1 // SPDX-License-Identifier: GPL-2.0-only 2 3 #include "util/cgroup.h" 4 #include "util/data.h" 5 #include "util/debug.h" 6 #include "util/dso.h" 7 #include "util/event.h" 8 #include "util/evlist.h" 9 #include "util/machine.h" 10 #include "util/map.h" 11 #include "util/map_symbol.h" 12 #include "util/branch.h" 13 #include "util/memswap.h" 14 #include "util/namespaces.h" 15 #include "util/session.h" 16 #include "util/stat.h" 17 #include "util/symbol.h" 18 #include "util/synthetic-events.h" 19 #include "util/target.h" 20 #include "util/time-utils.h" 21 #include <linux/bitops.h> 22 #include <linux/kernel.h> 23 #include <linux/string.h> 24 #include <linux/zalloc.h> 25 #include <linux/perf_event.h> 26 #include <asm/bug.h> 27 #include <perf/evsel.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 <api/io_dir.h> 42 #include <sys/types.h> 43 #include <sys/stat.h> 44 #include <fcntl.h> 45 #include <unistd.h> 46 47 #define DEFAULT_PROC_MAP_PARSE_TIMEOUT 500 48 49 unsigned int proc_map_timeout = DEFAULT_PROC_MAP_PARSE_TIMEOUT; 50 51 int perf_tool__process_synth_event(const struct perf_tool *tool, 52 union perf_event *event, 53 struct machine *machine, 54 perf_event__handler_t process) 55 { 56 struct perf_sample synth_sample = { 57 .pid = -1, 58 .tid = -1, 59 .time = -1, 60 .stream_id = -1, 61 .cpu = -1, 62 .period = 1, 63 .cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK, 64 }; 65 66 return process(tool, event, &synth_sample, machine); 67 }; 68 69 /* 70 * Assumes that the first 4095 bytes of /proc/pid/stat contains 71 * the comm, tgid and ppid. 72 */ 73 static int perf_event__get_comm_ids(pid_t pid, pid_t tid, char *comm, size_t len, 74 pid_t *tgid, pid_t *ppid, bool *kernel) 75 { 76 char bf[4096]; 77 int fd; 78 size_t size = 0; 79 ssize_t n; 80 char *name, *tgids, *ppids, *vmpeak, *threads; 81 82 *tgid = -1; 83 *ppid = -1; 84 85 if (pid) 86 snprintf(bf, sizeof(bf), "/proc/%d/task/%d/status", pid, tid); 87 else 88 snprintf(bf, sizeof(bf), "/proc/%d/status", tid); 89 90 fd = open(bf, O_RDONLY); 91 if (fd < 0) { 92 pr_debug("couldn't open %s\n", bf); 93 return -1; 94 } 95 96 n = read(fd, bf, sizeof(bf) - 1); 97 close(fd); 98 if (n <= 0) { 99 pr_warning("Couldn't get COMM, tigd and ppid for pid %d\n", 100 tid); 101 return -1; 102 } 103 bf[n] = '\0'; 104 105 name = strstr(bf, "Name:"); 106 tgids = strstr(name ?: bf, "Tgid:"); 107 ppids = strstr(tgids ?: bf, "PPid:"); 108 vmpeak = strstr(ppids ?: bf, "VmPeak:"); 109 110 if (vmpeak) 111 threads = NULL; 112 else 113 threads = strstr(ppids ?: bf, "Threads:"); 114 115 if (name) { 116 char *nl; 117 118 name = skip_spaces(name + 5); /* strlen("Name:") */ 119 nl = strchr(name, '\n'); 120 if (nl) 121 *nl = '\0'; 122 123 size = strlen(name); 124 if (size >= len) 125 size = len - 1; 126 memcpy(comm, name, size); 127 comm[size] = '\0'; 128 } else { 129 pr_debug("Name: string not found for pid %d\n", tid); 130 } 131 132 if (tgids) { 133 tgids += 5; /* strlen("Tgid:") */ 134 *tgid = atoi(tgids); 135 } else { 136 pr_debug("Tgid: string not found for pid %d\n", tid); 137 } 138 139 if (ppids) { 140 ppids += 5; /* strlen("PPid:") */ 141 *ppid = atoi(ppids); 142 } else { 143 pr_debug("PPid: string not found for pid %d\n", tid); 144 } 145 146 if (!vmpeak && threads) 147 *kernel = true; 148 else 149 *kernel = false; 150 151 return 0; 152 } 153 154 static int perf_event__prepare_comm(union perf_event *event, pid_t pid, pid_t tid, 155 struct machine *machine, 156 pid_t *tgid, pid_t *ppid, bool *kernel) 157 { 158 size_t size; 159 160 *ppid = -1; 161 162 memset(&event->comm, 0, sizeof(event->comm)); 163 164 if (machine__is_host(machine)) { 165 if (perf_event__get_comm_ids(pid, tid, event->comm.comm, 166 sizeof(event->comm.comm), 167 tgid, ppid, kernel) != 0) { 168 return -1; 169 } 170 } else { 171 *tgid = machine->pid; 172 } 173 174 if (*tgid < 0) 175 return -1; 176 177 event->comm.pid = *tgid; 178 event->comm.header.type = PERF_RECORD_COMM; 179 180 size = strlen(event->comm.comm) + 1; 181 size = PERF_ALIGN(size, sizeof(u64)); 182 memset(event->comm.comm + size, 0, machine->id_hdr_size); 183 event->comm.header.size = (sizeof(event->comm) - 184 (sizeof(event->comm.comm) - size) + 185 machine->id_hdr_size); 186 event->comm.tid = tid; 187 188 return 0; 189 } 190 191 pid_t perf_event__synthesize_comm(const struct perf_tool *tool, 192 union perf_event *event, pid_t pid, 193 perf_event__handler_t process, 194 struct machine *machine) 195 { 196 pid_t tgid, ppid; 197 bool kernel_thread; 198 199 if (perf_event__prepare_comm(event, 0, pid, machine, &tgid, &ppid, 200 &kernel_thread) != 0) 201 return -1; 202 203 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 204 return -1; 205 206 return tgid; 207 } 208 209 static void perf_event__get_ns_link_info(pid_t pid, const char *ns, 210 struct perf_ns_link_info *ns_link_info) 211 { 212 struct stat64 st; 213 char proc_ns[128]; 214 215 sprintf(proc_ns, "/proc/%u/ns/%s", pid, ns); 216 if (stat64(proc_ns, &st) == 0) { 217 ns_link_info->dev = st.st_dev; 218 ns_link_info->ino = st.st_ino; 219 } 220 } 221 222 int perf_event__synthesize_namespaces(const struct perf_tool *tool, 223 union perf_event *event, 224 pid_t pid, pid_t tgid, 225 perf_event__handler_t process, 226 struct machine *machine) 227 { 228 u32 idx; 229 struct perf_ns_link_info *ns_link_info; 230 231 if (!tool || !tool->namespace_events) 232 return 0; 233 234 memset(&event->namespaces, 0, (sizeof(event->namespaces) + 235 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 236 machine->id_hdr_size)); 237 238 event->namespaces.pid = tgid; 239 event->namespaces.tid = pid; 240 241 event->namespaces.nr_namespaces = NR_NAMESPACES; 242 243 ns_link_info = event->namespaces.link_info; 244 245 for (idx = 0; idx < event->namespaces.nr_namespaces; idx++) 246 perf_event__get_ns_link_info(pid, perf_ns__name(idx), 247 &ns_link_info[idx]); 248 249 event->namespaces.header.type = PERF_RECORD_NAMESPACES; 250 251 event->namespaces.header.size = (sizeof(event->namespaces) + 252 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 253 machine->id_hdr_size); 254 255 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 256 return -1; 257 258 return 0; 259 } 260 261 static int perf_event__synthesize_fork(const struct perf_tool *tool, 262 union perf_event *event, 263 pid_t pid, pid_t tgid, pid_t ppid, 264 perf_event__handler_t process, 265 struct machine *machine) 266 { 267 memset(&event->fork, 0, sizeof(event->fork) + machine->id_hdr_size); 268 269 /* 270 * for main thread set parent to ppid from status file. For other 271 * threads set parent pid to main thread. ie., assume main thread 272 * spawns all threads in a process 273 */ 274 if (tgid == pid) { 275 event->fork.ppid = ppid; 276 event->fork.ptid = ppid; 277 } else { 278 event->fork.ppid = tgid; 279 event->fork.ptid = tgid; 280 } 281 event->fork.pid = tgid; 282 event->fork.tid = pid; 283 event->fork.header.type = PERF_RECORD_FORK; 284 event->fork.header.misc = PERF_RECORD_MISC_FORK_EXEC; 285 286 event->fork.header.size = (sizeof(event->fork) + machine->id_hdr_size); 287 288 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) 289 return -1; 290 291 return 0; 292 } 293 294 static bool read_proc_maps_line(struct io *io, __u64 *start, __u64 *end, 295 u32 *prot, u32 *flags, __u64 *offset, 296 u32 *maj, u32 *min, 297 __u64 *inode, 298 ssize_t pathname_size, char *pathname) 299 { 300 __u64 temp; 301 int ch; 302 char *start_pathname = pathname; 303 304 if (io__get_hex(io, start) != '-') 305 return false; 306 if (io__get_hex(io, end) != ' ') 307 return false; 308 309 /* map protection and flags bits */ 310 *prot = 0; 311 ch = io__get_char(io); 312 if (ch == 'r') 313 *prot |= PROT_READ; 314 else if (ch != '-') 315 return false; 316 ch = io__get_char(io); 317 if (ch == 'w') 318 *prot |= PROT_WRITE; 319 else if (ch != '-') 320 return false; 321 ch = io__get_char(io); 322 if (ch == 'x') 323 *prot |= PROT_EXEC; 324 else if (ch != '-') 325 return false; 326 ch = io__get_char(io); 327 if (ch == 's') 328 *flags = MAP_SHARED; 329 else if (ch == 'p') 330 *flags = MAP_PRIVATE; 331 else 332 return false; 333 if (io__get_char(io) != ' ') 334 return false; 335 336 if (io__get_hex(io, offset) != ' ') 337 return false; 338 339 if (io__get_hex(io, &temp) != ':') 340 return false; 341 *maj = temp; 342 if (io__get_hex(io, &temp) != ' ') 343 return false; 344 *min = temp; 345 346 ch = io__get_dec(io, inode); 347 if (ch != ' ') { 348 *pathname = '\0'; 349 return ch == '\n'; 350 } 351 do { 352 ch = io__get_char(io); 353 } while (ch == ' '); 354 while (true) { 355 if (ch < 0) 356 return false; 357 if (ch == '\0' || ch == '\n' || 358 (pathname + 1 - start_pathname) >= pathname_size) { 359 *pathname = '\0'; 360 return true; 361 } 362 *pathname++ = ch; 363 ch = io__get_char(io); 364 } 365 } 366 367 static void perf_record_mmap2__read_build_id(struct perf_record_mmap2 *event, 368 struct machine *machine, 369 bool is_kernel) 370 { 371 struct build_id bid = { .size = 0, }; 372 struct nsinfo *nsi; 373 struct nscookie nc; 374 struct dso *dso = NULL; 375 struct dso_id dso_id = dso_id_empty; 376 int rc; 377 378 if (is_kernel) { 379 rc = sysfs__read_build_id("/sys/kernel/notes", &bid); 380 goto out; 381 } 382 383 if (event->header.misc & PERF_RECORD_MISC_MMAP_BUILD_ID) { 384 build_id__init(&dso_id.build_id, event->build_id, event->build_id_size); 385 } else { 386 dso_id.maj = event->maj; 387 dso_id.min = event->min; 388 dso_id.ino = event->ino; 389 dso_id.ino_generation = event->ino_generation; 390 dso_id.mmap2_valid = true; 391 dso_id.mmap2_ino_generation_valid = true; 392 }; 393 394 dso = dsos__findnew_id(&machine->dsos, event->filename, &dso_id); 395 if (dso && dso__has_build_id(dso)) { 396 bid = *dso__bid(dso); 397 rc = 0; 398 goto out; 399 } 400 401 nsi = nsinfo__new(event->pid); 402 nsinfo__mountns_enter(nsi, &nc); 403 404 rc = filename__read_build_id(event->filename, &bid, /*block=*/false) > 0 ? 