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"
28 #include "loop-util.h"
29 #include "missing_loop.h"
30 #include "parse-util.h"
31 #include "path-util.h"
32 #include "random-util.h"
33 #include "stat-util.h"
34 #include "stdio-util.h"
35 #include "string-util.h"
36 #include "tmpfile-util.h"
38 static void cleanup_clear_loop_close(int *fd
) {
42 (void) ioctl(*fd
, LOOP_CLR_FD
);
43 (void) safe_close(*fd
);
46 static int loop_is_bound(int fd
) {
47 struct loop_info64 info
;
51 if (ioctl(fd
, LOOP_GET_STATUS64
, &info
) < 0) {
53 return false; /* not bound! */
58 return true; /* bound! */
61 static int get_current_uevent_seqnum(uint64_t *ret
) {
62 _cleanup_free_
char *p
= NULL
;
65 r
= read_full_virtual_file("/sys/kernel/uevent_seqnum", &p
, NULL
);
67 return log_debug_errno(r
, "Failed to read current uevent sequence number: %m");
69 r
= safe_atou64(strstrip(p
), ret
);
71 return log_debug_errno(r
, "Failed to parse current uevent sequence number: %s", p
);
76 static int open_lock_fd(int primary_fd
, int operation
) {
77 _cleanup_close_
int lock_fd
= -EBADF
;
79 assert(primary_fd
>= 0);
80 assert(IN_SET(operation
& ~LOCK_NB
, LOCK_SH
, LOCK_EX
));
82 lock_fd
= fd_reopen(primary_fd
, O_RDONLY
|O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
);
86 if (flock(lock_fd
, operation
) < 0)
89 return TAKE_FD(lock_fd
);
92 static int loop_configure_verify_direct_io(int fd
, const struct loop_config
*c
) {
96 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_DIRECT_IO
)) {
97 struct loop_info64 info
;
99 if (ioctl(fd
, LOOP_GET_STATUS64
, &info
) < 0)
100 return log_debug_errno(errno
, "Failed to issue LOOP_GET_STATUS64: %m");
102 #if HAVE_VALGRIND_MEMCHECK_H
103 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
106 /* On older kernels (<= 5.3) it was necessary to set the block size of the loopback block
107 * device to the logical block size of the underlying file system. Since there was no nice
108 * way to query the value, we are not bothering to do this however. On newer kernels the
109 * block size is propagated automatically and does not require intervention from us. We'll
110 * check here if enabling direct IO worked, to make this easily debuggable however.
112 * (Should anyone really care and actually wants direct IO on old kernels: it might be worth
113 * enabling direct IO with iteratively larger block sizes until it eventually works.) */
114 if (!FLAGS_SET(info
.lo_flags
, LO_FLAGS_DIRECT_IO
))
115 log_debug("Could not enable direct IO mode, proceeding in buffered IO mode.");
121 static int loop_configure_verify(int fd
, const struct loop_config
*c
) {
128 if (c
->block_size
!= 0) {
131 r
= blockdev_get_sector_size(fd
, &ssz
);
135 if (ssz
!= c
->block_size
) {
136 log_debug("LOOP_CONFIGURE didn't honour requested block size %" PRIu32
", got %" PRIu32
" instead. Ignoring.", c
->block_size
, ssz
);
141 if (c
->info
.lo_sizelimit
!= 0) {
142 /* Kernel 5.8 vanilla doesn't properly propagate the size limit into the
143 * block device. If it's used, let's immediately check if it had the desired
144 * effect hence. And if not use classic LOOP_SET_STATUS64. */
147 if (ioctl(fd
, BLKGETSIZE64
, &z
) < 0)
150 if (z
!= c
->info
.lo_sizelimit
) {
151 log_debug("LOOP_CONFIGURE is broken, doesn't honour .info.lo_sizelimit. Falling back to LOOP_SET_STATUS64.");
156 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_PARTSCAN
)) {
157 /* Kernel 5.8 vanilla doesn't properly propagate the partition scanning flag
158 * into the block device. Let's hence verify if things work correctly here
159 * before returning. */
161 r
= blockdev_partscan_enabled(fd
);
165 log_debug("LOOP_CONFIGURE is broken, doesn't honour LO_FLAGS_PARTSCAN. Falling back to LOOP_SET_STATUS64.");
170 r
= loop_configure_verify_direct_io(fd
, c
);
177 static int loop_configure_fallback(int fd
, const struct loop_config
*c
) {
178 struct loop_info64 info_copy
;
184 /* Only some of the flags LOOP_CONFIGURE can set are also settable via LOOP_SET_STATUS64, hence mask
187 info_copy
.lo_flags
&= LOOP_SET_STATUS_SETTABLE_FLAGS
;
189 /* Since kernel commit 5db470e229e22b7eda6e23b5566e532c96fb5bc3 (kernel v5.0) the LOOP_SET_STATUS64
190 * ioctl can return EAGAIN in case we change the info.lo_offset field, if someone else is accessing the
191 * block device while we try to reconfigure it. This is a pretty common case, since udev might
