]>
Commit | Line | Data |
---|---|---|
db9ecf05 | 1 | /* SPDX-License-Identifier: LGPL-2.1-or-later */ |
8c1be37e | 2 | |
10c1b188 LP |
3 | #if HAVE_VALGRIND_MEMCHECK_H |
4 | #include <valgrind/memcheck.h> | |
5 | #endif | |
6 | ||
dccca82b | 7 | #include <errno.h> |
8c1be37e | 8 | #include <fcntl.h> |
f1443709 LP |
9 | #include <linux/blkpg.h> |
10 | #include <linux/fs.h> | |
8c1be37e | 11 | #include <linux/loop.h> |
441ec804 | 12 | #include <sys/file.h> |
8c1be37e | 13 | #include <sys/ioctl.h> |
f2d9213f | 14 | #include <unistd.h> |
8c1be37e | 15 | |
021bf175 LP |
16 | #include "sd-device.h" |
17 | ||
8c1be37e | 18 | #include "alloc-util.h" |
86c1c1f3 | 19 | #include "blockdev-util.h" |
fcd8a19d | 20 | #include "data-fd-util.h" |
021bf175 | 21 | #include "device-util.h" |
7176f06c | 22 | #include "devnum-util.h" |
e8c7c4d9 | 23 | #include "env-util.h" |
b0a94268 | 24 | #include "errno-util.h" |
8c1be37e | 25 | #include "fd-util.h" |
f1443709 | 26 | #include "fileio.h" |
8c1be37e | 27 | #include "loop-util.h" |
86c1c1f3 | 28 | #include "missing_loop.h" |
f1443709 | 29 | #include "parse-util.h" |
e77cab82 | 30 | #include "path-util.h" |
b202ec20 | 31 | #include "random-util.h" |
3cc44114 | 32 | #include "stat-util.h" |
f1443709 | 33 | #include "stdio-util.h" |
f2d9213f | 34 | #include "string-util.h" |
021bf175 | 35 | #include "tmpfile-util.h" |
8c1be37e | 36 | |
e8af3bfd | 37 | static void cleanup_clear_loop_close(int *fd) { |
86c1c1f3 LP |
38 | if (*fd < 0) |
39 | return; | |
40 | ||
41 | (void) ioctl(*fd, LOOP_CLR_FD); | |
42 | (void) safe_close(*fd); | |
43 | } | |
44 | ||
021bf175 LP |
45 | static int loop_is_bound(int fd) { |
46 | struct loop_info64 info; | |
47 | ||
48 | assert(fd >= 0); | |
49 | ||
50 | if (ioctl(fd, LOOP_GET_STATUS64, &info) < 0) { | |
51 | if (errno == ENXIO) | |
52 | return false; /* not bound! */ | |
53 | ||
54 | return -errno; | |
55 | } | |
56 | ||
57 | return true; /* bound! */ | |
58 | } | |
59 | ||
31c75fcc LP |
60 | static int get_current_uevent_seqnum(uint64_t *ret) { |
61 | _cleanup_free_ char *p = NULL; | |
62 | int r; | |
63 | ||
64 | r = read_full_virtual_file("/sys/kernel/uevent_seqnum", &p, NULL); | |
65 | if (r < 0) | |
66 | return log_debug_errno(r, "Failed to read current uevent sequence number: %m"); | |
67 | ||
a145f8c0 | 68 | r = safe_atou64(strstrip(p), ret); |
31c75fcc LP |
69 | if (r < 0) |
70 | return log_debug_errno(r, "Failed to parse current uevent sequence number: %s", p); | |
71 | ||
72 | return 0; | |
73 | } | |
74 | ||
7f52206a | 75 | static int open_lock_fd(int primary_fd, int operation) { |
254d1313 | 76 | _cleanup_close_ int lock_fd = -EBADF; |
7f52206a LP |
77 | |
78 | assert(primary_fd >= 0); | |
10719a6f | 79 | assert(IN_SET(operation & ~LOCK_NB, LOCK_SH, LOCK_EX)); |
7f52206a | 80 | |
3e6b7d26 | 81 | lock_fd = fd_reopen(primary_fd, O_RDONLY|O_CLOEXEC|O_NONBLOCK|O_NOCTTY); |
7f52206a LP |
82 | if (lock_fd < 0) |
83 | return lock_fd; | |
10719a6f | 84 | |
7f52206a LP |
85 | if (flock(lock_fd, operation) < 0) |
86 | return -errno; | |
87 | ||
10719a6f | 88 | return TAKE_FD(lock_fd); |
7f52206a LP |
89 | } |
90 | ||
54ba7daf YW |
91 | static int loop_configure_verify_direct_io(int fd, const struct loop_config *c) { |
92 | assert(fd); | |
93 | assert(c); | |
94 | ||
95 | if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_DIRECT_IO)) { | |
96 | struct loop_info64 info; | |
97 | ||
98 | if (ioctl(fd, LOOP_GET_STATUS64, &info) < 0) | |
99 | return log_debug_errno(errno, "Failed to issue LOOP_GET_STATUS64: %m"); | |
100 | ||
101 | #if HAVE_VALGRIND_MEMCHECK_H | |
102 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
103 | #endif | |
104 | ||
105 | /* On older kernels (<= 5.3) it was necessary to set the block size of the loopback block | |
106 | * device to the logical block size of the underlying file system. Since there was no nice | |
107 | * way to query the value, we are not bothering to do this however. On newer kernels the | |
108 | * block size is propagated automatically and does not require intervention from us. We'll | |
109 | * check here if enabling direct IO worked, to make this easily debuggable however. | |
110 | * | |
111 | * (Should anyone really care and actually wants direct IO on old kernels: it might be worth | |
112 | * enabling direct IO with iteratively larger block sizes until it eventually works.) */ | |
113 | if (!FLAGS_SET(info.lo_flags, LO_FLAGS_DIRECT_IO)) | |
114 | log_debug("Could not enable direct IO mode, proceeding in buffered IO mode."); | |
115 | } | |
116 | ||
117 | return 0; | |
118 | } | |
119 | ||
120 | static int loop_configure_verify(int fd, const struct loop_config *c) { | |
121 | bool broken = false; | |
122 | int r; | |
123 | ||
124 | assert(fd >= 0); | |
125 | assert(c); | |
126 | ||
fd83c98e | 127 | if (c->block_size != 0) { |
65046b92 | 128 | uint32_t ssz; |
fd83c98e | 129 | |
65046b92 LP |
130 | r = blockdev_get_sector_size(fd, &ssz); |
131 | if (r < 0) | |
132 | return r; | |
fd83c98e | 133 | |
1163ddb3 | 134 | if (ssz != c->block_size) { |
65046b92 | 135 | log_debug("LOOP_CONFIGURE didn't honour requested block size %" PRIu32 ", got %" PRIu32 " instead. Ignoring.", c->block_size, ssz); |
1163ddb3 LP |
136 | broken = true; |
137 | } | |
fd83c98e AD |
138 | } |
139 | ||
54ba7daf YW |
140 | if (c->info.lo_sizelimit != 0) { |
141 | /* Kernel 5.8 vanilla doesn't properly propagate the size limit into the | |
142 | * block device. If it's used, let's immediately check if it had the desired | |
143 | * effect hence. And if not use classic LOOP_SET_STATUS64. */ | |
144 | uint64_t z; | |
145 | ||
