1 #include <linux/types.h> 2 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */ 3 #include <api/fs/fs.h> 4 #include "event.h" 5 #include "debug.h" 6 #include "hist.h" 7 #include "machine.h" 8 #include "sort.h" 9 #include "string.h" 10 #include "strlist.h" 11 #include "thread.h" 12 #include "thread_map.h" 13 #include "symbol/kallsyms.h" 14 #include "asm/bug.h" 15 #include "stat.h" 16 17 static const char *perf_event__names[] = { 18 [0] = "TOTAL", 19 [PERF_RECORD_MMAP] = "MMAP", 20 [PERF_RECORD_MMAP2] = "MMAP2", 21 [PERF_RECORD_LOST] = "LOST", 22 [PERF_RECORD_COMM] = "COMM", 23 [PERF_RECORD_EXIT] = "EXIT", 24 [PERF_RECORD_THROTTLE] = "THROTTLE", 25 [PERF_RECORD_UNTHROTTLE] = "UNTHROTTLE", 26 [PERF_RECORD_FORK] = "FORK", 27 [PERF_RECORD_READ] = "READ", 28 [PERF_RECORD_SAMPLE] = "SAMPLE", 29 [PERF_RECORD_AUX] = "AUX", 30 [PERF_RECORD_ITRACE_START] = "ITRACE_START", 31 [PERF_RECORD_LOST_SAMPLES] = "LOST_SAMPLES", 32 [PERF_RECORD_SWITCH] = "SWITCH", 33 [PERF_RECORD_SWITCH_CPU_WIDE] = "SWITCH_CPU_WIDE", 34 [PERF_RECORD_HEADER_ATTR] = "ATTR", 35 [PERF_RECORD_HEADER_EVENT_TYPE] = "EVENT_TYPE", 36 [PERF_RECORD_HEADER_TRACING_DATA] = "TRACING_DATA", 37 [PERF_RECORD_HEADER_BUILD_ID] = "BUILD_ID", 38 [PERF_RECORD_FINISHED_ROUND] = "FINISHED_ROUND", 39 [PERF_RECORD_ID_INDEX] = "ID_INDEX", 40 [PERF_RECORD_AUXTRACE_INFO] = "AUXTRACE_INFO", 41 [PERF_RECORD_AUXTRACE] = "AUXTRACE", 42 [PERF_RECORD_AUXTRACE_ERROR] = "AUXTRACE_ERROR", 43 [PERF_RECORD_THREAD_MAP] = "THREAD_MAP", 44 [PERF_RECORD_CPU_MAP] = "CPU_MAP", 45 [PERF_RECORD_STAT_CONFIG] = "STAT_CONFIG", 46 [PERF_RECORD_STAT] = "STAT", 47 [PERF_RECORD_STAT_ROUND] = "STAT_ROUND", 48 [PERF_RECORD_EVENT_UPDATE] = "EVENT_UPDATE", 49 [PERF_RECORD_TIME_CONV] = "TIME_CONV", 50 }; 51 52 const char *perf_event__name(unsigned int id) 53 { 54 if (id >= ARRAY_SIZE(perf_event__names)) 55 return "INVALID"; 56 if (!perf_event__names[id]) 57 return "UNKNOWN"; 58 return perf_event__names[id]; 59 } 60 61 static int perf_tool__process_synth_event(struct perf_tool *tool, 62 union perf_event *event, 63 struct machine *machine, 64 perf_event__handler_t process) 65 { 66 struct perf_sample synth_sample = { 67 .pid = -1, 68 .tid = -1, 69 .time = -1, 70 .stream_id = -1, 71 .cpu = -1, 72 .period = 1, 73 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK, 74 }; 75 76 return process(tool, event, &synth_sample, machine); 77 }; 78 79 /* 80 * Assumes that the first 4095 bytes of /proc/pid/stat contains 81 * the comm, tgid and ppid. 82 */ 83 static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len, 84 pid_t *tgid, pid_t *ppid) 85 { 86 char filename[PATH_MAX]; 87 char bf[4096]; 88 int fd; 89 size_t size = 0; 90 ssize_t n; 91 char *nl, *name, *tgids, *ppids; 92 93 *tgid = -1; 94 *ppid = -1; 95 96 snprintf(filename, sizeof(filename), "/proc/%d/status", pid); 97 98 fd = open(filename, O_RDONLY); 99 if (fd < 0) { 100 pr_debug("couldn't open %s\n", filename); 101 return -1; 102 } 103 104 n = read(fd, bf, sizeof(bf) - 1); 105 close(fd); 106 if (n <= 0) { 107 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n", 108 pid); 109 return -1; 110 } 111 bf[n] = '\0'; 112 113 name = strstr(bf, "Name:"); 114 tgids = strstr(bf, "Tgid:"); 115 ppids = strstr(bf, "PPid:"); 116 117 if (name) { 118 name += 5; /* strlen("Name:") */ 119 120 while (*name && isspace(*name)) 121 ++name; 122 123 nl = strchr(name, '\n'); 124 if (nl) 125 *nl = '\0'; 126 127 size = strlen(name); 128 if (size >= len) 129 size = len - 1; 130 memcpy(comm, name, size); 131 comm[size] = '\0'; 132 } else { 133 pr_debug("Name: string not found for pid %d\n", pid); 134 } 135 136 if (tgids) { 137 tgids += 5; /* strlen("Tgid:") */ 138 *tgid = atoi(tgids); 139 } else { 140 pr_debug("Tgid: string not found for pid %d\n", pid); 141 } 142 143 if (ppids) { 144 ppids += 5; /* strlen("PPid:") */ 145 *ppid = atoi(ppids); 146 } else { 147 pr_debug("PPid: string not found for pid %d\n", pid); 148 } 149 150 return 0; 151 } 152 153 static int perf_event__prepare_comm(union perf_event *event, pid_t pid, 154 struct machine *machine, 155 pid_t *tgid, pid_t *ppid) 156 { 157 size_t size; 158 159 *ppid = -1; 160 161 memset(&event->comm, 0, sizeof(event->comm)); 162 163 if (machine__is_host(machine)) { 164 if (perf_event__get_comm_ids(pid, event->comm.comm, 165 sizeof(event->comm.comm), 166 tgid, ppid) != 0) { 167 return -1; 168 } 169 } else { 170 *tgid = machine->pid; 171 } 172 173 if (*tgid < 0) 174 return -1; 175 176 event->comm.pid = *tgid; 177 event->comm.header.type = PERF_RECORD_COMM; 178 179 size = strlen(event->comm.comm) + 1; 180 size = PERF_ALIGN(size, sizeof(u64)); 181 memset(event->comm.comm + size, 0, machine->id_hdr_size); 182 event->comm.header.size = (sizeof(event->comm) - 183 (sizeof(event->comm.comm) - size) + 184 machine->id_hdr_size); 185 event->comm.tid = pid; 186 187 return 0; 188 } 189 190 pid_t perf_event__synthesize_comm(struct perf_tool *tool, 191 union perf_event *event, pid_t pid, 192 perf_event__handler_t process, 193 struct machine *machine) 194 { 195 pid_t tgid, ppid; 196 197 if (perf_event__prepare_comm(event, pid, machine, &tgid, &ppid) != 0) 198 return -1; 199 200 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 201 return -1; 202 203 return tgid; 204 } 205 206 static int perf_event__synthesize_fork(struct perf_tool *tool, 207 union perf_event *event, 208 pid_t pid, pid_t tgid, pid_t ppid, 209 perf_event__handler_t process, 210 struct machine *machine) 211 { 212 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size); 213 214 /* 215 * for main thread set parent to ppid from status file. For other 216 * threads set parent pid to main thread. ie., assume main thread 217 * spawns all threads in a process 218 */ 219 if (tgid == pid) { 220 event->fork.ppid = ppid; 221 event->fork.ptid = ppid; 222 } else { 223 event->fork.ppid = tgid; 224 event->fork.ptid = tgid; 225 } 226 event->fork.pid = tgid; 227 event->fork.tid = pid; 228 event->fork.header.type = PERF_RECORD_FORK; 229 230 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size); 231 232 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 233 return -1; 234 235 return 0; 236 } 237 238 int perf_event__synthesize_mmap_events(struct perf_tool *tool, 239 union perf_event *event, 240 pid_t pid, pid_t tgid, 241 perf_event__handler_t process, 242 struct machine *machine, 243 bool mmap_data, 244 unsigned int proc_map_timeout) 245 { 246 char filename[PATH_MAX]; 247 FILE *fp; 248 unsigned long long t; 249 bool truncation = false; 250 unsigned long long timeout = proc_map_timeout * 1000000ULL; 251 int rc = 0; 252 const char *hugetlbfs_mnt = hugetlbfs__mountpoint(); 253 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0; 254 255 if (machine__is_default_guest(machine)) 256 return 0; 257 258 snprintf(filename, sizeof(filename), "%s/proc/%d/maps", 259 machine->root_dir, pid); 260 261 fp = fopen(filename, "r"); 262 if (fp == NULL) { 263 /* 264 * We raced with a task exiting - just return: 265 */ 266 pr_debug("couldn't open %s\n", filename); 267 return -1; 268 } 269 270 event->header.type = PERF_RECORD_MMAP2; 271 t = rdclock(); 272 273 while (1) { 274 char bf[BUFSIZ]; 275 char prot[5]; 276 char execname[PATH_MAX]; 277 char anonstr[] = "//anon"; 278 unsigned int ino; 279 size_t size; 280 ssize_t n; 281 282 if (fgets(bf, sizeof(bf), fp) == NULL) 283 break; 284 285 if ((rdclock() - t) > timeout) { 286 pr_warning("Reading %s time out. " 287 "You may want to increase " 288 "the time limit by --proc-map-timeout\n", 289 filename); 290 truncation = true; 291 goto out; 292 } 293 294 /* ensure null termination since stack will be reused. */ 295 strcpy(execname, ""); 296 297 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */ 298 n = sscanf(bf, "%"PRIx64"-%"PRIx64" %s %"PRIx64" %x:%x %u %[^\n]\n", 299 &event->mmap2.start, &event->mmap2.len, prot, 300 &event->mmap2.pgoff, &event->mmap2.maj, 301 &event->mmap2.min, 302 &ino, execname); 303 304 /* 305 * Anon maps don't have the execname. 306 */ 307 if (n < 7) 308 continue; 309 310 event->mmap2.ino = (u64)ino; 311 312 /* 313 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c 314 */ 315 if (machine__is_host(machine)) 316 event->header.misc = PERF_RECORD_MISC_USER; 317 else 318 event->header.misc = PERF_RECORD_MISC_GUEST_USER; 319 320 /* map protection and flags bits */ 321 event->mmap2.prot = 0; 322 event->mmap2.flags = 0; 323 if (prot[0] == 'r') 324 event->mmap2.prot |= PROT_READ; 325 if (prot[1] == 'w') 326 event->mmap2.prot |= PROT_WRITE; 327 if (prot[2] == 'x') 328 event->mmap2.prot |= PROT_EXEC; 329 330 if (prot[3] == 's') 331 event->mmap2.flags |= MAP_SHARED; 332 else 333 event->mmap2.flags |= MAP_PRIVATE; 334 335 if (prot[2] != 'x') { 336 if (!mmap_data || prot[0] != 'r') 337 continue; 338 339 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA; 340 } 341 342 out: 343 if (truncation) 344 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT; 345 346 if (!strcmp(execname, "")) 347 strcpy(execname, anonstr); 348 349 if (hugetlbfs_mnt_len && 350 !strncmp(execname, hugetlbfs_mnt, hugetlbfs_mnt_len)) { 351 strcpy(execname, anonstr); 352 event->mmap2.flags |= MAP_HUGETLB; 353 } 354 355 size = strlen(execname) + 1; 356 memcpy(event->mmap2.filename, execname, size); 357 size = PERF_ALIGN(size, sizeof(u64)); 358 event->mmap2.len -= event->mmap.start; 359 event->mmap2.header.size = (sizeof(event->mmap2) - 360 (sizeof(event->mmap2.filename) - size)); 361 memset(event->mmap2.filename + size, 0, machine->id_hdr_size); 362 event->mmap2.header.size += machine->id_hdr_size; 363 event->mmap2.pid = tgid; 364 event->mmap2.tid = pid; 365 366 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 367 rc = -1; 368 break; 369 } 370 371 if (truncation) 372 break; 373 } 374 375 fclose(fp); 376 return rc; 377 } 378 379 int perf_event__synthesize_modules(struct