1 #include <errno.h> 2 #include <fcntl.h> 3 #include <inttypes.h> 4 #include <linux/kernel.h> 5 #include <linux/types.h> 6 #include <perf/cpumap.h> 7 #include <sys/types.h> 8 #include <sys/stat.h> 9 #include <unistd.h> 10 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */ 11 #include <linux/perf_event.h> 12 #include <linux/zalloc.h> 13 #include "cpumap.h" 14 #include "dso.h" 15 #include "event.h" 16 #include "debug.h" 17 #include "hist.h" 18 #include "machine.h" 19 #include "sort.h" 20 #include "string2.h" 21 #include "strlist.h" 22 #include "thread.h" 23 #include "thread_map.h" 24 #include "time-utils.h" 25 #include <linux/ctype.h> 26 #include "map.h" 27 #include "util/namespaces.h" 28 #include "symbol.h" 29 #include "symbol/kallsyms.h" 30 #include "asm/bug.h" 31 #include "stat.h" 32 #include "session.h" 33 #include "bpf-event.h" 34 #include "tool.h" 35 #include "../perf.h" 36 37 static const char *perf_event__names[] = { 38 [0] = "TOTAL", 39 [PERF_RECORD_MMAP] = "MMAP", 40 [PERF_RECORD_MMAP2] = "MMAP2", 41 [PERF_RECORD_LOST] = "LOST", 42 [PERF_RECORD_COMM] = "COMM", 43 [PERF_RECORD_EXIT] = "EXIT", 44 [PERF_RECORD_THROTTLE] = "THROTTLE", 45 [PERF_RECORD_UNTHROTTLE] = "UNTHROTTLE", 46 [PERF_RECORD_FORK] = "FORK", 47 [PERF_RECORD_READ] = "READ", 48 [PERF_RECORD_SAMPLE] = "SAMPLE", 49 [PERF_RECORD_AUX] = "AUX", 50 [PERF_RECORD_ITRACE_START] = "ITRACE_START", 51 [PERF_RECORD_LOST_SAMPLES] = "LOST_SAMPLES", 52 [PERF_RECORD_SWITCH] = "SWITCH", 53 [PERF_RECORD_SWITCH_CPU_WIDE] = "SWITCH_CPU_WIDE", 54 [PERF_RECORD_NAMESPACES] = "NAMESPACES", 55 [PERF_RECORD_KSYMBOL] = "KSYMBOL", 56 [PERF_RECORD_BPF_EVENT] = "BPF_EVENT", 57 [PERF_RECORD_HEADER_ATTR] = "ATTR", 58 [PERF_RECORD_HEADER_EVENT_TYPE] = "EVENT_TYPE", 59 [PERF_RECORD_HEADER_TRACING_DATA] = "TRACING_DATA", 60 [PERF_RECORD_HEADER_BUILD_ID] = "BUILD_ID", 61 [PERF_RECORD_FINISHED_ROUND] = "FINISHED_ROUND", 62 [PERF_RECORD_ID_INDEX] = "ID_INDEX", 63 [PERF_RECORD_AUXTRACE_INFO] = "AUXTRACE_INFO", 64 [PERF_RECORD_AUXTRACE] = "AUXTRACE", 65 [PERF_RECORD_AUXTRACE_ERROR] = "AUXTRACE_ERROR", 66 [PERF_RECORD_THREAD_MAP] = "THREAD_MAP", 67 [PERF_RECORD_CPU_MAP] = "CPU_MAP", 68 [PERF_RECORD_STAT_CONFIG] = "STAT_CONFIG", 69 [PERF_RECORD_STAT] = "STAT", 70 [PERF_RECORD_STAT_ROUND] = "STAT_ROUND", 71 [PERF_RECORD_EVENT_UPDATE] = "EVENT_UPDATE", 72 [PERF_RECORD_TIME_CONV] = "TIME_CONV", 73 [PERF_RECORD_HEADER_FEATURE] = "FEATURE", 74 [PERF_RECORD_COMPRESSED] = "COMPRESSED", 75 }; 76 77 const char *perf_event__name(unsigned int id) 78 { 79 if (id >= ARRAY_SIZE(perf_event__names)) 80 return "INVALID"; 81 if (!perf_event__names[id]) 82 return "UNKNOWN"; 83 return perf_event__names[id]; 84 } 85 86 struct process_symbol_args { 87 const char *name; 88 u64 start; 89 }; 90 91 static int find_symbol_cb(void *arg, const char *name, char type, 92 u64 start) 93 { 94 struct process_symbol_args *args = arg; 95 96 /* 97 * Must be a function or at least an alias, as in PARISC64, where "_text" is 98 * an 'A' to the same address as "_stext". 99 */ 100 if (!