1 // SPDX-License-Identifier: GPL-2.0-only 2 3 #include "util/debug.h" 4 #include "util/dso.h" 5 #include "util/event.h" 6 #include "util/evlist.h" 7 #include "util/machine.h" 8 #include "util/map.h" 9 #include "util/map_symbol.h" 10 #include "util/branch.h" 11 #include "util/memswap.h" 12 #include "util/namespaces.h" 13 #include "util/session.h" 14 #include "util/stat.h" 15 #include "util/symbol.h" 16 #include "util/synthetic-events.h" 17 #include "util/target.h" 18 #include "util/time-utils.h" 19 #include <linux/bitops.h> 20 #include <linux/kernel.h> 21 #include <linux/string.h> 22 #include <linux/zalloc.h> 23 #include <linux/perf_event.h> 24 #include <asm/bug.h> 25 #include <perf/evsel.h> 26 #include <internal/cpumap.h> 27 #include <perf/cpumap.h> 28 #include <internal/lib.h> // page_size 29 #include <internal/threadmap.h> 30 #include <perf/threadmap.h> 31 #include <symbol/kallsyms.h> 32 #include <dirent.h> 33 #include <errno.h> 34 #include <inttypes.h> 35 #include <stdio.h> 36 #include <string.h> 37 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */ 38 #include <api/fs/fs.h> 39 #include <sys/types.h> 40 #include <sys/stat.h> 41 #include <fcntl.h> 42 #include <unistd.h> 43 44 #define DEFAULT_PROC_MAP_PARSE_TIMEOUT 500 45 46 unsigned int proc_map_timeout = DEFAULT_PROC_MAP_PARSE_TIMEOUT; 47 48 int perf_tool__process_synth_event(struct perf_tool *tool, 49 union perf_event *event, 50 struct machine *machine, 51 perf_event__handler_t process) 52 { 53 struct perf_sample synth_sample = { 54 .pid = -1, 55 .tid = -1, 56 .time = -1, 57 .stream_id = -1, 58 .cpu = -1, 59 .period = 1, 60 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK, 61 }; 62 63 return process(tool, event, &synth_sample, machine); 64 }; 65 66 /* 67 * Assumes that the first 4095 bytes of /proc/pid/stat contains 68 * the comm, tgid and ppid. 69 */ 70 static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len, 71 pid_t *tgid, pid_t *ppid) 72 { 73 char filename[PATH_MAX]; 74 char bf[4096]; 75 int fd; 76 size_t size = 0; 77 ssize_t n; 78 char *name, *tgids, *ppids; 79 80 *tgid = -1; 81 *ppid = -1; 82 83 snprintf(filename, sizeof(filename), "/proc/%d/status", pid); 84 85 fd = open(filename, O_RDONLY); 86 if (fd < 0) { 87 pr_debug("couldn't open %s\n", filename); 88 return -1; 89 } 90 91 n = read(fd, bf, sizeof(bf) - 1); 92 close(fd); 93 if (n <= 0) { 94 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n", 95 pid); 96 return -1; 97 } 98 bf[n] = '\0'; 99 100 name = strstr(bf, "Name:"); 101 tgids = strstr(bf, "Tgid:"); 102 ppids = strstr(bf, "PPid:"); 103 104 if (name) { 105 char *nl; 106 107 name = skip_spaces(name + 5); /* strlen("Name:") */ 108 nl = strchr(name, '\n'); 109 if (nl) 110 *nl = '\0'; 111 112 size = strlen(name); 113 if (size >= len) 114 size = len - 1; 115 memcpy(comm, name, size); 116 comm[size] = '\0'; 117 } else { 118 pr_debug("Name: string not found for pid %d\n", pid); 119 } 120 121 if (tgids) { 122 tgids += 5; /* strlen("Tgid:") */ 123 *tgid = atoi(tgids); 124 } else { 125 pr_debug("Tgid: string not found for pid %d\n", pid); 126 } 127 128 if (ppids) { 129 ppids += 5; /* strlen("PPid:") */ 130 *ppid = atoi(ppids); 131 } else { 132 pr_debug("PPid: string not found for pid %d\n", pid); 133 } 134 135 return 0; 136 } 137 138 static int perf_event__prepare_comm(union perf_event *event, pid_t pid, 139 struct machine *machine, 140 pid_t *tgid, pid_t *ppid) 141 { 142 size_t size; 143 144 *ppid = -1; 145 146 memset(&event->comm, 0, sizeof(event->comm)); 147 148 if (machine__is_host(machine)) { 149 if (perf_event__get_comm_ids(pid, event->comm.comm, 150 sizeof(event->comm.comm), 151 tgid, ppid) != 0) { 152 return -1; 153 } 154 } else { 155 *tgid = machine->pid; 156 } 157 158 if (*tgid < 0) 159 return -1; 160 161 event->comm.pid = *tgid; 162 event->comm.header.type = PERF_RECORD_COMM; 163 164 size = strlen(event->comm.comm) + 1; 165 size = PERF_ALIGN(size, sizeof(u64)); 166 memset(event->comm.comm + size, 0, machine->id_hdr_size); 167 event->comm.header.size = (sizeof(event->comm) - 168 (sizeof(event->comm.comm) - size) + 169 machine->id_hdr_size); 170 event->comm.tid = pid; 171 172 return 0; 173 } 174 175 pid_t perf_event__synthesize_comm(struct perf_tool *tool, 176 union perf_event *event, pid_t pid, 177 perf_event__handler_t process, 178 struct machine *machine) 179 { 180 pid_t tgid, ppid; 181 182 if (perf_event__prepare_comm(event, pid, machine, &tgid, &ppid) != 0) 183 return -1; 184 185 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 186 return -1; 187 188 return tgid; 189 } 190 191 static void perf_event__get_ns_link_info(pid_t pid, const char *ns, 192 struct perf_ns_link_info *ns_link_info) 193 { 194 struct stat64 st; 195 char proc_ns[128]; 196 197 sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns); 198 if (stat64(proc_ns, &st) == 0) { 199 ns_link_info->dev = st.st_dev; 200 ns_link_info->ino = st.st_ino; 201 } 202 } 203 204 int perf_event__synthesize_namespaces(struct perf_tool *tool, 205 union perf_event *event, 206 pid_t pid, pid_t tgid, 207 perf_event__handler_t process, 208 struct machine *machine) 209 { 210 u32 idx; 211 struct perf_ns_link_info *ns_link_info; 212 213 if (!tool || !tool->namespace_events) 214 return 0; 215 216 memset(&event->namespaces, 0, (sizeof(event->namespaces) + 217 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 218 machine->id_hdr_size)); 219 220 event->namespaces.pid = tgid; 221 event->namespaces.tid = pid; 222 223 event->namespaces.nr_namespaces = NR_NAMESPACES; 224 225 ns_link_info = event->namespaces.link_info; 226 227 for (idx = 0; idx < event->namespaces.nr_namespaces; idx++) 228 perf_event__get_ns_link_info(pid, perf_ns__name(idx), 229 &ns_link_info[idx]); 230 231 event->namespaces.header.type = PERF_RECORD_NAMESPACES; 232 233 event->namespaces.header.size = (sizeof(event->namespaces) + 234 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 235 machine->id_hdr_size); 236 237 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 238 return -1; 239 240 return 0; 241 } 242 243 static int perf_event__synthesize_fork(struct perf_tool *tool, 244 union perf_event *event, 245 pid_t pid, pid_t tgid, pid_t ppid, 246 perf_event__handler_t