perf_tool *tool, 380 perf_event__handler_t process, 381 struct machine *machine) 382 { 383 int rc = 0; 384 struct map *pos; 385 struct map_groups *kmaps = &machine->kmaps; 386 struct maps *maps = &kmaps->maps[MAP__FUNCTION]; 387 union perf_event *event = zalloc((sizeof(event->mmap) + 388 machine->id_hdr_size)); 389 if (event == NULL) { 390 pr_debug("Not enough memory synthesizing mmap event " 391 "for kernel modules\n"); 392 return -1; 393 } 394 395 event->header.type = PERF_RECORD_MMAP; 396 397 /* 398 * kernel uses 0 for user space maps, see kernel/perf_event.c 399 * __perf_event_mmap 400 */ 401 if (machine__is_host(machine)) 402 event->header.misc = PERF_RECORD_MISC_KERNEL; 403 else 404 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 405 406 for (pos = maps__first(maps); pos; pos = map__next(pos)) { 407 size_t size; 408 409 if (__map__is_kernel(pos)) 410 continue; 411 412 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64)); 413 event->mmap.header.type = PERF_RECORD_MMAP; 414 event->mmap.header.size = (sizeof(event->mmap) - 415 (sizeof(event->mmap.filename) - size)); 416 memset(event->mmap.filename + size, 0, machine->id_hdr_size); 417 event->mmap.header.size += machine->id_hdr_size; 418 event->mmap.start = pos->start; 419 event->mmap.len = pos->end - pos->start; 420 event->mmap.pid = machine->pid; 421 422 memcpy(event->mmap.filename, pos->dso->long_name, 423 pos->dso->long_name_len + 1); 424 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 425 rc = -1; 426 break; 427 } 428 } 429 430 free(event); 431 return rc; 432 } 433 434 static int __event__synthesize_thread(union perf_event *comm_event, 435 union perf_event *mmap_event, 436 union perf_event *fork_event, 437 pid_t pid, int full, 438 perf_event__handler_t process, 439 struct perf_tool *tool, 440 struct machine *machine, 441 bool mmap_data, 442 unsigned int proc_map_timeout) 443 { 444 char filename[PATH_MAX]; 445 DIR *tasks; 446 struct dirent *dirent; 447 pid_t tgid, ppid; 448 int rc = 0; 449 450 /* special case: only send one comm event using passed in pid */ 451 if (!full) { 452 tgid = perf_event__synthesize_comm(tool, comm_event, pid, 453 process, machine); 454 455 if (tgid == -1) 456 return -1; 457 458 return perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 459 process, machine, mmap_data, 460 proc_map_timeout); 461 } 462 463 if (machine__is_default_guest(machine)) 464 return 0; 465 466 snprintf(filename, sizeof(filename), "%s/proc/%d/task", 467 machine->root_dir, pid); 468 469 tasks = opendir(filename); 470 if (tasks == NULL) { 471 pr_debug("couldn't open %s\n", filename); 472 return 0; 473 } 474 475 while ((dirent = readdir(tasks)) != NULL) { 476 char *end; 477 pid_t _pid; 478 479 _pid = strtol(dirent->d_name, &end, 10); 480 if (*end) 481 continue; 482 483 rc = -1; 484 if (perf_event__prepare_comm(comm_event, _pid, machine, 485 &tgid, &ppid) != 0) 486 break; 487 488 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid, 489 ppid, process, machine) < 0) 490 break; 491 /* 492 * Send the prepared comm event 493 */ 494 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0) 495 break; 496 497 rc = 0; 498 if (_pid == pid) { 499 /* process the parent's maps too */ 500 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 501 process, machine, mmap_data, proc_map_timeout); 502 if (rc) 503 break; 504 } 505 } 506 507 closedir(tasks); 508 return rc; 509 } 510 511 int perf_event__synthesize_thread_map(struct perf_tool *tool, 512 struct thread_map *threads, 513 perf_event__handler_t process, 514 struct machine *machine, 515 bool mmap_data, 516 unsigned int proc_map_timeout) 517 { 518 union perf_event *comm_event, *mmap_event, *fork_event; 519 int err = -1, thread, j; 520 521 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 522 if (comm_event == NULL) 523 goto out; 524 525 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 526 if (mmap_event == NULL) 527 goto out_free_comm; 528 529 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 530 if (fork_event == NULL) 531 goto out_free_mmap; 532 533 err = 0; 534 for (thread = 0; thread < threads->nr; ++thread) { 535 if (__event__synthesize_thread(comm_event, mmap_event, 536 fork_event, 537 thread_map__pid(threads, thread), 0, 538 process, tool, machine, 539 mmap_data, proc_map_timeout)) { 540 err = -1; 541 break; 542 } 543 544 /* 545 * comm.pid is set to thread group id by 546 * perf_event__synthesize_comm 547 */ 548 if ((int) comm_event->comm.pid != thread_map__pid(threads, thread)) { 549 bool need_leader = true; 550 551 /* is thread group leader in thread_map? */ 552 for (j = 0; j < threads->nr; ++j) { 553 if ((int) comm_event->comm.pid == thread_map__pid(threads, j)) { 554 need_leader = false; 555 break; 556 } 557 } 558 559 /* if not, generate events for it */ 560 if (need_leader && 561 __event__synthesize_thread(comm_event, mmap_event, 562 fork_event, 563 comm_event->comm.pid, 