| 1 |
/* SPDX-License-Identifier: LGPL-2.1-or-later */ |
| 2 |
|
| 3 |
#include <sys/ioctl.h> |
| 4 |
#include <sys/stat.h> |
| 5 |
#include <unistd.h> |
| 6 |
|
| 7 |
#include "btrfs-util.h" |
| 8 |
#include "chattr-util.h" |
| 9 |
#include "env-util.h" |
| 10 |
#include "errno-util.h" |
| 11 |
#include "fd-util.h" |
| 12 |
#include "fs-util.h" |
| 13 |
#include "install-file.h" |
| 14 |
#include "log.h" |
| 15 |
#include "rm-rf.h" |
| 16 |
#include "sync-util.h" |
| 17 |
#include "time-util.h" |
| 18 |
|
| 19 |
static int fs_make_very_read_only(int fd) { |
| 20 |
struct stat st; |
| 21 |
int r; |
| 22 |
|
| 23 |
assert(fd >= 0); |
| 24 |
|
| 25 |
/* Tries to make the specified fd "comprehensively" read-only. Primary use case for this is OS images, |
| 26 |
* i.e. either loopback files or larger directory hierarchies. Depending on the inode type and |
| 27 |
* backing file system this means something different: |
| 28 |
* |
| 29 |
* 1. If the fd refers to a btrfs subvolume we'll mark it read-only as a whole |
| 30 |
* 2. If the fd refers to any other directory we'll set the FS_IMMUTABLE_FL flag on it |
| 31 |
* 3. If the fd refers to a regular file we'll drop the w bits. |
| 32 |
* 4. If the fd refers to a block device, use BLKROSET to set read-only state |
| 33 |
* |
| 34 |
* You might wonder why not drop the x bits for directories. That's because we want to guarantee that |
| 35 |
* everything "inside" the image remains largely the way it is, in case you mount it. And since the |
| 36 |
* mode of the root dir of the image is pretty visible we don't want to modify it. btrfs subvol flags |
| 37 |
* and the FS_IMMUTABLE_FL otoh are much less visible. Changing the mode of regular files should be |
| 38 |
* OK though, since after all this is supposed to be used for disk images, i.e. the fs in the disk |
| 39 |
* image doesn't make the mode of the loopback file it is stored in visible. */ |
| 40 |
|
| 41 |
if (fstat(fd, &st) < 0) |
| 42 |
return -errno; |
| 43 |
|
| 44 |
switch (st.st_mode & S_IFMT) { |
| 45 |
|
| 46 |
case S_IFDIR: |
| 47 |
if (btrfs_might_be_subvol(&st)) { |
| 48 |
r = btrfs_subvol_set_read_only_fd(fd, true); |
| 49 |
if (r >= 0) |
| 50 |
return 0; |
| 51 |
if (!ERRNO_IS_NEG_IOCTL_NOT_SUPPORTED(r)) |
| 52 |
return r; |
| 53 |
} |
| 54 |
|
| 55 |
r = chattr_fd(fd, FS_IMMUTABLE_FL, FS_IMMUTABLE_FL); |
| 56 |
if (r < 0) |
| 57 |
return r; |
| 58 |
|
| 59 |
break; |
| 60 |
|
| 61 |
case S_IFREG: |
| 62 |
if ((st.st_mode & 0222) != 0) |
| 63 |
if (fchmod(fd, st.st_mode & 07555) < 0) |
| 64 |
return -errno; |
| 65 |
|
| 66 |
break; |
| 67 |
|
| 68 |
case S_IFBLK: { |
| 69 |
int ro = 1; |
| 70 |
|
| 71 |
if (ioctl(fd, BLKROSET, &ro) < 0) |
| 72 |
return -errno; |
| 73 |
|
| 74 |
break; |
| 75 |
} |
| 76 |
|
| 77 |
default: |
| 78 |
return -EBADFD; |
| 79 |
} |
| 80 |
|
| 81 |
return 0; |
| 82 |
} |
| 83 |
|
| 84 |
static int unlinkat_maybe_dir(int dirfd, const char *pathname) { |
| 85 |
|
| 86 |
/* Invokes unlinkat() for regular files first, and if this fails with EISDIR tries again with |
| 87 |
* AT_REMOVEDIR */ |
