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, aligned_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 aligned_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) - aligned_size)); 491 memset(event->mmap2.filename + size, 0, machine->id_hdr_size + 492 (aligned_size - size)); 493 event->mmap2.header.size += machine->id_hdr_size; 494 event->mmap2.pid = tgid; 495 event->mmap2.tid = pid; 496 497 if (symbol_conf.buildid_mmap2) 498 perf_record_mmap2__read_build_id(&event->mmap2, false); 499 500 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 501 rc = -1; 502 break; 503 } 504 505 if (truncation) 506 break; 507 } 508 509 close(io.fd); 510 return rc; 511 } 512 513 #ifdef HAVE_FILE_HANDLE 514 static int perf_event__synthesize_cgroup(struct perf_tool *tool, 515 union perf_event *event, 516 char *path, size_t mount_len, 517 perf_event__handler_t process, 518 struct machine *machine) 519 { 520 size_t event_size = sizeof(event->cgroup) - sizeof(event->cgroup.path); 521 size_t path_len = strlen(path) - mount_len + 1; 522 struct { 523 struct file_handle fh; 524 uint64_t cgroup_id; 525 } handle; 526 int mount_id; 527 528 while (path_len % sizeof(u64)) 529 path[mount_len + path_len++] = '\0'; 530 531 memset(&event->cgroup, 0, event_size); 532 533 event->cgroup.header.type = PERF_RECORD_CGROUP; 534 event->cgroup.header.size = event_size + path_len + machine->id_hdr_size; 535 536 handle.fh.handle_bytes = sizeof(handle.cgroup_id); 537 if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0) { 538 pr_debug("stat failed: %s\n", path); 539 return -1; 540 } 541 542 event->cgroup.id = handle.cgroup_id; 543 strncpy(event->cgroup.path, path + mount_len, path_len); 544 memset(event->cgroup.path + path_len, 0, machine->id_hdr_size); 545 546 if (perf_tool__process_synth_event(tool, event, machine, process) < 0) { 547 pr_debug("process synth event failed\n"); 548 return -1; 549 } 550 551 return 0; 552 } 553 554 static int perf_event__walk_cgroup_tree(struct perf_tool *tool, 555 union perf_event *event, 556 char *path, size_t mount_len, 557 perf_event__handler_t process, 558 struct machine *machine) 559 { 560 size_t pos = strlen(path); 561 DIR *d; 562 struct dirent *dent; 563 int ret = 0; 564 565 if (perf_event__synthesize_cgroup(tool, event, path, mount_len, 566 process, machine) < 0) 567 return -1; 568 569 d = opendir(path); 570 if (d == NULL) { 571 pr_debug("failed to open directory: %s\n", path); 572 return -1; 573 } 574 575 while ((dent = readdir(d)) != NULL) { 576 if (dent->d_type != DT_DIR) 577 continue; 578 if (!strcmp(dent->d_name, ".") || 579 !strcmp(dent->d_name, "..")) 580 continue; 581 582 /* any sane path should be less than PATH_MAX */ 583 if (strlen(path) + strlen(dent->d_name) + 1 >= PATH_MAX) 584 continue; 585 586 if (path[pos - 1] != '/') 587 strcat(path, "/"); 588 strcat(path, dent->d_name); 589 590 ret = perf_event__walk_cgroup_tree(tool, event, path, 591 mount_len, process, machine); 592 if (ret < 0) 593 break; 594 595 path[pos] = '\0'; 596 } 597 598 closedir(d); 599 return ret; 600 } 601 602 int perf_event__synthesize_cgroups(struct perf_tool *tool, 603 perf_event__handler_t process, 604 struct machine *machine) 605 { 606 union perf_event event; 607 char cgrp_root[PATH_MAX]; 608 size_t mount_len; /* length of mount point in the path */ 609 610 if (!tool || !tool->cgroup_events) 611 return 0; 612 613 if (cgroupfs_find_mountpoint(cgrp_root, PATH_MAX, "perf_event") < 0) { 614 pr_debug("cannot find cgroup mount point\n"); 615 return -1; 616 } 617 618 mount_len = strlen(cgrp_root); 619 /* make sure the path starts with a slash (after mount point) */ 620 strcat(cgrp_root, "/"); 621 622 if (perf_event__walk_cgroup_tree(tool, &event, cgrp_root, mount_len, 623 process, machine) < 0) 624 return -1; 625 626 return 0; 627 } 628 #else 629 int perf_event__synthesize_cgroups(struct perf_tool *tool __maybe_unused, 630 perf_event__handler_t process __maybe_unused, 631 struct machine *machine __maybe_unused) 632 { 633 return -1; 634 } 635 #endif 636 637 int perf_event__synthesize_modules(struct perf_tool *tool, perf_event__handler_t process, 638 struct machine *machine) 639 { 640 int rc = 0; 641 struct map *pos; 642 struct maps *maps = machine__kernel_maps(machine); 643 union perf_event *event; 644 size_t size = symbol_conf.buildid_mmap2 ? 645 sizeof(event->mmap2) : sizeof(event->mmap); 646 647 event = zalloc(size + machine->id_hdr_size); 648 if (event == NULL) { 649 pr_debug("Not enough memory synthesizing mmap event " 650 "for kernel modules\n"); 651 return -1; 652 } 653 654 /* 655 * kernel uses 0 for user space maps, see kernel/perf_event.c 656 * __perf_event_mmap 657 */ 658 if (machine__is_host(machine)) 659 event->header.misc = PERF_RECORD_MISC_KERNEL; 660 else 661 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 662 663 maps__for_each_entry(maps, pos) { 664 if (!