1 /* Copyright (c) 2012 The Chromium OS Authors. All rights reserved. 2 * Use of this source code is governed by a BSD-style license that can be 3 * found in the LICENSE file. 4 */ 5 6 #define _BSD_SOURCE 7 #define _DEFAULT_SOURCE 8 #define _GNU_SOURCE 9 10 #include <asm/unistd.h> 11 #include <ctype.h> 12 #include <dirent.h> 13 #include <errno.h> 14 #include <fcntl.h> 15 #include <grp.h> 16 #include <inttypes.h> 17 #include <limits.h> 18 #include <linux/capability.h> 19 #include <net/if.h> 20 #include <pwd.h> 21 #include <sched.h> 22 #include <signal.h> 23 #include <stdarg.h> 24 #include <stdbool.h> 25 #include <stddef.h> 26 #include <stdio.h> 27 #include <stdlib.h> 28 #include <string.h> 29 #include <sys/capability.h> 30 #include <sys/mount.h> 31 #include <sys/param.h> 32 #include <sys/prctl.h> 33 #include <sys/socket.h> 34 #include <sys/stat.h> 35 #include <sys/types.h> 36 #include <sys/user.h> 37 #include <sys/wait.h> 38 #include <syscall.h> 39 #include <unistd.h> 40 41 #include "libminijail.h" 42 #include "libminijail-private.h" 43 44 #include "signal_handler.h" 45 #include "syscall_filter.h" 46 #include "syscall_wrapper.h" 47 #include "util.h" 48 49 #ifdef HAVE_SECUREBITS_H 50 # include <linux/securebits.h> 51 #else 52 # define SECURE_ALL_BITS 0x55 53 # define SECURE_ALL_LOCKS (SECURE_ALL_BITS << 1) 54 #endif 55 /* For kernels < 4.3. */ 56 #define OLD_SECURE_ALL_BITS 0x15 57 #define OLD_SECURE_ALL_LOCKS (OLD_SECURE_ALL_BITS << 1) 58 59 /* 60 * Assert the value of SECURE_ALL_BITS at compile-time. 61 * Brillo devices are currently compiled against 4.4 kernel headers. Kernel 4.3 62 * added a new securebit. 63 * When a new securebit is added, the new SECURE_ALL_BITS mask will return EPERM 64 * when used on older kernels. The compile-time assert will catch this situation 65 * at compile time. 66 */ 67 #ifdef __BRILLO__ 68 _Static_assert(SECURE_ALL_BITS == 0x55, "SECURE_ALL_BITS == 0x55."); 69 #endif 70 71 /* Until these are reliably available in linux/prctl.h. */ 72 #ifndef PR_SET_SECCOMP 73 # define PR_SET_SECCOMP 22 74 #endif 75 76 #ifndef PR_ALT_SYSCALL 77 # define PR_ALT_SYSCALL 0x43724f53 78 #endif 79 80 /* Seccomp filter related flags. */ 81 #ifndef PR_SET_NO_NEW_PRIVS 82 # define PR_SET_NO_NEW_PRIVS 38 83 #endif 84 85 #ifndef SECCOMP_MODE_FILTER 86 # define SECCOMP_MODE_FILTER 2 /* uses user-supplied filter. */ 87 #endif 88 89 #ifndef SECCOMP_SET_MODE_STRICT 90 # define SECCOMP_SET_MODE_STRICT 0 91 #endif 92 #ifndef SECCOMP_SET_MODE_FILTER 93 # define SECCOMP_SET_MODE_FILTER 1 94 #endif 95 96 #ifndef SECCOMP_FILTER_FLAG_TSYNC 97 # define SECCOMP_FILTER_FLAG_TSYNC 1 98 #endif 99 /* End seccomp filter related flags. */ 100 101 /* New cgroup namespace might not be in linux-headers yet. */ 102 #ifndef CLONE_NEWCGROUP 103 # define CLONE_NEWCGROUP 0x02000000 104 #endif 105 106 #define MAX_CGROUPS 10 /* 10 different controllers supported by Linux. */ 107 108 /* Keyctl commands. */ 109 #define KEYCTL_JOIN_SESSION_KEYRING 1 110 111 struct mountpoint { 112 char *src; 113 char *dest; 114 char *type; 115 char *data; 116 int has_data; 117 unsigned long flags; 118 struct mountpoint *next; 119 }; 120 121 struct minijail { 122 /* 123 * WARNING: if you add a flag here you need to make sure it's 124 * accounted for in minijail_pre{enter|exec}() below. 125 */ 126 struct { 127 int uid : 1; 128 int gid : 1; 129 int inherit_suppl_gids : 1; 130 int set_suppl_gids : 1; 131 int keep_suppl_gids : 1; 132 int use_caps : 1; 133 int capbset_drop : 1; 134 int vfs : 1; 135 int enter_vfs : 1; 136 int skip_remount_private : 1; 137 int pids : 1; 138 int ipc : 1; 139 int net : 1; 140 int enter_net : 1; 141 int ns_cgroups : 1; 142 int userns : 1; 143 int disable_setgroups : 1; 144 int seccomp : 1; 145 int remount_proc_ro : 1; 146 int no_new_privs : 1; 147 int seccomp_filter : 1; 148 int seccomp_filter_tsync : 1; 149 int seccomp_filter_logging : 1; 150 int chroot : 1; 151 int pivot_root : 1; 152 int mount_tmp : 1; 153 int do_init : 1; 154 int pid_file : 1; 155 int cgroups : 1; 156 int alt_syscall : 1; 157 int reset_signal_mask : 1; 158 int close_open_fds : 1; 159 int new_session_keyring : 1; 160 } flags; 161 uid_t uid; 162 gid_t gid; 163 gid_t usergid; 164 char *user; 165 size_t suppl_gid_count; 166 gid_t *suppl_gid_list; 167 uint64_t caps; 168 uint64_t cap_bset; 169 pid_t initpid; 170 int mountns_fd; 171 int netns_fd; 172 char *chrootdir; 173 char *pid_file_path; 174 char *uidmap; 175 char *gidmap; 176 size_t filter_len; 177 struct sock_fprog *filter_prog; 178 char *alt_syscall_table; 179 struct mountpoint *mounts_head; 180 struct mountpoint *mounts_tail; 181 size_t mounts_count; 182 size_t tmpfs_size; 183 char *cgroups[MAX_CGROUPS]; 184 size_t cgroup_count; 185 }; 186 187 /* 188 * Strip out flags meant for the parent. 189 * We keep things that are not inherited across execve(2) (e.g. capabilities), 190 * or are easier to set after execve(2) (e.g. seccomp filters). 191 */ 192 void minijail_preenter(struct minijail *j) 193 { 194 j->flags.vfs = 0; 195 j->flags.enter_vfs = 0; 196 j->flags.skip_remount_private = 0; 197 j->flags.remount_proc_ro = 0; 198 j->flags.pids = 0; 199 j->flags.do_init = 0; 200 j->flags.pid_file = 0; 201 j->flags.cgroups = 0; 202 } 203 204 /* 205 * Strip out flags meant for the child. 206 * We keep things that are inherited across execve(2). 207 */ 208 void minijail_preexec(struct minijail *j) 209 { 210 int vfs = j->flags.vfs; 211 int enter_vfs = j->flags.enter_vfs; 212 int skip_remount_private = j->flags.skip_remount_private; 213 int remount_proc_ro = j->flags.remount_proc_ro; 214 int userns = j->flags.userns; 215 if (j->user) 216 free(j->user); 217 j->user = NULL; 218 if (j->suppl_gid_list) 219 free(j->suppl_gid_list); 220 j->suppl_gid_list = NULL; 221 memset(&j->flags, 0, sizeof(j->flags)); 222 /* Now restore anything we meant to keep. */ 223 j->flags.vfs = vfs; 224 j->flags.enter_vfs = enter_vfs; 225 j->flags.skip_remount_private = skip_remount_private; 226 j->flags.remount_proc_ro = remount_proc_ro; 227 j->flags.userns = userns; 228 /* Note, |pids| will already have been used before this call. */ 229 } 230 231 /* Minijail API. */ 232 233 struct minijail API *minijail_new(void) 234 { 235 return calloc(1, sizeof(struct minijail)); 236 } 237 238 void API minijail_change_uid(struct minijail *j, uid_t uid) 239 { 240 if (uid == 0) 241 die("useless change to uid 0"); 242 j->uid = uid; 243 j->flags.uid = 1; 244 } 245 246 void API minijail_change_gid(struct minijail *j, gid_t gid) 247 { 248 if (gid == 0) 249 die("useless change to gid 0"); 250 j->gid = gid; 251 j->flags.gid = 1; 252 } 253 254 void API minijail_set_supplementary_gids(struct minijail *j, size_t size, 255 const gid_t *list) 256 { 257 size_t i; 258 259 if (j->flags.inherit_suppl_gids) 260 die("cannot inherit *and* set supplementary groups"); 261 if (j->flags.keep_suppl_gids) 262 die("cannot keep *and* set supplementary groups"); 263 264 if (size == 0) { 265 /* Clear supplementary groups. */ 266 j->suppl_gid_list = NULL; 267 j->suppl_gid_count = 0; 268 j->flags.set_suppl_gids = 1; 269 return; 270 } 271 272 /* Copy the gid_t array. */ 273 j->suppl_gid_list = calloc(size, sizeof(gid_t)); 274 if (!j->suppl_gid_list) { 275 die("failed to allocate internal supplementary group array"); 276 } 277 for (i = 0; i < size; i++) { 278 j->suppl_gid_list[i] = list[i]; 279 } 280 j->suppl_gid_count = size; 281 j->flags.set_suppl_gids = 1; 282 } 283 284 void API minijail_keep_supplementary_gids(struct minijail *j) { 285 j->flags.keep_suppl_gids = 1; 286 } 287 288 int API minijail_change_user(struct minijail *j, const char *user) 289 { 290 char *buf = NULL; 291 struct passwd pw; 292 struct passwd *ppw = NULL; 293 ssize_t sz = sysconf(_SC_GETPW_R_SIZE_MAX); 294 if (sz == -1) 295 sz = 65536; /* your guess is as good as mine... */ 296 297 /* 298 * sysconf(_SC_GETPW_R_SIZE_MAX), under glibc, is documented to return 299 * the maximum needed size of the buffer, so we don't have to search. 300 */ 301 buf = malloc(sz); 302 if (!buf) 303 return -ENOMEM; 304 getpwnam_r(user, &pw, buf, sz, &ppw); 305 /* 306 * We're safe to free the buffer here. The strings inside |pw| point 307 * inside |buf|, but we don't use any of them; this leaves the pointers 308 * dangling but it's safe. |ppw| points at |pw| if getpwnam_r(3) 309 * succeeded. 310 */ 311 free(buf); 312 /* getpwnam_r(3) does *not* set errno when |ppw| is NULL. */ 313 if (!ppw) 314 return -1; 315 minijail_change_uid(j, ppw->pw_uid); 316 j->user = strdup(user); 317 if (!j->user) 318 return -ENOMEM; 319 j->usergid = ppw->pw_gid; 320 return 0; 321 } 322 323 int API minijail_change_group(struct minijail *j, const char *group) 324 { 325 char *buf = NULL; 326 struct group gr; 327 struct group *pgr = NULL; 328 ssize_t sz = sysconf(_SC_GETGR_R_SIZE_MAX); 329 if (sz == -1) 330 sz = 65536; /* and mine is as good as yours, really */ 331 332 /* 333 * sysconf(_SC_GETGR_R_SIZE_MAX), under glibc, is documented to return 334 * the maximum needed size of the buffer, so we don't have to search. 