process, 247 struct machine *machine) 248 { 249 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size); 250 251 /* 252 * for main thread set parent to ppid from status file. For other 253 * threads set parent pid to main thread. ie., assume main thread 254 * spawns all threads in a process 255 */ 256 if (tgid == pid) { 257 event->fork.ppid = ppid; 258 event->fork.ptid = ppid; 259 } else { 260 event->fork.ppid = tgid; 261 event->fork.ptid = tgid; 262 } 263 event->fork.pid = tgid; 264 event->fork.tid = pid; 265 event->fork.header.type = PERF_RECORD_FORK; 266 event->fork.header.misc = PERF_RECORD_MISC_FORK_EXEC; 267 268 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size); 269 270 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 271 return -1; 272 273 return 0; 274 } 275 276 int perf_event__synthesize_mmap_events(struct perf_tool *tool, 277 union perf_event *event, 278 pid_t pid, pid_t tgid, 279 perf_event__handler_t process, 280 struct machine *machine, 281 bool mmap_data) 282 { 283 char filename[PATH_MAX]; 284 FILE *fp; 285 unsigned long long t; 286 bool truncation = false; 287 unsigned long long timeout = proc_map_timeout * 1000000ULL; 288 int rc = 0; 289 const char *hugetlbfs_mnt = hugetlbfs__mountpoint(); 290 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0; 291 292 if (machine__is_default_guest(machine)) 293 return 0; 294 295 snprintf(filename, sizeof(filename), "%s/proc/%d/task/%d/maps", 296 machine->root_dir, pid, pid); 297 298 fp = fopen(filename, "r"); 299 if (fp == NULL) { 300 /* 301 * We raced with a task exiting - just return: 302 */ 303 pr_debug("couldn't open %s\n", filename); 304 return -1; 305 } 306 307 event->header.type = PERF_RECORD_MMAP2; 308 t = rdclock(); 309 310 while (1) { 311 char bf[BUFSIZ]; 312 char prot[5]; 313 char execname[PATH_MAX]; 314 char anonstr[] = "//anon"; 315 unsigned int ino; 316 size_t size; 317 ssize_t n; 318 319 if (fgets(bf, sizeof(bf), fp) == NULL) 320 break; 321 322 if ((rdclock() - t) > timeout) { 323 pr_warning("Reading %s time out. " 324 "You may want to increase " 325 "the time limit by --proc-map-timeout\n", 326 filename); 327 truncation = true; 328 goto out; 329 } 330 331 /* ensure null termination since stack will be reused. */ 332 strcpy(execname, ""); 333 334 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */ 335 n = sscanf(bf, "%"PRI_lx64"-%"PRI_lx64" %s %"PRI_lx64" %x:%x %u %[^\n]\n", 336 &event->mmap2.start, &event->mmap2.len, prot, 337 &event->mmap2.pgoff, &event->mmap2.maj, 338 &event->mmap2.min, 339 &ino, execname); 340 341 /* 342 * Anon maps don't have the execname. 343 */ 344 if (n < 7) 345 continue; 346 347 event->mmap2.ino = (u64)ino; 348 event->mmap2.ino_generation = 0; 349 350 /* 351 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c 352 */ 353 if (machine__is_host(machine)) 354 event->header.misc = PERF_RECORD_MISC_USER; 355 else 356 event->header.misc = PERF_RECORD_MISC_GUEST_USER; 357 358 /* map protection and flags bits */ 359 event->mmap2.prot = 0; 360 event->mmap2.flags = 0; 361 if (prot[0] == 'r') 362 event->mmap2.prot |= PROT_READ; 363 if (prot[1] == 'w') 364 event->mmap2.prot |= PROT_WRITE; 365 if (prot[2] == 'x') 366 event->mmap2.prot |= PROT_EXEC; 367 368 if (prot[3] == 's') 369 event->mmap2.flags |= MAP_SHARED; 370 else 371 event->mmap2.flags |= MAP_PRIVATE; 372 373 if (prot[2] != 'x') { 374 if (!mmap_data || prot[0] != 'r') 375 continue; 376 377 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA; 378 } 379 380 out: 381 if (truncation) 382 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT; 383 384 if (!strcmp(execname, "")) 385 strcpy(execname, anonstr); 386 387 if (hugetlbfs_mnt_len && 388 !strncmp(execname, hugetlbfs_mnt, hugetlbfs_mnt_len)) { 389 strcpy(execname, anonstr); 390 event->mmap2.flags |= MAP_HUGETLB; 391 } 392 393 size = strlen(execname) + 1; 394 memcpy(event->mmap2.filename, execname, size); 395 size = PERF_ALIGN(size, sizeof(u64)); 396 event->mmap2.len -= event->mmap.start; 397 event->mmap2.header.size = (sizeof(event->mmap2) - 398 (sizeof(event->mmap2.filename) - size)); 399 memset(event->mmap2.filename + size, 0, machine->id_hdr_size); 400 event->mmap2.header.size += machine->id_hdr_size; 401 event->mmap2.pid = tgid; 402 event->mmap2.tid = pid; 403 404 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 405 rc = -1; 406 break; 407 } 408 409 if (truncation) 410 break; 411 } 412 413 fclose(fp); 414 return rc; 415 } 416 417 int perf_event__synthesize_modules(struct perf_tool *tool, perf_event__handler_t process, 418 struct machine *machine) 419 { 420 int rc = 0; 421 struct map *pos; 422 struct maps *maps = machine__kernel_maps(machine); 423 union perf_event *event = zalloc((sizeof(event->mmap) + 424 machine->id_hdr_size)); 425 if (event == NULL) { 426 pr_debug("Not enough memory synthesizing mmap event " 427 "for kernel modules\n"); 428 return -1; 429 } 430 431 event->header.type = PERF_RECORD_MMAP; 432 433 /* 434 * kernel uses 0 for user space maps, see kernel/perf_event.c 435 * __perf_event_mmap 436 */ 437 if (machine__is_host(machine)) 438 event->header.misc = PERF_RECORD_MISC_KERNEL; 439 else 440 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 441 442 maps__for_each_entry(maps, pos) { 443 size_t size; 444 445 if (!__map__is_kmodule(pos)) 446 continue; 447 448 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64)); 449 event->mmap.header.type = PERF_RECORD_MMAP; 450 event->mmap.header.size = (sizeof(event->mmap) - 451 (sizeof(event->mmap.filename) - size)); 452 memset(event->mmap.filename + size, 0, machine->id_hdr_size); 453 event->mmap.header.size += machine->id_hdr_size; 454 event->mmap.start = pos->start; 455 event->mmap.len = pos->end - pos->start; 456 event->mmap.pid = machine->pid; 457 458 memcpy(event->mmap.filename, pos->dso->long_name, 459 pos->dso->long_name_len + 1); 460 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 461 rc = -1; 462 break; 463 } 464 } 465 466 free(event); 467 return rc; 468 } 469 470 static int __event__synthesize_thread(union perf_event *comm_event, 471 union perf_event *mmap_event, 472 union perf_event *fork_event, 473 union perf_event *namespaces_event, 474 pid_t pid, int full, perf_event__handler_t