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"
24 #include "errno-util.h"
27 #include "loop-util.h"
28 #include "missing_loop.h"
29 #include "parse-util.h"
30 #include "path-util.h"
31 #include "random-util.h"
32 #include "stat-util.h"
33 #include "stdio-util.h"
34 #include "string-util.h"
35 #include "tmpfile-util.h"
37 static void cleanup_clear_loop_close(int *fd
) {
41 (void) ioctl(*fd
, LOOP_CLR_FD
);
42 (void) safe_close(*fd
);
45 static int loop_is_bound(int fd
) {
46 struct loop_info64 info
;
50 if (ioctl(fd
, LOOP_GET_STATUS64
, &info
) < 0) {
52 return false; /* not bound! */
57 return true; /* bound! */
60 static int get_current_uevent_seqnum(uint64_t *ret
) {
61 _cleanup_free_
char *p
= NULL
;
64 r
= read_full_virtual_file("/sys/kernel/uevent_seqnum", &p
, NULL
);
66 return log_debug_errno(r
, "Failed to read current uevent sequence number: %m");
68 r
= safe_atou64(strstrip(p
), ret
);
70 return log_debug_errno(r
, "Failed to parse current uevent sequence number: %s", p
);
75 static int open_lock_fd(int primary_fd
, int operation
) {
76 _cleanup_close_
int lock_fd
= -EBADF
;
78 assert(primary_fd
>= 0);
79 assert(IN_SET(operation
& ~LOCK_NB
, LOCK_SH
, LOCK_EX
));
81 lock_fd
= fd_reopen(primary_fd
, O_RDONLY
|O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
);
85 if (flock(lock_fd
, operation
) < 0)
88 return TAKE_FD(lock_fd
);
91 static int loop_configure_verify_direct_io(int fd
, const struct loop_config
*c
) {
95 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_DIRECT_IO
)) {
96 struct loop_info64 info
;
98 if (ioctl(fd
, LOOP_GET_STATUS64
, &info
) < 0)
99 return log_debug_errno(errno
, "Failed to issue LOOP_GET_STATUS64: %m");
101 #if HAVE_VALGRIND_MEMCHECK_H
102 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
105 /* On older kernels (<= 5.3) it was necessary to set the block size of the loopback block
106 * device to the logical block size of the underlying file system. Since there was no nice
107 * way to query the value, we are not bothering to do this however. On newer kernels the
108 * block size is propagated automatically and does not require intervention from us. We'll
109 * check here if enabling direct IO worked, to make this easily debuggable however.
111 * (Should anyone really care and actually wants direct IO on old kernels: it might be worth
112 * enabling direct IO with iteratively larger block sizes until it eventually works.) */
113 if (!FLAGS_SET(info
.lo_flags
, LO_FLAGS_DIRECT_IO
))
114 log_debug("Could not enable direct IO mode, proceeding in buffered IO mode.");
120 static int loop_configure_verify(int fd
, const struct loop_config
*c
) {
127 if (c
->block_size
!= 0) {
130 if (ioctl(fd
, BLKSSZGET
, &z
) < 0)
134 if ((uint32_t) z
!= c
->block_size
)
135 log_debug("LOOP_CONFIGURE didn't honour requested block size %u, got %i instead. Ignoring.", c
->block_size
, z
);
138 if (c
->info
.lo_sizelimit
!= 0) {
139 /* Kernel 5.8 vanilla doesn't properly propagate the size limit into the
140 * block device. If it's used, let's immediately check if it had the desired
141 * effect hence. And if not use classic LOOP_SET_STATUS64. */
144 if (ioctl(fd
, BLKGETSIZE64
, &z
) < 0)
147 if (z
!= c
->info
.lo_sizelimit
) {
148 log_debug("LOOP_CONFIGURE is broken, doesn't honour .info.lo_sizelimit. Falling back to LOOP_SET_STATUS64.");
153 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_PARTSCAN
)) {
154 /* Kernel 5.8 vanilla doesn't properly propagate the partition scanning flag
155 * into the block device. Let's hence verify if things work correctly here
156 * before returning. */
158 r
= blockdev_partscan_enabled(fd
);
