src/nspawn/nspawn-patch-uid.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2
   3 #include <fcntl.h>
   4 #include <linux/magic.h>
   5 #include <sys/stat.h>
   6 #include <sys/statvfs.h>
   7 #include <sys/vfs.h>
   8 #include <unistd.h>
   9
  10 #include "acl-util.h"
  11 #include "dirent-util.h"
  12 #include "fd-util.h"
  13 #include "fs-util.h"
  14 #include "missing_magic.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         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  27         acl_t acl;
  28
  29         assert(fd >= 0);
  30         assert(ret);
  31
  32         if (name) {
  33                 _cleanup_close_ int child_fd = -1;
  34
  35                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  36                 if (child_fd < 0)
  37                         return -errno;
  38
  39                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  40                 acl = acl_get_file(procfs_path, type);
  41         } else if (type == ACL_TYPE_ACCESS)
  42                 acl = acl_get_fd(fd);
  43         else {
  44                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  45                 acl = acl_get_file(procfs_path, type);
  46         }
  47         if (!acl)
  48                 return -errno;
  49
  50         *ret = acl;
  51         return 0;
  52 }
  53
  54 static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
  55         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  56         int r;
  57
  58         assert(fd >= 0);
  59         assert(acl);
  60
  61         if (name) {
  62                 _cleanup_close_ int child_fd = -1;
  63
  64                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  65                 if (child_fd < 0)
  66                         return -errno;
  67
  68                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  69                 r = acl_set_file(procfs_path, type, acl);
  70         } else if (type == ACL_TYPE_ACCESS)
  71                 r = acl_set_fd(fd, acl);
  72         else {
  73                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  74                 r = acl_set_file(procfs_path, type, acl);
  75         }
  76         if (r < 0)
  77                 return -errno;
  78
  79         return 0;
  80 }
  81
  82 static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
  83         _cleanup_(acl_freep) acl_t copy = NULL;
  84         acl_entry_t i;
  85         int r;
  86
  87         assert(acl);
  88         assert(ret);
  89
  90         r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
  91         if (r < 0)
  92                 return -errno;
  93         while (r > 0) {
  94                 uid_t *old_uid, new_uid;
  95                 bool modify = false;
  96                 acl_tag_t tag;
  97
  98                 if (acl_get_tag_type(i, &tag) < 0)
  99                         return -errno;
 100
 101                 if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
 102
 103                         /* We don't distinguish here between uid_t and gid_t, let's make sure the compiler checks that
 104                          * this is actually OK */
 105                         assert_cc(sizeof(uid_t) == sizeof(gid_t));
 106
 107                         old_uid = acl_get_qualifier(i);
 108                         if (!old_uid)
 109                                 return -errno;
 110
 111                         new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
 112                         if (!uid_is_valid(new_uid))
 113                                 return -EINVAL;
 114
 115                         modify = new_uid != *old_uid;
 116                         if (modify && !copy) {
 117                                 int n;
 118
 119                                 /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
 120                                  * beginning, so that we copy all entries, starting from the first, this time. */
 121
 122                                 n = acl_entries(acl);
 123                                 if (n < 0)
 124                                         return -errno;
 125
 126                                 copy = acl_init(n);
 127                                 if (!copy)
 128                                         return -errno;
 129
 130                                 /* Seek back to the beginning */
 131                                 r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
 132                                 if (r < 0)
 133                                         return -errno;
 134                                 continue;
 135                         }
 136                 }
 137
 138                 if (copy) {
 139                         acl_entry_t new_entry;
 140
 141                         if (acl_create_entry(&copy, &new_entry) < 0)
 142                                 return -errno;
 143
 144                         if (acl_copy_entry(new_entry, i) < 0)
 145                                 return -errno;
 146
 147                         if (modify)
 148                                 if (acl_set_qualifier(new_entry, &new_uid) < 0)
 149                                         return -errno;
 150                 }
 151
 152                 r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
 153                 if (r < 0)
 154                         return -errno;
 155         }
 156
 157         *ret = TAKE_PTR(copy);
 158
 159         return !!*ret;
 160 }
 161
 162 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 163         _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
 164         bool changed = false;
 165         int r;
 166
 167         assert(fd >= 0);
 168         assert(st);
 169
 170         /* ACLs are not supported on symlinks, there's no point in trying */