192 * instantly start probing the device as soon as we attach an fd to it. Hence handle it in two ways:
193 * first, let's take the BSD lock to ensure that udev will not step in between the point in
194 * time where we attach the fd and where we reconfigure the device. Secondly, let's wait 50ms on
195 * EAGAIN and retry. The former should be an efficient mechanism to avoid we have to wait 50ms
196 * needlessly if we are just racing against udev. The latter is protection against all other cases,
197 * i.e. peers that do not take the BSD lock. */
199 for (unsigned n_attempts
= 0;;) {
200 if (ioctl(fd
, LOOP_SET_STATUS64
, &info_copy
) >= 0)
203 if (errno
!= EAGAIN
|| ++n_attempts
>= 64)
204 return log_debug_errno(errno
, "Failed to configure loopback block device: %m");
206 /* Sleep some random time, but at least 10ms, at most 250ms. Increase the delay the more
207 * failed attempts we see */
208 (void) usleep(UINT64_C(10) * USEC_PER_MSEC
+
209 random_u64_range(UINT64_C(240) * USEC_PER_MSEC
* n_attempts
/64));
212 /* Work around a kernel bug, where changing offset/size of the loopback device doesn't correctly
213 * invalidate the buffer cache. For details see:
215 * https://android.googlesource.com/platform/system/apex/+/bef74542fbbb4cd629793f4efee8e0053b360570
217 * This was fixed in kernel 5.0, see:
219 * https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=5db470e229e22b7eda6e23b5566e532c96fb5bc3
221 * We'll run the work-around here in the legacy LOOP_SET_STATUS64 codepath. In the LOOP_CONFIGURE
222 * codepath above it should not be necessary. */
223 if (c
->info
.lo_offset
!= 0 || c
->info
.lo_sizelimit
!= 0)
224 if (ioctl(fd
, BLKFLSBUF
, 0) < 0)
225 log_debug_errno(errno
, "Failed to issue BLKFLSBUF ioctl, ignoring: %m");
227 /* If a block size is requested then try to configure it. If that doesn't work, ignore errors, but
228 * afterwards, let's validate what is in effect, and if it doesn't match what we want, fail */
229 if (c
->block_size
!= 0) {
232 if (ioctl(fd
, LOOP_SET_BLOCK_SIZE
, (unsigned long) c
->block_size
) < 0)
233 log_debug_errno(errno
, "Failed to set sector size, ignoring: %m");
235 r
= blockdev_get_sector_size(fd
, &ssz
);
237 return log_debug_errno(r
, "Failed to read sector size: %m");
238 if (ssz
!= c
->block_size
)
239 return log_debug_errno(SYNTHETIC_ERRNO(EIO
), "Sector size of loopback device doesn't match what we requested, refusing.");
242 /* LO_FLAGS_DIRECT_IO is a flags we need to configure via explicit ioctls. */
243 if (FLAGS_SET(c
->info
.lo_flags
, LO_FLAGS_DIRECT_IO
))
244 if (ioctl(fd
, LOOP_SET_DIRECT_IO
, 1UL) < 0)
245 log_debug_errno(errno
, "Failed to enable direct IO mode, ignoring: %m");
247 return loop_configure_verify_direct_io(fd
, c
);
250 static int loop_configure(
254 const struct loop_config
*c
,
257 static bool loop_configure_broken
= false;
259 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
260 _cleanup_(cleanup_clear_loop_close
) int loop_with_fd
= -EBADF
; /* This must be declared before lock_fd. */
261 _cleanup_close_
int fd
= -EBADF
, lock_fd
= -EBADF
;
262 _cleanup_free_
char *node
= NULL
;
263 uint64_t diskseq
= 0, seqnum
= UINT64_MAX
;
264 usec_t timestamp
= USEC_INFINITY
;
272 if (asprintf(&node
, "/dev/loop%i", nr
) < 0)
275 r
= sd_device_new_from_devname(&dev
, node
);
279 r
= sd_device_get_devnum(dev
, &devno
);
283 fd
= sd_device_open(dev
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|open_flags
);
287 /* Let's lock the device before we do anything. We take the BSD lock on a second, separately opened
288 * fd for the device. udev after all watches for close() events (specifically IN_CLOSE_WRITE) on
289 * block devices to reprobe them, hence by having a separate fd we will later close() we can ensure
290 * we trigger udev after everything is done. If we'd lock our own fd instead and keep it open for a
291 * long time udev would possibly never run on it again, even though the fd is unlocked, simply
292 * because we never close() it. It also has the nice benefit we can use the _cleanup_close_ logic to
293 * automatically release the lock, after we are done. */
294 lock_fd
= open_lock_fd(fd
, LOCK_EX
);
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
);
306 /* Let's see if the device is really detached, i.e. currently has no associated partition block
307 * devices. On various kernels (such as 5.8) it is possible to have a loopback block device that
