]>
Commit | Line | Data |
---|---|---|
db9ecf05 | 1 | /* SPDX-License-Identifier: LGPL-2.1-or-later */ |
cd61c3bf | 2 | |
a8fbdf54 | 3 | #include <errno.h> |
ebd93cb6 | 4 | #include <fcntl.h> |
e08f94ac LP |
5 | #include <linux/fs.h> |
6 | #include <linux/loop.h> | |
65ddc2c5 | 7 | #include <linux/magic.h> |
a8fbdf54 TA |
8 | #include <stdio.h> |
9 | #include <stdlib.h> | |
e306723e | 10 | #include <sys/file.h> |
204f52e3 | 11 | #include <sys/ioctl.h> |
a8fbdf54 TA |
12 | #include <sys/stat.h> |
13 | #include <unistd.h> | |
546dbec5 | 14 | |
b5efdb8a | 15 | #include "alloc-util.h" |
cd61c3bf | 16 | #include "btrfs-util.h" |
f4351959 | 17 | #include "chase-symlinks.h" |
c8b3094d | 18 | #include "chattr-util.h" |
ebd93cb6 | 19 | #include "copy.h" |
a0956174 | 20 | #include "dirent-util.h" |
57f1b61b | 21 | #include "discover-image.h" |
c7664c07 | 22 | #include "dissect-image.h" |
686d13b9 | 23 | #include "env-file.h" |
b6e953f2 | 24 | #include "env-util.h" |
3ffd4af2 | 25 | #include "fd-util.h" |
f4f15635 | 26 | #include "fs-util.h" |
a8fbdf54 | 27 | #include "hashmap.h" |
e2054217 | 28 | #include "hostname-setup.h" |
c7664c07 | 29 | #include "id128-util.h" |
a8fbdf54 TA |
30 | #include "lockfile-util.h" |
31 | #include "log.h" | |
c7664c07 | 32 | #include "loop-util.h" |
546dbec5 | 33 | #include "macro.h" |
30535c16 | 34 | #include "mkdir.h" |
d8b4d14d | 35 | #include "nulstr-util.h" |
d58ad743 | 36 | #include "os-util.h" |
8e0b6570 | 37 | #include "path-util.h" |
c6878637 | 38 | #include "rm-rf.h" |
65ddc2c5 | 39 | #include "stat-util.h" |
8b43440b | 40 | #include "string-table.h" |
07630cea | 41 | #include "string-util.h" |
8e0b6570 | 42 | #include "strv.h" |
a8fbdf54 | 43 | #include "time-util.h" |
8e0b6570 | 44 | #include "utf8.h" |
89a5a90c | 45 | #include "xattr-util.h" |
cd61c3bf | 46 | |
5ef46e5f | 47 | static const char* const image_search_path[_IMAGE_CLASS_MAX] = { |
9bca4ae4 LP |
48 | [IMAGE_MACHINE] = "/etc/machines\0" /* only place symlinks here */ |
49 | "/run/machines\0" /* and here too */ | |
50 | "/var/lib/machines\0" /* the main place for images */ | |
51 | "/var/lib/container\0" /* legacy */ | |
52 | "/usr/local/lib/machines\0" | |
53 | "/usr/lib/machines\0", | |
5ef46e5f | 54 | |
9bca4ae4 LP |
55 | [IMAGE_PORTABLE] = "/etc/portables\0" /* only place symlinks here */ |
56 | "/run/portables\0" /* and here too */ | |
57 | "/var/lib/portables\0" /* the main place for images */ | |
58 | "/usr/local/lib/portables\0" | |
59 | "/usr/lib/portables\0", | |
60 | ||
61 | [IMAGE_EXTENSION] = "/etc/extensions\0" /* only place symlinks here */ | |
62 | "/run/extensions\0" /* and here too */ | |
63 | "/var/lib/extensions\0" /* the main place for images */ | |
64 | "/usr/local/lib/extensions\0" | |
65 | "/usr/lib/extensions\0", | |
5ef46e5f | 66 | }; |
c2ce6a3d | 67 | |
8301aa0b YW |
68 | static Image *image_free(Image *i) { |
69 | assert(i); | |
9614bb06 | 70 | |
cd61c3bf LP |
71 | free(i->name); |
72 | free(i->path); | |
c7664c07 LP |
73 | |
74 | free(i->hostname); | |
75 | strv_free(i->machine_info); | |
76 | strv_free(i->os_release); | |
bcf94222 | 77 | strv_free(i->extension_release); |
c7664c07 | 78 | |
6b430fdb | 79 | return mfree(i); |
cd61c3bf LP |
80 | } |
81 | ||
8301aa0b | 82 | DEFINE_TRIVIAL_REF_UNREF_FUNC(Image, image, image_free); |
b07ec5a1 YW |
83 | DEFINE_HASH_OPS_WITH_VALUE_DESTRUCTOR(image_hash_ops, char, string_hash_func, string_compare_func, |
84 | Image, image_unref); | |
9614bb06 | 85 | |
8e0b6570 LP |
86 | static char **image_settings_path(Image *image) { |
87 | _cleanup_strv_free_ char **l = NULL; | |
5980d463 | 88 | const char *fn; |
8e0b6570 LP |
89 | unsigned i = 0; |
90 | ||
91 | assert(image); | |
92 | ||
93 | l = new0(char*, 4); | |
94 | if (!l) | |
95 | return NULL; | |
96 | ||
97 | fn = strjoina(image->name, ".nspawn"); | |
98 | ||
b910cc72 LP |
99 | FOREACH_STRING(s, "/etc/systemd/nspawn", "/run/systemd/nspawn") { |
100 | l[i] = path_join(s, fn); | |
8e0b6570 LP |
101 | if (!l[i]) |
102 | return NULL; | |
103 | ||
104 | i++; | |
105 | } | |
106 | ||
107 | l[i] = file_in_same_dir(image->path, fn); | |
108 | if (!l[i]) | |
109 | return NULL; | |
110 | ||
ae2a15bc | 111 | return TAKE_PTR(l); |
8e0b6570 LP |
112 | } |
113 | ||
bafbac4e LP |
114 | static char *image_roothash_path(Image *image) { |
115 | const char *fn; | |
116 | ||
117 | assert(image); | |
118 | ||
119 | fn = strjoina(image->name, ".roothash"); | |
120 | ||
121 | return file_in_same_dir(image->path, fn); | |
122 | } | |
123 | ||
c2ce6a3d | 124 | static int image_new( |
cd61c3bf | 125 | ImageType t, |
5fc7f358 | 126 | const char *pretty, |
cd61c3bf | 127 | const char *path, |
5fc7f358 | 128 | const char *filename, |
cd61c3bf | 129 | bool read_only, |
10f9c755 | 130 | usec_t crtime, |
cd61c3bf | 131 | usec_t mtime, |
c2ce6a3d | 132 | Image **ret) { |
cd61c3bf LP |
133 | |
134 | _cleanup_(image_unrefp) Image *i = NULL; | |
cd61c3bf | 135 | |
cd61c3bf LP |
136 | assert(t >= 0); |
137 | assert(t < _IMAGE_TYPE_MAX); | |
5fc7f358 LP |
138 | assert(pretty); |
139 | assert(filename); | |
c2ce6a3d | 140 | assert(ret); |
cd61c3bf | 141 | |
c2108701 | 142 | i = new(Image, 1); |
cd61c3bf LP |
143 | if (!i) |
144 | return -ENOMEM; | |
145 | ||
c2108701 LP |
146 | *i = (Image) { |
147 | .n_ref = 1, | |
148 | .type = t, | |
149 | .read_only = read_only, | |
150 | .crtime = crtime, | |
151 | .mtime = mtime, | |
152 | .usage = UINT64_MAX, | |
153 | .usage_exclusive = UINT64_MAX, | |
154 | .limit = UINT64_MAX, | |
155 | .limit_exclusive = UINT64_MAX, | |
156 | }; | |
cd61c3bf | 157 | |
5fc7f358 | 158 | i->name = strdup(pretty); |
cd61c3bf LP |
159 | if (!i->name) |
160 | return -ENOMEM; | |
161 | ||
657ee2d8 | 162 | i->path = path_join(path, filename); |
5fc7f358 LP |
163 | if (!i->path) |
164 | return -ENOMEM; | |
165 | ||
4ff361cc | 166 | path_simplify(i->path); |
cd61c3bf | 167 | |
1cc6c93a | 168 | *ret = TAKE_PTR(i); |
c2ce6a3d | 169 | |
cd61c3bf LP |
170 | return 0; |