146 | if (ioctl(fd, BLKGETSIZE64, &z) < 0) | |
147 | return -errno; | |
148 | ||
149 | if (z != c->info.lo_sizelimit) { | |
fd83c98e | 150 | log_debug("LOOP_CONFIGURE is broken, doesn't honour .info.lo_sizelimit. Falling back to LOOP_SET_STATUS64."); |
54ba7daf YW |
151 | broken = true; |
152 | } | |
153 | } | |
154 | ||
155 | if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_PARTSCAN)) { | |
156 | /* Kernel 5.8 vanilla doesn't properly propagate the partition scanning flag | |
157 | * into the block device. Let's hence verify if things work correctly here | |
158 | * before returning. */ | |
159 | ||
160 | r = blockdev_partscan_enabled(fd); | |
161 | if (r < 0) | |
162 | return r; | |
163 | if (r == 0) { | |
164 | log_debug("LOOP_CONFIGURE is broken, doesn't honour LO_FLAGS_PARTSCAN. Falling back to LOOP_SET_STATUS64."); | |
165 | broken = true; | |
166 | } | |
167 | } | |
168 | ||
169 | r = loop_configure_verify_direct_io(fd, c); | |
170 | if (r < 0) | |
171 | return r; | |
172 | ||
173 | return !broken; | |
174 | } | |
175 | ||
176 | static int loop_configure_fallback(int fd, const struct loop_config *c) { | |
177 | struct loop_info64 info_copy; | |
1163ddb3 | 178 | int r; |
54ba7daf YW |
179 | |
180 | assert(fd >= 0); | |
181 | assert(c); | |
182 | ||
183 | /* Only some of the flags LOOP_CONFIGURE can set are also settable via LOOP_SET_STATUS64, hence mask | |
184 | * them out. */ | |
185 | info_copy = c->info; | |
186 | info_copy.lo_flags &= LOOP_SET_STATUS_SETTABLE_FLAGS; | |
187 | ||
188 | /* Since kernel commit 5db470e229e22b7eda6e23b5566e532c96fb5bc3 (kernel v5.0) the LOOP_SET_STATUS64 | |
fd83c98e | 189 | * ioctl can return EAGAIN in case we change the info.lo_offset field, if someone else is accessing the |
54ba7daf YW |
190 | * block device while we try to reconfigure it. This is a pretty common case, since udev might |
191 | * instantly start probing the device as soon as we attach an fd to it. Hence handle it in two ways: | |
192 | * first, let's take the BSD lock to ensure that udev will not step in between the point in | |
193 | * time where we attach the fd and where we reconfigure the device. Secondly, let's wait 50ms on | |
194 | * EAGAIN and retry. The former should be an efficient mechanism to avoid we have to wait 50ms | |
195 | * needlessly if we are just racing against udev. The latter is protection against all other cases, | |
196 | * i.e. peers that do not take the BSD lock. */ | |
197 | ||
198 | for (unsigned n_attempts = 0;;) { | |
199 | if (ioctl(fd, LOOP_SET_STATUS64, &info_copy) >= 0) | |
200 | break; | |
201 | ||
202 | if (errno != EAGAIN || ++n_attempts >= 64) | |
203 | return log_debug_errno(errno, "Failed to configure loopback block device: %m"); | |
204 | ||
205 | /* Sleep some random time, but at least 10ms, at most 250ms. Increase the delay the more | |
206 | * failed attempts we see */ | |
207 | (void) usleep(UINT64_C(10) * USEC_PER_MSEC + | |
208 | random_u64_range(UINT64_C(240) * USEC_PER_MSEC * n_attempts/64)); | |
209 | } | |
210 | ||
211 | /* Work around a kernel bug, where changing offset/size of the loopback device doesn't correctly | |
212 | * invalidate the buffer cache. For details see: | |
213 | * | |
214 | * https://android.googlesource.com/platform/system/apex/+/bef74542fbbb4cd629793f4efee8e0053b360570 | |
215 | * | |
216 | * This was fixed in kernel 5.0, see: | |
217 | * | |
218 | * https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5db470e229e22b7eda6e23b5566e532c96fb5bc3 | |
219 | * | |
220 | * We'll run the work-around here in the legacy LOOP_SET_STATUS64 codepath. In the LOOP_CONFIGURE | |
221 | * codepath above it should not be necessary. */ | |
222 | if (c->info.lo_offset != 0 || c->info.lo_sizelimit != 0) | |
223 | if (ioctl(fd, BLKFLSBUF, 0) < 0) | |
224 | log_debug_errno(errno, "Failed to issue BLKFLSBUF ioctl, ignoring: %m"); | |
225 | ||
1163ddb3 LP |
226 | /* If a block size is requested then try to configure it. If that doesn't work, ignore errors, but |
227 | * afterwards, let's validate what is in effect, and if it doesn't match what we want, fail */ | |
228 | if (c->block_size != 0) { | |
229 | uint32_t ssz; | |
230 | ||
231 | if (ioctl(fd, LOOP_SET_BLOCK_SIZE, (unsigned long) c->block_size) < 0) | |
232 | log_debug_errno(errno, "Failed to set sector size, ignoring: %m"); | |
233 | ||
234 | r = blockdev_get_sector_size(fd, &ssz); | |
235 | if (r < 0) | |
236 | return log_debug_errno(r, "Failed to read sector size: %m"); | |
237 | if (ssz != c->block_size) | |
238 | return log_debug_errno(SYNTHETIC_ERRNO(EIO), "Sector size of loopback device doesn't match what we requested, refusing."); | |
239 | } | |
240 | ||
54ba7daf YW |
241 | /* LO_FLAGS_DIRECT_IO is a flags we need to configure via explicit ioctls. */ |
242 | if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_DIRECT_IO)) | |
243 | if (ioctl(fd, LOOP_SET_DIRECT_IO, 1UL) < 0) | |
244 | log_debug_errno(errno, "Failed to enable direct IO mode, ignoring: %m"); | |
245 | ||
246 | return loop_configure_verify_direct_io(fd, c); | |
247 | } | |
248 | ||
95c50092 | 249 | static int loop_configure( |
021bf175 | 250 | int nr, |
da4fd288 YW |
251 | int open_flags, |
252 | int lock_op, | |
95c50092 | 253 | const struct loop_config *c, |
da4fd288 | 254 | LoopDevice **ret) { |
95c50092 | 255 | |
bb273a51 YW |
256 | static bool loop_configure_broken = false; |
257 | ||
da4fd288 | 258 | _cleanup_(sd_device_unrefp) sd_device *dev = NULL; |
254d1313 ZJS |
259 | _cleanup_(cleanup_clear_loop_close) int loop_with_fd = -EBADF; /* This must be declared before lock_fd. */ |