0, 564 process, tool, machine, 565 mmap_data, proc_map_timeout)) { 566 err = -1; 567 break; 568 } 569 } 570 } 571 free(fork_event); 572 out_free_mmap: 573 free(mmap_event); 574 out_free_comm: 575 free(comm_event); 576 out: 577 return err; 578 } 579 580 int perf_event__synthesize_threads(struct perf_tool *tool, 581 perf_event__handler_t process, 582 struct machine *machine, 583 bool mmap_data, 584 unsigned int proc_map_timeout) 585 { 586 DIR *proc; 587 char proc_path[PATH_MAX]; 588 struct dirent *dirent; 589 union perf_event *comm_event, *mmap_event, *fork_event; 590 int err = -1; 591 592 if (machine__is_default_guest(machine)) 593 return 0; 594 595 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 596 if (comm_event == NULL) 597 goto out; 598 599 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 600 if (mmap_event == NULL) 601 goto out_free_comm; 602 603 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 604 if (fork_event == NULL) 605 goto out_free_mmap; 606 607 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir); 608 proc = opendir(proc_path); 609 610 if (proc == NULL) 611 goto out_free_fork; 612 613 while ((dirent = readdir(proc)) != NULL) { 614 char *end; 615 pid_t pid = strtol(dirent->d_name, &end, 10); 616 617 if (*end) /* only interested in proper numerical dirents */ 618 continue; 619 /* 620 * We may race with exiting thread, so don't stop just because 621 * one thread couldn't be synthesized. 622 */ 623 __event__synthesize_thread(comm_event, mmap_event, fork_event, pid, 624 1, process, tool, machine, mmap_data, 625 proc_map_timeout); 626 } 627 628 err = 0; 629 closedir(proc); 630 out_free_fork: 631 free(fork_event); 632 out_free_mmap: 633 free(mmap_event); 634 out_free_comm: 635 free(comm_event); 636 out: 637 return err; 638 } 639 640 struct process_symbol_args { 641 const char *name; 642 u64 start; 643 }; 644 645 static int find_symbol_cb(void *arg, const char *name, char type, 646 u64 start) 647 { 648 struct process_symbol_args *args = arg; 649 650 /* 651 * Must be a function or at least an alias, as in PARISC64, where "_text" is 652 * an 'A' to the same address as "_stext". 653 */ 654 if (!(symbol_type__is_a(type, MAP__FUNCTION) || 655 type == 'A') || strcmp(name, args->name)) 656 return 0; 657 658 args->start = start; 659 return 1; 660 } 661 662 u64 kallsyms__get_function_start(const char *kallsyms_filename, 663 const char *symbol_name) 664 { 665 struct process_symbol_args args = { .name = symbol_name, }; 666 667 if (kallsyms__parse(kallsyms_filename, &args, find_symbol_cb) <= 0) 668 return 0; 669 670 return args.start; 671 } 672 673 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 674 perf_event__handler_t process, 675 struct machine *machine) 676 { 677 size_t size; 678 const char *mmap_name; 679 char name_buff[PATH_MAX]; 680 struct map *map = machine__kernel_map(machine); 681 struct kmap *kmap; 682 int err; 683 union perf_event *event; 684 685 if (symbol_conf.kptr_restrict) 686 return -1; 687 if (map == NULL) 688 return -1; 689 690 /* 691 * We should get this from /sys/kernel/sections/.text, but till that is 692 * available use this, and after it is use this as a fallback for older 693 * kernels. 694 */ 695 event = zalloc((sizeof(event->mmap) + machine->id_hdr_size)); 696 if (event == NULL) { 697 pr_debug("Not enough memory synthesizing mmap event " 698 "for kernel modules\n"); 699 return -1; 700 } 701 702 mmap_name = machine__mmap_name(machine, name_buff, sizeof(name_buff)); 703 if (machine__is_host(machine)) { 704 /* 705 * kernel uses PERF_RECORD_MISC_USER for user space maps, 706 * see kernel/perf_event.c __perf_event_mmap 707 */ 708 event->header.misc = PERF_RECORD_MISC_KERNEL; 709 } else { 710 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 711 } 712 713 kmap = map__kmap(map); 714 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename), 715 "%s%s", mmap_name, kmap->ref_reloc_sym->name) + 1; 716 size = PERF_ALIGN(size, sizeof(u64)); 717 event->mmap.header.type = PERF_RECORD_MMAP; 718 event->mmap.header.size = (sizeof(event->mmap) - 719 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size); 720 event->mmap.pgoff = kmap->ref_reloc_sym->addr; 721 event->mmap.start = map->start; 722 event->mmap.len = map->end - event->mmap.start; 723 event->mmap.pid = machine->pid; 724 725 err = perf_tool__process_synth_event(tool, event, machine, process); 726 free(event); 727 728 return err; 729 } 730 731 int perf_event__synthesize_thread_map2(struct perf_tool *tool, 732 struct thread_map *threads, 733 perf_event__handler_t process, 734 struct machine *machine) 735 { 736 union perf_event *event; 737 int i, err, size; 738 739 size = sizeof(event->thread_map); 740 size += threads->nr * sizeof(event->thread_map.entries[0]); 741 742 event = zalloc(size); 743 if (!event) 744 return -ENOMEM; 745 746 event->header.type = PERF_RECORD_THREAD_MAP; 747 event->header.size = size; 