]>
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" |
021bf175 | 20 | #include "device-util.h" |
7176f06c | 21 | #include "devnum-util.h" |
e8c7c4d9 | 22 | #include "env-util.h" |
b0a94268 | 23 | #include "errno-util.h" |
8c1be37e | 24 | #include "fd-util.h" |
f1443709 | 25 | #include "fileio.h" |
8c1be37e | 26 | #include "loop-util.h" |
86c1c1f3 | 27 | #include "missing_loop.h" |
f1443709 | 28 | #include "parse-util.h" |
b202ec20 | 29 | #include "random-util.h" |
3cc44114 | 30 | #include "stat-util.h" |
f1443709 | 31 | #include "stdio-util.h" |
f2d9213f | 32 | #include "string-util.h" |
021bf175 | 33 | #include "tmpfile-util.h" |
8c1be37e | 34 | |
e8af3bfd | 35 | static void cleanup_clear_loop_close(int *fd) { |
86c1c1f3 LP |
36 | if (*fd < 0) |
37 | return; | |
38 | ||
39 | (void) ioctl(*fd, LOOP_CLR_FD); | |
40 | (void) safe_close(*fd); | |
41 | } | |
42 | ||
021bf175 LP |
43 | static int loop_is_bound(int fd) { |
44 | struct loop_info64 info; | |
45 | ||
46 | assert(fd >= 0); | |
47 | ||
48 | if (ioctl(fd, LOOP_GET_STATUS64, &info) < 0) { | |
49 | if (errno == ENXIO) | |
50 | return false; /* not bound! */ | |
51 | ||
52 | return -errno; | |
53 | } | |
54 | ||
55 | return true; /* bound! */ | |
56 | } | |
57 | ||
31c75fcc LP |
58 | static int get_current_uevent_seqnum(uint64_t *ret) { |
59 | _cleanup_free_ char *p = NULL; | |
60 | int r; | |
61 | ||
62 | r = read_full_virtual_file("/sys/kernel/uevent_seqnum", &p, NULL); | |
63 | if (r < 0) | |
64 | return log_debug_errno(r, "Failed to read current uevent sequence number: %m"); | |
65 | ||
a145f8c0 | 66 | r = safe_atou64(strstrip(p), ret); |
31c75fcc LP |
67 | if (r < 0) |
68 | return log_debug_errno(r, "Failed to parse current uevent sequence number: %s", p); | |
69 | ||
70 | return 0; | |
71 | } | |
72 | ||
021bf175 LP |
73 | static int device_has_block_children(sd_device *d) { |
74 | _cleanup_(sd_device_enumerator_unrefp) sd_device_enumerator *e = NULL; | |
7ffc7f3f | 75 | const char *main_ss, *main_dt; |
021bf175 LP |
76 | int r; |
77 | ||
78 | assert(d); | |
79 | ||
80 | /* Checks if the specified device currently has block device children (i.e. partition block | |
81 | * devices). */ | |
82 | ||
7ffc7f3f | 83 | r = sd_device_get_subsystem(d, &main_ss); |
021bf175 LP |
84 | if (r < 0) |
85 | return r; | |
86 | ||
7ffc7f3f LP |
87 | if (!streq(main_ss, "block")) |
88 | return -EINVAL; | |
89 | ||
90 | r = sd_device_get_devtype(d, &main_dt); | |
021bf175 LP |
91 | if (r < 0) |
92 | return r; | |
93 | ||
7ffc7f3f | 94 | if (!streq(main_dt, "disk")) /* Refuse invocation on partition block device, insist on "whole" device */ |
021bf175 LP |
95 | return -EINVAL; |
96 | ||
97 | r = sd_device_enumerator_new(&e); | |
98 | if (r < 0) | |
99 | return r; | |
100 | ||
101 | r = sd_device_enumerator_allow_uninitialized(e); | |
102 | if (r < 0) | |
103 | return r; | |
104 | ||
105 | r = sd_device_enumerator_add_match_parent(e, d); | |
106 | if (r < 0) | |
107 | return r; | |
108 | ||
5c467ef4 YW |
109 | r = sd_device_enumerator_add_match_subsystem(e, "block", /* match = */ true); |
110 | if (r < 0) | |
111 | return r; | |
021bf175 | 112 | |
5c467ef4 YW |
113 | r = sd_device_enumerator_add_match_property(e, "DEVTYPE", "partition"); |
114 | if (r < 0) | |
115 | return r; | |
021bf175 | 116 | |
5c467ef4 | 117 | return !!sd_device_enumerator_get_device_first(e); |
021bf175 LP |
118 | } |
119 | ||
7f52206a LP |
120 | static int open_lock_fd(int primary_fd, int operation) { |
121 | int lock_fd; | |
122 | ||
123 | assert(primary_fd >= 0); | |
124 | ||
125 | lock_fd = fd_reopen(primary_fd, O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY); | |
126 | if (lock_fd < 0) | |
127 | return lock_fd; | |
128 | if (flock(lock_fd, operation) < 0) | |
129 | return -errno; | |
130 | ||
131 | return lock_fd; | |
132 | } | |
133 | ||
95c50092 LP |
134 | static int loop_configure( |
135 | int fd, | |
021bf175 | 136 | int nr, |
95c50092 | 137 | const struct loop_config *c, |
31c75fcc | 138 | bool *try_loop_configure, |
8ede1e86 | 139 | uint64_t *ret_seqnum_not_before, |
7f52206a LP |
140 | usec_t *ret_timestamp_not_before, |
141 | int *ret_lock_fd) { | |
95c50092 | 142 | |
021bf175 LP |
143 | _cleanup_(sd_device_unrefp) sd_device *d = NULL; |
144 | _cleanup_free_ char *sysname = NULL; | |
738f29cb | 145 | _cleanup_close_ int lock_fd = -1; |
e8c7c4d9 | 146 | struct loop_info64 info_copy; |
31c75fcc | 147 | uint64_t seqnum; |
8ede1e86 | 148 | usec_t timestamp; |
86c1c1f3 LP |
149 | int r; |
150 | ||
151 | assert(fd >= 0); | |
021bf175 | 152 | assert(nr >= 0); |
86c1c1f3 | 153 | assert(c); |
95c50092 LP |
154 | assert(try_loop_configure); |
155 | ||
021bf175 LP |
156 | if (asprintf(&sysname, "loop%i", nr) < 0) |
157 | return -ENOMEM; | |
158 | ||
159 | r = sd_device_new_from_subsystem_sysname(&d, "block", sysname); | |
160 | if (r < 0) | |
161 | return r; | |
162 | ||
163 | /* Let's lock the device before we do anything. We take the BSD lock on a second, separately opened | |
164 | * fd for the device. udev after all watches for close() events (specifically IN_CLOSE_WRITE) on | |
165 | * block devices to reprobe them, hence by having a separate fd we will later close() we can ensure | |
166 | * we trigger udev after everything is done. If we'd lock our own fd instead and keep it open for a | |