260 | _cleanup_close_ int fd = -EBADF, lock_fd = -EBADF; | |
da4fd288 YW |
261 | _cleanup_free_ char *node = NULL; |
262 | uint64_t diskseq = 0, seqnum = UINT64_MAX; | |
263 | usec_t timestamp = USEC_INFINITY; | |
264 | dev_t devno; | |
86c1c1f3 LP |
265 | int r; |
266 | ||
021bf175 | 267 | assert(nr >= 0); |
86c1c1f3 | 268 | assert(c); |
da4fd288 YW |
269 | assert(ret); |
270 | ||
271 | if (asprintf(&node, "/dev/loop%i", nr) < 0) | |
272 | return -ENOMEM; | |
273 | ||
274 | r = sd_device_new_from_devname(&dev, node); | |
275 | if (r < 0) | |
276 | return r; | |
277 | ||
278 | r = sd_device_get_devnum(dev, &devno); | |
279 | if (r < 0) | |
280 | return r; | |
281 | ||
282 | fd = sd_device_open(dev, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags); | |
283 | if (fd < 0) | |
284 | return fd; | |
95c50092 | 285 | |
021bf175 LP |
286 | /* Let's lock the device before we do anything. We take the BSD lock on a second, separately opened |
287 | * fd for the device. udev after all watches for close() events (specifically IN_CLOSE_WRITE) on | |
288 | * block devices to reprobe them, hence by having a separate fd we will later close() we can ensure | |
289 | * we trigger udev after everything is done. If we'd lock our own fd instead and keep it open for a | |
290 | * long time udev would possibly never run on it again, even though the fd is unlocked, simply | |
291 | * because we never close() it. It also has the nice benefit we can use the _cleanup_close_ logic to | |
292 | * automatically release the lock, after we are done. */ | |
7f52206a | 293 | lock_fd = open_lock_fd(fd, LOCK_EX); |
021bf175 LP |
294 | if (lock_fd < 0) |
295 | return lock_fd; | |
021bf175 | 296 | |
53274fdb YW |
297 | /* Let's see if backing file is really unattached. Someone may already attach a backing file without |
298 | * taking BSD lock. */ | |
299 | r = loop_is_bound(fd); | |
300 | if (r < 0) | |
301 | return r; | |
302 | if (r > 0) | |
303 | return -EBUSY; | |
304 | ||
021bf175 LP |
305 | /* Let's see if the device is really detached, i.e. currently has no associated partition block |
306 | * devices. On various kernels (such as 5.8) it is possible to have a loopback block device that | |
247738b4 LP |
307 | * superficially is detached but still has partition block devices associated for it. Let's then |
308 | * manually remove the partitions via BLKPG, and tell the caller we did that via EUCLEAN, so they try | |
309 | * again. */ | |
833106b8 | 310 | r = block_device_remove_all_partitions(dev, fd); |
021bf175 LP |
311 | if (r < 0) |
312 | return r; | |
833106b8 YW |
313 | if (r > 0) |
314 | /* Removed all partitions. Let's report this to the caller, to try again, and count this as | |
53274fdb | 315 | * an attempt. */ |
247738b4 | 316 | return -EUCLEAN; |
021bf175 | 317 | |
bb273a51 | 318 | if (!loop_configure_broken) { |
31c75fcc LP |
319 | /* Acquire uevent seqnum immediately before attaching the loopback device. This allows |
320 | * callers to ignore all uevents with a seqnum before this one, if they need to associate | |
321 | * uevent with this attachment. Doing so isn't race-free though, as uevents that happen in | |
322 | * the window between this reading of the seqnum, and the LOOP_CONFIGURE call might still be | |
323 | * mistaken as originating from our attachment, even though might be caused by an earlier | |
324 | * use. But doing this at least shortens the race window a bit. */ | |
325 | r = get_current_uevent_seqnum(&seqnum); | |
326 | if (r < 0) | |
327 | return r; | |
54ba7daf | 328 | |
8ede1e86 | 329 | timestamp = now(CLOCK_MONOTONIC); |
31c75fcc | 330 | |
95c50092 LP |
331 | if (ioctl(fd, LOOP_CONFIGURE, c) < 0) { |
332 | /* Do fallback only if LOOP_CONFIGURE is not supported, propagate all other | |
333 | * errors. Note that the kernel is weird: non-existing ioctls currently return EINVAL | |
334 | * rather than ENOTTY on loopback block devices. They should fix that in the kernel, | |
335 | * but in the meantime we accept both here. */ | |
336 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && errno != EINVAL) | |
337 | return -errno; | |
86c1c1f3 | 338 | |
bb273a51 | 339 | loop_configure_broken = true; |
95c50092 | 340 | } else { |
da4fd288 YW |
341 | loop_with_fd = TAKE_FD(fd); |
342 | ||
343 | r = loop_configure_verify(loop_with_fd, c); | |
54ba7daf | 344 | if (r < 0) |
da4fd288 | 345 | return r; |
54ba7daf | 346 | if (r == 0) { |
95c50092 | 347 | /* LOOP_CONFIGURE doesn't work. Remember that. */ |
bb273a51 | 348 | loop_configure_broken = true; |
95c50092 LP |
349 | |
350 | /* We return EBUSY here instead of retrying immediately with LOOP_SET_FD, | |
351 | * because LOOP_CLR_FD is async: if the operation cannot be executed right | |
352 | * away it just sets the autoclear flag on the device. This means there's a | |
353 | * good chance we cannot actually reuse the loopback device right-away. Hence | |
354 | * let's assume it's busy, avoid the trouble and let the calling loop call us | |
355 | * again with a new, likely unused device. */ | |
da4fd288 | 356 | return -EBUSY; |
bb2551bd | 357 | } |
95c50092 | 358 | } |
86c1c1f3 LP |
359 | } |
360 | ||
bb273a51 YW |
361 | if (loop_configure_broken) { |
362 | /* Let's read the seqnum again, to shorten the window. */ | |
363 | r = get_current_uevent_seqnum(&seqnum); | |
364 | if (r < 0) | |
365 | return r; | |
31c75fcc | 366 | |
bb273a51 | 367 | timestamp = now(CLOCK_MONOTONIC); |
738f29cb | 368 | |
bb273a51 YW |
369 | if (ioctl(fd, LOOP_SET_FD, c->fd) < 0) |
370 | return -errno; | |
86c1c1f3 | 371 | |
da4fd288 YW |
372 | loop_with_fd = TAKE_FD(fd); |
373 | ||
