui_progress__finish(); 2594 2595 if (!tool->no_warn) 2596 perf_session__warn_about_errors(session); 2597 2598 /* 2599 * We may switching perf.data output, make ordered_events 2600 * reusable. 2601 */ 2602 ordered_events__reinit(&session->ordered_events); 2603 2604 session->one_mmap = false; 2605 2606 session->active_decomp = &session->decomp_data; 2607 for (i = 0; i < nr_readers; i++) 2608 reader__release_decomp(&rd[i]); 2609 zfree(&rd); 2610 2611 return ret; 2612 } 2613 2614 int perf_session__process_events(struct perf_session *session) 2615 { 2616 if (perf_session__register_idle_thread(session) < 0) 2617 return -ENOMEM; 2618 2619 if (perf_data__is_pipe(session->data)) 2620 return __perf_session__process_pipe_events(session); 2621 2622 if (perf_data__is_dir(session->data) && session->data->dir.nr) 2623 return __perf_session__process_dir_events(session); 2624 2625 return __perf_session__process_events(session); 2626 } 2627 2628 bool perf_session__has_traces(struct perf_session *session, const char *msg) 2629 { 2630 struct evsel *evsel; 2631 2632 evlist__for_each_entry(session->evlist, evsel) { 2633 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT) 2634 return true; 2635 } 2636 2637 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 2638 return false; 2639 } 2640 2641 bool perf_session__has_switch_events(struct perf_session *session) 2642 { 2643 struct evsel *evsel; 2644 2645 evlist__for_each_entry(session->evlist, evsel) { 2646 if (evsel->core.attr.context_switch) 2647 return true; 2648 } 2649 2650 return false; 2651 } 2652 2653 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr) 2654 { 2655 char *bracket; 2656 struct ref_reloc_sym *ref; 2657 struct kmap *kmap; 2658 2659 ref = zalloc(sizeof(struct ref_reloc_sym)); 2660 if (ref == NULL) 2661 return -ENOMEM; 2662 2663 ref->name = strdup(symbol_name); 2664 if (ref->name == NULL) { 2665 free(ref); 2666 return -ENOMEM; 2667 } 2668 2669 bracket = strchr(ref->name, ']'); 2670 if (bracket) 2671 *bracket = '\0'; 2672 2673 ref->addr = addr; 2674 2675 kmap = map__kmap(map); 2676 if (kmap) 2677 kmap->ref_reloc_sym = ref; 2678 2679 return 0; 2680 } 2681 2682 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp) 2683 { 2684 return machines__fprintf_dsos(&session->machines, fp); 2685 } 2686 2687 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp, 2688 bool (skip)(struct dso *dso, int parm), int parm) 2689 { 2690 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm); 2691 } 2692 2693 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 2694 { 2695 size_t ret; 2696 const char *msg = ""; 2697 2698 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) 2699 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)"; 2700 2701 ret = fprintf(fp, "\nAggregated stats:%s\n", msg); 2702 2703 ret += events_stats__fprintf(&session->evlist->stats, fp); 2704 return ret; 2705 } 2706 2707 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 2708 { 2709 /* 2710 * FIXME: Here we have to actually print all the machines in this 2711 * session, not just the host... 2712 */ 2713 return machine__fprintf(&session->machines.host, fp); 2714 } 2715 2716 void perf_session__dump_kmaps(struct perf_session *session) 2717 { 2718 int save_verbose = verbose; 2719 2720 fflush(stdout); 2721 