libminijail.c revision 3c84df1c18b410cb33da3c9df010b59f960785a9
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 _GNU_SOURCE 8 9#include <asm/unistd.h> 10#include <ctype.h> 11#include <errno.h> 12#include <fcntl.h> 13#include <grp.h> 14#include <inttypes.h> 15#include <limits.h> 16#include <linux/capability.h> 17#include <pwd.h> 18#include <sched.h> 19#include <signal.h> 20#include <stdarg.h> 21#include <stddef.h> 22#include <stdio.h> 23#include <stdlib.h> 24#include <string.h> 25#include <syscall.h> 26#include <sys/capability.h> 27#include <sys/mount.h> 28#include <sys/param.h> 29#include <sys/prctl.h> 30#include <sys/stat.h> 31#include <sys/types.h> 32#include <sys/user.h> 33#include <sys/wait.h> 34#include <unistd.h> 35 36#include "libminijail.h" 37#include "libminijail-private.h" 38 39#include "signal.h" 40#include "syscall_filter.h" 41#include "util.h" 42 43#ifdef HAVE_SECUREBITS_H 44#include <linux/securebits.h> 45#else 46#define SECURE_ALL_BITS 0x15 47#define SECURE_ALL_LOCKS (SECURE_ALL_BITS << 1) 48#endif 49 50/* Until these are reliably available in linux/prctl.h */ 51#ifndef PR_SET_SECCOMP 52# define PR_SET_SECCOMP 22 53#endif 54 55/* For seccomp_filter using BPF. */ 56#ifndef PR_SET_NO_NEW_PRIVS 57# define PR_SET_NO_NEW_PRIVS 38 58#endif 59#ifndef SECCOMP_MODE_FILTER 60# define SECCOMP_MODE_FILTER 2 /* uses user-supplied filter. */ 61#endif 62 63#ifdef USE_SECCOMP_SOFTFAIL 64# define SECCOMP_SOFTFAIL 1 65#else 66# define SECCOMP_SOFTFAIL 0 67#endif 68 69struct binding { 70 char *src; 71 char *dest; 72 int writeable; 73 struct binding *next; 74}; 75 76struct minijail { 77 /* 78 * WARNING: if you add a flag here you need to make sure it's 79 * accounted for in minijail_pre{enter|exec}() below. 80 */ 81 struct { 82 int uid:1; 83 int gid:1; 84 int caps:1; 85 int vfs:1; 86 int enter_vfs:1; 87 int pids:1; 88 int net:1; 89 int seccomp:1; 90 int readonly:1; 91 int usergroups:1; 92 int ptrace:1; 93 int no_new_privs:1; 94 int seccomp_filter:1; 95 int log_seccomp_filter:1; 96 int chroot:1; 97 int mount_tmp:1; 98 } flags; 99 uid_t uid; 100 gid_t gid; 101 gid_t usergid; 102 char *user; 103 uint64_t caps; 104 pid_t initpid; 105 int mountns_fd; 106 int filter_len; 107 int binding_count; 108 char *chrootdir; 109 struct sock_fprog *filter_prog; 110 struct binding *bindings_head; 111 struct binding *bindings_tail; 112}; 113 114/* 115 * Strip out flags meant for the parent. 116 * We keep things that are not inherited across execve(2) (e.g. capabilities), 117 * or are easier to set after execve(2) (e.g. seccomp filters). 118 */ 119void minijail_preenter(struct minijail *j) 120{ 121 j->flags.vfs = 0; 122 j->flags.enter_vfs = 0; 123 j->flags.readonly = 0; 124 j->flags.pids = 0; 125} 126 127/* 128 * Strip out flags meant for the child. 129 * We keep things that are inherited across execve(2). 130 */ 131void minijail_preexec(struct minijail *j) 132{ 133 int vfs = j->flags.vfs; 134 int enter_vfs = j->flags.enter_vfs; 135 int readonly = j->flags.readonly; 136 if (j->user) 137 free(j->user); 138 j->user = NULL; 139 memset(&j->flags, 0, sizeof(j->flags)); 140 /* Now restore anything we meant to keep. */ 141 j->flags.vfs = vfs; 142 j->flags.enter_vfs = enter_vfs; 143 j->flags.readonly = readonly; 144 /* Note, |pids| will already have been used before this call. */ 145} 146 147/* Minijail API. */ 148 149struct minijail API *minijail_new(void) 150{ 151 return calloc(1, sizeof(struct minijail)); 152} 153 154void API minijail_change_uid(struct minijail *j, uid_t uid) 155{ 156 if (uid == 0) 157 die("useless change to uid 0"); 158 j->uid = uid; 159 j->flags.uid = 1; 160} 161 162void API minijail_change_gid(struct minijail *j, gid_t gid) 163{ 164 if (gid == 0) 165 die("useless change to gid 0"); 166 j->gid = gid; 167 j->flags.gid = 1; 168} 169 170int API minijail_change_user(struct minijail *j, const char *user) 171{ 172 char *buf = NULL; 173 struct passwd pw; 174 struct passwd *ppw = NULL; 175 ssize_t sz = sysconf(_SC_GETPW_R_SIZE_MAX); 176 if (sz == -1) 177 sz = 65536; /* your guess is as good as mine... */ 178 179 /* 180 * sysconf(_SC_GETPW_R_SIZE_MAX), under glibc, is documented to return 181 * the maximum needed size of the buffer, so we don't have to search. 182 */ 183 buf = malloc(sz); 184 if (!buf) 185 return -ENOMEM; 186 getpwnam_r(user, &pw, buf, sz, &ppw); 187 /* 188 * We're safe to free the buffer here. The strings inside pw point 189 * inside buf, but we don't use any of them; this leaves the pointers 190 * dangling but it's safe. ppw points at pw if getpwnam_r succeeded. 