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 int ret = perf_session__deliver_event(session, event->event, 135 session->tool, event->file_offset, 136 event->file_path); 137 138 if (ret) { 139 pr_err("%#" PRIx64 " [%#x]: ordered event processing failed (%d) for event of type: %s (%d)\n", 140 event->file_offset, event->event->header.size, ret, 141 perf_event__name(event->event->header.type), 142 event->event->header.type); 143 } 144 return ret; 145 } 146 147 struct perf_session *__perf_session__new(struct perf_data *data, 148 struct perf_tool *tool, 149 bool trace_event_repipe, 150 struct perf_env *host_env) 151 { 152 int ret = -ENOMEM; 153 struct perf_session *session = zalloc(sizeof(*session)); 154 155 if (!session) 156 goto out; 157 158 session->trace_event_repipe = trace_event_repipe; 159 session->tool = tool; 160 session->decomp_data.zstd_decomp = &session->zstd_data; 161 session->active_decomp = &session->decomp_data; 162 INIT_LIST_HEAD(&session->auxtrace_index); 163 machines__init(&session->machines); 164 ordered_events__init(&session->ordered_events, 165 ordered_events__deliver_event, NULL); 166 167 perf_env__init(&session->header.env); 168 if (data) { 169 ret = perf_data__open(data); 170 if (ret < 0) 171 goto out_delete; 172 173 session->data = data; 174 175 if (perf_data__is_read(data)) { 176 ret = perf_session__open(session); 177 if (ret < 0) 178 goto out_delete; 179 180 /* 181 * set session attributes that are present in perf.data 182 * but not in pipe-mode. 183 */ 184 if (!data->is_pipe) { 185 perf_session__set_id_hdr_size(session); 186 perf_session__set_comm_exec(session); 187 } 188 189 evlist__init_trace_event_sample_raw(session->evlist, &session->header.env); 190 191 /* Open the directory data. */ 192 if (data->is_dir) { 193 ret = perf_data__open_dir(data); 194 if (ret) 195 goto out_delete; 196 } 197 198 if (!symbol_conf.kallsyms_name && 199 !symbol_conf.vmlinux_name) 200 symbol_conf.kallsyms_name = perf_data__kallsyms_name(data); 201 } 202 } else { 203 assert(host_env != NULL); 204 session->machines.host.env = host_env; 205 } 206 if (session->evlist) 207 session->evlist->session = session; 208 209 session->machines.host.single_address_space = 210 perf_env__single_address_space(session->machines.host.env); 211 212 if (!data || perf_data__is_write(data)) { 213 /* 214 * In O_RDONLY mode this will be performed when reading the 215 * kernel MMAP event, in perf_event__process_mmap(). 216 */ 217 if (perf_session__create_kernel_maps(session) < 0) 218 pr_warning("Cannot read kernel map\n"); 219 } 220 221 /* 222 * In pipe-mode, evlist is empty until PERF_RECORD_HEADER_ATTR is 223 * processed, so evlist__sample_id_all is not meaningful here. 224 */ 225 if ((!data || !data->is_pipe) && tool && tool->ordering_requires_timestamps && 226 tool->ordered_events && !evlist__sample_id_all(session->evlist)) { 227 dump_printf("WARNING: No sample_id_all support, falling back to unordered processing\n"); 228 tool->ordered_events = false; 229 } 230 231 return session; 232 233 out_delete: 234 perf_session__delete(session); 235 out: 236 return ERR_PTR(ret); 237 } 238 239 static void perf_decomp__release_events(struct decomp *next) 240 { 241 struct decomp *decomp; 242 size_t mmap_len; 243 244 do { 245 decomp = next; 246 if (decomp == NULL) 247 break; 248 next = decomp->next; 249 mmap_len = decomp->mmap_len; 250 munmap(decomp, mmap_len); 251 } while (1); 252 } 253 254 void perf_session__delete(struct perf_session *session) 255 { 256 if (session == NULL) 257 return; 258 auxtrace__free(session); 259 auxtrace_index__free(&session->auxtrace_index); 260 debuginfo_cache__delete(); 261 perf_session__destroy_kernel_maps(session); 262 perf_decomp__release_events(session->decomp_data.decomp); 263 perf_env__exit(&session->header.env); 264 machines__exit(&session->machines); 265 if (session->data) { 266 if (perf_data__is_read(session->data)) 267 evlist__delete(session->evlist); 268 perf_data__close(session->data); 269 } 270 #ifdef HAVE_LIBTRACEEVENT 271 trace_event__cleanup(&session->tevent); 272 #endif 273 free(session); 274 } 275 276 static void swap_sample_id_all(union perf_event *event, void *data) 277 { 278 void *end = (void *) event + event->header.size; 279 int size = end - data; 280 281 BUG_ON(size % sizeof(u64)); 282 mem_bswap_64(data, size); 283 } 284 285 static void perf_event__all64_swap(union perf_event *event, 286 bool sample_id_all __maybe_unused) 287 { 288 struct perf_event_header *hdr = &event->header; 289 mem_bswap_64(hdr + 1, event->header.size - sizeof(*hdr)); 290 } 291 292 static void perf_event__comm_swap(union perf_event *event, bool sample_id_all) 293 { 294 event->comm.pid = bswap_32(event->comm.pid); 295 event->comm.tid = bswap_32(event->comm.tid); 296 297 if (sample_id_all) { 298 void *data = &event->comm.comm; 299 300 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 301 swap_sample_id_all(event, data); 302 } 303 } 304 305 static void perf_event__mmap_swap(union perf_event *event, 306 bool sample_id_all) 307 { 308 event->mmap.pid = bswap_32(event->mmap.pid); 309 event->mmap.tid = bswap_32(event->mmap.tid); 310 event->mmap.start = bswap_64(event->mmap.start); 311 event->mmap.len = bswap_64(event->mmap.len); 312 event->mmap.pgoff = bswap_64(event->mmap.pgoff); 313 314 if (sample_id_all) { 315 void *data = &event->mmap.filename; 316 317 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 318 swap_sample_id_all(event, data); 319 } 320 } 321 322 static void perf_event__mmap2_swap(union perf_event *event, 323 bool sample_id_all) 324 { 325 event->mmap2.pid = bswap_32(event->mmap2.pid); 326 event->mmap2.tid = bswap_32(event->mmap2.tid); 327 event->mmap2.start = bswap_64(event->mmap2.start); 328 event->mmap2.len = bswap_64(event->mmap2.len); 329 event->mmap2.pgoff = bswap_64(event->mmap2.pgoff); 330 331 if (!(event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID)) { 332 event->mmap2.maj = bswap_32(event->mmap2.maj); 333 event->mmap2.min = bswap_32(event->mmap2.min); 334 event->mmap2.ino = bswap_64(event->mmap2.ino); 335 event->mmap2.ino_generation = bswap_64(event->mmap2.ino_generation); 336 } 337 338 if (sample_id_all) { 339 void *data = &event->mmap2.filename; 340 341 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 342 swap_sample_id_all(event, data); 343 } 344 } 345 static void perf_event__task_swap(union perf_event *event, bool sample_id_all) 346 { 347 event->fork.pid = bswap_32(event->fork.pid); 348 event->fork.tid = bswap_32(event->fork.tid); 349 event->fork.ppid = bswap_32(event->fork.ppid); 350 event->fork.ptid = bswap_32(event->fork.ptid); 351 event->fork.time = bswap_64(event->fork.time); 352 353 if (sample_id_all) 354 swap_sample_id_all(event, &event->fork + 1); 355 } 356 357 static void perf_event__read_swap(union perf_event *event, bool sample_id_all) 358 { 359 event->read.pid = bswap_32(event->read.pid); 360 event->read.tid = bswap_32(event->read.tid); 361 event->read.value = bswap_64(event->read.value); 362 event->read.time_enabled = bswap_64(event->read.time_enabled); 363 event->read.time_running = bswap_64(event->read.time_running); 364 event->read.id = bswap_64(event->read.id); 365 366 if (sample_id_all) 367 swap_sample_id_all(event, &event->read + 1); 368 } 369 370 static void perf_event__aux_swap(union perf_event *event, bool sample_id_all) 371 { 372 event->aux.aux_offset = bswap_64(event->aux.aux_offset); 373 event->aux.aux_size = bswap_64(event->aux.aux_size); 374 event->aux.flags = bswap_64(event->aux.flags); 375 376 if (sample_id_all) 377 swap_sample_id_all(event, &event->aux + 1); 378 } 379 380 static void perf_event__itrace_start_swap(union perf_event *event, 381 bool sample_id_all) 382 { 383 event->itrace_start.pid = bswap_32(event->itrace_start.pid); 384 event->itrace_start.tid = bswap_32(event->itrace_start.tid); 385 386 if (sample_id_all) 387 swap_sample_id_all(event, &event->itrace_start + 1); 388 } 389 390 static void perf_event__switch_swap(union perf_event *event, bool sample_id_all) 391 { 392 if (event->header.type == PERF_RECORD_SWITCH_CPU_WIDE) { 393 event->context_switch.next_prev_pid = 394 bswap_32(event->context_switch.next_prev_pid); 395 event->context_switch.next_prev_tid = 396 bswap_32(event->context_switch.next_prev_tid); 397 } 398 399 if (sample_id_all) 400 swap_sample_id_all(event, &event->context_switch + 1); 401 } 402 403 static void perf_event__text_poke_swap(union perf_event *event, bool sample_id_all) 404 { 405 event->text_poke.addr = bswap_64(event->text_poke.addr); 406 event->text_poke.old_len = bswap_16(event->text_poke.old_len); 407 event->text_poke.new_len = bswap_16(event->text_poke.new_len); 408 409 if (sample_id_all) { 410 size_t len = sizeof(event->text_poke.old_len) + 411 sizeof(event->text_poke.new_len) + 412 event->text_poke.old_len + 413 event->text_poke.new_len; 414 void *data = &event->text_poke.old_len; 415 416 data += PERF_ALIGN(len, sizeof(u64)); 417 swap_sample_id_all(event, data); 418 } 419 } 420 421 static void perf_event__throttle_swap(union perf_event *event, 422 bool sample_id_all) 423 { 424 event->throttle.time = bswap_64(event->throttle.time); 425 event->throttle.id = bswap_64(event->throttle.id); 426 event->throttle.stream_id = bswap_64(event->throttle.stream_id); 427 428 if (sample_id_all) 429 swap_sample_id_all(event, &event->throttle + 1); 430 } 431 432 static void perf_event__namespaces_swap(union perf_event *event, 433 bool sample_id_all) 434 { 435 u64 i; 436 437 event->namespaces.pid = bswap_32(event->namespaces.pid); 438 event->namespaces.tid = bswap_32(event->namespaces.tid); 439 event->namespaces.nr_namespaces = bswap_64(event->namespaces.nr_namespaces); 440 441 for (i = 0; i < event->namespaces.nr_namespaces; i++) { 442 struct perf_ns_link_info *ns = &event->namespaces.link_info[i]; 443 444 ns->dev = bswap_64(ns->dev); 445 ns->ino = bswap_64(ns->ino); 446 } 447 448 if (sample_id_all) 449 swap_sample_id_all(event, &event->namespaces.link_info[i]); 450 } 451 452 static void perf_event__cgroup_swap(union perf_event *event, bool sample_id_all) 453 { 454 event->cgroup.id = bswap_64(event->cgroup.id); 455 456 if (sample_id_all) { 457 void *data = &event->cgroup.path; 458 459 data += PERF_ALIGN(strlen(data) + 1, sizeof(u64)); 460 swap_sample_id_all(event, data); 461 } 462 } 463 464 static u8 revbyte(u8 b) 465 { 466 int rev = (b >> 4) | ((b & 0xf) << 4); 467 rev = ((rev & 0xcc) >> 2) | ((rev & 0x33) << 2); 468 rev = ((rev & 0xaa) >> 1) | ((rev & 0x55) << 1); 469 return (u8) rev; 470 } 471 472 /* 473 * XXX this is hack in attempt to carry flags bitfield 474 * through endian village. ABI says: 475 * 476 * Bit-fields are allocated from right to left (least to most significant) 477 * on little-endian implementations and from left to right (most to least 478 * significant) on big-endian implementations. 479 * 480 * The above seems to be byte specific, so we need to reverse each 481 * byte of the bitfield. 