src/nspawn/nspawn-patch-uid.c

   1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
   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 "fileio.h"
  12 #include "fs-util.h"
  13 #include "missing_magic.h"
  14 #include "missing_syscall.h"
  15 #include "nspawn-def.h"
  16 #include "nspawn-patch-uid.h"
  17 #include "stat-util.h"
  18 #include "stdio-util.h"
  19 #include "string-util.h"
  20 #include "strv.h"
  21 #include "user-util.h"
  22
  23 #if HAVE_ACL
  24
  25 static int get_acl(int fd, const char *name, acl_type_t type, acl_t *ret) {
  26         acl_t acl;
  27
  28         assert(fd >= 0);
  29         assert(ret);
  30
  31         if (name) {
  32                 _cleanup_close_ int child_fd = -EBADF;
  33
  34                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  35                 if (child_fd < 0)
  36                         return -errno;
  37
  38                 acl = acl_get_file(FORMAT_PROC_FD_PATH(child_fd), type);
  39         } else if (type == ACL_TYPE_ACCESS)
  40                 acl = acl_get_fd(fd);
  41         else
  42                 acl = acl_get_file(FORMAT_PROC_FD_PATH(fd), type);
  43         if (!acl)
  44                 return -errno;
  45
  46         *ret = acl;
  47         return 0;
  48 }
  49
  50 static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
  51         int r;
  52
  53         assert(fd >= 0);
  54         assert(acl);
  55
  56         if (name) {
  57                 _cleanup_close_ int child_fd = -EBADF;
  58
  59                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  60                 if (child_fd < 0)
  61                         return -errno;
  62
  63                 r = acl_set_file(FORMAT_PROC_FD_PATH(child_fd), type, acl);
  64         } else if (type == ACL_TYPE_ACCESS)
  65                 r = acl_set_fd(fd, acl);
  66         else
  67                 r = acl_set_file(FORMAT_PROC_FD_PATH(fd), type, acl);
  68         if (r < 0)
  69                 return -errno;
  70
  71         return 0;
  72 }
  73
  74 static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
  75         _cleanup_(acl_freep) acl_t copy = NULL;
  76         acl_entry_t i;
  77         int r;
  78
  79         assert(acl);
  80         assert(ret);
  81
  82         r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
  83         if (r < 0)
  84                 return -errno;
  85         while (r > 0) {
  86                 uid_t *old_uid, new_uid;
  87                 bool modify = false;
  88                 acl_tag_t tag;
  89
  90                 if (acl_get_tag_type(i, &tag) < 0)
  91                         return -errno;
  92
  93                 if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
  94
  95                         /* We don't distinguish here between uid_t and gid_t, let's make sure the compiler checks that
  96                          * this is actually OK */
  97                         assert_cc(sizeof(uid_t) == sizeof(gid_t));
  98
  99                         old_uid = acl_get_qualifier(i);
 100                         if (!old_uid)
 101                                 return -errno;
 102
 103                         new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
 104                         if (!uid_is_valid(new_uid))
 105                                 return -EINVAL;
 106
 107                         modify = new_uid != *old_uid;
 108                         if (modify && !copy) {
 109                                 int n;
 110
 111                                 /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
 112                                  * beginning, so that we copy all entries, starting from the first, this time. */
 113
 114                                 n = acl_entries(acl);
 115                                 if (n < 0)
 116                                         return -errno;
 117
 118                                 copy = acl_init(n);
 119                                 if (!copy)
 120                                         return -errno;
 121
 122                                 /* Seek back to the beginning */
 123                                 r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
 124                                 if (r < 0)
 125                                         return -errno;
 126                                 continue;
 127                         }
 128                 }
 129
 130                 if (copy) {
 131                         acl_entry_t new_entry;
 132
 133                         if (acl_create_entry(&copy, &new_entry) < 0)
 134                                 return -errno;
 135
 136                         if (acl_copy_entry(new_entry, i) < 0)
 137                                 return -errno;
 138
 139                         if (modify)
 140                                 if (acl_set_qualifier(new_entry, &new_uid) < 0)
 141                                         return -errno;
 142                 }
 143
 144                 r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
 145                 if (r < 0)
 146                         return -errno;
 147         }
 148
 149         *ret = TAKE_PTR(copy);
 150
 151         return !!*ret;
 152 }
 153
 154 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 155         _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
 156         bool changed = false;
 157         int r;
 158
 159         assert(fd >= 0);
 160         assert(st);
 161
 162         /* ACLs are not supported on symlinks, there's no point in trying */
 163         if (S_ISLNK(st->st_mode))
 164                 return 0;
 165
 166         r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
 167         if (r == -EOPNOTSUPP)
 168                 return 0;
 169         if (r < 0)
 170                 return r;
 171
 172         r = shift_acl(acl, shift, &shifted);
 173         if (r < 0)
 174                 return r;
 175         if (r > 0) {
 176                 r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
 177                 if (r < 0)
 178                         return r;
 179
 180                 changed = true;
 181         }
 182
 183         if (S_ISDIR(st->st_mode)) {
 184                 acl_free(acl);
 185
 186                 if (shifted)
 187                         acl_free(shifted);
 188
 189                 acl = shifted = NULL;
 190
 191                 r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
 192                 if (r < 0)
 193                         return r;
 194
 195                 r = shift_acl(acl, shift, &shifted);
 196                 if (r < 0)
 197                         return r;
 198                 if (r > 0) {
 199                         r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
 200                         if (r < 0)
 201                                 return r;
 202
 203                         changed = true;
 204                 }
 205         }
 206
 207         return changed;
 208 }
 209
 210 #else
 211
 212 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 213         return 0;
 214 }
 215
 216 #endif
 217
 218 static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
 219         uid_t new_uid;
 220         gid_t new_gid;
 221         bool changed = false;
 222         int r;
 223
 224         assert(fd >= 0);
 225         assert(st);
 226
 227         new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