__map__is_kmodule(pos)) 665 continue; 666 667 if (symbol_conf.buildid_mmap2) { 668 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64)); 669 event->mmap2.header.type = PERF_RECORD_MMAP2; 670 event->mmap2.header.size = (sizeof(event->mmap2) - 671 (sizeof(event->mmap2.filename) - size)); 672 memset(event->mmap2.filename + size, 0, machine->id_hdr_size); 673 event->mmap2.header.size += machine->id_hdr_size; 674 event->mmap2.start = pos->start; 675 event->mmap2.len = pos->end - pos->start; 676 event->mmap2.pid = machine->pid; 677 678 memcpy(event->mmap2.filename, pos->dso->long_name, 679 pos->dso->long_name_len + 1); 680 681 perf_record_mmap2__read_build_id(&event->mmap2, false); 682 } else { 683 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64)); 684 event->mmap.header.type = PERF_RECORD_MMAP; 685 event->mmap.header.size = (sizeof(event->mmap) - 686 (sizeof(event->mmap.filename) - size)); 687 memset(event->mmap.filename + size, 0, machine->id_hdr_size); 688 event->mmap.header.size += machine->id_hdr_size; 689 event->mmap.start = pos->start; 690 event->mmap.len = pos->end - pos->start; 691 event->mmap.pid = machine->pid; 692 693 memcpy(event->mmap.filename, pos->dso->long_name, 694 pos->dso->long_name_len + 1); 695 } 696 697 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 698 rc = -1; 699 break; 700 } 701 } 702 703 free(event); 704 return rc; 705 } 706 707 static int filter_task(const struct dirent *dirent) 708 { 709 return isdigit(dirent->d_name[0]); 710 } 711 712 static int __event__synthesize_thread(union perf_event *comm_event, 713 union perf_event *mmap_event, 714 union perf_event *fork_event, 715 union perf_event *namespaces_event, 716 pid_t pid, int full, perf_event__handler_t process, 717 struct perf_tool *tool, struct machine *machine, bool mmap_data) 718 { 719 char filename[PATH_MAX]; 720 struct dirent **dirent; 721 pid_t tgid, ppid; 722 int rc = 0; 723 int i, n; 724 725 /* special case: only send one comm event using passed in pid */ 726 if (!full) { 727 tgid = perf_event__synthesize_comm(tool, comm_event, pid, 728 process, machine); 729 730 if (tgid == -1) 731 return -1; 732 733 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid, 734 tgid, process, machine) < 0) 735 return -1; 736 737 /* 738 * send mmap only for thread group leader 739 * see thread__init_maps() 740 */ 741 if (pid == tgid && 742 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 743 process, machine, mmap_data)) 744 return -1; 745 746 return 0; 747 } 748 749 if (machine__is_default_guest(machine)) 750 return 0; 751 752 snprintf(filename, sizeof(filename), "%s/proc/%d/task", 753 machine->root_dir, pid); 754 755 n = scandir(filename, &dirent, filter_task, alphasort); 756 if (n < 0) 757 return n; 758 759 for (i = 0; i < n; i++) { 760 char *end; 761 pid_t _pid; 762 bool kernel_thread = false; 763 764 _pid = strtol(dirent[i]->d_name, &end, 10); 765 if (*end) 766 continue; 767 768 rc = -1; 769 if (perf_event__prepare_comm(comm_event, pid, _pid, machine, 770 &tgid, &ppid, &kernel_thread) != 0) 771 break; 772 773 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid, 774 ppid, process, machine) < 0) 775 break; 776 777 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid, 778 tgid, process, machine) < 0) 779 break; 780 781 /* 782 * Send the prepared comm event 783 */ 784 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0) 785 break; 786 787 rc = 0; 788 if (_pid == pid && !kernel_thread) { 789 /* process the parent's maps too */ 790 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 791 process, machine, mmap_data); 792 if (rc) 793 break; 794 } 795 } 796 797 for (i = 0; i < n; i++) 798 zfree(&dirent[i]); 799 free(dirent); 800 801 return rc; 802 } 803 804 int perf_event__synthesize_thread_map(struct perf_tool *tool, 805 struct perf_thread_map *threads, 806 perf_event__handler_t process, 807 struct machine *machine, 808 bool mmap_data) 809 { 810 union perf_event *comm_event, *mmap_event, *fork_event; 811 union perf_event *namespaces_event; 812 int err = -1, thread, j; 813 814 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 815 if (comm_event == NULL) 816 goto out; 817 818 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 819 if (mmap_event == NULL) 820 goto out_free_comm; 821 822 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 823 if (fork_event == NULL) 824 goto out_free_mmap; 825 826 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 827 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 828 machine->id_hdr_size); 829 if (namespaces_event == NULL) 830 goto out_free_fork; 831 832 err = 0; 833 for (thread = 0; thread < threads->nr; ++thread) { 834 if (__event__synthesize_thread(comm_event, mmap_event, 835 fork_event, namespaces_event, 836 perf_thread_map__pid(threads, thread), 0, 837 process, tool, machine, 838 mmap_data)) { 839 err = -1; 840 break; 841 } 842 843 /* 844 * comm.pid is set to thread group id by 845 * perf_event__synthesize_comm 846 */ 847 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) { 848 bool need_leader = true; 849 850 /* is thread group leader in thread_map? */ 851 for (j = 0; j < threads->nr; ++j) { 852 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) { 853 need_leader = false; 854 break; 855 } 856 } 857 858 /* if not, generate events for it */ 859 if (need_leader && 860 __event__synthesize_thread(comm_event, mmap_event, 861 fork_event, namespaces_event, 862 comm_event->comm.pid, 0, 863 process, tool, machine, 864 mmap_data)) { 865 err = -1; 866 break; 867 } 868 } 869 } 870 free(namespaces_event); 871 out_free_fork: 872 free(fork_event); 873 out_free_mmap: 874 free(mmap_event); 875 out_free_comm: 876 free(comm_event); 877 out: 878 return err; 879 } 880 881 static int __perf_event__synthesize_threads(struct perf_tool *tool, 882 perf_event__handler_t process, 883 struct machine *machine, 884 bool mmap_data, 885 struct dirent **dirent, 886 int start, 887 int num) 888 { 889 union perf_event *comm_event, *mmap_event, *fork_event; 890 union perf_event *namespaces_event; 891 int err = -1; 892 char *end; 893 pid_t pid; 894 int i; 895 896 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 897 if (comm_event == NULL) 898 goto out; 899 900 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 901 if (mmap_event == NULL) 902 goto out_free_comm; 903 904 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 905 if (fork_event == NULL) 906 goto out_free_mmap; 907 908 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 909 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 910 machine->id_hdr_size); 911 if (namespaces_event == NULL) 912 goto out_free_fork; 913 914 for (i = start; i < start + num; i++) { 915 if (!isdigit(dirent[i]->d_name[0])) 916 continue; 917 918 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10); 919 /* only interested in proper numerical dirents */ 920 if (*end) 921 continue; 922 /* 923 * We may race with exiting thread, so don't stop just because 924 * one thread couldn't be synthesized. 