748 event->thread_map.nr = threads->nr; 749 750 for (i = 0; i < threads->nr; i++) { 751 struct thread_map_event_entry *entry = &event->thread_map.entries[i]; 752 char *comm = thread_map__comm(threads, i); 753 754 if (!comm) 755 comm = (char *) ""; 756 757 entry->pid = thread_map__pid(threads, i); 758 strncpy((char *) &entry->comm, comm, sizeof(entry->comm)); 759 } 760 761 err = process(tool, event, NULL, machine); 762 763 free(event); 764 return err; 765 } 766 767 static void synthesize_cpus(struct cpu_map_entries *cpus, 768 struct cpu_map *map) 769 { 770 int i; 771 772 cpus->nr = map->nr; 773 774 for (i = 0; i < map->nr; i++) 775 cpus->cpu[i] = map->map[i]; 776 } 777 778 static void synthesize_mask(struct cpu_map_mask *mask, 779 struct cpu_map *map, int max) 780 { 781 int i; 782 783 mask->nr = BITS_TO_LONGS(max); 784 mask->long_size = sizeof(long); 785 786 for (i = 0; i < map->nr; i++) 787 set_bit(map->map[i], mask->mask); 788 } 789 790 static size_t cpus_size(struct cpu_map *map) 791 { 792 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16); 793 } 794 795 static size_t mask_size(struct cpu_map *map, int *max) 796 { 797 int i; 798 799 *max = 0; 800 801 for (i = 0; i < map->nr; i++) { 802 /* bit possition of the cpu is + 1 */ 803 int bit = map->map[i] + 1; 804 805 if (bit > *max) 806 *max = bit; 807 } 808 809 return sizeof(struct cpu_map_mask) + BITS_TO_LONGS(*max) * sizeof(long); 810 } 811 812 void *cpu_map_data__alloc(struct cpu_map *map, size_t *size, u16 *type, int *max) 813 { 814 size_t size_cpus, size_mask; 815 bool is_dummy = cpu_map__empty(map); 816 817 /* 818 * Both array and mask data have variable size based 819 * on the number of cpus and their actual values. 820 * The size of the 'struct cpu_map_data' is: 821 * 822 * array = size of 'struct cpu_map_entries' + 823 * number of cpus * sizeof(u64) 824 * 825 * mask = size of 'struct cpu_map_mask' + 826 * maximum cpu bit converted to size of longs 827 * 828 * and finaly + the size of 'struct cpu_map_data'. 829 */ 830 size_cpus = cpus_size(map); 831 size_mask = mask_size(map, max); 832 833 if (is_dummy || (size_cpus < size_mask)) { 834 *size += size_cpus; 835 *type = PERF_CPU_MAP__CPUS; 836 } else { 837 *size += size_mask; 838 *type = PERF_CPU_MAP__MASK; 839 } 840 841 *size += sizeof(struct cpu_map_data); 842 return zalloc(*size); 843 } 844 845 void cpu_map_data__synthesize(struct cpu_map_data *data, struct cpu_map *map, 846 u16 type, int max) 847 { 848 data->type = type; 849 850 switch (type) { 851 case PERF_CPU_MAP__CPUS: 852 synthesize_cpus((struct cpu_map_entries *) data->data, map); 853 break; 854 case PERF_CPU_MAP__MASK: 855 synthesize_mask((struct cpu_map_mask *) data->data, map, max); 856 default: 857 break; 858 }; 859 } 860 861 static struct cpu_map_event* cpu_map_event__new(struct cpu_map *map) 862 { 863 size_t size = sizeof(struct cpu_map_event); 864 struct cpu_map_event *event; 865 int max; 866 u16 type; 867 868 event = cpu_map_data__alloc(map, &size, &type, &max); 869 if (!event) 870 return NULL; 871 872 event->header.type = PERF_RECORD_CPU_MAP; 873 event->header.size = size; 874 event->data.type = type; 875 876 cpu_map_data__synthesize(&event->data, map, type, max); 877 return event; 878 } 879 880 int perf_event__synthesize_cpu_map(struct perf_tool *tool, 881 struct cpu_map *map, 882 perf_event__handler_t process, 883 struct machine *machine) 884 { 885 struct cpu_map_event *event; 886 int err; 887 888 event = cpu_map_event__new(map); 889 if (!event) 890 return -ENOMEM; 891 892 err = process(tool, (union perf_event *) event, NULL, machine); 893 894 free(event); 895 return err; 896 } 897 898 int perf_event__synthesize_stat_config(struct perf_tool *tool, 899 struct perf_stat_config *config, 900 perf_event__handler_t process, 901 struct machine *machine) 902 { 903 struct stat_config_event *event; 904 int size, i = 0, err; 905 906 size = sizeof(*event); 907 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0])); 908 909 event = zalloc(size); 910 if (!event) 911 return -ENOMEM; 912 913 event->header.type = PERF_RECORD_STAT_CONFIG; 914 event->header.size = size; 915 event->nr = PERF_STAT_CONFIG_TERM__MAX; 916 917 #define ADD(__term, __val) \ 918 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \ 919 event->data[i].val = __val; \ 920 i++; 921 922 ADD(AGGR_MODE, config->aggr_mode) 923 ADD(INTERVAL, config->interval) 924 ADD(SCALE, config->scale) 925 926 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX, 927 "stat config terms unbalanced\n"); 928 #undef ADD 929 930 err = process(tool, (union perf_event *) event, NULL, machine); 931 932 free(event); 933 return err; 934 } 935 936 int perf_event__synthesize_stat(struct perf_tool *tool, 937 u32 cpu, u32 thread, u64 id, 938 struct perf_counts_values *count, 939 perf_event__handler_t process, 940 struct machine *machine) 941 { 942 struct stat_event event; 943 944 event.header.type = PERF_RECORD_STAT; 945 event.header.size = sizeof(event); 946 event.header.misc = 0; 947 948 event.id = id; 949 event.cpu = cpu; 950 event.thread = thread; 951 event.val = count->val; 952 event.ena = count->ena; 953 event.run = count->run; 954 955 return process(tool, (union perf_event *) &event, NULL, machine); 956 } 957 958 int perf_event__synthesize_stat_round(struct perf_tool *tool, 959 u64 evtime, u64 type, 960 perf_event__handler_t process, 961 struct machine *machine) 962 { 963 struct stat_round_event event; 964 965 event.header.type = PERF_RECORD_STAT_ROUND; 966 event.header.size = sizeof(event); 967 event.header.misc = 0; 968 969 event.time = evtime; 970 event.type = type; 971 972 return process(tool, (union perf_event *) &event, NULL, machine); 973 } 974 975 void perf_event__read_stat_config(struct perf_stat_config *config, 976 struct stat_config_event *event) 977 { 978 unsigned i; 979 980 for (i = 0; i < event->nr; i++) { 981 982 switch (event->data[i].tag) { 983 #define CASE(__term, __val) \ 984 case PERF_STAT_CONFIG_TERM__##__term: \ 985 config->__val = event->data[i].val; \ 986 break; 987 988 CASE(AGGR_MODE, aggr_mode) 989 CASE(SCALE, scale) 990 CASE(INTERVAL, interval) 991 #undef CASE 992 default: 993 pr_warning("unknown stat config term %" PRIu64 "\n", 994 event->data[i].tag); 995 } 996 } 997 } 998 999 size_t perf_event__fprintf_comm(union perf_event *event, FILE *fp) 1000 { 1001 const char *s; 1002 1003 if (event->header.misc & PERF_RECORD_MISC_COMM_EXEC) 1004 s = " exec"; 1005 else 1006 s = ""; 1007 1008 return fprintf(fp, "%s: %s:%d/%d\n", s, event->comm.comm, event->comm.pid, event->comm.tid); 1009 } 1010 1011 int perf_event__process_comm(struct perf_tool *tool __maybe_unused, 1012 union perf_event *event, 1013 struct perf_sample *sample, 1014 struct machine *machine) 1015 { 1016 return machine__process_comm_event(machine, event, sample); 1017 } 1018 1019 int perf_event__process_lost(struct perf_tool *tool __maybe_unused, 1020 union perf_event *event, 1021 struct perf_sample *sample, 1022 struct machine *machine) 1023 { 1024 return machine__process_lost_event(machine, event, sample); 1025 } 1026 1027 int perf_event__process_aux(struct perf_tool *tool __maybe_unused, 1028 union perf_event *event, 1029 struct perf_sample *sample __maybe_unused, 1030 struct machine *machine) 1031 { 1032 return machine__process_aux_event(machine, event); 1033 } 1034 1035 int perf_event__process_itrace_start(struct perf_tool *tool __maybe_unused, 1036 union perf_event *event, 1037 struct perf_sample *sample __maybe_unused, 1038 struct machine *machine) 1039 { 1040 return machine__process_itrace_start_event(machine, event); 1041 } 1042 1043 int perf_event__process_lost_samples(struct perf_tool *tool __maybe_unused, 1044 union perf_event *event, 1045 struct perf_sample *sample, 1046 struct machine *machine) 1047 { 1048 return machine__process_lost_samples_event(machine, event, sample); 1049 } 1050 1051 int perf_event__process_switch(struct perf_tool *tool __maybe_unused, 1052 union perf_event *event, 1053 struct perf_sample *sample __maybe_unused, 1054 struct machine *machine) 1055 { 1056 return machine__process_switch_event(machine, event); 1057 } 1058 1059 size_t perf_event__fprintf_mmap(union perf_event *event, FILE *fp) 1060 { 1061 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64 "]: %c %s\n", 1062 event->mmap.pid, event->mmap.tid, event->mmap.start, 1063 event->mmap.len, event->mmap.pgoff, 1064 (event->header.misc & PERF_RECORD_MISC_MMAP_DATA) ? 'r' : 'x', 1065 event->mmap.filename); 1066 } 1067 1068 size_t perf_event__fprintf_mmap2(union perf_event *event, FILE *fp) 1069 { 1070 return fprintf(fp, " %d/%d: [%#" PRIx64 "(%#" PRIx64 ") @ %#" PRIx64 1071 " %02x:%02x %"PRIu64" %"PRIu64"]: %c%c%c%c %s\n", 1072 event->mmap2.pid, event->mmap2.tid, event->mmap2.start, 1073 event->mmap2.len, event->mmap2.pgoff, event->mmap2.maj, 1074 event->mmap2.min, event->mmap2.ino, 1075 event->mmap2.ino_generation, 1076 (event->mmap2.prot & PROT_READ) ? 'r' : '-', 1077 (event->mmap2.prot & PROT_WRITE) ? 'w' : '-', 1078 (event->mmap2.prot & PROT_EXEC) ? 'x' : '-', 1079 (event->mmap2.flags & MAP_SHARED) ? 