 171         if (S_ISLNK(st->st_mode))
 172                 return 0;
 173
 174         r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
 175         if (r == -EOPNOTSUPP)
 176                 return 0;
 177         if (r < 0)
 178                 return r;
 179
 180         r = shift_acl(acl, shift, &shifted);
 181         if (r < 0)
 182                 return r;
 183         if (r > 0) {
 184                 r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
 185                 if (r < 0)
 186                         return r;
 187
 188                 changed = true;
 189         }
 190
 191         if (S_ISDIR(st->st_mode)) {
 192                 acl_free(acl);
 193                 acl_free(shifted);
 194
 195                 acl = shifted = NULL;
 196
 197                 r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
 198                 if (r < 0)
 199                         return r;
 200
 201                 r = shift_acl(acl, shift, &shifted);
 202                 if (r < 0)
 203                         return r;
 204                 if (r > 0) {
 205                         r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
 206                         if (r < 0)
 207                                 return r;
 208
 209                         changed = true;
 210                 }
 211         }
 212
 213         return changed;
 214 }
 215
 216 #else
 217
 218 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 219         return 0;
 220 }
 221
 222 #endif
 223
 224 static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
 225         uid_t new_uid;
 226         gid_t new_gid;
 227         bool changed = false;
 228         int r;
 229
 230         assert(fd >= 0);
 231         assert(st);
 232
 233         new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
 234         new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));
 235
 236         if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
 237                 return -EINVAL;
 238
 239         if (st->st_uid != new_uid || st->st_gid != new_gid) {
 240                 if (name)
 241                         r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
 242                 else
 243                         r = fchown(fd, new_uid, new_gid);
 244                 if (r < 0)
 245                         return -errno;
 246
 247                 /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
 248                 if (name) {
 249                         if (!S_ISLNK(st->st_mode))
 250                                 r = fchmodat(fd, name, st->st_mode, 0);
 251                         else /* AT_SYMLINK_NOFOLLOW is not available for fchmodat() */
 252                                 r = 0;
 253                 } else
 254                         r = fchmod(fd, st->st_mode);
 255                 if (r < 0)
 256                         return -errno;
 257
 258                 changed = true;
 259         }
 260
 261         r = patch_acls(fd, name, st, shift);
 262         if (r < 0)
 263                 return r;
 264
 265         return r > 0 || changed;
 266 }
 267
 268 /*
 269  * Check if the filesystem is fully compatible with user namespaces or
 270  * UID/GID patching. Some filesystems in this list can be fully mounted inside
 271  * user namespaces, however their inodes may relate to host resources or only
 272  * valid in the global user namespace, therefore no patching should be applied.
 273  */
 274 static int is_fs_fully_userns_compatible(const struct statfs *sfs) {
 275
 276         assert(sfs);
 277
 278         return F_TYPE_EQUAL(sfs->f_type, BINFMTFS_MAGIC) ||
 279                F_TYPE_EQUAL(sfs->f_type, CGROUP_SUPER_MAGIC) ||
 280                F_TYPE_EQUAL(sfs->f_type, CGROUP2_SUPER_MAGIC) ||
 281                F_TYPE_EQUAL(sfs->f_type, DEBUGFS_MAGIC) ||
 282                F_TYPE_EQUAL(sfs->f_type, DEVPTS_SUPER_MAGIC) ||
 283                F_TYPE_EQUAL(sfs->f_type, EFIVARFS_MAGIC) ||
 284                F_TYPE_EQUAL(sfs->f_type, HUGETLBFS_MAGIC) ||
 285                F_TYPE_EQUAL(sfs->f_type, MQUEUE_MAGIC) ||
 286                F_TYPE_EQUAL(sfs->f_type, PROC_SUPER_MAGIC) ||
 287                F_TYPE_EQUAL(sfs->f_type, PSTOREFS_MAGIC) ||
 288                F_TYPE_EQUAL(sfs->f_type, SELINUX_MAGIC) ||
 289                F_TYPE_EQUAL(sfs->f_type, SMACK_MAGIC) ||
 290                F_TYPE_EQUAL(sfs->f_type, SECURITYFS_MAGIC) ||
 291                F_TYPE_EQUAL(sfs->f_type, BPF_FS_MAGIC) ||
 292                F_TYPE_EQUAL(sfs->f_type, TRACEFS_MAGIC) ||
 293                F_TYPE_EQUAL(sfs->f_type, SYSFS_MAGIC);
 294 }
 295
 296 static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift, bool is_toplevel) {
 297         _cleanup_closedir_ DIR *d = NULL;
 298         bool changed = false;
 299         struct statfs sfs;
 300         int r;
 301
 302         assert(fd >= 0);
 303
 304         if (fstatfs(fd, &sfs) < 0)
 305                 return -errno;
 306
 307         /* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
 308          * shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
 309          * when we hit procfs, sysfs or some other special file systems. */
 310
 311         r = is_fs_fully_userns_compatible(&sfs);
 312         if (r < 0)
 313                 goto finish;
 314         if (r > 0) {
 315                 r = 0; /* don't recurse */
 316                 goto finish;
 317         }
 318
 319         /* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
 320         if ((sfs.f_flags & ST_RDONLY) ||
 321             access_fd(fd, W_OK) == -EROFS)
 322                 goto read_only;
 323
 324         if (S_ISDIR(st->st_mode)) {
 325                 struct dirent *de;
 326
 327                 if (!donate_fd) {
 328                         int copy;
 329
 330                         copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