process, 475 struct perf_tool *tool, struct machine *machine, bool mmap_data) 476 { 477 char filename[PATH_MAX]; 478 DIR *tasks; 479 struct dirent *dirent; 480 pid_t tgid, ppid; 481 int rc = 0; 482 483 /* special case: only send one comm event using passed in pid */ 484 if (!full) { 485 tgid = perf_event__synthesize_comm(tool, comm_event, pid, 486 process, machine); 487 488 if (tgid == -1) 489 return -1; 490 491 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid, 492 tgid, process, machine) < 0) 493 return -1; 494 495 /* 496 * send mmap only for thread group leader 497 * see thread__init_maps() 498 */ 499 if (pid == tgid && 500 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 501 process, machine, mmap_data)) 502 return -1; 503 504 return 0; 505 } 506 507 if (machine__is_default_guest(machine)) 508 return 0; 509 510 snprintf(filename, sizeof(filename), "%s/proc/%d/task", 511 machine->root_dir, pid); 512 513 tasks = opendir(filename); 514 if (tasks == NULL) { 515 pr_debug("couldn't open %s\n", filename); 516 return 0; 517 } 518 519 while ((dirent = readdir(tasks)) != NULL) { 520 char *end; 521 pid_t _pid; 522 523 _pid = strtol(dirent->d_name, &end, 10); 524 if (*end) 525 continue; 526 527 rc = -1; 528 if (perf_event__prepare_comm(comm_event, _pid, machine, 529 &tgid, &ppid) != 0) 530 break; 531 532 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid, 533 ppid, process, machine) < 0) 534 break; 535 536 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid, 537 tgid, process, machine) < 0) 538 break; 539 540 /* 541 * Send the prepared comm event 542 */ 543 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0) 544 break; 545 546 rc = 0; 547 if (_pid == pid) { 548 /* process the parent's maps too */ 549 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 550 process, machine, mmap_data); 551 if (rc) 552 break; 553 } 554 } 555 556 closedir(tasks); 557 return rc; 558 } 559 560 int perf_event__synthesize_thread_map(struct perf_tool *tool, 561 struct perf_thread_map *threads, 562 perf_event__handler_t process, 563 struct machine *machine, 564 bool mmap_data) 565 { 566 union perf_event *comm_event, *mmap_event, *fork_event; 567 union perf_event *namespaces_event; 568 int err = -1, thread, j; 569 570 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 571 if (comm_event == NULL) 572 goto out; 573 574 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 575 if (mmap_event == NULL) 576 goto out_free_comm; 577 578 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 579 if (fork_event == NULL) 580 goto out_free_mmap; 581 582 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 583 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 584 machine->id_hdr_size); 585 if (namespaces_event == NULL) 586 goto out_free_fork; 587 588 err = 0; 589 for (thread = 0; thread < threads->nr; ++thread) { 590 if (__event__synthesize_thread(comm_event, mmap_event, 591 fork_event, namespaces_event, 592 perf_thread_map__pid(threads, thread), 0, 593 process, tool, machine, 594 mmap_data)) { 595 err = -1; 596 break; 597 } 598 599 /* 600 * comm.pid is set to thread group id by 601 * perf_event__synthesize_comm 602 */ 603 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) { 604 bool need_leader = true; 605 606 /* is thread group leader in thread_map? */ 607 for (j = 0; j < threads->nr; ++j) { 608 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) { 609 need_leader = false; 610 break; 611 } 612 } 613 614 /* if not, generate events for it */ 615 if (need_leader && 616 __event__synthesize_thread(comm_event, mmap_event, 617 fork_event, namespaces_event, 618 comm_event->comm.pid, 0, 619 process, tool, machine, 620 mmap_data)) { 621 err = -1; 622 break; 623 } 624 } 625 } 626 free(namespaces_event); 627 out_free_fork: 628 free(fork_event); 629 out_free_mmap: 630 free(mmap_event); 631 out_free_comm: 632 free(comm_event); 633 out: 634 return err; 635 } 636 637 static int __perf_event__synthesize_threads(struct perf_tool *tool, 638 perf_event__handler_t process, 639 struct machine *machine, 640 bool mmap_data, 641 struct dirent **dirent, 642 int start, 643 int num) 644 { 645 union perf_event *comm_event, *mmap_event, *fork_event; 646 union perf_event *namespaces_event; 647 int err = -1; 648 char *end; 649 pid_t pid; 650 int i; 651 652 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 653 if (comm_event == NULL) 654 goto out; 655 656 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 657 if (mmap_event == NULL) 658 goto out_free_comm; 659 660 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 661 if (fork_event == NULL) 662 goto out_free_mmap; 663 664 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 665 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 666 machine->id_hdr_size); 667 if (namespaces_event == NULL) 668 goto out_free_fork; 669 670 for (i = start; i < start + num; i++) { 671 if (!isdigit(dirent[i]->d_name[0])) 672 continue; 673 674 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10); 675 /* only interested in proper numerical dirents */ 676 if (*end) 677 continue; 678 /* 679 * We may race with exiting thread, so don't stop just because 680 * one thread couldn't be synthesized. 681 */ 682 __event__synthesize_thread(comm_event, mmap_event, fork_event, 683 namespaces_event, pid, 1, process, 684 tool, machine, mmap_data); 685 } 686 err = 0; 687 688 free(namespaces_event); 689 out_free_fork: 690 free(fork_event); 691 out_free_mmap: 692 free(mmap_event); 693 out_free_comm: 694 free(comm_event); 695 out: 696 return err; 697 } 698 699 struct synthesize_threads_arg { 700 struct perf_tool *tool; 701 perf_event__handler_t process; 702 struct machine *machine; 703 bool mmap_data; 704 struct dirent **dirent; 705 int num; 706 int start; 707 }; 708 709 static void *synthesize_threads_worker(void *arg) 710 { 711 struct synthesize_threads_arg *args = arg; 712 713 __perf_event__synthesize_threads(args->tool, args->process, 714 args->machine, args->mmap_data, 715 args->dirent, 716 args->start, args->num); 717 return NULL; 718 } 719 720 int perf_event__synthesize_threads(struct perf_tool *tool, 721 perf_event__handler_t process, 722 struct machine *machine, 723 bool