src/nspawn/nspawn-patch-uid.c

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