(kallsyms__is_function(type) || 101 type == 'A') || strcmp(name, args->name)) 102 return 0; 103 104 args->start = start; 105 return 1; 106 } 107 108 int kallsyms__get_function_start(const char *kallsyms_filename, 109 const char *symbol_name, u64 *addr) 110 { 111 struct process_symbol_args args = { .name = symbol_name, }; 112 113 if (kallsyms__parse(kallsyms_filename, &args, find_symbol_cb) <= 0) 114 return -1; 115 116 *addr = args.start; 117 return 0; 118 } 119 120 void perf_event__read_stat_config(struct perf_stat_config *config, 121 struct perf_record_stat_config *event) 122 { 123 unsigned i; 124 125 for (i = 0; i < event->nr; i++) { 126 127 switch (event->data[i].tag) { 128 #define CASE(__term, __val) \ 129 case PERF_STAT_CONFIG_TERM__##__term: \ 130 config->__val = event->data[i].val; \ 131 break; 132 133 CASE(AGGR_MODE, aggr_mode) 134 CASE(SCALE, scale) 135 CASE(INTERVAL, interval) 136 #undef CASE 137 default: 138 pr_warning("unknown stat config term %" PRI_lu64 "\n", 139 event->data[i].tag); 140 } 141 } 142 } 143 144 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp) 145 { 146 const char *s; 147 148 if (event->header.misc & PERF_RECORD_MISC_COMM_EXEC) 149 s = " exec"; 150 else 151 s = ""; 152 153 return fprintf(fp, "%s: %s:%d/%d\n", s, event->comm.comm, event->comm.pid, event->comm.tid); 154 } 155 156 size_t perf_event__fprintf_namespaces(union perf_event *event, FILE *fp) 157 { 158 size_t ret = 0; 159 struct perf_ns_link_info *ns_link_info; 160 u32 nr_namespaces, idx; 161 162 ns_link_info = event->namespaces.link_info; 163 nr_namespaces = event->namespaces.nr_namespaces; 164 165 ret += fprintf(fp, " %d/%d - nr_namespaces: %u\n\t\t[", 166 event->namespaces.pid, 167 event->namespaces.tid, 168 nr_namespaces); 169 170 for (idx = 0; idx < nr_namespaces; idx++) { 171 if (idx && (idx % 4 == 0)) 172 ret += fprintf(fp, "\n\t\t "); 173 174 ret += fprintf(fp, "%u/%s: %" PRIu64 "/%#" PRIx64 "%s", idx, 175 perf_ns__name(idx), (u64)ns_link_info[idx].dev, 176 (u64)ns_link_info[idx].ino, 177 ((idx + 1) != nr_namespaces) ? ", " : "]\n"); 178 } 179 180 return ret; 181 } 182 183 int perf_event__process_comm(struct perf_tool *tool __maybe_unused, 184 union perf_event *event, 185 struct perf_sample *sample, 186 struct machine *machine) 187 { 188 return machine__process_comm_event(machine, event, sample); 189 } 190 191 int perf_event__process_namespaces(struct perf_tool *tool __maybe_unused, 192 union perf_event *event, 193 struct perf_sample *sample, 194 struct machine *machine) 195 { 196 return machine__process_namespaces_event(machine, event, sample); 197 } 198 199 int perf_event__process_lost(struct perf_tool *tool __maybe_unused, 200 union perf_event *event, 201 struct perf_sample *sample, 202 struct machine *machine) 203 { 204 return machine__process_lost_event(machine, event, sample); 205 } 206 207 int perf_event__process_aux(struct perf_tool *tool __maybe_unused, 208 union perf_event *event, 209 struct perf_sample *sample __maybe_unused, 210 struct machine *machine) 211 { 212 return machine__process_aux_event(machine, event); 213 } 214 215 int perf_event__process_itrace_start(struct perf_tool *tool __maybe_unused, 216 union perf_event *event, 217 struct perf_sample *sample __maybe_unused, 218 struct machine *machine) 219 { 220 return machine__process_itrace_start_event(machine, event); 221 } 222 223 int perf_event__process_lost_samples(struct perf_tool *tool __maybe_unused, 224 union