171 | } | |
172 | ||
4756c94e LP |
173 | static int extract_pretty(const char *path, const char *suffix, char **ret) { |
174 | _cleanup_free_ char *name = NULL; | |
175 | const char *p; | |
176 | size_t n; | |
177 | ||
178 | assert(path); | |
179 | assert(ret); | |
180 | ||
181 | p = last_path_component(path); | |
182 | n = strcspn(p, "/"); | |
183 | ||
184 | name = strndup(p, n); | |
185 | if (!name) | |
186 | return -ENOMEM; | |
187 | ||
188 | if (suffix) { | |
189 | char *e; | |
190 | ||
191 | e = endswith(name, suffix); | |
192 | if (!e) | |
193 | return -EINVAL; | |
194 | ||
195 | *e = 0; | |
196 | } | |
197 | ||
198 | if (!image_name_is_valid(name)) | |
199 | return -EINVAL; | |
200 | ||
201 | *ret = TAKE_PTR(name); | |
202 | return 0; | |
203 | } | |
204 | ||
5fc7f358 LP |
205 | static int image_make( |
206 | const char *pretty, | |
207 | int dfd, | |
208 | const char *path, | |
209 | const char *filename, | |
3a6ce860 | 210 | const struct stat *st, |
5fc7f358 LP |
211 | Image **ret) { |
212 | ||
2570578d | 213 | _cleanup_free_ char *pretty_buffer = NULL, *parent = NULL; |
3a6ce860 | 214 | struct stat stbuf; |
5fc7f358 | 215 | bool read_only; |
cd61c3bf LP |
216 | int r; |
217 | ||
3a6ce860 | 218 | assert(dfd >= 0 || dfd == AT_FDCWD); |
bcb846f3 | 219 | assert(path || dfd == AT_FDCWD); |
5fc7f358 | 220 | assert(filename); |
cd61c3bf | 221 | |
eb38edce | 222 | /* We explicitly *do* follow symlinks here, since we want to allow symlinking trees, raw files and block |
3a6ce860 LP |
223 | * devices into /var/lib/machines/, and treat them normally. |
224 | * | |
225 | * This function returns -ENOENT if we can't find the image after all, and -EMEDIUMTYPE if it's not a file we | |
226 | * recognize. */ | |
cd61c3bf | 227 | |
3a6ce860 LP |
228 | if (!st) { |
229 | if (fstatat(dfd, filename, &stbuf, 0) < 0) | |
230 | return -errno; | |
231 | ||
232 | st = &stbuf; | |
233 | } | |
cd61c3bf | 234 | |
2570578d LP |
235 | if (!path) { |
236 | if (dfd == AT_FDCWD) | |
237 | (void) safe_getcwd(&parent); | |
238 | else | |
239 | (void) fd_get_path(dfd, &parent); | |
240 | } | |
bcb846f3 | 241 | |
5fc7f358 LP |
242 | read_only = |
243 | (path && path_startswith(path, "/usr")) || | |
08ff5529 | 244 | (faccessat(dfd, filename, W_OK, AT_EACCESS) < 0 && errno == EROFS); |
86e339c8 | 245 | |
3a6ce860 | 246 | if (S_ISDIR(st->st_mode)) { |
01b72568 LP |
247 | _cleanup_close_ int fd = -1; |
248 | unsigned file_attr = 0; | |
02dcf215 | 249 | usec_t crtime = 0; |
cd61c3bf | 250 | |
c2ce6a3d | 251 | if (!ret) |
3a6ce860 | 252 | return 0; |
cd61c3bf | 253 | |
4756c94e LP |
254 | if (!pretty) { |
255 | r = extract_pretty(filename, NULL, &pretty_buffer); | |
256 | if (r < 0) | |
257 | return r; | |
258 | ||
259 | pretty = pretty_buffer; | |
260 | } | |
5fc7f358 | 261 | |
01b72568 LP |
262 | fd = openat(dfd, filename, O_CLOEXEC|O_NOCTTY|O_DIRECTORY); |
263 | if (fd < 0) | |
264 | return -errno; | |
265 | ||
674b04ff | 266 | if (btrfs_might_be_subvol(st)) { |
cd61c3bf | 267 | |
65ddc2c5 | 268 | r = fd_is_fs_type(fd, BTRFS_SUPER_MAGIC); |
21222ea5 LP |
269 | if (r < 0) |
270 | return r; | |
271 | if (r) { | |
10f9c755 | 272 | BtrfsSubvolInfo info; |
cd61c3bf | 273 | |
c2ce6a3d | 274 | /* It's a btrfs subvolume */ |
cd61c3bf | 275 | |
5bcd08db | 276 | r = btrfs_subvol_get_info_fd(fd, 0, &info); |
10f9c755 LP |
277 | if (r < 0) |
278 | return r; | |
c2ce6a3d LP |
279 | |
280 | r = image_new(IMAGE_SUBVOLUME, | |
5fc7f358 | 281 | pretty, |
c2ce6a3d | 282 | path, |
5fc7f358 LP |
283 | filename, |
284 | info.read_only || read_only, | |
10f9c755 | 285 | info.otime, |
c2ce6a3d | 286 | 0, |
c2ce6a3d LP |
287 | ret); |
288 | if (r < 0) | |
289 | return r; | |
290 | ||
5bcd08db LP |
291 | if (btrfs_quota_scan_ongoing(fd) == 0) { |
292 | BtrfsQuotaInfo quota; | |
b6b18498 | 293 | |
5bcd08db LP |
294 | r = btrfs_subvol_get_subtree_quota_fd(fd, 0, "a); |
295 | if (r >= 0) { | |
296 | (*ret)->usage = quota.referenced; | |
297 | (*ret)->usage_exclusive = quota.exclusive; | |
298 | ||
299 | (*ret)->limit = quota.referenced_max; | |
300 | (*ret)->limit_exclusive = quota.exclusive_max; | |
301 | } | |
b6b18498 LP |
302 | } |
303 | ||
3a6ce860 | 304 | return 0; |
cd61c3bf | 305 | } |
c2ce6a3d | 306 | } |
cd61c3bf | 307 | |
02dcf215 | 308 | /* Get directory creation time (not available everywhere, but that's OK */ |
12a7f04a | 309 | (void) fd_getcrtime(fd, &crtime); |
02dcf215 LP |
310 | |
311 | /* If the IMMUTABLE bit is set, we consider the directory read-only. Since the ioctl is not | |
01b72568 LP |
312 | * supported everywhere we ignore failures. */ |
313 | (void) read_attr_fd(fd, &file_attr); | |
cd61c3bf | 314 | |
01b72568 | 315 | /* It's just a normal directory. */ |
c2ce6a3d | 316 | r = image_new(IMAGE_DIRECTORY, |
5fc7f358 | 317 | pretty, |
c2ce6a3d | 318 | path, |
5fc7f358 | 319 | filename, |
01b72568 | 320 | read_only || (file_attr & FS_IMMUTABLE_FL), |
02dcf215 LP |
321 | crtime, |
322 | 0, /* we don't use mtime of stat() here, since it's not the time of last change of the tree, but only of the top-level dir */ | |
c2ce6a3d LP |
323 | ret); |
324 | if (r < 0) | |
325 | return r; | |
cd61c3bf | 326 | |
3a6ce860 | 327 | return 0; |
cd61c3bf | 328 | |
3a6ce860 | 329 | } else if (S_ISREG(st->st_mode) && endswith(filename, ".raw")) { |
10f9c755 | 330 | usec_t crtime = 0; |
cd61c3bf | 331 | |
aceac2f0 | 332 | /* It's a RAW disk image */ |
cd61c3bf | 333 | |
c2ce6a3d | 334 | if (!ret) |
3a6ce860 | 335 | return 0; |
cd61c3bf | 336 | |
c53e07e2 | 337 | (void) fd_getcrtime_at(dfd, filename, AT_SYMLINK_FOLLOW, &crtime); |
10f9c755 | 338 | |
4756c94e LP |
339 | if (!pretty) { |
340 | r = extract_pretty(filename, ".raw", &pretty_buffer); | |
341 | if (r < 0) | |
342 | return r; | |
343 | ||
344 | pretty = pretty_buffer; | |
345 | } | |
10f9c755 | 346 | |
aceac2f0 | 347 | r = image_new(IMAGE_RAW, |
5fc7f358 | 348 | pretty, |
c2ce6a3d | 349 | path, |
5fc7f358 | 350 | filename, |