374 | r = loop_configure_fallback(loop_with_fd, c); | |
bb273a51 | 375 | if (r < 0) |
da4fd288 | 376 | return r; |
bb273a51 | 377 | } |
e8c7c4d9 | 378 | |
da4fd288 YW |
379 | r = fd_get_diskseq(loop_with_fd, &diskseq); |
380 | if (r < 0 && r != -EOPNOTSUPP) | |
381 | return r; | |
31c75fcc | 382 | |
da4fd288 YW |
383 | switch (lock_op & ~LOCK_NB) { |
384 | case LOCK_EX: /* Already in effect */ | |
385 | break; | |
386 | case LOCK_SH: /* Downgrade */ | |
387 | if (flock(lock_fd, lock_op) < 0) | |
388 | return -errno; | |
389 | break; | |
390 | case LOCK_UN: /* Release */ | |
391 | lock_fd = safe_close(lock_fd); | |
392 | break; | |
393 | default: | |
394 | assert_not_reached(); | |
395 | } | |
396 | ||
397 | LoopDevice *d = new(LoopDevice, 1); | |
398 | if (!d) | |
399 | return -ENOMEM; | |
400 | ||
401 | *d = (LoopDevice) { | |
36d5eb0b | 402 | .n_ref = 1, |
da4fd288 YW |
403 | .fd = TAKE_FD(loop_with_fd), |
404 | .lock_fd = TAKE_FD(lock_fd), | |
405 | .node = TAKE_PTR(node), | |
406 | .nr = nr, | |
407 | .devno = devno, | |
408 | .dev = TAKE_PTR(dev), | |
409 | .diskseq = diskseq, | |
410 | .uevent_seqnum_not_before = seqnum, | |
411 | .timestamp_not_before = timestamp, | |
412 | }; | |
86c1c1f3 | 413 | |
da4fd288 YW |
414 | *ret = TAKE_PTR(d); |
415 | return 0; | |
e8af3bfd ZJS |
416 | } |
417 | ||
e8c7c4d9 | 418 | static int loop_device_make_internal( |
e77cab82 | 419 | const char *path, |
ed9eeb7b LP |
420 | int fd, |
421 | int open_flags, | |
422 | uint64_t offset, | |
423 | uint64_t size, | |
fd83c98e | 424 | uint32_t block_size, |
ed9eeb7b | 425 | uint32_t loop_flags, |
7f52206a | 426 | int lock_op, |
ed9eeb7b | 427 | LoopDevice **ret) { |
8c1be37e | 428 | |
da4fd288 | 429 | _cleanup_(loop_device_unrefp) LoopDevice *d = NULL; |
254d1313 | 430 | _cleanup_close_ int direct_io_fd = -EBADF, control = -EBADF; |
da4fd288 | 431 | _cleanup_free_ char *backing_file = NULL; |
86c1c1f3 | 432 | struct loop_config config; |
da4fd288 | 433 | int r, f_flags; |
8c1be37e | 434 | struct stat st; |
8c1be37e LP |
435 | |
436 | assert(fd >= 0); | |
437 | assert(ret); | |
438 | assert(IN_SET(open_flags, O_RDWR, O_RDONLY)); | |
439 | ||
440 | if (fstat(fd, &st) < 0) | |
441 | return -errno; | |
442 | ||
443 | if (S_ISBLK(st.st_mode)) { | |
1996ad28 | 444 | if (offset == 0 && IN_SET(size, 0, UINT64_MAX)) |
d7654742 LP |
445 | /* If this is already a block device and we are supposed to cover the whole of it |
446 | * then store an fd to the original open device node — and do not actually create an | |
1996ad28 | 447 | * unnecessary loopback device for it. */ |
de3b7f16 | 448 | return loop_device_open_from_fd(fd, open_flags, lock_op, ret); |
ed9eeb7b LP |
449 | } else { |
450 | r = stat_verify_regular(&st); | |
451 | if (r < 0) | |
452 | return r; | |
8c1be37e LP |
453 | } |
454 | ||
e77cab82 YW |
455 | if (path) { |
456 | r = path_make_absolute_cwd(path, &backing_file); | |
457 | if (r < 0) | |
458 | return r; | |
459 | ||
460 | path_simplify(backing_file); | |
461 | } else { | |
462 | r = fd_get_path(fd, &backing_file); | |
463 | if (r < 0) | |
464 | return r; | |
465 | } | |
466 | ||
e8c7c4d9 LP |
467 | f_flags = fcntl(fd, F_GETFL); |
468 | if (f_flags < 0) | |
469 | return -errno; | |
470 | ||
471 | if (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) != FLAGS_SET(f_flags, O_DIRECT)) { | |
472 | /* If LO_FLAGS_DIRECT_IO is requested, then make sure we have the fd open with O_DIRECT, as | |
473 | * that's required. Conversely, if it's off require that O_DIRECT is off too (that's because | |
474 | * new kernels will implicitly enable LO_FLAGS_DIRECT_IO if O_DIRECT is set). | |
475 | * | |
476 | * Our intention here is that LO_FLAGS_DIRECT_IO is the primary knob, and O_DIRECT derived | |
477 | * from that automatically. */ | |
478 | ||
479 | direct_io_fd = fd_reopen(fd, (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0)|O_CLOEXEC|O_NONBLOCK|open_flags); | |
480 | if (direct_io_fd < 0) { | |
481 | if (!FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO)) | |
482 | return log_debug_errno(errno, "Failed to reopen file descriptor without O_DIRECT: %m"); | |
483 | ||
484 | /* Some file systems might not support O_DIRECT, let's gracefully continue without it then. */ | |
485 | log_debug_errno(errno, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without."); | |
486 | loop_flags &= ~LO_FLAGS_DIRECT_IO; | |
487 | } else | |
488 | fd = direct_io_fd; /* From now on, operate on our new O_DIRECT fd */ | |
489 | } | |
490 | ||
8c1be37e LP |
491 | control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK); |
492 | if (control < 0) | |
493 | return -errno; | |
494 | ||
86c1c1f3 LP |
495 | config = (struct loop_config) { |
496 | .fd = fd, | |
fd83c98e | 497 | .block_size = block_size, |
86c1c1f3 LP |
498 | .info = { |
499 | /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */ | |
0950526a | 500 | .lo_flags = (loop_flags & ~LO_FLAGS_READ_ONLY) | ((open_flags & O_ACCMODE) == O_RDONLY ? LO_FLAGS_READ_ONLY : 0) | LO_FLAGS_AUTOCLEAR, |
86c1c1f3 LP |
501 | .lo_offset = offset, |
502 | .lo_sizelimit = size == UINT64_MAX ? 0 : size, | |
503 | }, | |
504 | }; | |
505 | ||
0f6519d4 LP |
506 | /* Loop around LOOP_CTL_GET_FREE, since at the moment we attempt to open the returned device it might |
507 | * be gone already, taken by somebody else racing against us. */ | |
e8af3bfd | 508 | for (unsigned n_attempts = 0;;) { |
da4fd288 | 509 | int nr; |
e8af3bfd | 510 | |
cc530466 LP |
511 | /* Let's take a lock on the control device first. On a busy system, where many programs |
512 | * attempt to allocate a loopback device at the same time, we might otherwise keep looping | |