'Internet' also says this might be implementation 482 * specific and we probably need proper fix and carry perf_event_attr 483 * bitfield flags in separate data file FEAT_ section. Thought this seems 484 * to work for now. 485 */ 486 static void swap_bitfield(u8 *p, unsigned len) 487 { 488 unsigned i; 489 490 for (i = 0; i < len; i++) { 491 *p = revbyte(*p); 492 p++; 493 } 494 } 495 496 /* exported for swapping attributes in file header */ 497 void perf_event__attr_swap(struct perf_event_attr *attr) 498 { 499 attr->type = bswap_32(attr->type); 500 attr->size = bswap_32(attr->size); 501 502 #define bswap_safe(f, n) \ 503 (attr->size > (offsetof(struct perf_event_attr, f) + \ 504 sizeof(attr->f) * (n))) 505 #define bswap_field(f, sz) \ 506 do { \ 507 if (bswap_safe(f, 0)) \ 508 attr->f = bswap_##sz(attr->f); \ 509 } while(0) 510 #define bswap_field_16(f) bswap_field(f, 16) 511 #define bswap_field_32(f) bswap_field(f, 32) 512 #define bswap_field_64(f) bswap_field(f, 64) 513 514 bswap_field_64(config); 515 bswap_field_64(sample_period); 516 bswap_field_64(sample_type); 517 bswap_field_64(read_format); 518 bswap_field_32(wakeup_events); 519 bswap_field_32(bp_type); 520 bswap_field_64(bp_addr); 521 bswap_field_64(bp_len); 522 bswap_field_64(branch_sample_type); 523 bswap_field_64(sample_regs_user); 524 bswap_field_32(sample_stack_user); 525 bswap_field_32(aux_watermark); 526 bswap_field_16(sample_max_stack); 527 bswap_field_32(aux_sample_size); 528 529 /* 530 * After read_format are bitfields. Check read_format because 531 * we are unable to use offsetof on bitfield. 532 */ 533 if (bswap_safe(read_format, 1)) 534 swap_bitfield((u8 *) (&attr->read_format + 1), 535 sizeof(u64)); 536 #undef bswap_field_64 537 #undef bswap_field_32 538 #undef bswap_field 539 #undef bswap_safe 540 } 541 542 static void perf_event__hdr_attr_swap(union perf_event *event, 543 bool sample_id_all __maybe_unused) 544 { 545 size_t size; 546 547 perf_event__attr_swap(&event->attr.attr); 548 549 size = event->header.size; 550 size -= perf_record_header_attr_id(event) - (void *)event; 551 mem_bswap_64(perf_record_header_attr_id(event), size); 552 } 553 554 static void perf_event__event_update_swap(union perf_event *event, 555 bool sample_id_all __maybe_unused) 556 { 557 event->event_update.type = bswap_64(event->event_update.type); 558 event->event_update.id = bswap_64(event->event_update.id); 559 } 560 561 static void perf_event__event_type_swap(union perf_event *event, 562 bool sample_id_all __maybe_unused) 563 { 564 event->event_type.event_type.event_id = 565 bswap_64(event->event_type.event_type.event_id); 566 } 567 568 static void perf_event__tracing_data_swap(union perf_event *event, 569 bool sample_id_all __maybe_unused) 570 { 571 event->tracing_data.size = bswap_32(event->tracing_data.size); 572 } 573 574 static void perf_event__auxtrace_info_swap(union perf_event *event, 575 bool sample_id_all __maybe_unused) 576 { 577 size_t size; 578 579 event->auxtrace_info.type = bswap_32(event->auxtrace_info.type); 580 581 size = event->header.size; 582 size -= (void *)&event->auxtrace_info.priv - (void *)event; 583 mem_bswap_64(event->auxtrace_info.priv, size); 584 } 585 586 static void perf_event__auxtrace_swap(union perf_event *event, 587 bool sample_id_all __maybe_unused) 588 { 589 event->auxtrace.size = bswap_64(event->auxtrace.size); 590 event->auxtrace.offset = bswap_64(event->auxtrace.offset); 591 event->auxtrace.reference = bswap_64(event->auxtrace.reference); 592 event->auxtrace.idx = bswap_32(event->auxtrace.idx); 593 event->auxtrace.tid = bswap_32(event->auxtrace.tid); 594 event->auxtrace.cpu = bswap_32(event->auxtrace.cpu); 595 } 596 597 static void perf_event__auxtrace_error_swap(union perf_event *event, 598 bool sample_id_all __maybe_unused) 599 { 600 event->auxtrace_error.type = bswap_32(event->auxtrace_error.type); 601 event->auxtrace_error.code = bswap_32(event->auxtrace_error.code); 602 event->auxtrace_error.cpu = bswap_32(event->auxtrace_error.cpu); 603 event->auxtrace_error.pid = bswap_32(event->auxtrace_error.pid); 604 event->auxtrace_error.tid = bswap_32(event->auxtrace_error.tid); 605 event->auxtrace_error.fmt = bswap_32(event->auxtrace_error.fmt); 606 event->auxtrace_error.ip = bswap_64(event->auxtrace_error.ip); 607 if (event->auxtrace_error.fmt) 608 event->auxtrace_error.time = bswap_64(event->auxtrace_error.time); 609 if (event->auxtrace_error.fmt >= 2) { 610 event->auxtrace_error.machine_pid = bswap_32(event->auxtrace_error.machine_pid); 611 event->auxtrace_error.vcpu = bswap_32(event->auxtrace_error.vcpu); 612 } 613 } 614 615 static void perf_event__thread_map_swap(union perf_event *event, 616 bool sample_id_all __maybe_unused) 617 { 618 unsigned i; 619 620 event->thread_map.nr = bswap_64(event->thread_map.nr); 621 622 for (i = 0; i < event->thread_map.nr; i++) 623 event->thread_map.entries[i].pid = bswap_64(event->thread_map.entries[i].pid); 624 } 625 626 static void perf_event__cpu_map_swap(union perf_event *event, 627 bool sample_id_all __maybe_unused) 628 { 629 struct perf_record_cpu_map_data *data = &event->cpu_map.data; 630 631 data->type = bswap_16(data->type); 632 633 switch (data->type) { 634 case PERF_CPU_MAP__CPUS: 635 data->cpus_data.nr = bswap_16(data->cpus_data.nr); 636 637 for (unsigned i = 0; i < data->cpus_data.nr; i++) 638 data->cpus_data.cpu[i] = bswap_16(data->cpus_data.cpu[i]); 639 break; 640 case PERF_CPU_MAP__MASK: 641 data->mask32_data.long_size = bswap_16(data->mask32_data.long_size); 642 643 switch (data->mask32_data.long_size) { 644 case 4: 645 data->mask32_data.nr = bswap_16(data->mask32_data.nr); 646 for (unsigned i = 0; i < data->mask32_data.nr; i++) 647 data->mask32_data.mask[i] = bswap_32(data->mask32_data.mask[i]); 648 break; 649 case 8: 650 data->mask64_data.nr = bswap_16(data->mask64_data.nr); 651 for (unsigned i = 0; i < data->mask64_data.nr; i++) 652 data->mask64_data.mask[i] = bswap_64(data->mask64_data.mask[i]); 653 break; 654 default: 655 pr_err("cpu_map swap: unsupported long size\n"); 656 } 657 break; 658 case PERF_CPU_MAP__RANGE_CPUS: 659 data->range_cpu_data.start_cpu = bswap_16(data->range_cpu_data.start_cpu); 660 data->range_cpu_data.end_cpu = bswap_16(data->range_cpu_data.end_cpu); 661 break; 662 default: 663 break; 664 } 665 } 666 667 static void perf_event__stat_config_swap(union perf_event *event, 668 bool sample_id_all __maybe_unused) 669 { 670 u64 size; 671 672 size = bswap_64(event->stat_config.nr) * sizeof(event->stat_config.data[0]); 673 size += 1; /* nr item itself */ 674 mem_bswap_64(&event->stat_config.nr, size); 675 } 676 677 static void perf_event__stat_swap(union perf_event *event, 678 bool sample_id_all __maybe_unused) 679 { 680 event->stat.id = bswap_64(event->stat.id); 681 event->stat.thread = bswap_32(event->stat.thread); 682 event->stat.cpu = bswap_32(event->stat.cpu); 683 event->stat.val = bswap_64(event->stat.val); 684 event->stat.ena = bswap_64(event->stat.ena); 685 event->stat.run = bswap_64(event->stat.run); 686 } 687 688 static void perf_event__stat_round_swap(union perf_event *event, 689 bool sample_id_all __maybe_unused) 690 { 691 event->stat_round.type = bswap_64(event->stat_round.type); 692 event->stat_round.time = bswap_64(event->stat_round.time); 693 } 694 695 static void perf_event__time_conv_swap(union perf_event *event, 696 bool sample_id_all __maybe_unused) 697 { 698 event->time_conv.time_shift = bswap_64(event->time_conv.time_shift); 699 event->time_conv.time_mult = bswap_64(event->time_conv.time_mult); 700 event->time_conv.time_zero = bswap_64(event->time_conv.time_zero); 701 702 if (event_contains(event->time_conv, time_cycles)) { 703 event->time_conv.time_cycles = bswap_64(event->time_conv.time_cycles); 704 event->time_conv.time_mask = bswap_64(event->time_conv.time_mask); 705 } 706 } 707 708 static void 709 perf_event__schedstat_cpu_swap(union perf_event *event __maybe_unused, 710 bool sample_id_all __maybe_unused) 711 { 712 /* FIXME */ 713 } 714 715 static void 716 perf_event__schedstat_domain_swap(union perf_event *event __maybe_unused, 717 bool sample_id_all __maybe_unused) 718 { 719 /* FIXME */ 720 } 721 722 typedef void (*perf_event__swap_op)(union perf_event *event, 723 bool sample_id_all); 724 725 static perf_event__swap_op perf_event__swap_ops[] = { 726 [PERF_RECORD_MMAP] = perf_event__mmap_swap, 727 [PERF_RECORD_MMAP2] = perf_event__mmap2_swap, 728 [PERF_RECORD_COMM] = perf_event__comm_swap, 729 [PERF_RECORD_FORK] = perf_event__task_swap, 730 [PERF_RECORD_EXIT] = perf_event__task_swap, 731 [PERF_RECORD_LOST] = perf_event__all64_swap, 732 [PERF_RECORD_READ] = perf_event__read_swap, 733 [PERF_RECORD_THROTTLE] = perf_event__throttle_swap, 734 [PERF_RECORD_UNTHROTTLE] = perf_event__throttle_swap, 735 [PERF_RECORD_SAMPLE] = perf_event__all64_swap, 736 [PERF_RECORD_AUX] = perf_event__aux_swap, 737 [PERF_RECORD_ITRACE_START] = perf_event__itrace_start_swap, 738 [PERF_RECORD_LOST_SAMPLES] = perf_event__all64_swap, 739 [PERF_RECORD_SWITCH] = perf_event__switch_swap, 740 [PERF_RECORD_SWITCH_CPU_WIDE] = perf_event__switch_swap, 741 [PERF_RECORD_NAMESPACES] = perf_event__namespaces_swap, 742 [PERF_RECORD_CGROUP] = perf_event__cgroup_swap, 743 [PERF_RECORD_TEXT_POKE] = perf_event__text_poke_swap, 744 [PERF_RECORD_AUX_OUTPUT_HW_ID] = perf_event__all64_swap, 745 [PERF_RECORD_CALLCHAIN_DEFERRED] = perf_event__all64_swap, 746 [PERF_RECORD_HEADER_ATTR] = perf_event__hdr_attr_swap, 747 [PERF_RECORD_HEADER_EVENT_TYPE] = perf_event__event_type_swap, 748 [PERF_RECORD_HEADER_TRACING_DATA] = perf_event__tracing_data_swap, 749 [PERF_RECORD_HEADER_BUILD_ID] = NULL, 750 [PERF_RECORD_ID_INDEX] = perf_event__all64_swap, 751 [PERF_RECORD_AUXTRACE_INFO] = perf_event__auxtrace_info_swap, 752 [PERF_RECORD_AUXTRACE] = perf_event__auxtrace_swap, 753 [PERF_RECORD_AUXTRACE_ERROR] = perf_event__auxtrace_error_swap, 754 [PERF_RECORD_THREAD_MAP] = perf_event__thread_map_swap, 755 [PERF_RECORD_CPU_MAP] = perf_event__cpu_map_swap, 756 [PERF_RECORD_STAT_CONFIG] = perf_event__stat_config_swap, 757 [PERF_RECORD_STAT] = perf_event__stat_swap, 758 [PERF_RECORD_STAT_ROUND] = perf_event__stat_round_swap, 759 [PERF_RECORD_EVENT_UPDATE] = perf_event__event_update_swap, 760 [PERF_RECORD_TIME_CONV] = perf_event__time_conv_swap, 761 [PERF_RECORD_SCHEDSTAT_CPU] = perf_event__schedstat_cpu_swap, 762 [PERF_RECORD_SCHEDSTAT_DOMAIN] = perf_event__schedstat_domain_swap, 763 [PERF_RECORD_HEADER_MAX] = NULL, 764 }; 765 766 /* 767 * When perf record finishes a pass on every buffers, it records this pseudo 768 * event. 769 * We record the max timestamp t found in the pass n. 770 * Assuming these timestamps are monotonic across cpus, we know that if 771 * a buffer still has events with timestamps below t, they will be all 772 * available and then read in the pass n + 1. 