162 log_debug("LOOP_CONFIGURE is broken, doesn't honour LO_FLAGS_PARTSCAN. Falling back to LOOP_SET_STATUS64.");
167 r
= loop_configure_verify_direct_io(fd
, c
);
174 static int loop_configure_fallback(int fd
, const struct loop_config
*c
) {
175 struct loop_info64 info_copy
;
180 /* Only some of the flags LOOP_CONFIGURE can set are also settable via LOOP_SET_STATUS64, hence mask
183 info_copy
.lo_flags
&= LOOP_SET_STATUS_SETTABLE_FLAGS
;
185 /* Since kernel commit 5db470e229e22b7eda6e23b5566e532c96fb5bc3 (kernel v5.0) the LOOP_SET_STATUS64
186 * ioctl can return EAGAIN in case we change the info.lo_offset field, if someone else is accessing the
187 * block device while we try to reconfigure it. This is a pretty common case, since udev might
188 * instantly start probing the device as soon as we attach an fd to it. Hence handle it in two ways:
189 * first, let's take the BSD lock to ensure that udev will not step in between the point in
190 * time where we attach the fd and where we reconfigure the device. Secondly, let's wait 50ms on
191 * EAGAIN and retry. The former should be an efficient mechanism to avoid we have to wait 50ms
192 * needlessly if we are just racing against udev. The latter is protection against all other cases,
193 * i.e. peers that do not take the BSD lock. */
195 for (unsigned n_attempts
= 0;;) {
196 if (ioctl(fd
, LOOP_SET_STATUS64
, &info_copy
) >= 0)
199 if (errno
!= EAGAIN
|| ++n_attempts
>= 64)
200 return log_debug_errno(errno
, "Failed to configure loopback block device: %m");
202 /* Sleep some random time, but at least 10ms, at most 250ms. Increase the delay the more
203 * failed attempts we see */
204 (void) usleep(UINT64_C(10) * USEC_PER_MSEC
+
205 random_u64_range(UINT64_C(240) * USEC_PER_MSEC
* n_attempts
/64));
208 /* Work around a kernel bug, where changing offset/size of the loopback device doesn't correctly
209 * invalidate the buffer cache. For details see:
211 * https://android.googlesource.com/platform/system/apex/+/bef74542fbbb4cd629793f4efee8e0053b360570
213 * This was fixed in kernel 5.0, see:
215 * https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5db470e229e22b7eda6e23b5566e532c96fb5bc3
217 * We'll run the work-around here in the legacy LOOP_SET_STATUS64 codepath. In the LOOP_CONFIGURE
218 * codepath above it should not be necessary. */
219 if (c
->info
.lo_offset
!= 0 || c
->info
.lo_sizelimit
!= 0)
220 if (ioctl(fd
, BLKFLSBUF
, 0) < 0)
221 log_debug_errno(errno
, "Failed to issue BLKFLSBUF ioctl, ignoring: %m");
223 /* LO_FLAGS_DIRECT_IO is a flags we need to configure via explicit ioctls. */
224 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_DIRECT_IO
))
225 if (ioctl(fd
, LOOP_SET_DIRECT_IO
, 1UL) < 0)
226 log_debug_errno(errno
, "Failed to enable direct IO mode, ignoring: %m");
228 return loop_configure_verify_direct_io(fd
, c
);
231 static int loop_configure(
235 const struct loop_config
*c
,
238 static bool loop_configure_broken
= false;
240 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
241 _cleanup_(cleanup_clear_loop_close
) int loop_with_fd
= -EBADF
; /* This must be declared before lock_fd. */
242 _cleanup_close_
int fd
= -EBADF
, lock_fd
= -EBADF
;
243 _cleanup_free_
char *node
= NULL
;
244 uint64_t diskseq
= 0, seqnum
= UINT64_MAX
;
245 usec_t timestamp
= USEC_INFINITY
;
253 if (asprintf(&node
, "/dev/loop%i", nr
) < 0)
256 r
= sd_device_new_from_devname(&dev
, node
);
260 r
= sd_device_get_devnum(dev
, &devno
);
264 fd
= sd_device_open(dev
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|open_flags
);
268 /* Let's lock the device before we do anything. We take the BSD lock on a second, separately opened
