1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
3 #if HAVE_VALGRIND_MEMCHECK_H
4 #include <valgrind/memcheck.h>
9 #include <linux/blkpg.h>
11 #include <linux/loop.h>
13 #include <sys/ioctl.h>
16 #include "sd-device.h"
18 #include "alloc-util.h"
19 #include "blockdev-util.h"
20 #include "data-fd-util.h"
21 #include "device-util.h"
22 #include "devnum-util.h"
23 #include "dissect-image.h"
25 #include "errno-util.h"
29 #include "loop-util.h"
30 #include "missing_loop.h"
31 #include "parse-util.h"
32 #include "path-util.h"
33 #include "random-util.h"
34 #include "stat-util.h"
35 #include "stdio-util.h"
36 #include "string-util.h"
37 #include "tmpfile-util.h"
39 static void cleanup_clear_loop_close(int *fd
) {
43 (void) ioctl(*fd
, LOOP_CLR_FD
);
44 (void) safe_close(*fd
);
47 static int loop_is_bound(int fd
) {
48 struct loop_info64 info
;
50 if (ioctl(ASSERT_FD(fd
), LOOP_GET_STATUS64
, &info
) < 0) {
52 return false; /* not bound! */
57 return true; /* bound! */
60 static int open_lock_fd(int primary_fd
, int operation
) {
61 _cleanup_close_
int lock_fd
= -EBADF
;
63 assert(IN_SET(operation
& ~LOCK_NB
, LOCK_SH
, LOCK_EX
));
65 lock_fd
= fd_reopen(ASSERT_FD(primary_fd
), O_RDONLY
|O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
);
69 if (flock(lock_fd
, operation
) < 0)
72 return TAKE_FD(lock_fd
);
75 static int loop_configure_verify_direct_io(int fd
, const struct loop_config
*c
) {
79 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_DIRECT_IO
)) {
80 struct loop_info64 info
;
82 if (ioctl(fd
, LOOP_GET_STATUS64
, &info
) < 0)
83 return log_debug_errno(errno
, "Failed to issue LOOP_GET_STATUS64: %m");
85 #if HAVE_VALGRIND_MEMCHECK_H
86 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
89 /* On older kernels (<= 5.3) it was necessary to set the block size of the loopback block
90 * device to the logical block size of the underlying file system. Since there was no nice
91 * way to query the value, we are not bothering to do this however. On newer kernels the
92 * block size is propagated automatically and does not require intervention from us. We'll
93 * check here if enabling direct IO worked, to make this easily debuggable however.
95 * (Should anyone really care and actually wants direct IO on old kernels: it might be worth
96 * enabling direct IO with iteratively larger block sizes until it eventually works.)
98 * On older kernels (e.g.: 5.10) when this is attempted on a file stored on a dm-crypt
99 * backed partition the kernel will start returning I/O errors when accessing the mounted
100 * loop device, so return a recognizable error that causes the operation to be started
101 * from scratch without the LO_FLAGS_DIRECT_IO flag. */
102 if (!FLAGS_SET(info
.lo_flags
, LO_FLAGS_DIRECT_IO
))
103 return log_debug_errno(
104 SYNTHETIC_ERRNO(ENOANO
),
105 "Could not enable direct IO mode, retrying in buffered IO mode.");
111 static int loop_configure_verify(int fd
, const struct loop_config
*c
) {
118 if (c
->block_size
!= 0) {
121 r
= blockdev_get_sector_size(fd
, &ssz
);
125 if (ssz
!= c
->block_size
) {
126 log_debug("LOOP_CONFIGURE didn't honour requested block size %" PRIu32
", got %" PRIu32
" instead. Ignoring.", c
->block_size
, ssz
);
131 if (c
->info
.lo_sizelimit
!= 0) {
132 /* Kernel 5.8 vanilla doesn't properly propagate the size limit into the
133 * block device. If it's used, let's immediately check if it had the desired
134 * effect hence. And if not use classic LOOP_SET_STATUS64. */
137 r
= blockdev_get_device_size(fd
, &z
);
141 if (z
!= c
->info
.lo_sizelimit
) {
142 log_debug("LOOP_CONFIGURE is broken, doesn't honour .info.lo_sizelimit. Falling back to LOOP_SET_STATUS64.");
147 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_PARTSCAN
)) {
148 /* Kernel 5.8 vanilla doesn't properly propagate the partition scanning flag
149 * into the block device. Let's hence verify if things work correctly here
150 * before returning. */
152 r
= blockdev_partscan_enabled(fd
);
156 log_debug("LOOP_CONFIGURE is broken, doesn't honour LO_FLAGS_PARTSCAN. Falling back to LOOP_SET_STATUS64.");
161 r
= loop_configure_verify_direct_io(fd
, c
);
168 static int loop_configure_fallback(int fd
, const struct loop_config
*c
) {
169 struct loop_info64 info_copy
;
175 /* Only some of the flags LOOP_CONFIGURE can set are also settable via LOOP_SET_STATUS64, hence mask
178 info_copy
.lo_flags
&= LOOP_SET_STATUS_SETTABLE_FLAGS
;
180 /* Since kernel commit 5db470e229e22b7eda6e23b5566e532c96fb5bc3 (kernel v5.0) the LOOP_SET_STATUS64
181 * ioctl can return EAGAIN in case we change the info.lo_offset field, if someone else is accessing the
182 * block device while we try to reconfigure it. This is a pretty common case, since udev might