0 : -1; 405 406 nsinfo__mountns_exit(&nc); 407 nsinfo__put(nsi); 408 409 out: 410 if (rc == 0) { 411 memcpy(event->build_id, bid.data, sizeof(bid.data)); 412 event->build_id_size = (u8) bid.size; 413 event->header.misc |= PERF_RECORD_MISC_MMAP_BUILD_ID; 414 event->__reserved_1 = 0; 415 event->__reserved_2 = 0; 416 417 if (dso && !dso__has_build_id(dso)) 418 dso__set_build_id(dso, &bid); 419 } else { 420 if (event->filename[0] == '/') { 421 pr_debug2("Failed to read build ID for %s\n", 422 event->filename); 423 } 424 } 425 dso__put(dso); 426 } 427 428 int perf_event__synthesize_mmap_events(const struct perf_tool *tool, 429 union perf_event *event, 430 pid_t pid, pid_t tgid, 431 perf_event__handler_t process, 432 struct machine *machine, 433 bool mmap_data) 434 { 435 unsigned long long t; 436 char bf[BUFSIZ]; 437 struct io io; 438 bool truncation = false; 439 unsigned long long timeout = proc_map_timeout * 1000000ULL; 440 int rc = 0; 441 const char *hugetlbfs_mnt = hugetlbfs__mountpoint(); 442 int hugetlbfs_mnt_len = hugetlbfs_mnt ? strlen(hugetlbfs_mnt) : 0; 443 444 if (machine__is_default_guest(machine)) 445 return 0; 446 447 snprintf(bf, sizeof(bf), "%s/proc/%d/task/%d/maps", 448 machine->root_dir, pid, pid); 449 450 io.fd = open(bf, O_RDONLY, 0); 451 if (io.fd < 0) { 452 /* 453 * We raced with a task exiting - just return: 454 */ 455 pr_debug("couldn't open %s\n", bf); 456 return -1; 457 } 458 io__init(&io, io.fd, bf, sizeof(bf)); 459 460 event->header.type = PERF_RECORD_MMAP2; 461 t = rdclock(); 462 463 while (!io.eof) { 464 static const char anonstr[] = "//anon"; 465 size_t size, aligned_size; 466 467 /* ensure null termination since stack will be reused. */ 468 event->mmap2.filename[0] = '\0'; 469 470 /* 00400000-0040c000 r-xp 00000000 fd:01 41038 /bin/cat */ 471 if (!read_proc_maps_line(&io, 472 &event->mmap2.start, 473 &event->mmap2.len, 474 &event->mmap2.prot, 475 &event->mmap2.flags, 476 &event->mmap2.pgoff, 477 &event->mmap2.maj, 478 &event->mmap2.min, 479 &event->mmap2.ino, 480 sizeof(event->mmap2.filename), 481 event->mmap2.filename)) 482 continue; 483 484 if ((rdclock() - t) > timeout) { 485 pr_warning("Reading %s/proc/%d/task/%d/maps time out. " 486 "You may want to increase " 487 "the time limit by --proc-map-timeout\n", 488 machine->root_dir, pid, pid); 489 truncation = true; 490 goto out; 491 } 492 493 event->mmap2.ino_generation = 0; 494 495 /* 496 * Just like the kernel, see __perf_event_mmap in kernel/perf_event.c 497 */ 498 if (machine__is_host(machine)) 499 event->header.misc = PERF_RECORD_MISC_USER; 500 else 501 event->header.misc = PERF_RECORD_MISC_GUEST_USER; 502 503 if ((event->mmap2.prot & PROT_EXEC) == 0) { 504 if (!mmap_data || (event->mmap2.prot & PROT_READ) == 0) 505 continue; 506 507 event->header.misc |= PERF_RECORD_MISC_MMAP_DATA; 508 } 509 510 out: 511 if (truncation) 512 event->header.misc |= PERF_RECORD_MISC_PROC_MAP_PARSE_TIMEOUT; 513 514 if (!strcmp(event->mmap2.filename, "")) 515 strcpy(event->mmap2.filename, anonstr); 516 517 if (hugetlbfs_mnt_len && 518 !strncmp(event->mmap2.filename, hugetlbfs_mnt, 519 hugetlbfs_mnt_len)) { 520 strcpy(event->mmap2.filename, anonstr); 521 event->mmap2.flags |= MAP_HUGETLB; 522 } 523 524 size = strlen(event->mmap2.filename) + 1; 525 aligned_size = PERF_ALIGN(size, sizeof(u64)); 526 event->mmap2.len -= event->mmap.start; 527 event->mmap2.header.size = (sizeof(event->mmap2) - 528 (sizeof(event->mmap2.filename) - aligned_size)); 529 memset(event->mmap2.filename + size, 0, machine->id_hdr_size + 530 (aligned_size - size)); 531 event->mmap2.header.size += machine->id_hdr_size; 532 event->mmap2.pid = tgid; 533 event->mmap2.tid = pid; 534 535 if (!symbol_conf.no_buildid_mmap2) 536 perf_record_mmap2__read_build_id(&event->mmap2, machine, false); 537 538 if (perf_tool__process_synth_event(tool, event, machine, process) != 0) { 539 rc = -1; 540 break; 541 } 542 543 if (truncation) 544 break; 545 } 546 547 close(io.fd); 548 return rc; 549 } 550 551 #ifdef HAVE_FILE_HANDLE 552 static int perf_event__synthesize_cgroup(const struct perf_tool *tool, 553 union perf_event *event, 554 char *path, size_t mount_len, 555 perf_event__handler_t process, 556 struct machine *machine) 557 { 558 size_t event_size = sizeof(event->cgroup) - sizeof(event->cgroup.path); 559 size_t path_len = strlen(path) - mount_len + 1; 560 struct { 561 struct file_handle fh; 562 uint64_t cgroup_id; 563 } handle; 564 int mount_id; 565 566 while (path_len % sizeof(u64)) 567 path[mount_len + path_len++] = '\0'; 568 569 memset(&event->cgroup, 0, event_size); 570 571 event->cgroup.header.type = PERF_RECORD_CGROUP; 572 event->cgroup.header.size = event_size + path_len + machine->id_hdr_size; 573 574 handle.fh.handle_bytes = sizeof(handle.cgroup_id); 575 if (name_to_handle_at(AT_FDCWD, path, &handle.fh, &mount_id, 0) < 0) { 576 pr_debug("stat failed: %s\n", path); 577 return -1; 578 } 579 580 event->cgroup.id = handle.cgroup_id; 581 strncpy(event->cgroup.path, path + mount_len, path_len); 582 memset(event->cgroup.path + path_len, 0, machine->id_hdr_size); 583 584 if (perf_tool__process_synth_event(tool, event, machine, process) < 0) { 585 pr_debug("process synth event failed\n"); 586 return -1; 587 } 588 589 return 0; 590 } 591 592 static int perf_event__walk_cgroup_tree(const struct perf_tool *tool, 593 union perf_event *event, 594 char *path, size_t mount_len, 595 perf_event__handler_t process, 596 struct machine *machine) 597 { 598 size_t pos = strlen(path); 599 DIR *d; 600 struct dirent *dent; 601 int ret = 0; 602 603 if (perf_event__synthesize_cgroup(tool, event, path, mount_len, 604 process, machine) < 0) 605 return -1; 606 607 d = opendir(path); 608 if (d == NULL) { 609 pr_debug("failed to open directory: %s\n", path); 610 return -1; 611 } 612 613 while ((dent = readdir(d)) != NULL) { 614 if (dent->d_type != DT_DIR) 615 continue; 616 if (!strcmp(dent->d_name, ".") || 617 !strcmp(dent->d_name, "..")) 618 continue; 619 620 /* any sane path should be less than PATH_MAX */ 621 if (strlen(path) + strlen(dent->d_name) + 1 >= PATH_MAX) 622 continue; 623 624 if (path[pos - 1] != '/') 625 strcat(path, "/"); 626 strcat(path, dent->d_name); 627 628 ret = perf_event__walk_cgroup_tree(tool, event, path, 629 mount_len, process, machine); 630 if (ret < 0) 631 break; 632 633 path[pos] = '\0'; 634 } 635 636 closedir(d); 637 return ret; 638 } 639 640 int perf_event__synthesize_cgroups(const struct perf_tool *tool, 641 perf_event__handler_t process, 642 struct machine *machine) 643 { 644 union perf_event event; 645 char cgrp_root[PATH_MAX]; 646 size_t mount_len; /* length of mount point in the path */ 647 648 if (!tool || !tool->cgroup_events) 649 return 0; 650 651 if (cgroupfs_find_mountpoint(cgrp_root, PATH_MAX, "perf_event") < 0) { 652 pr_debug("cannot find cgroup mount point\n"); 653 return -1; 654 } 655 656 mount_len = strlen(cgrp_root); 657 /* make sure the path starts with a slash (after mount point) */ 658 strcat(cgrp_root, "/"); 659 660 if (perf_event__walk_cgroup_tree(tool, &event, cgrp_root, mount_len, 661 process, machine) < 0) 662 return -1; 663 664 return 0; 665 } 666 #else 667 int perf_event__synthesize_cgroups(const struct perf_tool *tool __maybe_unused, 668 perf_event__handler_t process __maybe_unused, 669 struct machine *machine __maybe_unused) 670 { 671 return -1; 672 } 673 #endif 674 675 struct perf_event__synthesize_modules_maps_cb_args { 676 const struct perf_tool *tool; 677 perf_event__handler_t process; 678 struct machine *machine; 679 union perf_event *event; 680 }; 681 682 static int perf_event__synthesize_modules_maps_cb(struct map *map, void *data) 683 { 684 struct perf_event__synthesize_modules_maps_cb_args *args = data; 685 union perf_event *event = args->event; 686 struct dso *dso; 687 size_t size; 688 689 if (!