3a6ce860 | 351 | !(st->st_mode & 0222) || read_only, |
10f9c755 | 352 | crtime, |
3a6ce860 | 353 | timespec_load(&st->st_mtim), |
c2ce6a3d LP |
354 | ret); |
355 | if (r < 0) | |
356 | return r; | |
cd61c3bf | 357 | |
3a6ce860 LP |
358 | (*ret)->usage = (*ret)->usage_exclusive = st->st_blocks * 512; |
359 | (*ret)->limit = (*ret)->limit_exclusive = st->st_size; | |
b6b18498 | 360 | |
3a6ce860 | 361 | return 0; |
eb38edce | 362 | |
3a6ce860 | 363 | } else if (S_ISBLK(st->st_mode)) { |
eb38edce LP |
364 | _cleanup_close_ int block_fd = -1; |
365 | uint64_t size = UINT64_MAX; | |
366 | ||
367 | /* A block device */ | |
368 | ||
369 | if (!ret) | |
3a6ce860 | 370 | return 0; |
eb38edce | 371 | |
4756c94e LP |
372 | if (!pretty) { |
373 | r = extract_pretty(filename, NULL, &pretty_buffer); | |
374 | if (r < 0) | |
375 | return r; | |
376 | ||
377 | pretty = pretty_buffer; | |
378 | } | |
eb38edce LP |
379 | |
380 | block_fd = openat(dfd, filename, O_RDONLY|O_NONBLOCK|O_CLOEXEC|O_NOCTTY); | |
381 | if (block_fd < 0) | |
2570578d | 382 | log_debug_errno(errno, "Failed to open block device %s/%s, ignoring: %m", path ?: strnull(parent), filename); |
eb38edce | 383 | else { |
3a6ce860 LP |
384 | /* Refresh stat data after opening the node */ |
385 | if (fstat(block_fd, &stbuf) < 0) | |
eb38edce | 386 | return -errno; |
3a6ce860 LP |
387 | st = &stbuf; |
388 | ||
389 | if (!S_ISBLK(st->st_mode)) /* Verify that what we opened is actually what we think it is */ | |
eb38edce LP |
390 | return -ENOTTY; |
391 | ||
392 | if (!read_only) { | |
393 | int state = 0; | |
394 | ||
395 | if (ioctl(block_fd, BLKROGET, &state) < 0) | |
2570578d | 396 | log_debug_errno(errno, "Failed to issue BLKROGET on device %s/%s, ignoring: %m", path ?: strnull(parent), filename); |
eb38edce LP |
397 | else if (state) |
398 | read_only = true; | |
399 | } | |
400 | ||
401 | if (ioctl(block_fd, BLKGETSIZE64, &size) < 0) | |
2570578d | 402 | log_debug_errno(errno, "Failed to issue BLKGETSIZE64 on device %s/%s, ignoring: %m", path ?: strnull(parent), filename); |
eb38edce LP |
403 | |
404 | block_fd = safe_close(block_fd); | |
405 | } | |
406 | ||
407 | r = image_new(IMAGE_BLOCK, | |
408 | pretty, | |
409 | path, | |
410 | filename, | |
3a6ce860 | 411 | !(st->st_mode & 0222) || read_only, |
eb38edce LP |
412 | 0, |
413 | 0, | |
414 | ret); | |
415 | if (r < 0) | |
416 | return r; | |
417 | ||
ed0cb346 | 418 | if (!IN_SET(size, 0, UINT64_MAX)) |
eb38edce LP |
419 | (*ret)->usage = (*ret)->usage_exclusive = (*ret)->limit = (*ret)->limit_exclusive = size; |
420 | ||
3a6ce860 | 421 | return 0; |
c2ce6a3d | 422 | } |
cd61c3bf | 423 | |
3a6ce860 | 424 | return -EMEDIUMTYPE; |
c2ce6a3d | 425 | } |
cd61c3bf | 426 | |
d577d4a4 LP |
427 | int image_find(ImageClass class, |
428 | const char *name, | |
429 | const char *root, | |
430 | Image **ret) { | |
431 | ||
c2ce6a3d LP |
432 | const char *path; |
433 | int r; | |
cd61c3bf | 434 | |
5ef46e5f LP |
435 | assert(class >= 0); |
436 | assert(class < _IMAGE_CLASS_MAX); | |
c2ce6a3d | 437 | assert(name); |
cd61c3bf | 438 | |
c2ce6a3d LP |
439 | /* There are no images with invalid names */ |
440 | if (!image_name_is_valid(name)) | |
3a6ce860 | 441 | return -ENOENT; |
cd61c3bf | 442 | |
5ef46e5f | 443 | NULSTR_FOREACH(path, image_search_path[class]) { |
d577d4a4 | 444 | _cleanup_free_ char *resolved = NULL; |
c2ce6a3d | 445 | _cleanup_closedir_ DIR *d = NULL; |
3a6ce860 | 446 | struct stat st; |
d577d4a4 | 447 | int flags; |
cd61c3bf | 448 | |
d577d4a4 LP |
449 | r = chase_symlinks_and_opendir(path, root, CHASE_PREFIX_ROOT, &resolved, &d); |
450 | if (r == -ENOENT) | |
451 | continue; | |
452 | if (r < 0) | |
453 | return r; | |
cd61c3bf | 454 | |
d577d4a4 LP |
455 | /* As mentioned above, we follow symlinks on this fstatat(), because we want to permit people |
456 | * to symlink block devices into the search path. (For now, we disable that when operating | |
457 | * relative to some root directory.) */ | |
458 | flags = root ? AT_SYMLINK_NOFOLLOW : 0; | |
459 | if (fstatat(dirfd(d), name, &st, flags) < 0) { | |
aceac2f0 | 460 | _cleanup_free_ char *raw = NULL; |
5fc7f358 | 461 | |
3a6ce860 LP |
462 | if (errno != ENOENT) |
463 | return -errno; | |
464 | ||
b910cc72 | 465 | raw = strjoin(name, ".raw"); |
aceac2f0 | 466 | if (!raw) |
5fc7f358 LP |
467 | return -ENOMEM; |
468 | ||
d577d4a4 | 469 | if (fstatat(dirfd(d), raw, &st, flags) < 0) { |
3a6ce860 LP |
470 | if (errno == ENOENT) |
471 | continue; | |
472 | ||
473 | return -errno; | |
474 | } | |
475 | ||
476 | if (!S_ISREG(st.st_mode)) | |
5fc7f358 | 477 | continue; |
3a6ce860 | 478 | |
d577d4a4 | 479 | r = image_make(name, dirfd(d), resolved, raw, &st, ret); |
3a6ce860 LP |
480 | |
481 | } else { | |
482 | if (!S_ISDIR(st.st_mode) && !S_ISBLK(st.st_mode)) | |
483 | continue; | |
484 | ||
d577d4a4 | 485 | r = image_make(name, dirfd(d), resolved, name, &st, ret); |
5fc7f358 | 486 | } |
3a6ce860 LP |
487 | if (IN_SET(r, -ENOENT, -EMEDIUMTYPE)) |
488 | continue; | |
c2ce6a3d LP |
489 | if (r < 0) |
490 | return r; | |
cd61c3bf | 491 | |
cf604fd4 LP |
492 | if (ret) |
493 | (*ret)->discoverable = true; | |
494 | ||
c2ce6a3d LP |
495 | return 1; |
496 | } | |
497 | ||
cf604fd4 | 498 | if (class == IMAGE_MACHINE && streq(name, ".host")) { |
d577d4a4 | 499 | r = image_make(".host", AT_FDCWD, NULL, empty_to_root(root), NULL, ret); |
cf604fd4 LP |
500 | if (r < 0) |
501 | return r; | |
502 | ||
503 | if (ret) | |
504 | (*ret)->discoverable = true; | |
505 | ||
506 | return r; | |
507 | } | |
5fc7f358 | 508 | |
3a6ce860 | 509 | return -ENOENT; |
c2ce6a3d LP |
510 | }; |
511 | ||
2ddf182b | 512 | int image_from_path(const char *path, Image **ret) { |
cf604fd4 LP |
513 | |
514 | /* Note that we don't set the 'discoverable' field of the returned object, because we don't check here whether | |
515 | * the image is in the image search path. And if it is we don't know if the path we used is actually not | |