925 */ 926 __event__synthesize_thread(comm_event, mmap_event, fork_event, 927 namespaces_event, pid, 1, process, 928 tool, machine, mmap_data); 929 } 930 err = 0; 931 932 free(namespaces_event); 933 out_free_fork: 934 free(fork_event); 935 out_free_mmap: 936 free(mmap_event); 937 out_free_comm: 938 free(comm_event); 939 out: 940 return err; 941 } 942 943 struct synthesize_threads_arg { 944 struct perf_tool *tool; 945 perf_event__handler_t process; 946 struct machine *machine; 947 bool mmap_data; 948 struct dirent **dirent; 949 int num; 950 int start; 951 }; 952 953 static void *synthesize_threads_worker(void *arg) 954 { 955 struct synthesize_threads_arg *args = arg; 956 957 __perf_event__synthesize_threads(args->tool, args->process, 958 args->machine, args->mmap_data, 959 args->dirent, 960 args->start, args->num); 961 return NULL; 962 } 963 964 int perf_event__synthesize_threads(struct perf_tool *tool, 965 perf_event__handler_t process, 966 struct machine *machine, 967 bool mmap_data, 968 unsigned int nr_threads_synthesize) 969 { 970 struct synthesize_threads_arg *args = NULL; 971 pthread_t *synthesize_threads = NULL; 972 char proc_path[PATH_MAX]; 973 struct dirent **dirent; 974 int num_per_thread; 975 int m, n, i, j; 976 int thread_nr; 977 int base = 0; 978 int err = -1; 979 980 981 if (machine__is_default_guest(machine)) 982 return 0; 983 984 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir); 985 n = scandir(proc_path, &dirent, filter_task, alphasort); 986 if (n < 0) 987 return err; 988 989 if (nr_threads_synthesize == UINT_MAX) 990 thread_nr = sysconf(_SC_NPROCESSORS_ONLN); 991 else 992 thread_nr = nr_threads_synthesize; 993 994 if (thread_nr <= 1) { 995 err = __perf_event__synthesize_threads(tool, process, 996 machine, mmap_data, 997 dirent, base, n); 998 goto free_dirent; 999 } 1000 if (thread_nr > n) 1001 thread_nr = n; 1002 1003 synthesize_threads = calloc(sizeof(pthread_t), thread_nr); 1004 if (synthesize_threads == NULL) 1005 goto free_dirent; 1006 1007 args = calloc(sizeof(*args), thread_nr); 1008 if (args == NULL) 1009 goto free_threads; 1010 1011 num_per_thread = n / thread_nr; 1012 m = n % thread_nr; 1013 for (i = 0; i < thread_nr; i++) { 1014 args[i].tool = tool; 1015 args[i].process = process; 1016 args[i].machine = machine; 1017 args[i].mmap_data = mmap_data; 1018 args[i].dirent = dirent; 1019 } 1020 for (i = 0; i < m; i++) { 1021 args[i].num = num_per_thread + 1; 1022 args[i].start = i * args[i].num; 1023 } 1024 if (i != 0) 1025 base = args[i-1].start + args[i-1].num; 1026 for (j = i; j < thread_nr; j++) { 1027 args[j].num = num_per_thread; 1028 args[j].start = base + (j - i) * args[i].num; 1029 } 1030 1031 for (i = 0; i < thread_nr; i++) { 1032 if (pthread_create(&synthesize_threads[i], NULL, 1033 synthesize_threads_worker, &args[i])) 1034 goto out_join; 1035 } 1036 err = 0; 1037 out_join: 1038 for (i = 0; i < thread_nr; i++) 1039 pthread_join(synthesize_threads[i], NULL); 1040 free(args); 1041 free_threads: 1042 free(synthesize_threads); 1043 free_dirent: 1044 for (i = 0; i < n; i++) 1045 zfree(&dirent[i]); 1046 free(dirent); 1047 1048 return err; 1049 } 1050 1051 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused, 1052 perf_event__handler_t process __maybe_unused, 1053 struct machine *machine __maybe_unused) 1054 { 1055 return 0; 1056 } 1057 1058 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 1059 perf_event__handler_t process, 1060 struct machine *machine) 1061 { 1062 union perf_event *event; 1063 size_t size = symbol_conf.buildid_mmap2 ? 1064 sizeof(event->mmap2) : sizeof(event->mmap); 1065 struct map *map = machine__kernel_map(machine); 1066 struct kmap *kmap; 1067 int err; 1068 1069 if (map == NULL) 1070 return -1; 1071 1072 kmap = map__kmap(map); 1073 if (!kmap->ref_reloc_sym) 1074 return -1; 1075 1076 /* 1077 * We should get this from /sys/kernel/sections/.text, but till that is 1078 * available use this, and after it is use this as a fallback for older 1079 * kernels. 1080 */ 1081 event = zalloc(size + machine->id_hdr_size); 1082 if (event == NULL) { 1083 pr_debug("Not enough memory synthesizing mmap event " 1084 "for kernel modules\n"); 1085 return -1; 1086 } 1087 1088 if (machine__is_host(machine)) { 1089 /* 1090 * kernel uses PERF_RECORD_MISC_USER for user space maps, 1091 * see kernel/perf_event.c __perf_event_mmap 1092 */ 1093 event->header.misc = PERF_RECORD_MISC_KERNEL; 1094 } else { 1095 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 1096 } 1097 1098 if (symbol_conf.buildid_mmap2) { 1099 size = snprintf(event->mmap2.filename, sizeof(event->mmap2.filename), 1100 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1; 1101 size = PERF_ALIGN(size, sizeof(u64)); 1102 event->mmap2.header.type = PERF_RECORD_MMAP2; 1103 event->mmap2.header.size = (sizeof(event->mmap2) - 1104 (sizeof(event->mmap2.filename) - size) + machine->id_hdr_size); 1105 event->mmap2.pgoff = kmap->ref_reloc_sym->addr; 1106 event->mmap2.start = map->start; 1107 event->mmap2.len = map->end - event->mmap.start; 1108 event->mmap2.pid = machine->pid; 1109 1110 perf_record_mmap2__read_build_id(&event->mmap2, true); 1111 } else { 