773 * Hence when we start to read the pass n + 2, we can safely flush every 774 * events with timestamps below t. 775 * 776 * ============ PASS n ================= 777 * CPU 0 | CPU 1 778 * | 779 * cnt1 timestamps | cnt2 timestamps 780 * 1 | 2 781 * 2 | 3 782 * - | 4 <--- max recorded 783 * 784 * ============ PASS n + 1 ============== 785 * CPU 0 | CPU 1 786 * | 787 * cnt1 timestamps | cnt2 timestamps 788 * 3 | 5 789 * 4 | 6 790 * 5 | 7 <---- max recorded 791 * 792 * Flush every events below timestamp 4 793 * 794 * ============ PASS n + 2 ============== 795 * CPU 0 | CPU 1 796 * | 797 * cnt1 timestamps | cnt2 timestamps 798 * 6 | 8 799 * 7 | 9 800 * - | 10 801 * 802 * Flush every events below timestamp 7 803 * etc... 804 */ 805 int perf_event__process_finished_round(const struct perf_tool *tool __maybe_unused, 806 union perf_event *event __maybe_unused, 807 struct ordered_events *oe) 808 { 809 if (dump_trace) 810 fprintf(stdout, "\n"); 811 return ordered_events__flush(oe, OE_FLUSH__ROUND); 812 } 813 814 int perf_session__queue_event(struct perf_session *s, union perf_event *event, 815 u64 timestamp, u64 file_offset, const char *file_path) 816 { 817 return ordered_events__queue(&s->ordered_events, event, timestamp, file_offset, file_path); 818 } 819 820 static void callchain__lbr_callstack_printf(struct perf_sample *sample) 821 { 822 struct ip_callchain *callchain = sample->callchain; 823 struct branch_stack *lbr_stack = sample->branch_stack; 824 struct branch_entry *entries = perf_sample__branch_entries(sample); 825 u64 kernel_callchain_nr = callchain->nr; 826 unsigned int i; 827 828 for (i = 0; i < kernel_callchain_nr; i++) { 829 if (callchain->ips[i] == PERF_CONTEXT_USER) 830 break; 831 } 832 833 if ((i != kernel_callchain_nr) && lbr_stack->nr) { 834 u64 total_nr; 835 /* 836 * LBR callstack can only get user call chain, 837 * i is kernel call chain number, 838 * 1 is PERF_CONTEXT_USER. 839 * 840 * The user call chain is stored in LBR registers. 841 * LBR are pair registers. The caller is stored 842 * in "from" register, while the callee is stored 843 * in "to" register. 844 * For example, there is a call stack 845 * "A"->"B"->"C"->"D". 846 * The LBR registers will be recorded like 847 * "C"->"D", "B"->"C", "A"->"B". 848 * So only the first "to" register and all "from" 849 * registers are needed to construct the whole stack. 850 */ 851 total_nr = i + 1 + lbr_stack->nr + 1; 852 kernel_callchain_nr = i + 1; 853 854 printf("... LBR call chain: nr:%" PRIu64 "\n", total_nr); 855 856 for (i = 0; i < kernel_callchain_nr; i++) 857 printf("..... %2d: %016" PRIx64 "\n", 858 i, callchain->ips[i]); 859 860 printf("..... %2d: %016" PRIx64 "\n", 861 (int)(kernel_callchain_nr), entries[0].to); 862 for (i = 0; i < lbr_stack->nr; i++) 863 printf("..... %2d: %016" PRIx64 "\n", 864 (int)(i + kernel_callchain_nr + 1), entries[i].from); 865 } 866 } 867 868 static const char *callchain_context_str(u64 ip) 869 { 870 switch (ip) { 871 case PERF_CONTEXT_HV: 872 return " (PERF_CONTEXT_HV)"; 873 case PERF_CONTEXT_KERNEL: 874 return " (PERF_CONTEXT_KERNEL)"; 875 case PERF_CONTEXT_USER: 876 return " (PERF_CONTEXT_USER)"; 877 case PERF_CONTEXT_GUEST: 878 return " (PERF_CONTEXT_GUEST)"; 879 case PERF_CONTEXT_GUEST_KERNEL: 880 return " (PERF_CONTEXT_GUEST_KERNEL)"; 881 case PERF_CONTEXT_GUEST_USER: 882 return " (PERF_CONTEXT_GUEST_USER)"; 883 case PERF_CONTEXT_USER_DEFERRED: 884 return " (PERF_CONTEXT_USER_DEFERRED)"; 885 default: 886 return ""; 887 } 888 } 889 890 static void callchain__printf(struct evsel *evsel, 891 struct perf_sample *sample) 892 { 893 unsigned int i; 894 struct ip_callchain *callchain = sample->callchain; 895 896 if (evsel__has_branch_callstack(evsel)) 897 callchain__lbr_callstack_printf(sample); 898 899 printf("... FP chain: nr:%" PRIu64 "\n", callchain->nr); 900 901 for (i = 0; i < callchain->nr; i++) 902 printf("..... %2d: %016" PRIx64 "%s\n", 903 i, callchain->ips[i], 904 callchain_context_str(callchain->ips[i])); 905 906 if (sample->deferred_callchain) 907 printf("...... (deferred)\n"); 908 } 909 910 static void branch_stack__printf(struct perf_sample *sample, 911 struct evsel *evsel) 912 { 913 struct branch_entry *entries = perf_sample__branch_entries(sample); 914 bool callstack = evsel__has_branch_callstack(evsel); 915 u64 *branch_stack_cntr = sample->branch_stack_cntr; 916 uint64_t i; 917 918 if (!callstack) { 919 printf("%s: nr:%" PRIu64 "\n", "... branch stack", sample->branch_stack->nr); 920 } else { 921 /* the reason of adding 1 to nr is because after expanding 922 * branch stack it generates nr + 1 callstack records. e.g., 923 * B()->C() 924 * A()->B() 925 * the final callstack should be: 926 * C() 927 * B() 928 * A() 929 */ 930 printf("%s: nr:%" PRIu64 "\n", "... branch callstack", sample->branch_stack->nr+1); 931 } 932 933 for (i = 0; i < sample->branch_stack->nr; i++) { 934 struct branch_entry *e = &entries[i]; 935 936 if (!callstack) { 937 printf("..... %2"PRIu64": %016" PRIx64 " -> %016" PRIx64 " %hu cycles %s%s%s%s %x %s %s\n", 938 i, e->from, e->to, 939 (unsigned short)e->flags.cycles, 940 e->flags.mispred ? "M" : " ", 941 e->flags.predicted ? "P" : " ", 942 e->flags.abort ? "A" : " ", 943 e->flags.in_tx ? "T" : " ", 944 (unsigned)e->flags.reserved, 945 get_branch_type(e), 946 e->flags.spec ? branch_spec_desc(e->flags.spec) : ""); 947 } else { 948 if (i == 0) { 949 printf("..... %2"PRIu64": %016" PRIx64 "\n" 950 "..... %2"PRIu64": %016" PRIx64 "\n", 951 i, e->to, i+1, e->from); 952 } else { 953 printf("..... %2"PRIu64": %016" PRIx64 "\n", i+1, e->from); 954 } 955 } 956 } 957 958 if (branch_stack_cntr) { 959 unsigned int br_cntr_width, br_cntr_nr; 960 961 perf_env__find_br_cntr_info(evsel__env(evsel), &br_cntr_nr, &br_cntr_width); 962 printf("... branch stack counters: nr:%" PRIu64 " (counter width: %u max counter nr:%u)\n", 963 sample->branch_stack->nr, br_cntr_width, br_cntr_nr); 964 for (i = 0; i < sample->branch_stack->nr; i++) 965 printf("..... %2"PRIu64": %016" PRIx64 "\n", i, branch_stack_cntr[i]); 966 } 967 } 968 969 static void regs_dump__printf(u64 mask, u64 *regs, uint16_t e_machine, uint32_t e_flags) 970 { 971 unsigned rid, i = 0; 972 973 for_each_set_bit(rid, (unsigned long *) &mask, sizeof(mask) * 8) { 974 u64 val = regs[i++]; 975 976 printf(".... %-5s 0x%016" PRIx64 "\n", 977 perf_reg_name(rid, e_machine, e_flags), val); 978 } 979 } 980 981 static const char *regs_abi[] = { 982 [PERF_SAMPLE_REGS_ABI_NONE] = "none", 983 [PERF_SAMPLE_REGS_ABI_32] = "32-bit", 984 [PERF_SAMPLE_REGS_ABI_64] = "64-bit", 985 }; 986 987 static inline const char *regs_dump_abi(struct regs_dump *d) 988 { 989 if (d->abi > PERF_SAMPLE_REGS_ABI_64) 990 return "unknown"; 991 992 return regs_abi[d->abi]; 993 } 994 995 static void regs__printf(const char *type, struct regs_dump *regs, 996 uint16_t e_machine, uint32_t e_flags) 997 { 998 u64 mask = regs->mask; 999 1000 printf("... %s regs: mask 0x%" PRIx64 " ABI %s\n", 1001 type, 1002 mask, 1003 regs_dump_abi(regs)); 1004 1005 regs_dump__printf(mask, regs->regs, e_machine, e_flags); 1006 } 1007 1008 static void regs_user__printf(struct perf_sample *sample, uint16_t e_machine, uint32_t e_flags) 1009 { 1010 struct regs_dump *user_regs; 1011 1012 if (!sample->user_regs) 1013 return; 1014 1015 user_regs = perf_sample__user_regs(sample); 1016 1017 if (user_regs->regs) 1018 regs__printf("user", user_regs, e_machine, e_flags); 1019 } 1020 1021 static void regs_intr__printf(struct perf_sample *sample, uint16_t e_machine, uint32_t e_flags) 1022 { 1023 struct regs_dump *intr_regs; 1024 1025 if (!sample->intr_regs) 1026 return; 1027 1028 intr_regs = perf_sample__intr_regs(sample); 1029 1030 if (intr_regs->regs) 1031 regs__printf("intr", intr_regs, e_machine, e_flags); 1032 } 1033 1034 static void stack_user__printf(struct stack_dump *dump) 1035 { 1036 printf("... ustack: size %" PRIu64 ", offset 0x%x\n", 1037 dump->size, dump->offset); 1038 } 1039 1040 static void evlist__print_tstamp(struct evlist *evlist, union perf_event *event, struct perf_sample *sample) 1041 { 1042 u64 sample_type = __evlist__combined_sample_type(evlist); 1043 1044 if (event->header.type != PERF_RECORD_SAMPLE && 1045 !evlist__sample_id_all(evlist)) { 1046 fputs("-1 -1 ", stdout); 1047 return; 1048 } 1049 1050 if ((sample_type & PERF_SAMPLE_CPU)) 1051 printf("%u ", sample->cpu); 1052 1053 if (sample_type & PERF_SAMPLE_TIME) 1054 printf("%" PRIu64 " ", sample->time); 1055 } 1056 1057 static void sample_read__printf(struct perf_sample *sample, u64 read_format) 1058 { 1059 printf("... sample_read:\n"); 1060 1061 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1062 printf("...... time enabled %016" PRIx64 "\n", 1063 sample->read.time_enabled); 1064 1065 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1066 printf("...... time running %016" PRIx64 "\n", 1067 sample->read.time_running); 1068 1069 if (read_format & PERF_FORMAT_GROUP) { 1070 struct sample_read_value *value = sample->read.group.values; 1071 1072 printf(".... group nr %" PRIu64 "\n", sample->read.group.nr); 1073 1074 sample_read_group__for_each(value, sample->read.group.nr, read_format) { 1075 printf("..... id %016" PRIx64 1076 ", value %016" PRIx64, 1077 value->id, value->value); 1078 if (read_format & PERF_FORMAT_LOST) 1079 printf(", lost %" PRIu64, value->lost); 1080 printf("\n"); 1081 } 1082 } else { 1083 printf("..... id %016" PRIx64 ", value %016" PRIx64, 1084 sample->read.one.id, sample->read.one.value); 1085 if (read_format & PERF_FORMAT_LOST) 1086 printf(", lost %" PRIu64, sample->read.one.lost); 1087 printf("\n"); 1088 } 1089 } 1090 1091 static void dump_event(struct evlist *evlist, union perf_event *event, 1092 u64 file_offset, struct perf_sample *sample, 1093 const char *file_path) 1094 { 1095 if (!dump_trace) 1096 return; 1097 1098 printf("\n%#" PRIx64 "@%s [%#x]: event: %d\n", 1099 file_offset, file_path, event->header.size, event->header.type); 1100 1101 trace_event(event); 1102 if (event->header.type == PERF_RECORD_SAMPLE && evlist->trace_event_sample_raw) 1103 evlist->trace_event_sample_raw(evlist, event, sample); 1104 1105 if (sample) 1106 evlist__print_tstamp(evlist, event, sample); 1107 1108 printf("%#" PRIx64 " [%#x]: PERF_RECORD_%s", file_offset, 1109 event->header.size, perf_event__name(event->header.type)); 1110 } 1111 1112 char *get_page_size_name(u64 size, char *str) 1113 { 1114 if (!size || !unit_number__scnprintf(str, PAGE_SIZE_NAME_LEN, size)) 1115 snprintf(str, PAGE_SIZE_NAME_LEN, "%s", "N/A"); 1116 1117 return str; 1118 } 1119 1120 static void dump_sample(struct machine *machine, struct evsel *evsel, union perf_event *event, 1121 struct perf_sample *sample) 1122 { 1123 u64 sample_type; 1124 char str[PAGE_SIZE_NAME_LEN]; 1125 uint16_t e_machine = EM_NONE; 1126 uint32_t e_flags = 0; 1127 1128 if (!dump_trace) 1129 return; 1130 1131 sample_type = evsel->core.attr.sample_type; 1132 1133 if (sample_type & (PERF_SAMPLE_REGS_USER | PERF_SAMPLE_REGS_INTR)) { 1134 struct thread *thread = machine__find_thread(machine, sample->pid, sample->pid); 1135 1136 e_machine = thread__e_machine(thread, machine, &e_flags); 1137 } 1138 1139 printf("(IP, 0x%x): %d/%d: %#" PRIx64 " period: %" PRIu64 " addr: %#" PRIx64 "\n", 1140 event->header.misc, sample->pid, sample->tid, sample->ip, 1141 sample->period, sample->addr); 1142 1143 if (evsel__has_callchain(evsel)) 1144 callchain__printf(evsel, sample); 1145 1146 if (evsel__has_br_stack(evsel)) 1147 branch_stack__printf(sample, evsel); 1148 1149 if (sample_type & PERF_SAMPLE_REGS_USER) 1150 regs_user__printf(sample, e_machine, e_flags); 1151 1152 if (sample_type & PERF_SAMPLE_REGS_INTR) 1153 