513 | * around relatively heavy operations: asking for a free loopback device, then opening it, | |
514 | * validating it, attaching something to it. Let's serialize this whole operation, to make | |
515 | * unnecessary busywork less likely. Note that this is just something we do to optimize our | |
516 | * own code (and whoever else decides to use LOCK_EX locks for this), taking this lock is not | |
517 | * necessary, it just means it's less likely we have to iterate through this loop again and | |
4c1d50e6 LP |
518 | * again if our own code races against our own code. |
519 | * | |
520 | * Note: our lock protocol is to take the /dev/loop-control lock first, and the block device | |
521 | * lock second, if both are taken, and always in this order, to avoid ABBA locking issues. */ | |
cc530466 LP |
522 | if (flock(control, LOCK_EX) < 0) |
523 | return -errno; | |
524 | ||
0f6519d4 LP |
525 | nr = ioctl(control, LOOP_CTL_GET_FREE); |
526 | if (nr < 0) | |
527 | return -errno; | |
8c1be37e | 528 | |
da4fd288 YW |
529 | r = loop_configure(nr, open_flags, lock_op, &config, &d); |
530 | if (r >= 0) | |
531 | break; | |
cc5bae6c | 532 | |
da4fd288 YW |
533 | /* -ENODEV or friends: Somebody might've gotten the same number from the kernel, used the |
534 | * device, and called LOOP_CTL_REMOVE on it. Let's retry with a new number. | |
535 | * -EBUSY: a file descriptor is already bound to the loopback block device. | |
536 | * -EUCLEAN: some left-over partition devices that were cleaned up. */ | |
537 | if (!ERRNO_IS_DEVICE_ABSENT(errno) && !IN_SET(r, -EBUSY, -EUCLEAN)) | |
538 | return -errno; | |
01813148 | 539 | |
cc530466 LP |
540 | /* OK, this didn't work, let's try again a bit later, but first release the lock on the |
541 | * control device */ | |
542 | if (flock(control, LOCK_UN) < 0) | |
543 | return -errno; | |
544 | ||
e8af3bfd ZJS |
545 | if (++n_attempts >= 64) /* Give up eventually */ |
546 | return -EBUSY; | |
0f6519d4 | 547 | |
b202ec20 LP |
548 | /* Wait some random time, to make collision less likely. Let's pick a random time in the |
549 | * range 0ms…250ms, linearly scaled by the number of failed attempts. */ | |
b0dbffd8 LP |
550 | (void) usleep(random_u64_range(UINT64_C(10) * USEC_PER_MSEC + |
551 | UINT64_C(240) * USEC_PER_MSEC * n_attempts/64)); | |
0f6519d4 | 552 | } |
8c1be37e | 553 | |
da4fd288 | 554 | d->backing_file = TAKE_PTR(backing_file); |
8c1be37e | 555 | |
3b195f63 LP |
556 | log_debug("Successfully acquired %s, devno=%u:%u, nr=%i, diskseq=%" PRIu64, |
557 | d->node, | |
558 | major(d->devno), minor(d->devno), | |
559 | d->nr, | |
560 | d->diskseq); | |
561 | ||
da4fd288 YW |
562 | *ret = TAKE_PTR(d); |
563 | return 0; | |
8c1be37e LP |
564 | } |
565 | ||
e8c7c4d9 LP |
566 | static uint32_t loop_flags_mangle(uint32_t loop_flags) { |
567 | int r; | |
568 | ||
569 | r = getenv_bool("SYSTEMD_LOOP_DIRECT_IO"); | |
570 | if (r < 0 && r != -ENXIO) | |
571 | log_debug_errno(r, "Failed to parse $SYSTEMD_LOOP_DIRECT_IO, ignoring: %m"); | |
572 | ||
bfd08445 | 573 | return UPDATE_FLAG(loop_flags, LO_FLAGS_DIRECT_IO, r != 0); /* Turn on LO_FLAGS_DIRECT_IO by default, unless explicitly configured to off. */ |
e8c7c4d9 LP |
574 | } |
575 | ||
576 | int loop_device_make( | |
577 | int fd, | |
578 | int open_flags, | |
579 | uint64_t offset, | |
580 | uint64_t size, | |
fd83c98e | 581 | uint32_t block_size, |
e8c7c4d9 | 582 | uint32_t loop_flags, |
7f52206a | 583 | int lock_op, |
e8c7c4d9 LP |
584 | LoopDevice **ret) { |
585 | ||
586 | assert(fd >= 0); | |
587 | assert(ret); | |
e8c7c4d9 LP |
588 | |
589 | return loop_device_make_internal( | |
e77cab82 | 590 | NULL, |
e8c7c4d9 LP |
591 | fd, |
592 | open_flags, | |
593 | offset, | |
594 | size, | |
fd83c98e | 595 | block_size, |
bfd08445 | 596 | loop_flags_mangle(loop_flags), |
7f52206a | 597 | lock_op, |
e8c7c4d9 LP |
598 | ret); |
599 | } | |
600 | ||
79e8393a LP |
601 | int loop_device_make_by_path( |
602 | const char *path, | |
603 | int open_flags, | |
604 | uint32_t loop_flags, | |
7f52206a | 605 | int lock_op, |
79e8393a LP |
606 | LoopDevice **ret) { |
607 | ||
e8c7c4d9 | 608 | int r, basic_flags, direct_flags, rdwr_flags; |
254d1313 | 609 | _cleanup_close_ int fd = -EBADF; |
aa4d3aa3 | 610 | bool direct = false; |
8c1be37e LP |
611 | |
612 | assert(path); | |
613 | assert(ret); | |
b0a94268 | 614 | assert(open_flags < 0 || IN_SET(open_flags, O_RDWR, O_RDONLY)); |
8c1be37e | 615 | |
b0a94268 LP |
616 | /* Passing < 0 as open_flags here means we'll try to open the device writable if we can, retrying |
617 | * read-only if we cannot. */ | |
618 | ||
e8c7c4d9 LP |
619 | loop_flags = loop_flags_mangle(loop_flags); |
620 | ||
621 | /* Let's open with O_DIRECT if we can. But not all file systems support that, hence fall back to | |
622 | * non-O_DIRECT mode automatically, if it fails. */ | |
623 | ||
624 | basic_flags = O_CLOEXEC|O_NONBLOCK|O_NOCTTY; | |
625 | direct_flags = FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0; | |
626 | rdwr_flags = open_flags >= 0 ? open_flags : O_RDWR; | |
627 | ||
628 | fd = open(path, basic_flags|direct_flags|rdwr_flags); | |
629 | if (fd < 0 && direct_flags != 0) /* If we had O_DIRECT on, and things failed with that, let's immediately try again without */ | |
630 | fd = open(path, basic_flags|rdwr_flags); | |
aa4d3aa3 LP |
631 | else |
632 | direct = direct_flags != 0; | |
b0a94268 LP |
633 | if (fd < 0) { |
634 | r = -errno; | |
635 | ||
636 | /* Retry read-only? */ | |
637 | if (open_flags >= 0 || !(ERRNO_IS_PRIVILEGE(r) || r == -EROFS)) | |
638 | return r; | |