1112 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename), 1113 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1; 1114 size = PERF_ALIGN(size, sizeof(u64)); 1115 event->mmap.header.type = PERF_RECORD_MMAP; 1116 event->mmap.header.size = (sizeof(event->mmap) - 1117 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size); 1118 event->mmap.pgoff = kmap->ref_reloc_sym->addr; 1119 event->mmap.start = map->start; 1120 event->mmap.len = map->end - event->mmap.start; 1121 event->mmap.pid = machine->pid; 1122 } 1123 1124 err = perf_tool__process_synth_event(tool, event, machine, process); 1125 free(event); 1126 1127 return err; 1128 } 1129 1130 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 1131 perf_event__handler_t process, 1132 struct machine *machine) 1133 { 1134 int err; 1135 1136 err = __perf_event__synthesize_kernel_mmap(tool, process, machine); 1137 if (err < 0) 1138 return err; 1139 1140 return perf_event__synthesize_extra_kmaps(tool, process, machine); 1141 } 1142 1143 int perf_event__synthesize_thread_map2(struct perf_tool *tool, 1144 struct perf_thread_map *threads, 1145 perf_event__handler_t process, 1146 struct machine *machine) 1147 { 1148 union perf_event *event; 1149 int i, err, size; 1150 1151 size = sizeof(event->thread_map); 1152 size += threads->nr * sizeof(event->thread_map.entries[0]); 1153 1154 event = zalloc(size); 1155 if (!event) 1156 return -ENOMEM; 1157 1158 event->header.type = PERF_RECORD_THREAD_MAP; 1159 event->header.size = size; 1160 event->thread_map.nr = threads->nr; 1161 1162 for (i = 0; i < threads->nr; i++) { 1163 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i]; 1164 char *comm = perf_thread_map__comm(threads, i); 1165 1166 if (!comm) 1167 comm = (char *) ""; 1168 1169 entry->pid = perf_thread_map__pid(threads, i); 1170 strncpy((char *) &entry->comm, comm, sizeof(entry->comm)); 1171 } 1172 1173 err = process(tool, event, NULL, machine); 1174 1175 free(event); 1176 return err; 1177 } 1178 1179 static void synthesize_cpus(struct cpu_map_entries *cpus, 1180 struct perf_cpu_map *map) 1181 { 1182 int i; 1183 1184 cpus->nr = map->nr; 1185 1186 for (i = 0; i < map->nr; i++) 1187 cpus->cpu[i] = map->map[i]; 1188 } 1189 1190 static void synthesize_mask(struct perf_record_record_cpu_map *mask, 1191 struct perf_cpu_map *map, int max) 1192 { 1193 int i; 1194 1195 mask->nr = BITS_TO_LONGS(max); 1196 mask->long_size = sizeof(long); 1197 1198 for (i = 0; i < map->nr; i++) 1199 set_bit(map->map[i], mask->mask); 1200 } 1201 1202 static size_t cpus_size(struct perf_cpu_map *map) 1203 { 1204 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16); 1205 } 1206 1207 static size_t mask_size(struct perf_cpu_map *map, int *max) 1208 { 1209 int i; 1210 1211 *max = 0; 1212 1213 for (i = 0; i < map->nr; i++) { 1214 /* bit position of the cpu is + 1 */ 1215 int bit = map->map[i] + 1; 1216 1217 if (bit > *max) 1218 *max = bit; 1219 } 1220 1221 return sizeof(struct perf_record_record_cpu_map) + BITS_TO_LONGS(*max) * sizeof(long); 1222 } 1223 1224 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max) 1225 { 1226 size_t size_cpus, size_mask; 1227 bool is_dummy = perf_cpu_map__empty(map); 1228 1229 /* 1230 * Both array and mask data have variable size based 1231 * on the number of cpus and their actual values. 1232 * The size of the 'struct perf_record_cpu_map_data' is: 1233 * 1234 * array = size of 'struct cpu_map_entries' + 1235 * number of cpus * sizeof(u64) 1236 * 1237 * mask = size of 'struct perf_record_record_cpu_map' + 1238 * maximum cpu bit converted to size of longs 1239 * 1240 * and finally + the size of 'struct perf_record_cpu_map_data'. 1241 */ 1242 size_cpus = cpus_size(map); 1243 size_mask = mask_size(map, max); 1244 1245 if (is_dummy || (size_cpus < size_mask)) { 1246 *size += size_cpus; 1247 *type = PERF_CPU_MAP__CPUS; 1248 } else { 1249 *size += size_mask; 1250 *type = PERF_CPU_MAP__MASK; 1251 } 1252 1253 *size += sizeof(struct perf_record_cpu_map_data); 1254 *size = PERF_ALIGN(*size, sizeof(u64)); 1255 return zalloc(*size); 1256 } 1257 1258 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map, 1259 u16 type, int max) 1260 { 1261 data->type = type; 1262 1263 switch (type) { 1264 case PERF_CPU_MAP__CPUS: 1265 synthesize_cpus((struct cpu_map_entries *) data->data, map); 1266 break; 1267 case PERF_CPU_MAP__MASK: 1268 synthesize_mask((struct perf_record_record_cpu_map *)data->data, map, max); 1269 default: 1270 break; 1271 } 1272 } 1273 1274 static struct perf_record_cpu_map *cpu_map_event__new(struct perf_cpu_map *map) 1275 { 1276 size_t size = sizeof(struct perf_record_cpu_map); 1277 struct perf_record_cpu_map *event; 1278 int max; 1279 u16 type; 1280 1281 event = cpu_map_data__alloc(map, &size, &type, &max); 1282 if (!event) 1283 return NULL; 1284 1285 event->header.type = PERF_RECORD_CPU_MAP; 1286 event->header.size = size; 1287 event->data.type = type; 1288 1289 cpu_map_data__synthesize(&event->data, map, type, max); 1290 return event; 1291 } 1292 1293 int perf_event__synthesize_cpu_map(struct perf_tool *tool, 1294 struct perf_cpu_map *map, 1295 perf_event__handler_t process, 1296 struct machine *machine) 1297 { 1298 struct perf_record_cpu_map *event; 1299 int err; 1300 1301 event = cpu_map_event__new(map); 1302 if (!event) 1303 return -ENOMEM; 1304 1305 err = process(tool, (union perf_event *) event, NULL, machine); 1306 1307 free(event); 1308 return err; 1309 } 1310 1311 int perf_event__synthesize_stat_config(struct perf_tool *tool, 1312 struct perf_stat_config *config, 1313 perf_event__handler_t process, 1314 struct machine *machine) 1315 { 1316 struct perf_record_stat_config *event; 1317 int size, i = 