167 | * long time udev would possibly never run on it again, even though the fd is unlocked, simply | |
168 | * because we never close() it. It also has the nice benefit we can use the _cleanup_close_ logic to | |
169 | * automatically release the lock, after we are done. */ | |
7f52206a | 170 | lock_fd = open_lock_fd(fd, LOCK_EX); |
021bf175 LP |
171 | if (lock_fd < 0) |
172 | return lock_fd; | |
021bf175 LP |
173 | |
174 | /* Let's see if the device is really detached, i.e. currently has no associated partition block | |
175 | * devices. On various kernels (such as 5.8) it is possible to have a loopback block device that | |
176 | * superficially is detached but still has partition block devices associated for it. They only go | |
177 | * away when the device is reattached. (Yes, LOOP_CLR_FD doesn't work then, because officially | |
178 | * nothing is attached and LOOP_CTL_REMOVE doesn't either, since it doesn't care about partition | |
179 | * block devices. */ | |
180 | r = device_has_block_children(d); | |
181 | if (r < 0) | |
182 | return r; | |
183 | if (r > 0) { | |
184 | r = loop_is_bound(fd); | |
185 | if (r < 0) | |
186 | return r; | |
187 | if (r > 0) | |
188 | return -EBUSY; | |
189 | ||
190 | return -EUCLEAN; /* Bound but children? Tell caller to reattach something so that the | |
191 | * partition block devices are gone too. */ | |
192 | } | |
193 | ||
95c50092 | 194 | if (*try_loop_configure) { |
31c75fcc LP |
195 | /* Acquire uevent seqnum immediately before attaching the loopback device. This allows |
196 | * callers to ignore all uevents with a seqnum before this one, if they need to associate | |
197 | * uevent with this attachment. Doing so isn't race-free though, as uevents that happen in | |
198 | * the window between this reading of the seqnum, and the LOOP_CONFIGURE call might still be | |
199 | * mistaken as originating from our attachment, even though might be caused by an earlier | |
200 | * use. But doing this at least shortens the race window a bit. */ | |
201 | r = get_current_uevent_seqnum(&seqnum); | |
202 | if (r < 0) | |
203 | return r; | |
8ede1e86 | 204 | timestamp = now(CLOCK_MONOTONIC); |
31c75fcc | 205 | |
95c50092 LP |
206 | if (ioctl(fd, LOOP_CONFIGURE, c) < 0) { |
207 | /* Do fallback only if LOOP_CONFIGURE is not supported, propagate all other | |
208 | * errors. Note that the kernel is weird: non-existing ioctls currently return EINVAL | |
209 | * rather than ENOTTY on loopback block devices. They should fix that in the kernel, | |
210 | * but in the meantime we accept both here. */ | |
211 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && errno != EINVAL) | |
212 | return -errno; | |
86c1c1f3 | 213 | |
95c50092 LP |
214 | *try_loop_configure = false; |
215 | } else { | |
216 | bool good = true; | |
217 | ||
218 | if (c->info.lo_sizelimit != 0) { | |
219 | /* Kernel 5.8 vanilla doesn't properly propagate the size limit into the | |
220 | * block device. If it's used, let's immediately check if it had the desired | |
221 | * effect hence. And if not use classic LOOP_SET_STATUS64. */ | |
222 | uint64_t z; | |
223 | ||
224 | if (ioctl(fd, BLKGETSIZE64, &z) < 0) { | |
225 | r = -errno; | |
226 | goto fail; | |
227 | } | |
228 | ||
229 | if (z != c->info.lo_sizelimit) { | |
230 | log_debug("LOOP_CONFIGURE is broken, doesn't honour .lo_sizelimit. Falling back to LOOP_SET_STATUS64."); | |
231 | good = false; | |
232 | } | |
bb2551bd | 233 | } |
86c1c1f3 | 234 | |
95c50092 LP |
235 | if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_PARTSCAN)) { |
236 | /* Kernel 5.8 vanilla doesn't properly propagate the partition scanning flag | |
237 | * into the block device. Let's hence verify if things work correctly here | |
238 | * before returning. */ | |
239 | ||
240 | r = blockdev_partscan_enabled(fd); | |
241 | if (r < 0) | |
242 | goto fail; | |
243 | if (r == 0) { | |
244 | log_debug("LOOP_CONFIGURE is broken, doesn't honour LO_FLAGS_PARTSCAN. Falling back to LOOP_SET_STATUS64."); | |
245 | good = false; | |
246 | } | |
bb2551bd | 247 | } |
86c1c1f3 | 248 | |
95c50092 LP |
249 | if (!good) { |
250 | /* LOOP_CONFIGURE doesn't work. Remember that. */ | |
251 | *try_loop_configure = false; | |
252 | ||
253 | /* We return EBUSY here instead of retrying immediately with LOOP_SET_FD, | |
254 | * because LOOP_CLR_FD is async: if the operation cannot be executed right | |
255 | * away it just sets the autoclear flag on the device. This means there's a | |
256 | * good chance we cannot actually reuse the loopback device right-away. Hence | |
257 | * let's assume it's busy, avoid the trouble and let the calling loop call us | |
258 | * again with a new, likely unused device. */ | |
259 | r = -EBUSY; | |
bb2551bd | 260 | goto fail; |
bb2551bd | 261 | } |
bb2551bd | 262 | |
7f52206a | 263 | goto success; |
95c50092 | 264 | } |
86c1c1f3 LP |
265 | } |
266 | ||
31c75fcc LP |
267 | /* Let's read the seqnum again, to shorten the window. */ |
268 | r = get_current_uevent_seqnum(&seqnum); | |
269 | if (r < 0) | |
270 | return r; | |
8ede1e86 | 271 | timestamp = now(CLOCK_MONOTONIC); |
31c75fcc | 272 | |
738f29cb LP |
273 | /* Since kernel commit 5db470e229e22b7eda6e23b5566e532c96fb5bc3 (kernel v5.0) the LOOP_SET_STATUS64 |