__map__is_kmodule(map)) 690 return 0; 691 692 dso = map__dso(map); 693 if (!symbol_conf.no_buildid_mmap2) { 694 size = PERF_ALIGN(dso__long_name_len(dso) + 1, sizeof(u64)); 695 event->mmap2.header.type = PERF_RECORD_MMAP2; 696 event->mmap2.header.size = (sizeof(event->mmap2) - 697 (sizeof(event->mmap2.filename) - size)); 698 memset(event->mmap2.filename + size, 0, args->machine->id_hdr_size); 699 event->mmap2.header.size += args->machine->id_hdr_size; 700 event->mmap2.start = map__start(map); 701 event->mmap2.len = map__size(map); 702 event->mmap2.pid = args->machine->pid; 703 704 memcpy(event->mmap2.filename, dso__long_name(dso), dso__long_name_len(dso) + 1); 705 706 perf_record_mmap2__read_build_id(&event->mmap2, args->machine, false); 707 } else { 708 size = PERF_ALIGN(dso__long_name_len(dso) + 1, sizeof(u64)); 709 event->mmap.header.type = PERF_RECORD_MMAP; 710 event->mmap.header.size = (sizeof(event->mmap) - 711 (sizeof(event->mmap.filename) - size)); 712 memset(event->mmap.filename + size, 0, args->machine->id_hdr_size); 713 event->mmap.header.size += args->machine->id_hdr_size; 714 event->mmap.start = map__start(map); 715 event->mmap.len = map__size(map); 716 event->mmap.pid = args->machine->pid; 717 718 memcpy(event->mmap.filename, dso__long_name(dso), dso__long_name_len(dso) + 1); 719 } 720 721 if (perf_tool__process_synth_event(args->tool, event, args->machine, args->process) != 0) 722 return -1; 723 724 return 0; 725 } 726 727 int perf_event__synthesize_modules(const struct perf_tool *tool, perf_event__handler_t process, 728 struct machine *machine) 729 { 730 int rc; 731 struct maps *maps = machine__kernel_maps(machine); 732 struct perf_event__synthesize_modules_maps_cb_args args = { 733 .tool = tool, 734 .process = process, 735 .machine = machine, 736 }; 737 size_t size = symbol_conf.no_buildid_mmap2 738 ? sizeof(args.event->mmap) 739 : sizeof(args.event->mmap2); 740 741 args.event = zalloc(size + machine->id_hdr_size); 742 if (args.event == NULL) { 743 pr_debug("Not enough memory synthesizing mmap event " 744 "for kernel modules\n"); 745 return -1; 746 } 747 748 /* 749 * kernel uses 0 for user space maps, see kernel/perf_event.c 750 * __perf_event_mmap 751 */ 752 if (machine__is_host(machine)) 753 args.event->header.misc = PERF_RECORD_MISC_KERNEL; 754 else 755 args.event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 756 757 rc = maps__for_each_map(maps, perf_event__synthesize_modules_maps_cb, &args); 758 759 free(args.event); 760 return rc; 761 } 762 763 static int filter_task(const struct dirent *dirent) 764 { 765 return isdigit(dirent->d_name[0]); 766 } 767 768 static int __event__synthesize_thread(union perf_event *comm_event, 769 union perf_event *mmap_event, 770 union perf_event *fork_event, 771 union perf_event *namespaces_event, 772 pid_t pid, int full, perf_event__handler_t process, 773 const struct perf_tool *tool, struct machine *machine, 774 bool needs_mmap, bool mmap_data) 775 { 776 char filename[PATH_MAX]; 777 struct io_dir iod; 778 struct io_dirent64 *dent; 779 pid_t tgid, ppid; 780 int rc = 0; 781 782 /* special case: only send one comm event using passed in pid */ 783 if (!full) { 784 tgid = perf_event__synthesize_comm(tool, comm_event, pid, 785 process, machine); 786 787 if (tgid == -1) 788 return -1; 789 790 if (perf_event__synthesize_namespaces(tool, namespaces_event, pid, 791 tgid, process, machine) < 0) 792 return -1; 793 794 /* 795 * send mmap only for thread group leader 796 * see thread__init_maps() 797 */ 798 if (pid == tgid && needs_mmap && 799 perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 800 process, machine, mmap_data)) 801 return -1; 802 803 return 0; 804 } 805 806 if (machine__is_default_guest(machine)) 807 return 0; 808 809 snprintf(filename, sizeof(filename), "%s/proc/%d/task", 810 machine->root_dir, pid); 811 812 io_dir__init(&iod, open(filename, O_CLOEXEC | O_DIRECTORY | O_RDONLY)); 813 if (iod.dirfd < 0) 814 return -1; 815 816 while ((dent = io_dir__readdir(&iod)) != NULL) { 817 char *end; 818 pid_t _pid; 819 bool kernel_thread = false; 820 821 if (!isdigit(dent->d_name[0])) 822 continue; 823 824 _pid = strtol(dent->d_name, &end, 10); 825 if (*end) 826 continue; 827 828 /* some threads may exit just after scan, ignore it */ 829 if (perf_event__prepare_comm(comm_event, pid, _pid, machine, 830 &tgid, &ppid, &kernel_thread) != 0) 831 continue; 832 833 rc = -1; 834 if (perf_event__synthesize_fork(tool, fork_event, _pid, tgid, 835 ppid, process, machine) < 0) 836 break; 837 838 if (perf_event__synthesize_namespaces(tool, namespaces_event, _pid, 839 tgid, process, machine) < 0) 840 break; 841 842 /* 843 * Send the prepared comm event 844 */ 845 if (perf_tool__process_synth_event(tool, comm_event, machine, process) != 0) 846 break; 847 848 rc = 0; 849 if (_pid == pid && !kernel_thread && needs_mmap) { 850 /* process the parent's maps too */ 851 rc = perf_event__synthesize_mmap_events(tool, mmap_event, pid, tgid, 852 process, machine, mmap_data); 853 if (rc) 854 break; 855 } 856 } 857 858 close(iod.dirfd); 859 860 return rc; 861 } 862 863 int perf_event__synthesize_thread_map(const struct perf_tool *tool, 864 struct perf_thread_map *threads, 865 perf_event__handler_t process, 866 struct machine *machine, 867 bool needs_mmap, bool mmap_data) 868 { 869 union perf_event *comm_event, *mmap_event, *fork_event; 870 union perf_event *namespaces_event; 871 int err = -1, thread, j; 872 873 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 874 if (comm_event == NULL) 875 goto out; 876 877 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 878 if (mmap_event == NULL) 879 goto out_free_comm; 880 881 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 882 if (fork_event == NULL) 883 goto out_free_mmap; 884 885 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 886 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 887 machine->id_hdr_size); 888 if (namespaces_event == NULL) 889 goto out_free_fork; 890 891 err = 0; 892 for (thread = 0; thread < threads->nr; ++thread) { 893 if (__event__synthesize_thread(comm_event, mmap_event, 894 fork_event, namespaces_event, 895 perf_thread_map__pid(threads, thread), 0, 896 process, tool, machine, 897 needs_mmap, mmap_data)) { 898 err = -1; 899 break; 900 } 901 902 /* 903 * comm.pid is set to thread group id by 904 * perf_event__synthesize_comm 905 */ 906 if ((int) comm_event->comm.pid != perf_thread_map__pid(threads, thread)) { 907 bool need_leader = true; 908 909 /* is thread group leader in thread_map? */ 910 for (j = 0; j < threads->nr; ++j) { 911 if ((int) comm_event->comm.pid == perf_thread_map__pid(threads, j)) { 912 need_leader = false; 913 break; 914 } 915 } 916 917 /* if not, generate events for it */ 918 if (need_leader && 919 __event__synthesize_thread(comm_event, mmap_event, 920 fork_event, namespaces_event, 921 comm_event->comm.pid, 0, 922 process, tool, machine, 923 needs_mmap, mmap_data)) { 924 err = -1; 925 break; 926 } 927 } 928 } 929 free(namespaces_event); 930 out_free_fork: 931 free(fork_event); 932 out_free_mmap: 933 free(mmap_event); 934 out_free_comm: 935 free(comm_event); 936 out: 937 return err; 938 } 939 940 static int __perf_event__synthesize_threads(const struct perf_tool *tool, 941 perf_event__handler_t process, 942 struct machine *machine, 943 bool needs_mmap, 944 bool mmap_data, 945 struct dirent **dirent, 946 int start, 947 int num) 948 { 949 union perf_event *comm_event, *mmap_event, *fork_event; 950 union perf_event *namespaces_event; 951 int err = -1; 952 char *end; 953 pid_t pid; 954 int i; 955 956 comm_event = malloc(sizeof(comm_event->comm) + machine->id_hdr_size); 957 if (comm_event == NULL) 958 goto out; 959 960 mmap_event = malloc(sizeof(mmap_event->mmap2) + machine->id_hdr_size); 961 if (mmap_event == NULL) 962 goto out_free_comm; 963 964 fork_event = malloc(sizeof(fork_event->fork) + machine->id_hdr_size); 965 if (fork_event == NULL) 966 goto out_free_mmap; 967 968 namespaces_event = malloc(sizeof(namespaces_event->namespaces) + 969 (NR_NAMESPACES * sizeof(struct perf_ns_link_info)) + 970 machine->id_hdr_size); 971 if (namespaces_event == NULL) 972 goto out_free_fork; 973 974 for (i = start; i < start + num; i++) { 975 if (!isdigit(dirent[i]->d_name[0])) 976 continue; 977 978 pid = (pid_t)strtol(dirent[i]->d_name, &end, 10); 979 /* only interested in proper numerical dirents */ 980 if (*end) 981 continue; 982 /* 983 * We may race with exiting thread, so don't stop just because 984 * one thread couldn't be synthesized. 