mmap_data, 724 unsigned int nr_threads_synthesize) 725 { 726 struct synthesize_threads_arg *args = NULL; 727 pthread_t *synthesize_threads = NULL; 728 char proc_path[PATH_MAX]; 729 struct dirent **dirent; 730 int num_per_thread; 731 int m, n, i, j; 732 int thread_nr; 733 int base = 0; 734 int err = -1; 735 736 737 if (machine__is_default_guest(machine)) 738 return 0; 739 740 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir); 741 n = scandir(proc_path, &dirent, 0, alphasort); 742 if (n < 0) 743 return err; 744 745 if (nr_threads_synthesize == UINT_MAX) 746 thread_nr = sysconf(_SC_NPROCESSORS_ONLN); 747 else 748 thread_nr = nr_threads_synthesize; 749 750 if (thread_nr <= 1) { 751 err = __perf_event__synthesize_threads(tool, process, 752 machine, mmap_data, 753 dirent, base, n); 754 goto free_dirent; 755 } 756 if (thread_nr > n) 757 thread_nr = n; 758 759 synthesize_threads = calloc(sizeof(pthread_t), thread_nr); 760 if (synthesize_threads == NULL) 761 goto free_dirent; 762 763 args = calloc(sizeof(*args), thread_nr); 764 if (args == NULL) 765 goto free_threads; 766 767 num_per_thread = n / thread_nr; 768 m = n % thread_nr; 769 for (i = 0; i < thread_nr; i++) { 770 args[i].tool = tool; 771 args[i].process = process; 772 args[i].machine = machine; 773 args[i].mmap_data = mmap_data; 774 args[i].dirent = dirent; 775 } 776 for (i = 0; i < m; i++) { 777 args[i].num = num_per_thread + 1; 778 args[i].start = i * args[i].num; 779 } 780 if (i != 0) 781 base = args[i-1].start + args[i-1].num; 782 for (j = i; j < thread_nr; j++) { 783 args[j].num = num_per_thread; 784 args[j].start = base + (j - i) * args[i].num; 785 } 786 787 for (i = 0; i < thread_nr; i++) { 788 if (pthread_create(&synthesize_threads[i], NULL, 789 synthesize_threads_worker, &args[i])) 790 goto out_join; 791 } 792 err = 0; 793 out_join: 794 for (i = 0; i < thread_nr; i++) 795 pthread_join(synthesize_threads[i], NULL); 796 free(args); 797 free_threads: 798 free(synthesize_threads); 799 free_dirent: 800 for (i = 0; i < n; i++) 801 zfree(&dirent[i]); 802 free(dirent); 803 804 return err; 805 } 806 807 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused, 808 perf_event__handler_t process __maybe_unused, 809 struct machine *machine __maybe_unused) 810 { 811 return 0; 812 } 813 814 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 815 perf_event__handler_t process, 816 struct machine *machine) 817 { 818 size_t size; 819 struct map *map = machine__kernel_map(machine); 820 struct kmap *kmap; 821 int err; 822 union perf_event *event; 823 824 if (map == NULL) 825 return -1; 826 827 kmap = map__kmap(map); 828 if (!kmap->ref_reloc_sym) 829 return -1; 830 831 /* 832 * We should get this from /sys/kernel/sections/.text, but till that is 833 * available use this, and after it is use this as a fallback for older 834 * kernels. 835 */ 836 event = zalloc((sizeof(event->mmap) + machine->id_hdr_size)); 837 if (event == NULL) { 838 pr_debug("Not enough memory synthesizing mmap event " 839 "for kernel modules\n"); 840 return -1; 841 } 842 843 if (machine__is_host(machine)) { 844 /* 845 * kernel uses PERF_RECORD_MISC_USER for user space maps, 846 * see kernel/perf_event.c __perf_event_mmap 847 */ 848 event->header.misc = PERF_RECORD_MISC_KERNEL; 849 } else { 850 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 851 } 852 853 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename), 854 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1; 855 size = PERF_ALIGN(size, sizeof(u64)); 856 event->mmap.header.type = PERF_RECORD_MMAP; 857 event->mmap.header.size = (sizeof(event->mmap) - 858 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size); 859 event->mmap.pgoff = kmap->ref_reloc_sym->addr; 860 event->mmap.start = map->start; 861 event->mmap.len = map->end - event->mmap.start; 862 event->mmap.pid = machine->pid; 863 864 err = perf_tool__process_synth_event(tool, event, machine, process); 865 free(event); 866 867 return err; 868 } 869 870 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 871 perf_event__handler_t process, 872 struct machine *machine) 873 { 874 int err; 875 876 err = __perf_event__synthesize_kernel_mmap(tool, process, machine); 877 if (err < 0) 878 return err; 879 880 return perf_event__synthesize_extra_kmaps(tool, process, machine); 881 } 882 883 int perf_event__synthesize_thread_map2(struct perf_tool *tool, 884 struct perf_thread_map *threads, 885 perf_event__handler_t process, 886 struct machine *machine) 887 { 888 union perf_event *event; 889 int i, err, size; 890 891 size = sizeof(event->thread_map); 892 size += threads->nr * sizeof(event->thread_map.entries[0]); 893 894 event = zalloc(size); 895 if (!event) 896 return -ENOMEM; 897 898 event->header.type = PERF_RECORD_THREAD_MAP; 899 event->header.size = size; 900 event->thread_map.nr = threads->nr; 901 902 for (i = 0; i < threads->nr; i++) { 903 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i]; 904 char *comm = perf_thread_map__comm(threads, i); 905 906 if (!comm) 907 comm = (char *) ""; 908 909 entry->pid = perf_thread_map__pid(threads, i); 910 strncpy((char *) &entry->comm, comm, sizeof(entry->comm)); 911 } 912 913 err = process(tool, event, NULL, machine); 914 915 free(event); 916 return err; 917 } 918 919 static void synthesize_cpus(struct cpu_map_entries *cpus, 920 struct perf_cpu_map *map) 921 { 922 int i; 923 924 cpus->nr = map->nr; 925 926 for (i = 0; i < map->nr; i++) 927 cpus->cpu[i] = map->map[i]; 928 } 929 930 static void synthesize_mask(struct perf_record_record_cpu_map *mask, 931 struct perf_cpu_map *map, int max) 932 { 933 int i; 934 935 mask->nr = BITS_TO_LONGS(max); 936 mask->long_size = sizeof(long); 937 938 for (i = 0; i < map->nr; i++) 939 set_bit(map->map[i], mask->mask); 940 } 941 942 static size_t cpus_size(struct perf_cpu_map *map) 943 { 944 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16); 945 } 946 947 static size_t mask_size(struct perf_cpu_map *map, int *max) 948 { 949 int i; 950 951 *max = 0; 952 953 for (i = 0; i < map->nr; i++) { 954 /* bit possition of the cpu is + 1 */ 955 int bit = map->map[i] + 1; 956 957 if (bit > *max) 958 *max = bit; 959 } 960 961 return sizeof(struct perf_record_record_cpu_map) + BITS_TO_LONGS(*max) * sizeof(long); 962 } 963 964 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max) 965 { 966 size_t size_cpus, size_mask; 967 bool is_dummy = perf_cpu_map__empty(map); 968 969 /* 970 * Both array and mask data have variable size based 971 * on the number of cpus and their actual values. 