 331                         if (copy < 0) {
 332                                 r = -errno;
 333                                 goto finish;
 334                         }
 335
 336                         fd = copy;
 337                         donate_fd = true;
 338                 }
 339
 340                 d = fdopendir(fd);
 341                 if (!d) {
 342                         r = -errno;
 343                         goto finish;
 344                 }
 345                 fd = -1;
 346
 347                 FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
 348                         struct stat fst;
 349
 350                         if (dot_or_dot_dot(de->d_name))
 351                                 continue;
 352
 353                         if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
 354                                 r = -errno;
 355                                 goto finish;
 356                         }
 357
 358                         if (S_ISDIR(fst.st_mode)) {
 359                                 int subdir_fd;
 360
 361                                 subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 362                                 if (subdir_fd < 0) {
 363                                         r = -errno;
 364                                         goto finish;
 365
 366                                 }
 367
 368                                 r = recurse_fd(subdir_fd, true, &fst, shift, false);
 369                                 if (r < 0)
 370                                         goto finish;
 371                                 if (r > 0)
 372                                         changed = true;
 373
 374                         } else {
 375                                 r = patch_fd(dirfd(d), de->d_name, &fst, shift);
 376                                 if (r < 0)
 377                                         goto finish;
 378                                 if (r > 0)
 379                                         changed = true;
 380                         }
 381                 }
 382         }
 383
 384         /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
 385          * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
 386          * tree is properly chown()ed already. */
 387         r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
 388         if (r == -EROFS)
 389                 goto read_only;
 390         if (r > 0)
 391                 changed = true;
 392
 393         r = changed;
 394         goto finish;
 395
 396 read_only:
 397         if (!is_toplevel) {
 398                 _cleanup_free_ char *name = NULL;
 399
 400                 /* When we hit a ready-only subtree we simply skip it, but log about it. */
 401                 (void) fd_get_path(fd, &name);
 402                 log_debug("Skipping read-only file or directory %s.", strna(name));
 403                 r = changed;
 404         }
 405
 406 finish:
 407         if (donate_fd)
 408                 safe_close(fd);
 409
 410         return r;
 411 }
 412
 413 static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
 414         struct stat st;
 415         int r;
 416
 417         assert(fd >= 0);
 418
 419         /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
 420          * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
 421          * following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
 422          * UID within the container. */
 423
 424         if ((shift & 0xFFFF) != 0) {
 425                 /* We only support containers where the shift starts at a 2^16 boundary */
 426                 r = -EOPNOTSUPP;
 427                 goto finish;
 428         }
 429
 430         if (shift == UID_BUSY_BASE) {
 431                 r = -EINVAL;
 432                 goto finish;
 433         }
 434
 435         if (range != 0x10000) {
 436                 /* We only support containers with 16bit UID ranges for the patching logic */
 437                 r = -EOPNOTSUPP;
 438                 goto finish;
 439         }
 440
 441         if (fstat(fd, &st) < 0) {
 442                 r = -errno;
 443                 goto finish;
 444         }
 445
 446         if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
 447                 /* We only support containers where the uid/gid container ID match */
 448                 r = -EBADE;
 449                 goto finish;
 450         }
 451
 452         /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
 453          * that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
 454         if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
 455                 return 0;
 456
 457         /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
 458          * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
 459          * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
 460
 461         if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
 462                 if (fchown(fd,
 463                            UID_BUSY_BASE | (st.st_uid & ~UID_BUSY_MASK),
 464                            (gid_t) UID_BUSY_BASE | (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
 465                         r = -errno;
 466                         goto finish;
 467                 }
 468         }
 469
 470         return recurse_fd(fd, donate_fd, &st, shift, true);
 471
 472 finish:
 473         if (donate_fd)
 474                 safe_close(fd);
 475
 476         return r;
 477 }
 478
 479 int path_patch_uid(const char *path, uid_t shift, uid_t range) {
 480         int fd;
 481
 482         fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 483         if (fd < 0)
 484                 return -errno;
 485
 486         return fd_patch_uid_internal(fd, true, shift, range);
 487 }