985 */ 986 __event__synthesize_thread(comm_event, mmap_event, fork_event, 987 namespaces_event, pid, 1, process, 988 tool, machine, needs_mmap, mmap_data); 989 } 990 err = 0; 991 992 free(namespaces_event); 993 out_free_fork: 994 free(fork_event); 995 out_free_mmap: 996 free(mmap_event); 997 out_free_comm: 998 free(comm_event); 999 out: 1000 return err; 1001 } 1002 1003 struct synthesize_threads_arg { 1004 const struct perf_tool *tool; 1005 perf_event__handler_t process; 1006 struct machine *machine; 1007 bool needs_mmap; 1008 bool mmap_data; 1009 struct dirent **dirent; 1010 int num; 1011 int start; 1012 }; 1013 1014 static void *synthesize_threads_worker(void *arg) 1015 { 1016 struct synthesize_threads_arg *args = arg; 1017 1018 __perf_event__synthesize_threads(args->tool, args->process, 1019 args->machine, 1020 args->needs_mmap, args->mmap_data, 1021 args->dirent, 1022 args->start, args->num); 1023 return NULL; 1024 } 1025 1026 int perf_event__synthesize_threads(const struct perf_tool *tool, 1027 perf_event__handler_t process, 1028 struct machine *machine, 1029 bool needs_mmap, bool mmap_data, 1030 unsigned int nr_threads_synthesize) 1031 { 1032 struct synthesize_threads_arg *args = NULL; 1033 pthread_t *synthesize_threads = NULL; 1034 char proc_path[PATH_MAX]; 1035 struct dirent **dirent; 1036 int num_per_thread; 1037 int m, n, i, j; 1038 int thread_nr; 1039 int base = 0; 1040 int err = -1; 1041 1042 1043 if (machine__is_default_guest(machine)) 1044 return 0; 1045 1046 snprintf(proc_path, sizeof(proc_path), "%s/proc", machine->root_dir); 1047 n = scandir(proc_path, &dirent, filter_task, NULL); 1048 if (n < 0) 1049 return err; 1050 1051 if (nr_threads_synthesize == UINT_MAX) 1052 thread_nr = sysconf(_SC_NPROCESSORS_ONLN); 1053 else 1054 thread_nr = nr_threads_synthesize; 1055 1056 if (thread_nr <= 1) { 1057 err = __perf_event__synthesize_threads(tool, process, 1058 machine, 1059 needs_mmap, mmap_data, 1060 dirent, base, n); 1061 goto free_dirent; 1062 } 1063 if (thread_nr > n) 1064 thread_nr = n; 1065 1066 synthesize_threads = calloc(thread_nr, sizeof(pthread_t)); 1067 if (synthesize_threads == NULL) 1068 goto free_dirent; 1069 1070 args = calloc(thread_nr, sizeof(*args)); 1071 if (args == NULL) 1072 goto free_threads; 1073 1074 num_per_thread = n / thread_nr; 1075 m = n % thread_nr; 1076 for (i = 0; i < thread_nr; i++) { 1077 args[i].tool = tool; 1078 args[i].process = process; 1079 args[i].machine = machine; 1080 args[i].needs_mmap = needs_mmap; 1081 args[i].mmap_data = mmap_data; 1082 args[i].dirent = dirent; 1083 } 1084 for (i = 0; i < m; i++) { 1085 args[i].num = num_per_thread + 1; 1086 args[i].start = i * args[i].num; 1087 } 1088 if (i != 0) 1089 base = args[i-1].start + args[i-1].num; 1090 for (j = i; j < thread_nr; j++) { 1091 args[j].num = num_per_thread; 1092 args[j].start = base + (j - i) * args[i].num; 1093 } 1094 1095 for (i = 0; i < thread_nr; i++) { 1096 if (pthread_create(&synthesize_threads[i], NULL, 1097 synthesize_threads_worker, &args[i])) 1098 goto out_join; 1099 } 1100 err = 0; 1101 out_join: 1102 for (i = 0; i < thread_nr; i++) 1103 pthread_join(synthesize_threads[i], NULL); 1104 free(args); 1105 free_threads: 1106 free(synthesize_threads); 1107 free_dirent: 1108 for (i = 0; i < n; i++) 1109 zfree(&dirent[i]); 1110 free(dirent); 1111 1112 return err; 1113 } 1114 1115 int __weak perf_event__synthesize_extra_kmaps(const struct perf_tool *tool __maybe_unused, 1116 perf_event__handler_t process __maybe_unused, 1117 struct machine *machine __maybe_unused) 1118 { 1119 return 0; 1120 } 1121 1122 static int __perf_event__synthesize_kernel_mmap(const struct perf_tool *tool, 1123 perf_event__handler_t process, 1124 struct machine *machine) 1125 { 1126 union perf_event *event; 1127 size_t size = symbol_conf.no_buildid_mmap2 ? 1128 sizeof(event->mmap) : sizeof(event->mmap2); 1129 struct map *map = machine__kernel_map(machine); 1130 struct kmap *kmap; 1131 int err; 1132 1133 if (map == NULL) 1134 return -1; 1135 1136 kmap = map__kmap(map); 1137 if (!kmap->ref_reloc_sym) 1138 return -1; 1139 1140 /* 1141 * We should get this from /sys/kernel/sections/.text, but till that is 1142 * available use this, and after it is use this as a fallback for older 1143 * kernels. 1144 */ 1145 event = zalloc(size + machine->id_hdr_size); 1146 if (event == NULL) { 1147 pr_debug("Not enough memory synthesizing mmap event " 1148 "for kernel modules\n"); 1149 return -1; 1150 } 1151 1152 if (machine__is_host(machine)) { 1153 /* 1154 * kernel uses PERF_RECORD_MISC_USER for user space maps, 1155 * see kernel/perf_event.c __perf_event_mmap 1156 */ 1157 event->header.misc = PERF_RECORD_MISC_KERNEL; 1158 } else { 1159 event->header.misc = PERF_RECORD_MISC_GUEST_KERNEL; 1160 } 1161 1162 if (!symbol_conf.no_buildid_mmap2) { 1163 size = snprintf(event->mmap2.filename, sizeof(event->mmap2.filename), 1164 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1; 1165 size = PERF_ALIGN(size, sizeof(u64)); 1166 event->mmap2.header.type = PERF_RECORD_MMAP2; 1167 event->mmap2.header.size = (sizeof(event->mmap2) - 1168 (sizeof(event->mmap2.filename) - size) + machine->id_hdr_size); 1169 event->mmap2.pgoff = kmap->ref_reloc_sym->addr; 1170 event->mmap2.start = map__start(map); 1171 event->mmap2.len = map__end(map) - event->mmap.start; 1172 event->mmap2.pid = machine->pid; 1173 1174 perf_record_mmap2__read_build_id(&event->mmap2, machine, true); 1175 } else { 1176 size = snprintf(event->mmap.filename, sizeof(event->mmap.filename), 1177 "%s%s", machine->mmap_name, kmap->ref_reloc_sym->name) + 1; 1178 size = PERF_ALIGN(size, sizeof(u64)); 1179 event->mmap.header.type = PERF_RECORD_MMAP; 1180 event->mmap.header.size = (sizeof(event->mmap) - 1181 (sizeof(event->mmap.filename) - size) + machine->id_hdr_size); 1182 event->mmap.pgoff = kmap->ref_reloc_sym->addr; 1183 event->mmap.start = map__start(map); 1184 event->mmap.len = map__end(map) - event->mmap.start; 1185 event->mmap.pid = machine->pid; 1186 } 1187 1188 err = perf_tool__process_synth_event(tool, event, machine, process); 1189 free(event); 1190 1191 return err; 1192 } 1193 1194 int perf_event__synthesize_kernel_mmap(const struct perf_tool *tool, 1195 perf_event__handler_t process, 1196 struct machine *machine) 1197 { 1198 int err; 1199 1200 err = __perf_event__synthesize_kernel_mmap(tool, process, machine); 1201 if (err < 0) 1202 return err; 1203 1204 return perf_event__synthesize_extra_kmaps(tool, process, machine); 1205 } 1206 1207 int perf_event__synthesize_thread_map2(const struct perf_tool *tool, 1208 struct perf_thread_map *threads, 1209 perf_event__handler_t process, 1210 struct machine *machine) 1211 { 1212 union perf_event *event; 1213 int i, err, size; 1214 1215 size = sizeof(event->thread_map); 1216 size += threads->nr * sizeof(event->thread_map.entries[0]); 1217 1218 event = zalloc(size); 1219 if (!event) 1220 return -ENOMEM; 1221 1222 event->header.type = PERF_RECORD_THREAD_MAP; 1223 event->header.size = size; 1224 event->thread_map.nr = threads->nr; 1225 1226 for (i = 0; i < threads->nr; i++) { 1227 struct perf_record_thread_map_entry *entry = &event->thread_map.entries[i]; 1228 char *comm = perf_thread_map__comm(threads, i); 1229 1230 if (!comm) 1231 comm = (char *) ""; 1232 1233 entry->pid = perf_thread_map__pid(threads, i); 1234 strncpy((char *) &entry->comm, comm, sizeof(entry->comm)); 1235 } 1236 1237 err = process(tool, event, NULL, machine); 1238 1239 free(event); 1240 return err; 1241 } 1242 1243 struct synthesize_cpu_map_data { 1244 const struct perf_cpu_map *map; 1245 int nr; 1246 int min_cpu; 1247 int max_cpu; 1248 int has_any_cpu; 1249 int type; 1250 size_t size; 1251 struct perf_record_cpu_map_data *data; 1252 }; 1253 1254 static void synthesize_cpus(struct synthesize_cpu_map_data *data) 1255 { 1256 data->data->type = PERF_CPU_MAP__CPUS; 1257 data->data->cpus_data.nr = data->nr; 1258 for (int i = 0; i < data->nr; i++) 1259 data->data->cpus_data.cpu[i] = perf_cpu_map__cpu(data->map, i).cpu; 1260 } 1261 1262 static void synthesize_mask(struct synthesize_cpu_map_data *data) 1263 { 1264 int idx; 1265 struct perf_cpu cpu; 1266 1267 /* Due to padding, the 4bytes per entry mask variant is always smaller. */ 1268 data->data->type = PERF_CPU_MAP__MASK; 1269 data->data->mask32_data.nr = BITS_TO_U32(data->max_cpu); 1270 data->data->mask32_data.long_size = 4; 1271 1272 perf_cpu_map__for_each_cpu(cpu, idx, data->map) { 1273 int bit_word = cpu.cpu / 32; 1274 u32 bit_mask = 1U << (cpu.cpu & 31); 1275 1276 data->data->mask32_data.mask[bit_word] |= bit_mask; 1277 } 1278 } 1279 1280 static void synthesize_range_cpus(struct synthesize_cpu_map_data *data) 1281 { 1282 data->data->type = PERF_CPU_MAP__RANGE_CPUS; 1283 data->data->range_cpu_data.any_cpu = data->has_any_cpu; 1284 data->data->range_cpu_data.start_cpu = data->min_cpu; 1285 data->data->range_cpu_data.end_cpu = data->max_cpu; 1286 } 1287 1288 static void *cpu_map_data__alloc(struct synthesize_cpu_map_data *syn_data, 1289 size_t header_size) 1290 { 1291 size_t size_cpus, size_mask; 1292 1293 syn_data->nr = perf_cpu_map__nr(syn_data->map); 1294 syn_data->has_any_cpu = (perf_cpu_map__cpu(syn_data->map, 0).cpu == -1) ? 