639 | ||
e8c7c4d9 LP |
640 | fd = open(path, basic_flags|direct_flags|O_RDONLY); |
641 | if (fd < 0 && direct_flags != 0) /* as above */ | |
642 | fd = open(path, basic_flags|O_RDONLY); | |
aa4d3aa3 LP |
643 | else |
644 | direct = direct_flags != 0; | |
b0a94268 LP |
645 | if (fd < 0) |
646 | return r; /* Propagate original error */ | |
647 | ||
648 | open_flags = O_RDONLY; | |
649 | } else if (open_flags < 0) | |
650 | open_flags = O_RDWR; | |
8c1be37e | 651 | |
aa4d3aa3 LP |
652 | log_debug("Opened '%s' in %s access mode%s, with O_DIRECT %s%s.", |
653 | path, | |
654 | open_flags == O_RDWR ? "O_RDWR" : "O_RDONLY", | |
655 | open_flags != rdwr_flags ? " (O_RDWR was requested but not allowed)" : "", | |
656 | direct ? "enabled" : "disabled", | |
657 | direct != (direct_flags != 0) ? " (O_DIRECT was requested but not supported)" : ""); | |
658 | ||
fd83c98e | 659 | return loop_device_make_internal(path, fd, open_flags, 0, 0, 0, loop_flags, lock_op, ret); |
8c1be37e LP |
660 | } |
661 | ||
fcd8a19d LP |
662 | int loop_device_make_by_path_memory( |
663 | const char *path, | |
664 | int open_flags, | |
665 | uint32_t loop_flags, | |
666 | int lock_op, | |
667 | LoopDevice **ret) { | |
668 | ||
669 | _cleanup_close_ int fd = -EBADF, mfd = -EBADF; | |
670 | _cleanup_free_ char *fn = NULL; | |
671 | struct stat st; | |
672 | int r; | |
673 | ||
674 | assert(path); | |
675 | assert(IN_SET(open_flags, O_RDWR, O_RDONLY)); | |
676 | assert(ret); | |
677 | ||
678 | loop_flags &= ~LO_FLAGS_DIRECT_IO; /* memfds don't support O_DIRECT, hence LO_FLAGS_DIRECT_IO can't be used either */ | |
679 | ||
680 | fd = open(path, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|O_RDONLY); | |
681 | if (fd < 0) | |
682 | return -errno; | |
683 | ||
684 | if (fstat(fd, &st) < 0) | |
685 | return -errno; | |
686 | ||
687 | if (!S_ISREG(st.st_mode) && !S_ISBLK(st.st_mode)) | |
688 | return -EBADF; | |
689 | ||
690 | r = path_extract_filename(path, &fn); | |
691 | if (r < 0) | |
692 | return r; | |
693 | ||
694 | mfd = memfd_clone_fd(fd, fn, open_flags|O_CLOEXEC); | |
695 | if (mfd < 0) | |
696 | return mfd; | |
697 | ||
698 | fd = safe_close(fd); /* Let's close the original early */ | |
699 | ||
700 | return loop_device_make_internal(NULL, mfd, open_flags, 0, 0, 0, loop_flags, lock_op, ret); | |
701 | } | |
702 | ||
36d5eb0b | 703 | static LoopDevice* loop_device_free(LoopDevice *d) { |
5bb1d7fb | 704 | _cleanup_close_ int control = -EBADF; |
3a6ed1e1 LP |
705 | int r; |
706 | ||
8c1be37e LP |
707 | if (!d) |
708 | return NULL; | |
709 | ||
4c1d50e6 LP |
710 | /* Release any lock we might have on the device first. We want to open+lock the /dev/loop-control |
711 | * device below, but our lock protocol says that if both control and block device locks are taken, | |
712 | * the control lock needs to be taken first, the block device lock second — in order to avoid ABBA | |
713 | * locking issues. Moreover, we want to issue LOOP_CLR_FD on the block device further down, and that | |
714 | * would fail if we had another fd open to the device. */ | |
7f52206a LP |
715 | d->lock_fd = safe_close(d->lock_fd); |
716 | ||
4c1d50e6 LP |
717 | /* Let's open the control device early, and lock it, so that we can release our block device and |
718 | * delete it in a synchronized fashion, and allocators won't needlessly see the block device as free | |
719 | * while we are about to delete it. */ | |
7cb349f0 | 720 | if (!LOOP_DEVICE_IS_FOREIGN(d) && !d->relinquished) { |
4c1d50e6 LP |
721 | control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK); |
722 | if (control < 0) | |
723 | log_debug_errno(errno, "Failed to open loop control device, cannot remove loop device '%s', ignoring: %m", strna(d->node)); | |
724 | else if (flock(control, LOCK_EX) < 0) | |
725 | log_debug_errno(errno, "Failed to lock loop control device, ignoring: %m"); | |
726 | } | |
727 | ||
728 | /* Then let's release the loopback block device */ | |
8c1be37e | 729 | if (d->fd >= 0) { |
cae1e8fb LP |
730 | /* Implicitly sync the device, since otherwise in-flight blocks might not get written */ |
731 | if (fsync(d->fd) < 0) | |
732 | log_debug_errno(errno, "Failed to sync loop block device, ignoring: %m"); | |
733 | ||
7cb349f0 | 734 | if (!LOOP_DEVICE_IS_FOREIGN(d) && !d->relinquished) { |
3a6ed1e1 LP |
735 | /* We are supposed to clear the loopback device. Let's do this synchronously: lock |
736 | * the device, manually remove all partitions and then clear it. This should ensure | |
737 | * udev doesn't concurrently access the devices, and we can be reasonably sure that | |
738 | * once we are done here the device is cleared and all its partition children | |
739 | * removed. Note that we lock our primary device fd here (and not a separate locking | |
740 | * fd, as we do during allocation, since we want to keep the lock all the way through | |
741 | * the LOOP_CLR_FD, but that call would fail if we had more than one fd open.) */ | |
8c1be37e | 742 | |
3a6ed1e1 LP |
743 | if (flock(d->fd, LOCK_EX) < 0) |
744 | log_debug_errno(errno, "Failed to lock loop block device, ignoring: %m"); | |
745 | ||
46c3a288 | 746 | r = block_device_remove_all_partitions(d->dev, d->fd); |
3a6ed1e1 LP |
747 | if (r < 0) |
748 | log_debug_errno(r, "Failed to remove partitions of loopback block device, ignoring: %m"); | |
749 | ||
750 | if (ioctl(d->fd, LOOP_CLR_FD) < 0) | |
751 | log_debug_errno(errno, "Failed to clear loop device, ignoring: %m"); | |
8c1be37e LP |
752 | } |
753 | ||
754 | safe_close(d->fd); | |
755 | } | |
756 | ||
4c1d50e6 LP |
757 | /* Now that the block device is released, let's also try to remove it */ |