regs_intr__printf(sample, e_machine, e_flags); 1154 1155 if (sample_type & PERF_SAMPLE_STACK_USER) 1156 stack_user__printf(&sample->user_stack); 1157 1158 if (sample_type & PERF_SAMPLE_WEIGHT_TYPE) { 1159 printf("... weight: %" PRIu64 "", sample->weight); 1160 if (sample_type & PERF_SAMPLE_WEIGHT_STRUCT) { 1161 printf(",0x%"PRIx16"", sample->ins_lat); 1162 printf(",0x%"PRIx16"", sample->weight3); 1163 } 1164 printf("\n"); 1165 } 1166 1167 if (sample_type & PERF_SAMPLE_DATA_SRC) 1168 printf(" . data_src: 0x%"PRIx64"\n", sample->data_src); 1169 1170 if (sample_type & PERF_SAMPLE_PHYS_ADDR) 1171 printf(" .. phys_addr: 0x%"PRIx64"\n", sample->phys_addr); 1172 1173 if (sample_type & PERF_SAMPLE_DATA_PAGE_SIZE) 1174 printf(" .. data page size: %s\n", get_page_size_name(sample->data_page_size, str)); 1175 1176 if (sample_type & PERF_SAMPLE_CODE_PAGE_SIZE) 1177 printf(" .. code page size: %s\n", get_page_size_name(sample->code_page_size, str)); 1178 1179 if (sample_type & PERF_SAMPLE_TRANSACTION) 1180 printf("... transaction: %" PRIx64 "\n", sample->transaction); 1181 1182 if (sample_type & PERF_SAMPLE_READ) 1183 sample_read__printf(sample, evsel->core.attr.read_format); 1184 } 1185 1186 static void dump_deferred_callchain(struct evsel *evsel, union perf_event *event, 1187 struct perf_sample *sample) 1188 { 1189 if (!dump_trace) 1190 return; 1191 1192 printf("(IP, 0x%x): %d/%d: %#" PRIx64 "\n", 1193 event->header.misc, sample->pid, sample->tid, sample->deferred_cookie); 1194 1195 if (evsel__has_callchain(evsel)) 1196 callchain__printf(evsel, sample); 1197 } 1198 1199 static void dump_read(struct evsel *evsel, union perf_event *event) 1200 { 1201 struct perf_record_read *read_event = &event->read; 1202 u64 read_format; 1203 1204 if (!dump_trace) 1205 return; 1206 1207 printf(": %d %d %s %" PRI_lu64 "\n", event->read.pid, event->read.tid, 1208 evsel__name(evsel), event->read.value); 1209 1210 if (!evsel) 1211 return; 1212 1213 read_format = evsel->core.attr.read_format; 1214 1215 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1216 printf("... time enabled : %" PRI_lu64 "\n", read_event->time_enabled); 1217 1218 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1219 printf("... time running : %" PRI_lu64 "\n", read_event->time_running); 1220 1221 if (read_format & PERF_FORMAT_ID) 1222 printf("... id : %" PRI_lu64 "\n", read_event->id); 1223 1224 if (read_format & PERF_FORMAT_LOST) 1225 printf("... lost : %" PRI_lu64 "\n", read_event->lost); 1226 } 1227 1228 static struct machine *machines__find_for_cpumode(struct machines *machines, 1229 union perf_event *event, 1230 struct perf_sample *sample) 1231 { 1232 if (perf_guest && 1233 ((sample->cpumode == PERF_RECORD_MISC_GUEST_KERNEL) || 1234 (sample->cpumode == PERF_RECORD_MISC_GUEST_USER))) { 1235 u32 pid; 1236 1237 if (sample->machine_pid) 1238 pid = sample->machine_pid; 1239 else if (event->header.type == PERF_RECORD_MMAP 1240 || event->header.type == PERF_RECORD_MMAP2) 1241 pid = event->mmap.pid; 1242 else 1243 pid = sample->pid; 1244 1245 /* 1246 * Guest code machine is created as needed and does not use 1247 * DEFAULT_GUEST_KERNEL_ID. 1248 */ 1249 if (symbol_conf.guest_code) 1250 return machines__findnew(machines, pid); 1251 1252 return machines__find_guest(machines, pid); 1253 } 1254 1255 return &machines->host; 1256 } 1257 1258 static int deliver_sample_value(struct evlist *evlist, 1259 const struct perf_tool *tool, 1260 union perf_event *event, 1261 struct perf_sample *sample, 1262 struct sample_read_value *v, 1263 struct machine *machine, 1264 bool per_thread) 1265 { 1266 struct perf_sample_id *sid = evlist__id2sid(evlist, v->id); 1267 struct evsel *saved_evsel = sample->evsel; 1268 u64 *storage = NULL; 1269 int ret; 1270 1271 if (sid) { 1272 storage = perf_sample_id__get_period_storage(sid, sample->tid, per_thread); 1273 } 1274 1275 if (storage) { 1276 sample->id = v->id; 1277 sample->period = v->value - *storage; 1278 *storage = v->value; 1279 } 1280 1281 if (!storage || sid->evsel == NULL) { 1282 ++evlist->stats.nr_unknown_id; 1283 return 0; 1284 } 1285 1286 /* 1287 * There's no reason to deliver sample 1288 * for zero period, bail out. 1289 */ 1290 if (!sample->period) 1291 return 0; 1292 1293 sample->evsel = container_of(sid->evsel, struct evsel, core); 1294 ret = tool->sample(tool, event, sample, sample->evsel, machine); 1295 sample->evsel = saved_evsel; 1296 return ret; 1297 } 1298 1299 static int deliver_sample_group(struct evlist *evlist, 1300 const struct perf_tool *tool, 1301 union perf_event *event, 1302 struct perf_sample *sample, 1303 struct machine *machine, 1304 u64 read_format, 1305 bool per_thread) 1306 { 1307 int ret = -EINVAL; 1308 struct sample_read_value *v = sample->read.group.values; 1309 1310 if (tool->dont_split_sample_group) 1311 return deliver_sample_value(evlist, tool, event, sample, v, machine, 1312 per_thread); 1313 1314 sample_read_group__for_each(v, sample->read.group.nr, read_format) { 1315 ret = deliver_sample_value(evlist, tool, event, sample, v, 1316 machine, per_thread); 1317 if (ret) 1318 break; 1319 } 1320 1321 return ret; 1322 } 1323 1324 static int evlist__deliver_sample(struct evlist *evlist, const struct perf_tool *tool, 1325 union perf_event *event, struct perf_sample *sample, 1326 struct evsel *evsel, struct machine *machine) 1327 { 1328 /* We know evsel != NULL. */ 1329 u64 sample_type = evsel->core.attr.sample_type; 1330 u64 read_format = evsel->core.attr.read_format; 1331 bool per_thread = perf_evsel__attr_has_per_thread_sample_period(&evsel->core); 1332 1333 /* Standard sample delivery. */ 1334 if (!(sample_type & PERF_SAMPLE_READ)) 1335 return tool->sample(tool, event, sample, evsel, machine); 1336 1337 /* For PERF_SAMPLE_READ we have either single or group mode. */ 1338 if (read_format & PERF_FORMAT_GROUP) 1339 return deliver_sample_group(evlist, tool, event, sample, 1340 machine, read_format, per_thread); 1341 else 1342 return deliver_sample_value(evlist, tool, event, sample, 1343 &sample->read.one, machine, 1344 per_thread); 1345 } 1346 1347 /* 1348 * Samples with deferred callchains should wait for the next matching 1349 * PERF_RECORD_CALLCHAIN_RECORD entries. Keep the events in a list and 1350 * deliver them once it finds the callchains. 1351 */ 1352 struct deferred_event { 1353 struct list_head list; 1354 union perf_event *event; 1355 }; 1356 1357 /* 1358 * This is called when a deferred callchain record comes up. Find all matching 1359 * samples, merge the callchains and process them. 1360 */ 1361 static int evlist__deliver_deferred_callchain(struct evlist *evlist, 1362 const struct perf_tool *tool, 1363 union perf_event *event, 1364 struct perf_sample *sample, 1365 struct machine *machine) 1366 { 1367 struct deferred_event *de, *tmp; 1368 int ret = 0; 1369 1370 if (!tool->merge_deferred_callchains) { 1371 struct evsel *saved_evsel = sample->evsel; 1372 1373 sample->evsel = evlist__id2evsel(evlist, sample->id); 1374 ret = tool->callchain_deferred(tool, event, sample, 1375 sample->evsel, machine); 1376 sample->evsel = saved_evsel; 1377 return ret; 1378 } 1379 1380 list_for_each_entry_safe(de, tmp, &evlist->deferred_samples, list) { 1381 struct perf_sample orig_sample; 1382 1383 perf_sample__init(&orig_sample, /*all=*/false); 1384 ret = evlist__parse_sample(evlist, de->event, &orig_sample); 1385 if (ret < 0) { 1386 pr_err("failed to parse original sample\n"); 1387 perf_sample__exit(&orig_sample); 1388 break; 1389 } 1390 1391 if (sample->tid != orig_sample.tid) { 1392 perf_sample__exit(&orig_sample); 1393 continue; 1394 } 1395 1396 if (event->callchain_deferred.cookie == orig_sample.deferred_cookie) 1397 sample__merge_deferred_callchain(&orig_sample, sample); 1398 else 1399 orig_sample.deferred_callchain = false; 1400 1401 orig_sample.evsel = evlist__id2evsel(evlist, orig_sample.id); 1402 ret = evlist__deliver_sample(evlist, tool, de->event, 1403 &orig_sample, orig_sample.evsel, machine); 1404 1405 perf_sample__exit(&orig_sample); 1406 list_del(&de->list); 1407 free(de->event); 1408 free(de); 1409 1410 if (ret) 1411 break; 1412 } 1413 return ret; 1414 } 1415 1416 /* 1417 * This is called at the end of the data processing for the session. Flush the 1418 * remaining samples as there's no hope for matching deferred callchains. 1419 */ 1420 static int session__flush_deferred_samples(struct perf_session *session, 1421 const struct perf_tool *tool) 1422 { 1423 struct evlist *evlist = session->evlist; 1424 struct machine *machine = &session->machines.host; 1425 struct deferred_event *de, *tmp; 1426 int ret = 0; 1427 1428 list_for_each_entry_safe(de, tmp, &evlist->deferred_samples, list) { 1429 struct perf_sample sample; 1430 1431 perf_sample__init(&sample, /*all=*/false); 1432 ret = evlist__parse_sample(evlist, de->event, &sample); 1433 if (ret < 0) { 1434 pr_err("failed to parse original sample\n"); 1435 perf_sample__exit(&sample); 1436 break; 1437 } 1438 1439 sample.evsel = evlist__id2evsel(evlist, sample.id); 1440 ret = evlist__deliver_sample(evlist, tool, de->event, 1441 &sample, sample.evsel, machine); 1442 1443 perf_sample__exit(&sample); 1444 list_del(&de->list); 1445 free(de->event); 1446 free(de); 1447 1448 if (ret) 1449 break; 1450 } 1451 return ret; 1452 } 1453 1454 static int machines__deliver_event(struct machines *machines, 1455 struct evlist *evlist, 1456 union perf_event *event, 1457 struct perf_sample *sample, 1458 const struct perf_tool *tool, u64 file_offset, 1459 const char *file_path) 1460 { 1461 struct evsel *evsel; 1462 struct machine *machine; 1463 1464 dump_event(evlist, event, file_offset, sample, file_path); 1465 1466 if (!sample->evsel) 1467 sample->evsel = evlist__id2evsel(evlist, sample->id); 1468 else 1469 assert(sample->evsel == evlist__id2evsel(evlist, sample->id)); 1470 1471 evsel = sample->evsel; 1472 machine = machines__find_for_cpumode(machines, event, sample); 1473 1474 switch (event->header.type) { 1475 case PERF_RECORD_SAMPLE: 1476 if (evsel == NULL) { 1477 ++evlist->stats.nr_unknown_id; 1478 return 0; 1479 } 1480 if (machine == NULL) { 1481 ++evlist->stats.nr_unprocessable_samples; 1482 dump_sample(machine, evsel, event, sample); 1483 return 0; 1484 } 1485 dump_sample(machine, evsel, event, sample); 1486 if (sample->deferred_callchain && tool->merge_deferred_callchains) { 1487 struct deferred_event *de = malloc(sizeof(*de)); 1488 size_t sz = event->header.size; 1489 1490 if (de == NULL) 1491 return -ENOMEM; 1492 1493 de->event = malloc(sz); 1494 if (de->event == NULL) { 1495 free(de); 1496 return -ENOMEM; 1497 } 1498 memcpy(de->event, event, sz); 1499 list_add_tail(&de->list, &evlist->deferred_samples); 1500 return 0; 1501 } 1502 return evlist__deliver_sample(evlist, tool, event, sample, evsel, machine); 1503 case PERF_RECORD_MMAP: 1504 return tool->mmap(tool, event, sample, machine); 1505 case PERF_RECORD_MMAP2: 1506 if (event->header.misc & PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT) 1507 ++evlist->stats.nr_proc_map_timeout; 1508 return tool->mmap2(tool, event, sample, machine); 1509 case PERF_RECORD_COMM: 1510 return tool->comm(tool, event, sample, machine); 