274 | * ioctl can return EAGAIN in case we change the lo_offset field, if someone else is accessing the | |
275 | * block device while we try to reconfigure it. This is a pretty common case, since udev might | |
276 | * instantly start probing the device as soon as we attach an fd to it. Hence handle it in two ways: | |
273d76f4 | 277 | * first, let's take the BSD lock to ensure that udev will not step in between the point in |
738f29cb LP |
278 | * time where we attach the fd and where we reconfigure the device. Secondly, let's wait 50ms on |
279 | * EAGAIN and retry. The former should be an efficient mechanism to avoid we have to wait 50ms | |
280 | * needlessly if we are just racing against udev. The latter is protection against all other cases, | |
021bf175 | 281 | * i.e. peers that do not take the BSD lock. */ |
738f29cb | 282 | |
86c1c1f3 LP |
283 | if (ioctl(fd, LOOP_SET_FD, c->fd) < 0) |
284 | return -errno; | |
285 | ||
e8c7c4d9 LP |
286 | /* Only some of the flags LOOP_CONFIGURE can set are also settable via LOOP_SET_STATUS64, hence mask |
287 | * them out. */ | |
288 | info_copy = c->info; | |
289 | info_copy.lo_flags &= LOOP_SET_STATUS_SETTABLE_FLAGS; | |
290 | ||
738f29cb | 291 | for (unsigned n_attempts = 0;;) { |
e8c7c4d9 | 292 | if (ioctl(fd, LOOP_SET_STATUS64, &info_copy) >= 0) |
738f29cb LP |
293 | break; |
294 | if (errno != EAGAIN || ++n_attempts >= 64) { | |
9d72a3cf | 295 | r = log_debug_errno(errno, "Failed to configure loopback block device: %m"); |
738f29cb LP |
296 | goto fail; |
297 | } | |
298 | ||
b202ec20 LP |
299 | /* Sleep some random time, but at least 10ms, at most 250ms. Increase the delay the more |
300 | * failed attempts we see */ | |
301 | (void) usleep(UINT64_C(10) * USEC_PER_MSEC + | |
b0dbffd8 | 302 | random_u64_range(UINT64_C(240) * USEC_PER_MSEC * n_attempts/64)); |
e8af3bfd | 303 | } |
86c1c1f3 | 304 | |
b9a9748a LP |
305 | /* Work around a kernel bug, where changing offset/size of the loopback device doesn't correctly |
306 | * invalidate the buffer cache. For details see: | |
307 | * | |
308 | * https://android.googlesource.com/platform/system/apex/+/bef74542fbbb4cd629793f4efee8e0053b360570 | |
309 | * | |
310 | * This was fixed in kernel 5.0, see: | |
311 | * | |
312 | * https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5db470e229e22b7eda6e23b5566e532c96fb5bc3 | |
313 | * | |
314 | * We'll run the work-around here in the legacy LOOP_SET_STATUS64 codepath. In the LOOP_CONFIGURE | |
315 | * codepath above it should not be necessary. */ | |
316 | if (c->info.lo_offset != 0 || c->info.lo_sizelimit != 0) | |
317 | if (ioctl(fd, BLKFLSBUF, 0) < 0) | |
318 | log_debug_errno(errno, "Failed to issue BLKFLSBUF ioctl, ignoring: %m"); | |
319 | ||
e8c7c4d9 LP |
320 | /* LO_FLAGS_DIRECT_IO is a flags we need to configure via explicit ioctls. */ |
321 | if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_DIRECT_IO)) { | |
322 | unsigned long b = 1; | |
323 | ||
324 | if (ioctl(fd, LOOP_SET_DIRECT_IO, b) < 0) | |
325 | log_debug_errno(errno, "Failed to enable direct IO mode on loopback device /dev/loop%i, ignoring: %m", nr); | |
326 | } | |
327 | ||
7f52206a | 328 | success: |
31c75fcc LP |
329 | if (ret_seqnum_not_before) |
330 | *ret_seqnum_not_before = seqnum; | |
8ede1e86 LP |
331 | if (ret_timestamp_not_before) |
332 | *ret_timestamp_not_before = timestamp; | |
7f52206a LP |
333 | if (ret_lock_fd) |
334 | *ret_lock_fd = TAKE_FD(lock_fd); | |
31c75fcc | 335 | |
86c1c1f3 LP |
336 | return 0; |
337 | ||
338 | fail: | |
339 | (void) ioctl(fd, LOOP_CLR_FD); | |
340 | return r; | |
e8af3bfd ZJS |
341 | } |
342 | ||
021bf175 LP |
343 | static int attach_empty_file(int loop, int nr) { |
344 | _cleanup_close_ int fd = -1; | |
345 | ||
346 | /* So here's the thing: on various kernels (5.8 at least) loop block devices might enter a state | |
347 | * where they are detached but nonetheless have partitions, when used heavily. Accessing these | |
348 | * partitions results in immediatey IO errors. There's no pretty way to get rid of them | |
349 | * again. Neither LOOP_CLR_FD nor LOOP_CTL_REMOVE suffice (see above). What does work is to | |
350 | * reassociate them with a new fd however. This is what we do here hence: we associate the devices | |
377a9545 | 351 | * with an empty file (i.e. an image that definitely has no partitions). We then immediately clear it |
021bf175 LP |
352 | * again. This suffices to make the partitions go away. Ugly but appears to work. */ |
353 | ||
354 | log_debug("Found unattached loopback block device /dev/loop%i with partitions. Attaching empty file to remove them.", nr); | |
355 | ||
356 | fd = open_tmpfile_unlinkable(NULL, O_RDONLY); | |
357 | if (fd < 0) | |
358 | return fd; | |
359 | ||
360 | if (flock(loop, LOCK_EX) < 0) | |
361 | return -errno; | |
362 | ||
363 | if (ioctl(loop, LOOP_SET_FD, fd) < 0) | |
364 | return -errno; | |
365 | ||
366 | if (ioctl(loop, LOOP_SET_STATUS64, &(struct loop_info64) { | |
367 | .lo_flags = LO_FLAGS_READ_ONLY| | |
368 | LO_FLAGS_AUTOCLEAR| | |
369 | LO_FLAGS_PARTSCAN, /* enable partscan, so that the partitions really go away */ | |
370 | }) < 0) | |
371 | return -errno; | |
372 | ||
373 | if (ioctl(loop, LOOP_CLR_FD) < 0) | |
374 | return -errno; | |
375 | ||
376 | /* The caller is expected to immediately close the loopback device after this, so that the BSD lock | |