| 88 |
|
| 89 |
if (unlinkat(dirfd, pathname, 0) < 0) { |
| 90 |
if (errno != EISDIR) |
| 91 |
return -errno; |
| 92 |
|
| 93 |
if (unlinkat(dirfd, pathname, AT_REMOVEDIR) < 0) |
| 94 |
return -errno; |
| 95 |
} |
| 96 |
|
| 97 |
return 0; |
| 98 |
} |
| 99 |
|
| 100 |
static bool need_opath(InstallFileFlags flags) { |
| 101 |
/* Returns true when we need to pin the source file via an O_PATH fd */ |
| 102 |
return (flags & (INSTALL_FSYNC|INSTALL_FSYNC_FULL|INSTALL_SYNCFS|INSTALL_READ_ONLY)) != 0; |
| 103 |
} |
| 104 |
|
| 105 |
int install_file(int source_atfd, const char *source_name, |
| 106 |
int target_atfd, const char *target_name, |
| 107 |
InstallFileFlags flags) { |
| 108 |
|
| 109 |
_cleanup_close_ int rofd = -EBADF; |
| 110 |
int r; |
| 111 |
|
| 112 |
/* Moves a file or directory tree into place, with some bells and whistles: |
| 113 |
* |
| 114 |
* 1. Optionally syncs before/after to ensure file installation can be used as barrier |
| 115 |
* 2. Optionally marks the file/directory read-only using fs_make_very_read_only() |
| 116 |
* 3. Optionally operates in replacing or in non-replacing mode. |
| 117 |
* 4. If it replaces will remove the old tree if needed. |
| 118 |
*/ |
| 119 |
|
| 120 |
assert(source_atfd >= 0 || source_atfd == AT_FDCWD); |
| 121 |
assert(source_name); |
| 122 |
assert(target_atfd >= 0 || target_atfd == AT_FDCWD); |
| 123 |
|
| 124 |
/* If target_name is specified as NULL no renaming takes place. Instead it is assumed the file is |
| 125 |
* already in place, and only the syncing/read-only marking shall be applied. Note that with |
| 126 |
* target_name=NULL and flags=0 this call is a NOP */ |
| 127 |
|
| 128 |
if (need_opath(flags)) { |
| 129 |
_cleanup_close_ int pfd = -EBADF; |
| 130 |
struct stat st; |
| 131 |
|
| 132 |
/* Open an O_PATH fd for the source if we need to sync things or mark things read only. */ |
| 133 |
|
| 134 |
pfd = openat(source_atfd, source_name, O_PATH|O_CLOEXEC|O_NOFOLLOW); |
| 135 |
if (pfd < 0) |
| 136 |
return -errno; |
| 137 |
|
| 138 |
if (fstat(pfd, &st) < 0) |
| 139 |
return -errno; |
| 140 |
|
| 141 |
switch (st.st_mode & S_IFMT) { |
| 142 |
|
| 143 |
case S_IFREG: { |
| 144 |
_cleanup_close_ int regfd = -EBADF; |
| 145 |
|
| 146 |
regfd = fd_reopen(pfd, O_RDONLY|O_CLOEXEC); |
| 147 |
if (regfd < 0) { |
| 148 |
if (!FLAGS_SET(flags, INSTALL_GRACEFUL)) |
| 149 |
return log_debug_errno(regfd, "Failed to open referenced inode: %m"); |
| 150 |
|
| 151 |
log_debug_errno(regfd, "Failed to open referenced inode, ignoring: %m"); |
| 152 |
} else { |
| 153 |
if ((flags & (INSTALL_FSYNC_FULL|INSTALL_SYNCFS)) != 0) |
| 154 |
/* If this is just a regular file (as opposed to a fully populated |
| 155 |
* directory) let's downgrade INSTALL_SYNCFS to INSTALL_FSYNC_FULL, |
| 156 |
* after all this is going to be a single inode we install */ |
| 157 |
r = fsync_full(regfd); |
| 158 |
else if (flags & INSTALL_FSYNC) |
| 159 |
r = RET_NERRNO(fsync(regfd)); |
| 160 |
else |
| 161 |
r = 0; |
| 162 |
if (r < 0) { |
| 163 |
if (!FLAGS_SET(flags, INSTALL_GRACEFUL)) |
| 164 |