1511 case PERF_RECORD_NAMESPACES: 1512 return tool->namespaces(tool, event, sample, machine); 1513 case PERF_RECORD_CGROUP: 1514 return tool->cgroup(tool, event, sample, machine); 1515 case PERF_RECORD_FORK: 1516 return tool->fork(tool, event, sample, machine); 1517 case PERF_RECORD_EXIT: 1518 return tool->exit(tool, event, sample, machine); 1519 case PERF_RECORD_LOST: 1520 if (tool->lost == perf_event__process_lost) 1521 evlist->stats.total_lost += event->lost.lost; 1522 return tool->lost(tool, event, sample, machine); 1523 case PERF_RECORD_LOST_SAMPLES: 1524 if (event->header.misc & PERF_RECORD_MISC_LOST_SAMPLES_BPF) 1525 evlist->stats.total_dropped_samples += event->lost_samples.lost; 1526 else if (tool->lost_samples == perf_event__process_lost_samples) 1527 evlist->stats.total_lost_samples += event->lost_samples.lost; 1528 return tool->lost_samples(tool, event, sample, machine); 1529 case PERF_RECORD_READ: 1530 dump_read(evsel, event); 1531 return tool->read(tool, event, sample, evsel, machine); 1532 case PERF_RECORD_THROTTLE: 1533 return tool->throttle(tool, event, sample, machine); 1534 case PERF_RECORD_UNTHROTTLE: 1535 return tool->unthrottle(tool, event, sample, machine); 1536 case PERF_RECORD_AUX: 1537 if (tool->aux == perf_event__process_aux) { 1538 if (event->aux.flags & PERF_AUX_FLAG_TRUNCATED) 1539 evlist->stats.total_aux_lost += 1; 1540 if (event->aux.flags & PERF_AUX_FLAG_PARTIAL) 1541 evlist->stats.total_aux_partial += 1; 1542 if (event->aux.flags & PERF_AUX_FLAG_COLLISION) 1543 evlist->stats.total_aux_collision += 1; 1544 } 1545 return tool->aux(tool, event, sample, machine); 1546 case PERF_RECORD_ITRACE_START: 1547 return tool->itrace_start(tool, event, sample, machine); 1548 case PERF_RECORD_SWITCH: 1549 case PERF_RECORD_SWITCH_CPU_WIDE: 1550 return tool->context_switch(tool, event, sample, machine); 1551 case PERF_RECORD_KSYMBOL: 1552 return tool->ksymbol(tool, event, sample, machine); 1553 case PERF_RECORD_BPF_EVENT: 1554 return tool->bpf(tool, event, sample, machine); 1555 case PERF_RECORD_TEXT_POKE: 1556 return tool->text_poke(tool, event, sample, machine); 1557 case PERF_RECORD_AUX_OUTPUT_HW_ID: 1558 return tool->aux_output_hw_id(tool, event, sample, machine); 1559 case PERF_RECORD_CALLCHAIN_DEFERRED: 1560 dump_deferred_callchain(evsel, event, sample); 1561 return evlist__deliver_deferred_callchain(evlist, tool, event, 1562 sample, machine); 1563 default: 1564 ++evlist->stats.nr_unknown_events; 1565 return -1; 1566 } 1567 } 1568 1569 static int perf_session__deliver_event(struct perf_session *session, 1570 union perf_event *event, 1571 const struct perf_tool *tool, 1572 u64 file_offset, 1573 const char *file_path) 1574 { 1575 struct perf_sample sample; 1576 int ret; 1577 1578 perf_sample__init(&sample, /*all=*/false); 1579 ret = evlist__parse_sample(session->evlist, event, &sample); 1580 if (ret) { 1581 pr_err("Can't parse sample, err = %d\n", ret); 1582 goto out; 1583 } 1584 1585 ret = auxtrace__process_event(session, event, &sample, tool); 1586 if (ret < 0) 1587 goto out; 1588 if (ret > 0) { 1589 ret = 0; 1590 goto out; 1591 } 1592 1593 ret = machines__deliver_event(&session->machines, session->evlist, 1594 event, &sample, tool, file_offset, file_path); 1595 1596 if (dump_trace && sample.aux_sample.size) 1597 auxtrace__dump_auxtrace_sample(session, &sample); 1598 out: 1599 perf_sample__exit(&sample); 1600 return ret; 1601 } 1602 1603 static s64 perf_session__process_user_event(struct perf_session *session, 1604 union perf_event *event, 1605 u64 file_offset, 1606 const char *file_path) 1607 { 1608 struct ordered_events *oe = &session->ordered_events; 1609 const struct perf_tool *tool = session->tool; 1610 struct perf_sample sample; 1611 int fd = perf_data__fd(session->data); 1612 s64 err; 1613 1614 perf_sample__init(&sample, /*all=*/true); 1615 if ((event->header.type != PERF_RECORD_COMPRESSED && 1616 event->header.type != PERF_RECORD_COMPRESSED2) || 1617 perf_tool__compressed_is_stub(tool)) 1618 dump_event(session->evlist, event, file_offset, &sample, file_path); 1619 1620 /* These events are processed right away */ 1621 switch (event->header.type) { 1622 case PERF_RECORD_HEADER_ATTR: 1623 err = tool->attr(tool, event, &session->evlist); 1624 if (err == 0) { 1625 perf_session__set_id_hdr_size(session); 1626 perf_session__set_comm_exec(session); 1627 } 1628 break; 1629 case PERF_RECORD_EVENT_UPDATE: 1630 err = tool->event_update(tool, event, &session->evlist); 1631 break; 1632 case PERF_RECORD_HEADER_EVENT_TYPE: 1633 /* 1634 * Deprecated, but we need to handle it for sake 1635 * of old data files create in pipe mode. 1636 */ 1637 err = 0; 1638 break; 1639 case PERF_RECORD_HEADER_TRACING_DATA: 1640 /* 1641 * Setup for reading amidst mmap, but only when we 1642 * are in 'file' mode. The 'pipe' fd is in proper 1643 * place already. 1644 */ 1645 if (!perf_data__is_pipe(session->data)) 1646 lseek(fd, file_offset, SEEK_SET); 1647 err = tool->tracing_data(tool, session, event); 1648 break; 1649 case PERF_RECORD_HEADER_BUILD_ID: 1650 err = tool->build_id(tool, session, event); 1651 break; 1652 case PERF_RECORD_FINISHED_ROUND: 1653 err = tool->finished_round(tool, event, oe); 1654 break; 1655 case PERF_RECORD_ID_INDEX: 1656 err = tool->id_index(tool, session, event); 1657 break; 1658 case PERF_RECORD_AUXTRACE_INFO: 1659 err = tool->auxtrace_info(tool, session, event); 1660 break; 1661 case PERF_RECORD_AUXTRACE: 1662 /* 1663 * Setup for reading amidst mmap, but only when we 1664 * are in 'file' mode. The 'pipe' fd is in proper 1665 * place already. 1666 */ 1667 if (!perf_data__is_pipe(session->data)) 1668 lseek(fd, file_offset + event->header.size, SEEK_SET); 1669 err = tool->auxtrace(tool, session, event); 1670 break; 1671 case PERF_RECORD_AUXTRACE_ERROR: 1672 perf_session__auxtrace_error_inc(session, event); 1673 err = tool->auxtrace_error(tool, session, event); 1674 break; 1675 case PERF_RECORD_THREAD_MAP: 1676 err = tool->thread_map(tool, session, event); 1677 break; 1678 case PERF_RECORD_CPU_MAP: 1679 err = tool->cpu_map(tool, session, event); 1680 break; 1681 case PERF_RECORD_STAT_CONFIG: 1682 err = tool->stat_config(tool, session, event); 1683 break; 1684 case PERF_RECORD_STAT: 1685 err = tool->stat(tool, session, event); 1686 break; 1687 case PERF_RECORD_STAT_ROUND: 1688 err = tool->stat_round(tool, session, event); 1689 break; 1690 case PERF_RECORD_TIME_CONV: 1691 session->time_conv = event->time_conv; 1692 err = tool->time_conv(tool, session, event); 1693 break; 1694 case PERF_RECORD_HEADER_FEATURE: 1695 err = tool->feature(tool, session, event); 1696 break; 1697 case PERF_RECORD_COMPRESSED: 1698 case PERF_RECORD_COMPRESSED2: 1699 err = tool->compressed(tool, session, event, file_offset, file_path); 1700 if (err) 1701 dump_event(session->evlist, event, file_offset, &sample, file_path); 1702 break; 1703 case PERF_RECORD_FINISHED_INIT: 1704 err = tool->finished_init(tool, session, event); 1705 break; 1706 case PERF_RECORD_BPF_METADATA: 1707 err = tool->bpf_metadata(tool, session, event); 1708 break; 1709 case PERF_RECORD_SCHEDSTAT_CPU: 1710 err = tool->schedstat_cpu(tool, session, event); 1711 break; 1712 case PERF_RECORD_SCHEDSTAT_DOMAIN: 1713 err = tool->schedstat_domain(tool, session, event); 1714 break; 1715 default: 1716 err = -EINVAL; 1717 break; 1718 } 1719 perf_sample__exit(&sample); 1720 return err; 1721 } 1722 1723 int perf_session__deliver_synth_event(struct perf_session *session, 1724 union perf_event *event, 1725 struct perf_sample *sample) 1726 { 1727 struct evlist *evlist = session->evlist; 1728 const struct perf_tool *tool = session->tool; 1729 1730 events_stats__inc(&evlist->stats, event->header.type); 1731 1732 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1733 return perf_session__process_user_event(session, event, 0, NULL); 1734 1735 return machines__deliver_event(&session->machines, evlist, event, sample, tool, 0, NULL); 1736 } 1737 1738 int perf_session__deliver_synth_attr_event(struct perf_session *session, 1739 const struct perf_event_attr *attr, 1740 u64 id) 1741 { 1742 union { 1743 struct { 1744 struct perf_record_header_attr attr; 1745 u64 ids[1]; 1746 } attr_id; 1747 union perf_event ev; 1748 } ev = { 1749 .attr_id.attr.header.type = PERF_RECORD_HEADER_ATTR, 1750 .attr_id.attr.header.size = sizeof(ev.attr_id), 1751 .attr_id.ids[0] = id, 1752 }; 1753 1754 if (attr->size != sizeof(ev.attr_id.attr.attr)) { 1755 pr_debug("Unexpected perf_event_attr size\n"); 1756 return -EINVAL; 1757 } 1758 ev.attr_id.attr.attr = *attr; 1759 return perf_session__deliver_synth_event(session, &ev.ev, NULL); 1760 } 1761 1762 static void event_swap(union perf_event *event, bool sample_id_all) 1763 { 1764 perf_event__swap_op swap; 1765 1766 swap = perf_event__swap_ops[event->header.type]; 1767 if (swap) 1768 swap(event, sample_id_all); 1769 } 1770 1771 int perf_session__peek_event(struct perf_session *session, off_t file_offset, 1772 void *buf, size_t buf_sz, 1773 union perf_event **event_ptr, 1774 struct perf_sample *sample) 1775 { 1776 union perf_event *event; 1777 size_t hdr_sz, rest; 1778 int fd; 1779 1780 if (session->one_mmap && !session->header.needs_swap) { 1781 event = file_offset - session->one_mmap_offset + 1782 session->one_mmap_addr; 1783 goto out_parse_sample; 1784 } 1785 1786 if (perf_data__is_pipe(session->data)) 1787 return -1; 1788 1789 fd = perf_data__fd(session->data); 1790 hdr_sz = sizeof(struct perf_event_header); 1791 1792 if (buf_sz < hdr_sz) 1793 return -1; 1794 1795 if (lseek(fd, file_offset, SEEK_SET) == (off_t)-1 || 1796 readn(fd, buf, hdr_sz) != (ssize_t)hdr_sz) 1797 return -1; 1798 1799 event = (union perf_event *)buf; 1800 1801 if (session->header.needs_swap) 1802 perf_event_header__bswap(&event->header); 1803 1804 if (event->header.size < hdr_sz || event->header.size > buf_sz) 1805 return -1; 1806 1807 buf += hdr_sz; 1808 rest = event->header.size - hdr_sz; 1809 1810 if (readn(fd, buf, rest) != (ssize_t)rest) 1811 return -1; 1812 1813 if (session->header.needs_swap) 1814 event_swap(event, evlist__sample_id_all(session->evlist)); 1815 1816 out_parse_sample: 1817 1818 if (sample && event->header.type < PERF_RECORD_USER_TYPE_START && 1819 evlist__parse_sample(session->evlist, event, sample)) 1820 return -1; 1821 1822 *event_ptr = event; 1823 1824 return 0; 1825 } 1826 1827 int perf_session__peek_events(struct perf_session *session, u64 offset, 1828 u64 size, peek_events_cb_t cb, void *data) 1829 { 1830 u64 max_offset = offset + size; 1831 char buf[PERF_SAMPLE_MAX_SIZE]; 1832 union perf_event *event; 1833 int err; 1834 1835 do { 1836 err = perf_session__peek_event(session, offset, buf, 1837 PERF_SAMPLE_MAX_SIZE, &event, 1838 NULL); 1839 if (err) 1840 return err; 1841 1842 err = cb(session, event, offset, data); 1843 if (err) 1844 return err; 1845 1846 offset += event->header.size; 1847 if (event->header.type == PERF_RECORD_AUXTRACE) 1848 offset += event->auxtrace.size; 1849 1850 } while (offset < max_offset); 1851 1852 return err; 1853 } 1854 1855 static s64 perf_session__process_event(struct perf_session *session, 1856 union perf_event *event, u64 file_offset, 1857 const char *file_path) 1858 { 1859 struct evlist *evlist = session->evlist; 1860 const struct perf_tool *tool = session->tool; 1861 int ret; 1862 1863 if (session->header.needs_swap) 1864 event_swap(event, evlist__sample_id_all(evlist)); 1865 1866 if (event->header.type >= PERF_RECORD_HEADER_MAX) { 1867 /* perf should not support unaligned event, stop here. */ 1868 if (event->header.size % sizeof(u64)) 1869 return -EINVAL; 1870 1871 /* This perf is outdated and does not support the latest event type. */ 1872 ui__warning("Unsupported header type %u, please consider updating perf.