]>
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" |
22ee78a8 | 23 | #include "dissect-image.h" |
e8c7c4d9 | 24 | #include "env-util.h" |
b0a94268 | 25 | #include "errno-util.h" |
8c1be37e | 26 | #include "fd-util.h" |
972c8db5 | 27 | #include "fs-util.h" |
f1443709 | 28 | #include "fileio.h" |
8c1be37e | 29 | #include "loop-util.h" |
86c1c1f3 | 30 | #include "missing_loop.h" |
f1443709 | 31 | #include "parse-util.h" |
e77cab82 | 32 | #include "path-util.h" |
b202ec20 | 33 | #include "random-util.h" |
3cc44114 | 34 | #include "stat-util.h" |
f1443709 | 35 | #include "stdio-util.h" |
f2d9213f | 36 | #include "string-util.h" |
021bf175 | 37 | #include "tmpfile-util.h" |
8c1be37e | 38 | |
e8af3bfd | 39 | static void cleanup_clear_loop_close(int *fd) { |
86c1c1f3 LP |
40 | if (*fd < 0) |
41 | return; | |
42 | ||
43 | (void) ioctl(*fd, LOOP_CLR_FD); | |
44 | (void) safe_close(*fd); | |
45 | } | |
46 | ||
021bf175 LP |
47 | static int loop_is_bound(int fd) { |
48 | struct loop_info64 info; | |
49 | ||
8e398254 | 50 | if (ioctl(ASSERT_FD(fd), LOOP_GET_STATUS64, &info) < 0) { |
021bf175 LP |
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 | 77 | |
10719a6f | 78 | assert(IN_SET(operation & ~LOCK_NB, LOCK_SH, LOCK_EX)); |
7f52206a | 79 | |
8e398254 | 80 | lock_fd = fd_reopen(ASSERT_FD(primary_fd), O_RDONLY|O_CLOEXEC|O_NONBLOCK|O_NOCTTY); |
7f52206a LP |
81 | if (lock_fd < 0) |
82 | return lock_fd; | |
10719a6f | 83 | |
7f52206a LP |
84 | if (flock(lock_fd, operation) < 0) |
85 | return -errno; | |
86 | ||
10719a6f | 87 | return TAKE_FD(lock_fd); |
7f52206a LP |
88 | } |
89 | ||
54ba7daf | 90 | static int loop_configure_verify_direct_io(int fd, const struct loop_config *c) { |
ac110243 | 91 | assert(fd >= 0); |
54ba7daf YW |
92 | assert(c); |
93 | ||
94 | if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_DIRECT_IO)) { | |
95 | struct loop_info64 info; | |
96 | ||
97 | if (ioctl(fd, LOOP_GET_STATUS64, &info) < 0) | |
98 | return log_debug_errno(errno, "Failed to issue LOOP_GET_STATUS64: %m"); | |
99 | ||
100 | #if HAVE_VALGRIND_MEMCHECK_H | |
101 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
102 | #endif | |
103 | ||
104 | /* On older kernels (<= 5.3) it was necessary to set the block size of the loopback block | |
105 | * device to the logical block size of the underlying file system. Since there was no nice | |
106 | * way to query the value, we are not bothering to do this however. On newer kernels the | |
107 | * block size is propagated automatically and does not require intervention from us. We'll | |
108 | * check here if enabling direct IO worked, to make this easily debuggable however. | |
109 | * | |
110 | * (Should anyone really care and actually wants direct IO on old kernels: it might be worth | |
f5bb0a31 LB |
111 | * enabling direct IO with iteratively larger block sizes until it eventually works.) |
112 | * | |
113 | * On older kernels (e.g.: 5.10) when this is attempted on a file stored on a dm-crypt | |
114 | * backed partition the kernel will start returning I/O errors when accessing the mounted | |
115 | * loop device, so return a recognizable error that causes the operation to be started | |
116 | * from scratch without the LO_FLAGS_DIRECT_IO flag. */ | |
54ba7daf | 117 | if (!FLAGS_SET(info.lo_flags, LO_FLAGS_DIRECT_IO)) |
f5bb0a31 LB |
118 | return log_debug_errno( |
119 | SYNTHETIC_ERRNO(ENOANO), | |
120 | "Could not enable direct IO mode, retrying in buffered IO mode."); | |
54ba7daf YW |
121 | } |
122 | ||
123 | return 0; | |
124 | } | |
125 | ||
126 | static int loop_configure_verify(int fd, const struct loop_config *c) { | |
127 | bool broken = false; | |
128 | int r; | |
129 | ||
130 | assert(fd >= 0); | |
131 | assert(c); | |
132 | ||
fd83c98e | 133 | if (c->block_size != 0) { |
65046b92 | 134 | uint32_t ssz; |
fd83c98e | 135 | |
65046b92 LP |
136 | r = blockdev_get_sector_size(fd, &ssz); |
137 | if (r < 0) | |
138 | return r; | |
fd83c98e | 139 | |
1163ddb3 | 140 | if (ssz != c->block_size) { |
65046b92 | 141 | log_debug("LOOP_CONFIGURE didn't honour requested block size %" PRIu32 ", got %" PRIu32 " instead. Ignoring.", c->block_size, ssz); |
1163ddb3 LP |
142 | broken = true; |
143 | } | |
fd83c98e AD |
144 | } |
145 | ||
54ba7daf YW |
146 | if (c->info.lo_sizelimit != 0) { |
147 | /* Kernel 5.8 vanilla doesn't properly propagate the size limit into the | |
148 | * block device. If it's used, let's immediately check if it had the desired | |
149 | * effect hence. And if not use classic LOOP_SET_STATUS64. */ | |
150 | uint64_t z; | |
151 | ||
152 | if (ioctl(fd, BLKGETSIZE64, &z) < 0) | |
153 | return -errno; | |
154 | ||
155 | if (z != c->info.lo_sizelimit) { | |
fd83c98e | 156 | log_debug("LOOP_CONFIGURE is broken, doesn't honour .info.lo_sizelimit. Falling back to LOOP_SET_STATUS64."); |
54ba7daf YW |
157 | broken = true; |
158 | } | |
159 | } | |
160 | ||
161 | if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_PARTSCAN)) { | |
162 | /* Kernel 5.8 vanilla doesn't properly propagate the partition scanning flag | |
163 | * into the block device. Let's hence verify if things work correctly here | |
164 | * before returning. */ | |
165 | ||
166 | r = blockdev_partscan_enabled(fd); | |
167 | if (r < 0) | |
168 | return r; | |
169 | if (r == 0) { | |
170 | log_debug("LOOP_CONFIGURE is broken, doesn't honour LO_FLAGS_PARTSCAN. Falling back to LOOP_SET_STATUS64."); | |
171 | broken = true; | |
172 | } | |
173 | } | |
174 | ||
175 | r = loop_configure_verify_direct_io(fd, c); | |
176 | if (r < 0) | |
177 | return r; | |
178 | ||
179 | return !broken; | |
180 | } | |
181 | ||
182 | static int loop_configure_fallback(int fd, const struct loop_config *c) { | |
183 | struct loop_info64 info_copy; | |
1163ddb3 | 184 | int r; |
54ba7daf YW |
185 | |
186 | assert(fd >= 0); | |
187 | assert(c); | |
188 | ||
189 | /* Only some of the flags LOOP_CONFIGURE can set are also settable via LOOP_SET_STATUS64, hence mask | |
190 | * them out. */ | |
191 | info_copy = c->info; | |
192 | info_copy.lo_flags &= LOOP_SET_STATUS_SETTABLE_FLAGS; | |
193 | ||
194 | /* Since kernel commit 5db470e229e22b7eda6e23b5566e532c96fb5bc3 (kernel v5.0) the LOOP_SET_STATUS64 | |
fd83c98e | 195 | * ioctl can return EAGAIN in case we change the info.lo_offset field, if someone else is accessing the |
54ba7daf YW |
196 | * block device while we try to reconfigure it. This is a pretty common case, since udev might |
197 | * instantly start probing the device as soon as we attach an fd to it. Hence handle it in two ways: | |