0, err; 1318 1319 size = sizeof(*event); 1320 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0])); 1321 1322 event = zalloc(size); 1323 if (!event) 1324 return -ENOMEM; 1325 1326 event->header.type = PERF_RECORD_STAT_CONFIG; 1327 event->header.size = size; 1328 event->nr = PERF_STAT_CONFIG_TERM__MAX; 1329 1330 #define ADD(__term, __val) \ 1331 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \ 1332 event->data[i].val = __val; \ 1333 i++; 1334 1335 ADD(AGGR_MODE, config->aggr_mode) 1336 ADD(INTERVAL, config->interval) 1337 ADD(SCALE, config->scale) 1338 1339 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX, 1340 "stat config terms unbalanced\n"); 1341 #undef ADD 1342 1343 err = process(tool, (union perf_event *) event, NULL, machine); 1344 1345 free(event); 1346 return err; 1347 } 1348 1349 int perf_event__synthesize_stat(struct perf_tool *tool, 1350 u32 cpu, u32 thread, u64 id, 1351 struct perf_counts_values *count, 1352 perf_event__handler_t process, 1353 struct machine *machine) 1354 { 1355 struct perf_record_stat event; 1356 1357 event.header.type = PERF_RECORD_STAT; 1358 event.header.size = sizeof(event); 1359 event.header.misc = 0; 1360 1361 event.id = id; 1362 event.cpu = cpu; 1363 event.thread = thread; 1364 event.val = count->val; 1365 event.ena = count->ena; 1366 event.run = count->run; 1367 1368 return process(tool, (union perf_event *) &event, NULL, machine); 1369 } 1370 1371 int perf_event__synthesize_stat_round(struct perf_tool *tool, 1372 u64 evtime, u64 type, 1373 perf_event__handler_t process, 1374 struct machine *machine) 1375 { 1376 struct perf_record_stat_round event; 1377 1378 event.header.type = PERF_RECORD_STAT_ROUND; 1379 event.header.size = sizeof(event); 1380 event.header.misc = 0; 1381 1382 event.time = evtime; 1383 event.type = type; 1384 1385 return process(tool, (union perf_event *) &event, NULL, machine); 1386 } 1387 1388 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format) 1389 { 1390 size_t sz, result = sizeof(struct perf_record_sample); 1391 1392 if (type & PERF_SAMPLE_IDENTIFIER) 1393 result += sizeof(u64); 1394 1395 if (type & PERF_SAMPLE_IP) 1396 result += sizeof(u64); 1397 1398 if (type & PERF_SAMPLE_TID) 1399 result += sizeof(u64); 1400 1401 if (type & PERF_SAMPLE_TIME) 1402 result += sizeof(u64); 1403 1404 if (type & PERF_SAMPLE_ADDR) 1405 result += sizeof(u64); 1406 1407 if (type & PERF_SAMPLE_ID) 1408 result += sizeof(u64); 1409 1410 if (type & PERF_SAMPLE_STREAM_ID) 1411 result += sizeof(u64); 1412 1413 if (type & PERF_SAMPLE_CPU) 1414 result += sizeof(u64); 1415 1416 if (type & PERF_SAMPLE_PERIOD) 1417 result += sizeof(u64); 1418 1419 if (type & PERF_SAMPLE_READ) { 1420 result += sizeof(u64); 1421 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1422 result += sizeof(u64); 1423 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1424 result += sizeof(u64); 1425 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1426 if (read_format & PERF_FORMAT_GROUP) { 1427 sz = sample->read.group.nr * 1428 sizeof(struct sample_read_value); 1429 result += sz; 1430 } else { 1431 result += sizeof(u64); 1432 } 1433 } 1434 1435 if (type & PERF_SAMPLE_CALLCHAIN) { 1436 sz = (sample->callchain->nr + 1) * sizeof(u64); 1437 result += sz; 1438 } 1439 1440 if (type & PERF_SAMPLE_RAW) { 1441 result += sizeof(u32); 1442 result += sample->raw_size; 1443 } 1444 1445 if (type & PERF_SAMPLE_BRANCH_STACK) { 1446 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1447 /* nr, hw_idx */ 1448 sz += 2 * sizeof(u64); 1449 result += sz; 1450 } 1451 1452 if (type & PERF_SAMPLE_REGS_USER) { 1453 if (sample->user_regs.abi) { 1454 result += sizeof(u64); 1455 sz = hweight64(sample->user_regs.mask) * sizeof(u64); 1456 result += sz; 1457 } else { 1458 result += sizeof(u64); 1459 } 1460 } 1461 1462 if (type & PERF_SAMPLE_STACK_USER) { 1463 sz = sample->user_stack.size; 1464 result += sizeof(u64); 1465 if (sz) { 1466 result += sz; 1467 result += sizeof(u64); 1468 } 1469 } 1470 1471 if (type & PERF_SAMPLE_WEIGHT_TYPE) 1472 result += sizeof(u64); 1473 1474 if (type & PERF_SAMPLE_DATA_SRC) 1475 result += sizeof(u64); 1476 1477 if (type & PERF_SAMPLE_TRANSACTION) 1478 result += sizeof(u64); 1479 1480 if (type & PERF_SAMPLE_REGS_INTR) { 1481 if (sample->intr_regs.abi) { 1482 result += sizeof(u64); 1483 sz = hweight64(sample->intr_regs.mask) * sizeof(u64); 1484 result += sz; 1485 } else { 1486 result += sizeof(u64); 1487 } 1488 } 1489 1490 if (type & PERF_SAMPLE_PHYS_ADDR) 1491 result += sizeof(u64); 1492 1493 if (type & PERF_SAMPLE_CGROUP) 1494 result += sizeof(u64); 1495 1496 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) 1497 result += sizeof(u64); 1498 1499 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) 1500 result += sizeof(u64); 1501 1502 if (type & PERF_SAMPLE_AUX) { 1503 result += sizeof(u64); 1504 result += sample->aux_sample.size; 1505 } 1506 1507 return result; 1508 } 1509 1510 void __weak arch_perf_synthesize_sample_weight(const struct perf_sample *data, 1511 __u64 *array, u64 type __maybe_unused) 1512 { 1513 *array = data->weight; 1514 } 1515 1516 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format, 1517 const struct perf_sample *sample) 1518 { 1519 __u64 *array; 1520 size_t sz; 1521 /* 1522 * used for cross-endian analysis. See git commit 65014ab3 1523 * for why this goofiness is needed. 