perf_event *event, 225 struct perf_sample *sample, 226 struct machine *machine) 227 { 228 return machine__process_lost_samples_event(machine, event, sample); 229 } 230 231 int perf_event__process_switch(struct perf_tool *tool __maybe_unused, 232 union perf_event *event, 233 struct perf_sample *sample __maybe_unused, 234 struct machine *machine) 235 { 236 return machine__process_switch_event(machine, event); 237 } 238 239 int perf_event__process_ksymbol(struct perf_tool *tool __maybe_unused, 240 union perf_event *event, 241 struct perf_sample *sample __maybe_unused, 242 struct machine *machine) 243 { 244 return machine__process_ksymbol(machine, event, sample); 245 } 246 247 int perf_event__process_bpf(struct perf_tool *tool __maybe_unused, 248 union perf_event *event, 249 struct perf_sample *sample, 250 struct machine *machine) 251 { 252 return machine__process_bpf(machine, event, sample); 253 } 254 255 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp) 256 { 257 return fprintf(fp, " %d/%d: [%#" PRI_lx64 "(%#" PRI_lx64 ") @ %#" PRI_lx64 "]: %c %s\n", 258 event->mmap.pid, event->mmap.tid, event->mmap.start, 259 event->mmap.len, event->mmap.pgoff, 260 (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) ? 'r' : 'x', 261 event->mmap.filename); 262 } 263 264 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp) 265 { 266 return fprintf(fp, " %d/%d: [%#" PRI_lx64 "(%#" PRI_lx64 ") @ %#" PRI_lx64 267 " %02x:%02x %"PRI_lu64" %"PRI_lu64"]: %c%c%c%c %s\n", 268 event->mmap2.pid, event->mmap2.tid, event->mmap2.start, 269 event->mmap2.len, event->mmap2.pgoff, event->mmap2.maj, 270 event->mmap2.min, event->mmap2.ino, 271 event->mmap2.ino_generation, 272 (event->mmap2.prot & PROT_READ) ? 'r' : '-', 273 (event->mmap2.prot & PROT_WRITE) ? 'w' : '-', 274 (event->mmap2.prot & PROT_EXEC) ? 'x' : '-', 275 (event->mmap2.flags & MAP_SHARED) ? 's' : 'p', 276 event->mmap2.filename); 277 } 278 279 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp) 280 { 281 struct perf_thread_map *threads = thread_map__new_event(&event->thread_map); 282 size_t ret; 283 284 ret = fprintf(fp, " nr: "); 285 286 if (threads) 287 ret += thread_map__fprintf(threads, fp); 288 else 289 ret += fprintf(fp, "failed to get threads from event\n"); 290 291 perf_thread_map__put(threads); 292 return ret; 293 } 294 295 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp) 296 { 297 struct perf_cpu_map *cpus = cpu_map__new_data(&event->cpu_map.data); 298 size_t ret; 299 300 ret = fprintf(fp, ": "); 301 302 if (cpus) 303 ret += cpu_map__fprintf(cpus, fp); 304 else 305 ret += fprintf(fp, "failed to get cpumap from event\n"); 306 307 perf_cpu_map__put(cpus); 308 return ret; 309 } 310 311 int perf_event__process_mmap(struct perf_tool *tool __maybe_unused, 312 union perf_event *event, 313 struct perf_sample *sample, 314 struct machine *machine) 315 { 316 return machine__process_mmap_event(machine, event, sample); 317 } 318 319 int perf_event__process_mmap2(struct perf_tool *tool __maybe_unused, 320 union perf_event *event, 321 struct perf_sample *sample, 322 struct machine *machine) 323 { 324 return machine__process_mmap2_event(machine, event, sample); 325 } 326 327 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp) 328 { 