3fe91079 | 516 | * overridden by another, different image earlier in the search path */ |
cf604fd4 | 517 | |
2ddf182b LP |
518 | if (path_equal(path, "/")) |
519 | return image_make(".host", AT_FDCWD, NULL, "/", NULL, ret); | |
520 | ||
521 | return image_make(NULL, AT_FDCWD, NULL, path, NULL, ret); | |
522 | } | |
523 | ||
d577d4a4 | 524 | int image_find_harder(ImageClass class, const char *name_or_path, const char *root, Image **ret) { |
2ddf182b | 525 | if (image_name_is_valid(name_or_path)) |
d577d4a4 | 526 | return image_find(class, name_or_path, root, ret); |
2ddf182b LP |
527 | |
528 | return image_from_path(name_or_path, ret); | |
529 | } | |
530 | ||
d577d4a4 LP |
531 | int image_discover( |
532 | ImageClass class, | |
533 | const char *root, | |
534 | Hashmap *h) { | |
535 | ||
c2ce6a3d LP |
536 | const char *path; |
537 | int r; | |
538 | ||
5ef46e5f LP |
539 | assert(class >= 0); |
540 | assert(class < _IMAGE_CLASS_MAX); | |
c2ce6a3d LP |
541 | assert(h); |
542 | ||
5ef46e5f | 543 | NULSTR_FOREACH(path, image_search_path[class]) { |
d577d4a4 | 544 | _cleanup_free_ char *resolved = NULL; |
c2ce6a3d | 545 | _cleanup_closedir_ DIR *d = NULL; |
c2ce6a3d | 546 | |
d577d4a4 LP |
547 | r = chase_symlinks_and_opendir(path, root, CHASE_PREFIX_ROOT, &resolved, &d); |
548 | if (r == -ENOENT) | |
549 | continue; | |
550 | if (r < 0) | |
551 | return r; | |
c2ce6a3d LP |
552 | |
553 | FOREACH_DIRENT_ALL(de, d, return -errno) { | |
554 | _cleanup_(image_unrefp) Image *image = NULL; | |
1bf36bf9 | 555 | _cleanup_free_ char *truncated = NULL; |
3a6ce860 LP |
556 | const char *pretty; |
557 | struct stat st; | |
d577d4a4 | 558 | int flags; |
c2ce6a3d | 559 | |
1bf36bf9 | 560 | if (dot_or_dot_dot(de->d_name)) |
c2ce6a3d LP |
561 | continue; |
562 | ||
d577d4a4 LP |
563 | /* As mentioned above, we follow symlinks on this fstatat(), because we want to |
564 | * permit people to symlink block devices into the search path. */ | |
565 | flags = root ? AT_SYMLINK_NOFOLLOW : 0; | |
566 | if (fstatat(dirfd(d), de->d_name, &st, flags) < 0) { | |
3a6ce860 LP |
567 | if (errno == ENOENT) |
568 | continue; | |
569 | ||
570 | return -errno; | |
571 | } | |
572 | ||
573 | if (S_ISREG(st.st_mode)) { | |
574 | const char *e; | |
575 | ||
576 | e = endswith(de->d_name, ".raw"); | |
577 | if (!e) | |
578 | continue; | |
579 | ||
1bf36bf9 LP |
580 | truncated = strndup(de->d_name, e - de->d_name); |
581 | if (!truncated) | |
582 | return -ENOMEM; | |
583 | ||
584 | pretty = truncated; | |
3a6ce860 | 585 | } else if (S_ISDIR(st.st_mode) || S_ISBLK(st.st_mode)) |
1bf36bf9 | 586 | pretty = de->d_name; |
3a6ce860 LP |
587 | else |
588 | continue; | |
1bf36bf9 LP |
589 | |
590 | if (!image_name_is_valid(pretty)) | |
591 | continue; | |
592 | ||
593 | if (hashmap_contains(h, pretty)) | |
c2ce6a3d LP |
594 | continue; |
595 | ||
d577d4a4 | 596 | r = image_make(pretty, dirfd(d), resolved, de->d_name, &st, &image); |
3a6ce860 | 597 | if (IN_SET(r, -ENOENT, -EMEDIUMTYPE)) |
c2ce6a3d LP |
598 | continue; |
599 | if (r < 0) | |
600 | return r; | |
601 | ||
cf604fd4 LP |
602 | image->discoverable = true; |
603 | ||
c2ce6a3d LP |
604 | r = hashmap_put(h, image->name, image); |
605 | if (r < 0) | |
606 | return r; | |
607 | ||
608 | image = NULL; | |
cd61c3bf LP |
609 | } |
610 | } | |
611 | ||
5ef46e5f | 612 | if (class == IMAGE_MACHINE && !hashmap_contains(h, ".host")) { |
5fc7f358 LP |
613 | _cleanup_(image_unrefp) Image *image = NULL; |
614 | ||
d577d4a4 | 615 | r = image_make(".host", AT_FDCWD, NULL, empty_to_root("/"), NULL, &image); |
5fc7f358 LP |
616 | if (r < 0) |
617 | return r; | |
618 | ||
cf604fd4 LP |
619 | image->discoverable = true; |
620 | ||
5fc7f358 LP |
621 | r = hashmap_put(h, image->name, image); |
622 | if (r < 0) | |
623 | return r; | |
624 | ||
625 | image = NULL; | |
5fc7f358 LP |
626 | } |
627 | ||
cd61c3bf LP |
628 | return 0; |
629 | } | |
630 | ||
08682124 | 631 | int image_remove(Image *i) { |
8e766630 | 632 | _cleanup_(release_lock_file) LockFile global_lock = LOCK_FILE_INIT, local_lock = LOCK_FILE_INIT; |
8e0b6570 | 633 | _cleanup_strv_free_ char **settings = NULL; |
bafbac4e | 634 | _cleanup_free_ char *roothash = NULL; |
30535c16 LP |
635 | int r; |
636 | ||
08682124 LP |
637 | assert(i); |
638 | ||
d94c2b06 | 639 | if (IMAGE_IS_VENDOR(i) || IMAGE_IS_HOST(i)) |
08682124 LP |
640 | return -EROFS; |
641 | ||
8e0b6570 LP |
642 | settings = image_settings_path(i); |
643 | if (!settings) | |
644 | return -ENOMEM; | |
645 | ||
bafbac4e LP |
646 | roothash = image_roothash_path(i); |
647 | if (!roothash) | |
648 | return -ENOMEM; | |
649 | ||
30535c16 LP |
650 | /* Make sure we don't interfere with a running nspawn */ |
651 | r = image_path_lock(i->path, LOCK_EX|LOCK_NB, &global_lock, &local_lock); | |
652 | if (r < 0) | |
653 | return r; | |
654 | ||
ebd93cb6 LP |
655 | switch (i->type) { |
656 | ||
657 | case IMAGE_SUBVOLUME: | |
9fb0b9c7 LP |
658 | |
659 | /* Let's unlink first, maybe it is a symlink? If that works we are happy. Otherwise, let's get out the | |
660 | * big guns */ | |
661 | if (unlink(i->path) < 0) { | |
662 | r = btrfs_subvol_remove(i->path, BTRFS_REMOVE_RECURSIVE|BTRFS_REMOVE_QUOTA); | |
663 | if (r < 0) | |
664 | return r; | |
665 | } | |
666 | ||
8e0b6570 | 667 | break; |
ebd93cb6 LP |
668 | |
669 | case IMAGE_DIRECTORY: | |
01b72568 | 670 | /* Allow deletion of read-only directories */ |
db9a4254 | 671 | (void) chattr_path(i->path, 0, FS_IMMUTABLE_FL, NULL); |
8e0b6570 LP |
672 | r = rm_rf(i->path, REMOVE_ROOT|REMOVE_PHYSICAL|REMOVE_SUBVOLUME); |
673 | if (r < 0) | |
674 | return r; | |
675 | ||
676 | break; | |
01b72568 | 677 | |
eb38edce LP |
678 | case IMAGE_BLOCK: |
679 | ||
680 | /* If this is inside of /dev, then it's a real block device, hence let's not touch the device node | |
681 | * itself (but let's remove the stuff stored alongside it). If it's anywhere else, let's try to unlink | |
682 | * the thing (it's most likely a symlink after all). */ | |