183 * instantly start probing the device as soon as we attach an fd to it. Hence handle it in two ways:
184 * first, let's take the BSD lock to ensure that udev will not step in between the point in
185 * time where we attach the fd and where we reconfigure the device. Secondly, let's wait 50ms on
186 * EAGAIN and retry. The former should be an efficient mechanism to avoid we have to wait 50ms
187 * needlessly if we are just racing against udev. The latter is protection against all other cases,
188 * i.e. peers that do not take the BSD lock. */
190 for (unsigned n_attempts
= 0;;) {
191 if (ioctl(fd
, LOOP_SET_STATUS64
, &info_copy
) >= 0)
194 if (errno
!= EAGAIN
|| ++n_attempts
>= 64)
195 return log_debug_errno(errno
, "Failed to configure loopback block device: %m");
197 /* Sleep some random time, but at least 10ms, at most 250ms. Increase the delay the more
198 * failed attempts we see */
199 (void) usleep_safe(UINT64_C(10) * USEC_PER_MSEC
+
200 random_u64_range(UINT64_C(240) * USEC_PER_MSEC
* n_attempts
/64));
203 /* Work around a kernel bug, where changing offset/size of the loopback device doesn't correctly
204 * invalidate the buffer cache. For details see:
206 * https://android.googlesource.com/platform/system/apex/+/bef74542fbbb4cd629793f4efee8e0053b360570
208 * This was fixed in kernel 5.0, see:
210 * https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5db470e229e22b7eda6e23b5566e532c96fb5bc3
212 * We'll run the work-around here in the legacy LOOP_SET_STATUS64 codepath. In the LOOP_CONFIGURE
213 * codepath above it should not be necessary. */
214 if (c
->info
.lo_offset
!= 0 || c
->info
.lo_sizelimit
!= 0)
215 if (ioctl(fd
, BLKFLSBUF
, 0) < 0)
216 log_debug_errno(errno
, "Failed to issue BLKFLSBUF ioctl, ignoring: %m");
218 /* If a block size is requested then try to configure it. If that doesn't work, ignore errors, but
219 * afterwards, let's validate what is in effect, and if it doesn't match what we want, fail */
220 if (c
->block_size
!= 0) {
223 if (ioctl(fd
, LOOP_SET_BLOCK_SIZE
, (unsigned long) c
->block_size
) < 0)
224 log_debug_errno(errno
, "Failed to set sector size, ignoring: %m");
226 r
= blockdev_get_sector_size(fd
, &ssz
);
228 return log_debug_errno(r
, "Failed to read sector size: %m");
229 if (ssz
!= c
->block_size
)
230 return log_debug_errno(SYNTHETIC_ERRNO(EIO
), "Sector size of loopback device doesn't match what we requested, refusing.");
233 /* LO_FLAGS_DIRECT_IO is a flags we need to configure via explicit ioctls. */
234 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_DIRECT_IO
))
235 if (ioctl(fd
, LOOP_SET_DIRECT_IO
, 1UL) < 0)
236 log_debug_errno(errno
, "Failed to enable direct IO mode, ignoring: %m");
238 return loop_configure_verify_direct_io(fd
, c
);
241 static int loop_configure(
245 const struct loop_config
*c
,
248 static bool loop_configure_broken
= false;
250 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
251 _cleanup_(cleanup_clear_loop_close
) int loop_with_fd
= -EBADF
; /* This must be declared before lock_fd. */
252 _cleanup_close_
int fd
= -EBADF
, lock_fd
= -EBADF
;
253 _cleanup_free_
char *node
= NULL
;
254 uint64_t diskseq
= 0;
262 if (asprintf(&node
, "/dev/loop%i", nr
) < 0)
263 return log_oom_debug();
265 r
= sd_device_new_from_devname(&dev
, node
);
267 return log_debug_errno(r
, "Failed to create sd_device object for \"%s\": %m", node
);
269 r
= sd_device_get_devnum(dev
, &devno
);
271 return log_device_debug_errno(dev
, r
, "Failed to get devnum: %m");
273 fd
= sd_device_open(dev
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|open_flags
);
275 return log_device_debug_errno(dev
, fd
, "Failed to open device: %m");
277 /* Let's lock the device before we do anything. We take the BSD lock on a second, separately opened
278 * fd for the device. udev after all watches for close() events (specifically IN_CLOSE_WRITE) on
279 * block devices to reprobe them, hence by having a separate fd we will later close() we can ensure
280 * we trigger udev after everything is done. If we'd lock our own fd instead and keep it open for a
281 * long time udev would possibly never run on it again, even though the fd is unlocked, simply
282 * because we never close() it. It also has the nice benefit we can use the _cleanup_close_ logic to
283 * automatically release the lock, after we are done. */
284 lock_fd
= open_lock_fd(fd
, LOCK_EX
);
286 return log_device_debug_errno(dev
, lock_fd
, "Failed to acquire lock: %m");
288 log_device_debug(dev
, "Acquired exclusive lock.");
290 /* Let's see if backing file is really unattached. Someone may already attach a backing file without
291 * taking BSD lock. */
292 r
= loop_is_bound(fd
);
294 return log_device_debug_errno(dev
, r
, "Failed to check if the loopback block device is bound: %m");