308 * superficially is detached but still has partition block devices associated for it. Let's then
309 * manually remove the partitions via BLKPG, and tell the caller we did that via EUCLEAN, so they try
311 r
= block_device_remove_all_partitions(dev
, fd
);
315 /* Removed all partitions. Let's report this to the caller, to try again, and count this as
319 if (!loop_configure_broken
) {
320 /* Acquire uevent seqnum immediately before attaching the loopback device. This allows
321 * callers to ignore all uevents with a seqnum before this one, if they need to associate
322 * uevent with this attachment. Doing so isn't race-free though, as uevents that happen in
323 * the window between this reading of the seqnum, and the LOOP_CONFIGURE call might still be
324 * mistaken as originating from our attachment, even though might be caused by an earlier
325 * use. But doing this at least shortens the race window a bit. */
326 r
= get_current_uevent_seqnum(&seqnum
);
330 timestamp
= now(CLOCK_MONOTONIC
);
332 if (ioctl(fd
, LOOP_CONFIGURE
, c
) < 0) {
333 /* Do fallback only if LOOP_CONFIGURE is not supported, propagate all other
334 * errors. Note that the kernel is weird: non-existing ioctls currently return EINVAL
335 * rather than ENOTTY on loopback block devices. They should fix that in the kernel,
336 * but in the meantime we accept both here. */
337 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && errno
!= EINVAL
)
340 loop_configure_broken
= true;
342 loop_with_fd
= TAKE_FD(fd
);
344 r
= loop_configure_verify(loop_with_fd
, c
);
348 /* LOOP_CONFIGURE doesn't work. Remember that. */
349 loop_configure_broken
= true;
351 /* We return EBUSY here instead of retrying immediately with LOOP_SET_FD,
352 * because LOOP_CLR_FD is async: if the operation cannot be executed right
353 * away it just sets the autoclear flag on the device. This means there's a
354 * good chance we cannot actually reuse the loopback device right-away. Hence
355 * let's assume it's busy, avoid the trouble and let the calling loop call us
356 * again with a new, likely unused device. */
362 if (loop_configure_broken
) {
363 /* Let's read the seqnum again, to shorten the window. */
364 r
= get_current_uevent_seqnum(&seqnum
);
368 timestamp
= now(CLOCK_MONOTONIC
);
370 if (ioctl(fd
, LOOP_SET_FD
, c
->fd
) < 0)
373 loop_with_fd
= TAKE_FD(fd
);
375 r
= loop_configure_fallback(loop_with_fd
, c
);
380 r
= fd_get_diskseq(loop_with_fd
, &diskseq
);
381 if (r
< 0 && r
!= -EOPNOTSUPP
)
384 switch (lock_op
& ~LOCK_NB
) {
385 case LOCK_EX
: /* Already in effect */
387 case LOCK_SH
: /* Downgrade */
388 if (flock(lock_fd
, lock_op
) < 0)
391 case LOCK_UN
: /* Release */
392 lock_fd
= safe_close(lock_fd
);
395 assert_not_reached();
398 LoopDevice
*d
= new(LoopDevice
, 1);
404 .fd
= TAKE_FD(loop_with_fd
),
405 .lock_fd
= TAKE_FD(lock_fd
),
406 .node
= TAKE_PTR(node
),
409 .dev
= TAKE_PTR(dev
),
411 .uevent_seqnum_not_before
= seqnum
,
412 .timestamp_not_before
= timestamp
,
413 .sector_size
= c
->block_size
,
420 static int loop_device_make_internal(
426 uint32_t sector_size
,
431 _cleanup_(loop_device_unrefp
) LoopDevice
*d
= NULL
;
432 _cleanup_close_
int direct_io_fd
= -EBADF
, control
= -EBADF
;
433 _cleanup_free_
char *backing_file
= NULL
;
434 struct loop_config config
;
440 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
442 if (fstat(fd
, &st
) < 0)
445 if (S_ISBLK(st
.st_mode
)) {
446 if (offset
== 0 && IN_SET(size
, 0, UINT64_MAX
))
447 /* If this is already a block device and we are supposed to cover the whole of it
448 * then store an fd to the original open device node — and do not actually create an
449 * unnecessary loopback device for it. */
450 return loop_device_open_from_fd(fd
, open_flags
, lock_op
, ret
);
452 r
= stat_verify_regular(&st
);
458 r
= path_make_absolute_cwd(path
, &backing_file
);
462 path_simplify(backing_file
);
464 r
= fd_get_path(fd
, &backing_file
);
469 f_flags
= fcntl(fd
, F_GETFL
);
473 if (FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) != FLAGS_SET(f_flags
, O_DIRECT
)) {
474 /* If LO_FLAGS_DIRECT_IO is requested, then make sure we have the fd open with O_DIRECT, as
475 * that's required. Conversely, if it's off require that O_DIRECT is off too (that's because
476 * new kernels will implicitly enable LO_FLAGS_DIRECT_IO if O_DIRECT is set).