269 * fd for the device. udev after all watches for close() events (specifically IN_CLOSE_WRITE) on
270 * block devices to reprobe them, hence by having a separate fd we will later close() we can ensure
271 * we trigger udev after everything is done. If we'd lock our own fd instead and keep it open for a
272 * long time udev would possibly never run on it again, even though the fd is unlocked, simply
273 * because we never close() it. It also has the nice benefit we can use the _cleanup_close_ logic to
274 * automatically release the lock, after we are done. */
275 lock_fd
= open_lock_fd(fd
, LOCK_EX
);
279 /* Let's see if backing file is really unattached. Someone may already attach a backing file without
280 * taking BSD lock. */
281 r
= loop_is_bound(fd
);
287 /* Let's see if the device is really detached, i.e. currently has no associated partition block
288 * devices. On various kernels (such as 5.8) it is possible to have a loopback block device that
289 * superficially is detached but still has partition block devices associated for it. Let's then
290 * manually remove the partitions via BLKPG, and tell the caller we did that via EUCLEAN, so they try
292 r
= block_device_remove_all_partitions(dev
, fd
);
296 /* Removed all partitions. Let's report this to the caller, to try again, and count this as
300 if (!loop_configure_broken
) {
301 /* Acquire uevent seqnum immediately before attaching the loopback device. This allows
302 * callers to ignore all uevents with a seqnum before this one, if they need to associate
303 * uevent with this attachment. Doing so isn't race-free though, as uevents that happen in
304 * the window between this reading of the seqnum, and the LOOP_CONFIGURE call might still be
305 * mistaken as originating from our attachment, even though might be caused by an earlier
306 * use. But doing this at least shortens the race window a bit. */
307 r
= get_current_uevent_seqnum(&seqnum
);
311 timestamp
= now(CLOCK_MONOTONIC
);
313 if (ioctl(fd
, LOOP_CONFIGURE
, c
) < 0) {
314 /* Do fallback only if LOOP_CONFIGURE is not supported, propagate all other
315 * errors. Note that the kernel is weird: non-existing ioctls currently return EINVAL
316 * rather than ENOTTY on loopback block devices. They should fix that in the kernel,
317 * but in the meantime we accept both here. */
318 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && errno
!= EINVAL
)
321 loop_configure_broken
= true;
323 loop_with_fd
= TAKE_FD(fd
);
325 r
= loop_configure_verify(loop_with_fd
, c
);
329 /* LOOP_CONFIGURE doesn't work. Remember that. */
330 loop_configure_broken
= true;
332 /* We return EBUSY here instead of retrying immediately with LOOP_SET_FD,
333 * because LOOP_CLR_FD is async: if the operation cannot be executed right
334 * away it just sets the autoclear flag on the device. This means there's a
335 * good chance we cannot actually reuse the loopback device right-away. Hence
336 * let's assume it's busy, avoid the trouble and let the calling loop call us
337 * again with a new, likely unused device. */
343 if (loop_configure_broken
) {
344 /* Let's read the seqnum again, to shorten the window. */
345 r
= get_current_uevent_seqnum(&seqnum
);
349 timestamp
= now(CLOCK_MONOTONIC
);
351 if (ioctl(fd
, LOOP_SET_FD
, c
->fd
) < 0)
354 loop_with_fd
= TAKE_FD(fd
);
356 r
= loop_configure_fallback(loop_with_fd
, c
);
361 r
= fd_get_diskseq(loop_with_fd
, &diskseq
);
362 if (r
< 0 && r
!= -EOPNOTSUPP
)
365 switch (lock_op
& ~LOCK_NB
) {
366 case LOCK_EX
: /* Already in effect */
368 case LOCK_SH
: /* Downgrade */
369 if (flock(lock_fd
, lock_op
) < 0)
372 case LOCK_UN
: /* Release */
373 lock_fd
= safe_close(lock_fd
);