377 | * is released, and udev sees the changes. */ | |
378 | return 0; | |
379 | } | |
380 | ||
e8c7c4d9 | 381 | static int loop_device_make_internal( |
ed9eeb7b LP |
382 | int fd, |
383 | int open_flags, | |
384 | uint64_t offset, | |
385 | uint64_t size, | |
386 | uint32_t loop_flags, | |
7f52206a | 387 | int lock_op, |
ed9eeb7b | 388 | LoopDevice **ret) { |
8c1be37e | 389 | |
7f52206a | 390 | _cleanup_close_ int direct_io_fd = -1, lock_fd = -1; |
8c1be37e | 391 | _cleanup_free_ char *loopdev = NULL; |
95c50092 | 392 | bool try_loop_configure = true; |
86c1c1f3 | 393 | struct loop_config config; |
50d04699 | 394 | LoopDevice *d = NULL; |
31c75fcc | 395 | uint64_t seqnum = UINT64_MAX; |
8ede1e86 | 396 | usec_t timestamp = USEC_INFINITY; |
e8c7c4d9 | 397 | int nr = -1, r, f_flags; |
8c1be37e | 398 | struct stat st; |
8c1be37e LP |
399 | |
400 | assert(fd >= 0); | |
401 | assert(ret); | |
402 | assert(IN_SET(open_flags, O_RDWR, O_RDONLY)); | |
403 | ||
404 | if (fstat(fd, &st) < 0) | |
405 | return -errno; | |
406 | ||
407 | if (S_ISBLK(st.st_mode)) { | |
7f52206a LP |
408 | if (lock_op != LOCK_UN) { |
409 | lock_fd = open_lock_fd(fd, lock_op); | |
410 | if (lock_fd < 0) | |
411 | return lock_fd; | |
412 | } | |
413 | ||
86c1c1f3 | 414 | if (ioctl(fd, LOOP_GET_STATUS64, &config.info) >= 0) { |
b26c39ad | 415 | /* Oh! This is a loopback device? That's interesting! */ |
10c1b188 LP |
416 | |
417 | #if HAVE_VALGRIND_MEMCHECK_H | |
418 | /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */ | |
86c1c1f3 | 419 | VALGRIND_MAKE_MEM_DEFINED(&config.info, sizeof(config.info)); |
10c1b188 | 420 | #endif |
86c1c1f3 | 421 | nr = config.info.lo_number; |
b26c39ad LP |
422 | |
423 | if (asprintf(&loopdev, "/dev/loop%i", nr) < 0) | |
424 | return -ENOMEM; | |
425 | } | |
426 | ||
ed9eeb7b | 427 | if (offset == 0 && IN_SET(size, 0, UINT64_MAX)) { |
ba5450f4 | 428 | _cleanup_close_ int copy = -1; |
bcef1743 | 429 | uint64_t diskseq = 0; |
8c1be37e | 430 | |
d7654742 LP |
431 | /* If this is already a block device and we are supposed to cover the whole of it |
432 | * then store an fd to the original open device node — and do not actually create an | |
433 | * unnecessary loopback device for it. Note that we reopen the inode here, instead of | |
434 | * keeping just a dup() clone of it around, since we want to ensure that the O_DIRECT | |
435 | * flag of the handle we keep is off, we have our own file index, and have the right | |
436 | * read/write mode in effect. */ | |
437 | ||
438 | copy = fd_reopen(fd, open_flags|O_NONBLOCK|O_CLOEXEC|O_NOCTTY); | |
ed9eeb7b | 439 | if (copy < 0) |
d7654742 | 440 | return copy; |
8c1be37e | 441 | |
7e93a658 | 442 | r = fd_get_diskseq(copy, &diskseq); |
bcef1743 LB |
443 | if (r < 0 && r != -EOPNOTSUPP) |
444 | return r; | |
445 | ||
ed9eeb7b LP |
446 | d = new(LoopDevice, 1); |
447 | if (!d) | |
448 | return -ENOMEM; | |
ed9eeb7b | 449 | *d = (LoopDevice) { |
ba5450f4 | 450 | .fd = TAKE_FD(copy), |
7f52206a | 451 | .lock_fd = TAKE_FD(lock_fd), |
b26c39ad LP |
452 | .nr = nr, |
453 | .node = TAKE_PTR(loopdev), | |
ed9eeb7b | 454 | .relinquished = true, /* It's not allocated by us, don't destroy it when this object is freed */ |
f3859d5f | 455 | .devno = st.st_rdev, |
bcef1743 | 456 | .diskseq = diskseq, |
31c75fcc | 457 | .uevent_seqnum_not_before = UINT64_MAX, |
8ede1e86 | 458 | .timestamp_not_before = USEC_INFINITY, |
ed9eeb7b LP |
459 | }; |
460 | ||
461 | *ret = d; | |
462 | return d->fd; | |
463 | } | |
464 | } else { | |
465 | r = stat_verify_regular(&st); | |
466 | if (r < 0) | |
467 | return r; | |
8c1be37e LP |
468 | } |
469 | ||
e8c7c4d9 LP |
470 | f_flags = fcntl(fd, F_GETFL); |
471 | if (f_flags < 0) | |
472 | return -errno; | |
473 | ||
474 | if (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) != FLAGS_SET(f_flags, O_DIRECT)) { | |
475 | /* If LO_FLAGS_DIRECT_IO is requested, then make sure we have the fd open with O_DIRECT, as | |
476 | * that's required. Conversely, if it's off require that O_DIRECT is off too (that's because | |
477 | * new kernels will implicitly enable LO_FLAGS_DIRECT_IO if O_DIRECT is set). | |
478 | * | |
479 | * Our intention here is that LO_FLAGS_DIRECT_IO is the primary knob, and O_DIRECT derived | |
480 | * from that automatically. */ | |
481 | ||
482 | direct_io_fd = fd_reopen(fd, (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0)|O_CLOEXEC|O_NONBLOCK|open_flags); | |
483 | if (direct_io_fd < 0) { | |
484 | if (!FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO)) | |
485 | return log_debug_errno(errno, "Failed to reopen file descriptor without O_DIRECT: %m"); | |
486 | ||
487 | /* Some file systems might not support O_DIRECT, let's gracefully continue without it then. */ | |
488 | log_debug_errno(errno, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without."); | |
489 | loop_flags &= ~LO_FLAGS_DIRECT_IO; | |
490 | } else | |
491 | fd = direct_io_fd; /* From now on, operate on our new O_DIRECT fd */ | |
492 | } | |
493 | ||
e8af3bfd ZJS |
494 | _cleanup_close_ int control = -1; |
495 | _cleanup_(cleanup_clear_loop_close) int loop_with_fd = -1; | |
496 | ||
8c1be37e LP |
497 | control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK); |
498 | if (control < 0) | |
499 | return -errno; | |
