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 "util/cgroup.h" 20 #include <linux/bitops.h> 21 #include <linux/kernel.h> 22 #include <linux/string.h> 23 #include <linux/zalloc.h> 24 #include <linux/perf_event.h> 25 #include <asm/bug.h> 26 #include <perf/evsel.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 <api/io.h> 40 #include <sys/types.h> 41 #include <sys/stat.h> 42 #include <fcntl.h> 43 #include <unistd.h> 44 45 #define DEFAULT_PROC_MAP_PARSE_TIMEOUT 500 46 47 unsigned int proc_map_timeout = DEFAULT_PROC_MAP_PARSE_TIMEOUT; 48 49 int perf_tool__process_synth_event(struct perf_tool *tool, 50 union perf_event *event, 51 struct machine *machine, 52 perf_event__handler_t process) 53 { 54 struct perf_sample synth_sample = { 55 .pid = -1, 56 .tid = -1, 57 .time = -1, 58 .stream_id = -1, 59 .cpu = -1, 60 .period = 1, 61 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK, 62 }; 63 64 return process(tool, event, &synth_sample, machine); 65 }; 66 67 /* 68 * Assumes that the first 4095 bytes of /proc/pid/stat contains 69 * the comm, tgid and ppid. 70 */ 71 static int perf_event__get_comm_ids(pid_t pid, pid_t tid, char *comm, size_t len, 72 pid_t *tgid, pid_t *ppid, bool *kernel) 73 { 74 char bf[4096]; 75 int fd; 76 size_t size = 0; 77 ssize_t n; 78 char *name, *tgids, *ppids, *vmpeak, *threads; 79 80 *tgid = -1; 81 *ppid = -1; 82 83 if (pid) 84 snprintf(bf, sizeof(bf), "/proc/%d/task/%d/status", pid, tid); 85 else 86 snprintf(bf, sizeof(bf), "/proc/%d/status", tid); 87 88 fd = open(bf, O_RDONLY); 89 if (fd < 0) { 90 pr_debug("couldn't open %s\n", bf); 91 return -1; 92 } 93 94 n = read(fd, bf, sizeof(bf) - 1); 95 close(fd); 96 if (n <= 0) { 97 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n", 98 tid); 99 return -1; 100 } 101 bf[n] = '\0'; 102 103 name = strstr(bf, "Name:"); 104 tgids = strstr(name ?: bf, "Tgid:"); 105 ppids = strstr(tgids ?: bf, "PPid:"); 106 vmpeak = strstr(ppids ?: bf, "VmPeak:"); 107 108 if (vmpeak) 109 threads = NULL; 110 else 111 threads = strstr(ppids ?: bf, "Threads:"); 112 113 if (name) { 114 char *nl; 115 116 name = skip_spaces(name + 5); /* strlen("Name:") */ 117 nl = strchr(name, '\n'); 118 if (nl) 119 *nl = '\0'; 120 121 size = strlen(name); 122 if (size >= len) 123 size = len - 1; 124 memcpy(comm, name, size); 125 comm[size] = '\0'; 126 } else { 127 pr_debug("Name: string not found for pid %d\n", tid); 128 } 129 130 if (tgids) { 131 tgids += 5; /* strlen("Tgid:") */ 132 *tgid = atoi(tgids); 133 } else { 134 pr_debug("Tgid: string not found for pid %d\n", tid); 135 } 136 137 if (ppids) { 138 ppids += 5; /* strlen("PPid:") */ 139 *ppid = atoi(ppids); 140 } else { 141 pr_debug("PPid: string not found for pid %d\n", tid); 142 } 143 144 if (!vmpeak && threads) 145 *kernel = true; 146 else 147 *kernel = false; 148 149 return 0; 150 } 151 152 static int perf_event__prepare_comm(union perf_event *event, pid_t pid, pid_t tid, 153 struct machine *machine, 154 pid_t *tgid, pid_t *ppid, bool *kernel) 155 { 156 size_t size; 157 158 *ppid = -1; 159 160 memset(&event->comm, 0, sizeof(event->comm)); 161 162 if (machine__is_host(machine)) { 163 if (perf_event__get_comm_ids(pid, tid, event->comm.comm, 164 sizeof(event->comm.comm), 165 tgid, ppid, kernel) != 0) { 166 return -1; 167 } 168 } else { 169 *tgid = machine->pid; 170 } 171 172 if (*tgid < 0) 173 return -1; 174 175 event->comm.pid = *tgid; 176 event->comm.header.type = PERF_RECORD_COMM; 177 178 size = strlen(event->comm.comm) + 1; 179 size = PERF_ALIGN(size, sizeof(u64)); 180 memset(event->comm.comm + size, 0, machine->id_hdr_size); 181 event->comm.header.size = (sizeof(event->comm) - 182 (sizeof(event->comm.comm) - size) + 183 machine->id_hdr_size); 184 event->comm.tid = tid; 185 186 return 0; 187 } 188 189 pid_t perf_event__synthesize_comm(struct perf_tool *tool, 190 union perf_event *event, pid_t pid, 191 perf_event__handler_t process, 192 struct machine *machine) 193 { 194 pid_t tgid, ppid; 195 bool kernel_thread; 196 197 if (perf_event__prepare_comm(event, 0, pid, machine, &tgid, &ppid, 198 &kernel_thread) != 0) 199 return -1; 200 201 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 202 return -1; 203 204 return tgid; 205 } 206 207 static void perf_event__get_ns_link_info(pid_t pid, const char *ns, 208 struct perf_ns_link_info *ns_link_info) 209 { 210 struct stat64 st; 211 char proc_ns[128]; 212 213 sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns); 214 if (stat64(proc_ns, &st) == 0) { 215 ns_link_info->dev = st.st_dev; 216 ns_link_info->ino = st.st_ino; 217 } 218 } 219 220 int perf_event__synthesize_namespaces(struct perf_tool *tool, 221 union perf_event *event, 222 pid_t pid, pid_t tgid, 223 perf_event__handler_t process, 224 struct machine *machine) 225 { 226 u32 idx; 227 struct perf_ns_link_info *ns_link_info; 228 229 if (!tool || !tool->namespace_events) 230 return 0; 231 232 memset(&event->namespaces, 0, (sizeof(event->namespaces) + 233 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 234 machine->id_hdr_size)); 235 236 event->namespaces.pid = tgid; 237 event->namespaces.tid = pid; 238 239 event->namespaces.nr_namespaces = NR_NAMESPACES; 240 241 ns_link_info = event->namespaces.link_info; 242 243 for (idx = 0; idx < event->namespaces.nr_namespaces; idx++) 244 perf_event__get_ns_link_info(pid, perf_ns__name(idx), 245 &ns_link_info[idx]); 246 247 event->namespaces.header.type = PERF_RECORD_NAMESPACES; 248 249 event->namespaces.header.size = (sizeof(event->namespaces) + 250 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 251 machine->id_hdr_size); 252 253 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 254 return -1; 255 256 return 0; 257 } 258 259 static int perf_event__synthesize_fork(struct perf_tool *tool, 260 union perf_event *event, 261 pid_t pid, pid_t tgid, pid_t ppid, 262 perf_event__handler_t process, 263 struct machine *machine) 264 { 265 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size); 266 267 /* 268 * for main thread set parent to ppid from status file. For other 269 * threads set parent pid to main thread. ie., assume main thread 270 * spawns all threads in a process 271 */ 272 if (tgid == pid) { 273 event->fork.ppid = ppid; 274 event->fork.ptid = ppid; 275 } else { 276 event->fork.ppid = tgid; 277 event->fork.ptid = tgid; 278 } 279 event->fork.pid = tgid; 280 event->fork.tid = pid; 281 event->fork.header.type = PERF_RECORD_FORK; 282 event->fork.header.misc = PERF_RECORD_MISC_FORK_EXEC; 283 284 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size); 285 286 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 287 return -1; 288 289 return 0; 290 } 291 292 static bool read_proc_maps_line(struct io *io, __u64 *start, __u64 *end, 293 u32 *prot, u32 *flags, __u64 *offset, 294 u32 *maj, u32 *min, 295 __u64 *inode, 296 ssize_t pathname_size, char *pathname) 297 { 298 __u64 temp; 299 int ch; 300 char *start_pathname = pathname; 301 302 if (io__get_hex(io, start) != '-') 303 return false; 304 if (io__get_hex(io, end) != ' ') 305 return false; 306 307 /* map protection and flags bits */ 308 *prot = 0; 309 ch = io__get_char(io); 310 if (ch == 'r') 311 *prot |= PROT_READ; 312 else if (ch != '-') 313 return false; 314 ch = io__get_char(io); 315 if (ch == 'w') 316 *prot |= PROT_WRITE; 317 else if (ch != '-') 318 return false; 319 ch = io__get_char(io); 320 if (ch == 'x') 321 *prot |= PROT_EXEC; 322 else if (ch != '-') 323 return false; 324 ch = io__get_char(io); 325 if (ch == 's') 326 *flags = MAP_SHARED; 327 else if (ch == 'p') 328 *flags = MAP_PRIVATE; 329 else 330 return false; 331 if (io__get_char(io) != ' ') 332 return false; 333 334 if (io__get_hex(io, offset) != ' ') 335 return false; 336 337 if (io__get_hex(io, &temp) != ':') 338 return false; 339 *maj = temp; 340 if (io__get_hex(io, &temp) != ' ') 341 return false; 342 *min = temp; 343 344 ch = io__get_dec(io, inode); 345 if (ch != ' ') { 346 *pathname = '\0'; 347 return ch == '\n'; 348 } 349 do { 350 ch = io__get_char(io); 351 } while (ch == ' '); 352 while (true) { 353 if (ch < 0) 354 return false; 355 if (ch == '\0' || ch == '\n' || 356 (pathname + 1 - start_pathname) >= pathname_size) { 357 *pathname = '\0'; 358 return true; 359 } 360 *pathname++ = ch; 361 ch = io__get_char(io); 362 } 363 } 364 365 static void perf_record_mmap2__read_build_id(struct perf_record_mmap2 *event, 366 bool is_kernel) 367 { 368 struct build_id bid; 369 int rc; 370 371 if (is_kernel) 372 rc = sysfs__read_build_id("/sys/kernel/notes", &bid); 373 else 374 rc = filename__read_build_id(event->filename, &bid) > 0 ? 0 : -1; 375 376 if (rc == 0) { 377 memcpy(event->build_id, bid.data, sizeof(bid.data)); 378 event->build_id_size = (u8) bid.size; 379 event->header.misc |= PERF_RECORD_MISC_MMAP_BUILD_ID; 380 event->__reserved_1 = 0; 381 event->__reserved_2 = 0; 382 } else { 383 if (event->filename[0] == '/') { 384 pr_debug2("Failed to read build ID for %s\n", 385 event->filename); 386 } 387 } 388 } 389 390 int perf_event__synthesize_mmap_events(struct perf_tool *tool, 391 union perf_event *event, 392 pid_t pid, pid_t tgid, 393 perf_event__handler_t process, 394 struct machine *machine, 395 bool mmap_data) 396 { 397 unsigned long long t; 398 char bf[BUFSIZ]; 399 struct io io; 400 bool truncation = false; 401 unsigned long long timeout = proc_map_timeout * 1000000ULL; 402 int rc = 0; 403 const char *hugetlbfs_mnt = hugetlbfs__mountpoint(); 404 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0; 405 406 if (machine__is_default_guest(machine)) 407 return 0; 408 409 snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps", 410 machine->root_dir, pid, pid); 411 412 io.fd = open(bf, O_RDONLY, 0); 413 if (io.fd < 0) { 414 /* 415 * We raced with a task exiting - just return: 416 */ 417 pr_debug("couldn't open %s\n", bf); 418 return -1; 419 } 420 io__init(&io, io.fd, bf, sizeof(bf)); 421 422 event->header.type = PERF_RECORD_MMAP2; 423 t = rdclock(); 424 425 while (!io.eof) { 426 static const char anonstr[] = "//anon"; 427 size_t size; 428 429 /* ensure null termination since stack will be reused. */ 430 event->mmap2.filename[0] = '\0'; 431 432 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */ 433 if (!read_proc_maps_line(&io, 434 &event->mmap2.start, 435 &event->mmap2.len, 436 &event->mmap2.prot, 437 &event->mmap2.flags, 438 &event->mmap2.pgoff, 439 &event->mmap2.maj, 440 &event->mmap2.min, 441 &event->mmap2.ino, 442 sizeof(event->mmap2.filename), 443 event->mmap2.filename)) 444 continue; 445 446 if ((rdclock() - t) > timeout) { 447 pr_warning("Reading %s/proc/%d/task/%d/maps time out. " 448 "You may want to increase " 449 "the time limit by --proc-map-timeout\n", 450 machine->root_dir, pid, pid); 451 truncation = true; 452 goto out; 453 } 454 455 event->mmap2.ino_generation = 0; 456 457 /* 458 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c 459 */ 460 if (machine__is_host(machine)) 461 event->header.misc = PERF_RECORD_MISC_USER; 462 else 463 event->header.misc = PERF_RECORD_MISC_GUEST_USER; 464 465 if ((event->mmap2.prot & PROT_EXEC) == 0) { 466 if (!mmap_data || (event->mmap2.prot & PROT_READ) == 0) 467 continue; 468 469 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA; 470 } 471 472 out: 473 if (truncation) 474 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT; 475 476 if (!strcmp(event->mmap2.filename, "")) 477 strcpy(event->mmap2.filename, anonstr); 478 479 if (hugetlbfs_mnt_len && 480 !strncmp(event->mmap2.filename, hugetlbfs_mnt, 481 hugetlbfs_mnt_len)) { 482 strcpy(event->mmap2.filename, anonstr); 483 event->mmap2.flags |= MAP_HUGETLB; 484 } 485 486 size = strlen(event->mmap2.filename) + 1; 487 size = PERF_ALIGN(size, sizeof(u64)); 488 event->mmap2.len -= event->mmap.start; 489 event->mmap2.header.size = (sizeof(event->mmap2) - 490 (sizeof(event->mmap2.filename) - size)); 491 memset(event->mmap2.filename + size, 0, machine->id_hdr_size); 492 event->mmap2.header.size += machine->id_hdr_size; 493 event->mmap2.pid = tgid; 494 event->mmap2.tid = pid; 495 496 if (symbol_conf.buildid_mmap2) 497 perf_record_mmap2__read_build_id(&event->mmap2, false); 498 499 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 500 rc = -1; 501 break; 502 } 503 504 if (truncation) 505 break; 506 } 507 508 close(io.fd); 509 return rc; 510 } 511 512 #ifdef HAVE_FILE_HANDLE 513 static int perf_event__synthesize_cgroup(struct perf_tool *tool, 514 union perf_event *event, 515 char *path, size_t mount_len, 516 perf_event__handler_t process, 517 struct machine *machine) 518 { 519 size_t event_size = sizeof(event->cgroup) - sizeof(event->cgroup.path); 520 size_t path_len = strlen(path) - mount_len + 1; 521 struct { 522 struct file_handle fh; 523 uint64_t cgroup_id; 524 } handle; 525 int mount_id; 526 527 while (path_len % sizeof(u64)) 528 path[mount_len + path_len++] = '\0'; 529 530 memset(&event->cgroup, 0, event_size); 531 532 event->cgroup.header.type = PERF_RECORD_CGROUP; 533 event->cgroup.header.size = event_size + path_len + machine->id_hdr_size; 534 535 handle.fh.handle_bytes = sizeof(handle.cgroup_id); 536 if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0) { 537 pr_debug("stat failed: %s\n", path); 538 return -1; 539 } 540 541 event->cgroup.id = handle.cgroup_id; 542 strncpy(event->cgroup.path, path + mount_len, path_len); 543 memset(event->cgroup.path + path_len, 0, machine->id_hdr_size); 544 545 if (perf_tool__process_synth_event(tool, event, machine, process) < 0) { 546 pr_debug("process synth event failed\n"); 547 return -1; 548 } 549 550 return 0; 551 } 552 553 static int perf_event__walk_cgroup_tree(struct perf_tool *tool, 554 union perf_event *event, 555 char *path, size_t mount_len, 556 perf_event__handler_t process, 557 struct machine *machine) 558 { 559 size_t pos = strlen(path); 560 DIR *d; 561 struct dirent *dent; 562 int ret = 0; 563 564 if (perf_event__synthesize_cgroup(tool, event, path, mount_len, 565 process, machine) < 0) 566 return -1; 567 568 d = opendir(path); 569 if (d == NULL) { 570 pr_debug("failed to open directory: %s\n", path); 571 return -1; 572 } 573 574 while ((dent = readdir(d)) != NULL) { 575 if (dent->d_type != DT_DIR) 576 continue; 577 if (!strcmp(dent->d_name, ".") || 578 !strcmp(dent->d_name, "..")) 