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 <internal/cpumap.h> 28 #include <perf/cpumap.h> 29 #include <internal/lib.h> // page_size 30 #include <internal/threadmap.h> 31 #include <perf/threadmap.h> 32 #include <symbol/kallsyms.h> 33 #include <dirent.h> 34 #include <errno.h> 35 #include <inttypes.h> 36 #include <stdio.h> 37 #include <string.h> 38 #include <uapi/linux/mman.h> /* To get things like MAP_HUGETLB even on older libc headers */ 39 #include <api/fs/fs.h> 40 #include <api/io.h> 41 #include <sys/types.h> 42 #include <sys/stat.h> 43 #include <fcntl.h> 44 #include <unistd.h> 45 46 #define DEFAULT_PROC_MAP_PARSE_TIMEOUT 500 47 48 unsigned int proc_map_timeout = DEFAULT_PROC_MAP_PARSE_TIMEOUT; 49 50 int perf_tool__process_synth_event(struct perf_tool *tool, 51 union perf_event *event, 52 struct machine *machine, 53 perf_event__handler_t process) 54 { 55 struct perf_sample synth_sample = { 56 .pid = -1, 57 .tid = -1, 58 .time = -1, 59 .stream_id = -1, 60 .cpu = -1, 61 .period = 1, 62 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK, 63 }; 64 65 return process(tool, event, &synth_sample, machine); 66 }; 67 68 /* 69 * Assumes that the first 4095 bytes of /proc/pid/stat contains 70 * the comm, tgid and ppid. 71 */ 72 static int perf_event__get_comm_ids(pid_t pid, char *comm, size_t len, 73 pid_t *tgid, pid_t *ppid) 74 { 75 char bf[4096]; 76 int fd; 77 size_t size = 0; 78 ssize_t n; 79 char *name, *tgids, *ppids; 80 81 *tgid = -1; 82 *ppid = -1; 83 84 snprintf(bf, sizeof(bf), "/proc/%d/status", pid); 85 86 fd = open(bf, O_RDONLY); 87 if (fd < 0) { 88 pr_debug("couldn't open %s\n", bf); 89 return -1; 90 } 91 92 n = read(fd, bf, sizeof(bf) - 1); 93 close(fd); 94 if (n <= 0) { 95 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n", 96 pid); 97 return -1; 98 } 99 bf[n] = '\0'; 100 101 name = strstr(bf, "Name:"); 102 tgids = strstr(bf, "Tgid:"); 103 ppids = strstr(bf, "PPid:"); 104 105 if (name) { 106 char *nl; 107 108 name = skip_spaces(name + 5); /* strlen("Name:") */ 109 nl = strchr(name, '\n'); 110 if (nl) 111 *nl = '\0'; 112 113 size = strlen(name); 114 if (size >= len) 115 size = len - 1; 116 memcpy(comm, name, size); 117 comm[size] = '\0'; 118 } else { 119 pr_debug("Name: string not found for pid %d\n", pid); 120 } 121 122 if (tgids) { 123 tgids += 5; /* strlen("Tgid:") */ 124 *tgid = atoi(tgids); 125 } else { 126 pr_debug("Tgid: string not found for pid %d\n", pid); 127 } 128 129 if (ppids) { 130 ppids += 5; /* strlen("PPid:") */ 131 *ppid = atoi(ppids); 132 } else { 133 pr_debug("PPid: string not found for pid %d\n", pid); 134 } 135 136 return 0; 137 } 138 139 static int perf_event__prepare_comm(union perf_event *event, pid_t pid, 140 struct machine *machine, 141 pid_t *tgid, pid_t *ppid) 142 { 143 size_t size; 144 145 *ppid = -1; 146 147 memset(&event->comm, 0, sizeof(event->comm)); 148 149 if (machine__is_host(machine)) { 150 if (perf_event__get_comm_ids(pid, event->comm.comm, 151 sizeof(event->comm.comm), 152 tgid, ppid) != 0) { 153 return -1; 154 } 155 } else { 156 *tgid = machine->pid; 157 } 158 159 if (*tgid < 0) 160 return -1; 161 162 event->comm.pid = *tgid; 163 event->comm.header.type = PERF_RECORD_COMM; 164 165 size = strlen(event->comm.comm) + 1; 166 size = PERF_ALIGN(size, sizeof(u64)); 167 memset(event->comm.comm + size, 0, machine->id_hdr_size); 168 event->comm.header.size = (sizeof(event->comm) - 169 (sizeof(event->comm.comm) - size) + 170 machine->id_hdr_size); 171 event->comm.tid = pid; 172 173 return 0; 174 } 175 176 pid_t perf_event__synthesize_comm(struct perf_tool *tool, 177 union perf_event *event, pid_t pid, 178 perf_event__handler_t process, 179 struct machine *machine) 180 { 181 pid_t tgid, ppid; 182 183 if (perf_event__prepare_comm(event, pid, machine, &tgid, &ppid) != 0) 184 return -1; 185 186 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 187 return -1; 188 189 return tgid; 190 } 191 192 static void perf_event__get_ns_link_info(pid_t pid, const char *ns, 193 struct perf_ns_link_info *ns_link_info) 194 { 195 struct stat64 st; 196 char proc_ns[128]; 197 198 sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns); 199 if (stat64(proc_ns, &st) == 0) { 200 ns_link_info->dev = st.st_dev; 201 ns_link_info->ino = st.st_ino; 202 } 203 } 204 205 int perf_event__synthesize_namespaces(struct perf_tool *tool, 206 union perf_event *event, 207 pid_t pid, pid_t tgid, 208 perf_event__handler_t process, 209 struct machine *machine) 210 { 211 u32 idx; 212 struct perf_ns_link_info *ns_link_info; 213 214 if (!tool || !tool->namespace_events) 215 return 0; 216 217 memset(&event->namespaces, 0, (sizeof(event->namespaces) + 218 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 219 machine->id_hdr_size)); 220 221 event->namespaces.pid = tgid; 222 event->namespaces.tid = pid; 223 224 event->namespaces.nr_namespaces = NR_NAMESPACES; 225 226 ns_link_info = event->namespaces.link_info; 227 228 for (idx = 0; idx < event->namespaces.nr_namespaces; idx++) 229 perf_event__get_ns_link_info(pid, perf_ns__name(idx), 230 &ns_link_info[idx]); 231 232 event->namespaces.header.type = PERF_RECORD_NAMESPACES; 233 234 event->namespaces.header.size = (sizeof(event->namespaces) + 235 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 236 machine->id_hdr_size); 237 238 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 239 return -1; 240 241 return 0; 242 } 243 244 static int perf_event__synthesize_fork(struct perf_tool *tool, 245 union perf_event *event, 246 pid_t pid, pid_t tgid, pid_t ppid, 247 perf_event__handler_t process, 248 struct machine *machine) 249 { 250 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size); 251 252 /* 253 * for main thread set parent to ppid from status file. For other 254 * threads set parent pid to main thread. ie., assume main thread 255 * spawns all threads in a process 256 */ 257 if (tgid == pid) { 258 event->fork.ppid = ppid; 259 event->fork.ptid = ppid; 260 } else { 261 event->fork.ppid = tgid; 262 event->fork.ptid = tgid; 263 } 264 event->fork.pid = tgid; 265 event->fork.tid = pid; 266 event->fork.header.type = PERF_RECORD_FORK; 267 event->fork.header.misc = PERF_RECORD_MISC_FORK_EXEC; 268 269 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size); 270 271 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 272 return -1; 273 274 return 0; 275 } 276 277 static bool read_proc_maps_line(struct io *io, __u64 *start, __u64 *end, 278 u32 *prot, u32 *flags, __u64 *offset, 279 u32 *maj, u32 *min, 280 __u64 *inode, 281 ssize_t pathname_size, char *pathname) 282 { 283 __u64 temp; 284 int ch; 285 char *start_pathname = pathname; 286 287 if (io__get_hex(io, start) != '-') 288 return false; 289 if (io__get_hex(io, end) != ' ') 290 return false; 291 292 /* map protection