500 | ||
86c1c1f3 LP |
501 | config = (struct loop_config) { |
502 | .fd = fd, | |
503 | .info = { | |
504 | /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */ | |
0950526a | 505 | .lo_flags = (loop_flags & ~LO_FLAGS_READ_ONLY) | ((open_flags & O_ACCMODE) == O_RDONLY ? LO_FLAGS_READ_ONLY : 0) | LO_FLAGS_AUTOCLEAR, |
86c1c1f3 LP |
506 | .lo_offset = offset, |
507 | .lo_sizelimit = size == UINT64_MAX ? 0 : size, | |
508 | }, | |
509 | }; | |
510 | ||
0f6519d4 LP |
511 | /* Loop around LOOP_CTL_GET_FREE, since at the moment we attempt to open the returned device it might |
512 | * be gone already, taken by somebody else racing against us. */ | |
e8af3bfd ZJS |
513 | for (unsigned n_attempts = 0;;) { |
514 | _cleanup_close_ int loop = -1; | |
515 | ||
cc530466 LP |
516 | /* Let's take a lock on the control device first. On a busy system, where many programs |
517 | * attempt to allocate a loopback device at the same time, we might otherwise keep looping | |
518 | * around relatively heavy operations: asking for a free loopback device, then opening it, | |
519 | * validating it, attaching something to it. Let's serialize this whole operation, to make | |
520 | * unnecessary busywork less likely. Note that this is just something we do to optimize our | |
521 | * own code (and whoever else decides to use LOCK_EX locks for this), taking this lock is not | |
522 | * necessary, it just means it's less likely we have to iterate through this loop again and | |
523 | * again if our own code races against our own code. */ | |
524 | if (flock(control, LOCK_EX) < 0) | |
525 | return -errno; | |
526 | ||
0f6519d4 LP |
527 | nr = ioctl(control, LOOP_CTL_GET_FREE); |
528 | if (nr < 0) | |
529 | return -errno; | |
8c1be37e | 530 | |
0f6519d4 LP |
531 | if (asprintf(&loopdev, "/dev/loop%i", nr) < 0) |
532 | return -ENOMEM; | |
8c1be37e | 533 | |
0f6519d4 | 534 | loop = open(loopdev, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags); |
01813148 ZJS |
535 | if (loop < 0) { |
536 | /* Somebody might've gotten the same number from the kernel, used the device, | |
537 | * and called LOOP_CTL_REMOVE on it. Let's retry with a new number. */ | |
49043f81 | 538 | if (!ERRNO_IS_DEVICE_ABSENT(errno)) |
01813148 ZJS |
539 | return -errno; |
540 | } else { | |
7f52206a | 541 | r = loop_configure(loop, nr, &config, &try_loop_configure, &seqnum, ×tamp, &lock_fd); |
86c1c1f3 | 542 | if (r >= 0) { |
01813148 ZJS |
543 | loop_with_fd = TAKE_FD(loop); |
544 | break; | |
545 | } | |
021bf175 LP |
546 | if (r == -EUCLEAN) { |
547 | /* Make left-over partition disappear hack (see above) */ | |
548 | r = attach_empty_file(loop, nr); | |
549 | if (r < 0 && r != -EBUSY) | |
550 | return r; | |
551 | } else if (r != -EBUSY) | |
86c1c1f3 | 552 | return r; |
e8af3bfd | 553 | } |
01813148 | 554 | |
cc530466 LP |
555 | /* OK, this didn't work, let's try again a bit later, but first release the lock on the |
556 | * control device */ | |
557 | if (flock(control, LOCK_UN) < 0) | |
558 | return -errno; | |
559 | ||
e8af3bfd ZJS |
560 | if (++n_attempts >= 64) /* Give up eventually */ |
561 | return -EBUSY; | |
0f6519d4 | 562 | |
3e921057 LP |
563 | /* Now close the loop device explicitly. This will release any lock acquired by |
564 | * attach_empty_file() or similar, while we sleep below. */ | |
565 | loop = safe_close(loop); | |
0f6519d4 | 566 | loopdev = mfree(loopdev); |
b202ec20 LP |
567 | |
568 | /* Wait some random time, to make collision less likely. Let's pick a random time in the | |
569 | * range 0ms…250ms, linearly scaled by the number of failed attempts. */ | |
b0dbffd8 LP |
570 | (void) usleep(random_u64_range(UINT64_C(10) * USEC_PER_MSEC + |
571 | UINT64_C(240) * USEC_PER_MSEC * n_attempts/64)); | |
0f6519d4 | 572 | } |
8c1be37e | 573 | |
e8c7c4d9 LP |
574 | if (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO)) { |
575 | struct loop_info64 info; | |
576 | ||
577 | if (ioctl(loop_with_fd, LOOP_GET_STATUS64, &info) < 0) | |
578 | return -errno; | |
579 | ||
580 | #if HAVE_VALGRIND_MEMCHECK_H | |
581 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
582 | #endif | |
583 | ||
584 | /* On older kernels (<= 5.3) it was necessary to set the block size of the loopback block | |
585 | * device to the logical block size of the underlying file system. Since there was no nice | |
586 | * way to query the value, we are not bothering to do this however. On newer kernels the | |
587 | * block size is propagated automatically and does not require intervention from us. We'll | |
588 | * check here if enabling direct IO worked, to make this easily debuggable however. | |
589 | * | |
590 | * (Should anyone really care and actually wants direct IO on old kernels: it might be worth | |
591 | * enabling direct IO with iteratively larger block sizes until it eventually works.) */ | |
592 | if (!FLAGS_SET(info.lo_flags, LO_FLAGS_DIRECT_IO)) | |
593 | log_debug("Could not enable direct IO mode, proceeding in buffered IO mode."); | |
594 | } | |
595 | ||
f3859d5f LP |
596 | if (fstat(loop_with_fd, &st) < 0) |
597 | return -errno; | |
598 | assert(S_ISBLK(st.st_mode)); | |
599 | ||
bcef1743 | 600 | uint64_t diskseq = 0; |
7e93a658 | 601 | r = fd_get_diskseq(loop_with_fd, &diskseq); |
bcef1743 LB |
602 | if (r < 0 && r != -EOPNOTSUPP) |
