src/nspawn/nspawn-patch-uid.c

   1 /* SPDX-License-Identifier: LGPL-2.1+ */
   2 /***
   3   This file is part of systemd.
   4
   5   Copyright 2016 Lennart Poettering
   6 ***/
   7
   8 #include <fcntl.h>
   9 #include <linux/magic.h>
  10 #if HAVE_ACL
  11 #include <sys/acl.h>
  12 #endif
  13 #include <sys/stat.h>
  14 #include <sys/statvfs.h>
  15 #include <sys/vfs.h>
  16 #include <unistd.h>
  17
  18 #include "acl-util.h"
  19 #include "dirent-util.h"
  20 #include "fd-util.h"
  21 #include "fs-util.h"
  22 #include "missing.h"
  23 #include "nspawn-def.h"
  24 #include "nspawn-patch-uid.h"
  25 #include "stat-util.h"
  26 #include "stdio-util.h"
  27 #include "string-util.h"
  28 #include "strv.h"
  29 #include "user-util.h"
  30
  31 #if HAVE_ACL
  32
  33 static int get_acl(int fd, const char *name, acl_type_t type, acl_t *ret) {
  34         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  35         acl_t acl;
  36
  37         assert(fd >= 0);
  38         assert(ret);
  39
  40         if (name) {
  41                 _cleanup_close_ int child_fd = -1;
  42
  43                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  44                 if (child_fd < 0)
  45                         return -errno;
  46
  47                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  48                 acl = acl_get_file(procfs_path, type);
  49         } else if (type == ACL_TYPE_ACCESS)
  50                 acl = acl_get_fd(fd);
  51         else {
  52                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  53                 acl = acl_get_file(procfs_path, type);
  54         }
  55         if (!acl)
  56                 return -errno;
  57
  58         *ret = acl;
  59         return 0;
  60 }
  61
  62 static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
  63         char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
  64         int r;
  65
  66         assert(fd >= 0);
  67         assert(acl);
  68
  69         if (name) {
  70                 _cleanup_close_ int child_fd = -1;
  71
  72                 child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
  73                 if (child_fd < 0)
  74                         return -errno;
  75
  76                 xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
  77                 r = acl_set_file(procfs_path, type, acl);
  78         } else if (type == ACL_TYPE_ACCESS)
  79                 r = acl_set_fd(fd, acl);
  80         else {
  81                 xsprintf(procfs_path, "/proc/self/fd/%i", fd);
  82                 r = acl_set_file(procfs_path, type, acl);
  83         }
  84         if (r < 0)
  85                 return -errno;
  86
  87         return 0;
  88 }
  89
  90 static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
  91         _cleanup_(acl_freep) acl_t copy = NULL;
  92         acl_entry_t i;
  93         int r;
  94
  95         assert(acl);
  96         assert(ret);
  97
  98         r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
  99         if (r < 0)
 100                 return -errno;
 101         while (r > 0) {
 102                 uid_t *old_uid, new_uid;
 103                 bool modify = false;
 104                 acl_tag_t tag;
 105
 106                 if (acl_get_tag_type(i, &tag) < 0)
 107                         return -errno;
 108
 109                 if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
 110
 111                         /* We don't distuingish here between uid_t and gid_t, let's make sure the compiler checks that
 112                          * this is actually OK */
 113                         assert_cc(sizeof(uid_t) == sizeof(gid_t));
 114
 115                         old_uid = acl_get_qualifier(i);
 116                         if (!old_uid)
 117                                 return -errno;
 118
 119                         new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
 120                         if (!uid_is_valid(new_uid))
 121                                 return -EINVAL;
 122
 123                         modify = new_uid != *old_uid;
 124                         if (modify && !copy) {
 125                                 int n;
 126
 127                                 /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
 128                                  * beginning, so that we copy all entries, starting from the first, this time. */
 129
 130                                 n = acl_entries(acl);
 131                                 if (n < 0)
 132                                         return -errno;
 133
 134                                 copy = acl_init(n);
 135                                 if (!copy)
 136                                         return -errno;
 137
 138                                 /* Seek back to the beginning */
 139                                 r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
 140                                 if (r < 0)
 141                                         return -errno;
 142                                 continue;
 143                         }
 144                 }
 145
 146                 if (copy) {
 147                         acl_entry_t new_entry;
 148
 149                         if (acl_create_entry(&copy, &new_entry) < 0)
 150                                 return -errno;
 151
 152                         if (acl_copy_entry(new_entry, i) < 0)
 153                                 return -errno;
 154
 155                         if (modify)
 156                                 if (acl_set_qualifier(new_entry, &new_uid) < 0)
 157                                         return -errno;
 158                 }
 159
 160                 r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
 161                 if (r < 0)
 162                         return -errno;
 163         }
 164
 165         *ret = TAKE_PTR(copy);
 166
 167         return !!*ret;
 168 }
 169
 170 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 171         _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