972 * The size of the 'struct perf_record_cpu_map_data' is: 973 * 974 * array = size of 'struct cpu_map_entries' + 975 * number of cpus * sizeof(u64) 976 * 977 * mask = size of 'struct perf_record_record_cpu_map' + 978 * maximum cpu bit converted to size of longs 979 * 980 * and finaly + the size of 'struct perf_record_cpu_map_data'. 981 */ 982 size_cpus = cpus_size(map); 983 size_mask = mask_size(map, max); 984 985 if (is_dummy || (size_cpus < size_mask)) { 986 *size += size_cpus; 987 *type = PERF_CPU_MAP__CPUS; 988 } else { 989 *size += size_mask; 990 *type = PERF_CPU_MAP__MASK; 991 } 992 993 *size += sizeof(struct perf_record_cpu_map_data); 994 *size = PERF_ALIGN(*size, sizeof(u64)); 995 return zalloc(*size); 996 } 997 998 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map, 999 u16 type, int max) 1000 { 1001 data->type = type; 1002 1003 switch (type) { 1004 case PERF_CPU_MAP__CPUS: 1005 synthesize_cpus((struct cpu_map_entries *) data->data, map); 1006 break; 1007 case PERF_CPU_MAP__MASK: 1008 synthesize_mask((struct perf_record_record_cpu_map *)data->data, map, max); 1009 default: 1010 break; 1011 }; 1012 } 1013 1014 static struct perf_record_cpu_map *cpu_map_event__new(struct perf_cpu_map *map) 1015 { 1016 size_t size = sizeof(struct perf_record_cpu_map); 1017 struct perf_record_cpu_map *event; 1018 int max; 1019 u16 type; 1020 1021 event = cpu_map_data__alloc(map, &size, &type, &max); 1022 if (!event) 1023 return NULL; 1024 1025 event->header.type = PERF_RECORD_CPU_MAP; 1026 event->header.size = size; 1027 event->data.type = type; 1028 1029 cpu_map_data__synthesize(&event->data, map, type, max); 1030 return event; 1031 } 1032 1033 int perf_event__synthesize_cpu_map(struct perf_tool *tool, 1034 struct perf_cpu_map *map, 1035 perf_event__handler_t process, 1036 struct machine *machine) 1037 { 1038 struct perf_record_cpu_map *event; 1039 int err; 1040 1041 event = cpu_map_event__new(map); 1042 if (!event) 1043 return -ENOMEM; 1044 1045 err = process(tool, (union perf_event *) event, NULL, machine); 1046 1047 free(event); 1048 return err; 1049 } 1050 1051 int perf_event__synthesize_stat_config(struct perf_tool *tool, 1052 struct perf_stat_config *config, 1053 perf_event__handler_t process, 1054 struct machine *machine) 1055 { 1056 struct perf_record_stat_config *event; 1057 int size, i = 0, err; 1058 1059 size = sizeof(*event); 1060 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0])); 1061 1062 event = zalloc(size); 1063 if (!event) 1064 return -ENOMEM; 1065 1066 event->header.type = PERF_RECORD_STAT_CONFIG; 1067 event->header.size = size; 1068 event->nr = PERF_STAT_CONFIG_TERM__MAX; 1069 1070 #define ADD(__term, __val) \ 1071 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \ 1072 event->data[i].val = __val; \ 1073 i++; 1074 1075 ADD(AGGR_MODE, config->aggr_mode) 1076 ADD(INTERVAL, config->interval) 1077 ADD(SCALE, config->scale) 1078 1079 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX, 1080 "stat config terms unbalanced\n"); 1081 #undef ADD 1082 1083 err = process(tool, (union perf_event *) event, NULL, machine); 1084 1085 free(event); 1086 return err; 1087 } 1088 1089 int perf_event__synthesize_stat(struct perf_tool *tool, 1090 u32 cpu, u32 thread, u64 id, 1091 struct perf_counts_values *count, 1092 perf_event__handler_t process, 1093 struct machine *machine) 1094 { 1095 struct perf_record_stat event; 1096 1097 event.header.type = PERF_RECORD_STAT; 1098 event.header.size = sizeof(event); 1099 event.header.misc = 0; 1100 1101 event.id = id; 1102 event.cpu = cpu; 1103 event.thread = thread; 1104 event.val = count->val; 1105 event.ena = count->ena; 1106 event.run = count->run; 1107 1108 return process(tool, (union perf_event *) &event, NULL, machine); 1109 } 1110 1111 int perf_event__synthesize_stat_round(struct perf_tool *tool, 1112 u64 evtime, u64 type, 1113 perf_event__handler_t process, 1114 struct machine *machine) 1115 { 1116 struct perf_record_stat_round event; 1117 1118 event.header.type = PERF_RECORD_STAT_ROUND; 1119 event.header.size = sizeof(event); 1120 event.header.misc = 0; 1121 1122 event.time = evtime; 1123 event.type = type; 1124 1125 return process(tool, (union perf_event *) &event, NULL, machine); 1126 } 1127 1128 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format) 1129 { 1130 size_t sz, result = sizeof(struct perf_record_sample); 1131 1132 if (type & PERF_SAMPLE_IDENTIFIER) 1133 result += sizeof(u64); 1134 1135 if (type & PERF_SAMPLE_IP) 1136 result += sizeof(u64); 1137 1138 if (type & PERF_SAMPLE_TID) 1139 result += sizeof(u64); 1140 1141 if (type & PERF_SAMPLE_TIME) 1142 result += sizeof(u64); 1143 1144 if (type & PERF_SAMPLE_ADDR) 1145 result += sizeof(u64); 1146 1147 if (type & PERF_SAMPLE_ID) 1148 result += sizeof(u64); 1149 1150 if (type & PERF_SAMPLE_STREAM_ID) 1151 result += sizeof(u64); 1152 1153 if (type & PERF_SAMPLE_CPU) 1154 result += sizeof(u64); 1155 1156 if (type & PERF_SAMPLE_PERIOD) 1157 result += sizeof(u64); 1158 1159 if (type & PERF_SAMPLE_READ) { 1160 result += sizeof(u64); 1161 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1162 result += sizeof(u64); 1163 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1164 result += sizeof(u64); 1165 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1166 if (read_format & PERF_FORMAT_GROUP) { 1167 sz = sample->read.group.nr * 1168 sizeof(struct sample_read_value); 1169 result += sz; 1170 } else { 1171 result += sizeof(u64); 1172 } 1173 } 1174 1175 if (type & PERF_SAMPLE_CALLCHAIN) { 1176 sz = (sample->callchain->nr + 1) * sizeof(u64); 1177 result += sz; 1178 } 1179 1180 if (type & PERF_SAMPLE_RAW) { 1181 result += sizeof(u32); 1182 result += sample->raw_size; 1183 } 1184 1185 if (type & PERF_SAMPLE_BRANCH_STACK) { 1186 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1187 /* nr, hw_idx */ 1188 sz += 2 * sizeof(u64); 1189 result += sz; 1190 } 1191 1192 if (type & PERF_SAMPLE_REGS_USER) { 1193 if (sample->user_regs.abi) { 1194 result += sizeof(u64); 1195 sz = hweight64(sample->user_regs.mask) * sizeof(u64); 1196 result += sz; 1197 } else { 1198 result += sizeof(u64); 1199 } 1200 } 1201 1202 if (type & PERF_SAMPLE_STACK_USER) { 1203 sz = sample->user_stack.size; 1204 result += sizeof(u64); 1205 if (sz) { 1206 result += sz; 1207 result += sizeof(u64); 1208 } 1209 } 1210 1211 if (type & PERF_SAMPLE_WEIGHT) 1212 result += sizeof(u64); 1213 1214 if (type & PERF_SAMPLE_DATA_SRC) 1215 result += sizeof(u64); 1216 1217 if (type & PERF_SAMPLE_TRANSACTION) 1218 result += sizeof(u64); 1219 1220 if (type & PERF_SAMPLE_REGS_INTR) { 1221 if (sample->intr_regs.abi) { 1222 result += sizeof(u64); 1223 sz = hweight64(sample->intr_regs.mask) * sizeof(u64); 1224 result += sz; 1225 } else { 1226 result += sizeof(u64); 1227 } 1228 } 1229 1230 if (type & PERF_SAMPLE_PHYS_ADDR) 1231 result += sizeof(u64); 1232 1233 if (type & PERF_SAMPLE_AUX) { 1234 result += sizeof(u64); 1235 result += sample->aux_sample.size; 1236 } 1237 1238 return result; 1239 } 1240 1241 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format, 1242 const struct perf_sample *sample) 1243 { 1244 __u64 *array; 1245 size_t sz; 1246 /* 1247 * used for cross-endian analysis. See git commit 65014ab3 1248 * for why this goofiness is needed. 1249 */ 1250 union u64_swap u; 1251 1252 array = event->sample.array; 1253 1254 if (type & PERF_SAMPLE_IDENTIFIER) { 1255 *array = sample->id; 1256 array++; 1257 } 1258 1259 if (type & PERF_SAMPLE_IP) { 1260 *array = sample->ip; 1261 array++; 1262 } 1263 1264 if (type & PERF_SAMPLE_TID) { 1265 u.val32[0] = sample->pid; 1266 u.val32[1] = sample->tid; 1267 *array = u.val64; 1268 array++; 1269 } 1270 1271 if (type & PERF_SAMPLE_TIME) { 1272 *array = sample->time; 1273 array++; 1274 } 1275 1276 if (type & PERF_SAMPLE_ADDR) { 1277 *array = sample->addr; 1278 array++; 1279 } 1280 1281 if (type & PERF_SAMPLE_ID) { 1282 *array = sample->id; 1283 array++; 1284 } 1285 1286 if (type & PERF_SAMPLE_STREAM_ID) { 1287 *array = sample->stream_id; 1288 array++; 1289 } 1290 1291 if (type & PERF_SAMPLE_CPU) { 1292 u.val32[0] = sample->cpu; 1293 u.val32[1] = 0; 1294 *array = u.val64; 1295 array++; 1296 } 1297 1298 if (type & PERF_SAMPLE_PERIOD) { 1299 *array = sample->period; 1300 array++; 1301 } 1302 1303 if (type & PERF_SAMPLE_READ) { 1304 if (read_format & PERF_FORMAT_GROUP) 1305 *array = sample->read.group.nr; 1306 else 1307 *array = sample->read.one.value; 1308 array++; 1309 1310 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 1311 *array = sample->read.time_enabled; 1312 array++; 1313 } 1314 1315 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 1316 *array = sample->read.time_running; 1317 array++; 1318 } 1319 1320 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1321 if (read_format & PERF_FORMAT_GROUP) { 1322 sz = sample->read.group.nr * 1323 sizeof(struct sample_read_value); 1324 memcpy(array, sample->read.group.values, sz); 1325 array = (void *)array + sz; 1326 } else { 1327 *array = sample->read.one.id; 1328 array++; 1329 } 1330 } 1331 1332 if (type & PERF_SAMPLE_CALLCHAIN) { 1333 sz = (sample->callchain->nr + 1) * sizeof(u64); 1334 memcpy(array, sample->callchain, sz); 1335 array = (void *)array + sz; 1336 } 1337 1338 if (type & PERF_SAMPLE_RAW) { 1339 u.val32[0] = sample->raw_size; 1340 *array = u.val64; 1341 array = (void *)array + sizeof(u32); 1342 1343 memcpy(array, sample->raw_data, sample->raw_size); 1344 array = (void *)array + sample->raw_size; 1345 } 1346 1347 if (type & PERF_SAMPLE_BRANCH_STACK) { 1348 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1349 /* nr, hw_idx */ 1350 sz += 2 * sizeof(u64); 1351 memcpy(array, sample->branch_stack, sz); 1352 array = (void *)array + sz; 1353 } 1354 1355 if (type & PERF_SAMPLE_REGS_USER) { 1356 if (sample->user_regs.abi) { 1357 *array++ = sample->user_regs.abi; 1358 sz = hweight64(sample->user_regs.mask) * sizeof(u64); 1359 memcpy(array, sample->user_regs.regs, sz); 1360 array = (void *)array + sz; 1361 } else { 1362 *array++ = 0; 1363 } 1364 } 1365 1366 if (type & PERF_SAMPLE_STACK_USER) { 1367 sz = sample->user_stack.size; 1368 *array++ = sz; 1369 if (sz) { 1370 memcpy(array, sample->user_stack.data, sz); 1371 array = (void *)array + sz; 1372 *array++ = sz; 1373 } 1374 } 1375 1376 if (type & PERF_SAMPLE_WEIGHT) { 1377 *array = sample->weight; 1378 array++; 1379 } 1380 1381 if (type & PERF_SAMPLE_DATA_SRC) { 1382 *array = sample->data_src; 1383 array++; 1384 } 1385 1386 if (type & PERF_SAMPLE_TRANSACTION) { 1387 *array = sample->transaction; 1388 array++; 1389 } 1390 1391 if (type & PERF_SAMPLE_REGS_INTR) { 1392 if (sample->intr_regs.abi) { 1393 *array++ = sample->intr_regs.abi; 1394 sz = hweight64(sample->intr_regs.mask) * sizeof(u64); 1395 memcpy(array, sample->intr_regs.regs, sz); 1396 array = (void *)array + sz; 1397 } else { 1398 *array++ = 0; 1399 } 1400 } 1401 1402 if (type & PERF_SAMPLE_PHYS_ADDR) { 1403 *array = sample->phys_addr; 1404 array++; 1405 } 1406 1407 if (type & PERF_SAMPLE_AUX) { 1408 sz = sample->aux_sample.size; 1409 *array++ = sz; 1410 memcpy(array, sample->aux_sample.data, sz); 1411 array = (void *)array + sz; 1412 } 1413 1414 return 0; 1415 } 1416 1417 int perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process, 1418 struct evlist *evlist, struct machine *machine) 1419 { 1420 union perf_event *ev; 1421 struct evsel *evsel; 1422 size_t nr = 0, i = 0, sz, max_nr, n; 1423 int err; 1424 1425 pr_debug2("Synthesizing id index\n"); 1426 1427 max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) / 1428 sizeof(struct id_index_entry); 1429 1430 evlist__for_each_entry(evlist, evsel) 1431 nr += evsel->core.ids; 1432 1433 n = nr > max_nr ? max_nr : nr; 1434 sz = sizeof(struct perf_record_id_index) + n * sizeof(struct id_index_entry); 1435 ev = zalloc(sz); 1436 if (!ev) 1437 return -ENOMEM; 1438 1439 ev->id_index.header.type = PERF_RECORD_ID_INDEX; 1440 ev->id_index.header.size = sz; 1441 ev->id_index.nr = n; 1442 1443 evlist__for_each_entry(evlist, evsel) { 1444 u32 j; 1445 1446 for (j = 0; j < evsel->core.ids; j++) { 1447 struct