 228         new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));
 229
 230         if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
 231                 return -EINVAL;
 232
 233         if (st->st_uid != new_uid || st->st_gid != new_gid) {
 234                 if (name)
 235                         r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
 236                 else
 237                         r = fchown(fd, new_uid, new_gid);
 238                 if (r < 0)
 239                         return -errno;
 240
 241                 /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
 242                 if (name) {
 243                         /* It looks like older glibc (before 2016) did not support AT_SYMLINK_NOFOLLOW. */
 244                         if (!S_ISLNK(st->st_mode))
 245                                 r = RET_NERRNO(fchmodat(fd, name, st->st_mode, 0));
 246                         else {
 247                                 r = RET_NERRNO(fchmodat2(fd, name, st->st_mode, AT_SYMLINK_NOFOLLOW));
 248                                 if (IN_SET(r, -ENOSYS, -EPERM))
 249                                         r = 0;
 250                         }
 251                 } else
 252                         r = RET_NERRNO(fchmod(fd, st->st_mode));
 253                 if (r < 0)
 254                         return r;
 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                 if (!donate_fd) {
 324                         int copy;
 325
 326                         copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
 327                         if (copy < 0) {
 328                                 r = -errno;
 329                                 goto finish;
 330                         }
 331
 332                         fd = copy;
 333                         donate_fd = true;
 334                 }
 335
 336                 d = take_fdopendir(&fd);
 337                 if (!d) {
 338                         r = -errno;
 339                         goto finish;
 340                 }
 341
 342                 FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
 343                         struct stat fst;
 344
 345                         if (dot_or_dot_dot(de->d_name))
 346                                 continue;
 347
 348                         if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
 349                                 r = -errno;
 350                                 goto finish;
 351                         }
 352
 353                         if (S_ISDIR(fst.st_mode)) {
 354                                 int subdir_fd;
 355
 356                                 subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 357                                 if (subdir_fd < 0) {
 358                                         r = -errno;
 359                                         goto finish;
 360
 361                                 }
 362
 363                                 r = recurse_fd(subdir_fd, true, &fst, shift, false);
 364                                 if (r < 0)
 365                                         goto finish;
 366                                 if (r > 0)
 367                                         changed = true;
 368
 369                         } else {
 370                                 r = patch_fd(dirfd(d), de->d_name, &fst, shift);
 371                                 if (r < 0)
 372                                         goto finish;
 373                                 if (r > 0)
 374                                         changed = true;
 375                         }
 376                 }
 377         }
 378
 379         /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
 380          * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
 381          * tree is properly chown()ed already. */
 382         r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
 383         if (r == -EROFS)
 384                 goto read_only;
 385         if (r > 0)
 386                 changed = true;
 387
 388         r = changed;
 389         goto finish;
 390
 391 read_only:
 392         if (!is_toplevel) {
 393                 _cleanup_free_ char *name = NULL;
 394
 395                 /* When we hit a ready-only subtree we simply skip it, but log about it. */
 396                 (void) fd_get_path(fd, &name);
 397                 log_debug("Skipping read-only file or directory %s.", strna(name));
 398                 r = changed;
 399         }
 400
 401 finish:
 402         if (donate_fd)
 403                 safe_close(fd);
 404
 405         return r;
 406 }
 407
 408 static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
 409         struct stat st;
 410         int r;
 411
 412         assert(fd >= 0);
 413
 414         /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
 415          * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
 416          * following the concept that the upper 16-bit of a UID identify the container, and the lower 16-bit are the actual
 417          * UID within the container. */
 418
 419         if ((shift & 0xFFFF) != 0) {
 420                 /* We only support containers where the shift starts at a 2^16 boundary */
 421                 r = -EOPNOTSUPP;
 422                 goto finish;
 423         }
 424
 425         if (shift == UID_BUSY_BASE) {
 426                 r = -EINVAL;
 427                 goto finish;
 428         }
 429
 430         if (range != 0x10000) {
 431                 /* We only support containers with 16-bit UID ranges for the patching logic */
 432                 r = -EOPNOTSUPP;
 433                 goto finish;
 434         }
 435
 436         if (fstat(fd, &st) < 0) {
 437                 r = -errno;
 438                 goto finish;
 439         }
 440
 441         if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
 442                 /* We only support containers where the uid/gid container ID match */
 443                 r = -EBADE;
 444                 goto finish;
 445         }
 446
 447         /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
 448          * that if the top-level dir has the right upper 16-bit assigned, then everything below will have too... */
 449         if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
 450                 return 0;
 451
 452         /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
 453          * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
 454          * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
 455
 456         if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
 457                 if (fchown(fd,
 458                            UID_BUSY_BASE | (st.st_uid & ~UID_BUSY_MASK),
 459                            (gid_t) UID_BUSY_BASE | (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
 460                         r = -errno;
 461                         goto finish;
 462                 }
 463         }
 464
 465         return recurse_fd(fd, donate_fd, &st, shift, true);
 466
 467 finish:
 468         if (donate_fd)
 469                 safe_close(fd);
 470
 471         return r;
 472 }
 473
 474 int path_patch_uid(const char *path, uid_t shift, uid_t range) {
 475         int fd;
 476
 477         fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 478         if (fd < 0)
 479                 return -errno;
 480
 481         return fd_patch_uid_internal(fd, true, shift, range);
 482 }