191 */ 192 free(buf); 193 /* getpwnam_r(3) does *not* set errno when |ppw| is NULL. */ 194 if (!ppw) 195 return -1; 196 minijail_change_uid(j, ppw->pw_uid); 197 j->user = strdup(user); 198 if (!j->user) 199 return -ENOMEM; 200 j->usergid = ppw->pw_gid; 201 return 0; 202} 203 204int API minijail_change_group(struct minijail *j, const char *group) 205{ 206 char *buf = NULL; 207 struct group gr; 208 struct group *pgr = NULL; 209 ssize_t sz = sysconf(_SC_GETGR_R_SIZE_MAX); 210 if (sz == -1) 211 sz = 65536; /* and mine is as good as yours, really */ 212 213 /* 214 * sysconf(_SC_GETGR_R_SIZE_MAX), under glibc, is documented to return 215 * the maximum needed size of the buffer, so we don't have to search. 216 */ 217 buf = malloc(sz); 218 if (!buf) 219 return -ENOMEM; 220 getgrnam_r(group, &gr, buf, sz, &pgr); 221 /* 222 * We're safe to free the buffer here. The strings inside gr point 223 * inside buf, but we don't use any of them; this leaves the pointers 224 * dangling but it's safe. pgr points at gr if getgrnam_r succeeded. 225 */ 226 free(buf); 227 /* getgrnam_r(3) does *not* set errno when |pgr| is NULL. */ 228 if (!pgr) 229 return -1; 230 minijail_change_gid(j, pgr->gr_gid); 231 return 0; 232} 233 234void API minijail_use_seccomp(struct minijail *j) 235{ 236 j->flags.seccomp = 1; 237} 238 239void API minijail_no_new_privs(struct minijail *j) 240{ 241 j->flags.no_new_privs = 1; 242} 243 244void API minijail_use_seccomp_filter(struct minijail *j) 245{ 246 j->flags.seccomp_filter = 1; 247} 248 249void API minijail_log_seccomp_filter_failures(struct minijail *j) 250{ 251 j->flags.log_seccomp_filter = 1; 252} 253 254void API minijail_use_caps(struct minijail *j, uint64_t capmask) 255{ 256 j->caps = capmask; 257 j->flags.caps = 1; 258} 259 260void API minijail_namespace_vfs(struct minijail *j) 261{ 262 j->flags.vfs = 1; 263} 264 265void API minijail_namespace_enter_vfs(struct minijail *j, const char *ns_path) 266{ 267 int ns_fd = open(ns_path, O_RDONLY); 268 if (ns_fd < 0) { 269 pdie("failed to open namespace '%s'", ns_path); 270 } 271 j->mountns_fd = ns_fd; 272 j->flags.enter_vfs = 1; 273} 274 275void API minijail_namespace_pids(struct minijail *j) 276{ 277 j->flags.vfs = 1; 278 j->flags.readonly = 1; 279 j->flags.pids = 1; 280} 281 282void API minijail_namespace_net(struct minijail *j) 283{ 284 j->flags.net = 1; 285} 286 287void API minijail_remount_readonly(struct minijail *j) 288{ 289 j->flags.vfs = 1; 290 j->flags.readonly = 1; 291} 292 293void API minijail_inherit_usergroups(struct minijail *j) 294{ 295 j->flags.usergroups = 1; 296} 297 298void API minijail_disable_ptrace(struct minijail *j) 299{ 300 j->flags.ptrace = 1; 301} 302 303int API minijail_enter_chroot(struct minijail *j, const char *dir) 304{ 305 if (j->chrootdir) 306 return -EINVAL; 307 j->chrootdir = strdup(dir); 308 if (!j->chrootdir) 309 return -ENOMEM; 310 j->flags.chroot = 1; 311 return 0; 312} 313 314void API minijail_mount_tmp(struct minijail *j) 315{ 316 j->flags.mount_tmp = 1; 317} 318 319int API minijail_bind(struct minijail *j, const char *src, const char *dest, 320 int writeable) 321{ 322 struct binding *b; 323 324 if (*dest != '/') 325 return -EINVAL; 326 b = calloc(1, sizeof(*b)); 327 if (!b) 328 return -ENOMEM; 329 b->dest = strdup(dest); 330 if (!b->dest) 331 goto error; 332 b->src = strdup(src); 333 if (!b->src) 334 goto error; 335 b->writeable = writeable; 336 337 info("bind %s -> %s", src, dest); 338 339 /* 340 * Force vfs namespacing so the bind mounts don't leak out into the 341 * containing vfs namespace. 342 */ 343 minijail_namespace_vfs(j); 344 345 if (j->bindings_tail) 346 j->bindings_tail->next = b; 347 else 348 j->bindings_head = b; 349 j->bindings_tail = b; 350 j->binding_count++; 351 352 return 0; 353 354error: 355 free(b->src); 356 free(b->dest); 357 free(b); 358 return -ENOMEM; 359} 360 361void API minijail_parse_seccomp_filters(struct minijail *j, const char *path) 362{ 363 if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, NULL)) { 364 if ((errno == ENOSYS) && SECCOMP_SOFTFAIL) { 365 warn("not loading seccomp filter, seccomp not supported"); 366 return; 367 } 368 } 369 FILE *file = fopen(path, "r"); 370 if (!file) { 371 pdie("failed to open seccomp filter file '%s'", path); 372 } 373 374 struct sock_fprog *fprog = malloc(sizeof(struct sock_fprog)); 375 if (compile_filter(file, fprog, j->flags.log_seccomp_filter)) { 376 die("failed to compile seccomp filter BPF program in '%s'", 377 path); 378 } 379 380 j->filter_len = fprog->len; 381 j->filter_prog = fprog; 382 383 fclose(file); 384} 385 386struct marshal_state { 387 size_t available; 388 size_t total; 389 char *buf; 390}; 391 392void marshal_state_init(struct marshal_state *state, 393 