\n", 1873 event->header.type); 1874 /* Skip unsupported event by returning its size. */ 1875 return event->header.size; 1876 } 1877 1878 events_stats__inc(&evlist->stats, event->header.type); 1879 1880 if (event->header.type >= PERF_RECORD_USER_TYPE_START) 1881 return perf_session__process_user_event(session, event, file_offset, file_path); 1882 1883 if (tool->ordered_events) { 1884 u64 timestamp = -1ULL; 1885 1886 ret = evlist__parse_sample_timestamp(evlist, event, ×tamp); 1887 if (ret && ret != -1) 1888 return ret; 1889 1890 ret = perf_session__queue_event(session, event, timestamp, file_offset, file_path); 1891 if (ret != -ETIME) 1892 return ret; 1893 } 1894 1895 return perf_session__deliver_event(session, event, tool, file_offset, file_path); 1896 } 1897 1898 void perf_event_header__bswap(struct perf_event_header *hdr) 1899 { 1900 hdr->type = bswap_32(hdr->type); 1901 hdr->misc = bswap_16(hdr->misc); 1902 hdr->size = bswap_16(hdr->size); 1903 } 1904 1905 struct thread *perf_session__findnew(struct perf_session *session, pid_t pid) 1906 { 1907 return machine__findnew_thread(&session->machines.host, -1, pid); 1908 } 1909 1910 int perf_session__register_idle_thread(struct perf_session *session) 1911 { 1912 struct thread *thread = machine__idle_thread(&session->machines.host); 1913 1914 /* machine__idle_thread() got the thread, so put it */ 1915 thread__put(thread); 1916 return thread ? 0 : -1; 1917 } 1918 1919 static void 1920 perf_session__warn_order(const struct perf_session *session) 1921 { 1922 const struct ordered_events *oe = &session->ordered_events; 1923 struct evsel *evsel; 1924 bool should_warn = true; 1925 1926 evlist__for_each_entry(session->evlist, evsel) { 1927 if (evsel->core.attr.write_backward) 1928 should_warn = false; 1929 } 1930 1931 if (!should_warn) 1932 return; 1933 if (oe->nr_unordered_events != 0) 1934 ui__warning("%u out of order events recorded.\n", oe->nr_unordered_events); 1935 } 1936 1937 static void perf_session__warn_about_errors(const struct perf_session *session) 1938 { 1939 const struct events_stats *stats = &session->evlist->stats; 1940 1941 if (session->tool->lost == perf_event__process_lost && 1942 stats->nr_events[PERF_RECORD_LOST] != 0) { 1943 ui__warning("Processed %d events and lost %d chunks!\n\n" 1944 "Check IO/CPU overload!\n\n", 1945 stats->nr_events[0], 1946 stats->nr_events[PERF_RECORD_LOST]); 1947 } 1948 1949 if (session->tool->lost_samples == perf_event__process_lost_samples) { 1950 double drop_rate; 1951 1952 drop_rate = (double)stats->total_lost_samples / 1953 (double) (stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples); 1954 if (drop_rate > 0.05) { 1955 ui__warning("Processed %" PRIu64 " samples and lost %3.2f%%!\n\n", 1956 stats->nr_events[PERF_RECORD_SAMPLE] + stats->total_lost_samples, 1957 drop_rate * 100.0); 1958 } 1959 } 1960 1961 if (session->tool->aux == perf_event__process_aux && 1962 stats->total_aux_lost != 0) { 1963 ui__warning("AUX data lost %" PRIu64 " times out of %u!\n\n", 1964 stats->total_aux_lost, 1965 stats->nr_events[PERF_RECORD_AUX]); 1966 } 1967 1968 if (session->tool->aux == perf_event__process_aux && 1969 stats->total_aux_partial != 0) { 1970 bool vmm_exclusive = false; 1971 1972 (void)sysfs__read_bool("module/kvm_intel/parameters/vmm_exclusive", 1973 &vmm_exclusive); 1974 1975 ui__warning("AUX data had gaps in it %" PRIu64 " times out of %u!\n\n" 1976 "Are you running a KVM guest in the background?%s\n\n", 1977 stats->total_aux_partial, 1978 stats->nr_events[PERF_RECORD_AUX], 1979 vmm_exclusive ? 1980 "\nReloading kvm_intel module with vmm_exclusive=0\n" 1981 "will reduce the gaps to only guest's timeslices." : 1982 ""); 1983 } 1984 1985 if (session->tool->aux == perf_event__process_aux && 1986 stats->total_aux_collision != 0) { 1987 ui__warning("AUX data detected collision %" PRIu64 " times out of %u!\n\n", 1988 stats->total_aux_collision, 1989 stats->nr_events[PERF_RECORD_AUX]); 1990 } 1991 1992 if (stats->nr_unknown_events != 0) { 1993 ui__warning("Found %u unknown events!\n\n" 1994 "Is this an older tool processing a perf.data " 1995 "file generated by a more recent tool?\n\n" 1996 "If that is not the case, consider " 1997 "reporting to linux-kernel@vger.kernel.org.\n\n", 1998 stats->nr_unknown_events); 1999 } 2000 2001 if (stats->nr_unknown_id != 0) { 2002 ui__warning("%u samples with id not present in the header\n", 2003 stats->nr_unknown_id); 2004 } 2005 2006 if (stats->nr_invalid_chains != 0) { 2007 ui__warning("Found invalid callchains!\n\n" 2008 "%u out of %u events were discarded for this reason.\n\n" 2009 "Consider reporting to linux-kernel@vger.kernel.org.\n\n", 2010 stats->nr_invalid_chains, 2011 stats->nr_events[PERF_RECORD_SAMPLE]); 2012 } 2013 2014 if (stats->nr_unprocessable_samples != 0) { 2015 ui__warning("%u unprocessable samples recorded.\n" 2016 "Do you have a KVM guest running and not using 'perf kvm'?\n", 2017 stats->nr_unprocessable_samples); 2018 } 2019 2020 perf_session__warn_order(session); 2021 2022 events_stats__auxtrace_error_warn(stats); 2023 2024 if (stats->nr_proc_map_timeout != 0) { 2025 ui__warning("%d map information files for pre-existing threads were\n" 2026 "not processed, if there are samples for addresses they\n" 2027 "will not be resolved, you may find out which are these\n" 2028 "threads by running with -v and redirecting the output\n" 2029 "to a file.\n" 2030 "The time limit to process proc map is too short?\n" 2031 "Increase it by --proc-map-timeout\n", 2032 stats->nr_proc_map_timeout); 2033 } 2034 } 2035 2036 static int perf_session__flush_thread_stack(struct thread *thread, 2037 void *p __maybe_unused) 2038 { 2039 return thread_stack__flush(thread); 2040 } 2041 2042 static int perf_session__flush_thread_stacks(struct perf_session *session) 2043 { 2044 return machines__for_each_thread(&session->machines, 2045 perf_session__flush_thread_stack, 2046 NULL); 2047 } 2048 2049 volatile sig_atomic_t session_done; 2050 2051 static int __perf_session__process_decomp_events(struct perf_session *session); 2052 2053 static int __perf_session__process_pipe_events(struct perf_session *session) 2054 { 2055 struct ordered_events *oe = &session->ordered_events; 2056 const struct perf_tool *tool = session->tool; 2057 struct ui_progress prog; 2058 union perf_event *event; 2059 uint32_t size, cur_size = 0; 2060 void *buf = NULL; 2061 s64 skip = 0; 2062 u64 head; 2063 ssize_t err; 2064 void *p; 2065 bool update_prog = false; 2066 2067 /* 2068 * If it's from a file saving pipe data (by redirection), it would have 2069 * a file name other than "-". Then we can get the total size and show 2070 * the progress. 2071 */ 2072 if (strcmp(session->data->path, "-") && session->data->file.size) { 2073 ui_progress__init_size(&prog, session->data->file.size, 2074 "Processing events..."); 2075 update_prog = true; 2076 } 2077 2078 head = 0; 2079 cur_size = sizeof(union perf_event); 2080 2081 buf = malloc(cur_size); 2082 if (!buf) 2083 return -errno; 2084 ordered_events__set_copy_on_queue(oe, true); 2085 more: 2086 event = buf; 2087 err = perf_data__read(session->data, event, 2088 sizeof(struct perf_event_header)); 2089 if (err <= 0) { 2090 if (err == 0) 2091 goto done; 2092 2093 pr_err("failed to read event header\n"); 2094 goto out_err; 2095 } 2096 2097 if (session->header.needs_swap) 2098 perf_event_header__bswap(&event->header); 2099 2100 size = event->header.size; 2101 if (size < sizeof(struct perf_event_header)) { 2102 pr_err("bad event header size\n"); 2103 goto out_err; 2104 } 2105 2106 if (size > cur_size) { 2107 void *new = realloc(buf, size); 2108 if (!new) { 2109 pr_err("failed to allocate memory to read event\n"); 2110 goto out_err; 2111 } 2112 buf = new; 2113 cur_size = size; 2114 event = buf; 2115 } 2116 p = event; 2117 p += sizeof(struct perf_event_header); 2118 2119 if (size - sizeof(struct perf_event_header)) { 2120 err = perf_data__read(session->data, p, 2121 size - sizeof(struct perf_event_header)); 2122 if (err <= 0) { 2123 if (err == 0) { 2124 pr_err("unexpected end of event stream\n"); 2125 goto done; 2126 } 2127 2128 pr_err("failed to read event data\n"); 2129 goto out_err; 2130 } 2131 } 2132 2133 if ((skip = perf_session__process_event(session, event, head, "pipe")) < 0) { 2134 pr_err("%#" PRIx64 " [%#x]: piped event processing failed for event of type: %s (%d)\n", 2135 head, event->header.size, 2136 perf_event__name(event->header.type), 2137 event->header.type); 2138 err = -EINVAL; 2139 goto out_err; 2140 } 2141 2142 head += size; 2143 2144 if (skip > 0) 2145 head += skip; 2146 2147 err = __perf_session__process_decomp_events(session); 2148 if (err) 2149 goto out_err; 2150 2151 if (update_prog) 2152 ui_progress__update(&prog, size); 2153 2154 if (!session_done()) 2155 goto more; 2156 done: 2157 /* do the final flush for ordered samples */ 2158 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2159 if (err) 2160 goto out_err; 2161 err = session__flush_deferred_samples(session, tool); 2162 if (err) 2163 goto out_err; 2164 err = auxtrace__flush_events(session, tool); 2165 if (err) 2166 goto out_err; 2167 err = perf_session__flush_thread_stacks(session); 2168 out_err: 2169 free(buf); 2170 if (update_prog) 2171 ui_progress__finish(); 2172 if (!tool->no_warn) 2173 perf_session__warn_about_errors(session); 2174 ordered_events__free(&session->ordered_events); 2175 auxtrace__free_events(session); 2176 return err; 2177 } 2178 2179 static union perf_event * 2180 prefetch_event(char *buf, u64 head, size_t mmap_size, 2181 bool needs_swap, union perf_event *error) 2182 { 2183 union perf_event *event; 2184 u16 event_size; 2185 2186 /* 2187 * Ensure we have enough space remaining to read 2188 * the size of the event in the headers. 2189 */ 2190 if (head + sizeof(event->header) > mmap_size) 2191 return NULL; 2192 2193 event = (union perf_event *)(buf + head); 2194 if (needs_swap) 2195 perf_event_header__bswap(&event->header); 2196 2197 event_size = event->header.size; 2198 if (head + event_size <= mmap_size) 2199 return event; 2200 2201 /* We're not fetching the event so swap back again */ 2202 if (needs_swap) 2203 perf_event_header__bswap(&event->header); 2204 2205 /* Check if the event fits into the next mmapped buf. */ 2206 if (event_size <= mmap_size - head % page_size) { 2207 /* Remap buf and fetch again. */ 2208 return NULL; 2209 } 2210 2211 /* Invalid input. Event size should never exceed mmap_size. */ 2212 pr_debug("%s: head=%#" PRIx64 " event->header.size=%#x, mmap_size=%#zx:" 2213 " fuzzed or compressed perf.data?