src/nspawn/nspawn-patch-uid.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2 /***
   3   Copyright 2016 Lennart Poettering
   4 ***/
   5
   6 #include <fcntl.h>
   7 #include <linux/magic.h>
   8 #if HAVE_ACL
   9 #include <sys/acl.h>
  10 #endif
  11 #include <sys/stat.h>
  12 #include <sys/statvfs.h>
  13 #include <sys/vfs.h>
  14 #include <unistd.h>
  15
  16 #include "acl-util.h"
  17 #include "dirent-util.h"
  18 #include "fd-util.h"
  19 #include "fs-util.h"
  20 #include "missing.h"
  21 #include "nspawn-def.h"
  22 #include "nspawn-patch-uid.h"
  23 #include "stat-util.h"
  24 #include "stdio-util.h"
  25 #include "string-util.h"
  26 #include "strv.h"
  27 #include "user-util.h"
  28
  29 #if HAVE_ACL
  30
  31 static int get_acl(int fd, const char *name, acl_type_t type, acl_t *ret) {
  32         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  33         acl_t acl;
  34
  35         assert(fd >= 0);
  36         assert(ret);
  37
  38         if (name) {
  39                 _cleanup_close_ int child_fd = -1;
  40
  41                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  42                 if (child_fd < 0)
  43                         return -errno;
  44
  45                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  46                 acl = acl_get_file(procfs_path, type);
  47         } else if (type == ACL_TYPE_ACCESS)
  48                 acl = acl_get_fd(fd);
  49         else {
  50                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  51                 acl = acl_get_file(procfs_path, type);
  52         }
  53         if (!acl)
  54                 return -errno;
  55
  56         *ret = acl;
  57         return 0;
  58 }
  59
  60 static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
  61         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  62         int r;
  63
  64         assert(fd >= 0);
  65         assert(acl);
  66
  67         if (name) {
  68                 _cleanup_close_ int child_fd = -1;
  69
  70                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  71                 if (child_fd < 0)
  72                         return -errno;
  73
  74                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  75                 r = acl_set_file(procfs_path, type, acl);
  76         } else if (type == ACL_TYPE_ACCESS)
  77                 r = acl_set_fd(fd, acl);
  78         else {
  79                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  80                 r = acl_set_file(procfs_path, type, acl);
  81         }
  82         if (r < 0)
  83                 return -errno;
  84
  85         return 0;
  86 }
  87
  88 static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
  89         _cleanup_(acl_freep) acl_t copy = NULL;
  90         acl_entry_t i;
  91         int r;
  92
  93         assert(acl);
  94         assert(ret);
  95
  96         r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
  97         if (r < 0)
  98                 return -errno;
  99         while (r > 0) {
 100                 uid_t *old_uid, new_uid;
 101                 bool modify = false;
 102                 acl_tag_t tag;
 103
 104                 if (acl_get_tag_type(i, &tag) < 0)
 105                         return -errno;
 106
 107                 if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
 108
 109                         /* We don't distuingish here between uid_t and gid_t, let's make sure the compiler checks that
 110                          * this is actually OK */
 111                         assert_cc(sizeof(uid_t) == sizeof(gid_t));
 112
 113                         old_uid = acl_get_qualifier(i);
 114                         if (!old_uid)
 115                                 return -errno;
 116
 117                         new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
 118                         if (!uid_is_valid(new_uid))
 119                                 return -EINVAL;
 120
 121                         modify = new_uid != *old_uid;
 122                         if (modify && !copy) {
 123                                 int n;
 124
 125                                 /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
 126                                  * beginning, so that we copy all entries, starting from the first, this time. */
 127
 128                                 n = acl_entries(acl);
 129                                 if (n < 0)
 130                                         return -errno;
 131
 132                                 copy = acl_init(n);
 133                                 if (!copy)
 134                                         return -errno;
 135
 136                                 /* Seek back to the beginning */
 137                                 r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
 138                                 if (r < 0)
 139                                         return -errno;
 140                                 continue;
 141                         }
 142                 }
 143
 144                 if (copy) {
 145                         acl_entry_t new_entry;
 146
 147                         if (acl_create_entry(&copy, &new_entry) < 0)
 148                                 return -errno;
 149
 150                         if (acl_copy_entry(new_entry, i) < 0)
 151                                 return -errno;
 152
 153                         if (modify)
 154                                 if (acl_set_qualifier(new_entry, &new_uid) < 0)
 155                                         return -errno;
 156                 }
 157
 158                 r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
 159                 if (r < 0)
 160                         return -errno;
 161         }
 162
 163         *ret = TAKE_PTR(copy);
 164
 165         return !!*ret;
 166 }
 167
 168 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 169         _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