 172         bool changed = false;
 173         int r;
 174
 175         assert(fd >= 0);
 176         assert(st);
 177
 178         /* ACLs are not supported on symlinks, there's no point in trying */
 179         if (S_ISLNK(st->st_mode))
 180                 return 0;
 181
 182         r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
 183         if (r == -EOPNOTSUPP)
 184                 return 0;
 185         if (r < 0)
 186                 return r;
 187
 188         r = shift_acl(acl, shift, &shifted);
 189         if (r < 0)
 190                 return r;
 191         if (r > 0) {
 192                 r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
 193                 if (r < 0)
 194                         return r;
 195
 196                 changed = true;
 197         }
 198
 199         if (S_ISDIR(st->st_mode)) {
 200                 acl_free(acl);
 201                 acl_free(shifted);
 202
 203                 acl = shifted = NULL;
 204
 205                 r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
 206                 if (r < 0)
 207                         return r;
 208
 209                 r = shift_acl(acl, shift, &shifted);
 210                 if (r < 0)
 211                         return r;
 212                 if (r > 0) {
 213                         r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
 214                         if (r < 0)
 215                                 return r;
 216
 217                         changed = true;
 218                 }
 219         }
 220
 221         return changed;
 222 }
 223
 224 #else
 225
 226 static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
 227         return 0;
 228 }
 229
 230 #endif
 231
 232 static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
 233         uid_t new_uid;
 234         gid_t new_gid;
 235         bool changed = false;
 236         int r;
 237
 238         assert(fd >= 0);
 239         assert(st);
 240
 241         new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
 242         new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));
 243
 244         if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
 245                 return -EINVAL;
 246
 247         if (st->st_uid != new_uid || st->st_gid != new_gid) {
 248                 if (name)
 249                         r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
 250                 else
 251                         r = fchown(fd, new_uid, new_gid);
 252                 if (r < 0)
 253                         return -errno;
 254
 255                 /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
 256                 if (name) {
 257                         if (!S_ISLNK(st->st_mode))
 258                                 r = fchmodat(fd, name, st->st_mode, 0);
 259                         else /* AT_SYMLINK_NOFOLLOW is not available for fchmodat() */
 260                                 r = 0;
 261                 } else
 262                         r = fchmod(fd, st->st_mode);
 263                 if (r < 0)
 264                         return -errno;
 265
 266                 changed = true;
 267         }
 268
 269         r = patch_acls(fd, name, st, shift);
 270         if (r < 0)
 271                 return r;
 272
 273         return r > 0 || changed;
 274 }
 275
 276 /*
 277  * Check if the filesystem is fully compatible with user namespaces or
 278  * UID/GID patching. Some filesystems in this list can be fully mounted inside
 279  * user namespaces, however their inodes may relate to host resources or only
 280  * valid in the global user namespace, therefore no patching should be applied.
 281  */
 282 static int is_fs_fully_userns_compatible(const struct statfs *sfs) {
 283
 284         assert(sfs);
 285
 286         return F_TYPE_EQUAL(sfs->f_type, BINFMTFS_MAGIC) ||
 287                F_TYPE_EQUAL(sfs->f_type, CGROUP_SUPER_MAGIC) ||
 288                F_TYPE_EQUAL(sfs->f_type, CGROUP2_SUPER_MAGIC) ||
 289                F_TYPE_EQUAL(sfs->f_type, DEBUGFS_MAGIC) ||
 290                F_TYPE_EQUAL(sfs->f_type, DEVPTS_SUPER_MAGIC) ||
 291                F_TYPE_EQUAL(sfs->f_type, EFIVARFS_MAGIC) ||
 292                F_TYPE_EQUAL(sfs->f_type, HUGETLBFS_MAGIC) ||
 293                F_TYPE_EQUAL(sfs->f_type, MQUEUE_MAGIC) ||
 294                F_TYPE_EQUAL(sfs->f_type, PROC_SUPER_MAGIC) ||
 295                F_TYPE_EQUAL(sfs->f_type, PSTOREFS_MAGIC) ||
 296                F_TYPE_EQUAL(sfs->f_type, SELINUX_MAGIC) ||
 297                F_TYPE_EQUAL(sfs->f_type, SMACK_MAGIC) ||
 298                F_TYPE_EQUAL(sfs->f_type, SECURITYFS_MAGIC) ||
 299                F_TYPE_EQUAL(sfs->f_type, BPF_FS_MAGIC) ||
 300                F_TYPE_EQUAL(sfs->f_type, TRACEFS_MAGIC) ||
 301                F_TYPE_EQUAL(sfs->f_type, SYSFS_MAGIC);
 302 }
 303
 304 static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift, bool is_toplevel) {
 305         _cleanup_closedir_ DIR *d = NULL;
 306         bool changed = false;
 307         struct statfs sfs;
 308         int r;
 309
 310         assert(fd >= 0);
 311
 312         if (fstatfs(fd, &sfs) < 0)
 313                 return -errno;
 314
 315         /* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
 316          * shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
 317          * when we hit procfs, sysfs or some other special file systems. */
 318
 319         r = is_fs_fully_userns_compatible(&sfs);
 320         if (r < 0)
 321                 goto finish;
 322         if (r > 0) {
 323                 r = 0; /* don't recurse */
 324                 goto finish;
 325         }
 326
 327         /* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
 328         if ((sfs.f_flags & ST_RDONLY) ||
 329             access_fd(fd, W_OK) == -EROFS)
 330                 goto read_only;
 331
 332         if (S_ISDIR(st->st_mode)) {
 333                 struct dirent *de;
 334
 335                 if (!donate_fd) {
 336                         int copy;