296 return log_device_debug_errno(dev
, SYNTHETIC_ERRNO(EBUSY
),
297 "The loopback block device is already bound, ignoring.");
299 /* Let's see if the device is really detached, i.e. currently has no associated partition block
300 * devices. On various kernels (such as 5.8) it is possible to have a loopback block device that
301 * superficially is detached but still has partition block devices associated for it. Let's then
302 * manually remove the partitions via BLKPG, and tell the caller we did that via EUCLEAN, so they try
304 r
= block_device_remove_all_partitions(dev
, fd
);
306 return log_device_debug_errno(dev
, r
, "Failed to remove partitions on the loopback block device: %m");
308 /* Removed all partitions. Let's report this to the caller, to try again, and count this as
310 return log_device_debug_errno(dev
, SYNTHETIC_ERRNO(EUCLEAN
),
311 "Removed partitions on the loopback block device.");
313 if (!loop_configure_broken
) {
314 if (ioctl(fd
, LOOP_CONFIGURE
, c
) < 0) {
315 /* Do fallback only if LOOP_CONFIGURE is not supported, propagate all other
316 * errors. Note that the kernel is weird: non-existing ioctls currently return EINVAL
317 * rather than ENOTTY on loopback block devices. They should fix that in the kernel,
318 * but in the meantime we accept both here. */
319 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && errno
!= EINVAL
)
320 return log_device_debug_errno(dev
, errno
, "ioctl(LOOP_CONFIGURE) failed: %m");
322 loop_configure_broken
= true;
324 loop_with_fd
= TAKE_FD(fd
);
326 r
= loop_configure_verify(loop_with_fd
, c
);
328 return log_device_debug_errno(dev
, r
, "Failed to verify if loopback block device is correctly configured: %m");
330 /* LOOP_CONFIGURE doesn't work. Remember that. */
331 loop_configure_broken
= true;
333 /* We return EBUSY here instead of retrying immediately with LOOP_SET_FD,
334 * because LOOP_CLR_FD is async: if the operation cannot be executed right
335 * away it just sets the autoclear flag on the device. This means there's a
336 * good chance we cannot actually reuse the loopback device right-away. Hence
337 * let's assume it's busy, avoid the trouble and let the calling loop call us
338 * again with a new, likely unused device. */
344 if (loop_configure_broken
) {
345 if (ioctl(fd
, LOOP_SET_FD
, c
->fd
) < 0)
346 return log_device_debug_errno(dev
, errno
, "ioctl(LOOP_SET_FD) failed: %m");
348 loop_with_fd
= TAKE_FD(fd
);
350 r
= loop_configure_fallback(loop_with_fd
, c
);
355 r
= fd_get_diskseq(loop_with_fd
, &diskseq
);
356 if (r
< 0 && r
!= -EOPNOTSUPP
)
357 return log_device_debug_errno(dev
, r
, "Failed to get diskseq: %m");
359 switch (lock_op
& ~LOCK_NB
) {
360 case LOCK_EX
: /* Already in effect */
362 case LOCK_SH
: /* Downgrade */
363 if (flock(lock_fd
, lock_op
) < 0)
364 return log_device_debug_errno(dev
, errno
, "Failed to downgrade lock level: %m");
366 case LOCK_UN
: /* Release */
367 lock_fd
= safe_close(lock_fd
);
370 assert_not_reached();
373 uint64_t device_size
;
374 r
= blockdev_get_device_size(loop_with_fd
, &device_size
);
376 return log_device_debug_errno(dev
, r
, "Failed to get loopback device size: %m");
378 LoopDevice
*d
= new(LoopDevice
, 1);
380 return log_oom_debug();
384 .fd
= TAKE_FD(loop_with_fd
),
385 .lock_fd
= TAKE_FD(lock_fd
),
386 .node
= TAKE_PTR(node
),
389 .dev
= TAKE_PTR(dev
),
391 .sector_size
= c
->block_size
,
392 .device_size
= device_size
,
400 static int loop_device_make_internal(
406 uint32_t sector_size
,
411 _cleanup_(loop_device_unrefp
) LoopDevice
*d
= NULL
;
412 _cleanup_close_
int reopened_fd
= -EBADF
, control
= -EBADF
;
413 _cleanup_free_
char *backing_file
= NULL
;
414 struct loop_config config
;
419 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
421 if (fstat(ASSERT_FD(fd
), &st
) < 0)
424 if (S_ISBLK(st
.st_mode
)) {
425 if (offset
== 0 && IN_SET(size
, 0, UINT64_MAX
))
426 /* If this is already a block device and we are supposed to cover the whole of it
427 * then store an fd to the original open device node — and do not actually create an
428 * unnecessary loopback device for it. */
429 return loop_device_open_from_fd(fd
, open_flags
, lock_op
, ret
);
431 r
= stat_verify_regular(&st
);
437 r
= path_make_absolute_cwd(path
, &backing_file
);
441 path_simplify(backing_file
);
443 r
= fd_get_path(fd
, &backing_file
);
448 f_flags
= fcntl(fd
, F_GETFL
);
452 if (FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) != FLAGS_SET(f_flags
, O_DIRECT
)) {
453 /* If LO_FLAGS_DIRECT_IO is requested, then make sure we have the fd open with O_DIRECT, as
454 * that's required. Conversely, if it's off require that O_DIRECT is off too (that's because
455 * new kernels will implicitly enable LO_FLAGS_DIRECT_IO if O_DIRECT is set).