 170         bool changed = false;
 171         int r;
 172
 173         assert(fd >= 0);
 174         assert(st);
 175
 176         /* ACLs are not supported on symlinks, there's no point in trying */
 177         if (S_ISLNK(st->st_mode))
 178                 return 0;
 179
 180         r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
 181         if (r == -EOPNOTSUPP)
 182                 return 0;
 183         if (r < 0)
 184                 return r;
 185
 186         r = shift_acl(acl, shift, &shifted);
 187         if (r < 0)
 188                 return r;
 189         if (r > 0) {
 190                 r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
 191                 if (r < 0)
 192                         return r;
 193
 194                 changed = true;
 195         }
 196
 197         if (S_ISDIR(st->st_mode)) {
 198                 acl_free(acl);
 199                 acl_free(shifted);
 200
 201                 acl = shifted = NULL;
 202
 203                 r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
 204                 if (r < 0)
 205                         return r;
 206
 207                 r = shift_acl(acl, shift, &shifted);
 208                 if (r < 0)
 209                         return r;
 210                 if (r > 0) {
 211                         r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
 212                         if (r < 0)
 213                                 return r;
 214
 215                         changed = true;
 216                 }
 217         }
 218
 219         return changed;
 220 }
 221
 222 #else
 223
 224 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 225         return 0;
 226 }
 227
 228 #endif
 229
 230 static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
 231         uid_t new_uid;
 232         gid_t new_gid;
 233         bool changed = false;
 234         int r;
 235
 236         assert(fd >= 0);
 237         assert(st);
 238
 239         new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
 240         new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));
 241
 242         if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
 243                 return -EINVAL;
 244
 245         if (st->st_uid != new_uid || st->st_gid != new_gid) {
 246                 if (name)
 247                         r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
 248                 else
 249                         r = fchown(fd, new_uid, new_gid);
 250                 if (r < 0)
 251                         return -errno;
 252
 253                 /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
 254                 if (name) {
 255                         if (!S_ISLNK(st->st_mode))
 256                                 r = fchmodat(fd, name, st->st_mode, 0);
 257                         else /* AT_SYMLINK_NOFOLLOW is not available for fchmodat() */
 258                                 r = 0;
 259                 } else
 260                         r = fchmod(fd, st->st_mode);
 261                 if (r < 0)
 262                         return -errno;
 263
 264                 changed = true;
 265         }
 266
 267         r = patch_acls(fd, name, st, shift);
 268         if (r < 0)
 269                 return r;
 270
 271         return r > 0 || changed;
 272 }
 273
 274 /*
 275  * Check if the filesystem is fully compatible with user namespaces or
 276  * UID/GID patching. Some filesystems in this list can be fully mounted inside
 277  * user namespaces, however their inodes may relate to host resources or only
 278  * valid in the global user namespace, therefore no patching should be applied.
 279  */
 280 static int is_fs_fully_userns_compatible(const struct statfs *sfs) {
 281
 282         assert(sfs);
 283
 284         return F_TYPE_EQUAL(sfs->f_type, BINFMTFS_MAGIC) ||
 285                F_TYPE_EQUAL(sfs->f_type, CGROUP_SUPER_MAGIC) ||
 286                F_TYPE_EQUAL(sfs->f_type, CGROUP2_SUPER_MAGIC) ||
 287                F_TYPE_EQUAL(sfs->f_type, DEBUGFS_MAGIC) ||
 288                F_TYPE_EQUAL(sfs->f_type, DEVPTS_SUPER_MAGIC) ||
 289                F_TYPE_EQUAL(sfs->f_type, EFIVARFS_MAGIC) ||
 290                F_TYPE_EQUAL(sfs->f_type, HUGETLBFS_MAGIC) ||
 291                F_TYPE_EQUAL(sfs->f_type, MQUEUE_MAGIC) ||
 292                F_TYPE_EQUAL(sfs->f_type, PROC_SUPER_MAGIC) ||
 293                F_TYPE_EQUAL(sfs->f_type, PSTOREFS_MAGIC) ||
 294                F_TYPE_EQUAL(sfs->f_type, SELINUX_MAGIC) ||
 295                F_TYPE_EQUAL(sfs->f_type, SMACK_MAGIC) ||
 296                F_TYPE_EQUAL(sfs->f_type, SECURITYFS_MAGIC) ||
 297                F_TYPE_EQUAL(sfs->f_type, BPF_FS_MAGIC) ||
 298                F_TYPE_EQUAL(sfs->f_type, TRACEFS_MAGIC) ||
 299                F_TYPE_EQUAL(sfs->f_type, SYSFS_MAGIC);
 300 }
 301
 302 static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift, bool is_toplevel) {
 303         _cleanup_closedir_ DIR *d = NULL;
 304         bool changed = false;
 305         struct statfs sfs;
 306         int r;
 307
 308         assert(fd >= 0);
 309
 310         if (fstatfs(fd, &sfs) < 0)
 311                 return -errno;
 312
 313         /* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
 314          * shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
 315          * when we hit procfs, sysfs or some other special file systems. */
 316
 317         r = is_fs_fully_userns_compatible(&sfs);
 318         if (r < 0)
 319                 goto finish;
 320         if (r > 0) {
 321                 r = 0; /* don't recurse */
 322                 goto finish;
 323         }
 324
 325         /* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
 326         if ((sfs.f_flags & ST_RDONLY) ||
 327             access_fd(fd, W_OK) == -EROFS)
 328                 goto read_only;
 329
 330         if (S_ISDIR(st->st_mode)) {
 331                 struct dirent *de;
 332
 333                 if (!donate_fd) {
 334                         int copy;
 335