and flags bits */ 293 *prot = 0; 294 ch = io__get_char(io); 295 if (ch == 'r') 296 *prot |= PROT_READ; 297 else if (ch != '-') 298 return false; 299 ch = io__get_char(io); 300 if (ch == 'w') 301 *prot |= PROT_WRITE; 302 else if (ch != '-') 303 return false; 304 ch = io__get_char(io); 305 if (ch == 'x') 306 *prot |= PROT_EXEC; 307 else if (ch != '-') 308 return false; 309 ch = io__get_char(io); 310 if (ch == 's') 311 *flags = MAP_SHARED; 312 else if (ch == 'p') 313 *flags = MAP_PRIVATE; 314 else 315 return false; 316 if (io__get_char(io) != ' ') 317 return false; 318 319 if (io__get_hex(io, offset) != ' ') 320 return false; 321 322 if (io__get_hex(io, &temp) != ':') 323 return false; 324 *maj = temp; 325 if (io__get_hex(io, &temp) != ' ') 326 return false; 327 *min = temp; 328 329 ch = io__get_dec(io, inode); 330 if (ch != ' ') { 331 *pathname = '\0'; 332 return ch == '\n'; 333 } 334 do { 335 ch = io__get_char(io); 336 } while (ch == ' '); 337 while (true) { 338 if (ch < 0) 339 return false; 340 if (ch == '\0' || ch == '\n' || 341 (pathname + 1 - start_pathname) >= pathname_size) { 342 *pathname = '\0'; 343 return true; 344 } 345 *pathname++ = ch; 346 ch = io__get_char(io); 347 } 348 } 349 350 int perf_event__synthesize_mmap_events(struct perf_tool *tool, 351 union perf_event *event, 352 pid_t pid, pid_t tgid, 353 perf_event__handler_t process, 354 struct machine *machine, 355 bool mmap_data) 356 { 357 unsigned long long t; 358 char bf[BUFSIZ]; 359 struct io io; 360 bool truncation = false; 361 unsigned long long timeout = proc_map_timeout * 1000000ULL; 362 int rc = 0; 363 const char *hugetlbfs_mnt = hugetlbfs__mountpoint(); 364 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0; 365 366 if (machine__is_default_guest(machine)) 367 return 0; 368 369 snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps", 370 machine->root_dir, pid, pid); 371 372 io.fd = open(bf, O_RDONLY, 0); 373 if (io.fd < 0) { 374 /* 375 * We raced with a task exiting - just return: 376 */ 377 pr_debug("couldn't open %s\n", bf); 378 return -1; 379 } 380 io__init(&io, io.fd, bf, sizeof(bf)); 381 382 event->header.type = PERF_RECORD_MMAP2; 383 t = rdclock(); 384 385 while (!io.eof) { 386 static const char anonstr[] = "//anon"; 387 size_t size; 388 389 /* ensure null termination since stack will be reused. */ 390 event->mmap2.filename[0] = '\0'; 391 392 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */ 393 if (!read_proc_maps_line(&io, 394 &event->mmap2.start, 395 &event->mmap2.len, 396 &event->mmap2.prot, 397 &event->mmap2.flags, 398 &event->mmap2.pgoff, 399 &event->mmap2.maj, 400 &event->mmap2.min, 401 &event->mmap2.ino, 402 sizeof(event->mmap2.filename), 403 event->mmap2.filename)) 404 continue; 405 406 if ((rdclock() - t) > timeout) { 407 pr_warning("Reading %s/proc/%d/task/%d/maps time out. " 408 "You may want to increase " 409 "the time limit by --proc-map-timeout\n", 410 machine->root_dir, pid, pid); 411 truncation = true; 412 goto out; 413 } 414 415 event->mmap2.ino_generation = 0; 416 417 /* 418 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c 419 */ 420 if (machine__is_host(machine)) 421 event->header.misc = PERF_RECORD_MISC_USER; 422 else 423 event->header.misc = PERF_RECORD_MISC_GUEST_USER; 424 425 if ((event->mmap2.prot & PROT_EXEC) == 0) { 426 if (!mmap_data || (event->mmap2.prot & PROT_READ) == 0) 427 continue; 428 429 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA; 430 } 431 432 out: 433 if (truncation) 434 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT; 435 436 if (!strcmp(event->mmap2.filename, "")) 437 strcpy(event->mmap2.filename, anonstr); 438 439 if (hugetlbfs_mnt_len && 440 !strncmp(event->mmap2.filename, hugetlbfs_mnt, 441 hugetlbfs_mnt_len)) { 442 strcpy(event->mmap2.filename, anonstr); 443 event->mmap2.flags |= MAP_HUGETLB; 444 } 445 446 size = strlen(event->mmap2.filename) + 1; 447 size = PERF_ALIGN(size, sizeof(u64)); 448 event->mmap2.len -= event->mmap.start; 449 event->mmap2.header.size = (sizeof(event->mmap2) - 450 (sizeof(event->mmap2.filename) - size)); 451 memset(event->mmap2.filename + size, 0, machine->id_hdr_size); 452 event->mmap2.header.size += machine->id_hdr_size; 453 event->mmap2.pid = tgid; 454 event->mmap2.tid = pid; 455 456 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 457 rc = -1; 458 break; 459 } 460 461 if (truncation) 462 break; 463 } 464 465 close(io.fd); 466 return rc; 467 } 468 469 #ifdef HAVE_FILE_HANDLE 470 static int perf_event__synthesize_cgroup(struct perf_tool *tool, 471 union perf_event *event, 472 char *path, size_t mount_len, 473 perf_event__handler_t process, 474 struct machine *machine) 475 { 476 size_t event_size = sizeof(event->cgroup) - sizeof(event->cgroup.path); 477 size_t path_len = strlen(path) - mount_len + 1; 478 struct { 479 struct file_handle fh; 480 uint64_t cgroup_id; 481 } handle; 482 int mount_id; 483 484 while (path_len % sizeof(u64)) 485 path[mount_len + path_len++] = '\0'; 486 487 memset(&event->cgroup, 0, event_size); 488 489 event->cgroup.header.type = PERF_RECORD_CGROUP; 490 event->cgroup.header.size = event_size + path_len + machine->id_hdr_size; 491 492 handle.fh.handle_bytes = sizeof(handle.cgroup_id); 493 if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0) { 494 pr_debug("stat failed: %s\n", path); 495 return -1; 496 } 497 498 event->cgroup.id = handle.cgroup_id; 499 strncpy(event->cgroup.path, path + mount_len, path_len); 500 memset(event->cgroup.path + path_len, 0, machine->id_hdr_size); 501 502 if (perf_tool__process_synth_event(tool, event, machine, process) < 0) { 503 pr_debug("process synth event failed\n"); 504 return -1; 505 } 506 507 return 0; 508 } 509 510 static int perf_event__walk_cgroup_tree(struct perf_tool *tool, 511 union perf_event *event, 512 char *path, size_t mount_len, 513 perf_event__handler_t process, 514 struct machine *machine) 515 { 516 size_t pos = strlen(path); 517 DIR *d; 518 struct dirent *dent; 519 int ret = 0; 520 521 if (perf_event__synthesize_cgroup(tool, event, path, mount_len, 522 process, machine) < 0) 523 return -1; 524 525 d = opendir(path); 526 if (d == NULL) { 527 pr_debug("failed to open directory: %s\n", path); 528 return -1; 529 } 530 531 while ((dent = readdir(d)) != NULL) { 532 if (dent->d_type != DT_DIR) 533 continue; 534 if (!strcmp(dent->d_name, ".") || 535 !strcmp(dent->d_name, "..")) 