1 : 0; 1295 1296 syn_data->min_cpu = perf_cpu_map__cpu(syn_data->map, syn_data->has_any_cpu).cpu; 1297 syn_data->max_cpu = perf_cpu_map__max(syn_data->map).cpu; 1298 if (syn_data->max_cpu - syn_data->min_cpu + 1 == syn_data->nr - syn_data->has_any_cpu) { 1299 /* A consecutive range of CPUs can be encoded using a range. */ 1300 assert(sizeof(u16) + sizeof(struct perf_record_range_cpu_map) == sizeof(u64)); 1301 syn_data->type = PERF_CPU_MAP__RANGE_CPUS; 1302 syn_data->size = header_size + sizeof(u64); 1303 return zalloc(syn_data->size); 1304 } 1305 1306 size_cpus = sizeof(u16) + sizeof(struct cpu_map_entries) + syn_data->nr * sizeof(u16); 1307 /* Due to padding, the 4bytes per entry mask variant is always smaller. */ 1308 size_mask = sizeof(u16) + sizeof(struct perf_record_mask_cpu_map32) + 1309 BITS_TO_U32(syn_data->max_cpu) * sizeof(__u32); 1310 if (syn_data->has_any_cpu || size_cpus < size_mask) { 1311 /* Follow the CPU map encoding. */ 1312 syn_data->type = PERF_CPU_MAP__CPUS; 1313 syn_data->size = header_size + PERF_ALIGN(size_cpus, sizeof(u64)); 1314 return zalloc(syn_data->size); 1315 } 1316 /* Encode using a bitmask. */ 1317 syn_data->type = PERF_CPU_MAP__MASK; 1318 syn_data->size = header_size + PERF_ALIGN(size_mask, sizeof(u64)); 1319 return zalloc(syn_data->size); 1320 } 1321 1322 static void cpu_map_data__synthesize(struct synthesize_cpu_map_data *data) 1323 { 1324 switch (data->type) { 1325 case PERF_CPU_MAP__CPUS: 1326 synthesize_cpus(data); 1327 break; 1328 case PERF_CPU_MAP__MASK: 1329 synthesize_mask(data); 1330 break; 1331 case PERF_CPU_MAP__RANGE_CPUS: 1332 synthesize_range_cpus(data); 1333 break; 1334 default: 1335 break; 1336 } 1337 } 1338 1339 static struct perf_record_cpu_map *cpu_map_event__new(const struct perf_cpu_map *map) 1340 { 1341 struct synthesize_cpu_map_data syn_data = { .map = map }; 1342 struct perf_record_cpu_map *event; 1343 1344 1345 event = cpu_map_data__alloc(&syn_data, sizeof(struct perf_event_header)); 1346 if (!event) 1347 return NULL; 1348 1349 syn_data.data = &event->data; 1350 event->header.type = PERF_RECORD_CPU_MAP; 1351 event->header.size = syn_data.size; 1352 cpu_map_data__synthesize(&syn_data); 1353 return event; 1354 } 1355 1356 1357 int perf_event__synthesize_cpu_map(const struct perf_tool *tool, 1358 const struct perf_cpu_map *map, 1359 perf_event__handler_t process, 1360 struct machine *machine) 1361 { 1362 struct perf_record_cpu_map *event; 1363 int err; 1364 1365 event = cpu_map_event__new(map); 1366 if (!event) 1367 return -ENOMEM; 1368 1369 err = process(tool, (union perf_event *) event, NULL, machine); 1370 1371 free(event); 1372 return err; 1373 } 1374 1375 int perf_event__synthesize_stat_config(const struct perf_tool *tool, 1376 struct perf_stat_config *config, 1377 perf_event__handler_t process, 1378 struct machine *machine) 1379 { 1380 struct perf_record_stat_config *event; 1381 int size, i = 0, err; 1382 1383 size = sizeof(*event); 1384 size += (PERF_STAT_CONFIG_TERM__MAX * sizeof(event->data[0])); 1385 1386 event = zalloc(size); 1387 if (!event) 1388 return -ENOMEM; 1389 1390 event->header.type = PERF_RECORD_STAT_CONFIG; 1391 event->header.size = size; 1392 event->nr = PERF_STAT_CONFIG_TERM__MAX; 1393 1394 #define ADD(__term, __val) \ 1395 event->data[i].tag = PERF_STAT_CONFIG_TERM__##__term; \ 1396 event->data[i].val = __val; \ 1397 i++; 1398 1399 ADD(AGGR_MODE, config->aggr_mode) 1400 ADD(INTERVAL, config->interval) 1401 ADD(SCALE, config->scale) 1402 ADD(AGGR_LEVEL, config->aggr_level) 1403 1404 WARN_ONCE(i != PERF_STAT_CONFIG_TERM__MAX, 1405 "stat config terms unbalanced\n"); 1406 #undef ADD 1407 1408 err = process(tool, (union perf_event *) event, NULL, machine); 1409 1410 free(event); 1411 return err; 1412 } 1413 1414 int perf_event__synthesize_stat(const struct perf_tool *tool, 1415 struct perf_cpu cpu, u32 thread, u64 id, 1416 struct perf_counts_values *count, 1417 perf_event__handler_t process, 1418 struct machine *machine) 1419 { 1420 struct perf_record_stat event; 1421 1422 event.header.type = PERF_RECORD_STAT; 1423 event.header.size = sizeof(event); 1424 event.header.misc = 0; 1425 1426 event.id = id; 1427 event.cpu = cpu.cpu; 1428 event.thread = thread; 1429 event.val = count->val; 1430 event.ena = count->ena; 1431 event.run = count->run; 1432 1433 return process(tool, (union perf_event *) &event, NULL, machine); 1434 } 1435 1436 int perf_event__synthesize_stat_round(const struct perf_tool *tool, 1437 u64 evtime, u64 type, 1438 perf_event__handler_t process, 1439 struct machine *machine) 1440 { 1441 struct perf_record_stat_round event; 1442 1443 event.header.type = PERF_RECORD_STAT_ROUND; 1444 event.header.size = sizeof(event); 1445 event.header.misc = 0; 1446 1447 event.time = evtime; 1448 event.type = type; 1449 1450 return process(tool, (union perf_event *) &event, NULL, machine); 1451 } 1452 1453 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type, u64 read_format) 1454 { 1455 size_t sz, result = sizeof(struct perf_record_sample); 1456 1457 if (type & PERF_SAMPLE_IDENTIFIER) 1458 result += sizeof(u64); 1459 1460 if (type & PERF_SAMPLE_IP) 1461 result += sizeof(u64); 1462 1463 if (type & PERF_SAMPLE_TID) 1464 result += sizeof(u64); 1465 1466 if (type & PERF_SAMPLE_TIME) 1467 result += sizeof(u64); 1468 1469 if (type & PERF_SAMPLE_ADDR) 1470 result += sizeof(u64); 1471 1472 if (type & PERF_SAMPLE_ID) 1473 result += sizeof(u64); 1474 1475 if (type & PERF_SAMPLE_STREAM_ID) 1476 result += sizeof(u64); 1477 1478 if (type & PERF_SAMPLE_CPU) 1479 result += sizeof(u64); 1480 1481 if (type & PERF_SAMPLE_PERIOD) 1482 result += sizeof(u64); 1483 1484 if (type & PERF_SAMPLE_READ) { 1485 result += sizeof(u64); 1486 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) 1487 result += sizeof(u64); 1488 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) 1489 result += sizeof(u64); 1490 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1491 if (read_format & PERF_FORMAT_GROUP) { 1492 sz = sample_read_value_size(read_format); 1493 result += sz * sample->read.group.nr; 1494 } else { 1495 result += sizeof(u64); 1496 if (read_format & PERF_FORMAT_LOST) 1497 result += sizeof(u64); 1498 } 1499 } 1500 1501 if (type & PERF_SAMPLE_CALLCHAIN) { 1502 sz = (sample->callchain->nr + 1) * sizeof(u64); 1503 result += sz; 1504 } 1505 1506 if (type & PERF_SAMPLE_RAW) { 1507 result += sizeof(u32); 1508 result += sample->raw_size; 1509 } 1510 1511 if (type & PERF_SAMPLE_BRANCH_STACK) { 1512 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1513 /* nr, hw_idx */ 1514 sz += 2 * sizeof(u64); 1515 result += sz; 1516 } 1517 1518 if (type & PERF_SAMPLE_REGS_USER) { 1519 if (sample->user_regs && sample->user_regs->abi) { 1520 result += sizeof(u64); 1521 sz = hweight64(sample->user_regs->mask) * sizeof(u64); 1522 result += sz; 1523 } else { 1524 result += sizeof(u64); 1525 } 1526 } 1527 1528 if (type & PERF_SAMPLE_STACK_USER) { 1529 sz = sample->user_stack.size; 1530 result += sizeof(u64); 1531 if (sz) { 1532 result += sz; 1533 result += sizeof(u64); 1534 } 1535 } 1536 1537 if (type & PERF_SAMPLE_WEIGHT_TYPE) 1538 result += sizeof(u64); 1539 1540 if (type & PERF_SAMPLE_DATA_SRC) 1541 result += sizeof(u64); 1542 1543 if (type & PERF_SAMPLE_TRANSACTION) 1544 result += sizeof(u64); 1545 1546 if (type & PERF_SAMPLE_REGS_INTR) { 1547 if (sample->intr_regs && sample->intr_regs->abi) { 1548 result += sizeof(u64); 1549 sz = hweight64(sample->intr_regs->mask) * sizeof(u64); 1550 result += sz; 1551 } else { 1552 result += sizeof(u64); 1553 } 1554 } 1555 1556 if (type & PERF_SAMPLE_PHYS_ADDR) 1557 result += sizeof(u64); 1558 1559 if (type & PERF_SAMPLE_CGROUP) 1560 result += sizeof(u64); 1561 1562 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) 1563 result += sizeof(u64); 1564 1565 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) 1566 result += sizeof(u64); 1567 1568 if (type & PERF_SAMPLE_AUX) { 1569 result += sizeof(u64); 1570 result += sample->aux_sample.size; 1571 } 1572 1573 return result; 1574 } 1575 1576 static void perf_synthesize_sample_weight(const struct perf_sample *data, 1577 __u64 *array, u64 type __maybe_unused) 1578 { 1579 *array = data->weight; 1580 1581 if (type & PERF_SAMPLE_WEIGHT_STRUCT) { 1582 *array &= 0xffffffff; 1583 *array |= ((u64)data->ins_lat << 32); 1584 *array |= ((u64)data->weight3 << 48); 1585 } 1586 } 1587 1588 static __u64 *copy_read_group_values(__u64 *array, __u64 read_format, 1589 const struct perf_sample *sample) 1590 { 1591 size_t sz = sample_read_value_size(read_format); 1592 struct sample_read_value *v = sample->read.group.values; 1593 1594 sample_read_group__for_each(v, sample->read.group.nr, read_format) { 1595 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1596 memcpy(array, v, sz); 1597 array = (void *)array + sz; 1598 } 1599 return array; 1600 } 1601 1602 int perf_event__synthesize_sample(union perf_event *event, u64 type, u64 read_format, 1603 const struct perf_sample *sample) 1604 { 1605 __u64 *array; 1606 size_t sz; 1607 /* 1608 * used for cross-endian analysis. See git commit 65014ab3 1609 * for why this goofiness is needed. 