198 | * first, let's take the BSD lock to ensure that udev will not step in between the point in | |
199 | * time where we attach the fd and where we reconfigure the device. Secondly, let's wait 50ms on | |
200 | * EAGAIN and retry. The former should be an efficient mechanism to avoid we have to wait 50ms | |
201 | * needlessly if we are just racing against udev. The latter is protection against all other cases, | |
202 | * i.e. peers that do not take the BSD lock. */ | |
203 | ||
204 | for (unsigned n_attempts = 0;;) { | |
205 | if (ioctl(fd, LOOP_SET_STATUS64, &info_copy) >= 0) | |
206 | break; | |
207 | ||
208 | if (errno != EAGAIN || ++n_attempts >= 64) | |
209 | return log_debug_errno(errno, "Failed to configure loopback block device: %m"); | |
210 | ||
211 | /* Sleep some random time, but at least 10ms, at most 250ms. Increase the delay the more | |
212 | * failed attempts we see */ | |
4251512e | 213 | (void) usleep_safe(UINT64_C(10) * USEC_PER_MSEC + |
54ba7daf YW |
214 | random_u64_range(UINT64_C(240) * USEC_PER_MSEC * n_attempts/64)); |
215 | } | |
216 | ||
217 | /* Work around a kernel bug, where changing offset/size of the loopback device doesn't correctly | |
218 | * invalidate the buffer cache. For details see: | |
219 | * | |
220 | * https://android.googlesource.com/platform/system/apex/+/bef74542fbbb4cd629793f4efee8e0053b360570 | |
221 | * | |
222 | * This was fixed in kernel 5.0, see: | |
223 | * | |
224 | * https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5db470e229e22b7eda6e23b5566e532c96fb5bc3 | |
225 | * | |
226 | * We'll run the work-around here in the legacy LOOP_SET_STATUS64 codepath. In the LOOP_CONFIGURE | |
227 | * codepath above it should not be necessary. */ | |
228 | if (c->info.lo_offset != 0 || c->info.lo_sizelimit != 0) | |
229 | if (ioctl(fd, BLKFLSBUF, 0) < 0) | |
230 | log_debug_errno(errno, "Failed to issue BLKFLSBUF ioctl, ignoring: %m"); | |
231 | ||
1163ddb3 LP |
232 | /* If a block size is requested then try to configure it. If that doesn't work, ignore errors, but |
233 | * afterwards, let's validate what is in effect, and if it doesn't match what we want, fail */ | |
234 | if (c->block_size != 0) { | |
235 | uint32_t ssz; | |
236 | ||
237 | if (ioctl(fd, LOOP_SET_BLOCK_SIZE, (unsigned long) c->block_size) < 0) | |
238 | log_debug_errno(errno, "Failed to set sector size, ignoring: %m"); | |
239 | ||
240 | r = blockdev_get_sector_size(fd, &ssz); | |
241 | if (r < 0) | |
242 | return log_debug_errno(r, "Failed to read sector size: %m"); | |
243 | if (ssz != c->block_size) | |
244 | return log_debug_errno(SYNTHETIC_ERRNO(EIO), "Sector size of loopback device doesn't match what we requested, refusing."); | |
245 | } | |
246 | ||
54ba7daf YW |
247 | /* LO_FLAGS_DIRECT_IO is a flags we need to configure via explicit ioctls. */ |
248 | if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_DIRECT_IO)) | |
249 | if (ioctl(fd, LOOP_SET_DIRECT_IO, 1UL) < 0) | |
250 | log_debug_errno(errno, "Failed to enable direct IO mode, ignoring: %m"); | |
251 | ||
252 | return loop_configure_verify_direct_io(fd, c); | |
253 | } | |
254 | ||
95c50092 | 255 | static int loop_configure( |
021bf175 | 256 | int nr, |
da4fd288 YW |
257 | int open_flags, |
258 | int lock_op, | |
95c50092 | 259 | const struct loop_config *c, |
da4fd288 | 260 | LoopDevice **ret) { |
95c50092 | 261 | |
bb273a51 YW |
262 | static bool loop_configure_broken = false; |
263 | ||
da4fd288 | 264 | _cleanup_(sd_device_unrefp) sd_device *dev = NULL; |
254d1313 ZJS |
265 | _cleanup_(cleanup_clear_loop_close) int loop_with_fd = -EBADF; /* This must be declared before lock_fd. */ |
266 | _cleanup_close_ int fd = -EBADF, lock_fd = -EBADF; | |
da4fd288 YW |
267 | _cleanup_free_ char *node = NULL; |
268 | uint64_t diskseq = 0, seqnum = UINT64_MAX; | |
269 | usec_t timestamp = USEC_INFINITY; | |
270 | dev_t devno; | |
86c1c1f3 LP |
271 | int r; |
272 | ||
021bf175 | 273 | assert(nr >= 0); |
86c1c1f3 | 274 | assert(c); |
da4fd288 YW |
275 | assert(ret); |
276 | ||
277 | if (asprintf(&node, "/dev/loop%i", nr) < 0) | |
432f1fa8 | 278 | return log_oom_debug(); |
da4fd288 YW |
279 | |
280 | r = sd_device_new_from_devname(&dev, node); | |
281 | if (r < 0) | |
432f1fa8 | 282 | return log_debug_errno(r, "Failed to create sd_device object for \"%s\": %m", node); |
da4fd288 YW |
283 | |
284 | r = sd_device_get_devnum(dev, &devno); | |
285 | if (r < 0) | |
432f1fa8 | 286 | return log_device_debug_errno(dev, r, "Failed to get devnum: %m"); |
da4fd288 YW |
287 | |
288 | fd = sd_device_open(dev, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags); | |
289 | if (fd < 0) | |
432f1fa8 | 290 | return log_device_debug_errno(dev, fd, "Failed to open device: %m"); |
95c50092 | 291 | |
021bf175 LP |
292 | /* Let's lock the device before we do anything. We take the BSD lock on a second, separately opened |
293 | * fd for the device. udev after all watches for close() events (specifically IN_CLOSE_WRITE) on | |
294 | * block devices to reprobe them, hence by having a separate fd we will later close() we can ensure | |
295 | * we trigger udev after everything is done. If we'd lock our own fd instead and keep it open for a | |
296 | * long time udev would possibly never run on it again, even though the fd is unlocked, simply | |
297 | * because we never close() it. It also has the nice benefit we can use the _cleanup_close_ logic to | |
298 | * automatically release the lock, after we are done. */ | |
7f52206a | 299 | lock_fd = open_lock_fd(fd, LOCK_EX); |
021bf175 | 300 | if (lock_fd < 0) |
432f1fa8 YW |
301 | return log_device_debug_errno(dev, lock_fd, "Failed to acquire lock: %m"); |
302 | ||
303 | log_device_debug(dev, "Acquired exclusive lock."); | |
021bf175 | 304 | |
53274fdb YW |
305 | /* Let's see if backing file is really unattached. Someone may already attach a backing file without |
306 | * taking BSD lock. */ | |
307 | r = loop_is_bound(fd); | |
308 | if (r < 0) | |
432f1fa8 | 309 | return log_device_debug_errno(dev, r, "Failed to check if the loopback block device is bound: %m"); |
53274fdb | 310 | if (r > 0) |
432f1fa8 YW |
311 | return log_device_debug_errno(dev, SYNTHETIC_ERRNO(EBUSY), |
312 | "The loopback block device is already bound, ignoring."); | |
53274fdb | 313 | |
021bf175 LP |
314 | /* Let's see if the device is really detached, i.e. currently has no associated partition block |
315 | * devices. On various kernels (such as 5.8) it is possible to have a loopback block device that | |