335 */ 336 buf = malloc(sz); 337 if (!buf) 338 return -ENOMEM; 339 getgrnam_r(group, &gr, buf, sz, &pgr); 340 /* 341 * We're safe to free the buffer here. The strings inside gr point 342 * inside buf, but we don't use any of them; this leaves the pointers 343 * dangling but it's safe. pgr points at gr if getgrnam_r succeeded. 344 */ 345 free(buf); 346 /* getgrnam_r(3) does *not* set errno when |pgr| is NULL. */ 347 if (!pgr) 348 return -1; 349 minijail_change_gid(j, pgr->gr_gid); 350 return 0; 351 } 352 353 void API minijail_use_seccomp(struct minijail *j) 354 { 355 j->flags.seccomp = 1; 356 } 357 358 void API minijail_no_new_privs(struct minijail *j) 359 { 360 j->flags.no_new_privs = 1; 361 } 362 363 void API minijail_use_seccomp_filter(struct minijail *j) 364 { 365 j->flags.seccomp_filter = 1; 366 } 367 368 void API minijail_set_seccomp_filter_tsync(struct minijail *j) 369 { 370 if (j->filter_len > 0 && j->filter_prog != NULL) { 371 die("minijail_set_seccomp_filter_tsync() must be called " 372 "before minijail_parse_seccomp_filters()"); 373 } 374 j->flags.seccomp_filter_tsync = 1; 375 } 376 377 void API minijail_log_seccomp_filter_failures(struct minijail *j) 378 { 379 if (j->filter_len > 0 && j->filter_prog != NULL) { 380 die("minijail_log_seccomp_filter_failures() must be called " 381 "before minijail_parse_seccomp_filters()"); 382 } 383 j->flags.seccomp_filter_logging = 1; 384 } 385 386 void API minijail_use_caps(struct minijail *j, uint64_t capmask) 387 { 388 /* 389 * 'minijail_use_caps' configures a runtime-capabilities-only 390 * environment, including a bounding set matching the thread's runtime 391 * (permitted|inheritable|effective) sets. 392 * Therefore, it will override any existing bounding set configurations 393 * since the latter would allow gaining extra runtime capabilities from 394 * file capabilities. 395 */ 396 if (j->flags.capbset_drop) { 397 warn("overriding bounding set configuration"); 398 j->cap_bset = 0; 399 j->flags.capbset_drop = 0; 400 } 401 j->caps = capmask; 402 j->flags.use_caps = 1; 403 } 404 405 void API minijail_capbset_drop(struct minijail *j, uint64_t capmask) 406 { 407 if (j->flags.use_caps) { 408 /* 409 * 'minijail_use_caps' will have already configured a capability 410 * bounding set matching the (permitted|inheritable|effective) 411 * sets. Abort if the user tries to configure a separate 412 * bounding set. 'minijail_capbset_drop' and 'minijail_use_caps' 413 * are mutually exclusive. 414 */ 415 die("runtime capabilities already configured, can't drop " 416 "bounding set separately"); 417 } 418 j->cap_bset = capmask; 419 j->flags.capbset_drop = 1; 420 } 421 422 void API minijail_reset_signal_mask(struct minijail *j) 423 { 424 j->flags.reset_signal_mask = 1; 425 } 426 427 void API minijail_namespace_vfs(struct minijail *j) 428 { 429 j->flags.vfs = 1; 430 } 431 432 void API minijail_namespace_enter_vfs(struct minijail *j, const char *ns_path) 433 { 434 int ns_fd = open(ns_path, O_RDONLY | O_CLOEXEC); 435 if (ns_fd < 0) { 436 pdie("failed to open namespace '%s'", ns_path); 437 } 438 j->mountns_fd = ns_fd; 439 j->flags.enter_vfs = 1; 440 } 441 442 void API minijail_new_session_keyring(struct minijail *j) 443 { 444 j->flags.new_session_keyring = 1; 445 } 446 447 void API minijail_skip_remount_private(struct minijail *j) 448 { 449 j->flags.skip_remount_private = 1; 450 } 451 452 void API minijail_namespace_pids(struct minijail *j) 453 { 454 j->flags.vfs = 1; 455 j->flags.remount_proc_ro = 1; 456 j->flags.pids = 1; 457 j->flags.do_init = 1; 458 } 459 460 void API minijail_namespace_ipc(struct minijail *j) 461 { 462 j->flags.ipc = 1; 463 } 464 465 void API minijail_namespace_net(struct minijail *j) 466 { 467 j->flags.net = 1; 468 } 469 470 void API minijail_namespace_enter_net(struct minijail *j, const char *ns_path) 471 { 472 int ns_fd = open(ns_path, O_RDONLY | O_CLOEXEC); 473 if (ns_fd < 0) { 474 pdie("failed to open namespace '%s'", ns_path); 475 } 476 j->netns_fd = ns_fd; 477 j->flags.enter_net = 1; 478 } 479 480 void API minijail_namespace_cgroups(struct minijail *j) 481 { 482 j->flags.ns_cgroups = 1; 483 } 484 485 void API minijail_close_open_fds(struct minijail *j) 486 { 487 j->flags.close_open_fds = 1; 488 } 489 490 void API minijail_remount_proc_readonly(struct minijail *j) 491 { 492 j->flags.vfs = 1; 493 j->flags.remount_proc_ro = 1; 494 } 495 496 void API minijail_namespace_user(struct minijail *j) 497 { 498 j->flags.userns = 1; 499 } 500 501 void API minijail_namespace_user_disable_setgroups(struct minijail *j) 502 { 503 j->flags.disable_setgroups = 1; 504 } 505 506 int API minijail_uidmap(struct minijail *j, const char *uidmap) 507 { 508 j->uidmap = strdup(uidmap); 509 if (!j->uidmap) 510 return -ENOMEM; 511 char *ch; 512 for (ch = j->uidmap; *ch; ch++) { 513 if (*ch == ',') 514 *ch = '\n'; 515 } 516 return 0; 517 } 518 519 int API minijail_gidmap(struct minijail *j, const char *gidmap) 520 { 521 j->gidmap = strdup(gidmap); 522 if (!j->gidmap) 523 return -ENOMEM; 524 char *ch; 525 for (ch = j->gidmap; *ch; ch++) { 526 if (*ch == ',') 527 *ch = '\n'; 528 } 529 return 0; 530 } 531 532 void API minijail_inherit_usergroups(struct minijail *j) 533 { 534 j->flags.inherit_suppl_gids = 1; 535 } 536 537 void API minijail_run_as_init(struct minijail *j) 538 { 539 /* 540 * Since the jailed program will become 'init' in the new PID namespace, 541 * Minijail does not need to fork an 'init' process. 542 */ 543 j->flags.do_init = 0; 544 } 545 546 int API minijail_enter_chroot(struct minijail *j, const char *dir) 547 { 548 if (j->chrootdir) 549 return -EINVAL; 550 j->chrootdir = strdup(dir); 551 if (!j->chrootdir) 552 return -ENOMEM; 553 j->flags.chroot = 1; 554 return 0; 555 } 556 557 int API minijail_enter_pivot_root(struct minijail *j, const char *dir) 558 { 559 if (j->chrootdir) 560 return -EINVAL; 561 j->chrootdir = strdup(dir); 562 if (!j->chrootdir) 563 return -ENOMEM; 564 j->flags.pivot_root = 1; 565 return 0; 566 } 567 568 char API *minijail_get_original_path(struct minijail *j, 569 const char *path_inside_chroot) 570 { 571 struct mountpoint *b; 572 573 b = j->mounts_head; 574 while (b) { 575 /* 576 * If |path_inside_chroot| is the exact destination of a 577 * mount, then the original path is exactly the source of 578 * the mount. 579 * for example: "-b /some/path/exe,/chroot/path/exe" 580 * mount source = /some/path/exe, mount dest = 581 * /chroot/path/exe Then when getting the original path of 582 * "/chroot/path/exe", the source of that mount, 583 * "/some/path/exe" is what should be returned. 584 */ 585 if (!strcmp(b->dest, path_inside_chroot)) 586 return strdup(b->src); 587 588 /* 589 * If |path_inside_chroot| is within the destination path of a 590 * mount, take the suffix of the chroot path relative to the 591 * mount destination path, and append it to the mount source 592 * path. 593 */ 594 if (!strncmp(b->dest, path_inside_chroot, strlen(b->dest))) { 595 const char *relative_path = 596 path_inside_chroot + strlen(b->dest); 597 return path_join(b->src, relative_path); 598 } 599 b = b->next; 600 } 601 602 /* If there is a chroot path, append |path_inside_chroot| to that. */ 603 if (j->chrootdir) 604 return path_join(j->chrootdir, path_inside_chroot); 605 606 /* No chroot, so the path outside is the same as it is inside. */ 607 return strdup(path_inside_chroot); 608 } 609 610 size_t minijail_get_tmpfs_size(const struct minijail *j) 611 { 612 return j->tmpfs_size; 613 } 614 615 void API minijail_mount_tmp(struct minijail *j) 616 { 617 minijail_mount_tmp_size(j, 64 * 1024 * 1024); 618 } 619 620 void API minijail_mount_tmp_size(struct minijail *j, size_t size) 621 { 622 j->tmpfs_size = size; 623 j->flags.mount_tmp = 1; 624 } 625 626 int API minijail_write_pid_file(struct minijail *j, const char *path) 627 { 628 j->pid_file_path = strdup(path); 629 if (!j->pid_file_path) 630 return -ENOMEM; 631 j->flags.pid_file = 1; 632 return 0; 633 } 634 635 int API minijail_add_to_cgroup(struct minijail *j, const char *path) 636 { 637 if (j->cgroup_count >= MAX_CGROUPS) 638 return -ENOMEM; 639 j->cgroups[j->cgroup_count] = strdup(path); 640 if (!j->cgroups[j->cgroup_count]) 641 return -ENOMEM; 642 j->cgroup_count++; 643 j->flags.cgroups = 1; 644 return 0; 645 } 646 647 int API minijail_mount_with_data(struct minijail *j, const char *src, 648 const char *dest, const char *type, 649 unsigned long flags, const char *data) 650 { 651 struct mountpoint *m; 652 653 if (*dest != '/') 654 return -EINVAL; 655 m = calloc(1, sizeof(*m)); 656 if (!m) 657 return -ENOMEM; 658 m->dest = strdup(dest); 659 if (!m->dest) 660 goto error; 661 m->src = strdup(src); 662 if (!m->src) 663 goto error; 664 m->type = strdup(type); 665 if (!m->type) 666 goto error; 667 if (data) { 668 m->data = strdup(data); 669 if (!m->data) 670 goto error; 671 m->has_data = 1; 672 } 673 m->flags = flags; 674 675 info("mount %s -> %s type '%s'", src, dest, type); 676 677 /* 678 * Force vfs namespacing so the mounts don't leak out into the 679 * containing vfs namespace. 