457 * Our intention here is that LO_FLAGS_DIRECT_IO is the primary knob, and O_DIRECT derived
458 * from that automatically. */
460 reopened_fd
= fd_reopen(fd
, (FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) ? O_DIRECT
: 0)|O_CLOEXEC
|O_NONBLOCK
|open_flags
);
461 if (reopened_fd
< 0) {
462 if (!FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
))
463 return log_debug_errno(reopened_fd
, "Failed to reopen file descriptor without O_DIRECT: %m");
465 /* Some file systems might not support O_DIRECT, let's gracefully continue without it then. */
466 log_debug_errno(reopened_fd
, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without.");
467 loop_flags
&= ~LO_FLAGS_DIRECT_IO
;
469 fd
= reopened_fd
; /* From now on, operate on our new O_DIRECT fd */
472 control
= open("/dev/loop-control", O_RDWR
|O_CLOEXEC
|O_NOCTTY
|O_NONBLOCK
);
476 if (sector_size
== 0)
477 /* If no sector size is specified, default to the classic default */
479 else if (sector_size
== UINT32_MAX
) {
481 if (S_ISBLK(st
.st_mode
))
482 /* If the sector size is specified as UINT32_MAX we'll propagate the sector size of
483 * the underlying block device. */
484 r
= blockdev_get_sector_size(fd
, §or_size
);
486 _cleanup_close_
int non_direct_io_fd
= -EBADF
;
489 assert(S_ISREG(st
.st_mode
));
491 /* If sector size is specified as UINT32_MAX, we'll try to probe the right sector
492 * size of the image in question by looking for the GPT partition header at various
493 * offsets. This of course only works if the image already has a disk label.
495 * So here we actually want to read the file contents ourselves. This is quite likely
496 * not going to work if we managed to enable O_DIRECT, because in such a case there
497 * are some pretty strict alignment requirements to offset, size and target, but
498 * there's no way to query what alignment specifically is actually required. Hence,
499 * let's avoid the mess, and temporarily open an fd without O_DIRECT for the probing
502 if (FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
)) {
503 non_direct_io_fd
= fd_reopen(fd
, O_RDONLY
|O_CLOEXEC
|O_NONBLOCK
);
504 if (non_direct_io_fd
< 0)
505 return non_direct_io_fd
;
507 probe_fd
= non_direct_io_fd
;
511 r
= probe_sector_size(probe_fd
, §or_size
);
517 config
= (struct loop_config
) {
519 .block_size
= sector_size
,
521 /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */
522 .lo_flags
= (loop_flags
& ~LO_FLAGS_READ_ONLY
) | ((open_flags
& O_ACCMODE
) == O_RDONLY
? LO_FLAGS_READ_ONLY
: 0) | LO_FLAGS_AUTOCLEAR
,
524 .lo_sizelimit
= size
== UINT64_MAX
? 0 : size
,
528 /* Loop around LOOP_CTL_GET_FREE, since at the moment we attempt to open the returned device it might
529 * be gone already, taken by somebody else racing against us. */
530 for (unsigned n_attempts
= 0;;) {
534 /* Let's take a lock on the control device first. On a busy system, where many programs
535 * attempt to allocate a loopback device at the same time, we might otherwise keep looping
536 * around relatively heavy operations: asking for a free loopback device, then opening it,
537 * validating it, attaching something to it. Let's serialize this whole operation, to make
538 * unnecessary busywork less likely. Note that this is just something we do to optimize our
539 * own code (and whoever else decides to use LOCK_EX locks for this), taking this lock is not
540 * necessary, it just means it's less likely we have to iterate through this loop again and
541 * again if our own code races against our own code.
543 * Note: our lock protocol is to take the /dev/loop-control lock first, and the block device
544 * lock second, if both are taken, and always in this order, to avoid ABBA locking issues. */
545 if (flock(control
, LOCK_EX
) < 0)
548 nr
= ioctl(control
, LOOP_CTL_GET_FREE
);
552 r
= loop_configure(nr
, open_flags
, lock_op
, &config
, &d
);
556 /* -ENODEV or friends: Somebody might've gotten the same number from the kernel, used the
557 * device, and called LOOP_CTL_REMOVE on it. Let's retry with a new number.