247738b4 LP |
316 | * superficially is detached but still has partition block devices associated for it. Let's then |
317 | * manually remove the partitions via BLKPG, and tell the caller we did that via EUCLEAN, so they try | |
318 | * again. */ | |
833106b8 | 319 | r = block_device_remove_all_partitions(dev, fd); |
021bf175 | 320 | if (r < 0) |
432f1fa8 | 321 | return log_device_debug_errno(dev, r, "Failed to remove partitions on the loopback block device: %m"); |
833106b8 YW |
322 | if (r > 0) |
323 | /* Removed all partitions. Let's report this to the caller, to try again, and count this as | |
53274fdb | 324 | * an attempt. */ |
432f1fa8 YW |
325 | return log_device_debug_errno(dev, SYNTHETIC_ERRNO(EUCLEAN), |
326 | "Removed partitions on the loopback block device."); | |
021bf175 | 327 | |
bb273a51 | 328 | if (!loop_configure_broken) { |
31c75fcc LP |
329 | /* Acquire uevent seqnum immediately before attaching the loopback device. This allows |
330 | * callers to ignore all uevents with a seqnum before this one, if they need to associate | |
331 | * uevent with this attachment. Doing so isn't race-free though, as uevents that happen in | |
332 | * the window between this reading of the seqnum, and the LOOP_CONFIGURE call might still be | |
333 | * mistaken as originating from our attachment, even though might be caused by an earlier | |
334 | * use. But doing this at least shortens the race window a bit. */ | |
335 | r = get_current_uevent_seqnum(&seqnum); | |
336 | if (r < 0) | |
432f1fa8 | 337 | return log_device_debug_errno(dev, r, "Failed to get the current uevent seqnum: %m"); |
54ba7daf | 338 | |
8ede1e86 | 339 | timestamp = now(CLOCK_MONOTONIC); |
31c75fcc | 340 | |
95c50092 LP |
341 | if (ioctl(fd, LOOP_CONFIGURE, c) < 0) { |
342 | /* Do fallback only if LOOP_CONFIGURE is not supported, propagate all other | |
343 | * errors. Note that the kernel is weird: non-existing ioctls currently return EINVAL | |
344 | * rather than ENOTTY on loopback block devices. They should fix that in the kernel, | |
345 | * but in the meantime we accept both here. */ | |
346 | if (!ERRNO_IS_NOT_SUPPORTED(errno) && errno != EINVAL) | |
432f1fa8 | 347 | return log_device_debug_errno(dev, errno, "ioctl(LOOP_CONFIGURE) failed: %m"); |
86c1c1f3 | 348 | |
bb273a51 | 349 | loop_configure_broken = true; |
95c50092 | 350 | } else { |
da4fd288 YW |
351 | loop_with_fd = TAKE_FD(fd); |
352 | ||
353 | r = loop_configure_verify(loop_with_fd, c); | |
54ba7daf | 354 | if (r < 0) |
432f1fa8 | 355 | return log_device_debug_errno(dev, r, "Failed to verify if loopback block device is correctly configured: %m"); |
54ba7daf | 356 | if (r == 0) { |
95c50092 | 357 | /* LOOP_CONFIGURE doesn't work. Remember that. */ |
bb273a51 | 358 | loop_configure_broken = true; |
95c50092 LP |
359 | |
360 | /* We return EBUSY here instead of retrying immediately with LOOP_SET_FD, | |
361 | * because LOOP_CLR_FD is async: if the operation cannot be executed right | |
362 | * away it just sets the autoclear flag on the device. This means there's a | |
363 | * good chance we cannot actually reuse the loopback device right-away. Hence | |
364 | * let's assume it's busy, avoid the trouble and let the calling loop call us | |
365 | * again with a new, likely unused device. */ | |
da4fd288 | 366 | return -EBUSY; |
bb2551bd | 367 | } |
95c50092 | 368 | } |
86c1c1f3 LP |
369 | } |
370 | ||
bb273a51 YW |
371 | if (loop_configure_broken) { |
372 | /* Let's read the seqnum again, to shorten the window. */ | |
373 | r = get_current_uevent_seqnum(&seqnum); | |
374 | if (r < 0) | |
432f1fa8 | 375 | return log_device_debug_errno(dev, r, "Failed to get the current uevent seqnum: %m"); |
31c75fcc | 376 | |
bb273a51 | 377 | timestamp = now(CLOCK_MONOTONIC); |
738f29cb | 378 | |
bb273a51 | 379 | if (ioctl(fd, LOOP_SET_FD, c->fd) < 0) |
432f1fa8 | 380 | return log_device_debug_errno(dev, errno, "ioctl(LOOP_SET_FD) failed: %m"); |
86c1c1f3 | 381 | |
da4fd288 YW |
382 | loop_with_fd = TAKE_FD(fd); |
383 | ||
384 | r = loop_configure_fallback(loop_with_fd, c); | |
bb273a51 | 385 | if (r < 0) |
da4fd288 | 386 | return r; |
bb273a51 | 387 | } |
e8c7c4d9 | 388 | |
da4fd288 YW |
389 | r = fd_get_diskseq(loop_with_fd, &diskseq); |
390 | if (r < 0 && r != -EOPNOTSUPP) | |
432f1fa8 | 391 | return log_device_debug_errno(dev, r, "Failed to get diskseq: %m"); |
31c75fcc | 392 | |
da4fd288 YW |
393 | switch (lock_op & ~LOCK_NB) { |
394 | case LOCK_EX: /* Already in effect */ | |
395 | break; | |
396 | case LOCK_SH: /* Downgrade */ | |
397 | if (flock(lock_fd, lock_op) < 0) | |
432f1fa8 | 398 | return log_device_debug_errno(dev, errno, "Failed to downgrade lock level: %m"); |
da4fd288 YW |
399 | break; |
400 | case LOCK_UN: /* Release */ | |
401 | lock_fd = safe_close(lock_fd); | |
402 | break; | |
403 | default: | |
404 | assert_not_reached(); | |
405 | } | |
406 | ||
407 | LoopDevice *d = new(LoopDevice, 1); | |
408 | if (!d) | |
432f1fa8 | 409 | return log_oom_debug(); |
da4fd288 YW |
410 | |
411 | *d = (LoopDevice) { | |
36d5eb0b | 412 | .n_ref = 1, |
da4fd288 YW |
413 | .fd = TAKE_FD(loop_with_fd), |
414 | .lock_fd = TAKE_FD(lock_fd), | |
415 | .node = TAKE_PTR(node), | |
416 | .nr = nr, | |
417 | .devno = devno, | |
418 | .dev = TAKE_PTR(dev), | |
419 | .diskseq = diskseq, | |
420 | .uevent_seqnum_not_before = seqnum, | |
421 | .timestamp_not_before = timestamp, | |
22ee78a8 | 422 | .sector_size = c->block_size, |
da4fd288 | 423 | }; |
86c1c1f3 | 424 | |
da4fd288 YW |
425 | *ret = TAKE_PTR(d); |
426 | return 0; | |
e8af3bfd ZJS |
427 | } |
428 | ||
e8c7c4d9 | 429 | static int loop_device_make_internal( |
e77cab82 | 430 | const char *path, |
ed9eeb7b LP |
431 | int fd, |
432 | int open_flags, | |
433 | uint64_t offset, | |
434 | uint64_t size, | |
22ee78a8 | 435 | uint32_t sector_size, |
ed9eeb7b | 436 | uint32_t loop_flags, |
7f52206a | 437 | int lock_op, |
ed9eeb7b | 438 | LoopDevice **ret) { |
8c1be37e | 439 | |
da4fd288 | 440 | _cleanup_(loop_device_unrefp) LoopDevice *d = NULL; |
f5bb0a31 | 441 | _cleanup_close_ int reopened_fd = -EBADF, control = -EBADF; |
da4fd288 | 442 | _cleanup_free_ char *backing_file = NULL; |
86c1c1f3 | 443 | struct loop_config config; |
da4fd288 | 444 | int r, f_flags; |
8c1be37e | 445 | struct stat st; |
8c1be37e | 446 | |
8c1be37e LP |
447 | assert(ret); |
448 | assert(IN_SET(open_flags, O_RDWR, O_RDONLY)); | |
449 | ||
8e398254 | 450 | if (fstat(ASSERT_FD(fd), &st) < 0) |
8c1be37e LP |
451 | return -errno; |
452 | ||
453 | if (S_ISBLK(st.st_mode)) { | |