603 | return r; | |
604 | ||
7f52206a LP |
605 | switch (lock_op & ~LOCK_NB) { |
606 | case LOCK_EX: /* Already in effect */ | |
607 | break; | |
608 | case LOCK_SH: /* Downgrade */ | |
609 | if (flock(lock_fd, lock_op) < 0) | |
610 | return -errno; | |
611 | break; | |
612 | case LOCK_UN: /* Release */ | |
613 | lock_fd = safe_close(lock_fd); | |
614 | break; | |
615 | default: | |
616 | assert_not_reached(); | |
617 | } | |
618 | ||
8c1be37e | 619 | d = new(LoopDevice, 1); |
e8af3bfd ZJS |
620 | if (!d) |
621 | return -ENOMEM; | |
8c1be37e | 622 | *d = (LoopDevice) { |
e8af3bfd | 623 | .fd = TAKE_FD(loop_with_fd), |
7f52206a | 624 | .lock_fd = TAKE_FD(lock_fd), |
1cc6c93a | 625 | .node = TAKE_PTR(loopdev), |
8c1be37e | 626 | .nr = nr, |
f3859d5f | 627 | .devno = st.st_rdev, |
bcef1743 | 628 | .diskseq = diskseq, |
31c75fcc | 629 | .uevent_seqnum_not_before = seqnum, |
8ede1e86 | 630 | .timestamp_not_before = timestamp, |
8c1be37e LP |
631 | }; |
632 | ||
3b195f63 LP |
633 | log_debug("Successfully acquired %s, devno=%u:%u, nr=%i, diskseq=%" PRIu64, |
634 | d->node, | |
635 | major(d->devno), minor(d->devno), | |
636 | d->nr, | |
637 | d->diskseq); | |
638 | ||
8c1be37e | 639 | *ret = d; |
38bd449f | 640 | return d->fd; |
8c1be37e LP |
641 | } |
642 | ||
e8c7c4d9 LP |
643 | static uint32_t loop_flags_mangle(uint32_t loop_flags) { |
644 | int r; | |
645 | ||
646 | r = getenv_bool("SYSTEMD_LOOP_DIRECT_IO"); | |
647 | if (r < 0 && r != -ENXIO) | |
648 | log_debug_errno(r, "Failed to parse $SYSTEMD_LOOP_DIRECT_IO, ignoring: %m"); | |
649 | ||
bfd08445 | 650 | 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 |
651 | } |
652 | ||
653 | int loop_device_make( | |
654 | int fd, | |
655 | int open_flags, | |
656 | uint64_t offset, | |
657 | uint64_t size, | |
658 | uint32_t loop_flags, | |
7f52206a | 659 | int lock_op, |
e8c7c4d9 LP |
660 | LoopDevice **ret) { |
661 | ||
662 | assert(fd >= 0); | |
663 | assert(ret); | |
e8c7c4d9 LP |
664 | |
665 | return loop_device_make_internal( | |
666 | fd, | |
667 | open_flags, | |
668 | offset, | |
669 | size, | |
bfd08445 | 670 | loop_flags_mangle(loop_flags), |
7f52206a | 671 | lock_op, |
e8c7c4d9 LP |
672 | ret); |
673 | } | |
674 | ||
79e8393a LP |
675 | int loop_device_make_by_path( |
676 | const char *path, | |
677 | int open_flags, | |
678 | uint32_t loop_flags, | |
7f52206a | 679 | int lock_op, |
79e8393a LP |
680 | LoopDevice **ret) { |
681 | ||
e8c7c4d9 | 682 | int r, basic_flags, direct_flags, rdwr_flags; |
8c1be37e | 683 | _cleanup_close_ int fd = -1; |
aa4d3aa3 | 684 | bool direct = false; |
8c1be37e LP |
685 | |
686 | assert(path); | |
687 | assert(ret); | |
b0a94268 | 688 | assert(open_flags < 0 || IN_SET(open_flags, O_RDWR, O_RDONLY)); |
8c1be37e | 689 | |
b0a94268 LP |
690 | /* Passing < 0 as open_flags here means we'll try to open the device writable if we can, retrying |
691 | * read-only if we cannot. */ | |
692 | ||
e8c7c4d9 LP |
693 | loop_flags = loop_flags_mangle(loop_flags); |
694 | ||
695 | /* Let's open with O_DIRECT if we can. But not all file systems support that, hence fall back to | |
696 | * non-O_DIRECT mode automatically, if it fails. */ | |
697 | ||
698 | basic_flags = O_CLOEXEC|O_NONBLOCK|O_NOCTTY; | |
699 | direct_flags = FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0; | |
700 | rdwr_flags = open_flags >= 0 ? open_flags : O_RDWR; | |
701 | ||
702 | fd = open(path, basic_flags|direct_flags|rdwr_flags); | |
703 | if (fd < 0 && direct_flags != 0) /* If we had O_DIRECT on, and things failed with that, let's immediately try again without */ | |
704 | fd = open(path, basic_flags|rdwr_flags); | |
aa4d3aa3 LP |
705 | else |
706 | direct = direct_flags != 0; | |
b0a94268 LP |
707 | if (fd < 0) { |
708 | r = -errno; | |
709 | ||
710 | /* Retry read-only? */ | |
711 | if (open_flags >= 0 || !(ERRNO_IS_PRIVILEGE(r) || r == -EROFS)) | |
712 | return r; | |
713 | ||
e8c7c4d9 LP |
714 | fd = open(path, basic_flags|direct_flags|O_RDONLY); |
715 | if (fd < 0 && direct_flags != 0) /* as above */ | |
716 | fd = open(path, basic_flags|O_RDONLY); | |
aa4d3aa3 LP |
717 | else |
718 | direct = direct_flags != 0; | |
b0a94268 LP |
719 | if (fd < 0) |
720 | return r; /* Propagate original error */ | |
721 | ||
722 | open_flags = O_RDONLY; | |
723 | } else if (open_flags < 0) | |
724 | open_flags = O_RDWR; | |
8c1be37e | 725 | |
aa4d3aa3 LP |
726 | log_debug("Opened '%s' in %s access mode%s, with O_DIRECT %s%s.", |
727 | path, | |
728 | open_flags == O_RDWR ? "O_RDWR" : "O_RDONLY", | |
729 | open_flags != rdwr_flags ? " (O_RDWR was requested but not allowed)" : "", | |
730 | direct ? "enabled" : "disabled", | |
731 | direct != (direct_flags != 0) ? " (O_DIRECT was requested but not supported)" : ""); | |
732 | ||
7f52206a | 733 | return loop_device_make_internal(fd, open_flags, 0, 0, loop_flags, lock_op, ret); |
8c1be37e LP |
734 | } |
735 | ||
736 | LoopDevice* loop_device_unref(LoopDevice *d) { | |
3a6ed1e1 LP |
737 | int r; |
738 | ||
8c1be37e LP |
739 | if (!d) |
740 | return NULL; | |
741 | ||
7f52206a LP |
742 | d->lock_fd = safe_close(d->lock_fd); |
743 | ||
8c1be37e | 744 | if (d->fd >= 0) { |
cae1e8fb LP |
745 | /* Implicitly sync the device, since otherwise in-flight blocks might not get written */ |
746 | if (fsync(d->fd) < 0) | |