's' : 'p', 1080 event->mmap2.filename); 1081 } 1082 1083 size_t perf_event__fprintf_thread_map(union perf_event *event, FILE *fp) 1084 { 1085 struct thread_map *threads = thread_map__new_event(&event->thread_map); 1086 size_t ret; 1087 1088 ret = fprintf(fp, " nr: "); 1089 1090 if (threads) 1091 ret += thread_map__fprintf(threads, fp); 1092 else 1093 ret += fprintf(fp, "failed to get threads from event\n"); 1094 1095 thread_map__put(threads); 1096 return ret; 1097 } 1098 1099 size_t perf_event__fprintf_cpu_map(union perf_event *event, FILE *fp) 1100 { 1101 struct cpu_map *cpus = cpu_map__new_data(&event->cpu_map.data); 1102 size_t ret; 1103 1104 ret = fprintf(fp, ": "); 1105 1106 if (cpus) 1107 ret += cpu_map__fprintf(cpus, fp); 1108 else 1109 ret += fprintf(fp, "failed to get cpumap from event\n"); 1110 1111 cpu_map__put(cpus); 1112 return ret; 1113 } 1114 1115 int perf_event__process_mmap(struct perf_tool *tool __maybe_unused, 1116 union perf_event *event, 1117 struct perf_sample *sample, 1118 struct machine *machine) 1119 { 1120 return machine__process_mmap_event(machine, event, sample); 1121 } 1122 1123 int perf_event__process_mmap2(struct perf_tool *tool __maybe_unused, 1124 union perf_event *event, 1125 struct perf_sample *sample, 1126 struct machine *machine) 1127 { 1128 return machine__process_mmap2_event(machine, event, sample); 1129 } 1130 1131 size_t perf_event__fprintf_task(union perf_event *event, FILE *fp) 1132 { 1133 return fprintf(fp, "(%d:%d):(%d:%d)\n", 1134 event->fork.pid, event->fork.tid, 1135 event->fork.ppid, event->fork.ptid); 1136 } 1137 1138 int perf_event__process_fork(struct perf_tool *tool __maybe_unused, 1139 union perf_event *event, 1140 struct perf_sample *sample, 1141 struct machine *machine) 1142 { 1143 return machine__process_fork_event(machine, event, sample); 1144 } 1145 1146 int perf_event__process_exit(struct perf_tool *tool __maybe_unused, 1147 union perf_event *event, 1148 struct perf_sample *sample, 1149 struct machine *machine) 1150 { 1151 return machine__process_exit_event(machine, event, sample); 1152 } 1153 1154 size_t perf_event__fprintf_aux(union perf_event *event, FILE *fp) 1155 { 1156 return fprintf(fp, " offset: %#"PRIx64" size: %#"PRIx64" flags: %#"PRIx64" [%s%s]\n", 1157 event->aux.aux_offset, event->aux.aux_size, 1158 event->aux.flags, 1159 event->aux.flags & PERF_AUX_FLAG_TRUNCATED ? "T" : "", 1160 event->aux.flags & PERF_AUX_FLAG_OVERWRITE ? "O" : ""); 1161 } 1162 1163 size_t perf_event__fprintf_itrace_start(union perf_event *event, FILE *fp) 1164 { 1165 return fprintf(fp, " pid: %u tid: %u\n", 1166 event->itrace_start.pid, event->itrace_start.tid); 1167 } 1168 1169 size_t perf_event__fprintf_switch(union perf_event *event, FILE *fp) 1170 { 1171 bool out = event->header.misc & PERF_RECORD_MISC_SWITCH_OUT; 1172 const char *in_out = out ? "OUT" : "IN "; 1173 1174 if (event->header.type == PERF_RECORD_SWITCH) 1175 return fprintf(fp, " %s\n", in_out); 1176 1177 return fprintf(fp, " %s %s pid/tid: %5u/%-5u\n", 1178 in_out, out ? "next" : "prev", 1179 event->context_switch.next_prev_pid, 1180 event->context_switch.next_prev_tid); 1181 } 1182 1183 size_t perf_event__fprintf(union perf_event *event, FILE *fp) 1184 { 1185 size_t ret = fprintf(fp, "PERF_RECORD_%s", 1186 perf_event__name(event->header.type)); 1187 1188 switch (event->header.type) { 1189 case PERF_RECORD_COMM: 1190 ret += perf_event__fprintf_comm(event, fp); 1191 break; 1192 case PERF_RECORD_FORK: 1193 case PERF_RECORD_EXIT: 1194 ret += perf_event__fprintf_task(event, fp); 1195 break; 1196 case PERF_RECORD_MMAP: 1197 ret += perf_event__fprintf_mmap(event, fp); 1198 break; 1199 case PERF_RECORD_MMAP2: 1200 ret += perf_event__fprintf_mmap2(event, fp); 1201 break; 1202 case PERF_RECORD_AUX: 1203 ret += perf_event__fprintf_aux(event, fp); 1204 break; 1205 case PERF_RECORD_ITRACE_START: 1206 ret += perf_event__fprintf_itrace_start(event, fp); 1207 break; 1208 case PERF_RECORD_SWITCH: 1209 case PERF_RECORD_SWITCH_CPU_WIDE: 1210 ret += perf_event__fprintf_switch(event, fp); 1211 break; 1212 default: 1213 ret += fprintf(fp, "\n"); 1214 } 1215 1216 return ret; 1217 } 1218 1219 int perf_event__process(struct perf_tool *tool __maybe_unused, 1220 union perf_event *event, 1221 struct perf_sample *sample, 1222 struct machine *machine) 1223 { 1224 return machine__process_event(machine, event, sample); 1225 } 1226 1227 void thread__find_addr_map(struct thread *thread, u8 cpumode, 1228 enum map_type type, u64 addr, 1229 struct addr_location *al) 1230 { 1231 struct map_groups *mg = thread->mg; 1232 struct machine *machine = mg->machine; 1233 bool load_map = false; 1234 1235 al->machine = machine; 1236 al->thread = thread; 1237 al->addr = addr; 1238 al->cpumode = cpumode; 1239 al->filtered = 0; 1240 1241 if (machine == NULL) { 1242 al->map = NULL; 1243 return; 1244 } 1245 1246 if (cpumode == PERF_RECORD_MISC_KERNEL && perf_host) { 1247 al->level = 'k'; 1248 mg = &machine->kmaps; 1249 load_map = true; 1250 } else if (cpumode == PERF_RECORD_MISC_USER && perf_host) { 1251 al->level = '.'; 1252 } else if (cpumode == PERF_RECORD_MISC_GUEST_KERNEL && perf_guest) { 1253 al->level = 'g'; 1254 mg = &machine->kmaps; 1255 load_map = true; 1256 } else if (cpumode == PERF_RECORD_MISC_GUEST_USER && perf_guest) { 1257 al->level = 'u'; 1258 } else { 1259 al->level = 'H'; 1260 al->map = NULL; 1261 1262 if ((cpumode == PERF_RECORD_MISC_GUEST_USER || 1263 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) && 1264 !perf_guest) 1265 al->filtered |= (1 << HIST_FILTER__GUEST); 1266 if ((cpumode == PERF_RECORD_MISC_USER || 1267 cpumode == PERF_RECORD_MISC_KERNEL) && 1268 !perf_host) 1269 al->filtered |= (1 << HIST_FILTER__HOST); 1270 1271 return; 1272 } 1273 try_again: 1274 al->map = map_groups__find(mg, type, al->addr); 1275 if (al->map == NULL) { 1276 /* 1277 * If this is outside of all known maps, and is a negative 1278 * address, try to look it up in the kernel dso, as it might be 1279 * a vsyscall or vdso (which executes in user-mode). 