char *buf, size_t available) 394{ 395 state->available = available; 396 state->buf = buf; 397 state->total = 0; 398} 399 400void marshal_append(struct marshal_state *state, 401 char *src, size_t length) 402{ 403 size_t copy_len = MIN(state->available, length); 404 405 /* Up to |available| will be written. */ 406 if (copy_len) { 407 memcpy(state->buf, src, copy_len); 408 state->buf += copy_len; 409 state->available -= copy_len; 410 } 411 /* |total| will contain the expected length. */ 412 state->total += length; 413} 414 415void minijail_marshal_helper(struct marshal_state *state, 416 const struct minijail *j) 417{ 418 struct binding *b = NULL; 419 marshal_append(state, (char *)j, sizeof(*j)); 420 if (j->user) 421 marshal_append(state, j->user, strlen(j->user) + 1); 422 if (j->chrootdir) 423 marshal_append(state, j->chrootdir, strlen(j->chrootdir) + 1); 424 if (j->flags.seccomp_filter && j->filter_prog) { 425 struct sock_fprog *fp = j->filter_prog; 426 marshal_append(state, (char *)fp->filter, 427 fp->len * sizeof(struct sock_filter)); 428 } 429 for (b = j->bindings_head; b; b = b->next) { 430 marshal_append(state, b->src, strlen(b->src) + 1); 431 marshal_append(state, b->dest, strlen(b->dest) + 1); 432 marshal_append(state, (char *)&b->writeable, 433 sizeof(b->writeable)); 434 } 435} 436 437size_t API minijail_size(const struct minijail *j) 438{ 439 struct marshal_state state; 440 marshal_state_init(&state, NULL, 0); 441 minijail_marshal_helper(&state, j); 442 return state.total; 443} 444 445int minijail_marshal(const struct minijail *j, char *buf, size_t available) 446{ 447 struct marshal_state state; 448 marshal_state_init(&state, buf, available); 449 minijail_marshal_helper(&state, j); 450 return (state.total > available); 451} 452 453/* consumebytes: consumes @length bytes from a buffer @buf of length @buflength 454 * @length Number of bytes to consume 455 * @buf Buffer to consume from 456 * @buflength Size of @buf 457 * 458 * Returns a pointer to the base of the bytes, or NULL for errors. 459 */ 460void *consumebytes(size_t length, char **buf, size_t *buflength) 461{ 462 char *p = *buf; 463 if (length > *buflength) 464 return NULL; 465 *buf += length; 466 *buflength -= length; 467 return p; 468} 469 470/* consumestr: consumes a C string from a buffer @buf of length @length 471 * @buf Buffer to consume 472 * @length Length of buffer 473 * 474 * Returns a pointer to the base of the string, or NULL for errors. 475 */ 476char *consumestr(char **buf, size_t *buflength) 477{ 478 size_t len = strnlen(*buf, *buflength); 479 if (len == *buflength) 480 /* There's no null-terminator */ 481 return NULL; 482 return consumebytes(len + 1, buf, buflength); 483} 484 485int minijail_unmarshal(struct minijail *j, char *serialized, size_t length) 486{ 487 int i; 488 int count; 489 int ret = -EINVAL; 490 491 if (length < sizeof(*j)) 492 goto out; 493 memcpy((void *)j, serialized, sizeof(*j)); 494 serialized += sizeof(*j); 495 length -= sizeof(*j); 496 497 /* Potentially stale pointers not used as signals. */ 498 j->bindings_head = NULL; 499 j->bindings_tail = NULL; 500 j->filter_prog = NULL; 501 502 if (j->user) { /* stale pointer */ 503 char *user = consumestr(&serialized, &length); 504 if (!user) 505 goto clear_pointers; 506 j->user = strdup(user); 507 if (!j->user) 508 goto clear_pointers; 509 } 510 511 if (j->chrootdir) { /* stale pointer */ 512 char *chrootdir = consumestr(&serialized, &length); 513 if (!chrootdir) 514 goto bad_chrootdir; 515 j->chrootdir = strdup(chrootdir); 516 if (!j->chrootdir) 517 goto bad_chrootdir; 518 } 519 520 if (j->flags.seccomp_filter && j->filter_len > 0) { 521 size_t ninstrs = j->filter_len; 522 if (ninstrs > (SIZE_MAX / sizeof(struct sock_filter)) || 523 ninstrs > USHRT_MAX) 524 goto bad_filters; 525 526 size_t program_len = ninstrs * sizeof(struct sock_filter); 527 void *program = consumebytes(program_len, &serialized, &length); 528 if (!program) 529 goto bad_filters; 530 531 j->filter_prog = malloc(sizeof(struct sock_fprog)); 532 j->filter_prog->len = ninstrs; 533 j->filter_prog->filter = malloc(program_len); 534 memcpy(j->filter_prog->filter, program, program_len); 535 } 536 537 count = j->binding_count; 538 j->binding_count = 0; 539 for (i = 0; i < count; ++i) { 540 int *writeable; 541 const char *dest; 542 const char *src = consumestr(&serialized, &length); 543 if (!src) 544 goto bad_bindings; 545 dest = consumestr(&serialized, &length); 546 if (!dest) 547 goto bad_bindings; 548 writeable = consumebytes(sizeof(*writeable), &serialized, &length); 549 if (!writeable) 550 goto bad_bindings; 551 if (minijail_bind(j, src, dest, *writeable)) 552 goto bad_bindings; 553 } 554 555 return 0; 556 557bad_bindings: 558 if (j->flags.seccomp_filter && j->filter_len > 0) { 559 free(j->filter_prog->filter); 560 free(j->filter_prog); 561 } 562bad_filters: 563 if (j->chrootdir) 564 free(j->chrootdir); 565bad_chrootdir: 566 if (j->user) 567 free(j->user); 568clear_pointers: 569 j->user = NULL; 570 j->chrootdir = NULL; 571out: 572 return ret; 573} 574 575/* bind_one: Applies bindings from @b for @j, recursing as needed. 