 336                         copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
 337                         if (copy < 0) {
 338                                 r = -errno;
 339                                 goto finish;
 340                         }
 341
 342                         fd = copy;
 343                         donate_fd = true;
 344                 }
 345
 346                 d = fdopendir(fd);
 347                 if (!d) {
 348                         r = -errno;
 349                         goto finish;
 350                 }
 351                 fd = -1;
 352
 353                 FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
 354                         struct stat fst;
 355
 356                         if (dot_or_dot_dot(de->d_name))
 357                                 continue;
 358
 359                         if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
 360                                 r = -errno;
 361                                 goto finish;
 362                         }
 363
 364                         if (S_ISDIR(fst.st_mode)) {
 365                                 int subdir_fd;
 366
 367                                 subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 368                                 if (subdir_fd < 0) {
 369                                         r = -errno;
 370                                         goto finish;
 371
 372                                 }
 373
 374                                 r = recurse_fd(subdir_fd, true, &fst, shift, false);
 375                                 if (r < 0)
 376                                         goto finish;
 377                                 if (r > 0)
 378                                         changed = true;
 379
 380                         } else {
 381                                 r = patch_fd(dirfd(d), de->d_name, &fst, shift);
 382                                 if (r < 0)
 383                                         goto finish;
 384                                 if (r > 0)
 385                                         changed = true;
 386                         }
 387                 }
 388         }
 389
 390         /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
 391          * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
 392          * tree is properly chown()ed already. */
 393         r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
 394         if (r == -EROFS)
 395                 goto read_only;
 396         if (r > 0)
 397                 changed = true;
 398
 399         r = changed;
 400         goto finish;
 401
 402 read_only:
 403         if (!is_toplevel) {
 404                 _cleanup_free_ char *name = NULL;
 405
 406                 /* When we hit a ready-only subtree we simply skip it, but log about it. */
 407                 (void) fd_get_path(fd, &name);
 408                 log_debug("Skippping read-only file or directory %s.", strna(name));
 409                 r = changed;
 410         }
 411
 412 finish:
 413         if (donate_fd)
 414                 safe_close(fd);
 415
 416         return r;
 417 }
 418
 419 static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
 420         struct stat st;
 421         int r;
 422
 423         assert(fd >= 0);
 424
 425         /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
 426          * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
 427          * following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
 428          * UID within the container. */
 429
 430         if ((shift & 0xFFFF) != 0) {
 431                 /* We only support containers where the shift starts at a 2^16 boundary */
 432                 r = -EOPNOTSUPP;
 433                 goto finish;
 434         }
 435
 436         if (shift == UID_BUSY_BASE) {
 437                 r = -EINVAL;
 438                 goto finish;
 439         }
 440
 441         if (range != 0x10000) {
 442                 /* We only support containers with 16bit UID ranges for the patching logic */
 443                 r = -EOPNOTSUPP;
 444                 goto finish;
 445         }
 446
 447         if (fstat(fd, &st) < 0) {
 448                 r = -errno;
 449                 goto finish;
 450         }
 451
 452         if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
 453                 /* We only support containers where the uid/gid container ID match */
 454                 r = -EBADE;
 455                 goto finish;
 456         }
 457
 458         /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
 459          * that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
 460         if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
 461                 return 0;
 462
 463         /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
 464          * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
 465          * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
 466
 467         if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
 468                 if (fchown(fd,
 469                            UID_BUSY_BASE | (st.st_uid & ~UID_BUSY_MASK),
 470                            (gid_t) UID_BUSY_BASE | (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
 471                         r = -errno;
 472                         goto finish;
 473                 }
 474         }
 475
 476         return recurse_fd(fd, donate_fd, &st, shift, true);
 477
 478 finish:
 479         if (donate_fd)
 480                 safe_close(fd);
 481
 482         return r;
 483 }
 484
 485 int fd_patch_uid(int fd, uid_t shift, uid_t range) {
 486         return fd_patch_uid_internal(fd, false, shift, range);
 487 }
 488
 489 int path_patch_uid(const char *path, uid_t shift, uid_t range) {
 490         int fd;
 491
 492         fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 493         if (fd < 0)
 494                 return -errno;
 495
 496         return fd_patch_uid_internal(fd, true, shift, range);
 497 }