680 */ 681 minijail_namespace_vfs(j); 682 683 if (j->mounts_tail) 684 j->mounts_tail->next = m; 685 else 686 j->mounts_head = m; 687 j->mounts_tail = m; 688 j->mounts_count++; 689 690 return 0; 691 692 error: 693 free(m->type); 694 free(m->src); 695 free(m->dest); 696 free(m); 697 return -ENOMEM; 698 } 699 700 int API minijail_mount(struct minijail *j, const char *src, const char *dest, 701 const char *type, unsigned long flags) 702 { 703 return minijail_mount_with_data(j, src, dest, type, flags, NULL); 704 } 705 706 int API minijail_bind(struct minijail *j, const char *src, const char *dest, 707 int writeable) 708 { 709 unsigned long flags = MS_BIND; 710 711 if (!writeable) 712 flags |= MS_RDONLY; 713 714 return minijail_mount(j, src, dest, "", flags); 715 } 716 717 static void clear_seccomp_options(struct minijail *j) 718 { 719 j->flags.seccomp_filter = 0; 720 j->flags.seccomp_filter_tsync = 0; 721 j->flags.seccomp_filter_logging = 0; 722 j->filter_len = 0; 723 j->filter_prog = NULL; 724 j->flags.no_new_privs = 0; 725 } 726 727 static int seccomp_should_parse_filters(struct minijail *j) 728 { 729 if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, NULL) == -1) { 730 /* 731 * |errno| will be set to EINVAL when seccomp has not been 732 * compiled into the kernel. On certain platforms and kernel 733 * versions this is not a fatal failure. In that case, and only 734 * in that case, disable seccomp and skip loading the filters. 735 */ 736 if ((errno == EINVAL) && seccomp_can_softfail()) { 737 warn("not loading seccomp filters, seccomp filter not " 738 "supported"); 739 clear_seccomp_options(j); 740 return 0; 741 } 742 /* 743 * If |errno| != EINVAL or seccomp_can_softfail() is false, 744 * we can proceed. Worst case scenario minijail_enter() will 745 * abort() if seccomp fails. 746 */ 747 } 748 if (j->flags.seccomp_filter_tsync) { 749 /* Are the seccomp(2) syscall and the TSYNC option supported? */ 750 if (sys_seccomp(SECCOMP_SET_MODE_FILTER, 751 SECCOMP_FILTER_FLAG_TSYNC, NULL) == -1) { 752 int saved_errno = errno; 753 if (saved_errno == ENOSYS && seccomp_can_softfail()) { 754 warn("seccomp(2) syscall not supported"); 755 clear_seccomp_options(j); 756 return 0; 757 } else if (saved_errno == EINVAL && 758 seccomp_can_softfail()) { 759 warn( 760 "seccomp filter thread sync not supported"); 761 clear_seccomp_options(j); 762 return 0; 763 } 764 /* 765 * Similar logic here. If seccomp_can_softfail() is 766 * false, or |errno| != ENOSYS, or |errno| != EINVAL, 767 * we can proceed. Worst case scenario minijail_enter() 768 * will abort() if seccomp or TSYNC fail. 769 */ 770 } 771 } 772 return 1; 773 } 774 775 static int parse_seccomp_filters(struct minijail *j, FILE *policy_file) 776 { 777 struct sock_fprog *fprog = malloc(sizeof(struct sock_fprog)); 778 int use_ret_trap = 779 j->flags.seccomp_filter_tsync || j->flags.seccomp_filter_logging; 780 int allow_logging = j->flags.seccomp_filter_logging; 781 782 if (compile_filter(policy_file, fprog, use_ret_trap, allow_logging)) { 783 free(fprog); 784 return -1; 785 } 786 787 j->filter_len = fprog->len; 788 j->filter_prog = fprog; 789 return 0; 790 } 791 792 void API minijail_parse_seccomp_filters(struct minijail *j, const char *path) 793 { 794 if (!seccomp_should_parse_filters(j)) 795 return; 796 797 FILE *file = fopen(path, "r"); 798 if (!file) { 799 pdie("failed to open seccomp filter file '%s'", path); 800 } 801 802 if (parse_seccomp_filters(j, file) != 0) { 803 die("failed to compile seccomp filter BPF program in '%s'", 804 path); 805 } 806 fclose(file); 807 } 808 809 void API minijail_parse_seccomp_filters_from_fd(struct minijail *j, int fd) 810 { 811 if (!seccomp_should_parse_filters(j)) 812 return; 813 814 FILE *file = fdopen(fd, "r"); 815 if (!file) { 816 pdie("failed to associate stream with fd %d", fd); 817 } 818 819 if (parse_seccomp_filters(j, file) != 0) { 820 die("failed to compile seccomp filter BPF program from fd %d", 821 fd); 822 } 823 fclose(file); 824 } 825 826 int API minijail_use_alt_syscall(struct minijail *j, const char *table) 827 { 828 j->alt_syscall_table = strdup(table); 829 if (!j->alt_syscall_table) 830 return -ENOMEM; 831 j->flags.alt_syscall = 1; 832 return 0; 833 } 834 835 struct marshal_state { 836 size_t available; 837 size_t total; 838 char *buf; 839 }; 840 841 void marshal_state_init(struct marshal_state *state, char *buf, 842 size_t available) 843 { 844 state->available = available; 845 state->buf = buf; 846 state->total = 0; 847 } 848 849 void marshal_append(struct marshal_state *state, void *src, size_t length) 850 { 851 size_t copy_len = MIN(state->available, length); 852 853 /* Up to |available| will be written. */ 854 if (copy_len) { 855 memcpy(state->buf, src, copy_len); 856 state->buf += copy_len; 857 state->available -= copy_len; 858 } 859 /* |total| will contain the expected length. */ 860 state->total += length; 861 } 862 863 void marshal_mount(struct marshal_state *state, const struct mountpoint *m) 864 { 865 marshal_append(state, m->src, strlen(m->src) + 1); 866 marshal_append(state, m->dest, strlen(m->dest) + 1); 867 marshal_append(state, m->type, strlen(m->type) + 1); 868 marshal_append(state, (char *)&m->has_data, sizeof(m->has_data)); 869 if (m->has_data) 870 marshal_append(state, m->data, strlen(m->data) + 1); 871 marshal_append(state, (char *)&m->flags, sizeof(m->flags)); 872 } 873 874 void minijail_marshal_helper(struct marshal_state *state, 875 const struct minijail *j) 876 { 877 struct mountpoint *m = NULL; 878 size_t i; 879 880 marshal_append(state, (char *)j, sizeof(*j)); 881 if (j->user) 882 marshal_append(state, j->user, strlen(j->user) + 1); 883 if (j->suppl_gid_list) { 884 marshal_append(state, j->suppl_gid_list, 885 j->suppl_gid_count * sizeof(gid_t)); 886 } 887 if (j->chrootdir) 888 marshal_append(state, j->chrootdir, strlen(j->chrootdir) + 1); 889 if (j->alt_syscall_table) { 890 marshal_append(state, j->alt_syscall_table, 891 strlen(j->alt_syscall_table) + 1); 892 } 893 if (j->flags.seccomp_filter && j->filter_prog) { 894 struct sock_fprog *fp = j->filter_prog; 895 marshal_append(state, (char *)fp->filter, 896 fp->len * sizeof(struct sock_filter)); 897 } 898 for (m = j->mounts_head; m; m = m->next) { 899 marshal_mount(state, m); 900 } 901 for (i = 0; i < j->cgroup_count; ++i) 902 marshal_append(state, j->cgroups[i], strlen(j->cgroups[i]) + 1); 903 } 904 905 size_t API minijail_size(const struct minijail *j) 906 { 907 struct marshal_state state; 908 marshal_state_init(&state, NULL, 0); 909 minijail_marshal_helper(&state, j); 910 return state.total; 911 } 912 913 int minijail_marshal(const struct minijail *j, char *buf, size_t available) 914 { 915 struct marshal_state state; 916 marshal_state_init(&state, buf, available); 917 minijail_marshal_helper(&state, j); 918 return (state.total > available); 919 } 920 921 int minijail_unmarshal(struct minijail *j, char *serialized, size_t length) 922 { 923 size_t i; 924 size_t count; 925 int ret = -EINVAL; 926 927 if (length < sizeof(*j)) 928 goto out; 929 memcpy((void *)j, serialized, sizeof(*j)); 930 serialized += sizeof(*j); 931 length -= sizeof(*j); 932 933 /* Potentially stale pointers not used as signals. */ 934 j->pid_file_path = NULL; 935 j->uidmap = NULL; 936 j->gidmap = NULL; 937 j->mounts_head = NULL; 938 j->mounts_tail = NULL; 939 j->filter_prog = NULL; 940 941 if (j->user) { /* stale pointer */ 942 char *user = consumestr(&serialized, &length); 943 if (!user) 944 goto clear_pointers; 945 j->user = strdup(user); 946 if (!j->user) 947 goto clear_pointers; 948 } 949 950 if (j->suppl_gid_list) { /* stale pointer */ 951 if (j->suppl_gid_count > NGROUPS_MAX) { 952 goto bad_gid_list; 953 } 954 size_t gid_list_size = j->suppl_gid_count * sizeof(gid_t); 955 void *gid_list_bytes = 956 consumebytes(gid_list_size, &serialized, &length); 957 if (!gid_list_bytes) 958 goto bad_gid_list; 959 960 j->suppl_gid_list = calloc(j->suppl_gid_count, sizeof(gid_t)); 961 if (!j->suppl_gid_list) 962 goto bad_gid_list; 963 964 memcpy(j->suppl_gid_list, gid_list_bytes, gid_list_size); 965 } 966 967 if (j->chrootdir) { /* stale pointer */ 968 char *chrootdir = consumestr(&serialized, &length); 969 if (!chrootdir) 970 goto bad_chrootdir; 971 j->chrootdir = strdup(chrootdir); 972 if (!j->chrootdir) 973 goto bad_chrootdir; 974 } 975 976 