558 * -EBUSY: a file descriptor is already bound to the loopback block device.
559 * -EUCLEAN: some left-over partition devices that were cleaned up.
560 * -ENOANO: we tried to use LO_FLAGS_DIRECT_IO but the kernel rejected it. */
561 if (!ERRNO_IS_DEVICE_ABSENT(r
) && !IN_SET(r
, -EBUSY
, -EUCLEAN
, -ENOANO
))
564 /* OK, this didn't work, let's try again a bit later, but first release the lock on the
566 if (flock(control
, LOCK_UN
) < 0)
569 if (++n_attempts
>= 64) /* Give up eventually */
572 /* If we failed to enable direct IO mode, let's retry without it. We restart the process as
573 * on some combination of kernel version and storage filesystem, the kernel is very unhappy
574 * about a failed DIRECT_IO enablement and throws I/O errors. */
575 if (r
== -ENOANO
&& FLAGS_SET(config
.info
.lo_flags
, LO_FLAGS_DIRECT_IO
)) {
576 config
.info
.lo_flags
&= ~LO_FLAGS_DIRECT_IO
;
577 open_flags
&= ~O_DIRECT
;
579 int non_direct_io_fd
= fd_reopen(config
.fd
, O_CLOEXEC
|O_NONBLOCK
|open_flags
);
580 if (non_direct_io_fd
< 0)
581 return log_debug_errno(
583 "Failed to reopen file descriptor without O_DIRECT: %m");
585 safe_close(reopened_fd
);
586 fd
= config
.fd
= /* For cleanups */ reopened_fd
= non_direct_io_fd
;
589 /* Wait some random time, to make collision less likely. Let's pick a random time in the
590 * range 0ms…250ms, linearly scaled by the number of failed attempts. */
591 usec
= random_u64_range(UINT64_C(10) * USEC_PER_MSEC
+
592 UINT64_C(240) * USEC_PER_MSEC
* n_attempts
/64);
593 log_debug("Trying again after %s.", FORMAT_TIMESPAN(usec
, USEC_PER_MSEC
));
594 (void) usleep_safe(usec
);
597 d
->backing_file
= TAKE_PTR(backing_file
);
598 d
->backing_inode
= st
.st_ino
;
599 d
->backing_devno
= st
.st_dev
;
601 log_debug("Successfully acquired %s, devno=%u:%u, nr=%i, diskseq=%" PRIu64
,
603 major(d
->devno
), minor(d
->devno
),
611 static uint32_t loop_flags_mangle(uint32_t loop_flags
) {
614 r
= getenv_bool("SYSTEMD_LOOP_DIRECT_IO");
615 if (r
< 0 && r
!= -ENXIO
)
616 log_debug_errno(r
, "Failed to parse $SYSTEMD_LOOP_DIRECT_IO, ignoring: %m");
618 return UPDATE_FLAG(loop_flags
, LO_FLAGS_DIRECT_IO
, r
!= 0); /* Turn on LO_FLAGS_DIRECT_IO by default, unless explicitly configured to off. */
621 int loop_device_make(
626 uint32_t sector_size
,
634 return loop_device_make_internal(
641 loop_flags_mangle(loop_flags
),
646 int loop_device_make_by_path_at(
650 uint32_t sector_size
,
655 int r
, basic_flags
, direct_flags
, rdwr_flags
;
656 _cleanup_close_
int fd
= -EBADF
;
659 assert(dir_fd
>= 0 || dir_fd
== AT_FDCWD
);
662 assert(open_flags
< 0 || IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
664 /* Passing < 0 as open_flags here means we'll try to open the device writable if we can, retrying
665 * read-only if we cannot. */
667 loop_flags
= loop_flags_mangle(loop_flags
);
669 /* Let's open with O_DIRECT if we can. But not all file systems support that, hence fall back to
670 * non-O_DIRECT mode automatically, if it fails. */
672 basic_flags
= O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
;
673 direct_flags
= FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) ? O_DIRECT
: 0;
674 rdwr_flags
= open_flags
>= 0 ? open_flags
: O_RDWR
;
676 fd
= xopenat(dir_fd
, path
, basic_flags
|direct_flags
|rdwr_flags
);
677 if (fd
< 0 && direct_flags
!= 0) /* If we had O_DIRECT on, and things failed with that, let's immediately try again without */
678 fd
= xopenat(dir_fd
, path
, basic_flags
|rdwr_flags
);
680 direct
= direct_flags
!= 0;
684 /* Retry read-only? */
685 if (open_flags
>= 0 || !(ERRNO_IS_PRIVILEGE(r
) || r
== -EROFS
))
688 fd
= xopenat(dir_fd
, path
, basic_flags
|direct_flags
|O_RDONLY
);
689 if (fd
< 0 && direct_flags
!= 0) /* as above */
690 fd
= xopenat(dir_fd
, path
, basic_flags
|O_RDONLY
);
692 direct
= direct_flags
!= 0;
694 return r
; /* Propagate original error */
696 open_flags
= O_RDONLY
;
697 } else if (open_flags
< 0)
700 log_debug("Opened '%s' in %s access mode%s, with O_DIRECT %s%s.",
702 open_flags
== O_RDWR
? "O_RDWR" : "O_RDONLY",
703 open_flags
!= rdwr_flags
? " (O_RDWR was requested but not allowed)" : "",
704 direct
? "enabled" : "disabled",
705 direct
!= (direct_flags
!= 0) ? " (O_DIRECT was requested but not supported)" : "");
707 return loop_device_make_internal(
708 dir_fd
== AT_FDCWD
? path
: NULL
,
719 int loop_device_make_by_path_memory(
722 uint32_t sector_size
,
727 _cleanup_close_
int fd
= -EBADF
, mfd
= -EBADF
;
728 _cleanup_free_