576 * @j Minijail these bindings are for 577 * @b Head of list of bindings 578 * 579 * Returns 0 for success. 580 */ 581int bind_one(const struct minijail *j, struct binding *b) 582{ 583 int ret = 0; 584 char *dest = NULL; 585 if (ret) 586 return ret; 587 /* dest has a leading "/" */ 588 if (asprintf(&dest, "%s%s", j->chrootdir, b->dest) < 0) 589 return -ENOMEM; 590 ret = mount(b->src, dest, NULL, MS_BIND, NULL); 591 if (ret) 592 pdie("bind: %s -> %s", b->src, dest); 593 if (!b->writeable) { 594 ret = mount(b->src, dest, NULL, 595 MS_BIND | MS_REMOUNT | MS_RDONLY, NULL); 596 if (ret) 597 pdie("bind ro: %s -> %s", b->src, dest); 598 } 599 free(dest); 600 if (b->next) 601 return bind_one(j, b->next); 602 return ret; 603} 604 605int enter_chroot(const struct minijail *j) 606{ 607 int ret; 608 if (j->bindings_head && (ret = bind_one(j, j->bindings_head))) 609 return ret; 610 611 if (chroot(j->chrootdir)) 612 return -errno; 613 614 if (chdir("/")) 615 return -errno; 616 617 return 0; 618} 619 620int mount_tmp(void) 621{ 622 return mount("none", "/tmp", "tmpfs", 0, "size=64M,mode=777"); 623} 624 625int remount_readonly(void) 626{ 627 const char *kProcPath = "/proc"; 628 const unsigned int kSafeFlags = MS_NODEV | MS_NOEXEC | MS_NOSUID; 629 /* 630 * Right now, we're holding a reference to our parent's old mount of 631 * /proc in our namespace, which means using MS_REMOUNT here would 632 * mutate our parent's mount as well, even though we're in a VFS 633 * namespace (!). Instead, remove their mount from our namespace 634 * and make our own. 635 */ 636 if (umount(kProcPath)) 637 return -errno; 638 if (mount("", kProcPath, "proc", kSafeFlags | MS_RDONLY, "")) 639 return -errno; 640 return 0; 641} 642 643void drop_ugid(const struct minijail *j) 644{ 645 if (j->flags.usergroups) { 646 if (initgroups(j->user, j->usergid)) 647 pdie("initgroups"); 648 } else { 649 /* Only attempt to clear supplemental groups if we are changing 650 * users. */ 651 if ((j->uid || j->gid) && setgroups(0, NULL)) 652 pdie("setgroups"); 653 } 654 655 if (j->flags.gid && setresgid(j->gid, j->gid, j->gid)) 656 pdie("setresgid"); 657 658 if (j->flags.uid && setresuid(j->uid, j->uid, j->uid)) 659 pdie("setresuid"); 660} 661 662/* 663 * We specifically do not use cap_valid() as that only tells us the last 664 * valid cap we were *compiled* against (i.e. what the version of kernel 665 * headers says). If we run on a different kernel version, then it's not 666 * uncommon for that to be less (if an older kernel) or more (if a newer 667 * kernel). So suck up the answer via /proc. 668 */ 669static int run_cap_valid(unsigned int cap) 670{ 671 static unsigned int last_cap; 672 673 if (!last_cap) { 674 const char cap_file[] = "/proc/sys/kernel/cap_last_cap"; 675 FILE *fp = fopen(cap_file, "re"); 676 if (fscanf(fp, "%u", &last_cap) != 1) 677 pdie("fscanf(%s)", cap_file); 678 fclose(fp); 679 } 680 681 return cap <= last_cap; 682} 683 684void drop_caps(const struct minijail *j) 685{ 686 cap_t caps = cap_get_proc(); 687 cap_value_t flag[1]; 688 const uint64_t one = 1; 689 unsigned int i; 690 if (!caps) 691 die("can't get process caps"); 692 if (cap_clear_flag(caps, CAP_INHERITABLE)) 693 die("can't clear inheritable caps"); 694 if (cap_clear_flag(caps, CAP_EFFECTIVE)) 695 die("can't clear effective caps"); 696 if (cap_clear_flag(caps, CAP_PERMITTED)) 697 die("can't clear permitted caps"); 698 for (i = 0; i < sizeof(j->caps) * 8 && run_cap_valid(i); ++i) { 699 /* Keep CAP_SETPCAP for dropping bounding set bits. */ 700 if (i != CAP_SETPCAP && !(j->caps & (one << i))) 701 continue; 702 flag[0] = i; 703 if (cap_set_flag(caps, CAP_EFFECTIVE, 1, flag, CAP_SET)) 704 die("can't add effective cap"); 705 if (cap_set_flag(caps, CAP_PERMITTED, 1, flag, CAP_SET)) 706 die("can't add permitted cap"); 707 if (cap_set_flag(caps, CAP_INHERITABLE, 1, flag, CAP_SET)) 708 die("can't add inheritable cap"); 709 } 710 if (cap_set_proc(caps)) 711 die("can't apply initial cleaned capset"); 712 713 /* 714 * Instead of dropping bounding set first, do it here in case 715 * the caller had a more permissive bounding set which could 716 * have been used above to raise a capability that wasn't already 717 * present. This requires CAP_SETPCAP, so we raised/kept it above. 