if (j->alt_syscall_table) { /* stale pointer */ 977 char *alt_syscall_table = consumestr(&serialized, &length); 978 if (!alt_syscall_table) 979 goto bad_syscall_table; 980 j->alt_syscall_table = strdup(alt_syscall_table); 981 if (!j->alt_syscall_table) 982 goto bad_syscall_table; 983 } 984 985 if (j->flags.seccomp_filter && j->filter_len > 0) { 986 size_t ninstrs = j->filter_len; 987 if (ninstrs > (SIZE_MAX / sizeof(struct sock_filter)) || 988 ninstrs > USHRT_MAX) 989 goto bad_filters; 990 991 size_t program_len = ninstrs * sizeof(struct sock_filter); 992 void *program = consumebytes(program_len, &serialized, &length); 993 if (!program) 994 goto bad_filters; 995 996 j->filter_prog = malloc(sizeof(struct sock_fprog)); 997 if (!j->filter_prog) 998 goto bad_filters; 999 1000 j->filter_prog->len = ninstrs; 1001 j->filter_prog->filter = malloc(program_len); 1002 if (!j->filter_prog->filter) 1003 goto bad_filter_prog_instrs; 1004 1005 memcpy(j->filter_prog->filter, program, program_len); 1006 } 1007 1008 count = j->mounts_count; 1009 j->mounts_count = 0; 1010 for (i = 0; i < count; ++i) { 1011 unsigned long *flags; 1012 int *has_data; 1013 const char *dest; 1014 const char *type; 1015 const char *data = NULL; 1016 const char *src = consumestr(&serialized, &length); 1017 if (!src) 1018 goto bad_mounts; 1019 dest = consumestr(&serialized, &length); 1020 if (!dest) 1021 goto bad_mounts; 1022 type = consumestr(&serialized, &length); 1023 if (!type) 1024 goto bad_mounts; 1025 has_data = consumebytes(sizeof(*has_data), &serialized, 1026 &length); 1027 if (!has_data) 1028 goto bad_mounts; 1029 if (*has_data) { 1030 data = consumestr(&serialized, &length); 1031 if (!data) 1032 goto bad_mounts; 1033 } 1034 flags = consumebytes(sizeof(*flags), &serialized, &length); 1035 if (!flags) 1036 goto bad_mounts; 1037 if (minijail_mount_with_data(j, src, dest, type, *flags, data)) 1038 goto bad_mounts; 1039 } 1040 1041 count = j->cgroup_count; 1042 j->cgroup_count = 0; 1043 for (i = 0; i < count; ++i) { 1044 char *cgroup = consumestr(&serialized, &length); 1045 if (!cgroup) 1046 goto bad_cgroups; 1047 j->cgroups[i] = strdup(cgroup); 1048 if (!j->cgroups[i]) 1049 goto bad_cgroups; 1050 ++j->cgroup_count; 1051 } 1052 1053 return 0; 1054 1055 bad_cgroups: 1056 while (j->mounts_head) { 1057 struct mountpoint *m = j->mounts_head; 1058 j->mounts_head = j->mounts_head->next; 1059 free(m->data); 1060 free(m->type); 1061 free(m->dest); 1062 free(m->src); 1063 free(m); 1064 } 1065 for (i = 0; i < j->cgroup_count; ++i) 1066 free(j->cgroups[i]); 1067 bad_mounts: 1068 if (j->flags.seccomp_filter && j->filter_len > 0) { 1069 free(j->filter_prog->filter); 1070 free(j->filter_prog); 1071 } 1072 bad_filter_prog_instrs: 1073 if (j->filter_prog) 1074 free(j->filter_prog); 1075 bad_filters: 1076 if (j->alt_syscall_table) 1077 free(j->alt_syscall_table); 1078 bad_syscall_table: 1079 if (j->chrootdir) 1080 free(j->chrootdir); 1081 bad_chrootdir: 1082 if (j->suppl_gid_list) 1083 free(j->suppl_gid_list); 1084 bad_gid_list: 1085 if (j->user) 1086 free(j->user); 1087 clear_pointers: 1088 j->user = NULL; 1089 j->suppl_gid_list = NULL; 1090 j->chrootdir = NULL; 1091 j->alt_syscall_table = NULL; 1092 j->cgroup_count = 0; 1093 out: 1094 return ret; 1095 } 1096 1097 /* 1098 * setup_mount_destination: Ensures the mount target exists. 1099 * Creates it if needed and possible. 1100 */ 1101 static int setup_mount_destination(const char *source, const char *dest, 1102 uid_t uid, uid_t gid) 1103 { 1104 int rc; 1105 struct stat st_buf; 1106 1107 rc = stat(dest, &st_buf); 1108 if (rc == 0) /* destination exists */ 1109 return 0; 1110 1111 /* 1112 * Try to create the destination. 1113 * Either make a directory or touch a file depending on the source type. 1114 * If the source doesn't exist, assume it is a filesystem type such as 1115 * "tmpfs" and create a directory to mount it on. 1116 */ 1117 rc = stat(source, &st_buf); 1118 if (rc || S_ISDIR(st_buf.st_mode) || S_ISBLK(st_buf.st_mode)) { 1119 if (mkdir(dest, 0700)) 1120 return -errno; 1121 } else { 1122 int fd = open(dest, O_RDWR | O_CREAT, 0700); 1123 if (fd < 0) 1124 return -errno; 1125 close(fd); 1126 } 1127 return chown(dest, uid, gid); 1128 } 1129 1130 /* 1131 * mount_one: Applies mounts from @m for @j, recursing as needed. 1132 * @j Minijail these mounts are for 1133 * @m Head of list of mounts 1134 * 1135 * Returns 0 for success. 1136 */ 1137 static int mount_one(const struct minijail *j, struct mountpoint *m) 1138 { 1139 int ret; 1140 char *dest; 1141 int remount_ro = 0; 1142 1143 /* |dest| has a leading "/". */ 1144 if (asprintf(&dest, "%s%s", j->chrootdir, m->dest) < 0) 1145 return -ENOMEM; 1146 1147 if (setup_mount_destination(m->src, dest, j->uid, j->gid)) 1148 pdie("creating mount target '%s' failed", dest); 1149 1150 /* 1151 * R/O bind mounts have to be remounted since 'bind' and 'ro' 1152 * can't both be specified in the original bind mount. 1153 * Remount R/O after the initial mount. 1154 */ 1155 if ((m->flags & MS_BIND) && (m->flags & MS_RDONLY)) { 1156 remount_ro = 1; 1157 m->flags &= ~MS_RDONLY; 1158 } 1159 1160 ret = mount(m->src, dest, m->type, m->flags, m->data); 1161 if (ret) 1162 pdie("mount: %s -> %s", m->src, dest); 1163 1164 if (remount_ro) { 1165 m->flags |= MS_RDONLY; 1166 ret = mount(m->src, dest, NULL, 1167 m->flags | MS_REMOUNT, m->data); 1168 if (ret) 1169 pdie("bind ro: %s -> %s", m->src, dest); 1170 } 1171 1172 free(dest); 1173 if (m->next) 1174 return mount_one(j, m->next); 1175 return ret; 1176 } 1177 1178 static int enter_chroot(const struct minijail *j) 1179 { 1180 int ret; 1181 1182 if (j->mounts_head && (ret = mount_one(j, j->mounts_head))) 1183 return ret; 1184 1185 if (chroot(j->chrootdir)) 1186 return -errno; 1187 1188 if (chdir("/")) 1189 return -errno; 1190 1191 return 0; 1192 } 1193 1194 static int enter_pivot_root(const struct minijail *j) 1195 { 1196 int ret, oldroot, newroot; 1197 1198 if (j->mounts_head && (ret = mount_one(j, j->mounts_head))) 1199 return ret; 1200 1201 /* 1202 * Keep the fd for both old and new root. 1203 * It will be used in fchdir(2) later. 1204 */ 1205 oldroot = open("/", O_DIRECTORY | O_RDONLY | O_CLOEXEC); 1206 if (oldroot < 0) 1207 pdie("failed to open / for fchdir"); 1208 newroot = open(j->chrootdir, O_DIRECTORY | O_RDONLY | O_CLOEXEC); 1209 if (newroot < 0) 1210 pdie("failed to open %s for fchdir", j->chrootdir); 1211 1212 /* 1213 * To ensure j->chrootdir is the root of a filesystem, 1214 * do a self bind mount. 1215 */ 1216 if (mount(j->chrootdir, j->chrootdir, "bind", MS_BIND | MS_REC, "")) 1217 pdie("failed to bind mount '%s'", j->chrootdir); 1218 if (chdir(j->chrootdir)) 1219 return -errno; 1220 if (syscall(SYS_pivot_root, ".", ".")) 1221 pdie("pivot_root"); 1222 1223 /* 1224 * Now the old root is mounted on top of the new root. Use fchdir(2) to 1225 * change to the old root and unmount it. 1226 */ 1227 if (fchdir(oldroot)) 1228 pdie("failed to fchdir to old /"); 1229 1230 /* 1231 * If j->flags.skip_remount_private was enabled for minijail_enter(), 1232 * there could be a shared mount point under |oldroot|. In that case, 1233 * mounts under this shared mount point will be unmounted below, and 1234 * this unmounting will propagate to the original mount namespace 1235 * (because the mount point is shared). To prevent this unexpected 1236 * unmounting, remove these mounts from their peer groups by recursively 1237 * remounting them as MS_PRIVATE. 1238 */ 1239 if (mount(NULL, ".", NULL, MS_REC | MS_PRIVATE, NULL)) 1240 pdie("failed to mount(/, private) before umount(/)"); 1241 /* The old root might be busy, so use lazy unmount. */ 1242 if (umount2(".", MNT_DETACH)) 1243 pdie("umount(/)"); 1244 /* Change back to the new root. */ 1245 if (fchdir(newroot)) 1246 return -errno; 1247 if (close(oldroot)) 1248 return -errno; 1249 if (close(newroot)) 1250 return -errno; 1251 if (chroot("/")) 1252 return -errno; 1253 /* Set correct CWD for getcwd(3). */ 1254 if (chdir("/")) 1255 return -errno; 1256 1257 return 0; 1258 } 1259 1260 static int mount_tmp(const struct minijail *j) 1261 { 1262 const char fmt[] = "size=%zu,mode=1777"; 1263 /* Count for the user storing ULLONG_MAX literally + extra space. */ 1264 char data[sizeof(fmt) + sizeof("18446744073709551615ULL")]; 1265 int ret; 1266 1267 ret = snprintf(data, sizeof(data), fmt, j->tmpfs_size); 1268 1269 if (ret <= 0) 1270 pdie("tmpfs size spec error"); 1271 else if ((size_t)ret >= sizeof(data)) 1272 pdie("tmpfs size spec too large"); 1273 return mount("none", "/tmp", "tmpfs", MS_NODEV | MS_NOEXEC | MS_NOSUID, 1274 data); 1275 } 1276 1277 static int remount_proc_readonly(const struct minijail *j) 1278 { 1279 const char *kProcPath = "/proc"; 1280 const unsigned int kSafeFlags = MS_NODEV | MS_NOEXEC | MS_NOSUID; 1281 /* 1282 * Right now, we're holding a reference to our parent's old mount of 1283 * /proc in our namespace, which means using MS_REMOUNT here would 1284 * mutate our parent's mount as well, even though we're in a VFS 1285 * namespace (!). Instead, remove their mount from our namespace lazily 1286 * (MNT_DETACH) and make our own. 1287 */ 1288 if (umount2(kProcPath, MNT_DETACH)) { 1289 /* 1290 * If we are in a new user namespace, umount(2) will fail. 1291 * See http://man7.org/linux/man-pages/man7/user_namespaces.7.html 1292 */ 1293 if (j->flags.userns) { 1294 info("umount(/proc, MNT_DETACH) failed, " 1295 "this is expected when using user namespaces"); 1296 } else { 1297 return -errno; 1298 } 1299 } 1300 if (mount("proc", kProcPath, "proc", kSafeFlags | MS_RDONLY, "")) 1301 return -errno; 1302 return 0; 1303 } 1304 1305 static void kill_child_and_die(const struct minijail *j, const char *msg) 1306 { 1307 kill(j->initpid, SIGKILL); 1308 die("%s", msg); 1309 } 1310 1311 static void write_pid_file_or_die(const struct minijail *j) 1312 { 1313 if (write_pid_to_path(j->initpid, j->pid_file_path)) 1314 kill_child_and_die(j, "failed to write pid file"); 1315 } 1316 1317 static void add_to_cgroups_or_die(const struct minijail *j) 1318 { 1319 size_t i; 1320 1321 for (i = 0; i < j->cgroup_count; ++i) { 1322 if (write_pid_to_path(j->initpid, j->cgroups[i])) 1323 kill_child_and_die(j, "failed to add to cgroups"); 1324 } 1325 } 1326 1327 static void write_ugid_maps_or_die(const struct minijail *j) 1328 { 1329 if (j->uidmap && write_proc_file(j->initpid, j->uidmap, "uid_map") != 0) 1330 kill_child_and_die(j, "failed to write uid_map"); 1331 if (j->gidmap && j->flags.disable_setgroups) { 1332 /* Older kernels might not have the /proc/<pid>/setgroups files. */ 1333 int ret = write_proc_file(j->initpid, "deny", "setgroups"); 1334 if (ret != 0) { 1335 if (ret == -ENOENT) { 1336 /* See http://man7.org/linux/man-pages/man7/user_namespaces.7.html. */ 1337 warn("could not disable setgroups(2)"); 1338 } else 1339 kill_child_and_die(j, "failed to disable setgroups(2)"); 1340 } 1341 } 1342 if (j->gidmap && write_proc_file(j->initpid, j->gidmap, "gid_map") != 0) 1343 kill_child_and_die(j, "failed to write gid_map"); 1344 } 1345 1346 static void enter_user_namespace(const struct minijail *j) 1347 { 1348 if (j->uidmap && setresuid(0, 0, 0)) 1349 pdie("user_namespaces: setresuid(0, 0, 0) failed"); 1350 if (j->gidmap && setresgid(0, 0, 0)) 1351 pdie("user_namespaces: setresgid(0, 0, 0) failed"); 1352 } 1353 1354 static void parent_setup_complete(int *pipe_fds) 1355 { 1356 close(pipe_fds[0]); 1357 close(pipe_fds[1]); 1358 } 1359 1360 /* 1361 * wait_for_parent_setup: Called by the child process to wait for any 1362 * further parent-side setup to complete before continuing. 1363 */ 1364 static void wait_for_parent_setup(int *pipe_fds) 1365 { 1366 char buf; 1367 1368 close(pipe_fds[1]); 1369 1370 /* Wait for parent to complete setup and close the pipe. */ 1371 if (read(pipe_fds[0], &buf, 1) != 0) 1372 die("failed to sync with parent"); 1373 close(pipe_fds[0]); 1374 } 1375 1376 static void drop_ugid(const struct minijail *j) 1377 { 1378 if (j->flags.inherit_suppl_gids + j->flags.keep_suppl_gids + 1379 j->flags.set_suppl_gids > 1) { 1380 die("can only do one of inherit, keep, or set supplementary " 1381 "groups"); 1382 } 1383 1384 if (j->flags.inherit_suppl_gids) { 1385 if (initgroups(j->user, j->usergid)) 1386 pdie("initgroups(%s, %d) failed", j->user, j->usergid); 1387 } else if (j->flags.set_suppl_gids) { 1388 if (setgroups(j->suppl_gid_count, j->suppl_gid_list)) 1389 pdie("setgroups(suppl_gids) failed"); 1390 } else if (!j->flags.keep_suppl_gids) { 1391 /* 1392 * Only attempt to clear supplementary groups if we are changing 1393 * users or groups. 1394 */ 1395 if ((j->flags.uid || j->flags.gid) && setgroups(0, NULL)) 1396 pdie("setgroups(0, NULL) failed"); 1397 } 1398 1399 if (j->flags.gid && setresgid(j->gid, j->gid, j->gid)) 1400 pdie("setresgid(%d, %d, %d) failed", j->gid, j->gid, j->gid); 1401 1402 if (j->flags.uid && setresuid(j->uid, j->uid, j->uid)) 1403 pdie("setresuid(%d, %d, %d) failed", j->uid, j->uid, j->uid); 1404 } 1405 1406 /* 1407 * We specifically do not use cap_valid() as that only tells us the last 1408 * valid cap we were *compiled* against (i.e. what the version of kernel 1409 * headers says). If we run on a different kernel version, then it's not 1410 * uncommon for that to be less (if an older kernel) or more (if a newer 1411 * kernel). 1412 * Normally, we suck up the answer via /proc. On Android, not all processes are 1413 * guaranteed to be able to access '/proc/sys/kernel/cap_last_cap' so we 1414 * programmatically find the value by calling prctl(PR_CAPBSET_READ). 1415 */ 1416 static unsigned int get_last_valid_cap() 1417 { 1418 unsigned int last_valid_cap = 0; 1419 if (is_android()) { 1420 for (; prctl(PR_CAPBSET_READ, last_valid_cap, 0, 0, 0) >= 0; 1421 ++last_valid_cap); 1422 1423 /* |last_valid_cap| will be the first failing value. */ 1424 if (last_valid_cap > 0) { 1425 last_valid_cap--; 1426 } 1427 } else { 1428 const char cap_file[] = "/proc/sys/kernel/cap_last_cap"; 1429 FILE *fp = fopen(cap_file, "re"); 1430 if (fscanf(fp, "%u", &last_valid_cap) != 1) 1431 pdie("fscanf(%s)", cap_file); 1432 fclose(fp); 1433 } 1434 return last_valid_cap; 1435 } 1436 1437 static void drop_capbset(uint64_t keep_mask, unsigned int last_valid_cap) 1438 { 1439 const uint64_t one = 1; 1440 unsigned int i; 1441 for (i = 0; i < sizeof(keep_mask) * 8 && i <= last_valid_cap; ++i) { 1442 if (keep_mask & (one << i)) 1443 continue; 1444 if (prctl(PR_CAPBSET_DROP, i)) 1445 pdie("could not drop capability from bounding set"); 1446 } 1447 } 1448 1449 static void drop_caps(const struct minijail *j, unsigned int last_valid_cap) 1450 { 1451 if (!j->flags.use_caps) 1452 return; 1453 1454 cap_t caps = cap_get_proc(); 1455 cap_value_t flag[1]; 1456 const uint64_t one = 1; 1457 unsigned int i; 1458 if (!caps) 1459 die("can't get process caps"); 1460 if (cap_clear_flag(caps, CAP_INHERITABLE)) 1461 die("can't clear inheritable caps"); 1462 if (cap_clear_flag(caps, CAP_EFFECTIVE)) 1463 die("can't clear effective caps"); 1464 if (cap_clear_flag(caps, CAP_PERMITTED)) 1465 die("can't clear permitted caps"); 1466 for (i = 0; i < sizeof(j->caps) * 8 && i <= last_valid_cap; ++i) { 1467 /* Keep CAP_SETPCAP for dropping bounding set bits. */ 1468 if (i != CAP_SETPCAP && !(j->caps & (one << i))) 1469 continue; 1470 flag[0] = i; 1471 if (cap_set_flag(caps, CAP_EFFECTIVE, 1, flag, CAP_SET)) 1472 die("can't add effective cap"); 1473 if (cap_set_flag(caps, CAP_PERMITTED, 1, flag, CAP_SET)) 1474 die("can't add permitted cap"); 1475 if (cap_set_flag(caps, CAP_INHERITABLE, 1, flag, CAP_SET)) 1476 die("can't add inheritable cap"); 1477 } 1478 if (cap_set_proc(caps)) 1479 die("can't apply initial cleaned capset"); 1480 1481 /* 1482 * Instead of dropping bounding set first, do it here in case 1483 * the caller had a more permissive bounding set which could 1484 * have been used above to raise a capability that wasn't already 1485 * present. This requires CAP_SETPCAP, so we raised/kept it above. 1486 */ 1487 drop_capbset(j->caps, last_valid_cap); 1488 1489 /* If CAP_SETPCAP wasn't specifically requested, now we remove it. */ 1490 if ((j->caps & (one << CAP_SETPCAP)) == 0) { 1491 flag[0] = CAP_SETPCAP; 1492 if (cap_set_flag(caps, CAP_EFFECTIVE, 1, flag, CAP_CLEAR)) 1493 die("can't clear effective cap"); 1494 if (cap_set_flag(caps, CAP_PERMITTED, 1, flag, CAP_CLEAR)) 1495 die("can't clear permitted cap"); 1496 if (cap_set_flag(caps, CAP_INHERITABLE, 1, flag, CAP_CLEAR)) 1497 die("can't clear inheritable cap"); 1498 } 1499 1500 if (cap_set_proc(caps)) 1501 die("can't apply final cleaned capset"); 1502 1503 cap_free(caps); 1504 } 1505 1506 static void set_seccomp_filter(const struct minijail *j) 1507 { 1508 /* 1509 * Set no_new_privs. See </kernel/seccomp.c> and </kernel/sys.c> 1510 * in the kernel source tree for an explanation of the parameters. 1511 */ 1512 if (j->flags.no_new_privs) { 1513 if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0)) 1514 pdie("prctl(PR_SET_NO_NEW_PRIVS)"); 1515 } 1516 1517 /* 1518 * Code running with ASan 1519 * (https://github.com/google/sanitizers/wiki/AddressSanitizer) 1520 * will make system calls not included in the syscall filter policy, 1521 * which will likely crash the program. Skip setting seccomp filter in 1522 * that case. 1523 * 'running_with_asan()' has no inputs and is completely defined at 1524 * build time, so this cannot be used by an attacker to skip setting 1525 * seccomp filter. 