1996ad28 | 454 | if (offset == 0 && IN_SET(size, 0, UINT64_MAX)) |
d7654742 LP |
455 | /* If this is already a block device and we are supposed to cover the whole of it |
456 | * then store an fd to the original open device node — and do not actually create an | |
1996ad28 | 457 | * unnecessary loopback device for it. */ |
de3b7f16 | 458 | return loop_device_open_from_fd(fd, open_flags, lock_op, ret); |
ed9eeb7b LP |
459 | } else { |
460 | r = stat_verify_regular(&st); | |
461 | if (r < 0) | |
462 | return r; | |
8c1be37e LP |
463 | } |
464 | ||
e77cab82 YW |
465 | if (path) { |
466 | r = path_make_absolute_cwd(path, &backing_file); | |
467 | if (r < 0) | |
468 | return r; | |
469 | ||
470 | path_simplify(backing_file); | |
471 | } else { | |
472 | r = fd_get_path(fd, &backing_file); | |
473 | if (r < 0) | |
474 | return r; | |
475 | } | |
476 | ||
e8c7c4d9 LP |
477 | f_flags = fcntl(fd, F_GETFL); |
478 | if (f_flags < 0) | |
479 | return -errno; | |
480 | ||
481 | if (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) != FLAGS_SET(f_flags, O_DIRECT)) { | |
482 | /* If LO_FLAGS_DIRECT_IO is requested, then make sure we have the fd open with O_DIRECT, as | |
483 | * that's required. Conversely, if it's off require that O_DIRECT is off too (that's because | |
484 | * new kernels will implicitly enable LO_FLAGS_DIRECT_IO if O_DIRECT is set). | |
485 | * | |
486 | * Our intention here is that LO_FLAGS_DIRECT_IO is the primary knob, and O_DIRECT derived | |
487 | * from that automatically. */ | |
488 | ||
f5bb0a31 LB |
489 | reopened_fd = fd_reopen(fd, (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0)|O_CLOEXEC|O_NONBLOCK|open_flags); |
490 | if (reopened_fd < 0) { | |
e8c7c4d9 | 491 | if (!FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO)) |
d579c42e | 492 | return log_debug_errno(reopened_fd, "Failed to reopen file descriptor without O_DIRECT: %m"); |
e8c7c4d9 LP |
493 | |
494 | /* Some file systems might not support O_DIRECT, let's gracefully continue without it then. */ | |
d579c42e | 495 | log_debug_errno(reopened_fd, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without."); |
e8c7c4d9 LP |
496 | loop_flags &= ~LO_FLAGS_DIRECT_IO; |
497 | } else | |
f5bb0a31 | 498 | fd = reopened_fd; /* From now on, operate on our new O_DIRECT fd */ |
e8c7c4d9 LP |
499 | } |
500 | ||
8c1be37e LP |
501 | control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK); |
502 | if (control < 0) | |
503 | return -errno; | |
504 | ||
22ee78a8 LP |
505 | if (sector_size == 0) |
506 | /* If no sector size is specified, default to the classic default */ | |
507 | sector_size = 512; | |
508 | else if (sector_size == UINT32_MAX) { | |
509 | ||
510 | if (S_ISBLK(st.st_mode)) | |
511 | /* If the sector size is specified as UINT32_MAX we'll propagate the sector size of | |
512 | * the underlying block device. */ | |
513 | r = blockdev_get_sector_size(fd, §or_size); | |
514 | else { | |
92651a7a | 515 | _cleanup_close_ int non_direct_io_fd = -EBADF; |
22ee78a8 LP |
516 | int probe_fd; |
517 | ||
518 | assert(S_ISREG(st.st_mode)); | |
519 | ||
520 | /* If sector size is specified as UINT32_MAX, we'll try to probe the right sector | |
521 | * size of the image in question by looking for the GPT partition header at various | |
522 | * offsets. This of course only works if the image already has a disk label. | |
523 | * | |
524 | * So here we actually want to read the file contents ourselves. This is quite likely | |
525 | * not going to work if we managed to enable O_DIRECT, because in such a case there | |
526 | * are some pretty strict alignment requirements to offset, size and target, but | |
527 | * there's no way to query what alignment specifically is actually required. Hence, | |
528 | * let's avoid the mess, and temporarily open an fd without O_DIRECT for the probing | |
529 | * logic. */ | |
530 | ||
531 | if (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO)) { | |
532 | non_direct_io_fd = fd_reopen(fd, O_RDONLY|O_CLOEXEC|O_NONBLOCK); | |
533 | if (non_direct_io_fd < 0) | |
534 | return non_direct_io_fd; | |
535 | ||
536 | probe_fd = non_direct_io_fd; | |
537 | } else | |
538 | probe_fd = fd; | |
539 | ||
540 | r = probe_sector_size(probe_fd, §or_size); | |
541 | } | |
542 | if (r < 0) | |
543 | return r; | |
544 | } | |
545 | ||
86c1c1f3 LP |
546 | config = (struct loop_config) { |
547 | .fd = fd, | |
22ee78a8 | 548 | .block_size = sector_size, |
86c1c1f3 LP |
549 | .info = { |
550 | /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */ | |
0950526a | 551 | .lo_flags = (loop_flags & ~LO_FLAGS_READ_ONLY) | ((open_flags & O_ACCMODE) == O_RDONLY ? LO_FLAGS_READ_ONLY : 0) | LO_FLAGS_AUTOCLEAR, |
86c1c1f3 LP |
552 | .lo_offset = offset, |
553 | .lo_sizelimit = size == UINT64_MAX ? 0 : size, | |
554 | }, | |
555 | }; | |
556 | ||
0f6519d4 LP |
557 | /* Loop around LOOP_CTL_GET_FREE, since at the moment we attempt to open the returned device it might |
558 | * be gone already, taken by somebody else racing against us. */ | |
e8af3bfd | 559 | for (unsigned n_attempts = 0;;) { |
432f1fa8 | 560 | usec_t usec; |
da4fd288 | 561 | int nr; |
e8af3bfd | 562 | |
cc530466 LP |
563 | /* Let's take a lock on the control device first. On a busy system, where many programs |
564 | * attempt to allocate a loopback device at the same time, we might otherwise keep looping | |
565 | * around relatively heavy operations: asking for a free loopback device, then opening it, | |
566 | * validating it, attaching something to it. Let's serialize this whole operation, to make | |
567 | * unnecessary busywork less likely. Note that this is just something we do to optimize our | |
568 | * own code (and whoever else decides to use LOCK_EX locks for this), taking this lock is not | |
569 | * necessary, it just means it's less likely we have to iterate through this loop again and | |
4c1d50e6 LP |
570 | * again if our own code races against our own code. |
571 | * | |
572 | * Note: our lock protocol is to take the /dev/loop-control lock first, and the block device | |
573 | * lock second, if both are taken, and always in this order, to avoid ABBA locking issues. */ | |
cc530466 LP |
574 | if (flock(control, LOCK_EX) < 0) |
575 | return -errno; | |
576 | ||
0f6519d4 LP |
577 | nr = ioctl(control, LOOP_CTL_GET_FREE); |
578 | if (nr < 0) | |
579 | return -errno; | |
8c1be37e | 580 | |
da4fd288 YW |
581 | r = loop_configure(nr, open_flags, lock_op, &config, &d); |
582 | if (r >= 0) | |
583 | break; | |
cc5bae6c | 584 | |
da4fd288 YW |