683 | ||
684 | if (path_startswith(i->path, "/dev")) | |
685 | break; | |
686 | ||
4831981d | 687 | _fallthrough_; |
aceac2f0 | 688 | case IMAGE_RAW: |
41d1ed05 LP |
689 | if (unlink(i->path) < 0) |
690 | return -errno; | |
8e0b6570 | 691 | break; |
ebd93cb6 LP |
692 | |
693 | default: | |
15411c0c | 694 | return -EOPNOTSUPP; |
ebd93cb6 | 695 | } |
8e0b6570 | 696 | |
de010b0b | 697 | STRV_FOREACH(j, settings) |
8e0b6570 LP |
698 | if (unlink(*j) < 0 && errno != ENOENT) |
699 | log_debug_errno(errno, "Failed to unlink %s, ignoring: %m", *j); | |
8e0b6570 | 700 | |
bafbac4e LP |
701 | if (unlink(roothash) < 0 && errno != ENOENT) |
702 | log_debug_errno(errno, "Failed to unlink %s, ignoring: %m", roothash); | |
703 | ||
8e0b6570 LP |
704 | return 0; |
705 | } | |
706 | ||
bafbac4e | 707 | static int rename_auxiliary_file(const char *path, const char *new_name, const char *suffix) { |
8e0b6570 LP |
708 | _cleanup_free_ char *rs = NULL; |
709 | const char *fn; | |
710 | ||
bafbac4e | 711 | fn = strjoina(new_name, suffix); |
8e0b6570 LP |
712 | |
713 | rs = file_in_same_dir(path, fn); | |
714 | if (!rs) | |
715 | return -ENOMEM; | |
716 | ||
717 | return rename_noreplace(AT_FDCWD, path, AT_FDCWD, rs); | |
ebd93cb6 LP |
718 | } |
719 | ||
720 | int image_rename(Image *i, const char *new_name) { | |
8e766630 | 721 | _cleanup_(release_lock_file) LockFile global_lock = LOCK_FILE_INIT, local_lock = LOCK_FILE_INIT, name_lock = LOCK_FILE_INIT; |
bafbac4e | 722 | _cleanup_free_ char *new_path = NULL, *nn = NULL, *roothash = NULL; |
8e0b6570 | 723 | _cleanup_strv_free_ char **settings = NULL; |
01b72568 | 724 | unsigned file_attr = 0; |
ebd93cb6 LP |
725 | int r; |
726 | ||
727 | assert(i); | |
728 | ||
729 | if (!image_name_is_valid(new_name)) | |
730 | return -EINVAL; | |
731 | ||
d94c2b06 | 732 | if (IMAGE_IS_VENDOR(i) || IMAGE_IS_HOST(i)) |
ebd93cb6 LP |
733 | return -EROFS; |
734 | ||
8e0b6570 LP |
735 | settings = image_settings_path(i); |
736 | if (!settings) | |
737 | return -ENOMEM; | |
738 | ||
bafbac4e LP |
739 | roothash = image_roothash_path(i); |
740 | if (!roothash) | |
741 | return -ENOMEM; | |
742 | ||
30535c16 LP |
743 | /* Make sure we don't interfere with a running nspawn */ |
744 | r = image_path_lock(i->path, LOCK_EX|LOCK_NB, &global_lock, &local_lock); | |
745 | if (r < 0) | |
746 | return r; | |
747 | ||
748 | /* Make sure nobody takes the new name, between the time we | |
749 | * checked it is currently unused in all search paths, and the | |
f8e2f4d6 | 750 | * time we take possession of it */ |
30535c16 LP |
751 | r = image_name_lock(new_name, LOCK_EX|LOCK_NB, &name_lock); |
752 | if (r < 0) | |
753 | return r; | |
754 | ||
d577d4a4 | 755 | r = image_find(IMAGE_MACHINE, new_name, NULL, NULL); |
3a6ce860 | 756 | if (r >= 0) |
ebd93cb6 | 757 | return -EEXIST; |
3a6ce860 LP |
758 | if (r != -ENOENT) |
759 | return r; | |
ebd93cb6 LP |
760 | |
761 | switch (i->type) { | |
762 | ||
ebd93cb6 | 763 | case IMAGE_DIRECTORY: |
01b72568 LP |
764 | /* Turn of the immutable bit while we rename the image, so that we can rename it */ |
765 | (void) read_attr_path(i->path, &file_attr); | |
766 | ||
767 | if (file_attr & FS_IMMUTABLE_FL) | |
db9a4254 | 768 | (void) chattr_path(i->path, 0, FS_IMMUTABLE_FL, NULL); |
01b72568 | 769 | |
4831981d | 770 | _fallthrough_; |
01b72568 | 771 | case IMAGE_SUBVOLUME: |
ebd93cb6 LP |
772 | new_path = file_in_same_dir(i->path, new_name); |
773 | break; | |
774 | ||
eb38edce LP |
775 | case IMAGE_BLOCK: |
776 | ||
777 | /* Refuse renaming raw block devices in /dev, the names are picked by udev after all. */ | |
778 | if (path_startswith(i->path, "/dev")) | |
779 | return -EROFS; | |
780 | ||
781 | new_path = file_in_same_dir(i->path, new_name); | |
782 | break; | |
783 | ||
aceac2f0 | 784 | case IMAGE_RAW: { |
ebd93cb6 LP |
785 | const char *fn; |
786 | ||
63c372cb | 787 | fn = strjoina(new_name, ".raw"); |
ebd93cb6 LP |
788 | new_path = file_in_same_dir(i->path, fn); |
789 | break; | |
790 | } | |
791 | ||
792 | default: | |
15411c0c | 793 | return -EOPNOTSUPP; |
ebd93cb6 LP |
794 | } |
795 | ||
796 | if (!new_path) | |
797 | return -ENOMEM; | |
798 | ||
799 | nn = strdup(new_name); | |
800 | if (!nn) | |
801 | return -ENOMEM; | |
802 | ||
f85ef957 AC |
803 | r = rename_noreplace(AT_FDCWD, i->path, AT_FDCWD, new_path); |
804 | if (r < 0) | |
805 | return r; | |
ebd93cb6 | 806 | |
01b72568 LP |
807 | /* Restore the immutable bit, if it was set before */ |
808 | if (file_attr & FS_IMMUTABLE_FL) | |
db9a4254 | 809 | (void) chattr_path(new_path, FS_IMMUTABLE_FL, FS_IMMUTABLE_FL, NULL); |
01b72568 | 810 | |
f9ecfd3b DL |
811 | free_and_replace(i->path, new_path); |
812 | free_and_replace(i->name, nn); | |
ebd93cb6 | 813 | |
8e0b6570 | 814 | STRV_FOREACH(j, settings) { |
bafbac4e | 815 | r = rename_auxiliary_file(*j, new_name, ".nspawn"); |
8e0b6570 LP |
816 | if (r < 0 && r != -ENOENT) |
817 | log_debug_errno(r, "Failed to rename settings file %s, ignoring: %m", *j); | |
818 | } | |
819 | ||
bafbac4e LP |
820 | r = rename_auxiliary_file(roothash, new_name, ".roothash"); |
821 | if (r < 0 && r != -ENOENT) | |
822 | log_debug_errno(r, "Failed to rename roothash file %s, ignoring: %m", roothash); | |
823 | ||
ebd93cb6 LP |
824 | return 0; |
825 | } | |
826 | ||
bafbac4e | 827 | static int clone_auxiliary_file(const char *path, const char *new_name, const char *suffix) { |
8e0b6570 LP |
828 | _cleanup_free_ char *rs = NULL; |
829 | const char *fn; | |
830 | ||
bafbac4e | 831 | fn = strjoina(new_name, suffix); |
8e0b6570 LP |
832 | |
833 | rs = file_in_same_dir(path, fn); | |
834 | if (!rs) | |
835 | return -ENOMEM; | |
836 | ||
8a016c74 | 837 | return copy_file_atomic(path, rs, 0664, 0, 0, COPY_REFLINK); |
8e0b6570 LP |
838 | } |
839 | ||
ebd93cb6 | 840 | int image_clone(Image *i, const char *new_name, bool read_only) { |
8e766630 | 841 | _cleanup_(release_lock_file) LockFile name_lock = LOCK_FILE_INIT; |