1526 */ 1527 if (j->flags.seccomp_filter && running_with_asan()) { 1528 warn("running with ASan, not setting seccomp filter"); 1529 return; 1530 } 1531 1532 if (j->flags.seccomp_filter) { 1533 if (j->flags.seccomp_filter_logging) { 1534 /* 1535 * If logging seccomp filter failures, 1536 * install the SIGSYS handler first. 1537 */ 1538 if (install_sigsys_handler()) 1539 pdie("failed to install SIGSYS handler"); 1540 warn("logging seccomp filter failures"); 1541 } else if (j->flags.seccomp_filter_tsync) { 1542 /* 1543 * If setting thread sync, 1544 * reset the SIGSYS signal handler so that 1545 * the entire thread group is killed. 1546 */ 1547 if (signal(SIGSYS, SIG_DFL) == SIG_ERR) 1548 pdie("failed to reset SIGSYS disposition"); 1549 info("reset SIGSYS disposition"); 1550 } 1551 } 1552 1553 /* 1554 * Install the syscall filter. 1555 */ 1556 if (j->flags.seccomp_filter) { 1557 if (j->flags.seccomp_filter_tsync) { 1558 if (sys_seccomp(SECCOMP_SET_MODE_FILTER, 1559 SECCOMP_FILTER_FLAG_TSYNC, 1560 j->filter_prog)) { 1561 pdie("seccomp(tsync) failed"); 1562 } 1563 } else { 1564 if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, 1565 j->filter_prog)) { 1566 pdie("prctl(seccomp_filter) failed"); 1567 } 1568 } 1569 } 1570 } 1571 1572 static void config_net_loopback(void) 1573 { 1574 static const char ifname[] = "lo"; 1575 int sock; 1576 struct ifreq ifr; 1577 1578 /* Make sure people don't try to add really long names. */ 1579 _Static_assert(sizeof(ifname) <= IFNAMSIZ, "interface name too long"); 1580 1581 sock = socket(AF_LOCAL, SOCK_DGRAM|SOCK_CLOEXEC, 0); 1582 if (sock < 0) 1583 pdie("socket(AF_LOCAL) failed"); 1584 1585 /* 1586 * Do the equiv of `ip link set up lo`. The kernel will assign 1587 * IPv4 (127.0.0.1) & IPv6 (::1) addresses automatically! 1588 */ 1589 strcpy(ifr.ifr_name, ifname); 1590 if (ioctl(sock, SIOCGIFFLAGS, &ifr) < 0) 1591 pdie("ioctl(SIOCGIFFLAGS) failed"); 1592 1593 /* The kernel preserves ifr.ifr_name for use. */ 1594 ifr.ifr_flags |= IFF_UP | IFF_RUNNING; 1595 if (ioctl(sock, SIOCSIFFLAGS, &ifr) < 0) 1596 pdie("ioctl(SIOCSIFFLAGS) failed"); 1597 1598 close(sock); 1599 } 1600 1601 void API minijail_enter(const struct minijail *j) 1602 { 1603 /* 1604 * If we're dropping caps, get the last valid cap from /proc now, 1605 * since /proc can be unmounted before drop_caps() is called. 1606 */ 1607 unsigned int last_valid_cap = 0; 1608 if (j->flags.capbset_drop || j->flags.use_caps) 1609 last_valid_cap = get_last_valid_cap(); 1610 1611 if (j->flags.pids) 1612 die("tried to enter a pid-namespaced jail;" 1613 " try minijail_run()?"); 1614 1615 if (j->flags.inherit_suppl_gids && !j->user) 1616 die("cannot inherit supplementary groups without setting a " 1617 "username"); 1618 1619 /* 1620 * We can't recover from failures if we've dropped privileges partially, 1621 * so we don't even try. If any of our operations fail, we abort() the 1622 * entire process. 1623 */ 1624 if (j->flags.enter_vfs && setns(j->mountns_fd, CLONE_NEWNS)) 1625 pdie("setns(CLONE_NEWNS) failed"); 1626 1627 if (j->flags.vfs) { 1628 if (unshare(CLONE_NEWNS)) 1629 pdie("unshare(CLONE_NEWNS) failed"); 1630 /* 1631 * Unless asked not to, remount all filesystems as private. 1632 * If they are shared, new bind mounts will creep out of our 1633 * namespace. 1634 * https://www.kernel.org/doc/Documentation/filesystems/sharedsubtree.txt 1635 */ 1636 if (!j->flags.skip_remount_private) { 1637 if (mount(NULL, "/", NULL, MS_REC | MS_PRIVATE, NULL)) 1638 pdie("mount(NULL, /, NULL, MS_REC | MS_PRIVATE," 1639 " NULL) failed"); 1640 } 1641 } 1642 1643 if (j->flags.ipc && unshare(CLONE_NEWIPC)) { 1644 pdie("unshare(CLONE_NEWIPC) failed"); 1645 } 1646 1647 if (j->flags.enter_net) { 1648 if (setns(j->netns_fd, CLONE_NEWNET)) 1649 pdie("setns(CLONE_NEWNET) failed"); 1650 } else if (j->flags.net) { 1651 if (unshare(CLONE_NEWNET)) 1652 pdie("unshare(CLONE_NEWNET) failed"); 1653 config_net_loopback(); 1654 } 1655 1656 if (j->flags.ns_cgroups && unshare(CLONE_NEWCGROUP)) 1657 pdie("unshare(CLONE_NEWCGROUP) failed"); 1658 1659 if (j->flags.new_session_keyring) { 1660 if (syscall(SYS_keyctl, KEYCTL_JOIN_SESSION_KEYRING, NULL) < 0) 1661 pdie("keyctl(KEYCTL_JOIN_SESSION_KEYRING) failed"); 1662 } 1663 1664 if (j->flags.chroot && enter_chroot(j)) 1665 pdie("chroot"); 1666 1667 if (j->flags.pivot_root && enter_pivot_root(j)) 1668 pdie("pivot_root"); 1669 1670 if (j->flags.mount_tmp && mount_tmp(j)) 1671 pdie("mount_tmp"); 1672 1673 if (j->flags.remount_proc_ro && remount_proc_readonly(j)) 1674 pdie("remount"); 1675 1676 /* 1677 * If we're only dropping capabilities from the bounding set, but not 1678 * from the thread's (permitted|inheritable|effective) sets, do it now. 1679 */ 1680 if (j->flags.capbset_drop) { 1681 drop_capbset(j->cap_bset, last_valid_cap); 1682 } 1683 1684 if (j->flags.use_caps) { 1685 /* 1686 * POSIX capabilities are a bit tricky. If we drop our 1687 * capability to change uids, our attempt to use setuid() 1688 * below will fail. Hang on to root caps across setuid(), then 1689 * lock securebits. 1690 */ 1691 if (prctl(PR_SET_KEEPCAPS, 1)) 1692 pdie("prctl(PR_SET_KEEPCAPS) failed"); 1693 1694 /* 1695 * Kernels 4.3+ define a new securebit 1696 * (SECURE_NO_CAP_AMBIENT_RAISE), so using the SECURE_ALL_BITS 1697 * and SECURE_ALL_LOCKS masks from newer kernel headers will 1698 * return EPERM on older kernels. Detect this, and retry with 1699 * the right mask for older (2.6.26-4.2) kernels. 1700 */ 1701 int securebits_ret = prctl(PR_SET_SECUREBITS, 1702 SECURE_ALL_BITS | SECURE_ALL_LOCKS); 1703 if (securebits_ret < 0) { 1704 if (errno == EPERM) { 1705 /* Possibly running on kernel < 4.3. */ 1706 securebits_ret = prctl( 1707 PR_SET_SECUREBITS, 1708 OLD_SECURE_ALL_BITS | OLD_SECURE_ALL_LOCKS); 1709 } 1710 } 1711 if (securebits_ret < 0) 1712 pdie("prctl(PR_SET_SECUREBITS) failed"); 1713 } 1714 1715 if (j->flags.no_new_privs) { 1716 /* 1717 * If we're setting no_new_privs, we can drop privileges 1718 * before setting seccomp filter. This way filter policies 1719 * don't need to allow privilege-dropping syscalls. 1720 */ 1721 drop_ugid(j); 1722 drop_caps(j, last_valid_cap); 1723 set_seccomp_filter(j); 1724 } else { 1725 /* 1726 * If we're not setting no_new_privs, 1727 * we need to set seccomp filter *before* dropping privileges. 1728 * WARNING: this means that filter policies *must* allow 1729 * setgroups()/setresgid()/setresuid() for dropping root and 1730 * capget()/capset()/prctl() for dropping caps. 1731 */ 1732 set_seccomp_filter(j); 1733 drop_ugid(j); 1734 drop_caps(j, last_valid_cap); 1735 } 1736 1737 /* 1738 * Select the specified alternate syscall table. The table must not 1739 * block prctl(2) if we're using seccomp as well. 1740 */ 1741 if (j->flags.alt_syscall) { 1742 if (prctl(PR_ALT_SYSCALL, 1, j->alt_syscall_table)) 1743 pdie("prctl(PR_ALT_SYSCALL) failed"); 1744 } 1745 1746 /* 1747 * seccomp has to come last since it cuts off all the other 1748 * privilege-dropping syscalls :) 1749 */ 1750 if (j->flags.seccomp && prctl(PR_SET_SECCOMP, 1)) { 1751 if ((errno == EINVAL) && seccomp_can_softfail()) { 1752 warn("seccomp not supported"); 1753 return; 1754 } 1755 pdie("prctl(PR_SET_SECCOMP) failed"); 1756 } 1757 } 1758 1759 /* TODO(wad): will visibility affect this variable? */ 1760 static int init_exitstatus = 0; 1761 1762 void init_term(int __attribute__ ((unused)) sig) 1763 { 1764 _exit(init_exitstatus); 1765 } 1766 1767 void init(pid_t rootpid) 1768 { 1769 pid_t pid; 1770 int status; 1771 /* So that we exit with the right status. */ 1772 signal(SIGTERM, init_term); 1773 /* TODO(wad): self jail with seccomp filters here. */ 1774 while ((pid = wait(&status)) > 0) { 1775 /* 1776 * This loop will only end when either there are no processes 1777 * left inside our pid namespace or we get a signal. 