747 | log_debug_errno(errno, "Failed to sync loop block device, ignoring: %m"); | |
748 | ||
a2ea3b2f | 749 | if (d->nr >= 0 && !d->relinquished) { |
3a6ed1e1 LP |
750 | /* We are supposed to clear the loopback device. Let's do this synchronously: lock |
751 | * the device, manually remove all partitions and then clear it. This should ensure | |
752 | * udev doesn't concurrently access the devices, and we can be reasonably sure that | |
753 | * once we are done here the device is cleared and all its partition children | |
754 | * removed. Note that we lock our primary device fd here (and not a separate locking | |
755 | * fd, as we do during allocation, since we want to keep the lock all the way through | |
756 | * the LOOP_CLR_FD, but that call would fail if we had more than one fd open.) */ | |
8c1be37e | 757 | |
3a6ed1e1 LP |
758 | if (flock(d->fd, LOCK_EX) < 0) |
759 | log_debug_errno(errno, "Failed to lock loop block device, ignoring: %m"); | |
760 | ||
761 | r = block_device_remove_all_partitions(d->fd); | |
762 | if (r < 0) | |
763 | log_debug_errno(r, "Failed to remove partitions of loopback block device, ignoring: %m"); | |
764 | ||
765 | if (ioctl(d->fd, LOOP_CLR_FD) < 0) | |
766 | log_debug_errno(errno, "Failed to clear loop device, ignoring: %m"); | |
8c1be37e LP |
767 | } |
768 | ||
769 | safe_close(d->fd); | |
770 | } | |
771 | ||
a2ea3b2f | 772 | if (d->nr >= 0 && !d->relinquished) { |
8c1be37e LP |
773 | _cleanup_close_ int control = -1; |
774 | ||
775 | control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK); | |
776 | if (control < 0) | |
f2d9213f ZJS |
777 | log_warning_errno(errno, |
778 | "Failed to open loop control device, cannot remove loop device %s: %m", | |
779 | strna(d->node)); | |
780 | else | |
781 | for (unsigned n_attempts = 0;;) { | |
782 | if (ioctl(control, LOOP_CTL_REMOVE, d->nr) >= 0) | |
783 | break; | |
784 | if (errno != EBUSY || ++n_attempts >= 64) { | |
785 | log_warning_errno(errno, "Failed to remove device %s: %m", strna(d->node)); | |
786 | break; | |
787 | } | |
cae1e8fb | 788 | (void) usleep(50 * USEC_PER_MSEC); |
f2d9213f | 789 | } |
8c1be37e LP |
790 | } |
791 | ||
792 | free(d->node); | |
5fecf46d | 793 | return mfree(d); |
8c1be37e | 794 | } |
a2ea3b2f LP |
795 | |
796 | void loop_device_relinquish(LoopDevice *d) { | |
797 | assert(d); | |
798 | ||
799 | /* Don't attempt to clean up the loop device anymore from this point on. Leave the clean-ing up to the kernel | |
800 | * itself, using the loop device "auto-clear" logic we already turned on when creating the device. */ | |
801 | ||
802 | d->relinquished = true; | |
803 | } | |
9dabc4fd | 804 | |
24d59aee DDM |
805 | void loop_device_unrelinquish(LoopDevice *d) { |
806 | assert(d); | |
807 | d->relinquished = false; | |
808 | } | |
809 | ||
7f52206a LP |
810 | int loop_device_open( |
811 | const char *loop_path, | |
812 | int open_flags, | |
813 | int lock_op, | |
814 | LoopDevice **ret) { | |
815 | ||
816 | _cleanup_close_ int loop_fd = -1, lock_fd = -1; | |
9dabc4fd | 817 | _cleanup_free_ char *p = NULL; |
b26c39ad | 818 | struct loop_info64 info; |
9dabc4fd LP |
819 | struct stat st; |
820 | LoopDevice *d; | |
b26c39ad | 821 | int nr; |
9dabc4fd LP |
822 | |
823 | assert(loop_path); | |
e8c7c4d9 | 824 | assert(IN_SET(open_flags, O_RDWR, O_RDONLY)); |
9dabc4fd LP |
825 | assert(ret); |
826 | ||
827 | loop_fd = open(loop_path, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags); | |
828 | if (loop_fd < 0) | |
829 | return -errno; | |
830 | ||
831 | if (fstat(loop_fd, &st) < 0) | |
832 | return -errno; | |
9dabc4fd LP |
833 | if (!S_ISBLK(st.st_mode)) |
834 | return -ENOTBLK; | |
835 | ||
10c1b188 LP |
836 | if (ioctl(loop_fd, LOOP_GET_STATUS64, &info) >= 0) { |
837 | #if HAVE_VALGRIND_MEMCHECK_H | |
838 | /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */ | |
839 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
840 | #endif | |
b26c39ad | 841 | nr = info.lo_number; |
10c1b188 | 842 | } else |
b26c39ad LP |
843 | nr = -1; |
844 | ||
7f52206a LP |
845 | if ((lock_op & ~LOCK_NB) != LOCK_UN) { |
846 | lock_fd = open_lock_fd(loop_fd, lock_op); | |
847 | if (lock_fd < 0) | |
848 | return lock_fd; | |
849 | } | |
850 | ||
9dabc4fd LP |
851 | p = strdup(loop_path); |
852 | if (!p) | |
853 | return -ENOMEM; | |
854 | ||
855 | d = new(LoopDevice, 1); | |
856 | if (!d) | |
857 | return -ENOMEM; | |
858 | ||
859 | *d = (LoopDevice) { | |
860 | .fd = TAKE_FD(loop_fd), | |
7f52206a | 861 | .lock_fd = TAKE_FD(lock_fd), |
b26c39ad | 862 | .nr = nr, |
9dabc4fd LP |
863 | .node = TAKE_PTR(p), |
864 | .relinquished = true, /* It's not ours, don't try to destroy it when this object is freed */ | |
79e8393a | 865 | .devno = st.st_dev, |
31c75fcc | 866 | .uevent_seqnum_not_before = UINT64_MAX, |
8ede1e86 | 867 | .timestamp_not_before = USEC_INFINITY, |
9dabc4fd LP |
868 | }; |
869 | ||
870 | *ret = d; | |
871 | return d->fd; | |
872 | } | |
873 | ||
f1443709 LP |
874 | static int resize_partition(int partition_fd, uint64_t offset, uint64_t size) { |
875 | char sysfs[STRLEN("/sys/dev/block/:/partition") + 2*DECIMAL_STR_MAX(dev_t) + 1]; | |
ca822829 | 876 | _cleanup_free_ char *buffer = NULL; |
f1443709 LP |
877 | uint64_t current_offset, current_size, partno; |
878 | _cleanup_close_ int whole_fd = -1; | |