fprintf(stderr, "Kernel and module maps:\n"); 2722 verbose = 0; /* Suppress verbose to print a summary only */ 2723 maps__fprintf(machine__kernel_maps(&session->machines.host), stderr); 2724 verbose = save_verbose; 2725 } 2726 2727 struct evsel *perf_session__find_first_evtype(struct perf_session *session, 2728 unsigned int type) 2729 { 2730 struct evsel *pos; 2731 2732 evlist__for_each_entry(session->evlist, pos) { 2733 if (pos->core.attr.type == type) 2734 return pos; 2735 } 2736 return NULL; 2737 } 2738 2739 int perf_session__cpu_bitmap(struct perf_session *session, 2740 const char *cpu_list, unsigned long *cpu_bitmap) 2741 { 2742 int i, err = -1; 2743 struct perf_cpu_map *map; 2744 int nr_cpus = min(perf_session__env(session)->nr_cpus_avail, MAX_NR_CPUS); 2745 struct perf_cpu cpu; 2746 2747 for (i = 0; i < PERF_TYPE_MAX; ++i) { 2748 struct evsel *evsel; 2749 2750 evsel = perf_session__find_first_evtype(session, i); 2751 if (!evsel) 2752 continue; 2753 2754 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) { 2755 pr_err("File does not contain CPU events. " 2756 "Remove -C option to proceed.\n"); 2757 return -1; 2758 } 2759 } 2760 2761 map = perf_cpu_map__new(cpu_list); 2762 if (map == NULL) { 2763 pr_err("Invalid cpu_list\n"); 2764 return -1; 2765 } 2766 2767 perf_cpu_map__for_each_cpu(cpu, i, map) { 2768 if (cpu.cpu >= nr_cpus) { 2769 pr_err("Requested CPU %d too large. " 2770 "Consider raising MAX_NR_CPUS\n", cpu.cpu); 2771 goto out_delete_map; 2772 } 2773 2774 __set_bit(cpu.cpu, cpu_bitmap); 2775 } 2776 2777 err = 0; 2778 2779 out_delete_map: 2780 perf_cpu_map__put(map); 2781 return err; 2782 } 2783 2784 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 2785 bool full) 2786 { 2787 if (session == NULL || fp == NULL) 2788 return; 2789 2790 fprintf(fp, "# ========\n"); 2791 perf_header__fprintf_info(session, fp, full); 2792 fprintf(fp, "# ========\n#\n"); 2793 } 2794 2795 static int perf_session__register_guest(struct perf_session *session, pid_t machine_pid) 2796 { 2797 struct machine *machine = machines__findnew(&session->machines, machine_pid); 2798 struct thread *thread; 2799 2800 if (!machine) 2801 return -ENOMEM; 2802 2803 machine->single_address_space = session->machines.host.single_address_space; 2804 2805 thread = machine__idle_thread(machine); 2806 if (!thread) 2807 return -ENOMEM; 2808 thread__put(thread); 2809 2810 machine->kallsyms_filename = perf_data__guest_kallsyms_name(session->data, machine_pid); 2811 2812 return 0; 2813 } 2814 2815 static int perf_session__set_guest_cpu(struct perf_session *session, pid_t pid, 2816 pid_t tid, int guest_cpu) 2817 { 2818 struct machine *machine = &session->machines.host; 2819 struct thread *thread = machine__findnew_thread(machine, pid, tid); 2820 2821 if (!thread) 2822 return -ENOMEM; 2823 thread__set_guest_cpu(thread, guest_cpu); 2824 thread__put(thread); 2825 2826 return 0; 2827 } 2828 2829 int perf_event__process_id_index(const struct perf_tool *tool __maybe_unused, 2830 struct perf_session *session, 2831 union perf_event *event) 2832 { 2833 struct evlist *evlist = session->evlist; 2834 struct perf_record_id_index *ie = &event->id_index; 2835 size_t sz = ie->header.size - sizeof(*ie); 2836 size_t i, nr, max_nr; 2837 size_t e1_sz = sizeof(struct