579 continue; 580 581 /* any sane path should be less than PATH_MAX */ 582 if (strlen(path) + strlen(dent->d_name) + 1 >= PATH_MAX) 583 continue; 584 585 if (path[pos - 1] != '/') 586 strcat(path, "/"); 587 strcat(path, dent->d_name); 588 589 ret = perf_event__walk_cgroup_tree(tool, event, path, 590 mount_len, process, machine); 591 if (ret < 0) 592 break; 593 594 path[pos] = '\0'; 595 } 596 597 closedir(d); 598 return ret; 599 } 600 601 int perf_event__synthesize_cgroups(struct perf_tool *tool, 602 perf_event__handler_t process, 603 struct machine *machine) 604 { 605 union perf_event event; 606 char cgrp_root[PATH_MAX]; 607 size_t mount_len; /* length of mount point in the path */ 608 609 if (!tool || !tool->cgroup_events) 610 return 0; 611 612 if (cgroupfs_find_mountpoint(cgrp_root, PATH_MAX, "perf_event") < 0) { 613 pr_debug("cannot find cgroup mount point\n"); 614 return -1; 615 } 616 617 mount_len = strlen(cgrp_root); 618 /* make sure the path starts with a slash (after mount point) */ 619 strcat(cgrp_root, "/"); 620 621 if (perf_event__walk_cgroup_tree(tool, &event, cgrp_root, mount_len, 622 process, machine) < 0) 623 return -1; 624 625 return 0; 626 } 627 #else 628 int perf_event__synthesize_cgroups(struct perf_tool *tool __maybe_unused, 629 perf_event__handler_t process __maybe_unused, 630 struct machine *machine __maybe_unused) 631 { 632 return -1; 633 } 634 #endif 635 636 int perf_event__synthesize_modules(struct perf_tool *tool, perf_event__handler_t process, 637 struct machine *machine) 638 { 639 int rc = 0; 640 struct map *pos; 641 struct maps *maps = machine__kernel_maps(machine); 642 union perf_event *event; 643 size_t size = symbol_conf.buildid_mmap2 ? 644 sizeof(event->mmap2) : sizeof(event->mmap); 645 646 event = zalloc(size + machine->id_hdr_size); 647 if (event == NULL) { 648 pr_debug("Not enough memory synthesizing mmap event " 649 "for kernel modules\n"); 650 return -1; 651 } 652 653 /* 654 * kernel uses 0 for user space maps, see kernel/perf_event.c 655 * __perf_event_mmap 656 */ 657 if (machine__is_host(machine)) 658 event->header.misc = PERF_RECORD_MISC_KERNEL; 659 else 660 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 661 662 maps__for_each_entry(maps, pos) { 663 if (!__map__is_kmodule(pos)) 664 continue; 665 666 if (symbol_conf.buildid_mmap2) { 667 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64)); 668 event->mmap2.header.type = PERF_RECORD_MMAP2; 669 event->mmap2.header.size = (sizeof(event->mmap2) - 670 (sizeof(event->mmap2.filename) - size)); 671 memset(event->mmap2.filename + size, 0, machine->id_hdr_size); 672 event->mmap2.header.size += machine->id_hdr_size; 673 event->mmap2.start = pos->start; 674 event->mmap2.len = pos->end - pos->start; 675 event->mmap2.pid = machine->pid; 676 677 memcpy(event->mmap2.filename, pos->dso->long_name, 678 pos->dso->long_name_len + 1); 679 680 perf_record_mmap2__read_build_id(&event->mmap2, false); 681 } else { 682 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64)); 683 event->mmap.header.type = PERF_RECORD_MMAP; 684 event->mmap.header.size = (sizeof(event->mmap) - 685 (sizeof(event->mmap.filename) - size)); 686 memset(event->mmap.filename + size, 0, machine->id_hdr_size); 687 event->mmap.header.size += machine->id_hdr_size; 688 event->mmap.start = pos->start; 689 event->mmap.len = pos->end - pos->start; 690 event->mmap.pid = machine->pid; 691 692 memcpy(event->mmap.filename, pos->dso->long_name, 693 pos->dso->long_name_len + 1); 694 } 695 696 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 697 rc = -1; 698 break; 699 } 700 } 701 702 free(event); 703 return rc; 704 } 705 706 static int filter_task(const struct dirent *dirent) 707 { 708 return isdigit(dirent->d_name[0]); 709 } 710 711 static int __event__synthesize_thread(union perf_event *comm_event, 712 union perf_event *mmap_event, 713 union perf_event *fork_event, 714 union perf_event *namespaces_event, 715 pid_t pid, int full, perf_event__handler_t process, 716 struct perf_tool *tool, struct machine *machine, bool mmap_data) 717 { 718 char filename[PATH_MAX]; 719 struct dirent **dirent; 720 pid_t tgid, ppid; 721 int rc = 0; 722 int i, n; 723 724 /* special case: only send one comm event using passed in pid */ 725 if (!full) { 726 tgid = perf_event__synthesize_comm(tool, comm_event, pid, 727 process, machine); 728 729 if (tgid == -1) 730 return -1; 731 732 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid, 733 tgid, process, machine) < 0) 734 return -1; 735 736 /* 737 * send mmap only for thread group leader 738 * see thread__init_maps() 739 */ 740 if (pid == tgid && 741 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 742 process, machine, mmap_data)) 743 return -1; 744 745 return 0; 746 } 747 748 if (machine__is_default_guest(machine)) 749 return 0; 750 751 snprintf(filename, sizeof(filename), "%s/proc/%d/task", 752 machine->root_dir, pid); 753 754 n = scandir(filename, &dirent, filter_task, alphasort); 755 if (n < 0) 756 return n; 757 758 for (i = 0; i < n; i++) { 759 char *end; 760 pid_t _pid; 761 bool kernel_thread; 762 763 _pid = strtol(dirent[i]->d_name, &end, 10); 764 if (*end) 765 continue; 766 767 rc = -1; 768 if (perf_event__prepare_comm(comm_event, pid, _pid, machine, 769 &tgid, &ppid, &kernel_thread) != 0) 770 break; 771 772 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid, 773 ppid, process, machine) < 0) 774 break; 775 776 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid, 777 tgid, process, machine) < 0) 778 break; 779 780 /* 781 * Send the prepared comm event 782 */ 783 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0) 784 break; 785 786 rc = 0; 787 if (_pid == pid && !kernel_thread) { 788 /* process the parent's maps too */ 789 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 790 process, machine, mmap_data); 791 if (rc) 792 break; 793 } 794 } 795 796 for (i = 0; i < n; i++) 797 zfree(&dirent[i]); 798 free(dirent); 799 800 return rc; 801 } 802 803 int perf_event__synthesize_thread_map(struct perf_tool *tool, 804 struct perf_thread_map *threads, 805 perf_event__handler_t process, 806 struct machine *machine, 807 bool mmap_data) 808 { 809 union perf_event *comm_event, *mmap_event, *fork_event; 810 union perf_event *namespaces_event; 811 int err = -1, thread, j; 812 813 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 814 if (comm_event == NULL) 815 goto out; 816 817 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 818 if (mmap_event == NULL) 819 goto out_free_comm; 820 821 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 822 if (fork_event == NULL) 823 goto out_free_mmap; 824 825 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 826 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 827 machine->id_hdr_size); 828 if (namespaces_event == NULL) 829 goto out_free_fork; 830 831 err = 0; 832 for (thread = 0; thread < threads->nr; ++thread) { 833 if (__event__synthesize_thread(comm_event, mmap_event, 834 fork_event, namespaces_event, 835 perf_thread_map__pid(threads, thread), 0, 836 process, tool, machine, 837 mmap_data)) { 838 err = -1; 839 break; 840 } 841 842 /* 843 * comm.pid is set to thread group