585 | /* -ENODEV or friends: Somebody might've gotten the same number from the kernel, used the |
586 | * device, and called LOOP_CTL_REMOVE on it. Let's retry with a new number. | |
587 | * -EBUSY: a file descriptor is already bound to the loopback block device. | |
f5bb0a31 LB |
588 | * -EUCLEAN: some left-over partition devices that were cleaned up. |
589 | * -ENOANO: we tried to use LO_FLAGS_DIRECT_IO but the kernel rejected it. */ | |
590 | if (!ERRNO_IS_DEVICE_ABSENT(r) && !IN_SET(r, -EBUSY, -EUCLEAN, -ENOANO)) | |
2421dd72 | 591 | return r; |
01813148 | 592 | |
cc530466 LP |
593 | /* OK, this didn't work, let's try again a bit later, but first release the lock on the |
594 | * control device */ | |
595 | if (flock(control, LOCK_UN) < 0) | |
596 | return -errno; | |
597 | ||
e8af3bfd ZJS |
598 | if (++n_attempts >= 64) /* Give up eventually */ |
599 | return -EBUSY; | |
0f6519d4 | 600 | |
f5bb0a31 LB |
601 | /* If we failed to enable direct IO mode, let's retry without it. We restart the process as |
602 | * on some combination of kernel version and storage filesystem, the kernel is very unhappy | |
603 | * about a failed DIRECT_IO enablement and throws I/O errors. */ | |
604 | if (r == -ENOANO && FLAGS_SET(config.info.lo_flags, LO_FLAGS_DIRECT_IO)) { | |
605 | config.info.lo_flags &= ~LO_FLAGS_DIRECT_IO; | |
606 | open_flags &= ~O_DIRECT; | |
607 | ||
608 | int non_direct_io_fd = fd_reopen(config.fd, O_CLOEXEC|O_NONBLOCK|open_flags); | |
609 | if (non_direct_io_fd < 0) | |
610 | return log_debug_errno( | |
611 | non_direct_io_fd, | |
612 | "Failed to reopen file descriptor without O_DIRECT: %m"); | |
613 | ||
614 | safe_close(reopened_fd); | |
615 | fd = config.fd = /* For cleanups */ reopened_fd = non_direct_io_fd; | |
616 | } | |
617 | ||
b202ec20 LP |
618 | /* Wait some random time, to make collision less likely. Let's pick a random time in the |
619 | * range 0ms…250ms, linearly scaled by the number of failed attempts. */ | |
432f1fa8 YW |
620 | usec = random_u64_range(UINT64_C(10) * USEC_PER_MSEC + |
621 | UINT64_C(240) * USEC_PER_MSEC * n_attempts/64); | |
622 | log_debug("Trying again after %s.", FORMAT_TIMESPAN(usec, USEC_PER_MSEC)); | |
4251512e | 623 | (void) usleep_safe(usec); |
0f6519d4 | 624 | } |
8c1be37e | 625 | |
da4fd288 | 626 | d->backing_file = TAKE_PTR(backing_file); |
4d2a9e3e LP |
627 | d->backing_inode = st.st_ino; |
628 | d->backing_devno = st.st_dev; | |
8c1be37e | 629 | |
3b195f63 LP |
630 | log_debug("Successfully acquired %s, devno=%u:%u, nr=%i, diskseq=%" PRIu64, |
631 | d->node, | |
632 | major(d->devno), minor(d->devno), | |
633 | d->nr, | |
634 | d->diskseq); | |
635 | ||
da4fd288 YW |
636 | *ret = TAKE_PTR(d); |
637 | return 0; | |
8c1be37e LP |
638 | } |
639 | ||
e8c7c4d9 LP |
640 | static uint32_t loop_flags_mangle(uint32_t loop_flags) { |
641 | int r; | |
642 | ||
643 | r = getenv_bool("SYSTEMD_LOOP_DIRECT_IO"); | |
644 | if (r < 0 && r != -ENXIO) | |
645 | log_debug_errno(r, "Failed to parse $SYSTEMD_LOOP_DIRECT_IO, ignoring: %m"); | |
646 | ||
bfd08445 | 647 | 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 |
648 | } |
649 | ||
650 | int loop_device_make( | |
651 | int fd, | |
652 | int open_flags, | |
653 | uint64_t offset, | |
654 | uint64_t size, | |
22ee78a8 | 655 | uint32_t sector_size, |
e8c7c4d9 | 656 | uint32_t loop_flags, |
7f52206a | 657 | int lock_op, |
e8c7c4d9 LP |
658 | LoopDevice **ret) { |
659 | ||
660 | assert(fd >= 0); | |
661 | assert(ret); | |
e8c7c4d9 LP |
662 | |
663 | return loop_device_make_internal( | |
e77cab82 | 664 | NULL, |
e8c7c4d9 LP |
665 | fd, |
666 | open_flags, | |
667 | offset, | |
668 | size, | |
22ee78a8 | 669 | sector_size, |
bfd08445 | 670 | loop_flags_mangle(loop_flags), |
7f52206a | 671 | lock_op, |
e8c7c4d9 LP |
672 | ret); |
673 | } | |
674 | ||
972c8db5 DDM |
675 | int loop_device_make_by_path_at( |
676 | int dir_fd, | |
79e8393a LP |
677 | const char *path, |
678 | int open_flags, | |
22ee78a8 | 679 | uint32_t sector_size, |
79e8393a | 680 | uint32_t loop_flags, |
7f52206a | 681 | int lock_op, |
79e8393a LP |
682 | LoopDevice **ret) { |
683 | ||
e8c7c4d9 | 684 | int r, basic_flags, direct_flags, rdwr_flags; |
254d1313 | 685 | _cleanup_close_ int fd = -EBADF; |
aa4d3aa3 | 686 | bool direct = false; |
8c1be37e | 687 | |
972c8db5 | 688 | assert(dir_fd >= 0 || dir_fd == AT_FDCWD); |
8c1be37e LP |
689 | assert(path); |
690 | assert(ret); | |
b0a94268 | 691 | assert(open_flags < 0 || IN_SET(open_flags, O_RDWR, O_RDONLY)); |
8c1be37e | 692 | |
b0a94268 LP |
693 | /* Passing < 0 as open_flags here means we'll try to open the device writable if we can, retrying |
694 | * read-only if we cannot. */ | |
695 | ||
e8c7c4d9 LP |
696 | loop_flags = loop_flags_mangle(loop_flags); |
697 | ||
698 | /* Let's open with O_DIRECT if we can. But not all file systems support that, hence fall back to | |
699 | * non-O_DIRECT mode automatically, if it fails. */ | |
700 | ||
701 | basic_flags = O_CLOEXEC|O_NONBLOCK|O_NOCTTY; | |
702 | direct_flags = FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0; | |
703 | rdwr_flags = open_flags >= 0 ? open_flags : O_RDWR; | |
704 | ||
420d2e31 | 705 | fd = xopenat(dir_fd, path, basic_flags|direct_flags|rdwr_flags, /* xopen_flags = */ 0, /* mode = */ 0); |
e8c7c4d9 | 706 | if (fd < 0 && direct_flags != 0) /* If we had O_DIRECT on, and things failed with that, let's immediately try again without */ |
420d2e31 | 707 | fd = xopenat(dir_fd, path, basic_flags|rdwr_flags, /* xopen_flags = */ 0, /* mode = */ 0); |
aa4d3aa3 LP |
708 | else |
709 | direct = direct_flags != 0; | |
b0a94268 LP |
710 | if (fd < 0) { |
711 | r = -errno; | |
712 | ||
713 | /* Retry read-only? */ | |
714 | if (open_flags >= 0 || !(ERRNO_IS_PRIVILEGE(r) || r == -EROFS)) | |
715 | return r; | |
716 | ||
420d2e31 | 717 | fd = xopenat(dir_fd, path, basic_flags|direct_flags|O_RDONLY, /* xopen_flags = */ 0, /* mode = */ 0); |
e8c7c4d9 | 718 | if (fd < 0 && direct_flags != 0) /* as above */ |
420d2e31 | 719 | fd = xopenat(dir_fd, path, basic_flags|O_RDONLY, /* xopen_flags = */ 0, /* mode = */ 0); |
aa4d3aa3 LP |
720 | else |
721 | direct = direct_flags != 0; | |
b0a94268 LP |
722 | if (fd < 0) |
723 | return r; /* Propagate original error */ | |
724 | ||
725 | open_flags = O_RDONLY; | |
726 | } else if (open_flags < 0) | |
727 | open_flags = O_RDWR; | |
8c1be37e | 728 | |
aa4d3aa3 LP |
729 | log_debug("Opened '%s' in %s access mode%s, with O_DIRECT %s%s.", |
730 | path, | |
731 | open_flags == O_RDWR ? "O_RDWR" : "O_RDONLY", | |