id_index_entry *e; 1448 struct perf_sample_id *sid; 1449 1450 if (i >= n) { 1451 err = process(tool, ev, NULL, machine); 1452 if (err) 1453 goto out_err; 1454 nr -= n; 1455 i = 0; 1456 } 1457 1458 e = &ev->id_index.entries[i++]; 1459 1460 e->id = evsel->core.id[j]; 1461 1462 sid = perf_evlist__id2sid(evlist, e->id); 1463 if (!sid) { 1464 free(ev); 1465 return -ENOENT; 1466 } 1467 1468 e->idx = sid->idx; 1469 e->cpu = sid->cpu; 1470 e->tid = sid->tid; 1471 } 1472 } 1473 1474 sz = sizeof(struct perf_record_id_index) + nr * sizeof(struct id_index_entry); 1475 ev->id_index.header.size = sz; 1476 ev->id_index.nr = nr; 1477 1478 err = process(tool, ev, NULL, machine); 1479 out_err: 1480 free(ev); 1481 1482 return err; 1483 } 1484 1485 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool, 1486 struct target *target, struct perf_thread_map *threads, 1487 perf_event__handler_t process, bool data_mmap, 1488 unsigned int nr_threads_synthesize) 1489 { 1490 if (target__has_task(target)) 1491 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap); 1492 else if (target__has_cpu(target)) 1493 return perf_event__synthesize_threads(tool, process, 1494 machine, data_mmap, 1495 nr_threads_synthesize); 1496 /* command specified */ 1497 return 0; 1498 } 1499 1500 int machine__synthesize_threads(struct machine *machine, struct target *target, 1501 struct perf_thread_map *threads, bool data_mmap, 1502 unsigned int nr_threads_synthesize) 1503 { 1504 return __machine__synthesize_threads(machine, NULL, target, threads, 1505 perf_event__process, data_mmap, 1506 nr_threads_synthesize); 1507 } 1508 1509 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id) 1510 { 1511 struct perf_record_event_update *ev; 1512 1513 size += sizeof(*ev); 1514 size = PERF_ALIGN(size, sizeof(u64)); 1515 1516 ev = zalloc(size); 1517 if (ev) { 1518 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1519 ev->header.size = (u16)size; 1520 ev->type = type; 1521 ev->id = id; 1522 } 1523 return ev; 1524 } 1525 1526 int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evsel *evsel, 1527 perf_event__handler_t process) 1528 { 1529 size_t size = strlen(evsel->unit); 1530 struct perf_record_event_update *ev; 1531 int err; 1532 1533 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]); 1534 if (ev == NULL) 1535 return -ENOMEM; 1536 1537 strlcpy(ev->data, evsel->unit, size + 1); 1538 err = process(tool, (union perf_event *)ev, NULL, NULL); 1539 free(ev); 1540 return err; 1541 } 1542 1543 int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsel *evsel, 1544 perf_event__handler_t process) 1545 { 1546 struct perf_record_event_update *ev; 1547 struct perf_record_event_update_scale *ev_data; 1548 int err; 1549 1550 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]); 1551 if (ev == NULL) 1552 return -ENOMEM; 1553 1554 ev_data = (struct perf_record_event_update_scale *)ev->data; 1555 ev_data->scale = evsel->scale; 1556 err = process(tool, (union perf_event *)ev, NULL, NULL); 1557 free(ev); 1558 return err; 1559 } 1560 1561 int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evsel *evsel, 1562 perf_event__handler_t process) 1563 { 1564 struct perf_record_event_update *ev; 1565 size_t len = strlen(evsel->name); 1566 int err; 1567 1568 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]); 1569 if (ev == NULL) 1570 return -ENOMEM; 1571 1572 strlcpy(ev->data, evsel->name, len + 1); 1573 err = process(tool, (union perf_event *)ev, NULL, NULL); 1574 free(ev); 1575 return err; 1576 } 1577 1578 int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel, 1579 perf_event__handler_t process) 1580 { 1581 size_t size = sizeof(struct perf_record_event_update); 1582 struct perf_record_event_update *ev; 1583 int max, err; 1584 u16 type; 1585 1586 if (!evsel->core.own_cpus) 1587 return 0; 1588 1589 ev = cpu_map_data__alloc(evsel->core.own_cpus, &size, &type, &max); 1590 if (!ev) 1591 return -ENOMEM; 1592 1593 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1594 ev->header.size = (u16)size; 1595 ev->type = PERF_EVENT_UPDATE__CPUS; 1596 ev->id = evsel->core.id[0]; 1597 1598 cpu_map_data__synthesize((struct perf_record_cpu_map_data *)ev->data, 1599 evsel->core.own_cpus, type, max); 1600 1601 err = process(tool, (union perf_event *)ev, NULL, NULL); 1602 free(ev); 1603 return err; 1604 } 1605 1606 int perf_event__synthesize_attrs(struct perf_tool *tool, struct evlist *evlist, 1607 perf_event__handler_t process) 1608 { 1609 struct evsel *evsel; 1610 int err = 0; 1611 1612 evlist__for_each_entry(evlist, evsel) { 1613 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids, 1614 evsel->core.id, process); 1615 if (err) { 1616 pr_debug("failed to create perf header attribute\n"); 1617 return err; 1618 } 1619 } 1620 1621 return err; 1622 } 1623 1624 static bool has_unit(struct evsel *evsel) 1625 { 1626 return evsel->unit && *evsel->unit; 1627 } 1628 1629 static bool has_scale(struct evsel *evsel) 1630 { 1631 return evsel->scale != 1; 1632 } 1633 1634 int perf_event__synthesize_extra_attr(struct perf_tool *tool, struct evlist *evsel_list, 1635 perf_event__handler_t process, bool is_pipe) 1636 { 1637 struct evsel *evsel; 1638 int err; 1639 1640 /* 1641 * Synthesize other events stuff not carried within 1642 * attr event - unit, scale, name 1643 */ 1644 evlist__for_each_entry(evsel_list, evsel) { 1645 if (!evsel->supported) 1646 continue; 1647 1648 /* 1649 * Synthesize unit and scale only if it's defined. 1650 */ 1651 if (has_unit(evsel)) { 1652 err = perf_event__synthesize_event_update_unit(tool, evsel, process); 1653 if (err < 0) { 1654 pr_err("Couldn't synthesize evsel unit.\n"); 1655 return err; 1656 } 1657 } 1658 1659 if (has_scale(evsel)) { 1660 err = perf_event__synthesize_event_update_scale(tool, evsel, process); 1661 if (err < 0) { 1662 pr_err("Couldn't synthesize evsel evsel.\n"); 1663 return err; 1664 } 1665 } 1666 1667 if (evsel->core.own_cpus) { 1668 err = perf_event__synthesize_event_update_cpus(tool, evsel, process); 1669 if (err < 0) { 1670 pr_err("Couldn't synthesize evsel cpus.\n"); 1671 return err; 1672 } 1673 } 1674 1675 /* 1676 * Name is needed only for pipe output, 1677 * perf.data carries event names. 