478 * Our intention here is that LO_FLAGS_DIRECT_IO is the primary knob, and O_DIRECT derived
479 * from that automatically. */
481 direct_io_fd
= fd_reopen(fd
, (FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) ? O_DIRECT
: 0)|O_CLOEXEC
|O_NONBLOCK
|open_flags
);
482 if (direct_io_fd
< 0) {
483 if (!FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
))
484 return log_debug_errno(errno
, "Failed to reopen file descriptor without O_DIRECT: %m");
486 /* Some file systems might not support O_DIRECT, let's gracefully continue without it then. */
487 log_debug_errno(errno
, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without.");
488 loop_flags
&= ~LO_FLAGS_DIRECT_IO
;
490 fd
= direct_io_fd
; /* From now on, operate on our new O_DIRECT fd */
493 control
= open("/dev/loop-control", O_RDWR
|O_CLOEXEC
|O_NOCTTY
|O_NONBLOCK
);
497 if (sector_size
== 0)
498 /* If no sector size is specified, default to the classic default */
500 else if (sector_size
== UINT32_MAX
) {
502 if (S_ISBLK(st
.st_mode
))
503 /* If the sector size is specified as UINT32_MAX we'll propagate the sector size of
504 * the underlying block device. */
505 r
= blockdev_get_sector_size(fd
, §or_size
);
507 _cleanup_close_
int non_direct_io_fd
= -EBADF
;
510 assert(S_ISREG(st
.st_mode
));
512 /* If sector size is specified as UINT32_MAX, we'll try to probe the right sector
513 * size of the image in question by looking for the GPT partition header at various
514 * offsets. This of course only works if the image already has a disk label.
516 * So here we actually want to read the file contents ourselves. This is quite likely
517 * not going to work if we managed to enable O_DIRECT, because in such a case there
518 * are some pretty strict alignment requirements to offset, size and target, but
519 * there's no way to query what alignment specifically is actually required. Hence,
520 * let's avoid the mess, and temporarily open an fd without O_DIRECT for the probing
523 if (FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
)) {
524 non_direct_io_fd
= fd_reopen(fd
, O_RDONLY
|O_CLOEXEC
|O_NONBLOCK
);
525 if (non_direct_io_fd
< 0)
526 return non_direct_io_fd
;
528 probe_fd
= non_direct_io_fd
;
532 r
= probe_sector_size(probe_fd
, §or_size
);
538 config
= (struct loop_config
) {
540 .block_size
= sector_size
,
542 /* Use the specified flags, but configure the read-only flag from the open flags, and force autoclear */
543 .lo_flags
= (loop_flags
& ~LO_FLAGS_READ_ONLY
) | ((open_flags
& O_ACCMODE
) == O_RDONLY
? LO_FLAGS_READ_ONLY
: 0) | LO_FLAGS_AUTOCLEAR
,
545 .lo_sizelimit
= size
== UINT64_MAX
? 0 : size
,
549 /* Loop around LOOP_CTL_GET_FREE, since at the moment we attempt to open the returned device it might
550 * be gone already, taken by somebody else racing against us. */
551 for (unsigned n_attempts
= 0;;) {
554 /* Let's take a lock on the control device first. On a busy system, where many programs
555 * attempt to allocate a loopback device at the same time, we might otherwise keep looping
556 * around relatively heavy operations: asking for a free loopback device, then opening it,
557 * validating it, attaching something to it. Let's serialize this whole operation, to make
558 * unnecessary busywork less likely. Note that this is just something we do to optimize our
559 * own code (and whoever else decides to use LOCK_EX locks for this), taking this lock is not
560 * necessary, it just means it's less likely we have to iterate through this loop again and
561 * again if our own code races against our own code.
