src/nspawn/nspawn-patch-uid.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #include <fcntl.h>
   4 #include <sys/statvfs.h>
   5 #include <sys/vfs.h>
   6 #include <unistd.h>
   7
   8 #include "acl-util.h"
   9 #include "dirent-util.h"
  10 #include "fd-util.h"
  11 #include "fs-util.h"
  12 #include "missing_magic.h"
  13 #include "nspawn-def.h"
  14 #include "nspawn-patch-uid.h"
  15 #include "stat-util.h"
  16 #include "stdio-util.h"
  17 #include "string-util.h"
  18 #include "strv.h"
  19 #include "user-util.h"
  20
  21 #if HAVE_ACL
  22
  23 static int get_acl(int fd, const char *name, acl_type_t type, acl_t *ret) {
  24         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  25         acl_t acl;
  26
  27         assert(fd >= 0);
  28         assert(ret);
  29
  30         if (name) {
  31                 _cleanup_close_ int child_fd = -1;
  32
  33                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  34                 if (child_fd < 0)
  35                         return -errno;
  36
  37                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  38                 acl = acl_get_file(procfs_path, type);
  39         } else if (type == ACL_TYPE_ACCESS)
  40                 acl = acl_get_fd(fd);
  41         else {
  42                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  43                 acl = acl_get_file(procfs_path, type);
  44         }
  45         if (!acl)
  46                 return -errno;
  47
  48         *ret = acl;
  49         return 0;
  50 }
  51
  52 static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
  53         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  54         int r;
  55
  56         assert(fd >= 0);
  57         assert(acl);
  58
  59         if (name) {
  60                 _cleanup_close_ int child_fd = -1;
  61
  62                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  63                 if (child_fd < 0)
  64                         return -errno;
  65
  66                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  67                 r = acl_set_file(procfs_path, type, acl);
  68         } else if (type == ACL_TYPE_ACCESS)
  69                 r = acl_set_fd(fd, acl);
  70         else {
  71                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  72                 r = acl_set_file(procfs_path, type, acl);
  73         }
  74         if (r < 0)
  75                 return -errno;
  76
  77         return 0;
  78 }
  79
  80 static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
  81         _cleanup_(acl_freep) acl_t copy = NULL;
  82         acl_entry_t i;
  83         int r;
  84
  85         assert(acl);
  86         assert(ret);
  87
  88         r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
  89         if (r < 0)
  90                 return -errno;
  91         while (r > 0) {
  92                 uid_t *old_uid, new_uid;
  93                 bool modify = false;
  94                 acl_tag_t tag;
  95
  96                 if (acl_get_tag_type(i, &tag) < 0)
  97                         return -errno;
  98
  99                 if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
 100
 101                         /* We don't distinguish here between uid_t and gid_t, let's make sure the compiler checks that
 102                          * this is actually OK */
 103                         assert_cc(sizeof(uid_t) == sizeof(gid_t));
 104
 105                         old_uid = acl_get_qualifier(i);
 106                         if (!old_uid)
 107                                 return -errno;
 108
 109                         new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
 110                         if (!uid_is_valid(new_uid))
 111                                 return -EINVAL;
 112
 113                         modify = new_uid != *old_uid;
 114                         if (modify && !copy) {
 115                                 int n;
 116
 117                                 /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
 118                                  * beginning, so that we copy all entries, starting from the first, this time. */
 119
 120                                 n = acl_entries(acl);
 121                                 if (n < 0)
 122                                         return -errno;
 123
 124                                 copy = acl_init(n);
 125                                 if (!copy)
 126                                         return -errno;
 127
 128                                 /* Seek back to the beginning */
 129                                 r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
 130                                 if (r < 0)
 131                                         return -errno;
 132                                 continue;
 133                         }
 134                 }
 135
 136                 if (copy) {
 137                         acl_entry_t new_entry;
 138
 139                         if (acl_create_entry(&copy, &new_entry) < 0)
 140                                 return -errno;
 141
 142                         if (acl_copy_entry(new_entry, i) < 0)
 143                                 return -errno;
 144
 145                         if (modify)
 146                                 if (acl_set_qualifier(new_entry, &new_uid) < 0)
 147                                         return -errno;
 148                 }
 149
 150                 r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
 151                 if (r < 0)
 152                         return -errno;
 153         }
 154
 155         *ret = TAKE_PTR(copy);
 156
 157         return !!*ret;
 158 }
 159
 160 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 161         _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
 162         bool changed = false;
 163         int r;
 164