char *fn
= NULL
;
733 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
736 loop_flags
&= ~LO_FLAGS_DIRECT_IO
; /* memfds don't support O_DIRECT, hence LO_FLAGS_DIRECT_IO can't be used either */
738 fd
= open(path
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|O_RDONLY
);
742 if (fstat(fd
, &st
) < 0)
745 if (!S_ISREG(st
.st_mode
) && !S_ISBLK(st
.st_mode
))
748 r
= path_extract_filename(path
, &fn
);
752 mfd
= memfd_clone_fd(fd
, fn
, open_flags
|O_CLOEXEC
);
756 fd
= safe_close(fd
); /* Let's close the original early */
758 return loop_device_make_internal(NULL
, mfd
, open_flags
, 0, 0, sector_size
, loop_flags
, lock_op
, ret
);
761 static LoopDevice
* loop_device_free(LoopDevice
*d
) {
762 _cleanup_close_
int control
= -EBADF
;
768 /* Release any lock we might have on the device first. We want to open+lock the /dev/loop-control
769 * device below, but our lock protocol says that if both control and block device locks are taken,
770 * the control lock needs to be taken first, the block device lock second — in order to avoid ABBA
771 * locking issues. Moreover, we want to issue LOOP_CLR_FD on the block device further down, and that
772 * would fail if we had another fd open to the device. */
773 d
->lock_fd
= safe_close(d
->lock_fd
);
775 /* Let's open the control device early, and lock it, so that we can release our block device and
776 * delete it in a synchronized fashion, and allocators won't needlessly see the block device as free
777 * while we are about to delete it. */
778 if (!LOOP_DEVICE_IS_FOREIGN(d
) && !d
->relinquished
) {
779 control
= open("/dev/loop-control", O_RDWR
|O_CLOEXEC
|O_NOCTTY
|O_NONBLOCK
);
781 log_debug_errno(errno
, "Failed to open loop control device, cannot remove loop device '%s', ignoring: %m", strna(d
->node
));
782 else if (flock(control
, LOCK_EX
) < 0)
783 log_debug_errno(errno
, "Failed to lock loop control device, ignoring: %m");
786 /* Then let's release the loopback block device */
788 /* Implicitly sync the device, since otherwise in-flight blocks might not get written */
789 if (fsync(d
->fd
) < 0)
790 log_debug_errno(errno
, "Failed to sync loop block device, ignoring: %m");
792 if (!LOOP_DEVICE_IS_FOREIGN(d
) && !d
->relinquished
) {
793 /* We are supposed to clear the loopback device. Let's do this synchronously: lock
794 * the device, manually remove all partitions and then clear it. This should ensure
795 * udev doesn't concurrently access the devices, and we can be reasonably sure that
796 * once we are done here the device is cleared and all its partition children
797 * removed. Note that we lock our primary device fd here (and not a separate locking
798 * fd, as we do during allocation, since we want to keep the lock all the way through
799 * the LOOP_CLR_FD, but that call would fail if we had more than one fd open.) */
801 if (flock(d
->fd
, LOCK_EX
) < 0)
802 log_debug_errno(errno
, "Failed to lock loop block device, ignoring: %m");
804 r
= block_device_remove_all_partitions(d
->dev
, d
->fd
);
806 log_debug_errno(r
, "Failed to remove partitions of loopback block device, ignoring: %m");
808 if (ioctl(d
->fd
, LOOP_CLR_FD
) < 0)
809 log_debug_errno(errno
, "Failed to clear loop device, ignoring: %m");
815 /* Now that the block device is released, let's also try to remove it */
817 useconds_t delay
= 5 * USEC_PER_MSEC
; /* A total delay of 5090 ms between 39 attempts,
818 * (4*5 + 5*10 + 5*20 + … + 3*640) = 5090. */
820 for (unsigned attempt
= 1;; attempt
++) {
821 if (ioctl(control
, LOOP_CTL_REMOVE
, d
->nr
) >= 0)
823 if (errno
!= EBUSY
|| attempt
> 38) {
824 log_debug_errno(errno
, "Failed to remove device %s: %m", strna(d
->node
));
827 if (attempt
% 5 == 0) {
828 log_debug("Device is still busy after %u attempts…", attempt
);
832 (void) usleep_safe(delay
);
837 sd_device_unref(d
->dev
);
838 free(d
->backing_file
);
842 DEFINE_TRIVIAL_REF_UNREF_FUNC(LoopDevice
, loop_device
, loop_device_free
);
844 void loop_device_relinquish(LoopDevice
*d
) {
847 /* Don't attempt to clean up the loop device anymore from this point on. Leave the clean-ing up to the kernel
848 * itself, using the loop device "auto-clear" logic we already turned on when creating the device. */
850 d
->relinquished
= true;
853 void loop_device_unrelinquish(LoopDevice
*d
) {
855 d
->relinquished
= false;
858 int loop_device_open(
864 _cleanup_close_
int fd
= -EBADF
, lock_fd
= -EBADF
;
865 _cleanup_free_
char *node
= NULL
, *backing_file
= NULL
;