8e0b6570 | 842 | _cleanup_strv_free_ char **settings = NULL; |
bafbac4e | 843 | _cleanup_free_ char *roothash = NULL; |
ebd93cb6 LP |
844 | const char *new_path; |
845 | int r; | |
846 | ||
847 | assert(i); | |
848 | ||
849 | if (!image_name_is_valid(new_name)) | |
850 | return -EINVAL; | |
851 | ||
8e0b6570 LP |
852 | settings = image_settings_path(i); |
853 | if (!settings) | |
854 | return -ENOMEM; | |
855 | ||
bafbac4e LP |
856 | roothash = image_roothash_path(i); |
857 | if (!roothash) | |
858 | return -ENOMEM; | |
859 | ||
30535c16 LP |
860 | /* Make sure nobody takes the new name, between the time we |
861 | * checked it is currently unused in all search paths, and the | |
f8e2f4d6 | 862 | * time we take possession of it */ |
30535c16 LP |
863 | r = image_name_lock(new_name, LOCK_EX|LOCK_NB, &name_lock); |
864 | if (r < 0) | |
865 | return r; | |
866 | ||
d577d4a4 | 867 | r = image_find(IMAGE_MACHINE, new_name, NULL, NULL); |
3a6ce860 | 868 | if (r >= 0) |
ebd93cb6 | 869 | return -EEXIST; |
3a6ce860 LP |
870 | if (r != -ENOENT) |
871 | return r; | |
ebd93cb6 LP |
872 | |
873 | switch (i->type) { | |
874 | ||
875 | case IMAGE_SUBVOLUME: | |
876 | case IMAGE_DIRECTORY: | |
9a50e3ca | 877 | /* If we can we'll always try to create a new btrfs subvolume here, even if the source is a plain |
13e785f7 | 878 | * directory. */ |
9a50e3ca | 879 | |
63c372cb | 880 | new_path = strjoina("/var/lib/machines/", new_name); |
ebd93cb6 | 881 | |
17cbb288 LP |
882 | r = btrfs_subvol_snapshot(i->path, new_path, |
883 | (read_only ? BTRFS_SNAPSHOT_READ_ONLY : 0) | | |
884 | BTRFS_SNAPSHOT_FALLBACK_COPY | | |
885 | BTRFS_SNAPSHOT_FALLBACK_DIRECTORY | | |
886 | BTRFS_SNAPSHOT_FALLBACK_IMMUTABLE | | |
887 | BTRFS_SNAPSHOT_RECURSIVE | | |
888 | BTRFS_SNAPSHOT_QUOTA); | |
889 | if (r >= 0) | |
9a50e3ca | 890 | /* Enable "subtree" quotas for the copy, if we didn't copy any quota from the source. */ |
8120ee28 | 891 | (void) btrfs_subvol_auto_qgroup(new_path, 0, true); |
5bcd08db | 892 | |
ebd93cb6 LP |
893 | break; |
894 | ||
aceac2f0 | 895 | case IMAGE_RAW: |
63c372cb | 896 | new_path = strjoina("/var/lib/machines/", new_name, ".raw"); |
ebd93cb6 | 897 | |
8a016c74 | 898 | r = copy_file_atomic(i->path, new_path, read_only ? 0444 : 0644, FS_NOCOW_FL, FS_NOCOW_FL, COPY_REFLINK|COPY_CRTIME); |
ebd93cb6 LP |
899 | break; |
900 | ||
eb38edce | 901 | case IMAGE_BLOCK: |
ebd93cb6 | 902 | default: |
15411c0c | 903 | return -EOPNOTSUPP; |
ebd93cb6 LP |
904 | } |
905 | ||
906 | if (r < 0) | |
907 | return r; | |
908 | ||
8e0b6570 | 909 | STRV_FOREACH(j, settings) { |
bafbac4e | 910 | r = clone_auxiliary_file(*j, new_name, ".nspawn"); |
8e0b6570 LP |
911 | if (r < 0 && r != -ENOENT) |
912 | log_debug_errno(r, "Failed to clone settings %s, ignoring: %m", *j); | |
913 | } | |
914 | ||
bafbac4e LP |
915 | r = clone_auxiliary_file(roothash, new_name, ".roothash"); |
916 | if (r < 0 && r != -ENOENT) | |
917 | log_debug_errno(r, "Failed to clone root hash file %s, ignoring: %m", roothash); | |
918 | ||
ebd93cb6 LP |
919 | return 0; |
920 | } | |
921 | ||
922 | int image_read_only(Image *i, bool b) { | |
8e766630 | 923 | _cleanup_(release_lock_file) LockFile global_lock = LOCK_FILE_INIT, local_lock = LOCK_FILE_INIT; |
ebd93cb6 | 924 | int r; |
c7664c07 | 925 | |
ebd93cb6 LP |
926 | assert(i); |
927 | ||
d94c2b06 | 928 | if (IMAGE_IS_VENDOR(i) || IMAGE_IS_HOST(i)) |
ebd93cb6 LP |
929 | return -EROFS; |
930 | ||
30535c16 LP |
931 | /* Make sure we don't interfere with a running nspawn */ |
932 | r = image_path_lock(i->path, LOCK_EX|LOCK_NB, &global_lock, &local_lock); | |
933 | if (r < 0) | |
934 | return r; | |
935 | ||
ebd93cb6 LP |
936 | switch (i->type) { |
937 | ||
938 | case IMAGE_SUBVOLUME: | |
5bcd08db LP |
939 | |
940 | /* Note that we set the flag only on the top-level | |
941 | * subvolume of the image. */ | |
942 | ||
ebd93cb6 LP |
943 | r = btrfs_subvol_set_read_only(i->path, b); |
944 | if (r < 0) | |
945 | return r; | |
01b72568 LP |
946 | |
947 | break; | |
948 | ||
949 | case IMAGE_DIRECTORY: | |
950 | /* For simple directory trees we cannot use the access | |
951 | mode of the top-level directory, since it has an | |
952 | effect on the container itself. However, we can | |
953 | use the "immutable" flag, to at least make the | |
954 | top-level directory read-only. It's not as good as | |
955 | a read-only subvolume, but at least something, and | |
13e785f7 | 956 | we can read the value back. */ |
01b72568 | 957 | |
db9a4254 | 958 | r = chattr_path(i->path, b ? FS_IMMUTABLE_FL : 0, FS_IMMUTABLE_FL, NULL); |
01b72568 LP |
959 | if (r < 0) |
960 | return r; | |
961 | ||
ebd93cb6 LP |
962 | break; |
963 | ||
aceac2f0 | 964 | case IMAGE_RAW: { |
ebd93cb6 LP |
965 | struct stat st; |
966 | ||
967 | if (stat(i->path, &st) < 0) | |
968 | return -errno; | |
969 | ||
970 | if (chmod(i->path, (st.st_mode & 0444) | (b ? 0000 : 0200)) < 0) | |
971 | return -errno; | |
f2068bcc LP |
972 | |
973 | /* If the images is now read-only, it's a good time to | |
974 | * defrag it, given that no write patterns will | |
975 | * fragment it again. */ | |
976 | if (b) | |
977 | (void) btrfs_defrag(i->path); | |
ebd93cb6 LP |
978 | break; |
979 | } | |
980 | ||
eb38edce LP |
981 | case IMAGE_BLOCK: { |
982 | _cleanup_close_ int fd = -1; | |
983 | struct stat st; | |
984 | int state = b; | |
985 | ||
986 | fd = open(i->path, O_CLOEXEC|O_RDONLY|O_NONBLOCK|O_NOCTTY); | |
987 | if (fd < 0) | |
988 | return -errno; | |
989 | ||
990 | if (fstat(fd, &st) < 0) | |
991 | return -errno; | |
992 | if (!S_ISBLK(st.st_mode)) | |
993 | return -ENOTTY; | |
994 | ||
995 | if (ioctl(fd, BLKROSET, &state) < 0) | |
996 | return -errno; | |
997 | ||
998 | break; | |
999 | } | |
1000 | ||
ebd93cb6 | 1001 | default: |
15411c0c | 1002 | return -EOPNOTSUPP; |
ebd93cb6 LP |
1003 | } |
1004 | ||
1005 | return 0; | |
08682124 LP |
1006 | } |
1007 | ||
30535c16 LP |
1008 | int image_path_lock(const char *path, int operation, LockFile *global, LockFile *local) { |