563 * Note: our lock protocol is to take the /dev/loop-control lock first, and the block device
564 * lock second, if both are taken, and always in this order, to avoid ABBA locking issues. */
565 if (flock(control
, LOCK_EX
) < 0)
568 nr
= ioctl(control
, LOOP_CTL_GET_FREE
);
572 r
= loop_configure(nr
, open_flags
, lock_op
, &config
, &d
);
576 /* -ENODEV or friends: Somebody might've gotten the same number from the kernel, used the
577 * device, and called LOOP_CTL_REMOVE on it. Let's retry with a new number.
578 * -EBUSY: a file descriptor is already bound to the loopback block device.
579 * -EUCLEAN: some left-over partition devices that were cleaned up. */
580 if (!ERRNO_IS_DEVICE_ABSENT(r
) && !IN_SET(r
, -EBUSY
, -EUCLEAN
))
583 /* OK, this didn't work, let's try again a bit later, but first release the lock on the
585 if (flock(control
, LOCK_UN
) < 0)
588 if (++n_attempts
>= 64) /* Give up eventually */
591 /* Wait some random time, to make collision less likely. Let's pick a random time in the
592 * range 0ms…250ms, linearly scaled by the number of failed attempts. */
593 (void) usleep(random_u64_range(UINT64_C(10) * USEC_PER_MSEC
+
594 UINT64_C(240) * USEC_PER_MSEC
* n_attempts
/64));
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(
649 uint32_t sector_size
,
654 int r
, basic_flags
, direct_flags
, rdwr_flags
;
655 _cleanup_close_
int fd
= -EBADF
;
660 assert(open_flags
< 0 || IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
662 /* Passing < 0 as open_flags here means we'll try to open the device writable if we can, retrying
663 * read-only if we cannot. */
665 loop_flags
= loop_flags_mangle(loop_flags
);
667 /* Let's open with O_DIRECT if we can. But not all file systems support that, hence fall back to
668 * non-O_DIRECT mode automatically, if it fails. */
670 basic_flags
= O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
;
671 direct_flags
= FLAGS_SET(loop_flags
, LO_FLAGS_DIRECT_IO
) ? O_DIRECT
: 0;
672 rdwr_flags
= open_flags
>= 0 ? open_flags
: O_RDWR
;
674 fd
= open(path
, basic_flags
|direct_flags
|rdwr_flags
);
675 if (fd
< 0 && direct_flags
!= 0) /* If we had O_DIRECT on, and things failed with that, let's immediately try again without */
676 fd
= open(path
, basic_flags
|rdwr_flags
);
678 direct
= direct_flags
!= 0;
682 /* Retry read-only? */
683 if (open_flags
>= 0 || !(ERRNO_IS_PRIVILEGE(r
) || r
== -EROFS
))
686 fd
= open(path
, basic_flags
|direct_flags
|O_RDONLY
);
687 if (fd
< 0 && direct_flags
!= 0) /* as above */
688 fd
= open(path
, basic_flags
|O_RDONLY
);
690 direct
= direct_flags
!= 0;
692 return r
; /* Propagate original error */
694 open_flags
= O_RDONLY
;
695 } else if (open_flags
< 0)
698 log_debug("Opened '%s' in %s access mode%s, with O_DIRECT %s%s.",
700 open_flags
== O_RDWR
? "O_RDWR" : "O_RDONLY",
701 open_flags
!= rdwr_flags
? " (O_RDWR was requested but not allowed)" : "",
702 direct
? "enabled" : "disabled",
703 direct
!= (direct_flags
!= 0) ? " (O_DIRECT was requested but not supported)" : "");
705 return loop_device_make_internal(path
, fd
, open_flags
, 0, 0, sector_size
, loop_flags
, lock_op
, ret
);
708 int loop_device_make_by_path_memory(
711 uint32_t sector_size
,
716 _cleanup_close_
int fd
= -EBADF
, mfd
= -EBADF
;
717 _cleanup_free_
char *fn
= NULL
;
722 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
725 loop_flags
&= ~LO_FLAGS_DIRECT_IO
; /* memfds don't support O_DIRECT, hence LO_FLAGS_DIRECT_IO can't be used either */
727 fd
= open(path
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|O_RDONLY
);
731 if (fstat(fd
, &st
) < 0)
734 if (!S_ISREG(st
.st_mode
) && !S_ISBLK(st
.st_mode
))
737 r
= path_extract_filename(path
, &fn
);
741 mfd
= memfd_clone_fd(fd
, fn
, open_flags
|O_CLOEXEC
);
745 fd
= safe_close(fd
); /* Let's close the original early */
747 return loop_device_make_internal(NULL
, mfd
, open_flags
, 0, 0, sector_size
, loop_flags
, lock_op
, ret
);
750 static LoopDevice
* loop_device_free(LoopDevice
*d