718 */ 719 for (i = 0; i < sizeof(j->caps) * 8 && run_cap_valid(i); ++i) { 720 if (j->caps & (one << i)) 721 continue; 722 if (prctl(PR_CAPBSET_DROP, i)) 723 pdie("prctl(PR_CAPBSET_DROP)"); 724 } 725 726 /* If CAP_SETPCAP wasn't specifically requested, now we remove it. */ 727 if ((j->caps & (one << CAP_SETPCAP)) == 0) { 728 flag[0] = CAP_SETPCAP; 729 if (cap_set_flag(caps, CAP_EFFECTIVE, 1, flag, CAP_CLEAR)) 730 die("can't clear effective cap"); 731 if (cap_set_flag(caps, CAP_PERMITTED, 1, flag, CAP_CLEAR)) 732 die("can't clear permitted cap"); 733 if (cap_set_flag(caps, CAP_INHERITABLE, 1, flag, CAP_CLEAR)) 734 die("can't clear inheritable cap"); 735 } 736 737 if (cap_set_proc(caps)) 738 die("can't apply final cleaned capset"); 739 740 cap_free(caps); 741} 742 743void set_seccomp_filter(const struct minijail *j) 744{ 745 /* 746 * Set no_new_privs. See </kernel/seccomp.c> and </kernel/sys.c> 747 * in the kernel source tree for an explanation of the parameters. 748 */ 749 if (j->flags.no_new_privs) { 750 if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0)) 751 pdie("prctl(PR_SET_NO_NEW_PRIVS)"); 752 } 753 754 /* 755 * If we're logging seccomp filter failures, 756 * install the SIGSYS handler first. 757 */ 758 if (j->flags.seccomp_filter && j->flags.log_seccomp_filter) { 759 if (install_sigsys_handler()) 760 pdie("install SIGSYS handler"); 761 warn("logging seccomp filter failures"); 762 } 763 764 /* 765 * Install the syscall filter. 766 */ 767 if (j->flags.seccomp_filter) { 768 if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, j->filter_prog)) { 769 if ((errno == ENOSYS) && SECCOMP_SOFTFAIL) { 770 warn("seccomp not supported"); 771 return; 772 } 773 pdie("prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER)"); 774 } 775 } 776} 777 778void API minijail_enter(const struct minijail *j) 779{ 780 if (j->flags.pids) 781 die("tried to enter a pid-namespaced jail;" 782 " try minijail_run()?"); 783 784 if (j->flags.usergroups && !j->user) 785 die("usergroup inheritance without username"); 786 787 /* 788 * We can't recover from failures if we've dropped privileges partially, 789 * so we don't even try. If any of our operations fail, we abort() the 790 * entire process. 791 */ 792 if (j->flags.enter_vfs && setns(j->mountns_fd, CLONE_NEWNS)) 793 pdie("setns(CLONE_NEWNS)"); 794 795 if (j->flags.vfs && unshare(CLONE_NEWNS)) 796 pdie("unshare(vfs)"); 797 798 if (j->flags.net && unshare(CLONE_NEWNET)) 799 pdie("unshare(net)"); 800 801 if (j->flags.chroot && enter_chroot(j)) 802 pdie("chroot"); 803 804 if (j->flags.mount_tmp && mount_tmp()) 805 pdie("mount_tmp"); 806 807 if (j->flags.readonly && remount_readonly()) 808 pdie("remount"); 809 810 if (j->flags.caps) { 811 /* 812 * POSIX capabilities are a bit tricky. If we drop our 813 * capability to change uids, our attempt to use setuid() 814 * below will fail. Hang on to root caps across setuid(), then 815 * lock securebits. 816 */ 817 if (prctl(PR_SET_KEEPCAPS, 1)) 818 pdie("prctl(PR_SET_KEEPCAPS)"); 819 if (prctl 820 (PR_SET_SECUREBITS, SECURE_ALL_BITS | SECURE_ALL_LOCKS)) 821 pdie("prctl(PR_SET_SECUREBITS)"); 822 } 823 824 /* 825 * If we're setting no_new_privs, we can drop privileges 826 * before setting seccomp filter. This way filter policies 827 * don't need to allow privilege-dropping syscalls. 828 */ 829 if (j->flags.no_new_privs) { 830 drop_ugid(j); 831 if (j->flags.caps) 832 drop_caps(j); 833 834 set_seccomp_filter(j); 835 } else { 836 /* 837 * If we're not setting no_new_privs, 838 * we need to set seccomp filter *before* dropping privileges. 839 * WARNING: this means that filter policies *must* allow 840 * setgroups()/setresgid()/setresuid() for dropping root and 841 * capget()/capset()/prctl() for dropping caps. 842 */ 843 set_seccomp_filter(j); 844 845 drop_ugid(j); 846 if (j->flags.caps) 847 drop_caps(j); 848 } 849 850 /* 851 * seccomp has to come last since it cuts off all the other 852 * privilege-dropping syscalls :) 853 */ 854 if (j->flags.seccomp && prctl(PR_SET_SECCOMP, 1)) { 855 if ((errno == ENOSYS) && SECCOMP_SOFTFAIL) { 856 warn("seccomp not supported"); 857 return; 858 } 859 pdie("prctl(PR_SET_SECCOMP)"); 860 } 861} 862 863/* TODO(wad) will visibility affect this variable? */ 864static int init_exitstatus = 0; 865 866void init_term(int __attribute__ ((unused)) sig) 867{ 868 _exit(init_exitstatus); 869} 870 871int init(pid_t rootpid) 872{ 873 pid_t pid; 874 int status; 875 /* so that we exit with the right status */ 876 signal(SIGTERM, init_term); 877 /* TODO(wad) self jail with seccomp_filters here. */ 878 while ((pid = wait(&status)) > 0) { 879 /* 880 * This loop will only end when either there are no processes 881 * left inside our pid namespace or we get a signal. 