1610 */ 1611 union u64_swap u; 1612 1613 array = event->sample.array; 1614 1615 if (type & PERF_SAMPLE_IDENTIFIER) { 1616 *array = sample->id; 1617 array++; 1618 } 1619 1620 if (type & PERF_SAMPLE_IP) { 1621 *array = sample->ip; 1622 array++; 1623 } 1624 1625 if (type & PERF_SAMPLE_TID) { 1626 u.val32[0] = sample->pid; 1627 u.val32[1] = sample->tid; 1628 *array = u.val64; 1629 array++; 1630 } 1631 1632 if (type & PERF_SAMPLE_TIME) { 1633 *array = sample->time; 1634 array++; 1635 } 1636 1637 if (type & PERF_SAMPLE_ADDR) { 1638 *array = sample->addr; 1639 array++; 1640 } 1641 1642 if (type & PERF_SAMPLE_ID) { 1643 *array = sample->id; 1644 array++; 1645 } 1646 1647 if (type & PERF_SAMPLE_STREAM_ID) { 1648 *array = sample->stream_id; 1649 array++; 1650 } 1651 1652 if (type & PERF_SAMPLE_CPU) { 1653 u.val32[0] = sample->cpu; 1654 u.val32[1] = 0; 1655 *array = u.val64; 1656 array++; 1657 } 1658 1659 if (type & PERF_SAMPLE_PERIOD) { 1660 *array = sample->period; 1661 array++; 1662 } 1663 1664 if (type & PERF_SAMPLE_READ) { 1665 if (read_format & PERF_FORMAT_GROUP) 1666 *array = sample->read.group.nr; 1667 else 1668 *array = sample->read.one.value; 1669 array++; 1670 1671 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) { 1672 *array = sample->read.time_enabled; 1673 array++; 1674 } 1675 1676 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) { 1677 *array = sample->read.time_running; 1678 array++; 1679 } 1680 1681 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */ 1682 if (read_format & PERF_FORMAT_GROUP) { 1683 array = copy_read_group_values(array, read_format, 1684 sample); 1685 } else { 1686 *array = sample->read.one.id; 1687 array++; 1688 1689 if (read_format & PERF_FORMAT_LOST) { 1690 *array = sample->read.one.lost; 1691 array++; 1692 } 1693 } 1694 } 1695 1696 if (type & PERF_SAMPLE_CALLCHAIN) { 1697 sz = (sample->callchain->nr + 1) * sizeof(u64); 1698 memcpy(array, sample->callchain, sz); 1699 array = (void *)array + sz; 1700 } 1701 1702 if (type & PERF_SAMPLE_RAW) { 1703 u32 *array32 = (void *)array; 1704 1705 *array32 = sample->raw_size; 1706 array32++; 1707 1708 memcpy(array32, sample->raw_data, sample->raw_size); 1709 array = (void *)(array32 + (sample->raw_size / sizeof(u32))); 1710 1711 /* make sure the array is 64-bit aligned */ 1712 BUG_ON(((long)array) % sizeof(u64)); 1713 } 1714 1715 if (type & PERF_SAMPLE_BRANCH_STACK) { 1716 sz = sample->branch_stack->nr * sizeof(struct branch_entry); 1717 /* nr, hw_idx */ 1718 sz += 2 * sizeof(u64); 1719 memcpy(array, sample->branch_stack, sz); 1720 array = (void *)array + sz; 1721 } 1722 1723 if (type & PERF_SAMPLE_REGS_USER) { 1724 if (sample->user_regs && sample->user_regs->abi) { 1725 *array++ = sample->user_regs->abi; 1726 sz = hweight64(sample->user_regs->mask) * sizeof(u64); 1727 memcpy(array, sample->user_regs->regs, sz); 1728 array = (void *)array + sz; 1729 } else { 1730 *array++ = 0; 1731 } 1732 } 1733 1734 if (type & PERF_SAMPLE_STACK_USER) { 1735 sz = sample->user_stack.size; 1736 *array++ = sz; 1737 if (sz) { 1738 memcpy(array, sample->user_stack.data, sz); 1739 array = (void *)array + sz; 1740 *array++ = sz; 1741 } 1742 } 1743 1744 if (type & PERF_SAMPLE_WEIGHT_TYPE) { 1745 perf_synthesize_sample_weight(sample, array, type); 1746 array++; 1747 } 1748 1749 if (type & PERF_SAMPLE_DATA_SRC) { 1750 *array = sample->data_src; 1751 array++; 1752 } 1753 1754 if (type & PERF_SAMPLE_TRANSACTION) { 1755 *array = sample->transaction; 1756 array++; 1757 } 1758 1759 if (type & PERF_SAMPLE_REGS_INTR) { 1760 if (sample->intr_regs && sample->intr_regs->abi) { 1761 *array++ = sample->intr_regs->abi; 1762 sz = hweight64(sample->intr_regs->mask) * sizeof(u64); 1763 memcpy(array, sample->intr_regs->regs, sz); 1764 array = (void *)array + sz; 1765 } else { 1766 *array++ = 0; 1767 } 1768 } 1769 1770 if (type & PERF_SAMPLE_PHYS_ADDR) { 1771 *array = sample->phys_addr; 1772 array++; 1773 } 1774 1775 if (type & PERF_SAMPLE_CGROUP) { 1776 *array = sample->cgroup; 1777 array++; 1778 } 1779 1780 if (type & PERF_SAMPLE_DATA_PAGE_SIZE) { 1781 *array = sample->data_page_size; 1782 array++; 1783 } 1784 1785 if (type & PERF_SAMPLE_CODE_PAGE_SIZE) { 1786 *array = sample->code_page_size; 1787 array++; 1788 } 1789 1790 if (type & PERF_SAMPLE_AUX) { 1791 sz = sample->aux_sample.size; 1792 *array++ = sz; 1793 memcpy(array, sample->aux_sample.data, sz); 1794 array = (void *)array + sz; 1795 } 1796 1797 return 0; 1798 } 1799 1800 int perf_event__synthesize_id_sample(__u64 *array, u64 type, const struct perf_sample *sample) 1801 { 1802 __u64 *start = array; 1803 1804 /* 1805 * used for cross-endian analysis. See git commit 65014ab3 1806 * for why this goofiness is needed. 1807 */ 1808 union u64_swap u; 1809 1810 if (type & PERF_SAMPLE_TID) { 1811 u.val32[0] = sample->pid; 1812 u.val32[1] = sample->tid; 1813 *array = u.val64; 1814 array++; 1815 } 1816 1817 if (type & PERF_SAMPLE_TIME) { 1818 *array = sample->time; 1819 array++; 1820 } 1821 1822 if (type & PERF_SAMPLE_ID) { 1823 *array = sample->id; 1824 array++; 1825 } 1826 1827 if (type & PERF_SAMPLE_STREAM_ID) { 1828 *array = sample->stream_id; 1829 array++; 1830 } 1831 1832 if (type & PERF_SAMPLE_CPU) { 1833 u.val32[0] = sample->cpu; 1834 u.val32[1] = 0; 1835 *array = u.val64; 1836 array++; 1837 } 1838 1839 if (type & PERF_SAMPLE_IDENTIFIER) { 1840 *array = sample->id; 1841 array++; 1842 } 1843 1844 return (void *)array - (void *)start; 1845 } 1846 1847 int __perf_event__synthesize_id_index(const struct perf_tool *tool, perf_event__handler_t process, 1848 struct evlist *evlist, struct machine *machine, size_t from) 1849 { 1850 union perf_event *ev; 1851 struct evsel *evsel; 1852 size_t nr = 0, i = 0, sz, max_nr, n, pos; 1853 size_t e1_sz = sizeof(struct id_index_entry); 1854 size_t e2_sz = sizeof(struct id_index_entry_2); 1855 size_t etot_sz = e1_sz + e2_sz; 1856 bool e2_needed = false; 1857 int err; 1858 1859 max_nr = (UINT16_MAX - sizeof(struct perf_record_id_index)) / etot_sz; 1860 1861 pos = 0; 1862 evlist__for_each_entry(evlist, evsel) { 1863 if (pos++ < from) 1864 continue; 1865 nr += evsel->core.ids; 1866 } 1867 1868 if (!nr) 1869 return 0; 1870 1871 pr_debug2("Synthesizing id index\n"); 1872 1873 n = nr > max_nr ? max_nr : nr; 1874 sz = sizeof(struct perf_record_id_index) + n * etot_sz; 1875 ev = zalloc(sz); 1876 if (!ev) 1877 return -ENOMEM; 1878 1879 sz = sizeof(struct perf_record_id_index) + n * e1_sz; 1880 1881 ev->id_index.header.type = PERF_RECORD_ID_INDEX; 1882 ev->id_index.nr = n; 1883 1884 pos = 0; 1885 evlist__for_each_entry(evlist, evsel) { 1886 u32 j; 1887 1888 if (pos++ < from) 1889 continue; 1890 for (j = 0; j < evsel->core.ids; j++, i++) { 1891 struct id_index_entry *e; 1892 struct id_index_entry_2 *e2; 1893 struct perf_sample_id *sid; 1894 1895 if (i >= n) { 1896 ev->id_index.header.size = sz + (e2_needed ? n * e2_sz : 0); 1897 err = process(tool, ev, NULL, machine); 1898 if (err) 1899 goto out_err; 1900 nr -= n; 1901 i = 0; 1902 e2_needed = false; 1903 } 1904 1905 e = &ev->id_index.entries[i]; 1906 1907 e->id = evsel->core.id[j]; 1908 1909 sid = evlist__id2sid(evlist, e->id); 1910 if (!sid) { 1911 free(ev); 1912 return -ENOENT; 1913 } 1914 1915 e->idx = sid->idx; 1916 e->cpu = sid->cpu.cpu; 1917 e->tid = sid->tid; 1918 1919 if (sid->machine_pid) 1920 e2_needed = true; 1921 1922 e2 = (void *)ev + sz; 1923 e2[i].machine_pid = sid->machine_pid; 1924 e2[i].vcpu = sid->vcpu.cpu; 1925 } 1926 } 1927 1928 sz = sizeof(struct perf_record_id_index) + nr * e1_sz; 1929 ev->id_index.header.size = sz + (e2_needed ? nr * e2_sz : 0); 1930 ev->id_index.nr = nr; 1931 1932 err = process(tool, ev, NULL, machine); 1933 out_err: 1934 free(ev); 1935 1936 return err; 1937 } 1938 1939 int perf_event__synthesize_id_index(const struct perf_tool *tool, perf_event__handler_t process, 1940 struct evlist *evlist, struct machine *machine) 1941 { 1942 return __perf_event__synthesize_id_index(tool, process, evlist, machine, 0); 1943 } 1944 1945 int __machine__synthesize_threads(struct machine *machine, const struct perf_tool *tool, 1946 struct target *target, struct perf_thread_map *threads, 1947 perf_event__handler_t process, bool needs_mmap, 1948 bool data_mmap, unsigned int nr_threads_synthesize) 1949 { 1950 /* 1951 * When perf runs in non-root PID namespace, and the namespace's proc FS 1952 * is not mounted, nsinfo__is_in_root_namespace() returns false. 