732 | open_flags != rdwr_flags ? " (O_RDWR was requested but not allowed)" : "", | |
733 | direct ? "enabled" : "disabled", | |
734 | direct != (direct_flags != 0) ? " (O_DIRECT was requested but not supported)" : ""); | |
735 | ||
972c8db5 DDM |
736 | return loop_device_make_internal( |
737 | dir_fd == AT_FDCWD ? path : NULL, | |
738 | fd, | |
739 | open_flags, | |
740 | /* offset = */ 0, | |
741 | /* size = */ 0, | |
742 | sector_size, | |
743 | loop_flags, | |
744 | lock_op, | |
745 | ret); | |
8c1be37e LP |
746 | } |
747 | ||
fcd8a19d LP |
748 | int loop_device_make_by_path_memory( |
749 | const char *path, | |
750 | int open_flags, | |
22ee78a8 | 751 | uint32_t sector_size, |
fcd8a19d LP |
752 | uint32_t loop_flags, |
753 | int lock_op, | |
754 | LoopDevice **ret) { | |
755 | ||
756 | _cleanup_close_ int fd = -EBADF, mfd = -EBADF; | |
757 | _cleanup_free_ char *fn = NULL; | |
758 | struct stat st; | |
759 | int r; | |
760 | ||
761 | assert(path); | |
762 | assert(IN_SET(open_flags, O_RDWR, O_RDONLY)); | |
763 | assert(ret); | |
764 | ||
765 | loop_flags &= ~LO_FLAGS_DIRECT_IO; /* memfds don't support O_DIRECT, hence LO_FLAGS_DIRECT_IO can't be used either */ | |
766 | ||
767 | fd = open(path, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|O_RDONLY); | |
768 | if (fd < 0) | |
769 | return -errno; | |
770 | ||
771 | if (fstat(fd, &st) < 0) | |
772 | return -errno; | |
773 | ||
774 | if (!S_ISREG(st.st_mode) && !S_ISBLK(st.st_mode)) | |
775 | return -EBADF; | |
776 | ||
777 | r = path_extract_filename(path, &fn); | |
778 | if (r < 0) | |
779 | return r; | |
780 | ||
781 | mfd = memfd_clone_fd(fd, fn, open_flags|O_CLOEXEC); | |
782 | if (mfd < 0) | |
783 | return mfd; | |
784 | ||
785 | fd = safe_close(fd); /* Let's close the original early */ | |
786 | ||
22ee78a8 | 787 | return loop_device_make_internal(NULL, mfd, open_flags, 0, 0, sector_size, loop_flags, lock_op, ret); |
fcd8a19d LP |
788 | } |
789 | ||
36d5eb0b | 790 | static LoopDevice* loop_device_free(LoopDevice *d) { |
5bb1d7fb | 791 | _cleanup_close_ int control = -EBADF; |
3a6ed1e1 LP |
792 | int r; |
793 | ||
8c1be37e LP |
794 | if (!d) |
795 | return NULL; | |
796 | ||
4c1d50e6 LP |
797 | /* Release any lock we might have on the device first. We want to open+lock the /dev/loop-control |
798 | * device below, but our lock protocol says that if both control and block device locks are taken, | |
799 | * the control lock needs to be taken first, the block device lock second — in order to avoid ABBA | |
800 | * locking issues. Moreover, we want to issue LOOP_CLR_FD on the block device further down, and that | |
801 | * would fail if we had another fd open to the device. */ | |
7f52206a LP |
802 | d->lock_fd = safe_close(d->lock_fd); |
803 | ||
4c1d50e6 LP |
804 | /* Let's open the control device early, and lock it, so that we can release our block device and |
805 | * delete it in a synchronized fashion, and allocators won't needlessly see the block device as free | |
806 | * while we are about to delete it. */ | |
7cb349f0 | 807 | if (!LOOP_DEVICE_IS_FOREIGN(d) && !d->relinquished) { |
4c1d50e6 LP |
808 | control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK); |
809 | if (control < 0) | |
810 | log_debug_errno(errno, "Failed to open loop control device, cannot remove loop device '%s', ignoring: %m", strna(d->node)); | |
811 | else if (flock(control, LOCK_EX) < 0) | |
812 | log_debug_errno(errno, "Failed to lock loop control device, ignoring: %m"); | |
813 | } | |
814 | ||
815 | /* Then let's release the loopback block device */ | |
8c1be37e | 816 | if (d->fd >= 0) { |
cae1e8fb LP |
817 | /* Implicitly sync the device, since otherwise in-flight blocks might not get written */ |
818 | if (fsync(d->fd) < 0) | |
819 | log_debug_errno(errno, "Failed to sync loop block device, ignoring: %m"); | |
820 | ||
7cb349f0 | 821 | if (!LOOP_DEVICE_IS_FOREIGN(d) && !d->relinquished) { |
3a6ed1e1 LP |
822 | /* We are supposed to clear the loopback device. Let's do this synchronously: lock |
823 | * the device, manually remove all partitions and then clear it. This should ensure | |
824 | * udev doesn't concurrently access the devices, and we can be reasonably sure that | |
825 | * once we are done here the device is cleared and all its partition children | |
826 | * removed. Note that we lock our primary device fd here (and not a separate locking | |
827 | * fd, as we do during allocation, since we want to keep the lock all the way through | |
828 | * the LOOP_CLR_FD, but that call would fail if we had more than one fd open.) */ | |
8c1be37e | 829 | |
3a6ed1e1 LP |
830 | if (flock(d->fd, LOCK_EX) < 0) |
831 | log_debug_errno(errno, "Failed to lock loop block device, ignoring: %m"); | |
832 | ||
46c3a288 | 833 | r = block_device_remove_all_partitions(d->dev, d->fd); |
3a6ed1e1 LP |
834 | if (r < 0) |
835 | log_debug_errno(r, "Failed to remove partitions of loopback block device, ignoring: %m"); | |
836 | ||
837 | if (ioctl(d->fd, LOOP_CLR_FD) < 0) | |
838 | log_debug_errno(errno, "Failed to clear loop device, ignoring: %m"); | |
8c1be37e LP |
839 | } |
840 | ||
841 | safe_close(d->fd); | |
842 | } | |
843 | ||
4c1d50e6 | 844 | /* Now that the block device is released, let's also try to remove it */ |
afbe20b7 | 845 | if (control >= 0) { |
6483bcef ZJS |
846 | useconds_t delay = 5 * USEC_PER_MSEC; /* A total delay of 5090 ms between 39 attempts, |
847 | * (4*5 + 5*10 + 5*20 + … + 3*640) = 5090. */ | |
afbe20b7 ZJS |
848 | |
849 | for (unsigned attempt = 1;; attempt++) { | |
4c1d50e6 LP |
850 | if (ioctl(control, LOOP_CTL_REMOVE, d->nr) >= 0) |
851 | break; | |
afbe20b7 | 852 | if (errno != EBUSY || attempt > 38) { |
4c1d50e6 LP |
853 | log_debug_errno(errno, "Failed to remove device %s: %m", strna(d->node)); |
854 | break; | |
f2d9213f | 855 | } |
afbe20b7 ZJS |
856 | if (attempt % 5 == 0) { |
857 | log_debug("Device is still busy after %u attempts…", attempt); | |
858 | delay *= 2; | |
859 | } | |
860 | ||
4251512e | 861 | (void) usleep_safe(delay); |
4c1d50e6 | 862 | } |
afbe20b7 | 863 | } |
8c1be37e LP |
864 | |
865 | free(d->node); | |
cc5bae6c | 866 | sd_device_unref(d->dev); |
e77cab82 | 867 | free(d->backing_file); |
5fecf46d | 868 | return mfree(d); |
8c1be37e | 869 | } |
a2ea3b2f | 870 | |
36d5eb0b YW |
871 | DEFINE_TRIVIAL_REF_UNREF_FUNC(LoopDevice, loop_device, loop_device_free); |
872 | ||
a2ea3b2f LP |
873 | void loop_device_relinquish(LoopDevice *d) { |
874 | assert(d); | |
875 | ||
876 | /* Don't attempt to clean up the loop device anymore from this point on. Leave the clean-ing up to the kernel | |