376 assert_not_reached();
379 LoopDevice
*d
= new(LoopDevice
, 1);
385 .fd
= TAKE_FD(loop_with_fd
),
386 .lock_fd
= TAKE_FD(lock_fd
),
387 .node
= TAKE_PTR(node
),
390 .dev
= TAKE_PTR(dev
),
392 .uevent_seqnum_not_before
= seqnum
,
393 .timestamp_not_before
= timestamp
,
400 static int loop_device_make_internal(
411 _cleanup_(loop_device_unrefp
) LoopDevice
*d
= NULL
;
412 _cleanup_close_
int direct_io_fd
= -EBADF
, control
= -EBADF
;
413 _cleanup_free_
char *backing_file
= NULL
;
414 struct loop_config config
;
420 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
422 if (fstat(fd
, &st
) < 0)
425 if (S_ISBLK(st
.st_mode
)) {
426 if (offset
== 0 && IN_SET(size
, 0, UINT64_MAX
))
427 /* If this is already a block device and we are supposed to cover the whole of it
428 * then store an fd to the original open device node — and do not actually create an
429 * unnecessary loopback device for it. */
430 return loop_device_open_from_fd(fd
, open_flags
, lock_op
, ret
);
432 r
= stat_verify_regular(&st
);
438 r
= path_make_absolute_cwd(path
, &backing_file
);
442 path_simplify(backing_file
);
444 r
= fd_get_path(fd
, &backing_file
);
449 f_flags
= fcntl(fd
, F_GETFL
);
453 if (FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) != FLAGS_SET(f_flags
, O_DIRECT
)) {
454 /* If LO_FLAGS_DIRECT_IO is requested, then make sure we have the fd open with O_DIRECT, as
455 * that's required. Conversely, if it's off require that O_DIRECT is off too (that's because
456 * new kernels will implicitly enable LO_FLAGS_DIRECT_IO if O_DIRECT is set).
458 * Our intention here is that LO_FLAGS_DIRECT_IO is the primary knob, and O_DIRECT derived
459 * from that automatically. */
461 direct_io_fd
= fd_reopen(fd
, (FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) ? O_DIRECT
: 0)|O_CLOEXEC
|O_NONBLOCK
|open_flags
);
462 if (direct_io_fd
< 0) {
463 if (!FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
))
464 return log_debug_errno(errno
, "Failed to reopen file descriptor without O_DIRECT: %m");
466 /* Some file systems might not support O_DIRECT, let's gracefully continue without it then. */
467 log_debug_errno(errno
, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without.");
468 loop_flags
&= ~LO_FLAGS_DIRECT_IO
;
470 fd
= direct_io_fd
; /* From now on, operate on our new O_DIRECT fd */
473 control
= open("/dev/loop-control", O_RDWR
|O_CLOEXEC
|O_NOCTTY
|O_NONBLOCK
);
477 config
= (struct loop_config
) {
479 .block_size
= block_size
,
481 /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */
482 .lo_flags
= (loop_flags
& ~LO_FLAGS_READ_ONLY
) | ((open_flags
& O_ACCMODE
) == O_RDONLY
? LO_FLAGS_READ_ONLY
: 0) | LO_FLAGS_AUTOCLEAR
,
484 .lo_sizelimit
= size
== UINT64_MAX
? 0 : size
,
488 /* Loop around LOOP_CTL_GET_FREE, since at the moment we attempt to open the returned device it might
489 * be gone already, taken by somebody else racing against us. */
490 for (unsigned n_attempts
= 0;;) {
493 /* Let's take a lock on the control device first. On a busy system, where many programs
494 * attempt to allocate a loopback device at the same time, we might otherwise keep looping
495 * around relatively heavy operations: asking for a free loopback device, then opening it,
496 * validating it, attaching something to it. Let's serialize this whole operation, to make
497 * unnecessary busywork less likely. Note that this is just something we do to optimize our
498 * own code (and whoever else decides to use LOCK_EX locks for this), taking this lock is not
499 * necessary, it just means it's less likely we have to iterate through this loop again and
500 * again if our own code races against our own code.