) {
751 _cleanup_close_
int control
= -EBADF
;
757 /* Release any lock we might have on the device first. We want to open+lock the /dev/loop-control
758 * device below, but our lock protocol says that if both control and block device locks are taken,
759 * the control lock needs to be taken first, the block device lock second — in order to avoid ABBA
760 * locking issues. Moreover, we want to issue LOOP_CLR_FD on the block device further down, and that
761 * would fail if we had another fd open to the device. */
762 d
->lock_fd
= safe_close(d
->lock_fd
);
764 /* Let's open the control device early, and lock it, so that we can release our block device and
765 * delete it in a synchronized fashion, and allocators won't needlessly see the block device as free
766 * while we are about to delete it. */
767 if (!LOOP_DEVICE_IS_FOREIGN(d
) && !d
->relinquished
) {
768 control
= open("/dev/loop-control", O_RDWR
|O_CLOEXEC
|O_NOCTTY
|O_NONBLOCK
);
770 log_debug_errno(errno
, "Failed to open loop control device, cannot remove loop device '%s', ignoring: %m", strna(d
->node
));
771 else if (flock(control
, LOCK_EX
) < 0)
772 log_debug_errno(errno
, "Failed to lock loop control device, ignoring: %m");
775 /* Then let's release the loopback block device */
777 /* Implicitly sync the device, since otherwise in-flight blocks might not get written */
778 if (fsync(d
->fd
) < 0)
779 log_debug_errno(errno
, "Failed to sync loop block device, ignoring: %m");
781 if (!LOOP_DEVICE_IS_FOREIGN(d
) && !d
->relinquished
) {
782 /* We are supposed to clear the loopback device. Let's do this synchronously: lock
783 * the device, manually remove all partitions and then clear it. This should ensure
784 * udev doesn't concurrently access the devices, and we can be reasonably sure that
785 * once we are done here the device is cleared and all its partition children
786 * removed. Note that we lock our primary device fd here (and not a separate locking
787 * fd, as we do during allocation, since we want to keep the lock all the way through
788 * the LOOP_CLR_FD, but that call would fail if we had more than one fd open.) */
790 if (flock(d
->fd
, LOCK_EX
) < 0)
791 log_debug_errno(errno
, "Failed to lock loop block device, ignoring: %m");
793 r
= block_device_remove_all_partitions(d
->dev
, d
->fd
);
795 log_debug_errno(r
, "Failed to remove partitions of loopback block device, ignoring: %m");
797 if (ioctl(d
->fd
, LOOP_CLR_FD
) < 0)
798 log_debug_errno(errno
, "Failed to clear loop device, ignoring: %m");
804 /* Now that the block device is released, let's also try to remove it */
806 for (unsigned n_attempts
= 0;;) {
807 if (ioctl(control
, LOOP_CTL_REMOVE
, d
->nr
) >= 0)
809 if (errno
!= EBUSY
|| ++n_attempts
>= 64) {
810 log_debug_errno(errno
, "Failed to remove device %s: %m", strna(d
->node
));
813 (void) usleep(50 * USEC_PER_MSEC
);
817 sd_device_unref(d
->dev
);
818 free(d
->backing_file
);
822 DEFINE_TRIVIAL_REF_UNREF_FUNC(LoopDevice
, loop_device
, loop_device_free
);
824 void loop_device_relinquish(LoopDevice
*d
) {
827 /* Don't attempt to clean up the loop device anymore from this point on. Leave the clean-ing up to the kernel
828 * itself, using the loop device "auto-clear" logic we already turned on when creating the device. */
830 d
->relinquished
= true;
833 void loop_device_unrelinquish(LoopDevice
*d
) {
835 d
->relinquished
= false;
838 int loop_device_open(
844 _cleanup_close_
int fd
= -EBADF
, lock_fd
= -EBADF
;
845 _cleanup_free_
char *node
= NULL
, *backing_file
= NULL
;
846 dev_t devnum
, backing_devno
= 0;
847 struct loop_info64 info
;
848 ino_t backing_inode
= 0;
849 uint64_t diskseq
= 0;
855 assert(IN_SET(open_flags
, O_RDWR
, O_RDONLY
));
858 /* Even if fd is provided through the argument in loop_device_open_from_fd(), we reopen the inode
859 * here, instead of keeping just a dup() clone of it around, since we want to ensure that the
860 * O_DIRECT flag of the handle we keep is off, we have our own file index, and have the right