1009 | _cleanup_free_ char *p = NULL; | |
1010 | LockFile t = LOCK_FILE_INIT; | |
1011 | struct stat st; | |
f25bed67 | 1012 | bool exclusive; |
30535c16 LP |
1013 | int r; |
1014 | ||
1015 | assert(path); | |
1016 | assert(global); | |
1017 | assert(local); | |
1018 | ||
f25bed67 LP |
1019 | /* Locks an image path. This actually creates two locks: one "local" one, next to the image path |
1020 | * itself, which might be shared via NFS. And another "global" one, in /run, that uses the | |
1021 | * device/inode number. This has the benefit that we can even lock a tree that is a mount point, | |
1022 | * correctly. */ | |
30535c16 | 1023 | |
30535c16 LP |
1024 | if (!path_is_absolute(path)) |
1025 | return -EINVAL; | |
1026 | ||
f25bed67 LP |
1027 | switch (operation & (LOCK_SH|LOCK_EX)) { |
1028 | case LOCK_SH: | |
1029 | exclusive = false; | |
1030 | break; | |
1031 | case LOCK_EX: | |
1032 | exclusive = true; | |
1033 | break; | |
1034 | default: | |
1035 | return -EINVAL; | |
1036 | } | |
1037 | ||
b6e953f2 LP |
1038 | if (getenv_bool("SYSTEMD_NSPAWN_LOCK") == 0) { |
1039 | *local = *global = (LockFile) LOCK_FILE_INIT; | |
1040 | return 0; | |
1041 | } | |
1042 | ||
f25bed67 LP |
1043 | /* Prohibit taking exclusive locks on the host image. We can't allow this, since we ourselves are |
1044 | * running off it after all, and we don't want any images to manipulate the host image. We make an | |
1045 | * exception for shared locks however: we allow those (and make them NOPs since there's no point in | |
1046 | * taking them if there can't be exclusive locks). Strictly speaking these are questionable as well, | |
1047 | * since it means changes made to the host might propagate to the container as they happen (and a | |
1048 | * shared lock kinda suggests that no changes happen at all while it is in place), but it's too | |
1049 | * useful not to allow read-only containers off the host root, hence let's support this, and trust | |
1050 | * the user to do the right thing with this. */ | |
1051 | if (path_equal(path, "/")) { | |
1052 | if (exclusive) | |
1053 | return -EBUSY; | |
1054 | ||
1055 | *local = *global = (LockFile) LOCK_FILE_INIT; | |
1056 | return 0; | |
1057 | } | |
b6e953f2 | 1058 | |
30535c16 | 1059 | if (stat(path, &st) >= 0) { |
eb38edce LP |
1060 | if (S_ISBLK(st.st_mode)) |
1061 | r = asprintf(&p, "/run/systemd/nspawn/locks/block-%u:%u", major(st.st_rdev), minor(st.st_rdev)); | |
1062 | else if (S_ISDIR(st.st_mode) || S_ISREG(st.st_mode)) | |
1063 | r = asprintf(&p, "/run/systemd/nspawn/locks/inode-%lu:%lu", (unsigned long) st.st_dev, (unsigned long) st.st_ino); | |
1064 | else | |
1065 | return -ENOTTY; | |
eb38edce | 1066 | if (r < 0) |
30535c16 LP |
1067 | return -ENOMEM; |
1068 | } | |
1069 | ||
f25bed67 LP |
1070 | /* For block devices we don't need the "local" lock, as the major/minor lock above should be |
1071 | * sufficient, since block devices are host local anyway. */ | |
1072 | if (!path_startswith(path, "/dev/")) { | |
eb38edce | 1073 | r = make_lock_file_for(path, operation, &t); |
8be17c9b | 1074 | if (r < 0) { |
f25bed67 | 1075 | if (!exclusive && r == -EROFS) |
771b7ead | 1076 | log_debug_errno(r, "Failed to create shared lock for '%s', ignoring: %m", path); |
8be17c9b LT |
1077 | else |
1078 | return r; | |
1079 | } | |
eb38edce | 1080 | } |
30535c16 LP |
1081 | |
1082 | if (p) { | |
6e5dcce4 | 1083 | (void) mkdir_p("/run/systemd/nspawn/locks", 0700); |
30535c16 LP |
1084 | |
1085 | r = make_lock_file(p, operation, global); | |
1086 | if (r < 0) { | |
1087 | release_lock_file(&t); | |
1088 | return r; | |
1089 | } | |
546dbec5 LP |
1090 | } else |
1091 | *global = (LockFile) LOCK_FILE_INIT; | |
30535c16 LP |
1092 | |
1093 | *local = t; | |
1094 | return 0; | |
1095 | } | |
1096 | ||
cb81cd80 | 1097 | int image_set_limit(Image *i, uint64_t referenced_max) { |
d6ce17c7 LP |
1098 | assert(i); |
1099 | ||
d94c2b06 | 1100 | if (IMAGE_IS_VENDOR(i) || IMAGE_IS_HOST(i)) |
d6ce17c7 LP |
1101 | return -EROFS; |
1102 | ||
1103 | if (i->type != IMAGE_SUBVOLUME) | |
15411c0c | 1104 | return -EOPNOTSUPP; |
d6ce17c7 | 1105 | |
5bcd08db LP |
1106 | /* We set the quota both for the subvolume as well as for the |
1107 | * subtree. The latter is mostly for historical reasons, since | |
1108 | * we didn't use to have a concept of subtree quota, and hence | |
1109 | * only modified the subvolume quota. */ | |
1110 | ||
1111 | (void) btrfs_qgroup_set_limit(i->path, 0, referenced_max); | |
1112 | (void) btrfs_subvol_auto_qgroup(i->path, 0, true); | |
1113 | return btrfs_subvol_set_subtree_quota_limit(i->path, 0, referenced_max); | |
d6ce17c7 LP |
1114 | } |
1115 | ||
c7664c07 | 1116 | int image_read_metadata(Image *i) { |
8e766630 | 1117 | _cleanup_(release_lock_file) LockFile global_lock = LOCK_FILE_INIT, local_lock = LOCK_FILE_INIT; |
c7664c07 LP |
1118 | int r; |
1119 | ||
1120 | assert(i); | |
1121 | ||
1122 | r = image_path_lock(i->path, LOCK_SH|LOCK_NB, &global_lock, &local_lock); | |
1123 | if (r < 0) | |
1124 | return r; | |
1125 | ||
1126 | switch (i->type) { | |
1127 | ||
1128 | case IMAGE_SUBVOLUME: | |
1129 | case IMAGE_DIRECTORY: { | |
bcf94222 | 1130 | _cleanup_strv_free_ char **machine_info = NULL, **os_release = NULL, **extension_release = NULL; |
c7664c07 LP |
1131 | sd_id128_t machine_id = SD_ID128_NULL; |
1132 | _cleanup_free_ char *hostname = NULL; | |
1133 | _cleanup_free_ char *path = NULL; | |
1134 | ||
a5648b80 | 1135 | r = chase_symlinks("/etc/hostname", i->path, CHASE_PREFIX_ROOT|CHASE_TRAIL_SLASH, &path, NULL); |
c7664c07 LP |
1136 | if (r < 0 && r != -ENOENT) |
1137 | log_debug_errno(r, "Failed to chase /etc/hostname in image %s: %m", i->name); | |
1138 | else if (r >= 0) { | |
1139 | r = read_etc_hostname(path, &hostname); | |
1140 | if (r < 0) | |
1141 | log_debug_errno(errno, "Failed to read /etc/hostname of image %s: %m", i->name); | |
1142 | } | |
1143 | ||
1144 | path = mfree(path); | |
