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