882 */ 883 if (pid == rootpid) 884 init_exitstatus = status; 885 } 886 if (!WIFEXITED(init_exitstatus)) 887 _exit(MINIJAIL_ERR_INIT); 888 _exit(WEXITSTATUS(init_exitstatus)); 889} 890 891int API minijail_from_fd(int fd, struct minijail *j) 892{ 893 size_t sz = 0; 894 size_t bytes = read(fd, &sz, sizeof(sz)); 895 char *buf; 896 int r; 897 if (sizeof(sz) != bytes) 898 return -EINVAL; 899 if (sz > USHRT_MAX) /* Arbitrary sanity check */ 900 return -E2BIG; 901 buf = malloc(sz); 902 if (!buf) 903 return -ENOMEM; 904 bytes = read(fd, buf, sz); 905 if (bytes != sz) { 906 free(buf); 907 return -EINVAL; 908 } 909 r = minijail_unmarshal(j, buf, sz); 910 free(buf); 911 return r; 912} 913 914int API minijail_to_fd(struct minijail *j, int fd) 915{ 916 char *buf; 917 size_t sz = minijail_size(j); 918 ssize_t written; 919 int r; 920 921 if (!sz) 922 return -EINVAL; 923 buf = malloc(sz); 924 r = minijail_marshal(j, buf, sz); 925 if (r) { 926 free(buf); 927 return r; 928 } 929 /* Sends [size][minijail]. */ 930 written = write(fd, &sz, sizeof(sz)); 931 if (written != sizeof(sz)) { 932 free(buf); 933 return -EFAULT; 934 } 935 written = write(fd, buf, sz); 936 if (written < 0 || (size_t) written != sz) { 937 free(buf); 938 return -EFAULT; 939 } 940 free(buf); 941 return 0; 942} 943 944int setup_preload(void) 945{ 946 char *oldenv = getenv(kLdPreloadEnvVar) ? : ""; 947 char *newenv = malloc(strlen(oldenv) + 2 + strlen(PRELOADPATH)); 948 if (!newenv) 949 return -ENOMEM; 950 951 /* Only insert a separating space if we have something to separate... */ 952 sprintf(newenv, "%s%s%s", oldenv, strlen(oldenv) ? " " : "", 953 PRELOADPATH); 954 955 /* setenv() makes a copy of the string we give it */ 956 setenv(kLdPreloadEnvVar, newenv, 1); 957 free(newenv); 958 return 0; 959} 960 961int setup_pipe(int fds[2]) 962{ 963 int r = pipe(fds); 964 char fd_buf[11]; 965 if (r) 966 return r; 967 r = snprintf(fd_buf, sizeof(fd_buf), "%d", fds[0]); 968 if (r <= 0) 969 return -EINVAL; 970 setenv(kFdEnvVar, fd_buf, 1); 971 return 0; 972} 973 974int setup_pipe_end(int fds[2], size_t index) 975{ 976 if (index > 1) 977 return -1; 978 979 close(fds[1 - index]); 980 return fds[index]; 981} 982 983int setup_and_dupe_pipe_end(int fds[2], size_t index, int fd) 984{ 985 if (index > 1) 986 return -1; 987 988 close(fds[1 - index]); 989 /* dup2(2) the corresponding end of the pipe into |fd|. */ 990 return dup2(fds[index], fd); 991} 992 993int API minijail_run(struct minijail *j, const char *filename, 994 char *const argv[]) 995{ 996 return minijail_run_pid_pipes(j, filename, argv, 997 NULL, NULL, NULL, NULL); 998} 999 1000int API minijail_run_pid(struct minijail *j, const char *filename, 1001 char *const argv[], pid_t *pchild_pid) 1002{ 1003 return minijail_run_pid_pipes(j, filename, argv, pchild_pid, 1004 NULL, NULL, NULL); 1005} 1006 1007int API minijail_run_pipe(struct minijail *j, const char *filename, 1008 char *const argv[], int *pstdin_fd) 1009{ 1010 return minijail_run_pid_pipes(j, filename, argv, NULL, pstdin_fd, 1011 NULL, NULL); 1012} 1013 1014int API minijail_run_pid_pipe(struct minijail *j, const char *filename, 1015 char *const argv[], pid_t *pchild_pid, 1016 int *pstdin_fd) 1017{ 1018 return minijail_run_pid_pipes(j, filename, argv, pchild_pid, pstdin_fd, 1019 NULL, NULL); 1020} 1021 1022int API minijail_run_pid_pipes(struct minijail *j, const char *filename, 1023 char *const argv[], pid_t *pchild_pid, 1024 int *pstdin_fd, int *pstdout_fd, int *pstderr_fd) 1025{ 1026 char *oldenv, *oldenv_copy = NULL; 1027 pid_t child_pid; 1028 int pipe_fds[2]; 1029 int stdin_fds[2]; 1030 int stdout_fds[2]; 1031 int stderr_fds[2]; 1032 int ret; 1033 /* We need to remember this across the minijail_preexec() call. */ 1034 int pid_namespace = j->flags.pids; 1035 1036 oldenv = getenv(kLdPreloadEnvVar); 1037 if (oldenv) { 1038 oldenv_copy = strdup(oldenv); 1039 if (!oldenv_copy) 1040 return -ENOMEM; 1041 } 1042 1043 if (setup_preload()) 1044 return -EFAULT; 1045 1046 /* 1047 * Make the process group ID of this process equal to its PID, so that 1048 * both the Minijail process and the jailed process can be killed 1049 * together. 1050 * Don't fail on EPERM, since setpgid(0, 0) can only EPERM when 1051 * the process is already a process group leader. 