502 * Note: our lock protocol is to take the /dev/loop-control lock first, and the block device
503 * lock second, if both are taken, and always in this order, to avoid ABBA locking issues. */
504 if (flock(control
, LOCK_EX
) < 0)
507 nr
= ioctl(control
, LOOP_CTL_GET_FREE
);
511 r
= loop_configure(nr
, open_flags
, lock_op
, &config
, &d
);
515 /* -ENODEV or friends: Somebody might've gotten the same number from the kernel, used the
516 * device, and called LOOP_CTL_REMOVE on it. Let's retry with a new number.
517 * -EBUSY: a file descriptor is already bound to the loopback block device.
518 * -EUCLEAN: some left-over partition devices that were cleaned up. */
519 if (!ERRNO_IS_DEVICE_ABSENT(errno
) && !IN_SET(r
, -EBUSY
, -EUCLEAN
))
522 /* OK, this didn't work, let's try again a bit later, but first release the lock on the
524 if (flock(control
, LOCK_UN
) < 0)
527 if (++n_attempts
>= 64) /* Give up eventually */
530 /* Wait some random time, to make collision less likely. Let's pick a random time in the
531 * range 0ms…250ms, linearly scaled by the number of failed attempts. */
532 (void) usleep(random_u64_range(UINT64_C(10) * USEC_PER_MSEC
+
533 UINT64_C(240) * USEC_PER_MSEC
* n_attempts
/64));
536 d
->backing_file
= TAKE_PTR(backing_file
);
538 log_debug("Successfully acquired %s, devno=%u:%u, nr=%i, diskseq=%" PRIu64
,
540 major(d
->devno
), minor(d
->devno
),
548 static uint32_t loop_flags_mangle(uint32_t loop_flags
) {
551 r
= getenv_bool("SYSTEMD_LOOP_DIRECT_IO");
552 if (r
< 0 && r
!= -ENXIO
)
553 log_debug_errno(r
, "Failed to parse $SYSTEMD_LOOP_DIRECT_IO, ignoring: %m");
555 return UPDATE_FLAG(loop_flags
, LO_FLAGS_DIRECT_IO
, r
!= 0); /* Turn on LO_FLAGS_DIRECT_IO by default, unless explicitly configured to off. */
558 int loop_device_make(
571 return loop_device_make_internal(
578 loop_flags_mangle(loop_flags
),
583 int loop_device_make_by_path(
590 int r
, basic_flags
, direct_flags
, rdwr_flags
;
591 _cleanup_close_
int fd
= -EBADF
;
596 assert(open_flags
< 0 || IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
598 /* Passing < 0 as open_flags here means we'll try to open the device writable if we can, retrying
599 * read-only if we cannot. */
601 loop_flags
= loop_flags_mangle(loop_flags
);
603 /* Let's open with O_DIRECT if we can. But not all file systems support that, hence fall back to
604 * non-O_DIRECT mode automatically, if it fails. */
606 basic_flags
= O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
;
607 direct_flags
= FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) ? O_DIRECT
: 0;
608 rdwr_flags
= open_flags
>= 0 ? open_flags
: O_RDWR
;
610 fd
= open(path
, basic_flags
|direct_flags
|rdwr_flags
);
611 if (fd
< 0 && direct_flags
!= 0) /* If we had O_DIRECT on, and things failed with that, let's immediately try again without */
612 fd
= open(path
, basic_flags
|rdwr_flags
);
614 direct
= direct_flags
!= 0;
618 /* Retry read-only? */
619 if (open_flags
>= 0 || !(ERRNO_IS_PRIVILEGE(r
) || r
== -EROFS
))
622 fd
= open(path
, basic_flags
|direct_flags
|O_RDONLY
);
623 if (fd
< 0 && direct_flags
!= 0) /* as above */
624 fd
= open(path
, basic_flags
|O_RDONLY
);
626 direct
= direct_flags
!= 0;
628 return r
; /* Propagate original error */
630 open_flags
= O_RDONLY
;
631 } else if (open_flags
< 0)
634 log_debug("Opened '%s' in %s access mode%s, with O_DIRECT %s%s.",
636 open_flags
== O_RDWR
? "O_RDWR" : "O_RDONLY",