877 | * itself, using the loop device "auto-clear" logic we already turned on when creating the device. */ | |
878 | ||
879 | d->relinquished = true; | |
880 | } | |
9dabc4fd | 881 | |
24d59aee DDM |
882 | void loop_device_unrelinquish(LoopDevice *d) { |
883 | assert(d); | |
884 | d->relinquished = false; | |
885 | } | |
886 | ||
4f0ad43e YW |
887 | int loop_device_open( |
888 | sd_device *dev, | |
7f52206a LP |
889 | int open_flags, |
890 | int lock_op, | |
891 | LoopDevice **ret) { | |
892 | ||
254d1313 | 893 | _cleanup_close_ int fd = -EBADF, lock_fd = -EBADF; |
4f0ad43e | 894 | _cleanup_free_ char *node = NULL, *backing_file = NULL; |
4d2a9e3e | 895 | dev_t devnum, backing_devno = 0; |
b26c39ad | 896 | struct loop_info64 info; |
4d2a9e3e | 897 | ino_t backing_inode = 0; |
ffcb3324 | 898 | uint64_t diskseq = 0; |
9dabc4fd | 899 | LoopDevice *d; |
4f0ad43e | 900 | const char *s; |
a8d8a619 | 901 | int r, nr = -1; |
9dabc4fd | 902 | |
4f0ad43e | 903 | assert(dev); |
e8c7c4d9 | 904 | assert(IN_SET(open_flags, O_RDWR, O_RDONLY)); |
9dabc4fd LP |
905 | assert(ret); |
906 | ||
4f0ad43e YW |
907 | /* Even if fd is provided through the argument in loop_device_open_from_fd(), we reopen the inode |
908 | * here, instead of keeping just a dup() clone of it around, since we want to ensure that the | |
909 | * O_DIRECT flag of the handle we keep is off, we have our own file index, and have the right | |
910 | * read/write mode in effect. */ | |
911 | fd = sd_device_open(dev, O_CLOEXEC|O_NONBLOCK|O_NOCTTY|open_flags); | |
912 | if (fd < 0) | |
913 | return fd; | |
cc5bae6c | 914 | |
4f0ad43e YW |
915 | if ((lock_op & ~LOCK_NB) != LOCK_UN) { |
916 | lock_fd = open_lock_fd(fd, lock_op); | |
917 | if (lock_fd < 0) | |
918 | return lock_fd; | |
a8d8a619 YW |
919 | } |
920 | ||
4f0ad43e | 921 | if (ioctl(fd, LOOP_GET_STATUS64, &info) >= 0) { |
10c1b188 LP |
922 | #if HAVE_VALGRIND_MEMCHECK_H |
923 | /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */ | |
924 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
925 | #endif | |
b26c39ad | 926 | nr = info.lo_number; |
e77cab82 YW |
927 | |
928 | if (sd_device_get_sysattr_value(dev, "loop/backing_file", &s) >= 0) { | |
929 | backing_file = strdup(s); | |
930 | if (!backing_file) | |
931 | return -ENOMEM; | |
932 | } | |
4d2a9e3e LP |
933 | |
934 | backing_devno = info.lo_device; | |
935 | backing_inode = info.lo_inode; | |
a8d8a619 | 936 | } |
b26c39ad | 937 | |
4f0ad43e | 938 | r = fd_get_diskseq(fd, &diskseq); |
ffcb3324 YW |
939 | if (r < 0 && r != -EOPNOTSUPP) |
940 | return r; | |
941 | ||
22ee78a8 LP |
942 | uint32_t sector_size; |
943 | r = blockdev_get_sector_size(fd, §or_size); | |
944 | if (r < 0) | |
945 | return r; | |
946 | ||
4f0ad43e YW |
947 | r = sd_device_get_devnum(dev, &devnum); |
948 | if (r < 0) | |
949 | return r; | |
7f52206a | 950 | |
4f0ad43e | 951 | r = sd_device_get_devname(dev, &s); |
cc5bae6c YW |
952 | if (r < 0) |
953 | return r; | |
954 | ||
4f0ad43e YW |
955 | node = strdup(s); |
956 | if (!node) | |
cc5bae6c | 957 | return -ENOMEM; |
9dabc4fd LP |
958 | |
959 | d = new(LoopDevice, 1); | |
960 | if (!d) | |
961 | return -ENOMEM; | |
962 | ||
963 | *d = (LoopDevice) { | |
36d5eb0b | 964 | .n_ref = 1, |
a8d8a619 | 965 | .fd = TAKE_FD(fd), |
7f52206a | 966 | .lock_fd = TAKE_FD(lock_fd), |
b26c39ad | 967 | .nr = nr, |
4f0ad43e YW |
968 | .node = TAKE_PTR(node), |
969 | .dev = sd_device_ref(dev), | |
e77cab82 | 970 | .backing_file = TAKE_PTR(backing_file), |
4d2a9e3e LP |
971 | .backing_inode = backing_inode, |
972 | .backing_devno = backing_devno, | |
9dabc4fd | 973 | .relinquished = true, /* It's not ours, don't try to destroy it when this object is freed */ |
4f0ad43e | 974 | .devno = devnum, |
ffcb3324 | 975 | .diskseq = diskseq, |
31c75fcc | 976 | .uevent_seqnum_not_before = UINT64_MAX, |
8ede1e86 | 977 | .timestamp_not_before = USEC_INFINITY, |
22ee78a8 | 978 | .sector_size = sector_size, |
9dabc4fd LP |
979 | }; |
980 | ||
981 | *ret = d; | |
4f0ad43e YW |
982 | return 0; |
983 | } | |
984 | ||
985 | int loop_device_open_from_fd( | |
986 | int fd, | |
987 | int open_flags, | |
988 | int lock_op, | |
989 | LoopDevice **ret) { | |
990 | ||
991 | _cleanup_(sd_device_unrefp) sd_device *dev = NULL; | |
992 | int r; | |
993 | ||
8e398254 | 994 | r = block_device_new_from_fd(ASSERT_FD(fd), 0, &dev); |
4f0ad43e YW |
995 | if (r < 0) |
996 | return r; | |
997 | ||
998 | return loop_device_open(dev, open_flags, lock_op, ret); | |
999 | } | |
1000 | ||
1001 | int loop_device_open_from_path( | |
1002 | const char *path, | |
1003 | int open_flags, | |
1004 | int lock_op, | |
1005 | LoopDevice **ret) { | |
1006 | ||
1007 | _cleanup_(sd_device_unrefp) sd_device *dev = NULL; | |
1008 | int r; | |
1009 | ||
1010 | assert(path); | |
1011 | ||
1012 | r = block_device_new_from_path(path, 0, &dev); | |
1013 | if (r < 0) | |
1014 | return r; | |
1015 | ||
1016 | return loop_device_open(dev, open_flags, lock_op, ret); | |
9dabc4fd LP |
1017 | } |
1018 | ||
f1443709 LP |
1019 | static int resize_partition(int partition_fd, uint64_t offset, uint64_t size) { |
1020 | char sysfs[STRLEN("/sys/dev/block/:/partition") + 2*DECIMAL_STR_MAX(dev_t) + 1]; | |
ca822829 | 1021 | _cleanup_free_ char *buffer = NULL; |
f1443709 | 1022 | uint64_t current_offset, current_size, partno; |
254d1313 | 1023 | _cleanup_close_ int whole_fd = -EBADF; |
f1443709 LP |
1024 | struct stat st; |
1025 | dev_t devno; | |
1026 | int r; | |
1027 | ||
f1443709 LP |
1028 | /* Resizes the partition the loopback device refer to (assuming it refers to one instead of an actual |
1029 | * loopback device), and changes the offset, if needed. This is a fancy wrapper around | |
1030 | * BLKPG_RESIZE_PARTITION. */ | |
1031 | ||
8e398254 | 1032 | if (fstat(ASSERT_FD(partition_fd), &st) < 0) |
f1443709 LP |
1033 | return -errno; |
1034 | ||
1035 | assert(S_ISBLK(st.st_mode)); | |
1036 | ||
ed13feff | 1037 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/partition", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
1038 | r = read_one_line_file(sysfs, &buffer); |
1039 | if (r == -ENOENT) /* not a partition, cannot resize */ | |
1040 | return -ENOTTY; | |
1041 | if (r < 0) | |
1042 | return r; | |
1043 | r = safe_atou64(buffer, &partno); | |
1044 | if (r < 0) | |
1045 | return r; | |
1046 | ||
ed13feff | 1047 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/start", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
1048 | |