id by 844 * perf_event__synthesize_comm 845 */ 846 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) { 847 bool need_leader = true; 848 849 /* is thread group leader in thread_map? */ 850 for (j = 0; j < threads->nr; ++j) { 851 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) { 852 need_leader = false; 853 break; 854 } 855 } 856 857 /* if not, generate events for it */ 858 if (need_leader && 859 __event__synthesize_thread(comm_event, mmap_event, 860 fork_event, namespaces_event, 861 comm_event->comm.pid, 0, 862 process, tool, machine, 863 mmap_data)) { 864 err = -1; 865 break; 866 } 867 } 868 } 869 free(namespaces_event); 870 out_free_fork: 871 free(fork_event); 872 out_free_mmap: 873 free(mmap_event); 874 out_free_comm: 875 free(comm_event); 876 out: 877 return err; 878 } 879 880 static int __perf_event__synthesize_threads(struct perf_tool *tool, 881 perf_event__handler_t process, 882 struct machine *machine, 883 bool mmap_data, 884 struct dirent **dirent, 885 int start, 886 int num) 887 { 888 union perf_event *comm_event, *mmap_event, *fork_event; 889 union perf_event *namespaces_event; 890 int err = -1; 891 char *end; 892 pid_t pid; 893 int i; 894 895 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 896 if (comm_event == NULL) 897 goto out; 898 899 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 900 if (mmap_event == NULL) 901 goto out_free_comm; 902 903 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 904 if (fork_event == NULL) 905 goto out_free_mmap; 906 907 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 908 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 909 machine->id_hdr_size); 910 if (namespaces_event == NULL) 911 goto out_free_fork; 912 913 for (i = start; i < start + num; i++) { 914 if (!isdigit(dirent[i]->d_name[0])) 915 continue; 916 917 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10); 918 /* only interested in proper numerical dirents */ 919 if (*end) 920 continue; 921 /* 922 * We may race with exiting thread, so don't stop just because 923 * one thread couldn't be synthesized. 924 */ 925 __event__synthesize_thread(comm_event, mmap_event, fork_event, 926 namespaces_event, pid, 1, process, 927 tool, machine, mmap_data); 928 } 929 err = 0; 930 931 free(namespaces_event); 932 out_free_fork: 933 free(fork_event); 934 out_free_mmap: 935 free(mmap_event); 936 out_free_comm: 937 free(comm_event); 938 out: 939 return err; 940 } 941 942 struct synthesize_threads_arg { 943 struct perf_tool *tool; 944 perf_event__handler_t process; 945 struct machine *machine; 946 bool mmap_data; 947 struct dirent **dirent; 948 int num; 949 int start; 950 }; 951 952 static void *synthesize_threads_worker(void *arg) 953 { 954 struct synthesize_threads_arg *args = arg; 955 956 __perf_event__synthesize_threads(args->tool, args->process, 957 args->machine, args->mmap_data, 958 args->dirent, 959 args->start, args->num); 960 return NULL; 961 } 962 963 int perf_event__synthesize_threads(struct perf_tool *tool, 964 perf_event__handler_t process, 965 struct machine *machine, 966 bool mmap_data, 967 unsigned int nr_threads_synthesize) 968 { 969 struct synthesize_threads_arg *args = NULL; 970 pthread_t *synthesize_threads = NULL; 971 char proc_path[PATH_MAX]; 972 struct dirent **dirent; 973 int num_per_thread; 974 int m, n, i, j; 975 int thread_nr; 976 int base = 0; 977 int err = -1; 978 979 980 if (machine__is_default_guest(machine)) 981 return 0; 982 983 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir); 984 n = scandir(proc_path, &dirent, filter_task, alphasort); 985 if (n < 0) 986 return err; 987 988 if (nr_threads_synthesize == UINT_MAX) 989 thread_nr = sysconf(_SC_NPROCESSORS_ONLN); 990 else 991 thread_nr = nr_threads_synthesize; 992 993 if (thread_nr <= 1) { 994 err = __perf_event__synthesize_threads(tool, process, 995 machine, mmap_data, 996 dirent, base, n); 997 goto free_dirent; 998 } 999 if (thread_nr > n) 1000 thread_nr = n; 1001 1002 synthesize_threads = calloc(sizeof(pthread_t), thread_nr); 1003 if (synthesize_threads == NULL) 1004 goto free_dirent; 1005 1006 args = calloc(sizeof(*args), thread_nr); 1007 if (args == NULL) 1008 goto free_threads; 1009 1010 num_per_thread = n / thread_nr; 1011 m = n % thread_nr; 1012 for (i = 0; i < thread_nr; i++) { 1013 args[i].tool = tool; 1014 args[i].process = process; 1015 args[i].machine = machine; 1016 args[i].mmap_data = mmap_data; 1017 args[i].dirent = dirent; 1018 } 1019 for (i = 0; i < m; i++) { 1020 args[i].num = num_per_thread + 1; 1021 args[i].start = i * args[i].num; 1022 } 1023 if (i != 0) 1024 base = args[i-1].start + args[i-1].num; 1025 for (j = i; j < thread_nr; j++) { 1026 args[j].num = num_per_thread; 1027 args[j].start = base + (j - i) * args[i].num; 1028 } 1029 1030 for (i = 0; i < thread_nr; i++) { 1031 if (pthread_create(&synthesize_threads[i], NULL, 1032 synthesize_threads_worker, &args[i])) 1033 goto out_join; 1034 } 1035 err = 0; 1036 out_join: 1037 for (i = 0; i < thread_nr; i++) 1038 pthread_join(synthesize_threads[i], NULL); 1039 free(args); 1040 free_threads: 1041 free(synthesize_threads); 1042 free_dirent: 1043 for (i = 0; i < n; i++) 1044 zfree(&dirent[i]); 1045 free(dirent); 1046 1047 return err; 1048 } 1049 1050 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused, 1051 perf_event__handler_t process __maybe_unused, 1052 struct machine *machine __maybe_unused) 1053 { 1054 return 0; 1055 } 1056 1057 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 1058 perf_event__handler_t process, 1059 struct machine *machine) 1060 { 1061 union perf_event *event; 1062 size_t size = symbol_conf.buildid_mmap2 ? 1063 sizeof(event->mmap2) : sizeof(event->mmap); 1064 struct map *map = machine__kernel_map(machine); 1065 struct kmap *kmap; 1066 int err; 1067 1068 if (map == NULL) 1069 return -1; 1070 1071 kmap = map__kmap(map); 1072 if (!kmap->ref_reloc_sym) 1073 return -1; 1074 1075 /* 1076 * We should get this from /sys/kernel/sections/.text, but till that is 1077 * available use this, and after it is use this as a fallback for older 1078 * kernels. 