536 continue; 537 538 /* any sane path should be less than PATH_MAX */ 539 if (strlen(path) + strlen(dent->d_name) + 1 >= PATH_MAX) 540 continue; 541 542 if (path[pos - 1] != '/') 543 strcat(path, "/"); 544 strcat(path, dent->d_name); 545 546 ret = perf_event__walk_cgroup_tree(tool, event, path, 547 mount_len, process, machine); 548 if (ret < 0) 549 break; 550 551 path[pos] = '\0'; 552 } 553 554 closedir(d); 555 return ret; 556 } 557 558 int perf_event__synthesize_cgroups(struct perf_tool *tool, 559 perf_event__handler_t process, 560 struct machine *machine) 561 { 562 union perf_event event; 563 char cgrp_root[PATH_MAX]; 564 size_t mount_len; /* length of mount point in the path */ 565 566 if (cgroupfs_find_mountpoint(cgrp_root, PATH_MAX, "perf_event") < 0) { 567 pr_debug("cannot find cgroup mount point\n"); 568 return -1; 569 } 570 571 mount_len = strlen(cgrp_root); 572 /* make sure the path starts with a slash (after mount point) */ 573 strcat(cgrp_root, "/"); 574 575 if (perf_event__walk_cgroup_tree(tool, &event, cgrp_root, mount_len, 576 process, machine) < 0) 577 return -1; 578 579 return 0; 580 } 581 #else 582 int perf_event__synthesize_cgroups(struct perf_tool *tool __maybe_unused, 583 perf_event__handler_t process __maybe_unused, 584 struct machine *machine __maybe_unused) 585 { 586 return -1; 587 } 588 #endif 589 590 int perf_event__synthesize_modules(struct perf_tool *tool, perf_event__handler_t process, 591 struct machine *machine) 592 { 593 int rc = 0; 594 struct map *pos; 595 struct maps *maps = machine__kernel_maps(machine); 596 union perf_event *event = zalloc((sizeof(event->mmap) + 597 machine->id_hdr_size)); 598 if (event == NULL) { 599 pr_debug("Not enough memory synthesizing mmap event " 600 "for kernel modules\n"); 601 return -1; 602 } 603 604 event->header.type = PERF_RECORD_MMAP; 605 606 /* 607 * kernel uses 0 for user space maps, see kernel/perf_event.c 608 * __perf_event_mmap 609 */ 610 if (machine__is_host(machine)) 611 event->header.misc = PERF_RECORD_MISC_KERNEL; 612 else 613 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 614 615 maps__for_each_entry(maps, pos) { 616 size_t size; 617 618 if (!__map__is_kmodule(pos)) 619 continue; 620 621 size = PERF_ALIGN(pos->dso->long_name_len + 1, sizeof(u64)); 622 event->mmap.header.type = PERF_RECORD_MMAP; 623 event->mmap.header.size = (sizeof(event->mmap) - 624 (sizeof(event->mmap.filename) - size)); 625 memset(event->mmap.filename + size, 0, machine->id_hdr_size); 626 event->mmap.header.size += machine->id_hdr_size; 627 event->mmap.start = pos->start; 628 event->mmap.len = pos->end - pos->start; 629 event->mmap.pid = machine->pid; 630 631 memcpy(event->mmap.filename, pos->dso->long_name, 632 pos->dso->long_name_len + 1); 633 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 634 rc = -1; 635 break; 636 } 637 } 638 639 free(event); 640 return rc; 641 } 642 643 static int __event__synthesize_thread(union perf_event *comm_event, 644 union perf_event *mmap_event, 645 union perf_event *fork_event, 646 union perf_event *namespaces_event, 647 pid_t pid, int full, perf_event__handler_t process, 648 struct perf_tool *tool, struct machine *machine, bool mmap_data) 649 { 650 char filename[PATH_MAX]; 651 DIR *tasks; 652 struct dirent *dirent; 653 pid_t tgid, ppid; 654 int rc = 0; 655 656 /* special case: only send one comm event using passed in pid */ 657 if (!full) { 658 tgid = perf_event__synthesize_comm(tool, comm_event, pid, 659 process, machine); 660 661 if (tgid == -1) 662 return -1; 663 664 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid, 665 tgid, process, machine) < 0) 666 return -1; 667 668 /* 669 * send mmap only for thread group leader 670 * see thread__init_maps() 671 */ 672 if (pid == tgid && 673 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 674 process, machine, mmap_data)) 675 return -1; 676 677 return 0; 678 } 679 680 if (machine__is_default_guest(machine)) 681 return 0; 682 683 snprintf(filename, sizeof(filename), "%s/proc/%d/task", 684 machine->root_dir, pid); 685 686 tasks = opendir(filename); 687 if (tasks == NULL) { 688 pr_debug("couldn't open %s\n", filename); 689 return 0; 690 } 691 692 while ((dirent = readdir(tasks)) != NULL) { 693 char *end; 694 pid_t _pid; 695 696 _pid = strtol(dirent->d_name, &end, 10); 697 if (*end) 698 continue; 699 700 rc = -1; 701 if (perf_event__prepare_comm(comm_event, _pid, machine, 702 &tgid, &ppid) != 0) 703 break; 704 705 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid, 706 ppid, process, machine) < 0) 707 break; 708 709 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid, 710 tgid, process, machine) < 0) 711 break; 712 713 /* 714 * Send the prepared comm event 715 */ 716 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0) 717 break; 718 719 rc = 0; 720 if (_pid == pid) { 721 /* process the parent's maps too */ 722 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 723 process, machine, mmap_data); 724 if (rc) 725 break; 726 } 727 } 728 729 closedir(tasks); 730 return rc; 731 } 732 733 int perf_event__synthesize_thread_map(struct perf_tool *tool, 734 struct perf_thread_map *threads, 735 perf_event__handler_t process, 736 struct machine *machine, 737 bool mmap_data) 738 { 739 union perf_event *comm_event, *mmap_event, *fork_event; 740 union perf_event *namespaces_event; 741 int err = -1, thread, j; 742 743 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 744 if (comm_event == NULL) 745 goto out; 746 747 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 748 if (mmap_event == NULL) 749 goto out_free_comm; 750 751 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 752 if (fork_event == NULL) 753 goto out_free_mmap; 754 755 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 756 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 757 machine->id_hdr_size); 758 if (namespaces_event == NULL) 759 goto out_free_fork; 760 761 err = 0; 762 for (thread = 0; thread < threads->nr; ++thread) { 763 if (__event__synthesize_thread(comm_event, mmap_event, 764 fork_event, namespaces_event, 765 perf_thread_map__pid(threads, thread), 0, 766 process, tool, machine, 767 mmap_data)) { 768 err = -1; 769 break; 770 } 771 772 /* 773 * comm.pid is set to thread group id by 774 * perf_event__synthesize_comm 775 */ 776 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) { 777 bool need_leader = true; 778 779 /* is thread group leader in thread_map? */ 780 for (j = 0; j < threads->nr; ++j) { 781 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) { 782 need_leader = false; 783 break; 784 } 785 } 786 787 /* if not, generate events for it */ 788 if (need_leader && 789 __event__synthesize_thread(comm_event, mmap_event, 790 fork_event, namespaces_event, 791 comm_event->comm.pid, 0, 792 process, tool, machine, 793 mmap_data)) { 794 err = -1; 795 break; 796 } 797 } 798 } 799 free(namespaces_event); 800 out_free_fork: 801 free(fork_event); 802 out_free_mmap: 803 free(mmap_event); 804 out_free_comm: 805 free(comm_event); 806 out: 807 return err; 808 } 809 810 static int __perf_event__synthesize_threads(struct perf_tool *tool, 811 perf_event__handler_t process, 812 struct machine *machine, 813 bool mmap_data, 814 struct dirent **dirent, 815 int start, 816 int num) 817 { 818 union perf_event *comm_event, *mmap_event, *fork_event; 819 union perf_event *namespaces_event; 820 int err = -1; 821 char *end; 822 pid_t pid; 823 int i; 824 825 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 826 if (comm_event == NULL) 827 goto out; 828 829 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 830 if (mmap_event == NULL) 831 goto out_free_comm; 832 833 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 834 if (fork_event == NULL) 835 goto out_free_mmap; 836 837 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 838 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 839 machine->id_hdr_size); 840 if (namespaces_event == NULL) 841 goto out_free_fork; 842 843 for (i = start; i < start + num; i++) { 844 if (!isdigit(dirent[i]->d_name[0])) 845 continue; 846 847 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10); 848 /* only interested in proper numerical dirents */ 849 if (*end) 850 continue; 851 /* 852 * We may race with exiting thread, so don't stop just because 853 * one thread couldn't be synthesized. 854 */ 855 __event__synthesize_thread(comm_event, mmap_event, fork_event, 856 namespaces_event, pid, 1, process, 857 tool, machine, mmap_data); 858 } 859 err = 0; 860 861 free(namespaces_event); 862 out_free_fork: 863 free(fork_event); 864 out_free_mmap: 865 free(mmap_event); 866 out_free_comm: 867 free(comm_event); 868 out: 869 return err; 870 } 871 872 struct synthesize_threads_arg { 873 struct perf_tool *tool; 874 perf_event__handler_t process; 875 struct machine *machine; 876 bool mmap_data; 877 struct dirent **dirent; 878 int num; 879 int start; 880 }; 881 882 static void *synthesize_threads_worker(void *arg) 883 { 884 struct synthesize_threads_arg *args = arg; 885 886 __perf_event__synthesize_threads(args->tool, args->process, 887 args->machine, args->mmap_data, 888 args->dirent, 889 args->start, args->num); 890 return NULL; 891 } 892 893 int perf_event__synthesize_threads(struct perf_tool *tool, 894 perf_event__handler_t process, 895 struct machine *machine, 896 bool mmap_data, 897 unsigned int nr_threads_synthesize) 898 { 899 struct synthesize_threads_arg *args = NULL; 900 pthread_t *synthesize_threads = NULL; 901 char proc_path[PATH_MAX]; 902 struct dirent **dirent; 903 int num_per_thread; 904 int m, n, i, j; 905 int thread_nr; 906 int base = 0; 907 int err = -1; 908 909 910 if (machine__is_default_guest(machine)) 911 return 0; 912 913 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir); 914 n = scandir(proc_path, &dirent, 0, alphasort); 915 if (n < 0) 916 return err; 917 918 if (nr_threads_synthesize == UINT_MAX) 919 thread_nr = sysconf(_SC_NPROCESSORS_ONLN); 920 else 921 thread_nr = nr_threads_synthesize; 922 923 if (thread_nr <= 1) { 924 err = __perf_event__synthesize_threads(tool, process, 925 machine, mmap_data, 926 dirent, base, n); 927 goto free_dirent; 928 } 929 if (thread_nr > n) 930 thread_nr = n; 931 932 synthesize_threads = calloc(sizeof(pthread_t), thread_nr); 933 if (synthesize_threads == NULL) 934 goto free_dirent; 935 936 args = calloc(sizeof(*args), thread_nr); 937 if (args == NULL) 938 goto free_threads; 939 940 num_per_thread = n / thread_nr; 941 m = n % thread_nr; 942 for (i = 0; i < thread_nr; i++) { 943 args[i].tool = tool; 944 args[i].process = process; 945 args[i].machine = machine; 946 args[i].mmap_data = mmap_data; 947 args[i].dirent = dirent; 948 } 949 for (i = 0; i < m; i++) { 950 args[i].num = num_per_thread + 1; 951 args[i].start = i * args[i].num; 952 } 953 if (i != 0) 954 base = args[i-1].start + args[i-1].num; 955 for (j = i; j < thread_nr; j++) { 956 args[j].num = num_per_thread; 957 args[j].start = base + (j - i) * args[i].num; 958 } 959 960 for (i = 0; i < thread_nr; i++) { 961 if (pthread_create(&synthesize_threads[i], NULL, 962 synthesize_threads_worker, &args[i])) 963 goto out_join; 964 } 965 err = 0; 966 out_join: 967 for (i = 0; i < thread_nr; i++) 968 pthread_join(synthesize_threads[i], NULL); 969 free(args); 970 free_threads: 971 free(synthesize_threads); 972 free_dirent: 973 for (i = 0; i < n; i++) 974 zfree(&dirent[i]); 975 free(dirent); 976 977 return err; 978 } 979 980 int __weak perf_event__synthesize_extra_kmaps(struct perf_tool *tool __maybe_unused, 981 perf_event__handler_t process __maybe_unused, 982 struct machine *machine __maybe_unused) 983 { 984 return 0; 985 } 986 987 static int __perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 988 perf_event__handler_t process, 989 struct machine *machine) 990 { 991 size_t size; 992 struct map *map = machine__kernel_map(machine); 993 struct kmap *kmap; 994 int err; 995 union perf_event *event; 996 997 if (map == NULL) 998 return -1; 999 1000 kmap = map__kmap(map); 1001 if (!kmap->ref_reloc_sym) 1002 return -1; 1003 1004 /* 1005 * We should get this from /sys/kernel/sections/.text, but till that is 1006 * available use this, and after it is use this as a fallback for older 1007 * kernels. 1008 */ 1009 event = zalloc((sizeof(event->mmap) + machine->id_hdr_size)); 1010 if (event == NULL) { 1011 pr_debug("Not enough memory synthesizing mmap event " 1012 "for kernel modules\n"); 1013 return -1; 1014 } 1015 1016 if (machine__is_host(machine)) { 1017 /* 1018 * kernel uses PERF_RECORD_MISC_USER for user space maps, 1019 * see kernel/perf_event.c __perf_event_mmap 1020 */ 1021 event->header.misc = PERF_RECORD_MISC_KERNEL; 1022 } else { 1023 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 1024 } 1025 1026 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename), 1027 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1; 1028 size = PERF_ALIGN(size, sizeof(u64)); 1029 event->mmap.header.type = PERF_RECORD_MMAP; 1030 event->mmap.header.size = (sizeof(event->mmap) - 1031 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size); 1032 event->mmap.pgoff = kmap->ref_reloc_sym->addr; 1033 event->mmap.start = map->start; 1034 event->mmap.len = map->end - event->mmap.start; 1035 event->mmap.pid = machine->pid; 1036 1037 err = perf_tool__process_synth_event(tool, event, machine, process); 1038 free(event); 1039 1040 return err; 1041 } 1042 1043 int perf_event__synthesize_kernel_mmap(struct perf_tool *tool, 1044 perf_event__handler_t process, 1045 struct machine *machine) 1046 { 1047 int err; 1048 1049 err = __perf_event__synthesize_kernel_mmap(tool, process, machine); 1050 if (err < 0) 1051 return err; 1052 1053 return perf_event__synthesize_extra_kmaps(tool, process, machine); 1054 } 1055 1056 int perf_event__synthesize_thread_map2(struct perf_tool *tool, 1057 struct perf_thread_map *threads, 1058 perf_event__handler_t process, 1059 struct machine *machine) 1060 { 1061 union perf_event *event; 1062 int i, err, size; 1063 1064 size = sizeof(event->thread_map); 1065 size += threads->nr * sizeof(event->thread_map.entries[0]); 1066 1067 event = zalloc(size); 1068 if (!event) 1069 return -ENOMEM; 1070 1071 event->header.type = PERF_RECORD_THREAD_MAP; 1072 event->header.size = size; 1073 event->thread_map.nr = threads->nr; 1074 1075 for (i = 0; i < threads->nr; i++) { 1076 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i]; 1077 char *comm = perf_thread_map__comm(threads, i); 1078 1079 if (!comm) 1080 comm = (char *) ""; 1081 1082 entry->pid = perf_thread_map__pid(threads, i); 1083 strncpy((char *) &entry->comm, comm, sizeof(entry->comm)); 1084 } 1085 1086 err = process(tool, event, NULL, machine); 1087 1088 free(event); 1089 return err; 1090 } 1091 1092 static void synthesize_cpus(struct cpu_map_entries *cpus, 1093 struct perf_cpu_map *map) 1094 { 1095 int i; 1096 1097 cpus->nr = map->nr; 1098 1099 for (i = 0; i < map->nr; i++) 1100 cpus->cpu[i] = map->map[i]; 1101 } 1102 1103 static void synthesize_mask(struct perf_record_record_cpu_map *mask, 1104 struct perf_cpu_map *map, int max) 1105 { 1106 int i; 1107 1108 mask->nr = BITS_TO_LONGS(max); 1109 mask->long_size = sizeof(long); 1110 1111 for (i = 0; i < map->nr; i++) 1112 set_bit(map->map[i], mask->mask); 1113 } 1114 1115 static size_t cpus_size(struct perf_cpu_map *map) 1116 { 1117 return sizeof(struct cpu_map_entries) + map->nr * sizeof(u16); 1118 } 1119 1120 static size_t mask_size(struct perf_cpu_map *map, int *max) 1121 { 1122 int i; 1123 1124 *max = 0; 1125 1126 for (i = 0; i < map->nr; i++) { 1127 /* bit possition of the cpu is + 1 */ 1128 int bit = map->map[i] + 1; 1129 1130 if (bit > *max) 1131 *max = bit; 1132 } 1133 1134 return sizeof(struct perf_record_record_cpu_map) + BITS_TO_LONGS(*max) * sizeof(long); 1135 } 1136 1137 void *cpu_map_data__alloc(struct perf_cpu_map *map, size_t *size, u16 *type, int *max) 1138 { 1139 size_t size_cpus, size_mask; 1140 bool is_dummy = perf_cpu_map__empty(map); 1141 1142 /* 1143 * Both array and mask data have variable size based 1144 * on the number of cpus and their actual values. 1145 * The size of the 'struct perf_record_cpu_map_data' is: 1146 * 1147 * array = size of 'struct cpu_map_entries' + 1148 * number of cpus * sizeof(u64) 1149 * 1150 * mask = size of 'struct perf_record_record_cpu_map' + 1151 * maximum cpu bit converted to size of longs 1152 * 1153 * and finaly + the size of 'struct perf_record_cpu_map_data'. 1154 */ 1155 size_cpus = cpus_size(map); 1156 size_mask = mask_size(map, max); 1157 1158 if (is_dummy || (size_cpus < size_mask)) { 1159 *size += size_cpus; 1160 *type = PERF_CPU_MAP__CPUS; 1161 } else { 1162 *size += size_mask; 1163 *type = PERF_CPU_MAP__MASK; 1164 } 1165 1166 *size += sizeof(struct perf_record_cpu_map_data); 1167 *size = PERF_ALIGN(*size, sizeof(u64)); 1168 return zalloc(*size); 1169 } 1170 1171 void cpu_map_data__synthesize(struct perf_record_cpu_map_data *data, struct perf_cpu_map *map, 1172 u16 type, int max) 1173 { 1174 data->type = type; 1175 1176 switch (type) { 1177 case PERF_CPU_MAP__CPUS: 1178 synthesize_cpus((struct cpu_map_entries *) data->data, map); 1179 break; 1180 case PERF_CPU_MAP__MASK: 1181 synthesize_mask((struct perf_record_record_cpu_map *)data->data, map, max); 1182 default: 1183 break; 1184 } 1185 } 1186 1187 static struct perf_record_cpu_map *cpu_map_event__new(struct perf_cpu_map *map) 1188 { 1189 size_t size = sizeof(struct perf_record_cpu_map); 1190 struct perf_record_cpu_map *event; 1191 int max; 1192 u16 type; 1193 1194 event = cpu_map_data__alloc(map, &size, &type, &max); 1195 if (!event) 1196 return NULL; 1197 1198 event->header.type = PERF_RECORD_CPU_MAP; 1199 event->header.size = size; 1200 event->data.type = type; 1201 1202 cpu_map_data__synthesize(&event->data, map, type, max); 1203 return event; 1204 } 1205 1206 int perf_event__synthesize_cpu_map(struct perf_tool *tool, 1207 struct perf_cpu_map *map, 1208 perf_event__handler_t process, 1209 struct machine *machine) 1210 { 1211 struct perf_record_cpu_map *event; 1212 int err; 1213 1214 event = cpu_map_event__new(map); 1215 if (!event) 1216 return -ENOMEM; 1217 1218 err = process(tool, (union perf_event *) event, NULL, machine); 1219 1220 free(event); 1221 return err; 1222 } 1223 1224 int perf_event__synthesize_stat_config(struct perf_tool *tool, 1225 struct perf_stat_config *config, 1226 perf_event__handler_t process, 1227 struct machine *machine) 1228 { 1229 struct perf_record_stat_config *event; 1230 int size, i = 0, err; 1231 1232 size = sizeof(*event); 1233 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0])); 1234 1235 event = zalloc(size); 1236 if (!event) 1237 return -ENOMEM; 1238 1239 event->header.type = PERF_RECORD_STAT_CONFIG; 1240 event->header.size = size; 1241 event->nr = PERF_STAT_CONFIG_TERM__MAX; 1242 1243 #define ADD(__term, __val) \ 1244 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \ 1245 event->data[i].val = __val; \ 1246 i++; 1247 1248 ADD(AGGR_MODE, config->aggr_mode) 1249 ADD(INTERVAL, config->interval) 1250 ADD(SCALE, config->scale) 1251 1252 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX, 1253 "stat config terms unbalanced\n"); 1254 #undef ADD 1255 1256 err = process(tool, (union perf_event *) event, NULL, machine); 1257 1258 free(event); 1259 return err; 1260 } 1261 1262 int perf_event__synthesize_stat(struct perf_tool *tool, 1263 u32 cpu, u32 thread, u64 id, 1264 struct perf_counts_values *count, 1265 perf_event__handler_t process, 1266 struct machine *machine) 1267 { 1268 struct perf_record_stat event; 1269 1270 event.header.type = PERF_RECORD_STAT; 1271 event.header.size = sizeof(event); 1272 event.header.misc = 0; 1273 1274 event.id = id; 1275 event.cpu = cpu; 1276 event.thread = thread; 1277 event.val = count->val; 1278 event.ena = count->ena; 1279 event.run = count->run; 1280 1281 return process(tool, (union perf_event *) &event, NULL, machine); 1282 } 1283 1284 int perf_event__synthesize_stat_round(struct perf_tool *tool, 1285 u64 evtime, u64 type, 1286 perf_event__handler_t process, 1287 struct machine *machine) 1288 { 1289 struct perf_record_stat_round event; 1290 1291 event.header.type = PERF_RECORD_STAT_ROUND; 1292 event.header.size = sizeof(event); 1293 event.header.misc = 0; 1294 1295 event.time = evtime; 1296 event.type = type; 1297 1298 return process(tool, (union perf_event *) &event, NULL, machine); 1299 } 1300 1301 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format) 1302 { 1303 size_t sz, result = sizeof(struct perf_record_sample); 1304 1305 if (type & PERF_SAMPLE_IDENTIFIER) 1306 result += sizeof(u64); 1307 1308 if (type & PERF_SAMPLE_IP) 1309 result += sizeof(u64); 1310 1311 if (type & PERF_SAMPLE_TID) 1312 result += sizeof(u64); 1313 1314 if (type & PERF_SAMPLE_TIME) 1315 result += sizeof(u64); 1316 1317 if (type & PERF_SAMPLE_ADDR) 1318 result += sizeof(u64); 1319 1320 if (type & PERF_SAMPLE_ID) 1321 result += sizeof(u64); 1322 1323 if (type & PERF_SAMPLE_STREAM_ID) 1324 result += sizeof(u64); 1325 1326 if (type & PERF_SAMPLE_CPU) 1327 result += sizeof(u64); 1328 1329 if (type & PERF_SAMPLE_PERIOD) 1330 result += sizeof(u64); 1331 1332 if (type & PERF_SAMPLE_READ) { 1333 result += sizeof(u64); 1334 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1335 result += sizeof(u64); 1336 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1337 result += sizeof(u64); 1338 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1339 if (read_format & PERF_FORMAT_GROUP) { 1340 sz = sample->read.group.nr * 1341 sizeof(struct sample_read_value); 1342 result += sz; 1343 } else { 1344 result += sizeof(u64); 1345 } 1346 } 1347 1348 if (type & PERF_SAMPLE_CALLCHAIN) { 1349 sz = (sample->callchain->nr + 1) * sizeof(u64); 1350 result += sz; 1351 } 1352 1353 if (type & PERF_SAMPLE_RAW) { 1354 result += sizeof(u32); 1355 result += sample->raw_size; 1356 } 1357 1358 if (type & PERF_SAMPLE_BRANCH_STACK) { 1359 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1360 /* nr, hw_idx */ 1361 sz += 2 * sizeof(u64); 1362 result += sz; 1363 } 1364 1365 if (type & PERF_SAMPLE_REGS_USER) { 1366 if (sample->user_regs.abi) { 1367 result += sizeof(u64); 1368 sz = hweight64(sample->user_regs.mask) * sizeof(u64); 1369 result += sz; 1370 } else { 1371 result += sizeof(u64); 1372 } 1373 } 1374 1375 if (type & PERF_SAMPLE_STACK_USER) { 1376 sz = sample->user_stack.size; 1377 result += sizeof(u64); 1378 if (sz) { 1379 result += sz; 1380 result += sizeof(u64); 1381 } 1382 } 1383 1384 if (type & PERF_SAMPLE_WEIGHT) 1385 result += sizeof(u64); 1386 1387 if (type & PERF_SAMPLE_DATA_SRC) 1388 result += sizeof(u64); 1389 1390 if (type & PERF_SAMPLE_TRANSACTION) 1391 result += sizeof(u64); 1392 1393 if (type & PERF_SAMPLE_REGS_INTR) { 1394 if (sample->intr_regs.abi) { 1395 result += sizeof(u64); 1396 sz = hweight64(sample->intr_regs.mask) * sizeof(u64); 1397 result += sz; 1398 } else { 1399 result += sizeof(u64); 1400 } 1401 } 1402 1403 if (type & PERF_SAMPLE_PHYS_ADDR) 1404 result += sizeof(u64); 1405 1406 if (type & PERF_SAMPLE_CGROUP) 1407 result += sizeof(u64); 1408 1409 if (type & PERF_SAMPLE_AUX) { 1410 result += sizeof(u64); 1411 result += sample->aux_sample.size; 1412 } 1413 1414 return result; 1415 } 1416 1417 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format, 1418 const struct perf_sample *sample) 1419 { 1420 __u64 *array; 1421 size_t sz; 1422 /* 1423 * used for cross-endian analysis. See git commit 65014ab3 1424 * for why this goofiness is needed. 1425 */ 1426 union u64_swap u; 1427 1428 array = event->sample.array; 1429 1430 if (type & PERF_SAMPLE_IDENTIFIER) { 1431 *array = sample->id; 1432 array++; 1433 } 1434 1435 if (type & PERF_SAMPLE_IP) { 1436 *array = sample->ip; 1437 array++; 1438 } 1439 1440 if (type & PERF_SAMPLE_TID) { 1441 u.val32[0] = sample->pid; 1442 u.val32[1] = sample->tid; 1443 *array = u.val64; 1444 array++; 1445 } 1446 1447 if (type & PERF_SAMPLE_TIME) { 1448 *array = sample->time; 1449 array++; 1450 } 1451 1452 if (type & PERF_SAMPLE_ADDR) { 1453 *array = sample->addr; 1454 array++; 1455 } 1456 1457 if (type & PERF_SAMPLE_ID) { 1458 *array = sample->id; 1459 array++; 1460 } 1461 1462 if (type & PERF_SAMPLE_STREAM_ID) { 1463 *array = sample->stream_id; 1464 array++; 1465 } 1466 1467 if (type & PERF_SAMPLE_CPU) { 1468 u.val32[0] = sample->cpu; 1469 u.val32[1] = 0; 1470 *array = u.val64; 1471 array++; 1472 } 1473 1474 if (type & PERF_SAMPLE_PERIOD) { 1475 *array = sample->period; 1476 array++; 1477 } 1478 1479 if (type & PERF_SAMPLE_READ) { 1480 if (read_format & PERF_FORMAT_GROUP) 1481 *array = sample->read.group.nr; 1482 else 1483 *array = sample->read.one.value; 1484 array++; 1485 1486 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 1487 *array = sample->read.time_enabled; 1488 array++; 1489 } 1490 1491 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 1492 *array = sample->read.time_running; 1493 array++; 1494 } 1495 1496 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1497 if (read_format & PERF_FORMAT_GROUP) { 1498 sz = sample->read.group.nr * 1499 sizeof(struct sample_read_value); 1500 memcpy(array, sample->read.group.values, sz); 1501 array = (void *)array + sz; 1502 } else { 1503 *array = sample->read.one.id; 1504 array++; 1505 } 1506 } 1507 1508 if (type & PERF_SAMPLE_CALLCHAIN) { 1509 sz = (sample->callchain->nr + 1) * sizeof(u64); 1510 memcpy(array, sample->callchain, sz); 1511 array = (void *)array + sz; 1512 } 1513 1514 if (type & PERF_SAMPLE_RAW) { 1515 u.val32[0] = sample->raw_size; 1516 *array = u.val64; 1517 array = (void *)array + sizeof(u32); 1518 1519 memcpy(array, sample->raw_data, sample->raw_size); 1520 array = (void *)array + sample->raw_size; 1521 } 1522 1523 if (type & PERF_SAMPLE_BRANCH_STACK) { 1524 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1525 /* nr, hw_idx */ 1526 sz += 2 * sizeof(u64); 1527 memcpy(array, sample->branch_stack, sz); 1528 array = (void *)array + sz; 1529 } 1530 1531 if (type & PERF_SAMPLE_REGS_USER) { 1532 if (sample->user_regs.abi) { 1533 *array++ = sample->user_regs.abi; 1534 sz = hweight64(sample->user_regs.mask) * sizeof(u64); 1535 memcpy(array, sample->user_regs.regs, sz); 1536 array = (void *)array + sz; 1537 } else { 1538 *array++ = 0; 1539 } 1540 } 1541 1542 if (type & PERF_SAMPLE_STACK_USER) { 1543 sz = sample->user_stack.size; 1544 *array++ = sz; 1545 if (sz) { 1546 memcpy(array, sample->user_stack.data, sz); 1547 array = (void *)array + sz; 1548 *array++ = sz; 1549 } 1550 } 1551 1552 if (type & PERF_SAMPLE_WEIGHT) { 1553 *array = sample->weight; 1554 array++; 1555 } 1556 1557 if (type & PERF_SAMPLE_DATA_SRC) { 1558 *array = sample->data_src; 1559 array++; 1560 } 1561 1562 if (type & PERF_SAMPLE_TRANSACTION) { 1563 *array = sample->transaction; 1564 array++; 1565 } 1566 1567 if (type & PERF_SAMPLE_REGS_INTR) { 1568 if (sample->intr_regs.abi) { 1569 *array++ = sample->intr_regs.abi; 1570 sz = hweight64(sample->intr_regs.mask) * sizeof(u64); 1571 memcpy(array, sample->intr_regs.regs, sz); 1572 array = (void *)array + sz; 1573 } else { 1574 *array++ = 0; 1575 } 1576 } 1577 1578 if (type & PERF_SAMPLE_PHYS_ADDR) { 1579 *array = sample->phys_addr; 1580 array++; 1581 } 1582 1583 if (type & PERF_SAMPLE_CGROUP) { 1584 *array = sample->cgroup; 1585 array++; 1586 } 1587 1588 if (type & PERF_SAMPLE_AUX) { 1589 sz = sample->aux_sample.size; 1590 *array++ = sz; 1591 memcpy(array, sample->aux_sample.data, sz); 1592 array = (void *)array + sz; 1593 } 1594 1595 return 0; 1596 } 