637 open_flags
!= rdwr_flags
? " (O_RDWR was requested but not allowed)" : "",
638 direct
? "enabled" : "disabled",
639 direct
!= (direct_flags
!= 0) ? " (O_DIRECT was requested but not supported)" : "");
641 return loop_device_make_internal(path
, fd
, open_flags
, 0, 0, 0, loop_flags
, lock_op
, ret
);
644 int loop_device_make_by_path_memory(
651 _cleanup_close_
int fd
= -EBADF
, mfd
= -EBADF
;
652 _cleanup_free_
char *fn
= NULL
;
657 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
660 loop_flags
&= ~LO_FLAGS_DIRECT_IO
; /* memfds don't support O_DIRECT, hence LO_FLAGS_DIRECT_IO can't be used either */
662 fd
= open(path
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|O_RDONLY
);
666 if (fstat(fd
, &st
) < 0)
669 if (!S_ISREG(st
.st_mode
) && !S_ISBLK(st
.st_mode
))
672 r
= path_extract_filename(path
, &fn
);
676 mfd
= memfd_clone_fd(fd
, fn
, open_flags
|O_CLOEXEC
);
680 fd
= safe_close(fd
); /* Let's close the original early */
682 return loop_device_make_internal(NULL
, mfd
, open_flags
, 0, 0, 0, loop_flags
, lock_op
, ret
);
685 static LoopDevice
* loop_device_free(LoopDevice
*d
) {
686 _cleanup_close_
int control
= -1;
692 /* Release any lock we might have on the device first. We want to open+lock the /dev/loop-control
693 * device below, but our lock protocol says that if both control and block device locks are taken,
694 * the control lock needs to be taken first, the block device lock second — in order to avoid ABBA
695 * locking issues. Moreover, we want to issue LOOP_CLR_FD on the block device further down, and that
696 * would fail if we had another fd open to the device. */
697 d
->lock_fd
= safe_close(d
->lock_fd
);
699 /* Let's open the control device early, and lock it, so that we can release our block device and
700 * delete it in a synchronized fashion, and allocators won't needlessly see the block device as free
701 * while we are about to delete it. */
702 if (!LOOP_DEVICE_IS_FOREIGN(d
) && !d
->relinquished
) {
703 control
= open("/dev/loop-control", O_RDWR
|O_CLOEXEC
|O_NOCTTY
|O_NONBLOCK
);
705 log_debug_errno(errno
, "Failed to open loop control device, cannot remove loop device '%s', ignoring: %m", strna(d
->node
));
706 else if (flock(control
, LOCK_EX
) < 0)
707 log_debug_errno(errno
, "Failed to lock loop control device, ignoring: %m");
710 /* Then let's release the loopback block device */
712 /* Implicitly sync the device, since otherwise in-flight blocks might not get written */
713 if (fsync(d
->fd
) < 0)
714 log_debug_errno(errno
, "Failed to sync loop block device, ignoring: %m");
716 if (!LOOP_DEVICE_IS_FOREIGN(d
) && !d
->relinquished
) {
717 /* We are supposed to clear the loopback device. Let's do this synchronously: lock
718 * the device, manually remove all partitions and then clear it. This should ensure
719 * udev doesn't concurrently access the devices, and we can be reasonably sure that
720 * once we are done here the device is cleared and all its partition children
721 * removed. Note that we lock our primary device fd here (and not a separate locking
722 * fd, as we do during allocation, since we want to keep the lock all the way through
723 * the LOOP_CLR_FD, but that call would fail if we had more than one fd open.) */
725 if (flock(d
->fd
, LOCK_EX
) < 0)
726 log_debug_errno(errno
, "Failed to lock loop block device, ignoring: %m");
728 r
= block_device_remove_all_partitions(d
->dev