1079 */ 1080 event = zalloc(size + machine->id_hdr_size); 1081 if (event == NULL) { 1082 pr_debug("Not enough memory synthesizing mmap event " 1083 "for kernel modules\n"); 1084 return -1; 1085 } 1086 1087 if (machine__is_host(machine)) { 1088 /* 1089 * kernel uses PERF_RECORD_MISC_USER for user space maps, 1090 * see kernel/perf_event.c __perf_event_mmap 1091 */ 1092 event->header.misc = PERF_RECORD_MISC_KERNEL; 1093 } else { 1094 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 1095 } 1096 1097 if (symbol_conf.buildid_mmap2) { 1098 size = snprintf(event->mmap2.filename, sizeof(event->mmap2.filename), 1099 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1; 1100 size = PERF_ALIGN(size, sizeof(u64)); 1101 event->mmap2.header.type = PERF_RECORD_MMAP2; 1102 event->mmap2.header.size = (sizeof(event->mmap2) - 1103 (sizeof(event->mmap2.filename) - size) + machine->id_hdr_size); 1104 event->mmap2.pgoff = kmap->ref_reloc_sym->addr; 1105 event->mmap2.start = map->start; 1106 event->mmap2.len = map->end - event->mmap.start; 1107 event->mmap2.pid = machine->pid; 1108 1109 perf_record_mmap2__read_build_id(&event->mmap2, true); 1110 } else { 1111 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename), 1112 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1; 1113 size = PERF_ALIGN(size, sizeof(u64)); 1114 event->mmap.header.type = PERF_RECORD_MMAP; 1115 event->mmap.header.size = (sizeof(event->mmap) - 1116 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size); 1117 event->mmap.pgoff = kmap->ref_reloc_sym->addr; 1118 event->mmap.start = map->start; 1119 event->mmap.len = map->end - event->mmap.start; 1120 event->mmap.pid = machine->pid; 1121 } 1122 1123 err = perf_tool__process_synth_event(tool, event, machine, process); 1124 free(event); 1125 1126 return err; 1127 } 1128 1129 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 1130 perf_event__handler_t process, 1131 struct machine *machine) 1132 { 1133 int err; 1134 1135 err = __perf_event__synthesize_kernel_mmap(tool, process, machine); 1136 if (err < 0) 1137 return err; 1138 1139 return perf_event__synthesize_extra_kmaps(tool, process, machine); 1140 } 1141 1142 int perf_event__synthesize_thread_map2(struct perf_tool *tool, 1143 struct perf_thread_map *threads, 1144 perf_event__handler_t process, 1145 struct machine *machine) 1146 { 1147 union perf_event *event; 1148 int i, err, size; 1149 1150 size = sizeof(event->thread_map); 1151 size += threads->nr * sizeof(event->thread_map.entries[0]); 1152 1153 event = zalloc(size); 1154 if (!event) 1155 return -ENOMEM; 1156 1157 event->header.type = PERF_RECORD_THREAD_MAP; 1158 event->header.size = size; 1159 event->thread_map.nr = threads->nr; 1160 1161 for (i = 0; i < threads->nr; i++) { 1162 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i]; 1163 char *comm = perf_thread_map__comm(threads, i); 1164 1165 if (!comm) 1166 comm = (char *) ""; 1167 1168 entry->pid = perf_thread_map__pid(threads, i); 1169 strncpy((char *) &entry->comm, comm, sizeof(entry->comm)); 1170 } 1171 1172 err = process(tool, event, NULL, machine); 1173 1174 free(event); 1175 return err; 1176 } 1177 1178 static void synthesize_cpus(struct cpu_map_entries *cpus, 1179 struct perf_cpu_map *map) 1180 { 1181 int i; 1182 1183 cpus->nr = map->nr; 1184 1185 for (i = 0; i < map->nr; i++) 1186 cpus->cpu[i] = map->map[i]; 1187 } 1188 1189 static void synthesize_mask(struct perf_record_record_cpu_map *mask, 1190 struct perf_cpu_map *map, int max) 1191 { 1192 int i; 1193 1194 mask->nr = BITS_TO_LONGS(max); 1195 mask->long_size = sizeof(long); 1196 1197 for (i = 0; i < map->nr; i++) 1198 set_bit(map->map[i], mask->mask); 1199 } 1200 1201 static size_t cpus_size(struct perf_cpu_map *map) 1202 { 1203 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16); 1204 } 1205 1206 static size_t mask_size(struct perf_cpu_map *map, int *max) 1207 { 1208 int i; 1209 1210 *max = 0; 1211 1212 for (i = 0; i < map->nr; i++) { 1213 /* bit possition of the cpu is + 1 */ 1214 int bit = map->map[i] + 1; 1215 1216 if (bit > *max) 1217 *max = bit; 1218 } 1219 1220 return sizeof(struct perf_record_record_cpu_map) + BITS_TO_LONGS(*max) * sizeof(long); 1221 } 1222 1223 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max) 1224 { 1225 size_t size_cpus, size_mask; 1226 bool is_dummy = perf_cpu_map__empty(map); 1227 1228 /* 1229 * Both array and mask data have variable size based 1230 * on the number of cpus and their actual values. 1231 * The size of the 'struct perf_record_cpu_map_data' is: 1232 * 1233 * array = size of 'struct cpu_map_entries' + 1234 * number of cpus * sizeof(u64) 1235 * 1236 * mask = size of 'struct perf_record_record_cpu_map' + 1237 * maximum cpu bit converted to size of longs 1238 * 1239 * and finaly + the size of 'struct perf_record_cpu_map_data'. 1240 */ 1241 size_cpus = cpus_size(map); 1242 size_mask = mask_size(map, max); 1243 1244 if (is_dummy || (size_cpus < size_mask)) { 1245 *size += size_cpus; 1246 *type = PERF_CPU_MAP__CPUS; 1247 } else { 1248 *size += size_mask; 1249 *type = PERF_CPU_MAP__MASK; 1250 } 1251 1252 *size += sizeof(struct perf_record_cpu_map_data); 1253 *size = PERF_ALIGN(*size, sizeof(u64)); 1254 return zalloc(*size); 1255 } 1256 1257 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map, 1258 u16 type, int max) 1259 { 1260 data->type = type; 1261 1262 switch (type) { 1263 case PERF_CPU_MAP__CPUS: 1264 synthesize_cpus((struct cpu_map_entries *) data->data, map); 1265 break; 1266 case PERF_CPU_MAP__MASK: 1267 synthesize_mask((struct perf_record_record_cpu_map *)data->data, map, max); 1268 default: 1269 break; 1270 } 1271 } 1272 1273 static struct perf_record_cpu_map *cpu_map_event__new(struct perf_cpu_map *map) 1274 { 1275 size_t size = sizeof(struct perf_record_cpu_map); 1276 struct perf_record_cpu_map *event; 1277 int max; 1278 u16 type; 1279 1280 event = cpu_map_data__alloc(map, &size, &type, &max); 1281 if (!event) 1282 return NULL; 1283 1284 event->header.type = PERF_RECORD_CPU_MAP; 1285 event->header.size = size; 1286 event->data.type = type; 1287 1288 cpu_map_data__synthesize(&event->data, map, type, max); 1289 return event; 1290 } 1291 1292 int perf_event__synthesize_cpu_map(struct perf_tool *tool, 1293 struct perf_cpu_map *map, 1294 perf_event__handler_t process, 1295 struct machine *machine) 1296 { 1297 struct perf_record_cpu_map *event; 1298 int err; 1299 1300 event = cpu_map_event__new(map); 1301 if (!event) 1302 return -ENOMEM; 1303 1304 err = process(tool, (union perf_event *) event, NULL, machine); 1305 1306 free(event); 1307 return err; 1308 } 1309 1310 int perf_event__synthesize_stat_config(struct perf_tool *tool, 1311 struct perf_stat_config *config, 1312 perf_event__handler_t process, 1313 struct machine *machine) 1314 { 1315 struct perf_record_stat_config *event; 1316 int size, i = 0, err; 1317 1318 size = sizeof(*event); 1319 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0])); 1320 1321 event = zalloc(size); 1322 if (!event) 1323 return -ENOMEM; 1324 1325 event->header.type = PERF_RECORD_STAT_CONFIG; 1326 event->header.size = size; 1327 event->nr = PERF_STAT_CONFIG_TERM__MAX; 1328 1329 #define ADD(__term, __val) \ 1330 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \ 1331 event->data[i].val = __val; \ 1332 i++; 1333 1334 ADD(AGGR_MODE, config->aggr_mode) 1335 ADD(INTERVAL, config->interval) 1336 ADD(SCALE, config->scale) 1337 1338 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX, 1339 "stat config terms unbalanced\n"); 1340 #undef ADD 1341 1342 err = process(tool, (union perf_event *) event, NULL, machine); 1343 1344 free(event); 1345 return err; 1346 } 1347 1348 int perf_event__synthesize_stat(struct perf_tool *tool, 1349 u32 cpu, u32 thread, u64 id, 1350 struct perf_counts_values *count, 1351 perf_event__handler_t process, 1352 struct machine *machine) 1353 { 1354 struct perf_record_stat event; 1355 1356 event.header.type = PERF_RECORD_STAT; 1357 event.header.size = sizeof(event); 1358 event.header.misc = 0; 1359 1360 event.id = id; 1361 event.cpu = cpu; 1362 event.thread = thread; 1363 event.val = count->val; 1364 event.ena = count->ena; 1365 event.run = count->run; 