329 return fprintf(fp, "(%d:%d):(%d:%d)\n", 330 event->fork.pid, event->fork.tid, 331 event->fork.ppid, event->fork.ptid); 332 } 333 334 int perf_event__process_fork(struct perf_tool *tool __maybe_unused, 335 union perf_event *event, 336 struct perf_sample *sample, 337 struct machine *machine) 338 { 339 return machine__process_fork_event(machine, event, sample); 340 } 341 342 int perf_event__process_exit(struct perf_tool *tool __maybe_unused, 343 union perf_event *event, 344 struct perf_sample *sample, 345 struct machine *machine) 346 { 347 return machine__process_exit_event(machine, event, sample); 348 } 349 350 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp) 351 { 352 return fprintf(fp, " offset: %#"PRI_lx64" size: %#"PRI_lx64" flags: %#"PRI_lx64" [%s%s%s]\n", 353 event->aux.aux_offset, event->aux.aux_size, 354 event->aux.flags, 355 event->aux.flags & PERF_AUX_FLAG_TRUNCATED ? "T" : "", 356 event->aux.flags & PERF_AUX_FLAG_OVERWRITE ? "O" : "", 357 event->aux.flags & PERF_AUX_FLAG_PARTIAL ? "P" : ""); 358 } 359 360 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp) 361 { 362 return fprintf(fp, " pid: %u tid: %u\n", 363 event->itrace_start.pid, event->itrace_start.tid); 364 } 365 366 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp) 367 { 368 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT; 369 const char *in_out = !out ? "IN " : 370 !(event->header.misc & PERF_RECORD_MISC_SWITCH_OUT_PREEMPT) ? 371 "OUT " : "OUT preempt"; 372 373 if (event->header.type == PERF_RECORD_SWITCH) 374 return fprintf(fp, " %s\n", in_out); 375 376 return fprintf(fp, " %s %s pid/tid: %5u/%-5u\n", 377 in_out, out ? "next" : "prev", 378 event->context_switch.next_prev_pid, 379 event->context_switch.next_prev_tid); 380 } 381 382 static size_t perf_event__fprintf_lost(union perf_event *event, FILE *fp) 383 { 384 return fprintf(fp, " lost %" PRI_lu64 "\n", event->lost.lost); 385 } 386 387 size_t perf_event__fprintf_ksymbol(union perf_event *event, FILE *fp) 388 { 389 return fprintf(fp, " addr %" PRI_lx64 " len %u type %u flags 0x%x name %s\n", 390 event->ksymbol.addr, event->ksymbol.len, 391 event->ksymbol.ksym_type, 392 event->ksymbol.flags, event->ksymbol.name); 393 } 394 395 size_t perf_event__fprintf_bpf(union perf_event *event, FILE *fp) 396 { 397 return fprintf(fp, " type %u, flags %u, id %u\n", 398 event->bpf.type, event->bpf.flags, event->bpf.id); 399 } 400 401 size_t perf_event__fprintf(union perf_event *event, FILE *fp) 402 { 403 size_t ret = fprintf(fp, "PERF_RECORD_%s", 404 perf_event__name(event->header.type)); 405 406 switch (event->header.type) { 407 case PERF_RECORD_COMM: 408 ret += perf_event__fprintf_comm(event, fp); 409 break; 410 case PERF_RECORD_FORK: 411 case PERF_RECORD_EXIT: 412 ret += perf_event__fprintf_task(event, fp); 413 break; 414 case PERF_RECORD_MMAP: 415 ret += perf_event__fprintf_mmap(event, fp); 416 break; 417 case PERF_RECORD_NAMESPACES: 418 ret += perf_event__fprintf_namespaces(event, fp); 419 break; 420 case PERF_RECORD_MMAP2: 421 ret += perf_event__fprintf_mmap2(event, fp); 422 break; 423 case PERF_RECORD_AUX: 424 ret += perf_event__fprintf_aux(event, fp); 425 break; 426 case PERF_RECORD_ITRACE_START: 427 ret += perf_event__fprintf_itrace_start(event, fp); 