return log_debug_errno(r, "Failed to sync source inode: %m"); |
| 165 |
|
| 166 |
log_debug_errno(r, "Failed to sync source inode, ignoring: %m"); |
| 167 |
} |
| 168 |
|
| 169 |
if (flags & INSTALL_READ_ONLY) |
| 170 |
rofd = TAKE_FD(regfd); |
| 171 |
} |
| 172 |
|
| 173 |
break; |
| 174 |
} |
| 175 |
|
| 176 |
case S_IFDIR: { |
| 177 |
_cleanup_close_ int dfd = -EBADF; |
| 178 |
|
| 179 |
dfd = fd_reopen(pfd, O_RDONLY|O_DIRECTORY|O_CLOEXEC); |
| 180 |
if (dfd < 0) { |
| 181 |
if (!FLAGS_SET(flags, INSTALL_GRACEFUL)) |
| 182 |
return log_debug_errno(dfd, "Failed to open referenced inode: %m"); |
| 183 |
|
| 184 |
log_debug_errno(dfd, "Failed to open referenced inode, ignoring: %m"); |
| 185 |
} else { |
| 186 |
if (flags & INSTALL_SYNCFS) |
| 187 |
r = RET_NERRNO(syncfs(dfd)); |
| 188 |
else if (flags & INSTALL_FSYNC_FULL) |
| 189 |
r = fsync_full(dfd); |
| 190 |
else if (flags & INSTALL_FSYNC) |
| 191 |
r = RET_NERRNO(fsync(dfd)); |
| 192 |
else |
| 193 |
r = 0; |
| 194 |
if (r < 0) { |
| 195 |
if (!FLAGS_SET(flags, INSTALL_GRACEFUL)) |
| 196 |
return log_debug_errno(r, "Failed to sync source inode: %m"); |
| 197 |
|
| 198 |
log_debug_errno(r, "Failed to sync source inode, ignoring: %m"); |
| 199 |
} |
| 200 |
|
| 201 |
if (flags & INSTALL_READ_ONLY) |
| 202 |
rofd = TAKE_FD(dfd); |
| 203 |
} |
| 204 |
|
| 205 |
break; |
| 206 |
} |
| 207 |
|
| 208 |
default: |
| 209 |
/* Other inodes: char/block device inodes, fifos, symlinks, sockets don't need |
| 210 |
* syncing themselves, as they only exist in the directory, and have no contents on |
| 211 |
* disk */ |
| 212 |
|
| 213 |
if (target_name && (flags & (INSTALL_FSYNC_FULL|INSTALL_SYNCFS)) != 0) { |
| 214 |
r = fsync_directory_of_file(pfd); |
| 215 |
if (r < 0) { |
| 216 |
if (!FLAGS_SET(flags, INSTALL_GRACEFUL)) |
| 217 |
return log_debug_errno(r, "Failed to sync source inode: %m"); |
| 218 |
|
| 219 |
log_debug_errno(r, "Failed to sync source inode, ignoring: %m"); |
| 220 |
} |
| 221 |
} |
| 222 |
|
| 223 |
break; |
| 224 |
} |
| 225 |
} |
| 226 |
|
| 227 |
if (target_name) { |
| 228 |
/* Rename the file */ |
| 229 |
|
| 230 |
if (flags & INSTALL_REPLACE) { |
| 231 |
/* First, try a simple renamat(), maybe that's enough */ |
| 232 |
if (renameat(source_atfd, source_name, target_atfd, target_name) < 0) { |
| 233 |
_cleanup_close_ int dfd = -EBADF; |
| 234 |
|
| 235 |
if (!IN_SET(errno, EEXIST, ENOTDIR, ENOTEMPTY, EISDIR, EBUSY)) |
| 236 |
return -errno; |
| 237 |
|
| 238 |
/* Hmm, the target apparently existed already. Let's try to use |
| 239 |
* RENAME_EXCHANGE. But let's first open the inode if it's a directory, so |
| 240 |
* that we can later remove its contents if it's a directory. Why do this |
| 241 |
* before the rename()? Mostly because if we have trouble opening the thing |
| 242 |
* we want to know before we start actually modifying the file system. */ |
| 243 |
|
| 244 |
dfd = openat(target_atfd, target_name, O_RDONLY|O_DIRECTORY|O_CLOEXEC, 0); |
| 245 |
if (dfd < 0 && errno != ENOTDIR) |
| 246 |
return -errno; |
| 247 |
|
| 248 |