1524 */ 1525 union u64_swap u; 1526 1527 array = event->sample.array; 1528 1529 if (type & PERF_SAMPLE_IDENTIFIER) { 1530 *array = sample->id; 1531 array++; 1532 } 1533 1534 if (type & PERF_SAMPLE_IP) { 1535 *array = sample->ip; 1536 array++; 1537 } 1538 1539 if (type & PERF_SAMPLE_TID) { 1540 u.val32[0] = sample->pid; 1541 u.val32[1] = sample->tid; 1542 *array = u.val64; 1543 array++; 1544 } 1545 1546 if (type & PERF_SAMPLE_TIME) { 1547 *array = sample->time; 1548 array++; 1549 } 1550 1551 if (type & PERF_SAMPLE_ADDR) { 1552 *array = sample->addr; 1553 array++; 1554 } 1555 1556 if (type & PERF_SAMPLE_ID) { 1557 *array = sample->id; 1558 array++; 1559 } 1560 1561 if (type & PERF_SAMPLE_STREAM_ID) { 1562 *array = sample->stream_id; 1563 array++; 1564 } 1565 1566 if (type & PERF_SAMPLE_CPU) { 1567 u.val32[0] = sample->cpu; 1568 u.val32[1] = 0; 1569 *array = u.val64; 1570 array++; 1571 } 1572 1573 if (type & PERF_SAMPLE_PERIOD) { 1574 *array = sample->period; 1575 array++; 1576 } 1577 1578 if (type & PERF_SAMPLE_READ) { 1579 if (read_format & PERF_FORMAT_GROUP) 1580 *array = sample->read.group.nr; 1581 else 1582 *array = sample->read.one.value; 1583 array++; 1584 1585 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 1586 *array = sample->read.time_enabled; 1587 array++; 1588 } 1589 1590 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 1591 *array = sample->read.time_running; 1592 array++; 1593 } 1594 1595 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1596 if (read_format & PERF_FORMAT_GROUP) { 1597 sz = sample->read.group.nr * 1598 sizeof(struct sample_read_value); 1599 memcpy(array, sample->read.group.values, sz); 1600 array = (void *)array + sz; 1601 } else { 1602 *array = sample->read.one.id; 1603 array++; 1604 } 1605 } 1606 1607 if (type & PERF_SAMPLE_CALLCHAIN) { 1608 sz = (sample->callchain->nr + 1) * sizeof(u64); 1609 memcpy(array, sample->callchain, sz); 1610 array = (void *)array + sz; 1611 } 1612 1613 if (type & PERF_SAMPLE_RAW) { 1614 u.val32[0] = sample->raw_size; 1615 *array = u.val64; 1616 array = (void *)array + sizeof(u32); 1617 1618 memcpy(array, sample->raw_data, sample->raw_size); 1619 array = (void *)array + sample->raw_size; 1620 } 1621 1622 if (type & PERF_SAMPLE_BRANCH_STACK) { 1623 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1624 /* nr, hw_idx */ 1625 sz += 2 * sizeof(u64); 1626 memcpy(array, sample->branch_stack, sz); 1627 array = (void *)array + sz; 1628 } 1629 1630 if (type & PERF_SAMPLE_REGS_USER) { 1631 if (sample->user_regs.abi) { 1632 *array++ = sample->user_regs.abi; 1633 sz = hweight64(sample->user_regs.mask) * sizeof(u64); 1634 memcpy(array, sample->user_regs.regs, sz); 1635 array = (void *)array + sz; 1636 } else { 1637 *array++ = 0; 1638 } 1639 } 1640 1641 if (type & PERF_SAMPLE_STACK_USER) { 1642 sz = sample->user_stack.size; 1643 *array++ = sz; 1644 if (sz) { 1645 memcpy(array, sample->user_stack.data, sz); 1646 array = (void *)array + sz; 1647 *array++ = sz; 1648 } 1649 } 1650 1651 if (type & PERF_SAMPLE_WEIGHT_TYPE) { 1652 arch_perf_synthesize_sample_weight(sample, array, type); 1653 array++; 1654 } 1655 1656 if (type & PERF_SAMPLE_DATA_SRC) { 1657 *array = sample->data_src; 1658 array++; 1659 } 1660 1661 if (type & PERF_SAMPLE_TRANSACTION) { 1662 *array = sample->transaction; 1663 array++; 1664 } 1665 1666 if (type & PERF_SAMPLE_REGS_INTR) { 1667 if (sample->intr_regs.abi) { 1668 *array++ = sample->intr_regs.abi; 1669 sz = hweight64(sample->intr_regs.mask) * sizeof(u64); 1670 memcpy(array, sample->intr_regs.regs, sz); 1671 array = (void *)array + sz; 1672 } else { 1673 *array++ = 0; 1674 } 1675 } 1676 1677 if (type & PERF_SAMPLE_PHYS_ADDR) { 1678 *array = sample->phys_addr; 1679 array++; 1680 } 1681 1682 if (type & PERF_SAMPLE_CGROUP) { 1683 *array = sample->cgroup; 1684 array++; 1685 } 1686 1687 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) { 1688 *array = sample->data_page_size; 1689 array++; 1690 } 1691 1692 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) { 1693 *array = sample->code_page_size; 1694 array++; 1695 } 1696 1697 if (type & PERF_SAMPLE_AUX) { 1698 sz = sample->aux_sample.size; 1699 *array++ = sz; 1700 memcpy(array, sample->aux_sample.data, sz); 1701 array = (void *)array + sz; 1702 } 1703 1704 return 0; 1705 } 1706 1707 int perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process, 1708 struct evlist *evlist, struct machine *machine) 1709 { 1710 union perf_event *ev; 1711 struct evsel *evsel; 1712 size_t nr = 0, i = 0, sz, max_nr, n; 1713 int err; 1714 1715 pr_debug2("Synthesizing id index\n"); 1716 1717 max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) / 1718 sizeof(struct id_index_entry); 1719 1720 evlist__for_each_entry(evlist, evsel) 1721 nr += evsel->core.ids; 1722 1723 n = nr > max_nr ? max_nr : nr; 1724 sz = sizeof(struct perf_record_id_index) + n * sizeof(struct id_index_entry); 1725 ev = zalloc(sz); 1726 if (!ev) 1727 return -ENOMEM; 1728 1729 ev->id_index.header.type = PERF_RECORD_ID_INDEX; 1730 ev->id_index.header.size = sz; 1731 ev->id_index.nr = n; 1732 1733 evlist__for_each_entry(evlist, evsel) { 1734 u32 j; 1735 1736 for (j = 0; j < evsel->core.ids; j++) { 1737 struct id_index_entry *e; 1738 struct perf_sample_id *sid; 1739 1740 if (i >= n) { 1741 err = process(tool, ev, NULL, machine); 1742 if (err) 1743 goto out_err; 1744 nr -= n; 1745 i = 0; 1746 } 1747 1748 e = &ev->id_index.entries[i++]; 1749 1750 e->id = evsel->core.id[j]; 1751 1752 sid = evlist__id2sid(evlist, e->id); 1753 if (!sid) { 1754 free(ev); 1755 return -ENOENT; 1756 } 1757 1758 e->idx = sid->idx; 1759 e->cpu = sid->cpu; 1760 e->tid = sid->tid; 1761 } 1762 } 1763 1764 sz = sizeof(struct perf_record_id_index) + nr * sizeof(struct id_index_entry); 1765 ev->id_index.header.size = sz; 1766 ev->id_index.nr = nr; 1767 1768 err = process(tool, ev, NULL, machine); 1769 out_err: 1770 free(ev); 1771 1772 return err; 1773 } 1774 1775 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool, 1776 