, d
->fd
);
730 log_debug_errno(r
, "Failed to remove partitions of loopback block device, ignoring: %m");
732 if (ioctl(d
->fd
, LOOP_CLR_FD
) < 0)
733 log_debug_errno(errno
, "Failed to clear loop device, ignoring: %m");
739 /* Now that the block device is released, let's also try to remove it */
741 for (unsigned n_attempts
= 0;;) {
742 if (ioctl(control
, LOOP_CTL_REMOVE
, d
->nr
) >= 0)
744 if (errno
!= EBUSY
|| ++n_attempts
>= 64) {
745 log_debug_errno(errno
, "Failed to remove device %s: %m", strna(d
->node
));
748 (void) usleep(50 * USEC_PER_MSEC
);
752 sd_device_unref(d
->dev
);
753 free(d
->backing_file
);
757 DEFINE_TRIVIAL_REF_UNREF_FUNC(LoopDevice
, loop_device
, loop_device_free
);
759 void loop_device_relinquish(LoopDevice
*d
) {
762 /* Don't attempt to clean up the loop device anymore from this point on. Leave the clean-ing up to the kernel
763 * itself, using the loop device "auto-clear" logic we already turned on when creating the device. */
765 d
->relinquished
= true;
768 void loop_device_unrelinquish(LoopDevice
*d
) {
770 d
->relinquished
= false;
773 int loop_device_open(
779 _cleanup_close_
int fd
= -EBADF
, lock_fd
= -EBADF
;
780 _cleanup_free_
char *node
= NULL
, *backing_file
= NULL
;
781 struct loop_info64 info
;
782 uint64_t diskseq
= 0;
789 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
792 /* Even if fd is provided through the argument in loop_device_open_from_fd(), we reopen the inode
793 * here, instead of keeping just a dup() clone of it around, since we want to ensure that the
794 * O_DIRECT flag of the handle we keep is off, we have our own file index, and have the right
795 * read/write mode in effect. */
796 fd
= sd_device_open(dev
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|open_flags
);
800 if ((lock_op
& ~LOCK_NB
) != LOCK_UN
) {
801 lock_fd
= open_lock_fd(fd
, lock_op
);
806 if (ioctl(fd
, LOOP_GET_STATUS64
, &info
) >= 0) {
807 #if HAVE_VALGRIND_MEMCHECK_H
808 /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
809 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
813 if (sd_device_get_sysattr_value(dev
, "loop/backing_file", &s
) >= 0) {
814 backing_file
= strdup(s
);
820 r
= fd_get_diskseq(fd
, &diskseq
);
821 if (r
< 0 && r
!= -EOPNOTSUPP
)
824 r
= sd_device_get_devnum(dev
, &devnum
);
828 r
= sd_device_get_devname(dev
, &s
);
836 d
= new(LoopDevice
, 1);
843 .lock_fd
= TAKE_FD(lock_fd
),
845 .node
= TAKE_PTR(node
),
846 .dev
= sd_device_ref(dev
),
847 .backing_file
= TAKE_PTR(backing_file
),
848 .relinquished
= true, /* It's not ours, don't try to destroy it when this object is freed */
851 .uevent_seqnum_not_before
= UINT64_MAX
,
852 .timestamp_not_before
= USEC_INFINITY
,
859 int loop_device_open_from_fd(
865 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
870 r
= block_device_new_from_fd(fd
, 0, &dev
);
874 return loop_device_open(dev
, open_flags
, lock_op
, ret
);
877 int loop_device_open_from_path(
883 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
888 r
= block_device_new_from_path(path
, 0, &dev
);
892 return loop_device_open(dev
, open_flags
, lock_op
, ret
);
895 static int resize_partition(int partition_fd
, uint64_t offset
, uint64_t size
) {
896 char sysfs
[STRLEN("/sys/dev/block/:/partition") + 2*DECIMAL_STR_MAX(dev_t
) + 1];
897 _cleanup_free_
char *buffer
= NULL
;
898 uint64_t current_offset
, current_size
, partno
;
899 _cleanup_close_
int whole_fd