1597 1598 int perf_event__synthesize_id_index(struct perf_tool *tool, perf_event__handler_t process, 1599 struct evlist *evlist, struct machine *machine) 1600 { 1601 union perf_event *ev; 1602 struct evsel *evsel; 1603 size_t nr = 0, i = 0, sz, max_nr, n; 1604 int err; 1605 1606 pr_debug2("Synthesizing id index\n"); 1607 1608 max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) / 1609 sizeof(struct id_index_entry); 1610 1611 evlist__for_each_entry(evlist, evsel) 1612 nr += evsel->core.ids; 1613 1614 n = nr > max_nr ? max_nr : nr; 1615 sz = sizeof(struct perf_record_id_index) + n * sizeof(struct id_index_entry); 1616 ev = zalloc(sz); 1617 if (!ev) 1618 return -ENOMEM; 1619 1620 ev->id_index.header.type = PERF_RECORD_ID_INDEX; 1621 ev->id_index.header.size = sz; 1622 ev->id_index.nr = n; 1623 1624 evlist__for_each_entry(evlist, evsel) { 1625 u32 j; 1626 1627 for (j = 0; j < evsel->core.ids; j++) { 1628 struct id_index_entry *e; 1629 struct perf_sample_id *sid; 1630 1631 if (i >= n) { 1632 err = process(tool, ev, NULL, machine); 1633 if (err) 1634 goto out_err; 1635 nr -= n; 1636 i = 0; 1637 } 1638 1639 e = &ev->id_index.entries[i++]; 1640 1641 e->id = evsel->core.id[j]; 1642 1643 sid = perf_evlist__id2sid(evlist, e->id); 1644 if (!sid) { 1645 free(ev); 1646 return -ENOENT; 1647 } 1648 1649 e->idx = sid->idx; 1650 e->cpu = sid->cpu; 1651 e->tid = sid->tid; 1652 } 1653 } 1654 1655 sz = sizeof(struct perf_record_id_index) + nr * sizeof(struct id_index_entry); 1656 ev->id_index.header.size = sz; 1657 ev->id_index.nr = nr; 1658 1659 err = process(tool, ev, NULL, machine); 1660 out_err: 1661 free(ev); 1662 1663 return err; 1664 } 1665 1666 int __machine__synthesize_threads(struct machine *machine, struct perf_tool *tool, 1667 struct target *target, struct perf_thread_map *threads, 1668 perf_event__handler_t process, bool data_mmap, 1669 unsigned int nr_threads_synthesize) 1670 { 1671 if (target__has_task(target)) 1672 return perf_event__synthesize_thread_map(tool, threads, process, machine, data_mmap); 1673 else if (target__has_cpu(target)) 1674 return perf_event__synthesize_threads(tool, process, 1675 machine, data_mmap, 1676 nr_threads_synthesize); 1677 /* command specified */ 1678 return 0; 1679 } 1680 1681 int machine__synthesize_threads(struct machine *machine, struct target *target, 1682 struct perf_thread_map *threads, bool data_mmap, 1683 unsigned int nr_threads_synthesize) 1684 { 1685 return __machine__synthesize_threads(machine, NULL, target, threads, 1686 perf_event__process, data_mmap, 1687 nr_threads_synthesize); 1688 } 1689 1690 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id) 1691 { 1692 struct perf_record_event_update *ev; 1693 1694 size += sizeof(*ev); 1695 size = PERF_ALIGN(size, sizeof(u64)); 1696 1697 ev = zalloc(size); 1698 if (ev) { 1699 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1700 ev->header.size = (u16)size; 1701 ev->type = type; 1702 ev->id = id; 1703 } 1704 return ev; 1705 } 1706 1707 int perf_event__synthesize_event_update_unit(struct perf_tool *tool, struct evsel *evsel, 1708 perf_event__handler_t process) 1709 { 1710 size_t size = strlen(evsel->unit); 1711 struct perf_record_event_update *ev; 1712 int err; 1713 1714 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]); 1715 if (ev == NULL) 1716 return -ENOMEM; 1717 1718 strlcpy(ev->data, evsel->unit, size + 1); 1719 err = process(tool, (union perf_event *)ev, NULL, NULL); 1720 free(ev); 1721 return err; 1722 } 1723 1724 int perf_event__synthesize_event_update_scale(struct perf_tool *tool, struct evsel *evsel, 1725 perf_event__handler_t process) 1726 { 1727 struct perf_record_event_update *ev; 1728 struct perf_record_event_update_scale *ev_data; 1729 int err; 1730 1731 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]); 1732 if (ev == NULL) 1733 return -ENOMEM; 1734 1735 ev_data = (struct perf_record_event_update_scale *)ev->data; 1736 ev_data->scale = evsel->scale; 1737 err = process(tool, (union perf_event *)ev, NULL, NULL); 1738 free(ev); 1739 return err; 1740 } 1741 1742 int perf_event__synthesize_event_update_name(struct perf_tool *tool, struct evsel *evsel, 1743 perf_event__handler_t process) 1744 { 1745 struct perf_record_event_update *ev; 1746 size_t len = strlen(evsel->name); 1747 int err; 1748 1749 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]); 1750 if (ev == NULL) 1751 return -ENOMEM; 1752 1753 strlcpy(ev->data, evsel->name, len + 1); 1754 err = process(tool, (union perf_event *)ev, NULL, NULL); 1755 free(ev); 1756 return err; 1757 } 1758 1759 int perf_event__synthesize_event_update_cpus(struct perf_tool *tool, struct evsel *evsel, 1760 perf_event__handler_t process) 1761 { 1762 size_t size = sizeof(struct perf_record_event_update); 1763 struct perf_record_event_update *ev; 1764 int max, err; 1765 u16 type; 1766 1767 if (!evsel->core.own_cpus) 1768 return 0; 1769 1770 ev = cpu_map_data__alloc(evsel->core.own_cpus, &size, &type, &max); 1771 if (!ev) 1772 return -ENOMEM; 1773 1774 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1775 ev->header.size = (u16)size; 1776 ev->type = PERF_EVENT_UPDATE__CPUS; 1777 ev->id = evsel->core.id[0]; 1778 1779 cpu_map_data__synthesize((struct perf_record_cpu_map_data *)ev->data, 1780 evsel->core.own_cpus, type, max); 1781 1782 err = process(tool, (union perf_event *)ev, NULL, NULL); 1783 free(ev); 1784 return err; 1785 } 1786 1787 int perf_event__synthesize_attrs(struct perf_tool *tool, struct evlist *evlist, 1788 perf_event__handler_t process) 1789 { 1790 struct evsel *evsel; 1791 int err = 0; 1792 1793 evlist__for_each_entry(evlist, evsel) { 1794 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids, 1795 evsel->core.id, process); 1796 if (err) { 1797 pr_debug("failed to create perf header attribute\n"); 1798 return err; 1799 } 1800 } 1801 1802 return err; 1803 } 1804 1805 static bool has_unit(struct evsel *evsel) 1806 { 1807 return evsel->unit && *evsel->unit; 1808 } 1809 1810 static bool has_scale(struct evsel *evsel) 1811 { 1812 return evsel->scale != 1; 1813 } 1814 1815 int perf_event__synthesize_extra_attr(struct perf_tool *tool, struct evlist *evsel_list, 1816 perf_event__handler_t process, bool is_pipe) 1817 { 1818 struct evsel *evsel; 1819 int err; 1820 1821 /* 1822 * Synthesize other events stuff not carried within 1823 * attr event - unit, scale, name 1824 */ 1825 evlist__for_each_entry(evsel_list, evsel) { 1826 if (!evsel->supported) 1827 continue; 1828 1829 /* 1830 * Synthesize unit and scale only if it's defined. 1831 */ 1832 if (has_unit(evsel)) { 1833 err = perf_event__synthesize_event_update_unit(tool, evsel, process); 1834 if (err < 0) { 1835 pr_err("Couldn't synthesize evsel unit.