id_index_entry); 2838 size_t e2_sz = sizeof(struct id_index_entry_2); 2839 size_t etot_sz = e1_sz + e2_sz; 2840 struct id_index_entry_2 *e2; 2841 pid_t last_pid = 0; 2842 2843 max_nr = sz / e1_sz; 2844 nr = ie->nr; 2845 if (nr > max_nr) { 2846 printf("Too big: nr %zu max_nr %zu\n", nr, max_nr); 2847 return -EINVAL; 2848 } 2849 2850 if (sz >= nr * etot_sz) { 2851 max_nr = sz / etot_sz; 2852 if (nr > max_nr) { 2853 printf("Too big2: nr %zu max_nr %zu\n", nr, max_nr); 2854 return -EINVAL; 2855 } 2856 e2 = (void *)ie + sizeof(*ie) + nr * e1_sz; 2857 } else { 2858 e2 = NULL; 2859 } 2860 2861 if (dump_trace) 2862 fprintf(stdout, " nr: %zu\n", nr); 2863 2864 for (i = 0; i < nr; i++, (e2 ? e2++ : 0)) { 2865 struct id_index_entry *e = &ie->entries[i]; 2866 struct perf_sample_id *sid; 2867 int ret; 2868 2869 if (dump_trace) { 2870 fprintf(stdout, " ... id: %"PRI_lu64, e->id); 2871 fprintf(stdout, " idx: %"PRI_lu64, e->idx); 2872 fprintf(stdout, " cpu: %"PRI_ld64, e->cpu); 2873 fprintf(stdout, " tid: %"PRI_ld64, e->tid); 2874 if (e2) { 2875 fprintf(stdout, " machine_pid: %"PRI_ld64, e2->machine_pid); 2876 fprintf(stdout, " vcpu: %"PRI_lu64"\n", e2->vcpu); 2877 } else { 2878 fprintf(stdout, "\n"); 2879 } 2880 } 2881 2882 sid = evlist__id2sid(evlist, e->id); 2883 if (!sid) 2884 return -ENOENT; 2885 2886 sid->idx = e->idx; 2887 sid->cpu.cpu = e->cpu; 2888 sid->tid = e->tid; 2889 2890 if (!e2) 2891 continue; 2892 2893 sid->machine_pid = e2->machine_pid; 2894 sid->vcpu.cpu = e2->vcpu; 2895 2896 if (!sid->machine_pid) 2897 continue; 2898 2899 if (sid->machine_pid != last_pid) { 2900 ret = perf_session__register_guest(session, sid->machine_pid); 2901 if (ret) 2902 return ret; 2903 last_pid = sid->machine_pid; 2904 perf_guest = true; 2905 } 2906 2907 ret = perf_session__set_guest_cpu(session, sid->machine_pid, e->tid, e2->vcpu); 2908 if (ret) 2909 return ret; 2910 } 2911 return 0; 2912 } 2913 2914 int perf_session__dsos_hit_all(struct perf_session *session) 2915 { 2916 struct rb_node *nd; 2917 int err; 2918 2919 err = machine__hit_all_dsos(&session->machines.host); 2920 if (err) 2921 return err; 2922 2923 for (nd = rb_first_cached(&session->machines.guests); nd; 2924 nd = rb_next(nd)) { 2925 struct machine *pos = rb_entry(nd, struct machine, rb_node); 2926 2927 err = machine__hit_all_dsos(pos); 2928 if (err) 2929 return err; 2930 } 2931 2932 return 0; 2933 } 2934 2935 struct perf_env *perf_session__env(struct perf_session *session) 2936 { 2937 return &session->header.env; 2938 } 2939 2940 static int perf_session__e_machine_cb(struct thread *thread, 2941 void *arg __maybe_unused) 2942 { 2943 uint16_t *result = arg; 2944 struct machine *machine = maps__machine(thread__maps(thread)); 2945 2946 *result = thread__e_machine(thread, machine); 2947 return *result != EM_NONE ? 1 : 0; 2948 } 2949 2950 /* 2951 * Note, a machine may have mixed 32-bit and 64-bit processes and so mixed 2952 * e_machines. Use thread__e_machine when this matters. 2953 */ 2954 uint16_t perf_session__e_machine(struct perf_session *session) 2955 { 2956 uint16_t e_machine = EM_NONE; 2957 2958 machines__for_each_thread(&session->machines, 2959 perf_session__e_machine_cb, 2960 &e_machine); 2961 2962 return e_machine == EM_NONE ? EM_HOST : e_machine; 2963 } 2964