 337
 338                         copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
 339                         if (copy < 0) {
 340                                 r = -errno;
 341                                 goto finish;
 342                         }
 343
 344                         fd = copy;
 345                         donate_fd = true;
 346                 }
 347
 348                 d = fdopendir(fd);
 349                 if (!d) {
 350                         r = -errno;
 351                         goto finish;
 352                 }
 353                 fd = -1;
 354
 355                 FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
 356                         struct stat fst;
 357
 358                         if (dot_or_dot_dot(de->d_name))
 359                                 continue;
 360
 361                         if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
 362                                 r = -errno;
 363                                 goto finish;
 364                         }
 365
 366                         if (S_ISDIR(fst.st_mode)) {
 367                                 int subdir_fd;
 368
 369                                 subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 370                                 if (subdir_fd < 0) {
 371                                         r = -errno;
 372                                         goto finish;
 373
 374                                 }
 375
 376                                 r = recurse_fd(subdir_fd, true, &fst, shift, false);
 377                                 if (r < 0)
 378                                         goto finish;
 379                                 if (r > 0)
 380                                         changed = true;
 381
 382                         } else {
 383                                 r = patch_fd(dirfd(d), de->d_name, &fst, shift);
 384                                 if (r < 0)
 385                                         goto finish;
 386                                 if (r > 0)
 387                                         changed = true;
 388                         }
 389                 }
 390         }
 391
 392         /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
 393          * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
 394          * tree is properly chown()ed already. */
 395         r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
 396         if (r == -EROFS)
 397                 goto read_only;
 398         if (r > 0)
 399                 changed = true;
 400
 401         r = changed;
 402         goto finish;
 403
 404 read_only:
 405         if (!is_toplevel) {
 406                 _cleanup_free_ char *name = NULL;
 407
 408                 /* When we hit a ready-only subtree we simply skip it, but log about it. */
 409                 (void) fd_get_path(fd, &name);
 410                 log_debug("Skippping read-only file or directory %s.", strna(name));
 411                 r = changed;
 412         }
 413
 414 finish:
 415         if (donate_fd)
 416                 safe_close(fd);
 417
 418         return r;
 419 }
 420
 421 static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
 422         struct stat st;
 423         int r;
 424
 425         assert(fd >= 0);
 426
 427         /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
 428          * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
 429          * following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
 430          * UID within the container. */
 431
 432         if ((shift & 0xFFFF) != 0) {
 433                 /* We only support containers where the shift starts at a 2^16 boundary */
 434                 r = -EOPNOTSUPP;
 435                 goto finish;
 436         }
 437
 438         if (shift == UID_BUSY_BASE) {
 439                 r = -EINVAL;
 440                 goto finish;
 441         }
 442
 443         if (range != 0x10000) {
 444                 /* We only support containers with 16bit UID ranges for the patching logic */
 445                 r = -EOPNOTSUPP;
 446                 goto finish;
 447         }
 448
 449         if (fstat(fd, &st) < 0) {
 450                 r = -errno;
 451                 goto finish;
 452         }
 453
 454         if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
 455                 /* We only support containers where the uid/gid container ID match */
 456                 r = -EBADE;
 457                 goto finish;
 458         }
 459
 460         /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
 461          * that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
 462         if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
 463                 return 0;
 464
 465         /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
 466          * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
 467          * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
 468
 469         if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
 470                 if (fchown(fd,
 471                            UID_BUSY_BASE | (st.st_uid & ~UID_BUSY_MASK),
 472                            (gid_t) UID_BUSY_BASE | (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
 473                         r = -errno;
 474                         goto finish;
 475                 }
 476         }
 477
 478         return recurse_fd(fd, donate_fd, &st, shift, true);
 479
 480 finish:
 481         if (donate_fd)
 482                 safe_close(fd);
 483
 484         return r;
 485 }
 486
 487 int fd_patch_uid(int fd, uid_t shift, uid_t range) {
 488         return fd_patch_uid_internal(fd, false, shift, range);
 489 }
 490
 491 int path_patch_uid(const char *path, uid_t shift, uid_t range) {
 492         int fd;
 493
 494         fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
 495         if (fd < 0)
 496                 return -errno;
 497
 498         return fd_patch_uid_internal(fd, true, shift, range);
 499 }