866 dev_t devnum
, backing_devno
= 0;
867 struct loop_info64 info
;
868 ino_t backing_inode
= 0;
869 uint64_t diskseq
= 0;
875 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
878 /* Even if fd is provided through the argument in loop_device_open_from_fd(), we reopen the inode
879 * here, instead of keeping just a dup() clone of it around, since we want to ensure that the
880 * O_DIRECT flag of the handle we keep is off, we have our own file index, and have the right
881 * read/write mode in effect. */
882 fd
= sd_device_open(dev
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|open_flags
);
886 if ((lock_op
& ~LOCK_NB
) != LOCK_UN
) {
887 lock_fd
= open_lock_fd(fd
, lock_op
);
892 if (ioctl(fd
, LOOP_GET_STATUS64
, &info
) >= 0) {
893 #if HAVE_VALGRIND_MEMCHECK_H
894 /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
895 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
899 if (sd_device_get_sysattr_value(dev
, "loop/backing_file", &s
) >= 0) {
900 backing_file
= strdup(s
);
905 backing_devno
= info
.lo_device
;
906 backing_inode
= info
.lo_inode
;
909 r
= fd_get_diskseq(fd
, &diskseq
);
910 if (r
< 0 && r
!= -EOPNOTSUPP
)
913 uint32_t sector_size
;
914 r
= blockdev_get_sector_size(fd
, §or_size
);
918 uint64_t device_size
;
919 r
= blockdev_get_device_size(fd
, &device_size
);
923 r
= sd_device_get_devnum(dev
, &devnum
);
927 r
= sd_device_get_devname(dev
, &s
);
935 d
= new(LoopDevice
, 1);
942 .lock_fd
= TAKE_FD(lock_fd
),
944 .node
= TAKE_PTR(node
),
945 .dev
= sd_device_ref(dev
),
946 .backing_file
= TAKE_PTR(backing_file
),
947 .backing_inode
= backing_inode
,
948 .backing_devno
= backing_devno
,
949 .relinquished
= true, /* It's not ours, don't try to destroy it when this object is freed */
952 .sector_size
= sector_size
,
953 .device_size
= device_size
,
961 int loop_device_open_from_fd(
967 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
970 r
= block_device_new_from_fd(ASSERT_FD(fd
), 0, &dev
);
974 return loop_device_open(dev
, open_flags
, lock_op
, ret
);
977 int loop_device_open_from_path(
983 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
988 r
= block_device_new_from_path(path
, 0, &dev
);
992 return loop_device_open(dev
, open_flags
, lock_op
, ret
);
995 static int resize_partition(int partition_fd
, uint64_t offset
, uint64_t size
) {
996 char sysfs
[STRLEN("/sys/dev/block/:/partition") + 2*DECIMAL_STR_MAX(dev_t
) + 1];
997 _cleanup_free_
char *buffer
= NULL
;
998 uint64_t current_offset
, current_size
, partno
;
999 _cleanup_close_
int whole_fd
= -EBADF
;
1004 /* Resizes the partition the loopback device refer to (assuming it refers to one instead of an actual
1005 * loopback device), and changes the offset, if needed. This is a fancy wrapper around
1006 * BLKPG_RESIZE_PARTITION. */
1008 if (fstat(ASSERT_FD(partition_fd
), &st
) < 0)
1011 assert(S_ISBLK(st
.st_mode
));
1013 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/partition", DEVNUM_FORMAT_VAL(st
.st_rdev
));
1014 r
= read_one_line_file(sysfs
, &buffer
);
1015 if (r
== -ENOENT
) /* not a partition, cannot resize */
1019 r
= safe_atou64(buffer
, &partno
);
1023 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/start", DEVNUM_FORMAT_VAL(st
.st_rdev
));
1025 buffer
= mfree(buffer
);
1026 r
= read_one_line_file(sysfs
, &buffer
);
1029 r
= safe_atou64(buffer
, ¤t_offset
);
1032 if (current_offset
> UINT64_MAX
/512U)
1034 current_offset
*= 512U;
1036 r
= blockdev_get_device_size(partition_fd
, ¤t_size
);
1040 if (size
== UINT64_MAX
&& offset
== UINT64_MAX
)
1042 if (current_size
== size
&& current_offset
== offset
)
1045 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/../dev", DEVNUM_FORMAT_VAL(st
.st_rdev
));
1047 buffer
= mfree(buffer
);
1048 r
= read_one_line_file(sysfs
, &buffer
);
1051 r
= parse_devnum(buffer
, &devno
);
1055 whole_fd
= r
= device_open_from_devnum(S_IFBLK
, devno
, O_RDWR
|O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
, NULL
);
1059 return block_device_resize_partition(
1062 offset
== UINT64_MAX
? current_offset
: offset
,
1063 size
== UINT64_MAX
? current_size
: size
);
1066 int loop_device_refresh_size(LoopDevice
*d
, uint64_t offset
, uint64_t size
) {
1067 struct loop_info64 info
;
1072 /* Changes the offset/start of the loop device relative to the beginning of the underlying file or
1073 * block device. If this loop device actually refers to a partition and not a loopback device, we'll
1074 * try to adjust the partition offsets instead.