428 break; 429 case PERF_RECORD_SWITCH: 430 case PERF_RECORD_SWITCH_CPU_WIDE: 431 ret += perf_event__fprintf_switch(event, fp); 432 break; 433 case PERF_RECORD_LOST: 434 ret += perf_event__fprintf_lost(event, fp); 435 break; 436 case PERF_RECORD_KSYMBOL: 437 ret += perf_event__fprintf_ksymbol(event, fp); 438 break; 439 case PERF_RECORD_BPF_EVENT: 440 ret += perf_event__fprintf_bpf(event, fp); 441 break; 442 default: 443 ret += fprintf(fp, "\n"); 444 } 445 446 return ret; 447 } 448 449 int perf_event__process(struct perf_tool *tool __maybe_unused, 450 union perf_event *event, 451 struct perf_sample *sample, 452 struct machine *machine) 453 { 454 return machine__process_event(machine, event, sample); 455 } 456 457 struct map *thread__find_map(struct thread *thread, u8 cpumode, u64 addr, 458 struct addr_location *al) 459 { 460 struct map_groups *mg = thread->mg; 461 struct machine *machine = mg->machine; 462 bool load_map = false; 463 464 al->machine = machine; 465 al->thread = thread; 466 al->addr = addr; 467 al->cpumode = cpumode; 468 al->filtered = 0; 469 470 if (machine == NULL) { 471 al->map = NULL; 472 return NULL; 473 } 474 475 if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) { 476 al->level = 'k'; 477 mg = &machine->kmaps; 478 load_map = true; 479 } else if (cpumode == PERF_RECORD_MISC_USER && perf_host) { 480 al->level = '.'; 481 } else if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) { 482 al->level = 'g'; 483 mg = &machine->kmaps; 484 load_map = true; 485 } else if (cpumode == PERF_RECORD_MISC_GUEST_USER && perf_guest) { 486 al->level = 'u'; 487 } else { 488 al->level = 'H'; 489 al->map = NULL; 490 491 if ((cpumode == PERF_RECORD_MISC_GUEST_USER || 492 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) && 493 !perf_guest) 494 al->filtered |= (1 << HIST_FILTER__GUEST); 495 if ((cpumode == PERF_RECORD_MISC_USER || 496 cpumode == PERF_RECORD_MISC_KERNEL) && 497 !perf_host) 498 al->filtered |= (1 << HIST_FILTER__HOST); 499 500 return NULL; 501 } 502 503 al->map = map_groups__find(mg, al->addr); 504 if (al->map != NULL) { 505 /* 506 * Kernel maps might be changed when loading symbols so loading 507 * must be done prior to using kernel maps. 508 */ 509 if (load_map) 510 map__load(al->map); 511 al->addr = al->map->map_ip(al->map, al->addr); 512 } 513 514 return al->map; 515 } 516 517 /* 518 * For branch stacks or branch samples, the sample cpumode might not be correct 519 * because it applies only to the sample 'ip' and not necessary to 'addr' or 520 * branch stack addresses. If possible, use a fallback to deal with those cases. 521 */ 522 struct map *thread__find_map_fb(struct thread *thread, u8 cpumode, u64 addr, 523 struct addr_location *al) 524 { 525 struct map *map = thread__find_map(thread, cpumode, addr, al); 526 struct machine *machine = thread->mg->machine; 527 u8 addr_cpumode = machine__addr_cpumode(machine, cpumode, addr); 528 529 if (map || addr_cpumode == cpumode) 530 return map; 531 532 return thread__find_map(thread, addr_cpumode, addr, al); 533 } 534 535 struct symbol *thread__find_symbol(struct thread *thread, u8 cpumode, 536 u64 addr, struct addr_location *al) 537 { 538 al->sym = NULL; 539 if (thread__find_map(thread, cpumode, addr, al)) 