1366 1367 return process(tool, (union perf_event *) &event, NULL, machine); 1368 } 1369 1370 int perf_event__synthesize_stat_round(struct perf_tool *tool, 1371 u64 evtime, u64 type, 1372 perf_event__handler_t process, 1373 struct machine *machine) 1374 { 1375 struct perf_record_stat_round event; 1376 1377 event.header.type = PERF_RECORD_STAT_ROUND; 1378 event.header.size = sizeof(event); 1379 event.header.misc = 0; 1380 1381 event.time = evtime; 1382 event.type = type; 1383 1384 return process(tool, (union perf_event *) &event, NULL, machine); 1385 } 1386 1387 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format) 1388 { 1389 size_t sz, result = sizeof(struct perf_record_sample); 1390 1391 if (type & PERF_SAMPLE_IDENTIFIER) 1392 result += sizeof(u64); 1393 1394 if (type & PERF_SAMPLE_IP) 1395 result += sizeof(u64); 1396 1397 if (type & PERF_SAMPLE_TID) 1398 result += sizeof(u64); 1399 1400 if (type & PERF_SAMPLE_TIME) 1401 result += sizeof(u64); 1402 1403 if (type & PERF_SAMPLE_ADDR) 1404 result += sizeof(u64); 1405 1406 if (type & PERF_SAMPLE_ID) 1407 result += sizeof(u64); 1408 1409 if (type & PERF_SAMPLE_STREAM_ID) 1410 result += sizeof(u64); 1411 1412 if (type & PERF_SAMPLE_CPU) 1413 result += sizeof(u64); 1414 1415 if (type & PERF_SAMPLE_PERIOD) 1416 result += sizeof(u64); 1417 1418 if (type & PERF_SAMPLE_READ) { 1419 result += sizeof(u64); 1420 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1421 result += sizeof(u64); 1422 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1423 result += sizeof(u64); 1424 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1425 if (read_format & PERF_FORMAT_GROUP) { 1426 sz = sample->read.group.nr * 1427 sizeof(struct sample_read_value); 1428 result += sz; 1429 } else { 1430 result += sizeof(u64); 1431 } 1432 } 1433 1434 if (type & PERF_SAMPLE_CALLCHAIN) { 1435 sz = (sample->callchain->nr + 1) * sizeof(u64); 1436 result += sz; 1437 } 1438 1439 if (type & PERF_SAMPLE_RAW) { 1440 result += sizeof(u32); 1441 result += sample->raw_size; 1442 } 1443 1444 if (type & PERF_SAMPLE_BRANCH_STACK) { 1445 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1446 /* nr, hw_idx */ 1447 sz += 2 * sizeof(u64); 1448 result += sz; 1449 } 1450 1451 if (type & PERF_SAMPLE_REGS_USER) { 1452 if (sample->user_regs.abi) { 1453 result += sizeof(u64); 1454 sz = hweight64(sample->user_regs.mask) * sizeof(u64); 1455 result += sz; 1456 } else { 1457 result += sizeof(u64); 1458 } 1459 } 1460 1461 if (type & PERF_SAMPLE_STACK_USER) { 1462 sz = sample->user_stack.size; 1463 result += sizeof(u64); 1464 if (sz) { 1465 result += sz; 1466 result += sizeof(u64); 1467 } 1468 } 1469 1470 if (type & PERF_SAMPLE_WEIGHT_TYPE) 1471 result += sizeof(u64); 1472 1473 if (type & PERF_SAMPLE_DATA_SRC) 1474 result += sizeof(u64); 1475 1476 if (type & PERF_SAMPLE_TRANSACTION) 1477 result += sizeof(u64); 1478 1479 if (type & PERF_SAMPLE_REGS_INTR) { 1480 if (sample->intr_regs.abi) { 1481 result += sizeof(u64); 1482 sz = hweight64(sample->intr_regs.mask) * sizeof(u64); 1483 result += sz; 1484 } else { 1485 result += sizeof(u64); 1486 } 1487 } 1488 1489 if (type & PERF_SAMPLE_PHYS_ADDR) 1490 result += sizeof(u64); 1491 1492 if (type & PERF_SAMPLE_CGROUP) 1493 result += sizeof(u64); 1494 1495 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) 1496 result += sizeof(u64); 1497 1498 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) 1499 result += sizeof(u64); 1500 1501 if (type & PERF_SAMPLE_AUX) { 1502 result += sizeof(u64); 1503 result += sample->aux_sample.size; 1504 } 1505 1506 return result; 1507 } 1508 1509 void __weak arch_perf_synthesize_sample_weight(const struct perf_sample *data, 1510 __u64 *array, u64 type __maybe_unused) 1511 { 1512 *array = data->weight; 1513 } 1514 1515 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format, 1516 const struct perf_sample *sample) 1517 { 1518 __u64 *array; 1519 size_t sz; 1520 /* 1521 * used for cross-endian analysis. See git commit 65014ab3 1522 * for why this goofiness is needed. 1523 */ 1524 union u64_swap u; 1525 1526 array = event->sample.array; 1527 1528 if (type & PERF_SAMPLE_IDENTIFIER) { 1529 *array = sample->id; 1530 array++; 1531 } 1532 1533 if (type & PERF_SAMPLE_IP) { 1534 *array = sample->ip; 1535 array++; 1536 } 1537 1538 if (type & PERF_SAMPLE_TID) { 1539 u.val32[0] = sample->pid; 1540 u.val32[1] = sample->tid; 1541 *array = u.val64; 1542 array++; 1543 } 1544 1545 if (type & PERF_SAMPLE_TIME) { 1546 *array = sample->time; 1547 array++; 1548 } 1549 1550 if (type & PERF_SAMPLE_ADDR) { 1551 *array = sample->addr; 1552 array++; 1553 } 1554 1555 if (type & PERF_SAMPLE_ID) { 1556 *array = sample->id; 1557 array++; 1558 } 1559 1560 if (type & PERF_SAMPLE_STREAM_ID) { 1561 *array = sample->stream_id; 1562 array++; 1563 } 1564 1565 if (type & PERF_SAMPLE_CPU) { 1566 u.val32[0] = sample->cpu; 1567 u.val32[1] = 0; 1568 *array = u.val64; 1569 array++; 1570 } 1571 1572 if (type & PERF_SAMPLE_PERIOD) { 1573 *array = sample->period; 1574 array++; 1575 } 1576 1577 if (type & PERF_SAMPLE_READ) { 1578 if (read_format & PERF_FORMAT_GROUP) 1579 *array = sample->read.group.nr; 1580 else 1581 *array = sample->read.one.value; 1582 array++; 1583 1584 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 1585 *array = sample->read.time_enabled; 1586 array++; 1587 } 1588 1589 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 1590 *array = sample->read.time_running; 1591 array++; 1592 } 1593 1594 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1595 if (read_format & PERF_FORMAT_GROUP) { 1596 sz = sample->read.group.nr * 1597 sizeof(struct sample_read_value); 1598 memcpy(array, sample->read.group.values, sz); 1599 array = (void *)array + sz; 1600 } else { 1601 *array = sample->read.one.id; 1602 array++; 1603 } 1604 } 1605 1606 if (type & PERF_SAMPLE_CALLCHAIN) { 1607 sz = (sample->callchain->nr + 1) * sizeof(u64); 1608 memcpy(array, sample->callchain, sz); 1609 array = (void *)array + sz; 1610 } 1611 1612 if (type & PERF_SAMPLE_RAW) { 1613 u.val32[0] = sample->raw_size; 1614 *array = u.val64; 1615 array = (void *)array + sizeof(u32); 1616 1617 memcpy(array, sample->raw_data, sample->raw_size); 1618 array = (void *)array + sample->raw_size; 1619 } 1620 1621 if (type & PERF_SAMPLE_BRANCH_STACK) { 1622 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1623 /* nr, hw_idx */ 1624 sz += 2 * sizeof(u64); 1625 memcpy(array, sample->branch_stack, sz); 1626 array = (void *)array + sz; 1627 } 1628 1629 if (type & PERF_SAMPLE_REGS_USER) { 1630 if (sample->user_regs.abi) { 1631 *array++ = sample->user_regs.abi; 1632 sz = hweight64(sample->user_regs.mask) * sizeof(u64); 1633 memcpy(array, sample->user_regs.regs, sz); 1634 array = (void *)array + sz; 1635 } else { 1636 *array++ = 0; 1637 } 1638 } 1639 1640 if (type & PERF_SAMPLE_STACK_USER) { 1641 sz = sample->user_stack.size; 1642 *array++ = sz; 1643 if (sz) { 1644 memcpy(array, sample->user_stack.data, sz); 1645 array = (void *)array + sz; 1646 *array++ = sz; 1647 } 1648 } 1649 1650 if (type & PERF_SAMPLE_WEIGHT_TYPE) { 1651 arch_perf_synthesize_sample_weight(sample, array, type); 