struct target *target, struct perf_thread_map *threads, 1777 perf_event__handler_t process, bool data_mmap, 1778 unsigned int nr_threads_synthesize) 1779 { 1780 if (target__has_task(target)) 1781 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap); 1782 else if (target__has_cpu(target)) 1783 return perf_event__synthesize_threads(tool, process, 1784 machine, data_mmap, 1785 nr_threads_synthesize); 1786 /* command specified */ 1787 return 0; 1788 } 1789 1790 int machine__synthesize_threads(struct machine *machine, struct target *target, 1791 struct perf_thread_map *threads, bool data_mmap, 1792 unsigned int nr_threads_synthesize) 1793 { 1794 return __machine__synthesize_threads(machine, NULL, target, threads, 1795 perf_event__process, data_mmap, 1796 nr_threads_synthesize); 1797 } 1798 1799 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id) 1800 { 1801 struct perf_record_event_update *ev; 1802 1803 size += sizeof(*ev); 1804 size = PERF_ALIGN(size, sizeof(u64)); 1805 1806 ev = zalloc(size); 1807 if (ev) { 1808 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1809 ev->header.size = (u16)size; 1810 ev->type = type; 1811 ev->id = id; 1812 } 1813 return ev; 1814 } 1815 1816 int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evsel *evsel, 1817 perf_event__handler_t process) 1818 { 1819 size_t size = strlen(evsel->unit); 1820 struct perf_record_event_update *ev; 1821 int err; 1822 1823 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]); 1824 if (ev == NULL) 1825 return -ENOMEM; 1826 1827 strlcpy(ev->data, evsel->unit, size + 1); 1828 err = process(tool, (union perf_event *)ev, NULL, NULL); 1829 free(ev); 1830 return err; 1831 } 1832 1833 int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsel *evsel, 1834 perf_event__handler_t process) 1835 { 1836 struct perf_record_event_update *ev; 1837 struct perf_record_event_update_scale *ev_data; 1838 int err; 1839 1840 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]); 1841 if (ev == NULL) 1842 return -ENOMEM; 1843 1844 ev_data = (struct perf_record_event_update_scale *)ev->data; 1845 ev_data->scale = evsel->scale; 1846 err = process(tool, (union perf_event *)ev, NULL, NULL); 1847 free(ev); 1848 return err; 1849 } 1850 1851 int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evsel *evsel, 1852 perf_event__handler_t process) 1853 { 1854 struct perf_record_event_update *ev; 1855 size_t len = strlen(evsel->name); 1856 int err; 1857 1858 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]); 1859 if (ev == NULL) 1860 return -ENOMEM; 1861 1862 strlcpy(ev->data, evsel->name, len + 1); 1863 err = process(tool, (union perf_event *)ev, NULL, NULL); 1864 free(ev); 1865 return err; 1866 } 1867 1868 int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel, 1869 perf_event__handler_t process) 1870 { 1871 size_t size = sizeof(struct perf_record_event_update); 1872 struct perf_record_event_update *ev; 1873 int max, err; 1874 u16 type; 1875 1876 if (!evsel->core.own_cpus) 1877 return 0; 1878 1879 ev = cpu_map_data__alloc(evsel->core.own_cpus, &size, &type, &max); 1880 if (!ev) 1881 return -ENOMEM; 1882 1883 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1884 ev->header.size = (u16)size; 1885 ev->type = PERF_EVENT_UPDATE__CPUS; 1886 ev->id = evsel->core.id[0]; 1887 1888 cpu_map_data__synthesize((struct perf_record_cpu_map_data *)ev->data, 1889 evsel->core.own_cpus, type, max); 1890 1891 err = process(tool, (union perf_event *)ev, NULL, NULL); 1892 free(ev); 1893 return err; 1894 } 1895 1896 int perf_event__synthesize_attrs(struct perf_tool *tool, struct evlist *evlist, 1897 perf_event__handler_t process) 1898 { 1899 struct evsel *evsel; 1900 int err = 0; 1901 1902 evlist__for_each_entry(evlist, evsel) { 1903 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids, 1904 evsel->core.id, process); 1905 if (err) { 1906 pr_debug("failed to create perf header attribute\n"); 1907 return err; 1908 } 1909 } 1910 1911 return err; 1912 } 1913 1914 static bool has_unit(struct evsel *evsel) 1915 { 1916 return evsel->unit && *evsel->unit; 1917 } 1918 1919 static bool has_scale(struct evsel *evsel) 1920 { 1921 return evsel->scale != 1; 1922 } 1923 1924 int perf_event__synthesize_extra_attr(struct perf_tool *tool, struct evlist *evsel_list, 1925 perf_event__handler_t process, bool is_pipe) 1926 { 1927 struct evsel *evsel; 1928 int err; 1929 1930 /* 1931 * Synthesize other events stuff not carried within 1932 * attr event - unit, scale, name 1933 */ 1934 evlist__for_each_entry(evsel_list, evsel) { 1935 if (!evsel->supported) 1936 continue; 1937 1938 /* 1939 * Synthesize unit and scale only if it's defined. 1940 */ 1941 if (has_unit(evsel)) { 1942 err = perf_event__synthesize_event_update_unit(tool, evsel, process); 1943 if (err < 0) { 1944 pr_err("Couldn't synthesize evsel unit.\n"); 1945 return err; 1946 } 1947 } 1948 1949 if (has_scale(evsel)) { 1950 err = perf_event__synthesize_event_update_scale(tool, evsel, process); 1951 if (err < 0) { 1952 pr_err("Couldn't synthesize evsel evsel.\n"); 1953 return err; 1954 } 1955 } 1956 1957 if (evsel->core.own_cpus) { 1958 err = perf_event__synthesize_event_update_cpus(tool, evsel, process); 1959 if (err < 0) { 1960 pr_err("Couldn't synthesize evsel cpus.\n"); 1961 return err; 1962 } 1963 } 1964 1965 /* 1966 * Name is needed only for pipe output, 1967 * perf.data carries event names. 1968 */ 1969 if (is_pipe) { 1970 err = perf_event__synthesize_event_update_name(tool, evsel, process); 1971 if (err < 0) { 1972 pr_err("Couldn't synthesize evsel name.