\n", __func__, head, event_size, mmap_size); 2214 2215 return error; 2216 } 2217 2218 static union perf_event * 2219 fetch_mmaped_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 2220 { 2221 return prefetch_event(buf, head, mmap_size, needs_swap, ERR_PTR(-EINVAL)); 2222 } 2223 2224 static union perf_event * 2225 fetch_decomp_event(u64 head, size_t mmap_size, char *buf, bool needs_swap) 2226 { 2227 return prefetch_event(buf, head, mmap_size, needs_swap, NULL); 2228 } 2229 2230 static int __perf_session__process_decomp_events(struct perf_session *session) 2231 { 2232 s64 skip; 2233 u64 size; 2234 struct decomp *decomp = session->active_decomp->decomp_last; 2235 2236 if (!decomp) 2237 return 0; 2238 2239 while (decomp->head < decomp->size && !session_done()) { 2240 union perf_event *event = fetch_decomp_event(decomp->head, decomp->size, decomp->data, 2241 session->header.needs_swap); 2242 2243 if (!event) 2244 break; 2245 2246 size = event->header.size; 2247 2248 if (size < sizeof(struct perf_event_header) || 2249 (skip = perf_session__process_event(session, event, decomp->file_pos, 2250 decomp->file_path)) < 0) { 2251 pr_err("%#" PRIx64 " [%#x]: decompress event processing failed for event of type: %s (%d)\n", 2252 decomp->file_pos + decomp->head, event->header.size, 2253 perf_event__name(event->header.type), 2254 event->header.type); 2255 return -EINVAL; 2256 } 2257 2258 if (skip) 2259 size += skip; 2260 2261 decomp->head += size; 2262 } 2263 2264 return 0; 2265 } 2266 2267 /* 2268 * On 64bit we can mmap the data file in one go. No need for tiny mmap 2269 * slices. On 32bit we use 32MB. 2270 */ 2271 #if BITS_PER_LONG == 64 2272 #define MMAP_SIZE ULLONG_MAX 2273 #define NUM_MMAPS 1 2274 #else 2275 #define MMAP_SIZE (32 * 1024 * 1024ULL) 2276 #define NUM_MMAPS 128 2277 #endif 2278 2279 struct reader; 2280 2281 typedef s64 (*reader_cb_t)(struct perf_session *session, 2282 union perf_event *event, 2283 u64 file_offset, 2284 const char *file_path); 2285 2286 struct reader { 2287 int fd; 2288 const char *path; 2289 u64 data_size; 2290 u64 data_offset; 2291 reader_cb_t process; 2292 bool in_place_update; 2293 char *mmaps[NUM_MMAPS]; 2294 size_t mmap_size; 2295 int mmap_idx; 2296 char *mmap_cur; 2297 u64 file_pos; 2298 u64 file_offset; 2299 u64 head; 2300 u64 size; 2301 bool done; 2302 struct zstd_data zstd_data; 2303 struct decomp_data decomp_data; 2304 }; 2305 2306 static int 2307 reader__init(struct reader *rd, bool *one_mmap) 2308 { 2309 u64 data_size = rd->data_size; 2310 char **mmaps = rd->mmaps; 2311 2312 rd->head = rd->data_offset; 2313 data_size += rd->data_offset; 2314 2315 rd->mmap_size = MMAP_SIZE; 2316 if (rd->mmap_size > data_size) { 2317 rd->mmap_size = data_size; 2318 if (one_mmap) 2319 *one_mmap = true; 2320 } 2321 2322 memset(mmaps, 0, sizeof(rd->mmaps)); 2323 2324 if (zstd_init(&rd->zstd_data, 0)) 2325 return -1; 2326 rd->decomp_data.zstd_decomp = &rd->zstd_data; 2327 2328 return 0; 2329 } 2330 2331 static void 2332 reader__release_decomp(struct reader *rd) 2333 { 2334 perf_decomp__release_events(rd->decomp_data.decomp); 2335 zstd_fini(&rd->zstd_data); 2336 } 2337 2338 static int 2339 reader__mmap(struct reader *rd, struct perf_session *session) 2340 { 2341 int mmap_prot, mmap_flags; 2342 char *buf, **mmaps = rd->mmaps; 2343 u64 page_offset; 2344 2345 mmap_prot = PROT_READ; 2346 mmap_flags = MAP_SHARED; 2347 2348 if (rd->in_place_update) { 2349 mmap_prot |= PROT_WRITE; 2350 } else if (session->header.needs_swap) { 2351 mmap_prot |= PROT_WRITE; 2352 mmap_flags = MAP_PRIVATE; 2353 } 2354 2355 if (mmaps[rd->mmap_idx]) { 2356 munmap(mmaps[rd->mmap_idx], rd->mmap_size); 2357 mmaps[rd->mmap_idx] = NULL; 2358 } 2359 2360 page_offset = page_size * (rd->head / page_size); 2361 rd->file_offset += page_offset; 2362 rd->head -= page_offset; 2363 2364 buf = mmap(NULL, rd->mmap_size, mmap_prot, mmap_flags, rd->fd, 2365 rd->file_offset); 2366 if (buf == MAP_FAILED) { 2367 pr_err("failed to mmap file\n"); 2368 return -errno; 2369 } 2370 mmaps[rd->mmap_idx] = rd->mmap_cur = buf; 2371 rd->mmap_idx = (rd->mmap_idx + 1) & (ARRAY_SIZE(rd->mmaps) - 1); 2372 rd->file_pos = rd->file_offset + rd->head; 2373 if (session->one_mmap) { 2374 session->one_mmap_addr = buf; 2375 session->one_mmap_offset = rd->file_offset; 2376 } 2377 2378 return 0; 2379 } 2380 2381 enum { 2382 READER_OK, 2383 READER_NODATA, 2384 }; 2385 2386 static int 2387 reader__read_event(struct reader *rd, struct perf_session *session, 2388 struct ui_progress *prog) 2389 { 2390 u64 size; 2391 int err = READER_OK; 2392 union perf_event *event; 2393 s64 skip; 2394 2395 event = fetch_mmaped_event(rd->head, rd->mmap_size, rd->mmap_cur, 2396 session->header.needs_swap); 2397 if (IS_ERR(event)) 2398 return PTR_ERR(event); 2399 2400 if (!event) 2401 return READER_NODATA; 2402 2403 size = event->header.size; 2404 2405 skip = -EINVAL; 2406 2407 if (size < sizeof(struct perf_event_header) || 2408 (skip = rd->process(session, event, rd->file_pos, rd->path)) < 0) { 2409 errno = -skip; 2410 pr_err("%#" PRIx64 " [%#x]: processing failed for event of type: %s (%d) [%m]\n", 2411 rd->file_offset + rd->head, event->header.size, 2412 perf_event__name(event->header.type), 2413 event->header.type); 2414 err = skip; 2415 goto out; 2416 } 2417 2418 if (skip) 2419 size += skip; 2420 2421 rd->size += size; 2422 rd->head += size; 2423 rd->file_pos += size; 2424 2425 err = __perf_session__process_decomp_events(session); 2426 if (err) 2427 goto out; 2428 2429 ui_progress__update(prog, size); 2430 2431 out: 2432 return err; 2433 } 2434 2435 static inline bool 2436 reader__eof(struct reader *rd) 2437 { 2438 return (rd->file_pos >= rd->data_size + rd->data_offset); 2439 } 2440 2441 static int 2442 reader__process_events(struct reader *rd, struct perf_session *session, 2443 struct ui_progress *prog) 2444 { 2445 int err; 2446 2447 err = reader__init(rd, &session->one_mmap); 2448 if (err) 2449 goto out; 2450 2451 session->active_decomp = &rd->decomp_data; 2452 2453 remap: 2454 err = reader__mmap(rd, session); 2455 if (err) 2456 goto out; 2457 2458 more: 2459 err = reader__read_event(rd, session, prog); 2460 if (err < 0) 2461 goto out; 2462 else if (err == READER_NODATA) 2463 goto remap; 2464 2465 if (session_done()) 2466 goto out; 2467 2468 if (!reader__eof(rd)) 2469 goto more; 2470 2471 out: 2472 session->active_decomp = &session->decomp_data; 2473 return err; 2474 } 2475 2476 static s64 process_simple(struct perf_session *session, 2477 union perf_event *event, 2478 u64 file_offset, 2479 const char *file_path) 2480 { 2481 return perf_session__process_event(session, event, file_offset, file_path); 2482 } 2483 2484 static int __perf_session__process_events(struct perf_session *session) 2485 { 2486 struct reader rd = { 2487 .fd = perf_data__fd(session->data), 2488 .path = session->data->file.path, 2489 .data_size = session->header.data_size, 2490 .data_offset = session->header.data_offset, 2491 .process = process_simple, 2492 .in_place_update = session->data->in_place_update, 2493 }; 2494 struct ordered_events *oe = &session->ordered_events; 2495 const struct perf_tool *tool = session->tool; 2496 struct ui_progress prog; 2497 int err; 2498 2499 if (rd.data_size == 0) 2500 return -1; 2501 2502 ui_progress__init_size(&prog, rd.data_size, "Processing events..."); 2503 2504 err = reader__process_events(&rd, session, &prog); 2505 if (err) 2506 goto out_err; 2507 /* do the final flush for ordered samples */ 2508 err = ordered_events__flush(oe, OE_FLUSH__FINAL); 2509 if (err) 2510 goto out_err; 2511 err = auxtrace__flush_events(session, tool); 2512 if (err) 2513 goto out_err; 2514 err = session__flush_deferred_samples(session, tool); 2515 if (err) 2516 goto out_err; 2517 err = perf_session__flush_thread_stacks(session); 2518 out_err: 2519 ui_progress__finish(); 2520 if (!tool->no_warn) 2521 perf_session__warn_about_errors(session); 2522 /* 2523 * We may switching perf.data output, make ordered_events 2524 * reusable. 2525 */ 2526 ordered_events__reinit(&session->ordered_events); 2527 auxtrace__free_events(session); 2528 reader__release_decomp(&rd); 2529 session->one_mmap = false; 2530 return err; 2531 } 2532 2533 /* 2534 * Processing 2 MB of data from each reader in sequence, 2535 * because that's the way the ordered events sorting works 2536 * most efficiently. 2537 */ 2538 #define READER_MAX_SIZE (2 * 1024 * 1024) 2539 2540 /* 2541 * This function reads, merge and process directory data. 2542 * It assumens the version 1 of directory data, where each 2543 * data file holds per-cpu data, already sorted by kernel. 2544 */ 2545 static int __perf_session__process_dir_events(struct perf_session *session) 2546 { 2547 struct perf_data *data = session->data; 2548 const struct perf_tool *tool = session->tool; 2549 int i, ret, readers, nr_readers; 2550 struct ui_progress prog; 2551 u64 total_size = perf_data__size(session->data); 2552 struct reader *rd; 2553 2554 ui_progress__init_size(&prog, total_size, "Processing events..."); 2555 2556 nr_readers = 1; 2557 for (i = 0; i < data->dir.nr; i++) { 2558 if (data->dir.files[i].size) 2559 nr_readers++; 2560 } 2561 2562 rd = calloc(nr_readers, sizeof(struct reader)); 2563 if (!rd) 2564 return -ENOMEM; 2565 2566 rd[0] = (struct reader) { 2567 .fd = perf_data__fd(session->data), 2568 .path = session->data->file.path, 2569 .data_size = session->header.data_size, 2570 .data_offset = session->header.data_offset, 2571 .process = process_simple, 2572 .in_place_update = session->data->in_place_update, 2573 }; 2574 ret = reader__init(&rd[0], NULL); 2575 if (ret) 2576 goto out_err; 2577 ret = reader__mmap(&rd[0], session); 2578 if (ret) 2579 goto out_err; 2580 readers = 1; 2581 2582 for (i = 0; i < data->dir.nr; i++) { 2583 if (!data->dir.files[i].size) 2584 continue; 2585 rd[readers] = (struct reader) { 2586 .fd = perf_data_file__fd(&data->dir.files[i]), 2587 .path = data->dir.files[i].path, 2588 .data_size = data->dir.files[i].size, 2589 .data_offset = 0, 2590 .process = process_simple, 2591 .in_place_update = session->data->in_place_update, 2592 }; 2593 ret = reader__init(&rd[readers], NULL); 2594 if (ret) 2595 goto out_err; 2596 ret = reader__mmap(&rd[readers], session); 2597 if (ret) 2598 goto out_err; 2599 readers++; 2600 } 2601 2602 i = 0; 2603 while (readers) { 2604 if (session_done()) 2605 break; 2606 2607 if (rd[i].done) { 2608 i = (i + 1) % nr_readers; 2609 continue; 2610 } 2611 if (reader__eof(&rd[i])) { 2612 rd[i].done = true; 2613 readers--; 2614 continue; 2615 } 2616 2617 session->active_decomp = &rd[i].decomp_data; 2618 ret = reader__read_event(&rd[i], session, &prog); 2619 if (ret < 0) { 2620 goto out_err; 2621 } else if (ret == READER_NODATA) { 2622 ret = reader__mmap(&rd[i], session); 2623 if (ret) 2624 goto out_err; 2625 } 2626 2627 if (rd[i].size >= READER_MAX_SIZE) { 2628 rd[i].size = 0; 2629 i = (i + 1) % nr_readers; 2630 } 2631 } 2632 2633 ret = ordered_events__flush(&session->ordered_events, OE_FLUSH__FINAL); 2634 if (ret) 2635 goto out_err; 2636 2637 ret = session__flush_deferred_samples(session, tool); 2638 if (ret) 2639 goto out_err; 2640 2641 ret = perf_session__flush_thread_stacks(session); 2642 out_err: 2643 ui_progress__finish(); 2644 2645 if (!tool->no_warn) 2646 perf_session__warn_about_errors(session); 2647 2648 /* 2649 * We may switching perf.data output, make ordered_events 2650 * reusable. 