1052 */ 1053 if (setpgid(0 /* use calling PID */, 0 /* make PGID = PID */)) { 1054 if (errno != EPERM) { 1055 pdie("setpgid(0, 0)"); 1056 } 1057 } 1058 1059 /* 1060 * Before we fork(2) and execve(2) the child process, we need to open 1061 * a pipe(2) to send the minijail configuration over. 1062 */ 1063 if (setup_pipe(pipe_fds)) 1064 return -EFAULT; 1065 1066 /* 1067 * If we want to write to the child process' standard input, 1068 * create the pipe(2) now. 1069 */ 1070 if (pstdin_fd) { 1071 if (pipe(stdin_fds)) 1072 return -EFAULT; 1073 } 1074 1075 /* 1076 * If we want to read from the child process' standard output, 1077 * create the pipe(2) now. 1078 */ 1079 if (pstdout_fd) { 1080 if (pipe(stdout_fds)) 1081 return -EFAULT; 1082 } 1083 1084 /* 1085 * If we want to read from the child process' standard error, 1086 * create the pipe(2) now. 1087 */ 1088 if (pstderr_fd) { 1089 if (pipe(stderr_fds)) 1090 return -EFAULT; 1091 } 1092 1093 /* Use sys_clone() if and only if we're creating a pid namespace. 1094 * 1095 * tl;dr: WARNING: do not mix pid namespaces and multithreading. 1096 * 1097 * In multithreaded programs, there are a bunch of locks inside libc, 1098 * some of which may be held by other threads at the time that we call 1099 * minijail_run_pid(). If we call fork(), glibc does its level best to 1100 * ensure that we hold all of these locks before it calls clone() 1101 * internally and drop them after clone() returns, but when we call 1102 * sys_clone(2) directly, all that gets bypassed and we end up with a 1103 * child address space where some of libc's important locks are held by 1104 * other threads (which did not get cloned, and hence will never release 1105 * those locks). This is okay so long as we call exec() immediately 1106 * after, but a bunch of seemingly-innocent libc functions like setenv() 1107 * take locks. 1108 * 1109 * Hence, only call sys_clone() if we need to, in order to get at pid 1110 * namespacing. If we follow this path, the child's address space might 1111 * have broken locks; you may only call functions that do not acquire 1112 * any locks. 1113 * 1114 * Unfortunately, fork() acquires every lock it can get its hands on, as 1115 * previously detailed, so this function is highly likely to deadlock 1116 * later on (see "deadlock here") if we're multithreaded. 1117 * 1118 * We might hack around this by having the clone()d child (init of the 1119 * pid namespace) return directly, rather than leaving the clone()d 1120 * process hanging around to be init for the new namespace (and having 1121 * its fork()ed child return in turn), but that process would be crippled 1122 * with its libc locks potentially broken. We might try fork()ing in the 1123 * parent before we clone() to ensure that we own all the locks, but 1124 * then we have to have the forked child hanging around consuming 1125 * resources (and possibly having file descriptors / shared memory 1126 * regions / etc attached). We'd need to keep the child around to avoid 1127 * having its children get reparented to init. 1128 * 1129 * TODO(ellyjones): figure out if the "forked child hanging around" 1130 * problem is fixable or not. It would be nice if we worked in this 1131 * case. 1132 */ 1133 if (pid_namespace) 1134 child_pid = syscall(SYS_clone, CLONE_NEWPID | SIGCHLD, NULL); 1135 else 1136 child_pid = fork(); 1137 1138 if (child_pid < 0) { 1139 free(oldenv_copy); 1140 die("failed to fork child"); 1141 } 1142 1143 if (child_pid) { 1144 /* Restore parent's LD_PRELOAD. */ 1145 if (oldenv_copy) { 1146 setenv(kLdPreloadEnvVar, oldenv_copy, 1); 1147 free(oldenv_copy); 1148 } else { 1149 unsetenv(kLdPreloadEnvVar); 1150 } 1151 unsetenv(kFdEnvVar); 1152 1153 j->initpid = child_pid; 1154 1155 /* Send marshalled minijail. */ 1156 close(pipe_fds[0]); /* read endpoint */ 1157 ret = minijail_to_fd(j, pipe_fds[1]); 1158 close(pipe_fds[1]); /* write endpoint */ 1159 if (ret) { 1160 kill(j->initpid, SIGKILL); 1161 die("failed to send marshalled minijail"); 1162 } 1163 1164 if (pchild_pid) 1165 *pchild_pid = child_pid; 1166 1167 /* 1168 * If we want to write to the child process' standard input, 1169 * set up the write end of the pipe. 1170 */ 1171 if (pstdin_fd) 1172 *pstdin_fd = setup_pipe_end(stdin_fds, 1173 1 /* write end */); 1174 1175 /* 1176 * If we want to read from the child process' standard output, 1177 * set up the read end of the pipe. 1178 */ 1179 if (pstdout_fd) 1180 *pstdout_fd = setup_pipe_end(stdout_fds, 1181 0 /* read end */); 1182 1183 /* 1184 * If we want to read from the child process' standard error, 1185 * set up the read end of the pipe. 1186 */ 1187 if (pstderr_fd) 1188 *pstderr_fd = setup_pipe_end(stderr_fds, 1189 0 /* read end */); 1190 1191 return 0; 1192 } 1193 free(oldenv_copy); 1194 1195 /* 1196 * If we want to write to the jailed