1678 */ 1679 if (is_pipe) { 1680 err = perf_event__synthesize_event_update_name(tool, evsel, process); 1681 if (err < 0) { 1682 pr_err("Couldn't synthesize evsel name.\n"); 1683 return err; 1684 } 1685 } 1686 } 1687 return 0; 1688 } 1689 1690 int perf_event__synthesize_attr(struct perf_tool *tool, struct perf_event_attr *attr, 1691 u32 ids, u64 *id, perf_event__handler_t process) 1692 { 1693 union perf_event *ev; 1694 size_t size; 1695 int err; 1696 1697 size = sizeof(struct perf_event_attr); 1698 size = PERF_ALIGN(size, sizeof(u64)); 1699 size += sizeof(struct perf_event_header); 1700 size += ids * sizeof(u64); 1701 1702 ev = zalloc(size); 1703 1704 if (ev == NULL) 1705 return -ENOMEM; 1706 1707 ev->attr.attr = *attr; 1708 memcpy(ev->attr.id, id, ids * sizeof(u64)); 1709 1710 ev->attr.header.type = PERF_RECORD_HEADER_ATTR; 1711 ev->attr.header.size = (u16)size; 1712 1713 if (ev->attr.header.size == size) 1714 err = process(tool, ev, NULL, NULL); 1715 else 1716 err = -E2BIG; 1717 1718 free(ev); 1719 1720 return err; 1721 } 1722 1723 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, struct evlist *evlist, 1724 perf_event__handler_t process) 1725 { 1726 union perf_event ev; 1727 struct tracing_data *tdata; 1728 ssize_t size = 0, aligned_size = 0, padding; 1729 struct feat_fd ff; 1730 1731 /* 1732 * We are going to store the size of the data followed 1733 * by the data contents. Since the fd descriptor is a pipe, 1734 * we cannot seek back to store the size of the data once 1735 * we know it. Instead we: 1736 * 1737 * - write the tracing data to the temp file 1738 * - get/write the data size to pipe 1739 * - write the tracing data from the temp file 1740 * to the pipe 1741 */ 1742 tdata = tracing_data_get(&evlist->core.entries, fd, true); 1743 if (!tdata) 1744 return -1; 1745 1746 memset(&ev, 0, sizeof(ev)); 1747 1748 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA; 1749 size = tdata->size; 1750 aligned_size = PERF_ALIGN(size, sizeof(u64)); 1751 padding = aligned_size - size; 1752 ev.tracing_data.header.size = sizeof(ev.tracing_data); 1753 ev.tracing_data.size = aligned_size; 1754 1755 process(tool, &ev, NULL, NULL); 1756 1757 /* 1758 * The put function will copy all the tracing data 1759 * stored in temp file to the pipe. 1760 */ 1761 tracing_data_put(tdata); 1762 1763 ff = (struct feat_fd){ .fd = fd }; 1764 if (write_padded(&ff, NULL, 0, padding)) 1765 return -1; 1766 1767 return aligned_size; 1768 } 1769 1770 int perf_event__synthesize_build_id(struct perf_tool *tool, struct dso *pos, u16 misc, 1771 perf_event__handler_t process, struct machine *machine) 1772 { 1773 union perf_event ev; 1774 size_t len; 1775 1776 if (!pos->hit) 1777 return 0; 1778 1779 memset(&ev, 0, sizeof(ev)); 1780 1781 len = pos->long_name_len + 1; 1782 len = PERF_ALIGN(len, NAME_ALIGN); 1783 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id)); 1784 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID; 1785 ev.build_id.header.misc = misc; 1786 ev.build_id.pid = machine->pid; 1787 ev.build_id.header.size = sizeof(ev.build_id) + len; 1788 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len); 1789 1790 return process(tool, &ev, NULL, machine); 1791 } 1792 1793 int perf_event__synthesize_stat_events(struct perf_stat_config *config, struct perf_tool *tool, 1794 struct evlist *evlist, perf_event__handler_t process, bool attrs) 1795 { 1796 int err; 1797 1798 if (attrs) { 1799 err = perf_event__synthesize_attrs(tool, evlist, process); 1800 if (err < 0) { 1801 pr_err("Couldn't synthesize attrs.\n"); 1802 return err; 1803 } 1804 } 1805 1806 err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs); 1807 err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL); 1808 if (err < 0) { 1809 pr_err("Couldn't synthesize thread map.\n"); 1810 return err; 1811 } 1812 1813 err = perf_event__synthesize_cpu_map(tool, evlist->core.cpus, process, NULL); 1814 if (err < 0) { 1815 pr_err("Couldn't synthesize thread map.\n"); 1816 return err; 1817 } 1818 1819 err = perf_event__synthesize_stat_config(tool, config, process, NULL); 1820 if (err < 0) { 1821 pr_err("Couldn't synthesize config.\n"); 1822 return err; 1823 } 1824 1825 return 0; 1826 } 1827 1828 int __weak perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused, 1829 struct perf_tool *tool __maybe_unused, 1830 perf_event__handler_t process __maybe_unused, 1831 struct machine *machine __maybe_unused) 1832 { 1833 return 0; 1834 } 1835 1836 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE]; 1837 1838 int perf_event__synthesize_features(struct perf_tool *tool, struct perf_session *session, 1839 struct evlist *evlist, perf_event__handler_t process) 1840 { 1841 struct perf_header *header = &session->header; 1842 struct perf_record_header_feature *fe; 1843 struct feat_fd ff; 1844 size_t sz, sz_hdr; 1845 int feat, ret; 1846 1847 sz_hdr = sizeof(fe->header); 1848 sz = sizeof(union perf_event); 1849 /* get a nice alignment */ 1850 sz = PERF_ALIGN(sz, page_size); 1851 1852 memset(&ff, 0, sizeof(ff)); 1853 1854 ff.buf = malloc(sz); 1855 if (!ff.buf) 1856 return -ENOMEM; 1857 1858 ff.size = sz - sz_hdr; 1859 ff.ph = &session->header; 1860 1861 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) { 1862 if (!feat_ops[feat].synthesize) { 1863 pr_debug("No record header feature for header :%d\n", feat); 1864 continue; 1865 } 1866 1867 ff.offset = sizeof(*fe); 1868 1869 ret = feat_ops[feat].write(&ff, evlist); 1870 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) { 1871 pr_debug("Error writing feature\n"); 1872 continue; 1873 } 1874 /* ff.buf may have changed due to realloc in do_write() */ 1875 fe = ff.buf; 1876 memset(fe, 0, sizeof(*fe)); 1877 1878 fe->feat_id = feat; 1879 fe->header.type = PERF_RECORD_HEADER_FEATURE; 1880 fe->header.size = ff.offset; 1881 1882 ret = process(tool, ff.buf, NULL, NULL); 1883 if (ret) { 1884 free(ff.buf); 1885 return ret; 1886 } 1887 } 1888 1889 /* Send HEADER_LAST_FEATURE mark. */ 1890 fe = ff.buf; 1891 fe->feat_id = HEADER_LAST_FEATURE; 1892 fe->header.type = PERF_RECORD_HEADER_FEATURE; 1893 fe->header.size = sizeof(*fe); 1894 1895 ret = process(tool, ff.buf, NULL, NULL); 1896 1897 free(ff.buf); 1898 return ret; 1899 } 1900