861 * read/write mode in effect. */
862 fd
= sd_device_open(dev
, O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
|open_flags
);
866 if ((lock_op
& ~LOCK_NB
) != LOCK_UN
) {
867 lock_fd
= open_lock_fd(fd
, lock_op
);
872 if (ioctl(fd
, LOOP_GET_STATUS64
, &info
) >= 0) {
873 #if HAVE_VALGRIND_MEMCHECK_H
874 /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
875 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
879 if (sd_device_get_sysattr_value(dev
, "loop/backing_file", &s
) >= 0) {
880 backing_file
= strdup(s
);
885 backing_devno
= info
.lo_device
;
886 backing_inode
= info
.lo_inode
;
889 r
= fd_get_diskseq(fd
, &diskseq
);
890 if (r
< 0 && r
!= -EOPNOTSUPP
)
893 uint32_t sector_size
;
894 r
= blockdev_get_sector_size(fd
, §or_size
);
898 r
= sd_device_get_devnum(dev
, &devnum
);
902 r
= sd_device_get_devname(dev
, &s
);
910 d
= new(LoopDevice
, 1);
917 .lock_fd
= TAKE_FD(lock_fd
),
919 .node
= TAKE_PTR(node
),
920 .dev
= sd_device_ref(dev
),
921 .backing_file
= TAKE_PTR(backing_file
),
922 .backing_inode
= backing_inode
,
923 .backing_devno
= backing_devno
,
924 .relinquished
= true, /* It's not ours, don't try to destroy it when this object is freed */
927 .uevent_seqnum_not_before
= UINT64_MAX
,
928 .timestamp_not_before
= USEC_INFINITY
,
929 .sector_size
= sector_size
,
936 int loop_device_open_from_fd(
942 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
947 r
= block_device_new_from_fd(fd
, 0, &dev
);
951 return loop_device_open(dev
, open_flags
, lock_op
, ret
);
954 int loop_device_open_from_path(
960 _cleanup_(sd_device_unrefp
) sd_device
*dev
= NULL
;
965 r
= block_device_new_from_path(path
, 0, &dev
);
969 return loop_device_open(dev
, open_flags
, lock_op
, ret
);
972 static int resize_partition(int partition_fd
, uint64_t offset
, uint64_t size
) {
973 char sysfs
[STRLEN("/sys/dev/block/:/partition") + 2*DECIMAL_STR_MAX(dev_t
) + 1];
974 _cleanup_free_
char *buffer
= NULL
;
975 uint64_t current_offset
, current_size
, partno
;
976 _cleanup_close_
int whole_fd
= -EBADF
;
981 assert(partition_fd
>= 0);
983 /* Resizes the partition the loopback device refer to (assuming it refers to one instead of an actual
984 * loopback device), and changes the offset, if needed. This is a fancy wrapper around
985 * BLKPG_RESIZE_PARTITION. */
987 if (fstat(partition_fd
, &st
) < 0)
990 assert(S_ISBLK(st
.st_mode
));
992 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/partition", DEVNUM_FORMAT_VAL(st
.st_rdev
));
993 r
= read_one_line_file(sysfs
, &buffer
);
994 if (r
== -ENOENT
) /* not a partition, cannot resize */
998 r
= safe_atou64(buffer
, &partno
);
1002 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/start", DEVNUM_FORMAT_VAL(st
.st_rdev
));
1004 buffer
= mfree(buffer
);
1005 r
= read_one_line_file(sysfs
, &buffer
);
1008 r
= safe_atou64(buffer
, ¤t_offset
);
1011 if (current_offset
> UINT64_MAX
/512U)
1013 current_offset
*= 512U;
1015 if (ioctl(partition_fd
, BLKGETSIZE64
, ¤t_size
) < 0)
1018 if (size
== UINT64_MAX
&& offset
== UINT64_MAX
)
1020 if (current_size
== size
&& current_offset
== offset
)
1023 xsprintf(sysfs
, "/sys/dev/block/" DEVNUM_FORMAT_STR
"/../dev", DEVNUM_FORMAT_VAL(st
.st_rdev
));
1025 buffer
= mfree(buffer
);
1026 r
= read_one_line_file(sysfs
, &buffer
);
1029 r
= parse_devnum(buffer
, &devno
);
1033 whole_fd
= r
= device_open_from_devnum(S_IFBLK
, devno
, O_RDWR
|O_CLOEXEC
|O_NONBLOCK
|O_NOCTTY
, NULL
);
1037 return block_device_resize_partition(
1040 offset
== UINT64_MAX
? current_offset
: offset
,
1041 size
== UINT64_MAX
? current_size
: size
);
1044 int loop_device_refresh_size(LoopDevice
*d
, uint64_t offset
, uint64_t size
) {
1045 struct loop_info64 info
;
1050 /* Changes the offset/start of the loop device relative to the beginning of the underlying file or
1051 * block device. If this loop device actually refers to a partition and not a loopback device, we'll
1052 * try to adjust the partition offsets instead.