process' standard input, 1197 * set up the read end of the pipe. 1198 */ 1199 if (pstdin_fd) { 1200 if (setup_and_dupe_pipe_end(stdin_fds, 0 /* read end */, 1201 STDIN_FILENO) < 0) 1202 die("failed to set up stdin pipe"); 1203 } 1204 1205 /* 1206 * If we want to read from the jailed process' standard output, 1207 * set up the write end of the pipe. 1208 */ 1209 if (pstdout_fd) { 1210 if (setup_and_dupe_pipe_end(stdout_fds, 1 /* write end */, 1211 STDOUT_FILENO) < 0) 1212 die("failed to set up stdout pipe"); 1213 } 1214 1215 /* 1216 * If we want to read from the jailed process' standard error, 1217 * set up the write end of the pipe. 1218 */ 1219 if (pstderr_fd) { 1220 if (setup_and_dupe_pipe_end(stderr_fds, 1 /* write end */, 1221 STDERR_FILENO) < 0) 1222 die("failed to set up stderr pipe"); 1223 } 1224 1225 /* Strip out flags that cannot be inherited across execve. */ 1226 minijail_preexec(j); 1227 /* Jail this process and its descendants... */ 1228 minijail_enter(j); 1229 1230 if (pid_namespace) { 1231 /* 1232 * pid namespace: this process will become init inside the new 1233 * namespace, so fork off a child to actually run the program 1234 * (we don't want all programs we might exec to have to know 1235 * how to be init). 1236 * 1237 * If we're multithreaded, we'll probably deadlock here. See 1238 * WARNING above. 1239 */ 1240 child_pid = fork(); 1241 if (child_pid < 0) 1242 _exit(child_pid); 1243 else if (child_pid > 0) 1244 init(child_pid); /* never returns */ 1245 } 1246 1247 /* 1248 * If we aren't pid-namespaced: 1249 * calling process 1250 * -> execve()-ing process 1251 * If we are: 1252 * calling process 1253 * -> init()-ing process 1254 * -> execve()-ing process 1255 */ 1256 _exit(execve(filename, argv, environ)); 1257} 1258 1259int API minijail_run_static(struct minijail *j, const char *filename, 1260 char *const argv[]) 1261{ 1262 pid_t child_pid; 1263 int pid_namespace = j->flags.pids; 1264 1265 if (j->flags.caps) 1266 die("caps not supported with static targets"); 1267 1268 if (pid_namespace) 1269 child_pid = syscall(SYS_clone, CLONE_NEWPID | SIGCHLD, NULL); 1270 else 1271 child_pid = fork(); 1272 1273 if (child_pid < 0) { 1274 die("failed to fork child"); 1275 } 1276 if (child_pid > 0 ) { 1277 j->initpid = child_pid; 1278 return 0; 1279 } 1280 1281 /* 1282 * We can now drop this child into the sandbox 1283 * then execve the target. 1284 */ 1285 1286 j->flags.pids = 0; 1287 minijail_enter(j); 1288 1289 if (pid_namespace) { 1290 /* 1291 * pid namespace: this process will become init inside the new 1292 * namespace, so fork off a child to actually run the program 1293 * (we don't want all programs we might exec to have to know 1294 * how to be init). 1295 * 1296 * If we're multithreaded, we'll probably deadlock here. See 1297 * WARNING above. 1298 */ 1299 child_pid = fork(); 1300 if (child_pid < 0) 1301 _exit(child_pid); 1302 else if (child_pid > 0) 1303 init(child_pid); /* never returns */ 1304 } 1305 1306 _exit(execve(filename, argv, environ)); 1307} 1308 1309int API minijail_kill(struct minijail *j) 1310{ 1311 int st; 1312 if (kill(j->initpid, SIGTERM)) 1313 return -errno; 1314 if (waitpid(j->initpid, &st, 0) < 0) 1315 return -errno; 1316 return st; 1317} 1318 1319int API minijail_wait(struct minijail *j) 1320{ 1321 int st; 1322 if (waitpid(j->initpid, &st, 0) < 0) 1323 return -errno; 1324 1325 if (!WIFEXITED(st)) { 1326 int error_status = st; 1327 if (WIFSIGNALED(st)) { 1328 int signum = WTERMSIG(st); 1329 warn("child process %d received signal %d", 1330 j->initpid, signum); 1331 /* 1332 * We return MINIJAIL_ERR_JAIL if the process received 1333 * SIGSYS, which happens when a syscall is blocked by 1334 * seccomp filters. 1335 * If not, we do what bash(1) does: 1336 * $? = 128 + signum 1337 */ 1338 if (signum == SIGSYS) { 1339 error_status = MINIJAIL_ERR_JAIL; 1340 } else { 1341 error_status = 128 + signum; 1342 } 1343 } 1344 return error_status; 1345 } 1346 1347 int exit_status = WEXITSTATUS(st); 1348 if (exit_status != 0) 1349 info("child process %d exited with status %d", 1350 j->initpid, exit_status); 1351 1352 return exit_status; 1353} 1354 1355void API minijail_destroy(struct minijail *j) 1356{ 1357 if (j->flags.seccomp_filter && j->filter_prog) { 1358 free(j->filter_prog->filter); 1359 free(j->filter_prog); 1360 } 1361 while (j->bindings_head) { 1362 struct binding *b = j->bindings_head; 1363 j->bindings_head = j->bindings_head->next; 1364 free(b->dest); 1365 free(b->src); 1366 free(b); 1367 } 1368 j->bindings_tail = NULL; 1369 if (j->user) 1370 free(j->user); 1371 if (j->chrootdir) 1372 free(j->chrootdir); 1373 free(j); 1374} 1375