if (renameat2(source_atfd, source_name, target_atfd, target_name, RENAME_EXCHANGE) < 0) { |
| 249 |
|
| 250 |
if (!ERRNO_IS_NOT_SUPPORTED(errno) && errno != EINVAL) |
| 251 |
return -errno; |
| 252 |
|
| 253 |
/* The exchange didn't work, let's remove the target first, and try again */ |
| 254 |
|
| 255 |
if (dfd >= 0) |
| 256 |
(void) rm_rf_children(TAKE_FD(dfd), REMOVE_PHYSICAL|REMOVE_SUBVOLUME|REMOVE_CHMOD, NULL); |
| 257 |
|
| 258 |
r = unlinkat_maybe_dir(target_atfd, target_name); |
| 259 |
if (r < 0) |
| 260 |
return log_debug_errno(r, "Failed to remove target directory: %m"); |
| 261 |
|
| 262 |
if (renameat(source_atfd, source_name, target_atfd, target_name) < 0) |
| 263 |
return -errno; |
| 264 |
} else { |
| 265 |
/* The exchange worked, hence let's remove the source (i.e. the old target) */ |
| 266 |
if (dfd >= 0) |
| 267 |
(void) rm_rf_children(TAKE_FD(dfd), REMOVE_PHYSICAL|REMOVE_SUBVOLUME|REMOVE_CHMOD, NULL); |
| 268 |
|
| 269 |
r = unlinkat_maybe_dir(source_atfd, source_name); |
| 270 |
if (r < 0) |
| 271 |
return log_debug_errno(r, "Failed to remove replaced target directory: %m"); |
| 272 |
} |
| 273 |
} |
| 274 |
} else { |
| 275 |
r = rename_noreplace(source_atfd, source_name, target_atfd, target_name); |
| 276 |
if (r < 0) |
| 277 |
return r; |
| 278 |
} |
| 279 |
} |
| 280 |
|
| 281 |
if (rofd >= 0) { |
| 282 |
r = fs_make_very_read_only(rofd); |
| 283 |
if (r < 0) { |
| 284 |
if (!FLAGS_SET(flags, INSTALL_GRACEFUL)) |
| 285 |
return log_debug_errno(r, "Failed to make inode read-only: %m"); |
| 286 |
|
| 287 |
log_debug_errno(r, "Failed to make inode read-only, ignoring: %m"); |
| 288 |
} |
| 289 |
} |
| 290 |
|
| 291 |
if ((flags & (INSTALL_FSYNC_FULL|INSTALL_SYNCFS)) != 0) { |
| 292 |
if (target_name) |
| 293 |
r = fsync_parent_at(target_atfd, target_name); |
| 294 |
else |
| 295 |
r = fsync_parent_at(source_atfd, source_name); |
| 296 |
if (r < 0) { |
| 297 |
if (!FLAGS_SET(flags, INSTALL_GRACEFUL)) |
| 298 |
return log_debug_errno(r, "Failed to sync inode: %m"); |
| 299 |
|
| 300 |
log_debug_errno(r, "Failed to sync inode, ignoring: %m"); |
| 301 |
} |
| 302 |
} |
| 303 |
|
| 304 |
return 0; |
| 305 |
} |
| 306 |
|
| 307 |
usec_t parse_source_date_epoch(void) { |
| 308 |
static usec_t cache; |
| 309 |
static bool cached = false; |
| 310 |
int r; |
| 311 |
|
| 312 |
if (cached) |
| 313 |
return cache; |
| 314 |
|
| 315 |
uint64_t t; |
| 316 |
r = secure_getenv_uint64("SOURCE_DATE_EPOCH", &t); |
| 317 |
if (r >= 0) { |
| 318 |
if (MUL_SAFE(&cache, t, USEC_PER_SEC)) { |
| 319 |
cached = true; |
| 320 |
return cache; |
| 321 |
} |
| 322 |
|
| 323 |
r = -ERANGE; |
| 324 |
} |
| 325 |
if (r != -ENXIO) |
| 326 |
log_debug_errno(r, "Failed to parse $SOURCE_DATE_EPOCH, ignoring: %m"); |
| 327 |
|
| 328 |
cached = true; |
| 329 |
return (cache = USEC_INFINITY); |
| 330 |
} |
| 331 |
|
| 332 |
usec_t source_date_epoch_or_now(void) { |
| 333 |
usec_t epoch; |
| 334 |
|
| 335 |
epoch = parse_source_date_epoch(); |
| 336 |
if (epoch != USEC_INFINITY) |
| 337 |
return epoch; |
| 338 |
|
| 339 |
return now(CLOCK_REALTIME); |
| 340 |
} |