= -EBADF
;
904 assert(partition_fd
>= 0);
906 /* Resizes the partition the loopback device refer to (assuming it refers to one instead of an actual
907 * loopback device), and changes the offset, if needed. This is a fancy wrapper around
908 * BLKPG_RESIZE_PARTITION. */
910 if (fstat(partition_fd
, &st
) < 0)
913 assert(S_ISBLK(st
.st_mode
));
915 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/partition", DEVNUM_FORMAT_VAL(st
.st_rdev
));
916 r
= read_one_line_file(sysfs
, &buffer
);
917 if (r
== -ENOENT
) /* not a partition, cannot resize */
921 r
= safe_atou64(buffer
, &partno
);
925 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/start", DEVNUM_FORMAT_VAL(st
.st_rdev
));
927 buffer
= mfree(buffer
);
928 r
= read_one_line_file(sysfs
, &buffer
);
931 r
= safe_atou64(buffer
, ¤t_offset
);
934 if (current_offset
> UINT64_MAX
/512U)
936 current_offset
*= 512U;
938 if (ioctl(partition_fd
, BLKGETSIZE64
, ¤t_size
) < 0)
941 if (size
== UINT64_MAX
&& offset
== UINT64_MAX
)
943 if (current_size
== size
&& current_offset
== offset
)
946 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/../dev", DEVNUM_FORMAT_VAL(st
.st_rdev
));
948 buffer
= mfree(buffer
);
949 r
= read_one_line_file(sysfs
, &buffer
);
952 r
= parse_devnum(buffer
, &devno
);
956 whole_fd
= r
= device_open_from_devnum(S_IFBLK
, devno
, O_RDWR
|O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
, NULL
);
960 return block_device_resize_partition(
963 offset
== UINT64_MAX
? current_offset
: offset
,
964 size
== UINT64_MAX
? current_size
: size
);
967 int loop_device_refresh_size(LoopDevice
*d
, uint64_t offset
, uint64_t size
) {
968 struct loop_info64 info
;
973 /* Changes the offset/start of the loop device relative to the beginning of the underlying file or
974 * block device. If this loop device actually refers to a partition and not a loopback device, we'll
975 * try to adjust the partition offsets instead.
977 * If either offset or size is UINT64_MAX we won't change that parameter. */
979 if (d
->nr
< 0) /* not a loopback device */
980 return resize_partition(d
->fd
, offset
, size
);
982 if (ioctl(d
->fd
, LOOP_GET_STATUS64
, &info
) < 0)
985 #if HAVE_VALGRIND_MEMCHECK_H
986 /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
987 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
990 if (size
== UINT64_MAX
&& offset
== UINT64_MAX
)
992 if (info
.lo_sizelimit
== size
&& info
.lo_offset
== offset
)
995 if (size
!= UINT64_MAX
)
996 info
.lo_sizelimit
= size
;
997 if (offset
!= UINT64_MAX
)
998 info
.lo_offset
= offset
;
1000 return RET_NERRNO(ioctl(d
->fd
, LOOP_SET_STATUS64
, &info
));
1003 int loop_device_flock(LoopDevice
*d
, int operation
) {
1004 assert(IN_SET(operation
& ~LOCK_NB
, LOCK_UN
, LOCK_SH
, LOCK_EX
));
1007 /* When unlocking just close the lock fd */
1008 if ((operation
& ~LOCK_NB
) == LOCK_UN
) {
1009 d
->lock_fd
= safe_close(d
->lock_fd
);
1013 /* If we had no lock fd so far, create one and lock it right-away */
1014 if (d
->lock_fd
< 0) {
1017 d
->lock_fd
= open_lock_fd(d
->fd
, operation
);
1024 /* Otherwise change the current lock mode on the existing fd */
1025 return RET_NERRNO(flock(d
->lock_fd
, operation
));
1028 int loop_device_sync(LoopDevice
*d
) {
1032 /* We also do this implicitly in loop_device_unref(). Doing this explicitly here has the benefit that
1033 * we can check the return value though. */
1035 return RET_NERRNO(fsync(d
->fd
));