1652 array++; 1653 } 1654 1655 if (type & PERF_SAMPLE_DATA_SRC) { 1656 *array = sample->data_src; 1657 array++; 1658 } 1659 1660 if (type & PERF_SAMPLE_TRANSACTION) { 1661 *array = sample->transaction; 1662 array++; 1663 } 1664 1665 if (type & PERF_SAMPLE_REGS_INTR) { 1666 if (sample->intr_regs.abi) { 1667 *array++ = sample->intr_regs.abi; 1668 sz = hweight64(sample->intr_regs.mask) * sizeof(u64); 1669 memcpy(array, sample->intr_regs.regs, sz); 1670 array = (void *)array + sz; 1671 } else { 1672 *array++ = 0; 1673 } 1674 } 1675 1676 if (type & PERF_SAMPLE_PHYS_ADDR) { 1677 *array = sample->phys_addr; 1678 array++; 1679 } 1680 1681 if (type & PERF_SAMPLE_CGROUP) { 1682 *array = sample->cgroup; 1683 array++; 1684 } 1685 1686 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) { 1687 *array = sample->data_page_size; 1688 array++; 1689 } 1690 1691 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) { 1692 *array = sample->code_page_size; 1693 array++; 1694 } 1695 1696 if (type & PERF_SAMPLE_AUX) { 1697 sz = sample->aux_sample.size; 1698 *array++ = sz; 1699 memcpy(array, sample->aux_sample.data, sz); 1700 array = (void *)array + sz; 1701 } 1702 1703 return 0; 1704 } 1705 1706 int perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process, 1707 struct evlist *evlist, struct machine *machine) 1708 { 1709 union perf_event *ev; 1710 struct evsel *evsel; 1711 size_t nr = 0, i = 0, sz, max_nr, n; 1712 int err; 1713 1714 pr_debug2("Synthesizing id index\n"); 1715 1716 max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) / 1717 sizeof(struct id_index_entry); 1718 1719 evlist__for_each_entry(evlist, evsel) 1720 nr += evsel->core.ids; 1721 1722 n = nr > max_nr ? max_nr : nr; 1723 sz = sizeof(struct perf_record_id_index) + n * sizeof(struct id_index_entry); 1724 ev = zalloc(sz); 1725 if (!ev) 1726 return -ENOMEM; 1727 1728 ev->id_index.header.type = PERF_RECORD_ID_INDEX; 1729 ev->id_index.header.size = sz; 1730 ev->id_index.nr = n; 1731 1732 evlist__for_each_entry(evlist, evsel) { 1733 u32 j; 1734 1735 for (j = 0; j < evsel->core.ids; j++) { 1736 struct id_index_entry *e; 1737 struct perf_sample_id *sid; 1738 1739 if (i >= n) { 1740 err = process(tool, ev, NULL, machine); 1741 if (err) 1742 goto out_err; 1743 nr -= n; 1744 i = 0; 1745 } 1746 1747 e = &ev->id_index.entries[i++]; 1748 1749 e->id = evsel->core.id[j]; 1750 1751 sid = evlist__id2sid(evlist, e->id); 1752 if (!sid) { 1753 free(ev); 1754 return -ENOENT; 1755 } 1756 1757 e->idx = sid->idx; 1758 e->cpu = sid->cpu; 1759 e->tid = sid->tid; 1760 } 1761 } 1762 1763 sz = sizeof(struct perf_record_id_index) + nr * sizeof(struct id_index_entry); 1764 ev->id_index.header.size = sz; 1765 ev->id_index.nr = nr; 1766 1767 err = process(tool, ev, NULL, machine); 1768 out_err: 1769 free(ev); 1770 1771 return err; 1772 } 1773 1774 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool, 1775 struct target *target, struct perf_thread_map *threads, 1776 perf_event__handler_t process, bool data_mmap, 1777 unsigned int nr_threads_synthesize) 1778 { 1779 if (target__has_task(target)) 1780 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap); 1781 else if (target__has_cpu(target)) 1782 return perf_event__synthesize_threads(tool, process, 1783 machine, data_mmap, 1784 nr_threads_synthesize); 1785 /* command specified */ 1786 return 0; 1787 } 1788 1789 int machine__synthesize_threads(struct machine *machine, struct target *target, 1790 struct perf_thread_map *threads, bool data_mmap, 1791 unsigned int nr_threads_synthesize) 1792 { 1793 return __machine__synthesize_threads(machine, NULL, target, threads, 1794 perf_event__process, data_mmap, 1795 nr_threads_synthesize); 1796 } 1797 1798 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id) 1799 { 1800 struct perf_record_event_update *ev; 1801 1802 size += sizeof(*ev); 1803 size = PERF_ALIGN(size, sizeof(u64)); 1804 1805 ev = zalloc(size); 1806 if (ev) { 1807 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1808 ev->header.size = (u16)size; 1809 ev->type = type; 1810 ev->id = id; 1811 } 1812 return ev; 1813 } 1814 1815 int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evsel *evsel, 1816 perf_event__handler_t process) 1817 { 1818 size_t size = strlen(evsel->unit); 1819 struct perf_record_event_update *ev; 1820 int err; 1821 1822 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]); 1823 if (ev == NULL) 1824 return -ENOMEM; 1825 1826 strlcpy(ev->data, evsel->unit, size + 1); 1827 err = process(tool, (union perf_event *)ev, NULL, NULL); 1828 free(ev); 1829 return err; 1830 } 1831 1832 int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsel *evsel, 1833 perf_event__handler_t process) 1834 { 1835 struct perf_record_event_update *ev; 1836 struct perf_record_event_update_scale *ev_data; 1837 int err; 1838 1839 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]); 1840 if (ev == NULL) 1841 return -ENOMEM; 1842 1843 ev_data = (struct perf_record_event_update_scale *)ev->data; 1844 ev_data->scale = evsel->scale; 1845 err = process(tool, (union perf_event *)ev, NULL, NULL); 1846 free(ev); 1847 return err; 1848 } 1849 1850 int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evsel *evsel, 1851 perf_event__handler_t process) 1852 { 1853 struct perf_record_event_update *ev; 1854 size_t len = strlen(evsel->name); 1855 int err; 1856 1857 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]); 1858 if (ev == NULL) 1859 return -ENOMEM; 1860 1861 strlcpy(ev->data, evsel->name, len + 1); 1862 err = process(tool, (union perf_event *)ev, NULL, NULL); 1863 free(ev); 1864 return err; 1865 } 1866 1867 int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel, 1868 perf_event__handler_t process) 1869 { 1870 size_t size = sizeof(struct perf_record_event_update); 1871 struct perf_record_event_update *ev; 1872 int max, err; 1873 u16 type; 1874 1875 if (!evsel->core.own_cpus) 1876 return 0; 1877 1878 ev = cpu_map_data__alloc(evsel->core.own_cpus, &size, &type, &max); 1879 if (!ev) 1880 return -ENOMEM; 1881 1882 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1883 ev->header.size = (u16)size; 1884 ev->type = PERF_EVENT_UPDATE__CPUS; 1885 ev->id = evsel->core.id[0]; 1886 1887 cpu_map_data__synthesize((struct perf_record_cpu_map_data *)ev->data, 1888 evsel->core.own_cpus, type, max); 1889 1890 err = process(tool, (union perf_event *)ev, NULL, NULL); 1891 free(ev); 1892 return err; 1893 } 1894 1895 int perf_event__synthesize_attrs(struct perf_tool *tool, struct evlist *evlist, 1896 perf_event__handler_t process) 1897 { 1898 struct evsel *evsel; 1899 int err = 0; 1900 1901 evlist__for_each_entry(evlist, evsel) { 1902 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids, 1903 evsel->core.id, process); 1904 if (err) { 1905 pr_debug("failed to create perf header attribute\n"); 1906 return err; 1907 } 1908 } 1909 1910 return err; 1911 } 1912 1913 static bool has_unit(struct evsel *evsel) 1914 { 1915 return evsel->unit && *evsel->unit; 1916 } 1917 1918 static bool has_scale(struct evsel *evsel) 1919 { 1920 return evsel->scale != 1; 1921 } 1922 1923 int perf_event__synthesize_extra_attr(struct perf_tool *tool, struct evlist *evsel_list, 1924 perf_event__handler_t process, bool is_pipe) 1925 { 1926 struct evsel *evsel; 1927 int err; 1928 1929 /* 1930 * Synthesize other events stuff not carried within 1931 * attr event - unit, scale, name 1932 */ 1933 evlist__for_each_entry(evsel_list, evsel) { 1934 if (!evsel->supported) 1935 continue; 1936 1937 /* 1938 * Synthesize unit and scale only if it's defined. 