1778 */ 1779 if (pid == rootpid) 1780 init_exitstatus = status; 1781 } 1782 if (!WIFEXITED(init_exitstatus)) 1783 _exit(MINIJAIL_ERR_INIT); 1784 _exit(WEXITSTATUS(init_exitstatus)); 1785 } 1786 1787 int API minijail_from_fd(int fd, struct minijail *j) 1788 { 1789 size_t sz = 0; 1790 size_t bytes = read(fd, &sz, sizeof(sz)); 1791 char *buf; 1792 int r; 1793 if (sizeof(sz) != bytes) 1794 return -EINVAL; 1795 if (sz > USHRT_MAX) /* arbitrary sanity check */ 1796 return -E2BIG; 1797 buf = malloc(sz); 1798 if (!buf) 1799 return -ENOMEM; 1800 bytes = read(fd, buf, sz); 1801 if (bytes != sz) { 1802 free(buf); 1803 return -EINVAL; 1804 } 1805 r = minijail_unmarshal(j, buf, sz); 1806 free(buf); 1807 return r; 1808 } 1809 1810 int API minijail_to_fd(struct minijail *j, int fd) 1811 { 1812 char *buf; 1813 size_t sz = minijail_size(j); 1814 ssize_t written; 1815 int r; 1816 1817 if (!sz) 1818 return -EINVAL; 1819 buf = malloc(sz); 1820 r = minijail_marshal(j, buf, sz); 1821 if (r) { 1822 free(buf); 1823 return r; 1824 } 1825 /* Sends [size][minijail]. */ 1826 written = write(fd, &sz, sizeof(sz)); 1827 if (written != sizeof(sz)) { 1828 free(buf); 1829 return -EFAULT; 1830 } 1831 written = write(fd, buf, sz); 1832 if (written < 0 || (size_t) written != sz) { 1833 free(buf); 1834 return -EFAULT; 1835 } 1836 free(buf); 1837 return 0; 1838 } 1839 1840 int setup_preload(void) 1841 { 1842 #if defined(__ANDROID__) 1843 /* Don't use LDPRELOAD on Brillo. */ 1844 return 0; 1845 #else 1846 char *oldenv = getenv(kLdPreloadEnvVar) ? : ""; 1847 char *newenv = malloc(strlen(oldenv) + 2 + strlen(PRELOADPATH)); 1848 if (!newenv) 1849 return -ENOMEM; 1850 1851 /* Only insert a separating space if we have something to separate... */ 1852 sprintf(newenv, "%s%s%s", oldenv, strlen(oldenv) ? " " : "", 1853 PRELOADPATH); 1854 1855 /* setenv() makes a copy of the string we give it. */ 1856 setenv(kLdPreloadEnvVar, newenv, 1); 1857 free(newenv); 1858 return 0; 1859 #endif 1860 } 1861 1862 int setup_pipe(int fds[2]) 1863 { 1864 int r = pipe(fds); 1865 char fd_buf[11]; 1866 if (r) 1867 return r; 1868 r = snprintf(fd_buf, sizeof(fd_buf), "%d", fds[0]); 1869 if (r <= 0) 1870 return -EINVAL; 1871 setenv(kFdEnvVar, fd_buf, 1); 1872 return 0; 1873 } 1874 1875 int setup_pipe_end(int fds[2], size_t index) 1876 { 1877 if (index > 1) 1878 return -1; 1879 1880 close(fds[1 - index]); 1881 return fds[index]; 1882 } 1883 1884 int setup_and_dupe_pipe_end(int fds[2], size_t index, int fd) 1885 { 1886 if (index > 1) 1887 return -1; 1888 1889 close(fds[1 - index]); 1890 /* dup2(2) the corresponding end of the pipe into |fd|. */ 1891 return dup2(fds[index], fd); 1892 } 1893 1894 int close_open_fds(int *inheritable_fds, size_t size) 1895 { 1896 const char *kFdPath = "/proc/self/fd"; 1897 1898 DIR *d = opendir(kFdPath); 1899 struct dirent *dir_entry; 1900 1901 if (d == NULL) 1902 return -1; 1903 int dir_fd = dirfd(d); 1904 while ((dir_entry = readdir(d)) != NULL) { 1905 size_t i; 1906 char *end; 1907 bool should_close = true; 1908 const int fd = strtol(dir_entry->d_name, &end, 10); 1909 1910 if ((*end) != '\0') { 1911 continue; 1912 } 1913 /* 1914 * We might have set up some pipes that we want to share with 1915 * the parent process, and should not be closed. 1916 */ 1917 for (i = 0; i < size; ++i) { 1918 if (fd == inheritable_fds[i]) { 1919 should_close = false; 1920 break; 1921 } 1922 } 1923 /* Also avoid closing the directory fd. */ 1924 if (should_close && fd != dir_fd) 1925 close(fd); 1926 } 1927 closedir(d); 1928 return 0; 1929 } 1930 1931 int minijail_run_internal(struct minijail *j, const char *filename, 1932 char *const argv[], pid_t *pchild_pid, 1933 int *pstdin_fd, int *pstdout_fd, int *pstderr_fd, 1934 int use_preload); 1935 1936 int API minijail_run(struct minijail *j, const char *filename, 1937 char *const argv[]) 1938 { 1939 return minijail_run_internal(j, filename, argv, NULL, NULL, NULL, NULL, 1940 true); 1941 } 1942 1943 int API minijail_run_pid(struct minijail *j, const char *filename, 1944 char *const argv[], pid_t *pchild_pid) 1945 { 1946 return minijail_run_internal(j, filename, argv, pchild_pid, 1947 NULL, NULL, NULL, true); 1948 } 1949 1950 int API minijail_run_pipe(struct minijail *j, const char *filename, 1951 char *const argv[], int *pstdin_fd) 1952 { 1953 return minijail_run_internal(j, filename, argv, NULL, pstdin_fd, 1954 NULL, NULL, true); 1955 } 1956 1957 int API minijail_run_pid_pipes(struct minijail *j, const char *filename, 1958 char *const argv[], pid_t *pchild_pid, 1959 int *pstdin_fd, int *pstdout_fd, int *pstderr_fd) 1960 { 1961 return minijail_run_internal(j, filename, argv, pchild_pid, 1962 pstdin_fd, pstdout_fd, pstderr_fd, true); 1963 } 1964 1965 int API minijail_run_no_preload(struct minijail *j, const char *filename, 1966 char *const argv[]) 1967 { 1968 return minijail_run_internal(j, filename, argv, NULL, NULL, NULL, NULL, 1969 false); 1970 } 1971 1972 int API minijail_run_pid_pipes_no_preload(struct minijail *j, 1973 const char *filename, 1974 char *const argv[], 1975 pid_t *pchild_pid, 1976 int *pstdin_fd, int *pstdout_fd, 1977 int *pstderr_fd) 1978 { 1979 return minijail_run_internal(j, filename, argv, pchild_pid, 1980 pstdin_fd, pstdout_fd, pstderr_fd, false); 1981 } 1982 1983 int minijail_run_internal(struct minijail *j, const char *filename, 1984 char *const argv[], pid_t *pchild_pid, 1985 int *pstdin_fd, int *pstdout_fd, int *pstderr_fd, 1986 int use_preload) 1987 { 1988 char *oldenv, *oldenv_copy = NULL; 1989 pid_t child_pid; 1990 int pipe_fds[2]; 1991 int stdin_fds[2]; 1992 int stdout_fds[2]; 1993 int stderr_fds[2]; 1994 int child_sync_pipe_fds[2]; 1995 int sync_child = 0; 1996 int ret; 1997 /* We need to remember this across the minijail_preexec() call. */ 1998 int pid_namespace = j->flags.pids; 1999 int do_init = j->flags.do_init; 2000 2001 if (use_preload) { 2002 oldenv = getenv(kLdPreloadEnvVar); 2003 if (oldenv) { 2004 oldenv_copy = strdup(oldenv); 2005 if (!oldenv_copy) 2006 return -ENOMEM; 2007 } 2008 2009 if (setup_preload()) 2010 return -EFAULT; 2011 } 2012 2013 if (!use_preload) { 2014 if (j->flags.use_caps && j->caps != 0) 2015 die("non-empty capabilities are not supported without " 2016 "LD_PRELOAD"); 2017 } 2018 2019 /* 2020 * Make the process group ID of this process equal to its PID. 2021 * In the non-interactive case (e.g. when the parent process is started 2022 * from init) this ensures the parent process and the jailed process 2023 * can be killed together. 2024 * When the parent process is started from the console this ensures 2025 * the call to setsid(2) in the jailed process succeeds. 2026 * 2027 * Don't fail on EPERM, since setpgid(0, 0) can only EPERM when 2028 * the process is already a process group leader. 2029 */ 2030 if (setpgid(0 /* use calling PID */, 0 /* make PGID = PID */)) { 2031 if (errno != EPERM) { 2032 pdie("setpgid(0, 0) failed"); 2033 } 2034 } 2035 2036 if (use_preload) { 2037 /* 2038 * Before we fork(2) and execve(2) the child process, we need 2039 * to open a pipe(2) to send the minijail configuration over. 2040 */ 2041 if (setup_pipe(pipe_fds)) 2042 return -EFAULT; 2043 } 2044 2045 /* 2046 * If we want to write to the child process' standard input, 2047 * create the pipe(2) now. 2048 */ 2049 if (pstdin_fd) { 2050 if (pipe(stdin_fds)) 2051 return -EFAULT; 2052 } 2053 2054 /* 2055 * If we want to read from the child process' standard output, 2056 * create the pipe(2) now. 2057 */ 2058 if (pstdout_fd) { 2059 if (pipe(stdout_fds)) 2060 return -EFAULT; 2061 } 2062 2063 /* 2064 * If we want to read from the child process' standard error, 2065 * create the pipe(2) now. 2066 */ 2067 if (pstderr_fd) { 2068 if (pipe(stderr_fds)) 2069 return -EFAULT; 2070 } 2071 2072 /* 2073 * If we want to set up a new uid/gid map in the user namespace, 2074 * or if we need to add the child process to cgroups, create the pipe(2) 2075 * to sync between parent and child. 2076 */ 2077 if (j->flags.userns || j->flags.cgroups) { 2078 sync_child = 1; 2079 if (pipe(child_sync_pipe_fds)) 2080 return -EFAULT; 2081 } 2082 2083 /* 2084 * Use sys_clone() if and only if we're creating a pid namespace. 2085 * 2086 * tl;dr: WARNING: do not mix pid namespaces and multithreading. 2087 * 2088 * In multithreaded programs, there are a bunch of locks inside libc, 2089 * some of which may be held by other threads at the time that we call 2090 * minijail_run_pid(). If we call fork(), glibc does its level best to 2091 * ensure that we hold all of these locks before it calls clone() 2092 * internally and drop them after clone() returns, but when we call 2093 * sys_clone(2) directly, all that gets bypassed and we end up with a 2094 * child address space where some of libc's important locks are held by 2095 * other threads (which did not get cloned, and hence will never release 2096 * those locks). This is okay so long as we call exec() immediately 2097 * after, but a bunch of seemingly-innocent libc functions like setenv() 2098 * take locks. 2099 * 2100 * Hence, only call sys_clone() if we need to, in order to get at pid 2101 * namespacing. If we follow this path, the child's address space might 2102 * have broken locks; you may only call functions that do not acquire 2103 * any locks. 2104 * 2105 * Unfortunately, fork() acquires every lock it can get its hands on, as 2106 * previously detailed, so this function is highly likely to deadlock 2107 * later on (see "deadlock here") if we're multithreaded. 