758 | if (control >= 0) | |
759 | for (unsigned n_attempts = 0;;) { | |
760 | if (ioctl(control, LOOP_CTL_REMOVE, d->nr) >= 0) | |
761 | break; | |
762 | if (errno != EBUSY || ++n_attempts >= 64) { | |
763 | log_debug_errno(errno, "Failed to remove device %s: %m", strna(d->node)); | |
764 | break; | |
f2d9213f | 765 | } |
4c1d50e6 LP |
766 | (void) usleep(50 * USEC_PER_MSEC); |
767 | } | |
8c1be37e LP |
768 | |
769 | free(d->node); | |
cc5bae6c | 770 | sd_device_unref(d->dev); |
e77cab82 | 771 | free(d->backing_file); |
5fecf46d | 772 | return mfree(d); |
8c1be37e | 773 | } |
a2ea3b2f | 774 | |
36d5eb0b YW |
775 | DEFINE_TRIVIAL_REF_UNREF_FUNC(LoopDevice, loop_device, loop_device_free); |
776 | ||
a2ea3b2f LP |
777 | void loop_device_relinquish(LoopDevice *d) { |
778 | assert(d); | |
779 | ||
780 | /* Don't attempt to clean up the loop device anymore from this point on. Leave the clean-ing up to the kernel | |
781 | * itself, using the loop device "auto-clear" logic we already turned on when creating the device. */ | |
782 | ||
783 | d->relinquished = true; | |
784 | } | |
9dabc4fd | 785 | |
24d59aee DDM |
786 | void loop_device_unrelinquish(LoopDevice *d) { |
787 | assert(d); | |
788 | d->relinquished = false; | |
789 | } | |
790 | ||
4f0ad43e YW |
791 | int loop_device_open( |
792 | sd_device *dev, | |
7f52206a LP |
793 | int open_flags, |
794 | int lock_op, | |
795 | LoopDevice **ret) { | |
796 | ||
254d1313 | 797 | _cleanup_close_ int fd = -EBADF, lock_fd = -EBADF; |
4f0ad43e | 798 | _cleanup_free_ char *node = NULL, *backing_file = NULL; |
b26c39ad | 799 | struct loop_info64 info; |
ffcb3324 | 800 | uint64_t diskseq = 0; |
9dabc4fd | 801 | LoopDevice *d; |
4f0ad43e YW |
802 | const char *s; |
803 | dev_t devnum; | |
a8d8a619 | 804 | int r, nr = -1; |
9dabc4fd | 805 | |
4f0ad43e | 806 | assert(dev); |
e8c7c4d9 | 807 | assert(IN_SET(open_flags, O_RDWR, O_RDONLY)); |
9dabc4fd LP |
808 | assert(ret); |
809 | ||
4f0ad43e YW |
810 | /* Even if fd is provided through the argument in loop_device_open_from_fd(), we reopen the inode |
811 | * here, instead of keeping just a dup() clone of it around, since we want to ensure that the | |
812 | * O_DIRECT flag of the handle we keep is off, we have our own file index, and have the right | |
813 | * read/write mode in effect. */ | |
814 | fd = sd_device_open(dev, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags); | |
815 | if (fd < 0) | |
816 | return fd; | |
cc5bae6c | 817 | |
4f0ad43e YW |
818 | if ((lock_op & ~LOCK_NB) != LOCK_UN) { |
819 | lock_fd = open_lock_fd(fd, lock_op); | |
820 | if (lock_fd < 0) | |
821 | return lock_fd; | |
a8d8a619 YW |
822 | } |
823 | ||
4f0ad43e | 824 | if (ioctl(fd, LOOP_GET_STATUS64, &info) >= 0) { |
10c1b188 LP |
825 | #if HAVE_VALGRIND_MEMCHECK_H |
826 | /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */ | |
827 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
828 | #endif | |
b26c39ad | 829 | nr = info.lo_number; |
e77cab82 YW |
830 | |
831 | if (sd_device_get_sysattr_value(dev, "loop/backing_file", &s) >= 0) { | |
832 | backing_file = strdup(s); | |
833 | if (!backing_file) | |
834 | return -ENOMEM; | |
835 | } | |
a8d8a619 | 836 | } |
b26c39ad | 837 | |
4f0ad43e | 838 | r = fd_get_diskseq(fd, &diskseq); |
ffcb3324 YW |
839 | if (r < 0 && r != -EOPNOTSUPP) |
840 | return r; | |
841 | ||
4f0ad43e YW |
842 | r = sd_device_get_devnum(dev, &devnum); |
843 | if (r < 0) | |
844 | return r; | |
7f52206a | 845 | |
4f0ad43e | 846 | r = sd_device_get_devname(dev, &s); |
cc5bae6c YW |
847 | if (r < 0) |
848 | return r; | |
849 | ||
4f0ad43e YW |
850 | node = strdup(s); |
851 | if (!node) | |
cc5bae6c | 852 | return -ENOMEM; |
9dabc4fd LP |
853 | |
854 | d = new(LoopDevice, 1); | |
855 | if (!d) | |
856 | return -ENOMEM; | |
857 | ||
858 | *d = (LoopDevice) { | |
36d5eb0b | 859 | .n_ref = 1, |
a8d8a619 | 860 | .fd = TAKE_FD(fd), |
7f52206a | 861 | .lock_fd = TAKE_FD(lock_fd), |
b26c39ad | 862 | .nr = nr, |
4f0ad43e YW |
863 | .node = TAKE_PTR(node), |
864 | .dev = sd_device_ref(dev), | |
e77cab82 | 865 | .backing_file = TAKE_PTR(backing_file), |
9dabc4fd | 866 | .relinquished = true, /* It's not ours, don't try to destroy it when this object is freed */ |
4f0ad43e | 867 | .devno = devnum, |
ffcb3324 | 868 | .diskseq = diskseq, |
31c75fcc | 869 | .uevent_seqnum_not_before = UINT64_MAX, |
8ede1e86 | 870 | .timestamp_not_before = USEC_INFINITY, |
9dabc4fd LP |
871 | }; |
872 | ||
873 | *ret = d; | |
4f0ad43e YW |
874 | return 0; |
875 | } | |
876 | ||
877 | int loop_device_open_from_fd( | |
878 | int fd, | |
879 | int open_flags, | |
880 | int lock_op, | |
881 | LoopDevice **ret) { | |
882 | ||
883 | _cleanup_(sd_device_unrefp) sd_device *dev = NULL; | |
884 | int r; | |
885 | ||
886 | assert(fd >= 0); | |
887 | ||
888 | r = block_device_new_from_fd(fd, 0, &dev); | |
889 | if (r < 0) | |
890 | return r; | |
891 | ||
892 | return loop_device_open(dev, open_flags, lock_op, ret); | |
893 | } | |
894 | ||
895 | int loop_device_open_from_path( | |
896 | const char *path, | |
897 | int open_flags, | |
898 | int lock_op, | |
899 | LoopDevice **ret) { | |
900 | ||
901 | _cleanup_(sd_device_unrefp) sd_device *dev = NULL; | |
902 | int r; | |
903 | ||
904 | assert(path); | |
905 | ||
906 | r = block_device_new_from_path(path, 0, &dev); | |
907 | if (r < 0) | |
908 | return r; | |
909 | ||
910 | return loop_device_open(dev, open_flags, lock_op, ret); | |
9dabc4fd LP |
911 | } |
912 | ||
f1443709 LP |
913 | static int resize_partition(int partition_fd, uint64_t offset, uint64_t size) { |