1939 */ 1940 if (has_unit(evsel)) { 1941 err = perf_event__synthesize_event_update_unit(tool, evsel, process); 1942 if (err < 0) { 1943 pr_err("Couldn't synthesize evsel unit.\n"); 1944 return err; 1945 } 1946 } 1947 1948 if (has_scale(evsel)) { 1949 err = perf_event__synthesize_event_update_scale(tool, evsel, process); 1950 if (err < 0) { 1951 pr_err("Couldn't synthesize evsel evsel.\n"); 1952 return err; 1953 } 1954 } 1955 1956 if (evsel->core.own_cpus) { 1957 err = perf_event__synthesize_event_update_cpus(tool, evsel, process); 1958 if (err < 0) { 1959 pr_err("Couldn't synthesize evsel cpus.\n"); 1960 return err; 1961 } 1962 } 1963 1964 /* 1965 * Name is needed only for pipe output, 1966 * perf.data carries event names. 1967 */ 1968 if (is_pipe) { 1969 err = perf_event__synthesize_event_update_name(tool, evsel, process); 1970 if (err < 0) { 1971 pr_err("Couldn't synthesize evsel name.\n"); 1972 return err; 1973 } 1974 } 1975 } 1976 return 0; 1977 } 1978 1979 int perf_event__synthesize_attr(struct perf_tool *tool, struct perf_event_attr *attr, 1980 u32 ids, u64 *id, perf_event__handler_t process) 1981 { 1982 union perf_event *ev; 1983 size_t size; 1984 int err; 1985 1986 size = sizeof(struct perf_event_attr); 1987 size = PERF_ALIGN(size, sizeof(u64)); 1988 size += sizeof(struct perf_event_header); 1989 size += ids * sizeof(u64); 1990 1991 ev = zalloc(size); 1992 1993 if (ev == NULL) 1994 return -ENOMEM; 1995 1996 ev->attr.attr = *attr; 1997 memcpy(ev->attr.id, id, ids * sizeof(u64)); 1998 1999 ev->attr.header.type = PERF_RECORD_HEADER_ATTR; 2000 ev->attr.header.size = (u16)size; 2001 2002 if (ev->attr.header.size == size) 2003 err = process(tool, ev, NULL, NULL); 2004 else 2005 err = -E2BIG; 2006 2007 free(ev); 2008 2009 return err; 2010 } 2011 2012 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, struct evlist *evlist, 2013 perf_event__handler_t process) 2014 { 2015 union perf_event ev; 2016 struct tracing_data *tdata; 2017 ssize_t size = 0, aligned_size = 0, padding; 2018 struct feat_fd ff; 2019 2020 /* 2021 * We are going to store the size of the data followed 2022 * by the data contents. Since the fd descriptor is a pipe, 2023 * we cannot seek back to store the size of the data once 2024 * we know it. Instead we: 2025 * 2026 * - write the tracing data to the temp file 2027 * - get/write the data size to pipe 2028 * - write the tracing data from the temp file 2029 * to the pipe 2030 */ 2031 tdata = tracing_data_get(&evlist->core.entries, fd, true); 2032 if (!tdata) 2033 return -1; 2034 2035 memset(&ev, 0, sizeof(ev)); 2036 2037 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA; 2038 size = tdata->size; 2039 aligned_size = PERF_ALIGN(size, sizeof(u64)); 2040 padding = aligned_size - size; 2041 ev.tracing_data.header.size = sizeof(ev.tracing_data); 2042 ev.tracing_data.size = aligned_size; 2043 2044 process(tool, &ev, NULL, NULL); 2045 2046 /* 2047 * The put function will copy all the tracing data 2048 * stored in temp file to the pipe. 2049 */ 2050 tracing_data_put(tdata); 2051 2052 ff = (struct feat_fd){ .fd = fd }; 2053 if (write_padded(&ff, NULL, 0, padding)) 2054 return -1; 2055 2056 return aligned_size; 2057 } 2058 2059 int perf_event__synthesize_build_id(struct perf_tool *tool, struct dso *pos, u16 misc, 2060 perf_event__handler_t process, struct machine *machine) 2061 { 2062 union perf_event ev; 2063 size_t len; 2064 2065 if (!pos->hit) 2066 return 0; 2067 2068 memset(&ev, 0, sizeof(ev)); 2069 2070 len = pos->long_name_len + 1; 2071 len = PERF_ALIGN(len, NAME_ALIGN); 2072 memcpy(&ev.build_id.build_id, pos->bid.data, sizeof(pos->bid.data)); 2073 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID; 2074 ev.build_id.header.misc = misc; 2075 ev.build_id.pid = machine->pid; 2076 ev.build_id.header.size = sizeof(ev.build_id) + len; 2077 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len); 2078 2079 return process(tool, &ev, NULL, machine); 2080 } 2081 2082 int perf_event__synthesize_stat_events(struct perf_stat_config *config, struct perf_tool *tool, 2083 struct evlist *evlist, perf_event__handler_t process, bool attrs) 2084 { 2085 int err; 2086 2087 if (attrs) { 2088 err = perf_event__synthesize_attrs(tool, evlist, process); 2089 if (err < 0) { 2090 pr_err("Couldn't synthesize attrs.\n"); 2091 return err; 2092 } 2093 } 2094 2095 err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs); 2096 err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL); 2097 if (err < 0) { 2098 pr_err("Couldn't synthesize thread map.\n"); 2099 return err; 2100 } 2101 2102 err = perf_event__synthesize_cpu_map(tool, evlist->core.cpus, process, NULL); 2103 if (err < 0) { 2104 pr_err("Couldn't synthesize thread map.\n"); 2105 return err; 2106 } 2107 2108 err = perf_event__synthesize_stat_config(tool, config, process, NULL); 2109 if (err < 0) { 2110 pr_err("Couldn't synthesize config.\n"); 2111 return err; 2112 } 2113 2114 return 0; 2115 } 2116 2117 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE]; 2118 2119 int perf_event__synthesize_features(struct perf_tool *tool, struct perf_session *session, 2120 struct evlist *evlist, perf_event__handler_t process) 2121 { 2122 struct perf_header *header = &session->header; 2123 struct perf_record_header_feature *fe; 2124 struct feat_fd ff; 2125 size_t sz, sz_hdr; 2126 int feat, ret; 2127 2128 sz_hdr = sizeof(fe->header); 2129 sz = sizeof(union perf_event); 2130 /* get a nice alignment */ 2131 sz = PERF_ALIGN(sz, page_size); 2132 2133 memset(&ff, 0, sizeof(ff)); 2134 2135 ff.buf = malloc(sz); 2136 if (!ff.buf) 2137 return -ENOMEM; 2138 2139 ff.size = sz - sz_hdr; 2140 ff.ph = &session->header; 2141 2142 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) { 2143 if (!feat_ops[feat].synthesize) { 2144 pr_debug("No record header feature for header :%d\n", feat); 2145 continue; 2146 } 2147 2148 ff.offset = sizeof(*fe); 2149 2150 ret = feat_ops[feat].write(&ff, evlist); 2151 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) { 2152 pr_debug("Error writing feature\n"); 2153 continue; 2154 } 2155 /* ff.buf may have changed due to realloc in do_write() */ 2156 fe = ff.buf; 2157 memset(fe, 0, sizeof(*fe)); 2158 2159 fe->feat_id = feat; 2160 fe->header.type = PERF_RECORD_HEADER_FEATURE; 2161 fe->header.size = ff.offset; 2162 2163 ret = process(tool, ff.buf, NULL, NULL); 2164 if (ret) { 2165 free(ff.buf); 2166 return ret; 2167 } 2168 } 2169 2170 /* Send HEADER_LAST_FEATURE mark. */ 2171 fe = ff.buf; 2172 fe->feat_id = HEADER_LAST_FEATURE; 2173 fe->header.type = PERF_RECORD_HEADER_FEATURE; 2174 fe->header.size = sizeof(*fe); 2175 2176 ret = process(tool, ff.buf, NULL, NULL); 2177 2178 free(ff.buf); 2179 return ret; 2180 } 2181