1076 * If either offset or size is UINT64_MAX we won't change that parameter. */
1078 if (d
->nr
< 0) /* not a loopback device */
1079 return resize_partition(d
->fd
, offset
, size
);
1081 if (ioctl(d
->fd
, LOOP_GET_STATUS64
, &info
) < 0)
1084 #if HAVE_VALGRIND_MEMCHECK_H
1085 /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
1086 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
1089 if (size
== UINT64_MAX
&& offset
== UINT64_MAX
)
1091 if (info
.lo_sizelimit
== size
&& info
.lo_offset
== offset
)
1094 if (size
!= UINT64_MAX
)
1095 info
.lo_sizelimit
= size
;
1096 if (offset
!= UINT64_MAX
)
1097 info
.lo_offset
= offset
;
1099 return RET_NERRNO(ioctl(d
->fd
, LOOP_SET_STATUS64
, &info
));
1102 int loop_device_flock(LoopDevice
*d
, int operation
) {
1103 assert(IN_SET(operation
& ~LOCK_NB
, LOCK_UN
, LOCK_SH
, LOCK_EX
));
1106 /* When unlocking just close the lock fd */
1107 if ((operation
& ~LOCK_NB
) == LOCK_UN
) {
1108 d
->lock_fd
= safe_close(d
->lock_fd
);
1112 /* If we had no lock fd so far, create one and lock it right-away */
1113 if (d
->lock_fd
< 0) {
1114 d
->lock_fd
= open_lock_fd(ASSERT_FD(d
->fd
), operation
);
1121 /* Otherwise change the current lock mode on the existing fd */
1122 return RET_NERRNO(flock(d
->lock_fd
, operation
));
1125 int loop_device_sync(LoopDevice
*d
) {
1128 /* We also do this implicitly in loop_device_unref(). Doing this explicitly here has the benefit that
1129 * we can check the return value though. */
1131 return RET_NERRNO(fsync(ASSERT_FD(d
->fd
)));
1134 int loop_device_set_autoclear(LoopDevice
*d
, bool autoclear
) {
1135 struct loop_info64 info
;
1139 if (ioctl(ASSERT_FD(d
->fd
), LOOP_GET_STATUS64
, &info
) < 0)
1142 if (autoclear
== FLAGS_SET(info
.lo_flags
, LO_FLAGS_AUTOCLEAR
))
1145 SET_FLAG(info
.lo_flags
, LO_FLAGS_AUTOCLEAR
, autoclear
);
1147 if (ioctl(d
->fd
, LOOP_SET_STATUS64
, &info
) < 0)
1153 int loop_device_set_filename(LoopDevice
*d
, const char *name
) {
1154 struct loop_info64 info
;
1158 /* Sets the .lo_file_name of the loopback device. This is supposed to contain the path to the file
1159 * backing the block device, but is actually just a free-form string you can pass to the kernel. Most
1160 * tools that actually care for the backing file path use the sysfs attribute file loop/backing_file
1161 * which is a kernel generated string, subject to file system namespaces and such.
1163 * .lo_file_name is useful since userspace can select it freely when creating a loopback block
1164 * device, and we can use it for /dev/disk/by-loop-ref/ symlinks, and similar, so that apps can
1165 * recognize their own loopback files. */
1167 if (name
&& strlen(name
) >= sizeof(info
.lo_file_name
))
1170 if (ioctl(ASSERT_FD(d
->fd
), LOOP_GET_STATUS64
, &info
) < 0)
1173 if (strneq((char*) info
.lo_file_name
, strempty(name
), sizeof(info
.lo_file_name
)))
1177 strncpy((char*) info
.lo_file_name
, name
, sizeof(info
.lo_file_name
)-1);
1178 info
.lo_file_name
[sizeof(info
.lo_file_name
)-1] = 0;
1180 memzero(info
.lo_file_name
, sizeof(info
.lo_file_name
));
1182 if (ioctl(d
->fd
, LOOP_SET_STATUS64
, &info
) < 0)