879 | struct stat st; | |
880 | dev_t devno; | |
881 | int r; | |
882 | ||
883 | assert(partition_fd >= 0); | |
884 | ||
885 | /* Resizes the partition the loopback device refer to (assuming it refers to one instead of an actual | |
886 | * loopback device), and changes the offset, if needed. This is a fancy wrapper around | |
887 | * BLKPG_RESIZE_PARTITION. */ | |
888 | ||
889 | if (fstat(partition_fd, &st) < 0) | |
890 | return -errno; | |
891 | ||
892 | assert(S_ISBLK(st.st_mode)); | |
893 | ||
ed13feff | 894 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/partition", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
895 | r = read_one_line_file(sysfs, &buffer); |
896 | if (r == -ENOENT) /* not a partition, cannot resize */ | |
897 | return -ENOTTY; | |
898 | if (r < 0) | |
899 | return r; | |
900 | r = safe_atou64(buffer, &partno); | |
901 | if (r < 0) | |
902 | return r; | |
903 | ||
ed13feff | 904 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/start", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
905 | |
906 | buffer = mfree(buffer); | |
907 | r = read_one_line_file(sysfs, &buffer); | |
908 | if (r < 0) | |
909 | return r; | |
910 | r = safe_atou64(buffer, ¤t_offset); | |
911 | if (r < 0) | |
912 | return r; | |
913 | if (current_offset > UINT64_MAX/512U) | |
914 | return -EINVAL; | |
915 | current_offset *= 512U; | |
916 | ||
917 | if (ioctl(partition_fd, BLKGETSIZE64, ¤t_size) < 0) | |
918 | return -EINVAL; | |
919 | ||
920 | if (size == UINT64_MAX && offset == UINT64_MAX) | |
921 | return 0; | |
922 | if (current_size == size && current_offset == offset) | |
923 | return 0; | |
924 | ||
ed13feff | 925 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/../dev", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
926 | |
927 | buffer = mfree(buffer); | |
928 | r = read_one_line_file(sysfs, &buffer); | |
929 | if (r < 0) | |
930 | return r; | |
7176f06c | 931 | r = parse_devnum(buffer, &devno); |
f1443709 LP |
932 | if (r < 0) |
933 | return r; | |
934 | ||
ca822829 | 935 | whole_fd = r = device_open_from_devnum(S_IFBLK, devno, O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY, NULL); |
f1443709 LP |
936 | if (r < 0) |
937 | return r; | |
938 | ||
91e1ce1a LP |
939 | return block_device_resize_partition( |
940 | whole_fd, | |
941 | partno, | |
942 | offset == UINT64_MAX ? current_offset : offset, | |
943 | size == UINT64_MAX ? current_size : size); | |
f1443709 LP |
944 | } |
945 | ||
c37878fc LP |
946 | int loop_device_refresh_size(LoopDevice *d, uint64_t offset, uint64_t size) { |
947 | struct loop_info64 info; | |
ff27ef4b | 948 | |
9dabc4fd | 949 | assert(d); |
ff27ef4b | 950 | assert(d->fd >= 0); |
9dabc4fd | 951 | |
f1443709 LP |
952 | /* Changes the offset/start of the loop device relative to the beginning of the underlying file or |
953 | * block device. If this loop device actually refers to a partition and not a loopback device, we'll | |
954 | * try to adjust the partition offsets instead. | |
955 | * | |
956 | * If either offset or size is UINT64_MAX we won't change that parameter. */ | |
957 | ||
f1443709 LP |
958 | if (d->nr < 0) /* not a loopback device */ |
959 | return resize_partition(d->fd, offset, size); | |
960 | ||
c37878fc LP |
961 | if (ioctl(d->fd, LOOP_GET_STATUS64, &info) < 0) |
962 | return -errno; | |
963 | ||
10c1b188 LP |
964 | #if HAVE_VALGRIND_MEMCHECK_H |
965 | /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */ | |
966 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
967 | #endif | |
968 | ||
c37878fc LP |
969 | if (size == UINT64_MAX && offset == UINT64_MAX) |
970 | return 0; | |
971 | if (info.lo_sizelimit == size && info.lo_offset == offset) | |
972 | return 0; | |
973 | ||
974 | if (size != UINT64_MAX) | |
975 | info.lo_sizelimit = size; | |
976 | if (offset != UINT64_MAX) | |
977 | info.lo_offset = offset; | |
978 | ||
7c248223 | 979 | return RET_NERRNO(ioctl(d->fd, LOOP_SET_STATUS64, &info)); |
9dabc4fd | 980 | } |
441ec804 LP |
981 | |
982 | int loop_device_flock(LoopDevice *d, int operation) { | |
7f52206a | 983 | assert(IN_SET(operation & ~LOCK_NB, LOCK_UN, LOCK_SH, LOCK_EX)); |
441ec804 LP |
984 | assert(d); |
985 | ||
7f52206a LP |
986 | /* When unlocking just close the lock fd */ |
987 | if ((operation & ~LOCK_NB) == LOCK_UN) { | |
988 | d->lock_fd = safe_close(d->lock_fd); | |
989 | return 0; | |
990 | } | |
991 | ||
992 | /* If we had no lock fd so far, create one and lock it right-away */ | |
993 | if (d->lock_fd < 0) { | |
994 | assert(d->fd >= 0); | |
995 | ||
996 | d->lock_fd = open_lock_fd(d->fd, operation); | |
997 | if (d->lock_fd < 0) | |
998 | return d->lock_fd; | |
999 | ||
1000 | return 0; | |
1001 | } | |
441ec804 | 1002 | |
7f52206a LP |
1003 | /* Otherwise change the current lock mode on the existing fd */ |
1004 | return RET_NERRNO(flock(d->lock_fd, operation)); | |
441ec804 | 1005 | } |
8dbc208c LP |
1006 | |
1007 | int loop_device_sync(LoopDevice *d) { | |
1008 | assert(d); | |
ff27ef4b | 1009 | assert(d->fd >= 0); |
8dbc208c LP |
1010 | |
1011 | /* We also do this implicitly in loop_device_unref(). Doing this explicitly here has the benefit that | |
1012 | * we can check the return value though. */ | |
1013 | ||
7c248223 | 1014 | return RET_NERRNO(fsync(d->fd)); |
8dbc208c | 1015 | } |