1953 * In this case, the proc FS is coming for the parent namespace, thus 1954 * perf tool will wrongly gather process info from its parent PID 1955 * namespace. 1956 * 1957 * To avoid the confusion that the perf tool runs in a child PID 1958 * namespace but it synthesizes thread info from its parent PID 1959 * namespace, returns failure with warning. 1960 */ 1961 if (!nsinfo__is_in_root_namespace()) { 1962 pr_err("Perf runs in non-root PID namespace but it tries to "); 1963 pr_err("gather process info from its parent PID namespace.\n"); 1964 pr_err("Please mount the proc file system properly, e.g. "); 1965 pr_err("add the option '--mount-proc' for unshare command.\n"); 1966 return -EPERM; 1967 } 1968 1969 if (target__has_task(target)) 1970 return perf_event__synthesize_thread_map(tool, threads, process, machine, 1971 needs_mmap, data_mmap); 1972 else if (target__has_cpu(target)) 1973 return perf_event__synthesize_threads(tool, process, machine, 1974 needs_mmap, data_mmap, 1975 nr_threads_synthesize); 1976 /* command specified */ 1977 return 0; 1978 } 1979 1980 int machine__synthesize_threads(struct machine *machine, struct target *target, 1981 struct perf_thread_map *threads, bool needs_mmap, 1982 bool data_mmap, unsigned int nr_threads_synthesize) 1983 { 1984 return __machine__synthesize_threads(machine, NULL, target, threads, 1985 perf_event__process, needs_mmap, 1986 data_mmap, nr_threads_synthesize); 1987 } 1988 1989 static struct perf_record_event_update *event_update_event__new(size_t size, u64 type, u64 id) 1990 { 1991 struct perf_record_event_update *ev; 1992 1993 size += sizeof(*ev); 1994 size = PERF_ALIGN(size, sizeof(u64)); 1995 1996 ev = zalloc(size); 1997 if (ev) { 1998 ev->header.type = PERF_RECORD_EVENT_UPDATE; 1999 ev->header.size = (u16)size; 2000 ev->type = type; 2001 ev->id = id; 2002 } 2003 return ev; 2004 } 2005 2006 int perf_event__synthesize_event_update_unit(const struct perf_tool *tool, struct evsel *evsel, 2007 perf_event__handler_t process) 2008 { 2009 size_t size = strlen(evsel->unit); 2010 struct perf_record_event_update *ev; 2011 int err; 2012 2013 ev = event_update_event__new(size + 1, PERF_EVENT_UPDATE__UNIT, evsel->core.id[0]); 2014 if (ev == NULL) 2015 return -ENOMEM; 2016 2017 strlcpy(ev->unit, evsel->unit, size + 1); 2018 err = process(tool, (union perf_event *)ev, NULL, NULL); 2019 free(ev); 2020 return err; 2021 } 2022 2023 int perf_event__synthesize_event_update_scale(const struct perf_tool *tool, struct evsel *evsel, 2024 perf_event__handler_t process) 2025 { 2026 struct perf_record_event_update *ev; 2027 struct perf_record_event_update_scale *ev_data; 2028 int err; 2029 2030 ev = event_update_event__new(sizeof(*ev_data), PERF_EVENT_UPDATE__SCALE, evsel->core.id[0]); 2031 if (ev == NULL) 2032 return -ENOMEM; 2033 2034 ev->scale.scale = evsel->scale; 2035 err = process(tool, (union perf_event *)ev, NULL, NULL); 2036 free(ev); 2037 return err; 2038 } 2039 2040 int perf_event__synthesize_event_update_name(const struct perf_tool *tool, struct evsel *evsel, 2041 perf_event__handler_t process) 2042 { 2043 struct perf_record_event_update *ev; 2044 size_t len = strlen(evsel__name(evsel)); 2045 int err; 2046 2047 ev = event_update_event__new(len + 1, PERF_EVENT_UPDATE__NAME, evsel->core.id[0]); 2048 if (ev == NULL) 2049 return -ENOMEM; 2050 2051 strlcpy(ev->name, evsel->name, len + 1); 2052 err = process(tool, (union perf_event *)ev, NULL, NULL); 2053 free(ev); 2054 return err; 2055 } 2056 2057 int perf_event__synthesize_event_update_cpus(const struct perf_tool *tool, struct evsel *evsel, 2058 perf_event__handler_t process) 2059 { 2060 struct synthesize_cpu_map_data syn_data = { .map = evsel->core.pmu_cpus }; 2061 struct perf_record_event_update *ev; 2062 int err; 2063 2064 ev = cpu_map_data__alloc(&syn_data, sizeof(struct perf_event_header) + 2 * sizeof(u64)); 2065 if (!ev) 2066 return -ENOMEM; 2067 2068 syn_data.data = &ev->cpus.cpus; 2069 ev->header.type = PERF_RECORD_EVENT_UPDATE; 2070 ev->header.size = (u16)syn_data.size; 2071 ev->type = PERF_EVENT_UPDATE__CPUS; 2072 ev->id = evsel->core.id[0]; 2073 cpu_map_data__synthesize(&syn_data); 2074 2075 err = process(tool, (union perf_event *)ev, NULL, NULL); 2076 free(ev); 2077 return err; 2078 } 2079 2080 int perf_event__synthesize_attrs(const struct perf_tool *tool, struct evlist *evlist, 2081 perf_event__handler_t process) 2082 { 2083 struct evsel *evsel; 2084 int err = 0; 2085 2086 evlist__for_each_entry(evlist, evsel) { 2087 err = perf_event__synthesize_attr(tool, &evsel->core.attr, evsel->core.ids, 2088 evsel->core.id, process); 2089 if (err) { 2090 pr_debug("failed to create perf header attribute\n"); 2091 return err; 2092 } 2093 } 2094 2095 return err; 2096 } 2097 2098 static bool has_unit(struct evsel *evsel) 2099 { 2100 return evsel->unit && *evsel->unit; 2101 } 2102 2103 static bool has_scale(struct evsel *evsel) 2104 { 2105 return evsel->scale != 1; 2106 } 2107 2108 int perf_event__synthesize_extra_attr(const struct perf_tool *tool, struct evlist *evsel_list, 2109 perf_event__handler_t process, bool is_pipe) 2110 { 2111 struct evsel *evsel; 2112 int err; 2113 2114 /* 2115 * Synthesize other events stuff not carried within 2116 * attr event - unit, scale, name 2117 */ 2118 evlist__for_each_entry(evsel_list, evsel) { 2119 if (!evsel->supported) 2120 continue; 2121 2122 /* 2123 * Synthesize unit and scale only if it's defined. 2124 */ 2125 if (has_unit(evsel)) { 2126 err = perf_event__synthesize_event_update_unit(tool, evsel, process); 2127 if (err < 0) { 2128 pr_err("Couldn't synthesize evsel unit.\n"); 2129 return err; 2130 } 2131 } 2132 2133 if (has_scale(evsel)) { 2134 err = perf_event__synthesize_event_update_scale(tool, evsel, process); 2135 if (err < 0) { 2136 pr_err("Couldn't synthesize evsel evsel.\n"); 2137 return err; 2138 } 2139 } 2140 2141 if (evsel->core.pmu_cpus) { 2142 err = perf_event__synthesize_event_update_cpus(tool, evsel, process); 2143 if (err < 0) { 2144 pr_err("Couldn't synthesize evsel cpus.\n"); 2145 return err; 2146 } 2147 } 2148 2149 /* 2150 * Name is needed only for pipe output, 2151 * perf.data carries event names. 2152 */ 2153 if (is_pipe) { 2154 err = perf_event__synthesize_event_update_name(tool, evsel, process); 2155 if (err < 0) { 2156 pr_err("Couldn't synthesize evsel name.\n"); 2157 return err; 2158 } 2159 } 2160 } 2161 return 0; 2162 } 2163 2164 int perf_event__synthesize_attr(const struct perf_tool *tool, struct perf_event_attr *attr, 2165 u32 ids, u64 *id, perf_event__handler_t process) 2166 { 2167 union perf_event *ev; 2168 size_t size; 2169 int err; 2170 2171 size = sizeof(struct perf_event_attr); 2172 size = PERF_ALIGN(size, sizeof(u64)); 2173 size += sizeof(struct perf_event_header); 2174 size += ids * sizeof(u64); 2175 2176 ev = zalloc(size); 2177 2178 if (ev == NULL) 2179 return -ENOMEM; 2180 2181 ev->attr.attr = *attr; 2182 memcpy(perf_record_header_attr_id(ev), id, ids * sizeof(u64)); 2183 2184 ev->attr.header.type = PERF_RECORD_HEADER_ATTR; 2185 ev->attr.header.size = (u16)size; 2186 2187 if (ev->attr.header.size == size) 2188 err = process(tool, ev, NULL, NULL); 2189 else 2190 err = -E2BIG; 2191 2192 free(ev); 2193 2194 return err; 2195 } 2196 2197 #ifdef HAVE_LIBTRACEEVENT 2198 int perf_event__synthesize_tracing_data(const struct perf_tool *tool, int fd, struct evlist *evlist, 2199 perf_event__handler_t process) 2200 { 2201 union perf_event ev; 2202 struct tracing_data *tdata; 2203 ssize_t size = 0, aligned_size = 0, padding; 2204 struct feat_fd ff; 2205 2206 /* 2207 * We are going to store the size of the data followed 2208 * by the data contents. Since the fd descriptor is a pipe, 2209 * we cannot seek back to store the size of the data once 2210 * we know it. Instead we: 2211 * 2212 * - write the tracing data to the temp file 2213 * - get/write the data size to pipe 2214 * - write the tracing data from the temp file 2215 * to the pipe 2216 */ 2217 tdata = tracing_data_get(&evlist->core.entries, fd, true); 2218 if (!tdata) 2219 return -1; 2220 2221 memset(&ev, 0, sizeof(ev.tracing_data)); 2222 2223 ev.tracing_data.header.type = PERF_RECORD_HEADER_TRACING_DATA; 2224 size = tdata->size; 2225 aligned_size = PERF_ALIGN(size, sizeof(u64)); 2226 padding = aligned_size - size; 2227 ev.tracing_data.header.size = sizeof(ev.tracing_data); 2228 ev.tracing_data.size = aligned_size; 2229 2230 process(tool, &ev, NULL, NULL); 2231 2232 /* 2233 * The put function will copy all the tracing data 2234 * stored in temp file to the pipe. 2235 */ 2236 tracing_data_put(tdata); 2237 2238 ff = (struct feat_fd){ .fd = fd }; 2239 if (write_padded(&ff, NULL, 0, padding)) 2240 return -1; 2241 2242 return aligned_size; 2243 } 2244 #endif 2245 2246 int perf_event__synthesize_build_id(const struct perf_tool *tool, 2247 struct perf_sample *sample, 2248 struct machine *machine, 2249 perf_event__handler_t process, 2250 const struct evsel *evsel, 2251 __u16 misc, 2252 const struct build_id *bid, 2253 const char *filename) 2254 { 2255 union perf_event ev; 2256 size_t len; 2257 2258 len = sizeof(ev.build_id) + strlen(filename) + 1; 2259 len = PERF_ALIGN(len, sizeof(u64)); 2260 2261 memset(&ev, 0, len); 2262 2263 ev.build_id.size = bid->size; 2264 if (ev.build_id.size > sizeof(ev.build_id.build_id)) 2265 ev.build_id.size = sizeof(ev.build_id.build_id); 2266 memcpy(ev.build_id.build_id, bid->data, ev.build_id.size); 2267 ev.build_id.header.type = PERF_RECORD_HEADER_BUILD_ID; 2268 ev.build_id.header.misc = misc | PERF_RECORD_MISC_BUILD_ID_SIZE; 2269 ev.build_id.pid = machine->pid; 2270 ev.build_id.header.size = len; 2271 strcpy(ev.build_id.filename, filename); 2272 2273 if (evsel) { 2274 void *array = &ev; 2275 int ret; 2276 2277 array += ev.header.size; 2278 ret = perf_event__synthesize_id_sample(array, evsel->core.attr.sample_type, sample); 2279 if (ret < 0) 2280 return ret; 2281 2282 if (ret & 7) { 2283 pr_err("Bad id sample size %d\n", ret); 2284 return -EINVAL; 2285 } 2286 2287 ev.header.size += ret; 2288 } 2289 2290 return process(tool, &ev, sample, machine); 2291 } 2292 2293 int perf_event__synthesize_mmap2_build_id(const struct perf_tool *tool, 2294 struct perf_sample *sample, 2295 struct machine *machine, 2296 perf_event__handler_t process, 2297 const struct evsel *evsel, 2298 __u16 misc, 2299 __u32 pid, __u32 tid, 2300 __u64 start, __u64 len, __u64 pgoff, 2301 const struct build_id *bid, 2302 __u32 prot, __u32 flags, 2303 const char *filename) 2304 { 2305 union perf_event ev; 2306 size_t ev_len; 2307 void *array; 2308 int ret; 2309 2310 ev_len = sizeof(ev.mmap2) - sizeof(ev.mmap2.filename) + strlen(filename) + 1; 2311 ev_len = PERF_ALIGN(ev_len, sizeof(u64)); 2312 2313 memset(&ev, 0, ev_len); 2314 2315 ev.mmap2.header.type = PERF_RECORD_MMAP2; 2316 ev.mmap2.header.misc = misc | PERF_RECORD_MISC_MMAP_BUILD_ID; 2317 ev.mmap2.header.size = ev_len; 2318 2319 ev.mmap2.pid = pid; 2320 ev.mmap2.tid = tid; 2321 ev.mmap2.start = start; 2322 ev.mmap2.len = len; 2323 ev.mmap2.pgoff = pgoff; 2324 2325 ev.mmap2.build_id_size = bid->size; 2326 if (ev.mmap2.build_id_size > sizeof(ev.mmap2.build_id)) 2327 ev.build_id.size = sizeof(ev.mmap2.build_id); 2328 memcpy(ev.mmap2.build_id, bid->data, ev.mmap2.build_id_size); 2329 2330 ev.mmap2.prot = prot; 2331 ev.mmap2.flags = flags; 2332 2333 memcpy(ev.mmap2.filename, filename, min(strlen(filename), sizeof(ev.mmap.filename))); 2334 2335 array = &ev; 2336 array += ev.header.size; 2337 ret = perf_event__synthesize_id_sample(array, evsel->core.attr.sample_type, sample); 2338 if (ret < 0) 2339 return ret; 2340 2341 if (ret & 7) { 2342 pr_err("Bad id sample size %d\n", ret); 2343 return -EINVAL; 2344 } 2345 2346 ev.header.size += ret; 2347 2348 return process(tool, &ev, sample, machine); 2349 } 2350 2351 int perf_event__synthesize_stat_events(struct perf_stat_config *config, const struct perf_tool *tool, 2352 struct evlist *evlist, perf_event__handler_t process, bool attrs) 2353 { 2354 int err; 2355 2356 if (attrs) { 2357 err = perf_event__synthesize_attrs(tool, evlist, process); 2358 if (err < 0) { 2359 pr_err("Couldn't synthesize attrs.\n"); 2360 return err; 2361 } 2362 } 2363 2364 err = perf_event__synthesize_extra_attr(tool, evlist, process, attrs); 2365 err = perf_event__synthesize_thread_map2(tool, evlist->core.threads, process, NULL); 2366 if (err < 0) { 2367 pr_err("Couldn't synthesize thread map.\n"); 2368 return err; 2369 } 2370 2371 err = perf_event__synthesize_cpu_map(tool, evlist->core.user_requested_cpus, process, NULL); 2372 if (err < 0) { 2373 pr_err("Couldn't synthesize thread map.\n"); 2374 return err; 2375 } 2376 2377 err = perf_event__synthesize_stat_config(tool, config, process, NULL); 2378 if (err < 0) { 2379 pr_err("Couldn't synthesize config.\n"); 2380 return err; 2381 } 2382 2383 return 0; 2384 } 2385 2386 extern const struct perf_header_feature_ops feat_ops[HEADER_LAST_FEATURE]; 2387 2388 int perf_event__synthesize_features(const struct perf_tool *tool, struct perf_session *session, 2389 struct evlist *evlist, perf_event__handler_t process) 2390 { 2391 struct perf_header *header = &session->header; 2392 struct perf_record_header_feature *fe; 2393 struct feat_fd ff; 2394 size_t sz, sz_hdr; 2395 int feat, ret; 2396 2397 sz_hdr = sizeof(fe->header); 2398 sz = sizeof(union perf_event); 2399 /* get a nice alignment */ 2400 sz = PERF_ALIGN(sz, page_size); 2401 2402 memset(&ff, 0, sizeof(ff)); 2403 2404 ff.buf = malloc(sz); 2405 if (!ff.buf) 2406 return -ENOMEM; 2407 2408 ff.size = sz - sz_hdr; 2409 ff.ph = &session->header; 2410 2411 for_each_set_bit(feat, header->adds_features, HEADER_FEAT_BITS) { 2412 if (!feat_ops[feat].synthesize) { 2413 pr_debug("No record header feature for header :%d\n", feat); 2414 continue; 2415 } 2416 2417 ff.offset = sizeof(*fe); 2418 2419 ret = feat_ops[feat].write(&ff, evlist); 2420 if (ret || ff.offset <= (ssize_t)sizeof(*fe)) { 2421 pr_debug("Error writing feature\n"); 2422 continue; 2423 } 2424 /* ff.buf may have changed due to realloc in do_write() */ 2425 fe = ff.buf; 2426 memset(fe, 0, sizeof(*fe)); 2427 2428 fe->feat_id = feat; 2429 fe->header.type = PERF_RECORD_HEADER_FEATURE; 2430 fe->header.size = ff.offset; 2431 2432 ret = process(tool, ff.buf, NULL, NULL); 2433 if (ret) { 2434 free(ff.buf); 2435 return ret; 2436 } 2437 } 2438 2439 /* Send HEADER_LAST_FEATURE mark. */ 2440 fe = ff.buf; 2441 fe->feat_id = HEADER_LAST_FEATURE; 2442 fe->header.type = PERF_RECORD_HEADER_FEATURE; 2443 fe->header.size = sizeof(*fe); 2444 2445 ret = process(tool, ff.buf, NULL, NULL); 2446 2447 free(ff.buf); 2448 return ret; 2449 } 2450 2451 int perf_event__synthesize_for_pipe(const struct perf_tool *tool, 2452 struct perf_session *session, 2453 struct perf_data *data, 2454 perf_event__handler_t process) 2455 { 2456 int err; 2457 int ret = 0; 2458 struct evlist *evlist = session->evlist; 2459 2460 /* 2461 * We need to synthesize events first, because some 2462 * features works on top of them (on report side). 2463 */ 2464 err = perf_event__synthesize_attrs(tool, evlist, process); 2465 if (err < 0) { 2466 pr_err("Couldn't synthesize attrs.\n"); 2467 return err; 2468 } 2469 ret += err; 2470 2471 err = perf_event__synthesize_features(tool, session, evlist, process); 2472 if (err < 0) { 2473 pr_err("Couldn't synthesize features.\n"); 2474 return err; 2475 } 2476 ret += err; 2477 2478 #ifdef HAVE_LIBTRACEEVENT 2479 if (have_tracepoints(&evlist->core.entries)) { 2480 int fd = perf_data__fd(data); 2481 2482 /* 2483 * FIXME err <= 0 here actually means that 2484 * there were no tracepoints so its not really 2485 * an error, just that we don't need to 2486 * synthesize anything. We really have to 2487 * return this more properly and also 2488 * propagate errors that now are calling die() 2489 */ 2490 err = perf_event__synthesize_tracing_data(tool, fd, evlist, 2491 process); 2492 if (err <= 0) { 2493 pr_err("Couldn't record tracing data.\n"); 2494 return err; 2495 } 2496 ret += err; 2497 } 2498 #else 2499 (void)data; 2500 #endif 2501 2502 return ret; 2503 } 2504 2505 int parse_synth_opt(char *synth) 2506 { 2507 char *p, *q; 2508 int ret = 0; 2509 2510 if (synth == NULL) 2511 return -1; 2512 2513 for (q = synth; (p = strsep(&q, ",")); p = q) { 2514 if (!strcasecmp(p, "no") || !strcasecmp(p, "none")) 2515 return 0; 2516 2517 if (!strcasecmp(p, "all")) 2518 return PERF_SYNTH_ALL; 2519 2520 if (!strcasecmp(p, "task")) 2521 ret |= PERF_SYNTH_TASK; 2522 else if (!strcasecmp(p, "mmap")) 2523 ret |= PERF_SYNTH_TASK | PERF_SYNTH_MMAP; 2524 else if (!strcasecmp(p, "cgroup")) 2525 ret |= PERF_SYNTH_CGROUP; 2526 else 2527 return -1; 2528 } 2529 2530 return ret; 2531 } 2532