2651 */ 2652 ordered_events__reinit(&session->ordered_events); 2653 2654 session->one_mmap = false; 2655 2656 session->active_decomp = &session->decomp_data; 2657 for (i = 0; i < nr_readers; i++) 2658 reader__release_decomp(&rd[i]); 2659 zfree(&rd); 2660 2661 return ret; 2662 } 2663 2664 int perf_session__process_events(struct perf_session *session) 2665 { 2666 if (perf_session__register_idle_thread(session) < 0) 2667 return -ENOMEM; 2668 2669 if (perf_data__is_pipe(session->data)) 2670 return __perf_session__process_pipe_events(session); 2671 2672 if (perf_data__is_dir(session->data) && session->data->dir.nr) 2673 return __perf_session__process_dir_events(session); 2674 2675 return __perf_session__process_events(session); 2676 } 2677 2678 bool perf_session__has_traces(struct perf_session *session, const char *msg) 2679 { 2680 struct evsel *evsel; 2681 2682 evlist__for_each_entry(session->evlist, evsel) { 2683 if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT) 2684 return true; 2685 } 2686 2687 pr_err("No trace sample to read. Did you call 'perf %s'?\n", msg); 2688 return false; 2689 } 2690 2691 bool perf_session__has_switch_events(struct perf_session *session) 2692 { 2693 struct evsel *evsel; 2694 2695 evlist__for_each_entry(session->evlist, evsel) { 2696 if (evsel->core.attr.context_switch) 2697 return true; 2698 } 2699 2700 return false; 2701 } 2702 2703 int map__set_kallsyms_ref_reloc_sym(struct map *map, const char *symbol_name, u64 addr) 2704 { 2705 char *bracket, *name; 2706 struct ref_reloc_sym *ref; 2707 struct kmap *kmap; 2708 2709 ref = zalloc(sizeof(struct ref_reloc_sym)); 2710 if (ref == NULL) 2711 return -ENOMEM; 2712 2713 ref->name = name = strdup(symbol_name); 2714 if (ref->name == NULL) { 2715 free(ref); 2716 return -ENOMEM; 2717 } 2718 2719 bracket = strchr(name, ']'); 2720 if (bracket) 2721 *bracket = '\0'; 2722 2723 ref->addr = addr; 2724 2725 kmap = map__kmap(map); 2726 if (kmap) 2727 kmap->ref_reloc_sym = ref; 2728 2729 return 0; 2730 } 2731 2732 size_t perf_session__fprintf_dsos(struct perf_session *session, FILE *fp) 2733 { 2734 return machines__fprintf_dsos(&session->machines, fp); 2735 } 2736 2737 size_t perf_session__fprintf_dsos_buildid(struct perf_session *session, FILE *fp, 2738 bool (skip)(struct dso *dso, int parm), int parm) 2739 { 2740 return machines__fprintf_dsos_buildid(&session->machines, fp, skip, parm); 2741 } 2742 2743 size_t perf_session__fprintf_nr_events(struct perf_session *session, FILE *fp) 2744 { 2745 size_t ret; 2746 const char *msg = ""; 2747 2748 if (perf_header__has_feat(&session->header, HEADER_AUXTRACE)) 2749 msg = " (excludes AUX area (e.g. instruction trace) decoded / synthesized events)"; 2750 2751 ret = fprintf(fp, "\nAggregated stats:%s\n", msg); 2752 2753 ret += events_stats__fprintf(&session->evlist->stats, fp); 2754 return ret; 2755 } 2756 2757 size_t perf_session__fprintf(struct perf_session *session, FILE *fp) 2758 { 2759 size_t ret = machine__fprintf(&session->machines.host, fp); 2760 2761 for (struct rb_node *nd = rb_first_cached(&session->machines.guests); nd; nd = rb_next(nd)) { 2762 struct machine *pos = rb_entry(nd, struct machine, rb_node); 2763 2764 ret += machine__fprintf(pos, fp); 2765 } 2766 return ret; 2767 } 2768 2769 void perf_session__dump_kmaps(struct perf_session *session) 2770 { 2771 int save_verbose = verbose; 2772 2773 fflush(stdout); 2774 fprintf(stderr, "Kernel and module maps:\n"); 2775 verbose = 0; /* Suppress verbose to print a summary only */ 2776 maps__fprintf(machine__kernel_maps(&session->machines.host), stderr); 2777 verbose = save_verbose; 2778 } 2779 2780 struct evsel *perf_session__find_first_evtype(struct perf_session *session, 2781 unsigned int type) 2782 { 2783 struct evsel *pos; 2784 2785 evlist__for_each_entry(session->evlist, pos) { 2786 if (pos->core.attr.type == type) 2787 return pos; 2788 } 2789 return NULL; 2790 } 2791 2792 int perf_session__cpu_bitmap(struct perf_session *session, 2793 const char *cpu_list, unsigned long *cpu_bitmap) 2794 { 2795 unsigned int i; 2796 int err = -1; 2797 struct perf_cpu_map *map; 2798 int nr_cpus = min(perf_session__env(session)->nr_cpus_avail, MAX_NR_CPUS); 2799 struct perf_cpu cpu; 2800 2801 for (i = 0; i < PERF_TYPE_MAX; ++i) { 2802 struct evsel *evsel; 2803 2804 evsel = perf_session__find_first_evtype(session, i); 2805 if (!evsel) 2806 continue; 2807 2808 if (!(evsel->core.attr.sample_type & PERF_SAMPLE_CPU)) { 2809 pr_err("File does not contain CPU events. " 2810 "Remove -C option to proceed.\n"); 2811 return -1; 2812 } 2813 } 2814 2815 map = perf_cpu_map__new(cpu_list); 2816 if (map == NULL) { 2817 pr_err("Invalid cpu_list\n"); 2818 return -1; 2819 } 2820 2821 perf_cpu_map__for_each_cpu(cpu, i, map) { 2822 if (cpu.cpu >= nr_cpus) { 2823 pr_err("Requested CPU %d too large. " 2824 "Consider raising MAX_NR_CPUS\n", cpu.cpu); 2825 goto out_delete_map; 2826 } 2827 2828 __set_bit(cpu.cpu, cpu_bitmap); 2829 } 2830 2831 err = 0; 2832 2833 out_delete_map: 2834 perf_cpu_map__put(map); 2835 return err; 2836 } 2837 2838 void perf_session__fprintf_info(struct perf_session *session, FILE *fp, 2839 bool full) 2840 { 2841 if (session == NULL || fp == NULL) 2842 return; 2843 2844 fprintf(fp, "# ========\n"); 2845 perf_header__fprintf_info(session, fp, full); 2846 fprintf(fp, "# ========\n#\n"); 2847 } 2848 2849 static int perf_session__register_guest(struct perf_session *session, pid_t machine_pid) 2850 { 2851 struct machine *machine = machines__findnew(&session->machines, machine_pid); 2852 struct thread *thread; 2853 2854 if (!machine) 2855 return -ENOMEM; 2856 2857 machine->single_address_space = session->machines.host.single_address_space; 2858 2859 thread = machine__idle_thread(machine); 2860 if (!thread) 2861 return -ENOMEM; 2862 thread__put(thread); 2863 2864 machine->kallsyms_filename = perf_data__guest_kallsyms_name(session->data, machine_pid); 2865 2866 return 0; 2867 } 2868 2869 static int perf_session__set_guest_cpu(struct perf_session *session, pid_t pid, 2870 pid_t tid, int guest_cpu) 2871 { 2872 struct machine *machine = &session->machines.host; 2873 struct thread *thread = machine__findnew_thread(machine, pid, tid); 2874 2875 if (!thread) 2876 return -ENOMEM; 2877 thread__set_guest_cpu(thread, guest_cpu); 2878 thread__put(thread); 2879 2880 return 0; 2881 } 2882 2883 int perf_event__process_id_index(const struct perf_tool *tool __maybe_unused, 2884 struct perf_session *session, 2885 union perf_event *event) 2886 { 2887 struct evlist *evlist = session->evlist; 2888 struct perf_record_id_index *ie = &event->id_index; 2889 size_t sz = ie->header.size - sizeof(*ie); 2890 size_t i, nr, max_nr; 2891 size_t e1_sz = sizeof(struct id_index_entry); 2892 size_t e2_sz = sizeof(struct id_index_entry_2); 2893 size_t etot_sz = e1_sz + e2_sz; 2894 struct id_index_entry_2 *e2; 2895 pid_t last_pid = 0; 2896 2897 max_nr = sz / e1_sz; 2898 nr = ie->nr; 2899 if (nr > max_nr) { 2900 printf("Too big: nr %zu max_nr %zu\n", nr, max_nr); 2901 return -EINVAL; 2902 } 2903 2904 if (sz >= nr * etot_sz) { 2905 max_nr = sz / etot_sz; 2906 if (nr > max_nr) { 2907 printf("Too big2: nr %zu max_nr %zu\n", nr, max_nr); 2908 return -EINVAL; 2909 } 2910 e2 = (void *)ie + sizeof(*ie) + nr * e1_sz; 2911 } else { 2912 e2 = NULL; 2913 } 2914 2915 if (dump_trace) 2916 fprintf(stdout, " nr: %zu\n", nr); 2917 2918 for (i = 0; i < nr; i++, (e2 ? e2++ : 0)) { 2919 struct id_index_entry *e = &ie->entries[i]; 2920 struct perf_sample_id *sid; 2921 int ret; 2922 2923 if (dump_trace) { 2924 fprintf(stdout, " ... id: %"PRI_lu64, e->id); 2925 fprintf(stdout, " idx: %"PRI_lu64, e->idx); 2926 fprintf(stdout, " cpu: %"PRI_ld64, e->cpu); 2927 fprintf(stdout, " tid: %"PRI_ld64, e->tid); 2928 if (e2) { 2929 fprintf(stdout, " machine_pid: %"PRI_ld64, e2->machine_pid); 2930 fprintf(stdout, " vcpu: %"PRI_lu64"\n", e2->vcpu); 2931 } else { 2932 fprintf(stdout, "\n"); 2933 } 2934 } 2935 2936 sid = evlist__id2sid(evlist, e->id); 2937 if (!sid) 2938 return -ENOENT; 2939 2940 sid->idx = e->idx; 2941 sid->cpu.cpu = e->cpu; 2942 sid->tid = e->tid; 2943 2944 if (!e2) 2945 continue; 2946 2947 sid->machine_pid = e2->machine_pid; 2948 sid->vcpu.cpu = e2->vcpu; 2949 2950 if (!sid->machine_pid) 2951 continue; 2952 2953 if (sid->machine_pid != last_pid) { 2954 ret = perf_session__register_guest(session, sid->machine_pid); 2955 if (ret) 2956 return ret; 2957 last_pid = sid->machine_pid; 2958 perf_guest = true; 2959 } 2960 2961 ret = perf_session__set_guest_cpu(session, sid->machine_pid, e->tid, e2->vcpu); 2962 if (ret) 2963 return ret; 2964 } 2965 return 0; 2966 } 2967 2968 int perf_session__dsos_hit_all(struct perf_session *session) 2969 { 2970 struct rb_node *nd; 2971 int err; 2972 2973 err = machine__hit_all_dsos(&session->machines.host); 2974 if (err) 2975 return err; 2976 2977 for (nd = rb_first_cached(&session->machines.guests); nd; 2978 nd = rb_next(nd)) { 2979 struct machine *pos = rb_entry(nd, struct machine, rb_node); 2980 2981 err = machine__hit_all_dsos(pos); 2982 if (err) 2983 return err; 2984 } 2985 2986 return 0; 2987 } 2988 2989 struct perf_env *perf_session__env(struct perf_session *session) 2990 { 2991 return &session->header.env; 2992 } 2993 2994 struct perf_session__e_machine_cb_args { 2995 uint32_t e_flags; 2996 uint16_t e_machine; 2997 }; 2998 2999 static int perf_session__e_machine_cb(struct thread *thread, void *_args) 3000 { 3001 struct perf_session__e_machine_cb_args *args = _args; 3002 3003 args->e_machine = thread__e_machine(thread, /*machine=*/NULL, &args->e_flags); 3004 return args->e_machine != EM_NONE ? 1 : 0; 3005 } 3006 3007 /* 3008 * Note, a machine may have mixed 32-bit and 64-bit processes and so mixed 3009 * e_machines. Use thread__e_machine when this matters. 3010 */ 3011 uint16_t perf_session__e_machine(struct perf_session *session, uint32_t *e_flags) 3012 { 3013 struct perf_session__e_machine_cb_args args = { 3014 .e_machine = EM_NONE, 3015 }; 3016 struct perf_env *env; 3017 3018 if (!session) { 3019 /* Default to assuming a host machine. */ 3020 if (e_flags) 3021 *e_flags = EF_HOST; 3022 3023 return EM_HOST; 3024 } 3025 3026 env = perf_session__env(session); 3027 if (env && env->e_machine != EM_NONE) { 3028 if (e_flags) 3029 *e_flags = env->e_flags; 3030 3031 return env->e_machine; 3032 } 3033 3034 machines__for_each_thread(&session->machines, 3035 perf_session__e_machine_cb, 3036 &args); 3037 3038 if (args.e_machine != EM_NONE) { 3039 if (env) { 3040 env->e_machine = args.e_machine; 3041 env->e_flags = args.e_flags; 3042 } 3043 if (e_flags) 3044 *e_flags = args.e_flags; 3045 3046 return args.e_machine; 3047 } 3048 3049 /* 3050 * Couldn't determine from the perf_env or current set of 3051 * threads. Default to the host. 3052 */ 3053 if (e_flags) 3054 *e_flags = EF_HOST; 3055 3056 return EM_HOST; 3057 } 3058