1049 | buffer = mfree(buffer); | |
1050 | r = read_one_line_file(sysfs, &buffer); | |
1051 | if (r < 0) | |
1052 | return r; | |
1053 | r = safe_atou64(buffer, ¤t_offset); | |
1054 | if (r < 0) | |
1055 | return r; | |
1056 | if (current_offset > UINT64_MAX/512U) | |
1057 | return -EINVAL; | |
1058 | current_offset *= 512U; | |
1059 | ||
1060 | if (ioctl(partition_fd, BLKGETSIZE64, ¤t_size) < 0) | |
1061 | return -EINVAL; | |
1062 | ||
1063 | if (size == UINT64_MAX && offset == UINT64_MAX) | |
1064 | return 0; | |
1065 | if (current_size == size && current_offset == offset) | |
1066 | return 0; | |
1067 | ||
ed13feff | 1068 | xsprintf(sysfs, "/sys/dev/block/" DEVNUM_FORMAT_STR "/../dev", DEVNUM_FORMAT_VAL(st.st_rdev)); |
f1443709 LP |
1069 | |
1070 | buffer = mfree(buffer); | |
1071 | r = read_one_line_file(sysfs, &buffer); | |
1072 | if (r < 0) | |
1073 | return r; | |
7176f06c | 1074 | r = parse_devnum(buffer, &devno); |
f1443709 LP |
1075 | if (r < 0) |
1076 | return r; | |
1077 | ||
ca822829 | 1078 | whole_fd = r = device_open_from_devnum(S_IFBLK, devno, O_RDWR|O_CLOEXEC|O_NONBLOCK|O_NOCTTY, NULL); |
f1443709 LP |
1079 | if (r < 0) |
1080 | return r; | |
1081 | ||
91e1ce1a LP |
1082 | return block_device_resize_partition( |
1083 | whole_fd, | |
1084 | partno, | |
1085 | offset == UINT64_MAX ? current_offset : offset, | |
1086 | size == UINT64_MAX ? current_size : size); | |
f1443709 LP |
1087 | } |
1088 | ||
c37878fc LP |
1089 | int loop_device_refresh_size(LoopDevice *d, uint64_t offset, uint64_t size) { |
1090 | struct loop_info64 info; | |
ff27ef4b | 1091 | |
9dabc4fd | 1092 | assert(d); |
ff27ef4b | 1093 | assert(d->fd >= 0); |
9dabc4fd | 1094 | |
f1443709 LP |
1095 | /* Changes the offset/start of the loop device relative to the beginning of the underlying file or |
1096 | * block device. If this loop device actually refers to a partition and not a loopback device, we'll | |
1097 | * try to adjust the partition offsets instead. | |
1098 | * | |
1099 | * If either offset or size is UINT64_MAX we won't change that parameter. */ | |
1100 | ||
f1443709 LP |
1101 | if (d->nr < 0) /* not a loopback device */ |
1102 | return resize_partition(d->fd, offset, size); | |
1103 | ||
c37878fc LP |
1104 | if (ioctl(d->fd, LOOP_GET_STATUS64, &info) < 0) |
1105 | return -errno; | |
1106 | ||
10c1b188 LP |
1107 | #if HAVE_VALGRIND_MEMCHECK_H |
1108 | /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */ | |
1109 | VALGRIND_MAKE_MEM_DEFINED(&info, sizeof(info)); | |
1110 | #endif | |
1111 | ||
c37878fc LP |
1112 | if (size == UINT64_MAX && offset == UINT64_MAX) |
1113 | return 0; | |
1114 | if (info.lo_sizelimit == size && info.lo_offset == offset) | |
1115 | return 0; | |
1116 | ||
1117 | if (size != UINT64_MAX) | |
1118 | info.lo_sizelimit = size; | |
1119 | if (offset != UINT64_MAX) | |
1120 | info.lo_offset = offset; | |
1121 | ||
7c248223 | 1122 | return RET_NERRNO(ioctl(d->fd, LOOP_SET_STATUS64, &info)); |
9dabc4fd | 1123 | } |
441ec804 LP |
1124 | |
1125 | int loop_device_flock(LoopDevice *d, int operation) { | |
7f52206a | 1126 | assert(IN_SET(operation & ~LOCK_NB, LOCK_UN, LOCK_SH, LOCK_EX)); |
441ec804 LP |
1127 | assert(d); |
1128 | ||
7f52206a LP |
1129 | /* When unlocking just close the lock fd */ |
1130 | if ((operation & ~LOCK_NB) == LOCK_UN) { | |
1131 | d->lock_fd = safe_close(d->lock_fd); | |
1132 | return 0; | |
1133 | } | |
1134 | ||
1135 | /* If we had no lock fd so far, create one and lock it right-away */ | |
1136 | if (d->lock_fd < 0) { | |
8e398254 | 1137 | d->lock_fd = open_lock_fd(ASSERT_FD(d->fd), operation); |
7f52206a LP |
1138 | if (d->lock_fd < 0) |
1139 | return d->lock_fd; | |
1140 | ||
1141 | return 0; | |
1142 | } | |
441ec804 | 1143 | |
7f52206a LP |
1144 | /* Otherwise change the current lock mode on the existing fd */ |
1145 | return RET_NERRNO(flock(d->lock_fd, operation)); | |
441ec804 | 1146 | } |
8dbc208c LP |
1147 | |
1148 | int loop_device_sync(LoopDevice *d) { | |
1149 | assert(d); | |
1150 | ||
1151 | /* We also do this implicitly in loop_device_unref(). Doing this explicitly here has the benefit that | |
1152 | * we can check the return value though. */ | |
1153 | ||
8e398254 | 1154 | return RET_NERRNO(fsync(ASSERT_FD(d->fd))); |
8dbc208c | 1155 | } |
d2430d50 LP |
1156 | |
1157 | int loop_device_set_autoclear(LoopDevice *d, bool autoclear) { | |
1158 | struct loop_info64 info; | |
1159 | ||
1160 | assert(d); | |
1161 | ||
8e398254 | 1162 | if (ioctl(ASSERT_FD(d->fd), LOOP_GET_STATUS64, &info) < 0) |
d2430d50 LP |
1163 | return -errno; |
1164 | ||
1165 | if (autoclear == FLAGS_SET(info.lo_flags, LO_FLAGS_AUTOCLEAR)) | |
1166 | return 0; | |
1167 | ||
1168 | SET_FLAG(info.lo_flags, LO_FLAGS_AUTOCLEAR, autoclear); | |
1169 | ||
1170 | if (ioctl(d->fd, LOOP_SET_STATUS64, &info) < 0) | |
1171 | return -errno; | |
1172 | ||
1173 | return 1; | |
1174 | } | |
999ac3e2 LP |
1175 | |
1176 | int loop_device_set_filename(LoopDevice *d, const char *name) { | |
1177 | struct loop_info64 info; | |
1178 | ||
1179 | assert(d); | |
1180 | ||
1181 | /* Sets the .lo_file_name of the loopback device. This is supposed to contain the path to the file | |
1182 | * backing the block device, but is actually just a free-form string you can pass to the kernel. Most | |
1183 | * tools that actually care for the backing file path use the sysfs attribute file loop/backing_file | |
7a05926f YW |
1184 | * which is a kernel generated string, subject to file system namespaces and such. |
1185 | * | |
1186 | * .lo_file_name is useful since userspace can select it freely when creating a loopback block | |
1187 | * device, and we can use it for /dev/disk/by-loop-ref/ symlinks, and similar, so that apps can | |
1188 | * recognize their own loopback files. */ | |
999ac3e2 LP |
1189 | |
1190 | if (name && strlen(name) >= sizeof(info.lo_file_name)) | |
1191 | return -ENOBUFS; | |
1192 | ||
8e398254 | 1193 | if (ioctl(ASSERT_FD(d->fd), LOOP_GET_STATUS64, &info) < 0) |
999ac3e2 LP |
1194 | return -errno; |
1195 | ||
1196 | if (strneq((char*) info.lo_file_name, strempty(name), sizeof(info.lo_file_name))) | |
1197 | return 0; | |
1198 | ||
1199 | if (name) { | |
1200 | strncpy((char*) info.lo_file_name, name, sizeof(info.lo_file_name)-1); | |
1201 | info.lo_file_name[sizeof(info.lo_file_name)-1] = 0; | |
1202 | } else | |
1203 | memzero(info.lo_file_name, sizeof(info.lo_file_name)); | |
1204 | ||
1205 | if (ioctl(d->fd, LOOP_SET_STATUS64, &info) < 0) | |
1206 | return -errno; | |
1207 | ||
1208 | return 1; | |
1209 | } |