2108 * 2109 * We might hack around this by having the clone()d child (init of the 2110 * pid namespace) return directly, rather than leaving the clone()d 2111 * process hanging around to be init for the new namespace (and having 2112 * its fork()ed child return in turn), but that process would be 2113 * crippled with its libc locks potentially broken. We might try 2114 * fork()ing in the parent before we clone() to ensure that we own all 2115 * the locks, but then we have to have the forked child hanging around 2116 * consuming resources (and possibly having file descriptors / shared 2117 * memory regions / etc attached). We'd need to keep the child around to 2118 * avoid having its children get reparented to init. 2119 * 2120 * TODO(ellyjones): figure out if the "forked child hanging around" 2121 * problem is fixable or not. It would be nice if we worked in this 2122 * case. 2123 */ 2124 if (pid_namespace) { 2125 int clone_flags = CLONE_NEWPID | SIGCHLD; 2126 if (j->flags.userns) 2127 clone_flags |= CLONE_NEWUSER; 2128 child_pid = syscall(SYS_clone, clone_flags, NULL); 2129 } else { 2130 child_pid = fork(); 2131 } 2132 2133 if (child_pid < 0) { 2134 if (use_preload) { 2135 free(oldenv_copy); 2136 } 2137 die("failed to fork child"); 2138 } 2139 2140 if (child_pid) { 2141 if (use_preload) { 2142 /* Restore parent's LD_PRELOAD. */ 2143 if (oldenv_copy) { 2144 setenv(kLdPreloadEnvVar, oldenv_copy, 1); 2145 free(oldenv_copy); 2146 } else { 2147 unsetenv(kLdPreloadEnvVar); 2148 } 2149 unsetenv(kFdEnvVar); 2150 } 2151 2152 j->initpid = child_pid; 2153 2154 if (j->flags.pid_file) 2155 write_pid_file_or_die(j); 2156 2157 if (j->flags.cgroups) 2158 add_to_cgroups_or_die(j); 2159 2160 if (j->flags.userns) 2161 write_ugid_maps_or_die(j); 2162 2163 if (sync_child) 2164 parent_setup_complete(child_sync_pipe_fds); 2165 2166 if (use_preload) { 2167 /* Send marshalled minijail. */ 2168 close(pipe_fds[0]); /* read endpoint */ 2169 ret = minijail_to_fd(j, pipe_fds[1]); 2170 close(pipe_fds[1]); /* write endpoint */ 2171 if (ret) { 2172 kill(j->initpid, SIGKILL); 2173 die("failed to send marshalled minijail"); 2174 } 2175 } 2176 2177 if (pchild_pid) 2178 *pchild_pid = child_pid; 2179 2180 /* 2181 * If we want to write to the child process' standard input, 2182 * set up the write end of the pipe. 2183 */ 2184 if (pstdin_fd) 2185 *pstdin_fd = setup_pipe_end(stdin_fds, 2186 1 /* write end */); 2187 2188 /* 2189 * If we want to read from the child process' standard output, 2190 * set up the read end of the pipe. 2191 */ 2192 if (pstdout_fd) 2193 *pstdout_fd = setup_pipe_end(stdout_fds, 2194 0 /* read end */); 2195 2196 /* 2197 * If we want to read from the child process' standard error, 2198 * set up the read end of the pipe. 2199 */ 2200 if (pstderr_fd) 2201 *pstderr_fd = setup_pipe_end(stderr_fds, 2202 0 /* read end */); 2203 2204 return 0; 2205 } 2206 /* Child process. */ 2207 free(oldenv_copy); 2208 2209 if (j->flags.reset_signal_mask) { 2210 sigset_t signal_mask; 2211 if (sigemptyset(&signal_mask) != 0) 2212 pdie("sigemptyset failed"); 2213 if (sigprocmask(SIG_SETMASK, &signal_mask, NULL) != 0) 2214 pdie("sigprocmask failed"); 2215 } 2216 2217 if (j->flags.close_open_fds) { 2218 const size_t kMaxInheritableFdsSize = 10; 2219 int inheritable_fds[kMaxInheritableFdsSize]; 2220 size_t size = 0; 2221 if (use_preload) { 2222 inheritable_fds[size++] = pipe_fds[0]; 2223 inheritable_fds[size++] = pipe_fds[1]; 2224 } 2225 if (sync_child) { 2226 inheritable_fds[size++] = child_sync_pipe_fds[0]; 2227 inheritable_fds[size++] = child_sync_pipe_fds[1]; 2228 } 2229 if (pstdin_fd) { 2230 inheritable_fds[size++] = stdin_fds[0]; 2231 inheritable_fds[size++] = stdin_fds[1]; 2232 } 2233 if (pstdout_fd) { 2234 inheritable_fds[size++] = stdout_fds[0]; 2235 inheritable_fds[size++] = stdout_fds[1]; 2236 } 2237 if (pstderr_fd) { 2238 inheritable_fds[size++] = stderr_fds[0]; 2239 inheritable_fds[size++] = stderr_fds[1]; 2240 } 2241 2242 if (close_open_fds(inheritable_fds, size) < 0) 2243 die("failed to close open file descriptors"); 2244 } 2245 2246 if (sync_child) 2247 wait_for_parent_setup(child_sync_pipe_fds); 2248 2249 if (j->flags.userns) 2250 enter_user_namespace(j); 2251 2252 /* 2253 * If we want to write to the jailed process' standard input, 2254 * set up the read end of the pipe. 2255 */ 2256 if (pstdin_fd) { 2257 if (setup_and_dupe_pipe_end(stdin_fds, 0 /* read end */, 2258 STDIN_FILENO) < 0) 2259 die("failed to set up stdin pipe"); 2260 } 2261 2262 /* 2263 * If we want to read from the jailed process' standard output, 2264 * set up the write end of the pipe. 2265 */ 2266 if (pstdout_fd) { 2267 if (setup_and_dupe_pipe_end(stdout_fds, 1 /* write end */, 2268 STDOUT_FILENO) < 0) 2269 die("failed to set up stdout pipe"); 2270 } 2271 2272 /* 2273 * If we want to read from the jailed process' standard error, 2274 * set up the write end of the pipe. 2275 */ 2276 if (pstderr_fd) { 2277 if (setup_and_dupe_pipe_end(stderr_fds, 1 /* write end */, 2278 STDERR_FILENO) < 0) 2279 die("failed to set up stderr pipe"); 2280 } 2281 2282 /* 2283 * If any of stdin, stdout, or stderr are TTYs, create a new session. 2284 * This prevents the jailed process from using the TIOCSTI ioctl 2285 * to push characters into the parent process terminal's input buffer, 2286 * therefore escaping the jail. 2287 */ 2288 if (isatty(STDIN_FILENO) || isatty(STDOUT_FILENO) || 2289 isatty(STDERR_FILENO)) { 2290 if (setsid() < 0) { 2291 pdie("setsid() failed"); 2292 } 2293 } 2294 2295 /* If running an init program, let it decide when/how to mount /proc. */ 2296 if (pid_namespace && !do_init) 2297 j->flags.remount_proc_ro = 0; 2298 2299 if (use_preload) { 2300 /* Strip out flags that cannot be inherited across execve(2). */ 2301 minijail_preexec(j); 2302 } else { 2303 /* 2304 * If not using LD_PRELOAD, do all jailing before execve(2). 2305 * Note that PID namespaces can only be entered on fork(2), 2306 * so that flag is still cleared. 2307 */ 2308 j->flags.pids = 0; 2309 } 2310 /* Jail this process, then execve(2) the target. */ 2311 minijail_enter(j); 2312 2313 if (pid_namespace && do_init) { 2314 /* 2315 * pid namespace: this process will become init inside the new 2316 * namespace. We don't want all programs we might exec to have 2317 * to know how to be init. Normally (do_init == 1) we fork off 2318 * a child to actually run the program. If |do_init == 0|, we 2319 * let the program keep pid 1 and be init. 2320 * 2321 * If we're multithreaded, we'll probably deadlock here. See 2322 * WARNING above. 2323 */ 2324 child_pid = fork(); 2325 if (child_pid < 0) { 2326 _exit(child_pid); 2327 } else if (child_pid > 0) { 2328 /* 2329 * Best effort. Don't bother checking the return value. 2330 */ 2331 prctl(PR_SET_NAME, "minijail-init"); 2332 init(child_pid); /* Never returns. */ 2333 } 2334 } 2335 2336 /* 2337 * If we aren't pid-namespaced, or the jailed program asked to be init: 2338 * calling process 2339 * -> execve()-ing process 2340 * If we are: 2341 * calling process 2342 * -> init()-ing process 2343 * -> execve()-ing process 2344 */ 2345 ret = execve(filename, argv, environ); 2346 if (ret == -1) { 2347 pwarn("execve(%s) failed", filename); 2348 } 2349 _exit(ret); 2350 } 2351 2352 int API minijail_kill(struct minijail *j) 2353 { 2354 int st; 2355 if (kill(j->initpid, SIGTERM)) 2356 return -errno; 2357 if (waitpid(j->initpid, &st, 0) < 0) 2358 return -errno; 2359 return st; 2360 } 2361 2362 int API minijail_wait(struct minijail *j) 2363 { 2364 int st; 2365 if (waitpid(j->initpid, &st, 0) < 0) 2366 return -errno; 2367 2368 if (!WIFEXITED(st)) { 2369 int error_status = st; 2370 if (WIFSIGNALED(st)) { 2371 int signum = WTERMSIG(st); 2372 warn("child process %d received signal %d", 2373 j->initpid, signum); 2374 /* 2375 * We return MINIJAIL_ERR_JAIL if the process received 2376 * SIGSYS, which happens when a syscall is blocked by 2377 * seccomp filters. 2378 * If not, we do what bash(1) does: 2379 * $? = 128 + signum 2380 */ 2381 if (signum == SIGSYS) { 2382 error_status = MINIJAIL_ERR_JAIL; 2383 } else { 2384 error_status = 128 + signum; 2385 } 2386 } 2387 return error_status; 2388 } 2389 2390 int exit_status = WEXITSTATUS(st); 2391 if (exit_status != 0) 2392 info("child process %d exited with status %d", 2393 j->initpid, exit_status); 2394 2395 return exit_status; 2396 } 2397 2398 void API minijail_destroy(struct minijail *j) 2399 { 2400 size_t i; 2401 2402 if (j->flags.seccomp_filter && j->filter_prog) { 2403 free(j->filter_prog->filter); 2404 free(j->filter_prog); 2405 } 2406 while (j->mounts_head) { 2407 struct mountpoint *m = j->mounts_head; 2408 j->mounts_head = j->mounts_head->next; 2409 free(m->data); 2410 free(m->type); 2411 free(m->dest); 2412 free(m->src); 2413 free(m); 2414 } 2415 j->mounts_tail = NULL; 2416 if (j->user) 2417 free(j->user); 2418 if (j->suppl_gid_list) 2419 free(j->suppl_gid_list); 2420 if (j->chrootdir) 2421 free(j->chrootdir); 2422 if (j->pid_file_path) 2423 free(j->pid_file_path); 2424 if (j->uidmap) 2425 free(j->uidmap); 2426 if (j->gidmap) 2427 free(j->gidmap); 2428 if (j->alt_syscall_table) 2429 free(j->alt_syscall_table); 2430 for (i = 0; i < j->cgroup_count; ++i) 2431 free(j->cgroups[i]); 2432 free(j); 2433 } 2434