540 al->sym = map__find_symbol(al->map, al->addr); 541 return al->sym; 542 } 543 544 struct symbol *thread__find_symbol_fb(struct thread *thread, u8 cpumode, 545 u64 addr, struct addr_location *al) 546 { 547 al->sym = NULL; 548 if (thread__find_map_fb(thread, cpumode, addr, al)) 549 al->sym = map__find_symbol(al->map, al->addr); 550 return al->sym; 551 } 552 553 /* 554 * Callers need to drop the reference to al->thread, obtained in 555 * machine__findnew_thread() 556 */ 557 int machine__resolve(struct machine *machine, struct addr_location *al, 558 struct perf_sample *sample) 559 { 560 struct thread *thread = machine__findnew_thread(machine, sample->pid, 561 sample->tid); 562 563 if (thread == NULL) 564 return -1; 565 566 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid); 567 thread__find_map(thread, sample->cpumode, sample->ip, al); 568 dump_printf(" ...... dso: %s\n", 569 al->map ? al->map->dso->long_name : 570 al->level == 'H' ? "[hypervisor]" : "<not found>"); 571 572 if (thread__is_filtered(thread)) 573 al->filtered |= (1 << HIST_FILTER__THREAD); 574 575 al->sym = NULL; 576 al->cpu = sample->cpu; 577 al->socket = -1; 578 al->srcline = NULL; 579 580 if (al->cpu >= 0) { 581 struct perf_env *env = machine->env; 582 583 if (env && env->cpu) 584 al->socket = env->cpu[al->cpu].socket_id; 585 } 586 587 if (al->map) { 588 struct dso *dso = al->map->dso; 589 590 if (symbol_conf.dso_list && 591 (!dso || !(strlist__has_entry(symbol_conf.dso_list, 592 dso->short_name) || 593 (dso->short_name != dso->long_name && 594 strlist__has_entry(symbol_conf.dso_list, 595 dso->long_name))))) { 596 al->filtered |= (1 << HIST_FILTER__DSO); 597 } 598 599 al->sym = map__find_symbol(al->map, al->addr); 600 } 601 602 if (symbol_conf.sym_list && 603 (!al->sym || !strlist__has_entry(symbol_conf.sym_list, 604 al->sym->name))) { 605 al->filtered |= (1 << HIST_FILTER__SYMBOL); 606 } 607 608 return 0; 609 } 610 611 /* 612 * The preprocess_sample method will return with reference counts for the 613 * in it, when done using (and perhaps getting ref counts if needing to 614 * keep a pointer to one of those entries) it must be paired with 615 * addr_location__put(), so that the refcounts can be decremented. 616 */ 617 void addr_location__put(struct addr_location *al) 618 { 619 thread__zput(al->thread); 620 } 621 622 bool is_bts_event(struct perf_event_attr *attr) 623 { 624 return attr->type == PERF_TYPE_HARDWARE && 625 (attr->config & PERF_COUNT_HW_BRANCH_INSTRUCTIONS) && 626 attr->sample_period == 1; 627 } 628 629 bool sample_addr_correlates_sym(struct perf_event_attr *attr) 630 { 631 if (attr->type == PERF_TYPE_SOFTWARE && 632 (attr->config == PERF_COUNT_SW_PAGE_FAULTS || 633 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MIN || 634 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)) 635 return true; 636 637 if (is_bts_event(attr)) 638 return true; 639 640 return false; 641 } 642 643 void thread__resolve(struct thread *thread, struct addr_location *al, 644 struct perf_sample *sample) 645 { 646 thread__find_map_fb(thread, sample->cpumode, sample->addr, al); 647 648 al->cpu = sample->cpu; 649 al->sym = NULL; 650 651 if (al->map) 652 al->sym = map__find_symbol(al->map, al->addr); 653 } 654