 165         assert(fd >= 0);
 166         assert(st);
 167
 168         /* ACLs are not supported on symlinks, there's no point in trying */
 169         if (S_ISLNK(st->st_mode))
 170                 return 0;
 171
 172         r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
 173         if (r == -EOPNOTSUPP)
 174                 return 0;
 175         if (r < 0)
 176                 return r;
 177
 178         r = shift_acl(acl, shift, &shifted);
 179         if (r < 0)
 180                 return r;
 181         if (r > 0) {
 182                 r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
 183                 if (r < 0)
 184                         return r;
 185
 186                 changed = true;
 187         }
 188
 189         if (S_ISDIR(st->st_mode)) {
 190                 acl_free(acl);
 191                 acl_free(shifted);
 192
 193                 acl = shifted = NULL;
 194
 195                 r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
 196                 if (r < 0)
 197                         return r;
 198
 199                 r = shift_acl(acl, shift, &shifted);
 200                 if (r < 0)
 201                         return r;
 202                 if (r > 0) {
 203                         r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
 204                         if (r < 0)
 205                                 return r;
 206
 207                         changed = true;
 208                 }
 209         }
 210
 211         return changed;
 212 }
 213
 214 #else
 215
 216 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 217         return 0;
 218 }
 219
 220 #endif
 221
 222 static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
 223         uid_t new_uid;
 224         gid_t new_gid;
 225         bool changed = false;
 226         int r;
 227
 228         assert(fd >= 0);
 229         assert(st);
 230
 231         new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
 232         new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));
 233
 234         if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
 235                 return -EINVAL;
 236
 237         if (st->st_uid != new_uid || st->st_gid != new_gid) {
 238                 if (name)
 239                         r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
 240                 else
 241                         r = fchown(fd, new_uid, new_gid);
 242                 if (r < 0)
 243                         return -errno;
 244
 245                 /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
 246                 if (name) {
 247                         if (!S_ISLNK(st->st_mode))
 248                                 r = fchmodat(fd, name, st->st_mode, 0);
 249                         else /* AT_SYMLINK_NOFOLLOW is not available for fchmodat() */
 250                                 r = 0;
 251                 } else
 252                         r = fchmod(fd, st->st_mode);
 253                 if (r < 0)
 254                         return -errno;
 255
 256                 changed = true;
 257         }
 258
 259         r = patch_acls(fd, name, st, shift);
 260         if (r < 0)
 261                 return r;
 262
 263         return r > 0 || changed;
 264 }
 265
 266 /*
 267  * Check if the filesystem is fully compatible with user namespaces or
 268  * UID/GID patching. Some filesystems in this list can be fully mounted inside
 269  * user namespaces, however their inodes may relate to host resources or only
 270  * valid in the global user namespace, therefore no patching should be applied.
 271  */
 272 static int is_fs_fully_userns_compatible(const struct statfs *sfs) {
 273
 274         assert(sfs);
 275
 276         return F_TYPE_EQUAL(sfs->f_type, BINFMTFS_MAGIC) ||
 277                F_TYPE_EQUAL(sfs->f_type, CGROUP_SUPER_MAGIC) ||
 278                F_TYPE_EQUAL(sfs->f_type, CGROUP2_SUPER_MAGIC) ||
 279                F_TYPE_EQUAL(sfs->f_type, DEBUGFS_MAGIC) ||
 280                F_TYPE_EQUAL(sfs->f_type, DEVPTS_SUPER_MAGIC) ||
 281                F_TYPE_EQUAL(sfs->f_type, EFIVARFS_MAGIC) ||
 282                F_TYPE_EQUAL(sfs->f_type, HUGETLBFS_MAGIC) ||
 283                F_TYPE_EQUAL(sfs->f_type, MQUEUE_MAGIC) ||
 284                F_TYPE_EQUAL(sfs->f_type, PROC_SUPER_MAGIC) ||
 285                F_TYPE_EQUAL(sfs->f_type, PSTOREFS_MAGIC) ||
 286                F_TYPE_EQUAL(sfs->f_type, SELINUX_MAGIC) ||
 287                F_TYPE_EQUAL(sfs->f_type, SMACK_MAGIC) ||
 288                F_TYPE_EQUAL(sfs->f_type, SECURITYFS_MAGIC) ||
 289                F_TYPE_EQUAL(sfs->f_type, BPF_FS_MAGIC) ||
 290                F_TYPE_EQUAL(sfs->f_type, TRACEFS_MAGIC) ||
 291                F_TYPE_EQUAL(sfs->f_type, SYSFS_MAGIC);
 292 }
 293
 294 static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift, bool is_toplevel) {
 295         _cleanup_closedir_ DIR *d = NULL;
 296         bool changed = false;
 297         struct statfs sfs;
 298         int r;
 299
 300         assert(fd >= 0);
 301
 302         if (fstatfs(fd, &sfs) < 0)
 303                 return -errno;
 304
 305         /* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
 306          * shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
 307          * when we hit procfs, sysfs or some other special file systems. */
 308
 309         r = is_fs_fully_userns_compatible(&sfs);
 310         if (r < 0)
 311                 goto finish;
 312         if (r > 0) {
 313                 r = 0; /* don't recurse */
 314                 goto finish;
 315         }
 316
 317         /* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
 318         if ((sfs.f_flags & ST_RDONLY) ||
 319             access_fd(fd, W_OK) == -EROFS)
 320                 goto read_only;
 321
 322         if (S_ISDIR(st->st_mode)) {
 323                 struct dirent *de;
 324
 325                 if (!donate_fd) {
 326                         int copy;