1280 * 1281 * XXX This is nasty, we should have a symbol list in the 1282 * "[vdso]" dso, but for now lets use the old trick of looking 1283 * in the whole kernel symbol list. 1284 */ 1285 if (cpumode == PERF_RECORD_MISC_USER && machine && 1286 mg != &machine->kmaps && 1287 machine__kernel_ip(machine, al->addr)) { 1288 mg = &machine->kmaps; 1289 load_map = true; 1290 goto try_again; 1291 } 1292 } else { 1293 /* 1294 * Kernel maps might be changed when loading symbols so loading 1295 * must be done prior to using kernel maps. 1296 */ 1297 if (load_map) 1298 map__load(al->map); 1299 al->addr = al->map->map_ip(al->map, al->addr); 1300 } 1301 } 1302 1303 void thread__find_addr_location(struct thread *thread, 1304 u8 cpumode, enum map_type type, u64 addr, 1305 struct addr_location *al) 1306 { 1307 thread__find_addr_map(thread, cpumode, type, addr, al); 1308 if (al->map != NULL) 1309 al->sym = map__find_symbol(al->map, al->addr); 1310 else 1311 al->sym = NULL; 1312 } 1313 1314 /* 1315 * Callers need to drop the reference to al->thread, obtained in 1316 * machine__findnew_thread() 1317 */ 1318 int machine__resolve(struct machine *machine, struct addr_location *al, 1319 struct perf_sample *sample) 1320 { 1321 struct thread *thread = machine__findnew_thread(machine, sample->pid, 1322 sample->tid); 1323 1324 if (thread == NULL) 1325 return -1; 1326 1327 dump_printf(" ... thread: %s:%d\n", thread__comm_str(thread), thread->tid); 1328 /* 1329 * Have we already created the kernel maps for this machine? 1330 * 1331 * This should have happened earlier, when we processed the kernel MMAP 1332 * events, but for older perf.data files there was no such thing, so do 1333 * it now. 1334 */ 1335 if (sample->cpumode == PERF_RECORD_MISC_KERNEL && 1336 machine__kernel_map(machine) == NULL) 1337 machine__create_kernel_maps(machine); 1338 1339 thread__find_addr_map(thread, sample->cpumode, MAP__FUNCTION, sample->ip, al); 1340 dump_printf(" ...... dso: %s\n", 1341 al->map ? al->map->dso->long_name : 1342 al->level == 'H' ? "[hypervisor]" : "<not found>"); 1343 1344 if (thread__is_filtered(thread)) 1345 al->filtered |= (1 << HIST_FILTER__THREAD); 1346 1347 al->sym = NULL; 1348 al->cpu = sample->cpu; 1349 al->socket = -1; 1350 1351 if (al->cpu >= 0) { 1352 struct perf_env *env = machine->env; 1353 1354 if (env && env->cpu) 1355 al->socket = env->cpu[al->cpu].socket_id; 1356 } 1357 1358 if (al->map) { 1359 struct dso *dso = al->map->dso; 1360 1361 if (symbol_conf.dso_list && 1362 (!dso || !(strlist__has_entry(symbol_conf.dso_list, 1363 dso->short_name) || 1364 (dso->short_name != dso->long_name && 1365 strlist__has_entry(symbol_conf.dso_list, 1366 dso->long_name))))) { 1367 al->filtered |= (1 << HIST_FILTER__DSO); 1368 } 1369 1370 al->sym = map__find_symbol(al->map, al->addr); 1371 } 1372 1373 if (symbol_conf.sym_list && 1374 (!al->sym || !strlist__has_entry(symbol_conf.sym_list, 1375 al->sym->name))) { 1376 al->filtered |= (1 << HIST_FILTER__SYMBOL); 1377 } 1378 1379 return 0; 1380 } 1381 1382 /* 1383 * The preprocess_sample method will return with reference counts for the 1384 * in it, when done using (and perhaps getting ref counts if needing to 1385 * keep a pointer to one of those entries) it must be paired with 1386 * addr_location__put(), so that the refcounts can be decremented. 1387 */ 1388 void addr_location__put(struct addr_location *al) 1389 { 1390 thread__zput(al->thread); 1391 } 1392 1393 bool is_bts_event(struct perf_event_attr *attr) 1394 { 1395 return attr->type == PERF_TYPE_HARDWARE && 1396 (attr->config & PERF_COUNT_HW_BRANCH_INSTRUCTIONS) && 1397 attr->sample_period == 1; 1398 } 1399 1400 bool sample_addr_correlates_sym(struct perf_event_attr *attr) 1401 { 1402 if (attr->type == PERF_TYPE_SOFTWARE && 1403 (attr->config == PERF_COUNT_SW_PAGE_FAULTS || 1404 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MIN || 1405 attr->config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)) 1406 return true; 1407 1408 if (is_bts_event(attr)) 1409 return true; 1410 1411 return false; 1412 } 1413 1414 void thread__resolve(struct thread *thread, struct addr_location *al, 1415 struct perf_sample *sample) 1416 { 1417 thread__find_addr_map(thread, sample->cpumode, MAP__FUNCTION, sample->addr, al); 1418 if (!al->map) 1419 thread__find_addr_map(thread, sample->cpumode, MAP__VARIABLE, 1420 sample->addr, al); 1421 1422 al->cpu = sample->cpu; 1423 al->sym = NULL; 1424 1425 if (al->map) 1426 al->sym = map__find_symbol(al->map, al->addr); 1427 } 1428