\n"); 1836 return err; 1837 } 1838 } 1839 1840 if (has_scale(evsel)) { 1841 err = perf_event__synthesize_event_update_scale(tool, evsel, process); 1842 if (err < 0) { 1843 pr_err("Couldn't synthesize evsel evsel.\n"); 1844 return err; 1845 } 1846 } 1847 1848 if (evsel->core.own_cpus) { 1849 err = perf_event__synthesize_event_update_cpus(tool, evsel, process); 1850 if (err < 0) { 1851 pr_err("Couldn't synthesize evsel cpus.\n"); 1852 return err; 1853 } 1854 } 1855 1856 /* 1857 * Name is needed only for pipe output, 1858 * perf.data carries event names. 1859 */ 1860 if (is_pipe) { 1861 err = perf_event__synthesize_event_update_name(tool, evsel, process); 1862 if (err < 0) { 1863 pr_err("Couldn't synthesize evsel name.\n"); 1864 return err; 1865 } 1866 } 1867 } 1868 return 0; 1869 } 1870 1871 int perf_event__synthesize_attr(struct perf_tool *tool, struct perf_event_attr *attr, 1872 u32 ids, u64 *id, perf_event__handler_t process) 1873 { 1874 union perf_event *ev; 1875 size_t size; 1876 int err; 1877 1878 size = sizeof(struct perf_event_attr); 1879 size = PERF_ALIGN(size, sizeof(u64)); 1880 size += sizeof(struct perf_event_header); 1881 size += ids * sizeof(u64); 1882 1883 ev = zalloc(size); 1884 1885 if (ev == NULL) 1886 return -ENOMEM; 1887 1888 ev->attr.attr = *attr; 1889 memcpy(ev->attr.id, id, ids * sizeof(u64)); 1890 1891 ev->attr.header.type = PERF_RECORD_HEADER_ATTR; 1892 ev->attr.header.size = (u16)size; 1893 1894 if (ev->attr.header.size == size) 1895 err = process(tool, ev, NULL, NULL); 1896 else 1897 err = -E2BIG; 1898 1899 free(ev); 1900 1901 return err; 1902 } 1903 1904 int perf_event__synthesize_tracing_data(struct perf_tool *tool, int fd, struct evlist *evlist, 1905 perf_event__handler_t process) 1906 { 1907 union perf_event ev; 1908 struct tracing_data *tdata; 1909 ssize_t size = 0, aligned_size = 0, padding; 1910 struct feat_fd ff; 1911 1912 /* 1913 * We are going to store the size of the data followed 1914 * by the data contents. Since the fd descriptor is a pipe, 1915 * we cannot seek back to store the size of the data once 1916 * we know it. Instead we: 1917 * 1918 * - write the tracing data to the temp file 1919 * - get/write the data size to pipe 1920 * - write the tracing data from the temp file 1921 * to the pipe 1922 */ 1923 tdata = tracing_data_get(&evlist->core.entries, fd, true); 1924 if (!tdata) 1925 return -1; 1926 1927 memset(&ev, 0, sizeof(ev)); 1928 1929 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA; 1930 size = tdata->size; 1931 aligned_size = PERF_ALIGN(size, sizeof(u64)); 1932 padding = aligned_size - size; 1933 ev.tracing_data.header.size = sizeof(ev.tracing_data); 1934 ev.tracing_data.size = aligned_size; 1935 1936 process(tool, &ev, NULL, NULL); 1937 1938 /* 1939 * The put function will copy all the tracing data 1940 * stored in temp file to the pipe. 1941 */ 1942 tracing_data_put(tdata); 1943 1944 ff = (struct feat_fd){ .fd = fd }; 1945 if (write_padded(&ff, NULL, 0, padding)) 1946 return -1; 1947 1948 return aligned_size; 1949 } 1950 1951 int perf_event__synthesize_build_id(struct perf_tool *tool, struct dso *pos, u16 misc, 1952 perf_event__handler_t process, struct machine *machine) 1953 { 1954 union perf_event ev; 1955 size_t len; 1956 1957 if (!pos->hit) 1958 return 0; 1959 1960 memset(&ev, 0, sizeof(ev)); 1961 1962 len = pos->long_name_len + 1; 1963 len = PERF_ALIGN(len, NAME_ALIGN); 1964 memcpy(&ev.build_id.build_id, pos->build_id, sizeof(pos->build_id)); 1965 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID; 1966 ev.build_id.header.misc = misc; 1967 ev.build_id.pid = machine->pid; 1968 ev.build_id.header.size = sizeof(ev.build_id) + len; 1969 memcpy(&ev.build_id.filename, pos->long_name, pos->long_name_len); 1970 1971 return process(tool, &ev, NULL, machine); 1972 } 1973 1974 int perf_event__synthesize_stat_events(struct perf_stat_config *config, struct perf_tool *tool, 1975 struct evlist *evlist, perf_event__handler_t process, bool attrs) 1976 { 1977 int err; 1978 1979 if (attrs) { 1980 err = perf_event__synthesize_attrs(tool, evlist, process); 1981 if (err < 0) { 1982 pr_err("Couldn't synthesize attrs.\n"); 1983 return err; 1984 } 1985 } 1986 1987 err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs); 1988 err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL); 1989 if (err < 0) { 1990 pr_err("Couldn't synthesize thread map.\n"); 1991 return err; 1992 } 1993 1994 err = perf_event__synthesize_cpu_map(tool, evlist->core.cpus, process, NULL); 1995 if (err < 0) { 1996 pr_err("Couldn't synthesize thread map.\n"); 1997 return err; 1998 } 1999 2000 err = perf_event__synthesize_stat_config(tool, config, process, NULL); 2001 if (err < 0) { 2002 pr_err("Couldn't synthesize config.\n"); 2003 return err; 2004 } 2005 2006 return 0; 2007 } 2008 2009 int __weak perf_event__synth_time_conv(const struct perf_event_mmap_page *pc __maybe_unused, 2010 struct perf_tool *tool __maybe_unused, 2011 perf_event__handler_t process __maybe_unused, 2012 struct machine *machine __maybe_unused) 2013 { 2014 return 0; 2015 } 2016 2017 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE]; 2018 2019 int perf_event__synthesize_features(struct perf_tool *tool, struct perf_session *session, 2020 struct evlist *evlist, perf_event__handler_t process) 2021 { 2022 struct perf_header *header = &session->header; 2023 struct perf_record_header_feature *fe; 2024 struct feat_fd ff; 2025 size_t sz, sz_hdr; 2026 int feat, ret; 2027 2028 sz_hdr = sizeof(fe->header); 2029 sz = sizeof(union perf_event); 2030 /* get a nice alignment */ 2031 sz = PERF_ALIGN(sz, page_size); 2032 2033 memset(&ff, 0, sizeof(ff)); 2034 2035 ff.buf = malloc(sz); 2036 if (!ff.buf) 2037 return -ENOMEM; 2038 2039 ff.size = sz - sz_hdr; 2040 ff.ph = &session->header; 2041 2042 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) { 2043 if (!feat_ops[feat].synthesize) { 2044 pr_debug("No record header feature for header :%d\n", feat); 2045 continue; 2046 } 2047 2048 ff.offset = sizeof(*fe); 2049 2050 ret = feat_ops[feat].write(&ff, evlist); 2051 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) { 2052 pr_debug("Error writing feature\n"); 2053 continue; 2054 } 2055 /* ff.buf may have changed due to realloc in do_write() */ 2056 fe = ff.buf; 2057 memset(fe, 0, sizeof(*fe)); 2058 2059 fe->feat_id = feat; 2060 fe->header.type = PERF_RECORD_HEADER_FEATURE; 2061 fe->header.size = ff.offset; 2062 2063 ret = process(tool, ff.buf, NULL, NULL); 2064 if (ret) { 2065 free(ff.buf); 2066 return ret; 2067 } 2068 } 2069 2070 /* Send HEADER_LAST_FEATURE mark. */ 2071 fe = ff.buf; 2072 fe->feat_id = HEADER_LAST_FEATURE; 2073 fe->header.type = PERF_RECORD_HEADER_FEATURE; 2074 fe->header.size = sizeof(*fe); 2075 2076 ret = process(tool, ff.buf, NULL, NULL); 2077 2078 free(ff.buf); 2079 return ret; 2080 } 2081