1145 | ||
a5648b80 | 1146 | r = chase_symlinks("/etc/machine-id", i->path, CHASE_PREFIX_ROOT|CHASE_TRAIL_SLASH, &path, NULL); |
c7664c07 LP |
1147 | if (r < 0 && r != -ENOENT) |
1148 | log_debug_errno(r, "Failed to chase /etc/machine-id in image %s: %m", i->name); | |
1149 | else if (r >= 0) { | |
1150 | _cleanup_close_ int fd = -1; | |
1151 | ||
1152 | fd = open(path, O_RDONLY|O_CLOEXEC|O_NOCTTY); | |
1153 | if (fd < 0) | |
1154 | log_debug_errno(errno, "Failed to open %s: %m", path); | |
1155 | else { | |
1156 | r = id128_read_fd(fd, ID128_PLAIN, &machine_id); | |
1157 | if (r < 0) | |
1158 | log_debug_errno(r, "Image %s contains invalid machine ID.", i->name); | |
1159 | } | |
1160 | } | |
1161 | ||
1162 | path = mfree(path); | |
1163 | ||
a5648b80 | 1164 | r = chase_symlinks("/etc/machine-info", i->path, CHASE_PREFIX_ROOT|CHASE_TRAIL_SLASH, &path, NULL); |
c7664c07 LP |
1165 | if (r < 0 && r != -ENOENT) |
1166 | log_debug_errno(r, "Failed to chase /etc/machine-info in image %s: %m", i->name); | |
1167 | else if (r >= 0) { | |
aa8fbc74 | 1168 | r = load_env_file_pairs(NULL, path, &machine_info); |
c7664c07 LP |
1169 | if (r < 0) |
1170 | log_debug_errno(r, "Failed to parse machine-info data of %s: %m", i->name); | |
1171 | } | |
1172 | ||
d58ad743 LP |
1173 | r = load_os_release_pairs(i->path, &os_release); |
1174 | if (r < 0) | |
1175 | log_debug_errno(r, "Failed to read os-release in image, ignoring: %m"); | |
c7664c07 | 1176 | |
bcf94222 LB |
1177 | r = load_extension_release_pairs(i->path, i->name, &extension_release); |
1178 | if (r < 0) | |
1179 | log_debug_errno(r, "Failed to read extension-release in image, ignoring: %m"); | |
1180 | ||
c7664c07 LP |
1181 | free_and_replace(i->hostname, hostname); |
1182 | i->machine_id = machine_id; | |
1183 | strv_free_and_replace(i->machine_info, machine_info); | |
1184 | strv_free_and_replace(i->os_release, os_release); | |
bcf94222 | 1185 | strv_free_and_replace(i->extension_release, extension_release); |
c7664c07 LP |
1186 | |
1187 | break; | |
1188 | } | |
1189 | ||
1190 | case IMAGE_RAW: | |
1191 | case IMAGE_BLOCK: { | |
1192 | _cleanup_(loop_device_unrefp) LoopDevice *d = NULL; | |
1193 | _cleanup_(dissected_image_unrefp) DissectedImage *m = NULL; | |
1194 | ||
7f52206a | 1195 | r = loop_device_make_by_path(i->path, O_RDONLY, LO_FLAGS_PARTSCAN, LOCK_SH, &d); |
41bc4849 LP |
1196 | if (r < 0) |
1197 | return r; | |
1198 | ||
d04faa4e LP |
1199 | r = dissect_image( |
1200 | d->fd, | |
1201 | NULL, NULL, | |
a3642997 | 1202 | d->diskseq, |
75dc190d | 1203 | d->uevent_seqnum_not_before, |
4a62257d | 1204 | d->timestamp_not_before, |
4b5de5dd LP |
1205 | DISSECT_IMAGE_GENERIC_ROOT | |
1206 | DISSECT_IMAGE_REQUIRE_ROOT | | |
1207 | DISSECT_IMAGE_RELAX_VAR_CHECK | | |
f6f4ec79 | 1208 | DISSECT_IMAGE_READ_ONLY | |
75dc190d LP |
1209 | DISSECT_IMAGE_USR_NO_ROOT, |
1210 | &m); | |
c7664c07 LP |
1211 | if (r < 0) |
1212 | return r; | |
1213 | ||
22847508 ZJS |
1214 | r = dissected_image_acquire_metadata(m, |
1215 | DISSECT_IMAGE_VALIDATE_OS | | |
1216 | DISSECT_IMAGE_VALIDATE_OS_EXT); | |
c7664c07 LP |
1217 | if (r < 0) |
1218 | return r; | |
1219 | ||
1220 | free_and_replace(i->hostname, m->hostname); | |
1221 | i->machine_id = m->machine_id; | |
1222 | strv_free_and_replace(i->machine_info, m->machine_info); | |
1223 | strv_free_and_replace(i->os_release, m->os_release); | |
bcf94222 | 1224 | strv_free_and_replace(i->extension_release, m->extension_release); |
c7664c07 LP |
1225 | |
1226 | break; | |
1227 | } | |
1228 | ||
1229 | default: | |
1230 | return -EOPNOTSUPP; | |
1231 | } | |
1232 | ||
1233 | i->metadata_valid = true; | |
1234 | ||
1235 | return 0; | |
1236 | } | |
1237 | ||
30535c16 | 1238 | int image_name_lock(const char *name, int operation, LockFile *ret) { |
99d97afc LP |
1239 | const char *p; |
1240 | ||
30535c16 LP |
1241 | assert(name); |
1242 | assert(ret); | |
1243 | ||
1244 | /* Locks an image name, regardless of the precise path used. */ | |
1245 | ||
99d97afc LP |
1246 | if (streq(name, ".host")) |
1247 | return -EBUSY; | |
1248 | ||
30535c16 LP |
1249 | if (!image_name_is_valid(name)) |
1250 | return -EINVAL; | |
1251 | ||
b6e953f2 LP |
1252 | if (getenv_bool("SYSTEMD_NSPAWN_LOCK") == 0) { |
1253 | *ret = (LockFile) LOCK_FILE_INIT; | |
1254 | return 0; | |
1255 | } | |
1256 | ||
6e5dcce4 | 1257 | (void) mkdir_p("/run/systemd/nspawn/locks", 0700); |
99d97afc LP |
1258 | |
1259 | p = strjoina("/run/systemd/nspawn/locks/name-", name); | |
30535c16 LP |
1260 | return make_lock_file(p, operation, ret); |
1261 | } | |
1262 | ||
d577d4a4 LP |
1263 | bool image_in_search_path( |
1264 | ImageClass class, | |
1265 | const char *root, | |
1266 | const char *image) { | |
1267 | ||
ace9ab19 LP |
1268 | const char *path; |
1269 | ||
1270 | assert(image); | |
1271 | ||
1272 | NULSTR_FOREACH(path, image_search_path[class]) { | |
d577d4a4 | 1273 | const char *p, *q; |
ace9ab19 LP |
1274 | size_t k; |
1275 | ||
d577d4a4 LP |
1276 | if (!empty_or_root(root)) { |
1277 | q = path_startswith(path, root); | |
1278 | if (!q) | |
1279 | continue; | |
1280 | } else | |
1281 | q = path; | |
1282 | ||
1283 | p = path_startswith(q, path); | |
ace9ab19 LP |
1284 | if (!p) |
1285 | continue; | |
1286 | ||
1287 | /* Make sure there's a filename following */ | |
1288 | k = strcspn(p, "/"); | |
1289 | if (k == 0) | |
1290 | continue; | |
1291 | ||
1292 | p += k; | |
1293 | ||
1294 | /* Accept trailing slashes */ | |
1295 | if (p[strspn(p, "/")] == 0) | |
1296 | return true; | |
1297 | ||
1298 | } | |
1299 | ||
1300 | return false; | |
1301 | } | |
1302 | ||
cd61c3bf LP |
1303 | static const char* const image_type_table[_IMAGE_TYPE_MAX] = { |
1304 | [IMAGE_DIRECTORY] = "directory", | |
1305 | [IMAGE_SUBVOLUME] = "subvolume", | |
aceac2f0 | 1306 | [IMAGE_RAW] = "raw", |
eb38edce | 1307 | [IMAGE_BLOCK] = "block", |
cd61c3bf LP |
1308 | }; |
1309 | ||
1310 | DEFINE_STRING_TABLE_LOOKUP(image_type, ImageType); |