\n"); 1973 return err; 1974 } 1975 } 1976 } 1977 return 0; 1978 } 1979 1980 int perf_event__synthesize_attr(struct perf_tool *tool, struct perf_event_attr *attr, 1981 u32 ids, u64 *id, perf_event__handler_t process) 1982 { 1983 union perf_event *ev; 1984 size_t size; 1985 int err; 1986 1987 size = sizeof(struct perf_event_attr); 1988 size = PERF_ALIGN(size, sizeof(u64)); 1989 size += sizeof(struct perf_event_header); 1990 size += ids * sizeof(u64); 1991 1992 ev = zalloc(size); 1993 1994 if (ev == NULL) 1995 return -ENOMEM; 1996 1997 ev->attr.attr = *attr; 1998 memcpy(ev->attr.id, id, ids * sizeof(u64)); 1999 2000 ev->attr.header.type = PERF_RECORD_HEADER_ATTR; 2001 ev->attr.header.size = (u16)size; 2002 2003 if (ev->attr.header.size == size) 2004 err = process(tool, ev, NULL, NULL); 2005 else 2006 err = -E2BIG; 2007 2008 free(ev); 2009 2010 return err; 2011 } 2012 2013 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, struct evlist *evlist, 2014 perf_event__handler_t process) 2015 { 2016 union perf_event ev; 2017 struct tracing_data *tdata; 2018 ssize_t size = 0, aligned_size = 0, padding; 2019 struct feat_fd ff; 2020 2021 /* 2022 * We are going to store the size of the data followed 2023 * by the data contents. Since the fd descriptor is a pipe, 2024 * we cannot seek back to store the size of the data once 2025 * we know it. Instead we: 2026 * 2027 * - write the tracing data to the temp file 2028 * - get/write the data size to pipe 2029 * - write the tracing data from the temp file 2030 * to the pipe 2031 */ 2032 tdata = tracing_data_get(&evlist->core.entries, fd, true); 2033 if (!tdata) 2034 return -1; 2035 2036 memset(&ev, 0, sizeof(ev)); 2037 2038 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA; 2039 size = tdata->size; 2040 aligned_size = PERF_ALIGN(size, sizeof(u64)); 2041 padding = aligned_size - size; 2042 ev.tracing_data.header.size = sizeof(ev.tracing_data); 2043 ev.tracing_data.size = aligned_size; 2044 2045 process(tool, &ev, NULL, NULL); 2046 2047 /* 2048 * The put function will copy all the tracing data 2049 * stored in temp file to the pipe. 2050 */ 2051 tracing_data_put(tdata); 2052 2053 ff = (struct feat_fd){ .fd = fd }; 2054 if (write_padded(&ff, NULL, 0, padding)) 2055 return -1; 2056 2057 return aligned_size; 2058 } 2059 2060 int perf_event__synthesize_build_id(struct perf_tool *tool, struct dso *pos, u16 misc, 2061 perf_event__handler_t process, struct machine *machine) 2062 { 2063 union perf_event ev; 2064 size_t len; 2065 2066 if (!pos->hit) 2067 return 0; 2068 2069 memset(&ev, 0, sizeof(ev)); 2070 2071 len = pos->long_name_len + 1; 2072 len = PERF_ALIGN(len, NAME_ALIGN); 2073 memcpy(&ev.build_id.build_id, pos->bid.data, sizeof(pos->bid.data)); 2074 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID; 2075 ev.build_id.header.misc = misc; 2076 ev.build_id.pid = machine->pid; 2077 ev.build_id.header.size = sizeof(ev.build_id) + len; 2078 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len); 2079 2080 return process(tool, &ev, NULL, machine); 2081 } 2082 2083 int perf_event__synthesize_stat_events(struct perf_stat_config *config, struct perf_tool *tool, 2084 struct evlist *evlist, perf_event__handler_t process, bool attrs) 2085 { 2086 int err; 2087 2088 if (attrs) { 2089 err = perf_event__synthesize_attrs(tool, evlist, process); 2090 if (err < 0) { 2091 pr_err("Couldn't synthesize attrs.\n"); 2092 return err; 2093 } 2094 } 2095 2096 err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs); 2097 err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL); 2098 if (err < 0) { 2099 pr_err("Couldn't synthesize thread map.\n"); 2100 return err; 2101 } 2102 2103 err = perf_event__synthesize_cpu_map(tool, evlist->core.cpus, process, NULL); 2104 if (err < 0) { 2105 pr_err("Couldn't synthesize thread map.\n"); 2106 return err; 2107 } 2108 2109 err = perf_event__synthesize_stat_config(tool, config, process, NULL); 2110 if (err < 0) { 2111 pr_err("Couldn't synthesize config.\n"); 2112 return err; 2113 } 2114 2115 return 0; 2116 } 2117 2118 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE]; 2119 2120 int perf_event__synthesize_features(struct perf_tool *tool, struct perf_session *session, 2121 struct evlist *evlist, perf_event__handler_t process) 2122 { 2123 struct perf_header *header = &session->header; 2124 struct perf_record_header_feature *fe; 2125 struct feat_fd ff; 2126 size_t sz, sz_hdr; 2127 int feat, ret; 2128 2129 sz_hdr = sizeof(fe->header); 2130 sz = sizeof(union perf_event); 2131 /* get a nice alignment */ 2132 sz = PERF_ALIGN(sz, page_size); 2133 2134 memset(&ff, 0, sizeof(ff)); 2135 2136 ff.buf = malloc(sz); 2137 if (!ff.buf) 2138 return -ENOMEM; 2139 2140 ff.size = sz - sz_hdr; 2141 ff.ph = &session->header; 2142 2143 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) { 2144 if (!feat_ops[feat].synthesize) { 2145 pr_debug("No record header feature for header :%d\n", feat); 2146 continue; 2147 } 2148 2149 ff.offset = sizeof(*fe); 2150 2151 ret = feat_ops[feat].write(&ff, evlist); 2152 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) { 2153 pr_debug("Error writing feature\n"); 2154 continue; 2155 } 2156 /* ff.buf may have changed due to realloc in do_write() */ 2157 fe = ff.buf; 2158 memset(fe, 0, sizeof(*fe)); 2159 2160 fe->feat_id = feat; 2161 fe->header.type = PERF_RECORD_HEADER_FEATURE; 2162 fe->header.size = ff.offset; 2163 2164 ret = process(tool, ff.buf, NULL, NULL); 2165 if (ret) { 2166 free(ff.buf); 2167 return ret; 2168 } 2169 } 2170 2171 /* Send HEADER_LAST_FEATURE mark. */ 2172 fe = ff.buf; 2173 fe->feat_id = HEADER_LAST_FEATURE; 2174 fe->header.type = PERF_RECORD_HEADER_FEATURE; 2175 fe->header.size = sizeof(*fe); 2176 2177 ret = process(tool, ff.buf, NULL, NULL); 2178 2179 free(ff.buf); 2180 return ret; 2181 } 2182