 327
 328                         copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
 329                         if (copy < 0) {
 330                                 r = -errno;
 331                                 goto finish;
 332                         }
 333
 334                         fd = copy;
 335                         donate_fd = true;
 336                 }
 337
 338                 d = fdopendir(fd);
 339                 if (!d) {
 340                         r = -errno;
 341                         goto finish;
 342                 }
 343                 fd = -1;
 344
 345                 FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
 346                         struct stat fst;
 347
 348                         if (dot_or_dot_dot(de->d_name))
 349                                 continue;
 350
 351                         if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
 352                                 r = -errno;
 353                                 goto finish;
 354                         }
 355
 356                         if (S_ISDIR(fst.st_mode)) {
 357                                 int subdir_fd;
 358
 359                                 subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 360                                 if (subdir_fd < 0) {
 361                                         r = -errno;
 362                                         goto finish;
 363
 364                                 }
 365
 366                                 r = recurse_fd(subdir_fd, true, &fst, shift, false);
 367                                 if (r < 0)
 368                                         goto finish;
 369                                 if (r > 0)
 370                                         changed = true;
 371
 372                         } else {
 373                                 r = patch_fd(dirfd(d), de->d_name, &fst, shift);
 374                                 if (r < 0)
 375                                         goto finish;
 376                                 if (r > 0)
 377                                         changed = true;
 378                         }
 379                 }
 380         }
 381
 382         /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
 383          * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
 384          * tree is properly chown()ed already. */
 385         r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
 386         if (r == -EROFS)
 387                 goto read_only;
 388         if (r > 0)
 389                 changed = true;
 390
 391         r = changed;
 392         goto finish;
 393
 394 read_only:
 395         if (!is_toplevel) {
 396                 _cleanup_free_ char *name = NULL;
 397
 398                 /* When we hit a ready-only subtree we simply skip it, but log about it. */
 399                 (void) fd_get_path(fd, &name);
 400                 log_debug("Skipping read-only file or directory %s.", strna(name));
 401                 r = changed;
 402         }
 403
 404 finish:
 405         if (donate_fd)
 406                 safe_close(fd);
 407
 408         return r;
 409 }
 410
 411 static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
 412         struct stat st;
 413         int r;
 414
 415         assert(fd >= 0);
 416
 417         /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
 418          * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
 419          * following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
 420          * UID within the container. */
 421
 422         if ((shift & 0xFFFF) != 0) {
 423                 /* We only support containers where the shift starts at a 2^16 boundary */
 424                 r = -EOPNOTSUPP;
 425                 goto finish;
 426         }
 427
 428         if (shift == UID_BUSY_BASE) {
 429                 r = -EINVAL;
 430                 goto finish;
 431         }
 432
 433         if (range != 0x10000) {
 434                 /* We only support containers with 16bit UID ranges for the patching logic */
 435                 r = -EOPNOTSUPP;
 436                 goto finish;
 437         }
 438
 439         if (fstat(fd, &st) < 0) {
 440                 r = -errno;
 441                 goto finish;
 442         }
 443
 444         if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
 445                 /* We only support containers where the uid/gid container ID match */
 446                 r = -EBADE;
 447                 goto finish;
 448         }
 449
 450         /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
 451          * that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
 452         if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
 453                 return 0;
 454
 455         /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
 456          * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
 457          * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
 458
 459         if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
 460                 if (fchown(fd,
 461                            UID_BUSY_BASE | (st.st_uid & ~UID_BUSY_MASK),
 462                            (gid_t) UID_BUSY_BASE | (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
 463                         r = -errno;
 464                         goto finish;
 465                 }
 466         }
 467
 468         return recurse_fd(fd, donate_fd, &st, shift, true);
 469
 470 finish:
 471         if (donate_fd)
 472                 safe_close(fd);
 473
 474         return r;
 475 }
 476
 477 int path_patch_uid(const char *path, uid_t shift, uid_t range) {
 478         int fd;
 479
 480         fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 481         if (fd < 0)
 482                 return -errno;
 483
 484         return fd_patch_uid_internal(fd, true, shift, range);
 485 }