1054 * If either offset or size is UINT64_MAX we won't change that parameter. */
1056 if (d
->nr
< 0) /* not a loopback device */
1057 return resize_partition(d
->fd
, offset
, size
);
1059 if (ioctl(d
->fd
, LOOP_GET_STATUS64
, &info
) < 0)
1062 #if HAVE_VALGRIND_MEMCHECK_H
1063 /* Valgrind currently doesn't know LOOP_GET_STATUS64. Remove this once it does */
1064 VALGRIND_MAKE_MEM_DEFINED(&info
, sizeof(info
));
1067 if (size
== UINT64_MAX
&& offset
== UINT64_MAX
)
1069 if (info
.lo_sizelimit
== size
&& info
.lo_offset
== offset
)
1072 if (size
!= UINT64_MAX
)
1073 info
.lo_sizelimit
= size
;
1074 if (offset
!= UINT64_MAX
)
1075 info
.lo_offset
= offset
;
1077 return RET_NERRNO(ioctl(d
->fd
, LOOP_SET_STATUS64
, &info
));
1080 int loop_device_flock(LoopDevice
*d
, int operation
) {
1081 assert(IN_SET(operation
& ~LOCK_NB
, LOCK_UN
, LOCK_SH
, LOCK_EX
));
1084 /* When unlocking just close the lock fd */
1085 if ((operation
& ~LOCK_NB
) == LOCK_UN
) {
1086 d
->lock_fd
= safe_close(d
->lock_fd
);
1090 /* If we had no lock fd so far, create one and lock it right-away */
1091 if (d
->lock_fd
< 0) {
1094 d
->lock_fd
= open_lock_fd(d
->fd
, operation
);
1101 /* Otherwise change the current lock mode on the existing fd */
1102 return RET_NERRNO(flock(d
->lock_fd
, operation
));
1105 int loop_device_sync(LoopDevice
*d
) {
1109 /* We also do this implicitly in loop_device_unref(). Doing this explicitly here has the benefit that
1110 * we can check the return value though. */
1112 return RET_NERRNO(fsync(d
->fd
));
1115 int loop_device_set_autoclear(LoopDevice
*d
, bool autoclear
) {
1116 struct loop_info64 info
;
1120 if (ioctl(d
->fd
, LOOP_GET_STATUS64
, &info
) < 0)
1123 if (autoclear
== FLAGS_SET(info
.lo_flags
, LO_FLAGS_AUTOCLEAR
))
1126 SET_FLAG(info
.lo_flags
, LO_FLAGS_AUTOCLEAR
, autoclear
);
1128 if (ioctl(d
->fd
, LOOP_SET_STATUS64
, &info
) < 0)
1134 int loop_device_set_filename(LoopDevice
*d
, const char *name
) {
1135 struct loop_info64 info
;
1139 /* Sets the .lo_file_name of the loopback device. This is supposed to contain the path to the file
1140 * backing the block device, but is actually just a free-form string you can pass to the kernel. Most
1141 * tools that actually care for the backing file path use the sysfs attribute file loop/backing_file
1142 * which is a kernel generated string, subject to file system namespaces and such.
1144 * .lo_file_name is useful since userspace can select it freely when creating a loopback block
1145 * device, and we can use it for /dev/loop/by-ref/ symlinks, and similar, so that apps can recognize
1146 * their own loopback files. */
1148 if (name
&& strlen(name
) >= sizeof(info
.lo_file_name
))
1151 if (ioctl(d
->fd
, LOOP_GET_STATUS64
, &info
) < 0)
1154 if (strneq((char*) info
.lo_file_name
, strempty(name
), sizeof(info
.lo_file_name
)))
1158 strncpy((char*) info
.lo_file_name
, name
, sizeof(info
.lo_file_name
)-1);
1159 info
.lo_file_name
[sizeof(info
.lo_file_name
)-1] = 0;
1161 memzero(info
.lo_file_name
, sizeof(info
.lo_file_name
));
1163 if (ioctl(d
->fd
, LOOP_SET_STATUS64
, &info
) < 0)