914 | char sysfs[STRLEN("/sys/dev/block/:/partition") + 2*DECIMAL_STR_MAX(dev_t) + 1]; | |
ca822829 | 915 | _cleanup_free_ char *buffer = NULL; |
f1443709 | 916 | uint64_t current_offset, current_size, partno; |
254d1313 | 917 | _cleanup_close_ int whole_fd = -EBADF; |
f1443709 LP |
918 | struct stat st; |
919 | dev_t devno; | |
920 | int r; | |
921 | ||
922 | assert(partition_fd >= 0); | |
923 | ||
924 | /* Resizes the partition the loopback device refer to (assuming it refers to one instead of an actual | |
925 | * loopback device), and changes the offset, if needed. This is a fancy wrapper around | |
926 | * BLKPG_RESIZE_PARTITION. */ | |
927 | ||
928 | if (fstat(partition_fd, &st) < 0) | |
929 | return -errno; | |
930 | ||
931 | assert(S_ISBLK(st.st_mode)); | |
932 | ||
ed13feff | 933 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/partition", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
934 | r = read_one_line_file(sysfs, &buffer); |
935 | if (r == -ENOENT) /* not a partition, cannot resize */ | |
936 | return -ENOTTY; | |
937 | if (r < 0) | |
938 | return r; | |
939 | r = safe_atou64(buffer, &partno); | |
940 | if (r < 0) | |
941 | return r; | |
942 | ||
ed13feff | 943 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/start", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
944 | |
945 | buffer = mfree(buffer); | |
946 | r = read_one_line_file(sysfs, &buffer); | |
947 | if (r < 0) | |
948 | return r; | |
949 | r = safe_atou64(buffer, ¤t_offset); | |
950 | if (r < 0) | |
951 | return r; | |
952 | if (current_offset > UINT64_MAX/512U) | |
953 | return -EINVAL; | |
954 | current_offset *= 512U; | |
955 | ||
956 | if (ioctl(partition_fd, BLKGETSIZE64, ¤t_size) < 0) | |
957 | return -EINVAL; | |
958 | ||
959 | if (size == UINT64_MAX && offset == UINT64_MAX) | |
960 | return 0; | |
961 | if (current_size == size && current_offset == offset) | |
962 | return 0; | |
963 | ||
ed13feff | 964 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/../dev", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
965 | |
966 | buffer = mfree(buffer); | |
967 | r = read_one_line_file(sysfs, &buffer); | |
968 | if (r < 0) | |
969 | return r; | |
7176f06c | 970 | r = parse_devnum(buffer, &devno); |
f1443709 LP |
971 | if (r < 0) |
972 | return r; | |
973 | ||
ca822829 | 974 | whole_fd = r = device_open_from_devnum(S_IFBLK, devno, O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY, NULL); |
f1443709 LP |
975 | if (r < 0) |
976 | return r; | |
977 | ||
91e1ce1a LP |
978 | return block_device_resize_partition( |
979 | whole_fd, | |
980 | partno, | |
981 | offset == UINT64_MAX ? current_offset : offset, | |
982 | size == UINT64_MAX ? current_size : size); | |
f1443709 LP |
983 | } |
984 | ||
c37878fc LP |
985 | int loop_device_refresh_size(LoopDevice *d, uint64_t offset, uint64_t size) { |
986 | struct loop_info64 info; | |
ff27ef4b | 987 | |
9dabc4fd | 988 | assert(d); |
ff27ef4b | 989 | assert(d->fd >= 0); |
9dabc4fd | 990 | |
f1443709 LP |
991 | /* Changes the offset/start of the loop device relative to the beginning of the underlying file or |
992 | * block device. If this loop device actually refers to a partition and not a loopback device, we'll | |
993 | * try to adjust the partition offsets instead. | |
994 | * | |
995 | * If either offset or size is UINT64_MAX we won't change that parameter. */ | |
996 | ||
f1443709 LP |
997 | if (d->nr < 0) /* not a loopback device */ |
998 | return resize_partition(d->fd, offset, size); | |
999 | ||
c37878fc LP |
1000 | if (ioctl(d->fd, LOOP_GET_STATUS64, &info) < 0) |
1001 | return -errno; | |
1002 | ||
10c1b188 LP |
1003 | #if HAVE_VALGRIND_MEMCHECK_H |
1004 | /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */ | |
1005 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
1006 | #endif | |
1007 | ||
c37878fc LP |
1008 | if (size == UINT64_MAX && offset == UINT64_MAX) |
1009 | return 0; | |
1010 | if (info.lo_sizelimit == size && info.lo_offset == offset) | |
1011 | return 0; | |
1012 | ||
1013 | if (size != UINT64_MAX) | |
1014 | info.lo_sizelimit = size; | |
1015 | if (offset != UINT64_MAX) | |
1016 | info.lo_offset = offset; | |
1017 | ||
7c248223 | 1018 | return RET_NERRNO(ioctl(d->fd, LOOP_SET_STATUS64, &info)); |
9dabc4fd | 1019 | } |
441ec804 LP |
1020 | |
1021 | int loop_device_flock(LoopDevice *d, int operation) { | |
7f52206a | 1022 | assert(IN_SET(operation & ~LOCK_NB, LOCK_UN, LOCK_SH, LOCK_EX)); |
441ec804 LP |
1023 | assert(d); |
1024 | ||
7f52206a LP |
1025 | /* When unlocking just close the lock fd */ |
1026 | if ((operation & ~LOCK_NB) == LOCK_UN) { | |
1027 | d->lock_fd = safe_close(d->lock_fd); | |
1028 | return 0; | |
1029 | } | |
1030 | ||
1031 | /* If we had no lock fd so far, create one and lock it right-away */ | |
1032 | if (d->lock_fd < 0) { | |
1033 | assert(d->fd >= 0); | |
1034 | ||
1035 | d->lock_fd = open_lock_fd(d->fd, operation); | |
1036 | if (d->lock_fd < 0) | |
1037 | return d->lock_fd; | |
1038 | ||
1039 | return 0; | |
1040 | } | |
441ec804 | 1041 | |
7f52206a LP |
1042 | /* Otherwise change the current lock mode on the existing fd */ |
1043 | return RET_NERRNO(flock(d->lock_fd, operation)); | |
441ec804 | 1044 | } |
8dbc208c LP |
1045 | |
1046 | int loop_device_sync(LoopDevice *d) { | |
1047 | assert(d); | |
ff27ef4b | 1048 | assert(d->fd >= 0); |
8dbc208c LP |
1049 | |
1050 | /* We also do this implicitly in loop_device_unref(). Doing this explicitly here has the benefit that | |
1051 | * we can check the return value though. */ | |
1052 | ||
7c248223 | 1053 | return RET_NERRNO(fsync(d->fd)); |
8dbc208c | 1054 | } |