}
static int loop_configure_verify_direct_io(int fd, const struct loop_config *c) {
- assert(fd);
+ assert(fd >= 0);
assert(c);
if (FLAGS_SET(c->info.lo_flags, LO_FLAGS_DIRECT_IO)) {
* check here if enabling direct IO worked, to make this easily debuggable however.
*
* (Should anyone really care and actually wants direct IO on old kernels: it might be worth
- * enabling direct IO with iteratively larger block sizes until it eventually works.) */
+ * enabling direct IO with iteratively larger block sizes until it eventually works.)
+ *
+ * On older kernels (e.g.: 5.10) when this is attempted on a file stored on a dm-crypt
+ * backed partition the kernel will start returning I/O errors when accessing the mounted
+ * loop device, so return a recognizable error that causes the operation to be started
+ * from scratch without the LO_FLAGS_DIRECT_IO flag. */
if (!FLAGS_SET(info.lo_flags, LO_FLAGS_DIRECT_IO))
- log_debug("Could not enable direct IO mode, proceeding in buffered IO mode.");
+ return log_debug_errno(
+ SYNTHETIC_ERRNO(ENOANO),
+ "Could not enable direct IO mode, retrying in buffered IO mode.");
}
return 0;
* effect hence. And if not use classic LOOP_SET_STATUS64. */
uint64_t z;
- if (ioctl(fd, BLKGETSIZE64, &z) < 0)
- return -errno;
+ r = blockdev_get_device_size(fd, &z);
+ if (r < 0)
+ return r;
if (z != c->info.lo_sizelimit) {
log_debug("LOOP_CONFIGURE is broken, doesn't honour .info.lo_sizelimit. Falling back to LOOP_SET_STATUS64.");
/* Sleep some random time, but at least 10ms, at most 250ms. Increase the delay the more
* failed attempts we see */
- (void) usleep(UINT64_C(10) * USEC_PER_MSEC +
+ (void) usleep_safe(UINT64_C(10) * USEC_PER_MSEC +
random_u64_range(UINT64_C(240) * USEC_PER_MSEC * n_attempts/64));
}
assert_not_reached();
}
+ uint64_t device_size;
+ r = blockdev_get_device_size(loop_with_fd, &device_size);
+ if (r < 0)
+ return log_device_debug_errno(dev, r, "Failed to get loopback device size: %m");
+
LoopDevice *d = new(LoopDevice, 1);
if (!d)
return log_oom_debug();
.uevent_seqnum_not_before = seqnum,
.timestamp_not_before = timestamp,
.sector_size = c->block_size,
+ .device_size = device_size,
+ .created = true,
};
*ret = TAKE_PTR(d);
LoopDevice **ret) {
_cleanup_(loop_device_unrefp) LoopDevice *d = NULL;
- _cleanup_close_ int direct_io_fd = -EBADF, control = -EBADF;
+ _cleanup_close_ int reopened_fd = -EBADF, control = -EBADF;
_cleanup_free_ char *backing_file = NULL;
struct loop_config config;
int r, f_flags;
* Our intention here is that LO_FLAGS_DIRECT_IO is the primary knob, and O_DIRECT derived
* from that automatically. */
- direct_io_fd = fd_reopen(fd, (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0)|O_CLOEXEC|O_NONBLOCK|open_flags);
- if (direct_io_fd < 0) {
+ reopened_fd = fd_reopen(fd, (FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0)|O_CLOEXEC|O_NONBLOCK|open_flags);
+ if (reopened_fd < 0) {
if (!FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO))
- return log_debug_errno(errno, "Failed to reopen file descriptor without O_DIRECT: %m");
+ return log_debug_errno(reopened_fd, "Failed to reopen file descriptor without O_DIRECT: %m");
/* Some file systems might not support O_DIRECT, let's gracefully continue without it then. */
- log_debug_errno(errno, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without.");
+ log_debug_errno(reopened_fd, "Failed to enable O_DIRECT for backing file descriptor for loopback device. Continuing without.");
loop_flags &= ~LO_FLAGS_DIRECT_IO;
} else
- fd = direct_io_fd; /* From now on, operate on our new O_DIRECT fd */
+ fd = reopened_fd; /* From now on, operate on our new O_DIRECT fd */
}
control = open("/dev/loop-control", O_RDWR|O_CLOEXEC|O_NOCTTY|O_NONBLOCK);
/* -ENODEV or friends: Somebody might've gotten the same number from the kernel, used the
* device, and called LOOP_CTL_REMOVE on it. Let's retry with a new number.
* -EBUSY: a file descriptor is already bound to the loopback block device.
- * -EUCLEAN: some left-over partition devices that were cleaned up. */
- if (!ERRNO_IS_DEVICE_ABSENT(r) && !IN_SET(r, -EBUSY, -EUCLEAN))
+ * -EUCLEAN: some left-over partition devices that were cleaned up.
+ * -ENOANO: we tried to use LO_FLAGS_DIRECT_IO but the kernel rejected it. */
+ if (!ERRNO_IS_DEVICE_ABSENT(r) && !IN_SET(r, -EBUSY, -EUCLEAN, -ENOANO))
return r;
/* OK, this didn't work, let's try again a bit later, but first release the lock on the
if (++n_attempts >= 64) /* Give up eventually */
return -EBUSY;
+ /* If we failed to enable direct IO mode, let's retry without it. We restart the process as
+ * on some combination of kernel version and storage filesystem, the kernel is very unhappy
+ * about a failed DIRECT_IO enablement and throws I/O errors. */
+ if (r == -ENOANO && FLAGS_SET(config.info.lo_flags, LO_FLAGS_DIRECT_IO)) {
+ config.info.lo_flags &= ~LO_FLAGS_DIRECT_IO;
+ open_flags &= ~O_DIRECT;
+
+ int non_direct_io_fd = fd_reopen(config.fd, O_CLOEXEC|O_NONBLOCK|open_flags);
+ if (non_direct_io_fd < 0)
+ return log_debug_errno(
+ non_direct_io_fd,
+ "Failed to reopen file descriptor without O_DIRECT: %m");
+
+ safe_close(reopened_fd);
+ fd = config.fd = /* For cleanups */ reopened_fd = non_direct_io_fd;
+ }
+
/* Wait some random time, to make collision less likely. Let's pick a random time in the
* range 0ms…250ms, linearly scaled by the number of failed attempts. */
usec = random_u64_range(UINT64_C(10) * USEC_PER_MSEC +
UINT64_C(240) * USEC_PER_MSEC * n_attempts/64);
log_debug("Trying again after %s.", FORMAT_TIMESPAN(usec, USEC_PER_MSEC));
- (void) usleep(usec);
+ (void) usleep_safe(usec);
}
d->backing_file = TAKE_PTR(backing_file);
direct_flags = FLAGS_SET(loop_flags, LO_FLAGS_DIRECT_IO) ? O_DIRECT : 0;
rdwr_flags = open_flags >= 0 ? open_flags : O_RDWR;
- fd = xopenat(dir_fd, path, basic_flags|direct_flags|rdwr_flags, 0);
+ fd = xopenat(dir_fd, path, basic_flags|direct_flags|rdwr_flags, /* xopen_flags = */ 0, /* mode = */ 0);
if (fd < 0 && direct_flags != 0) /* If we had O_DIRECT on, and things failed with that, let's immediately try again without */
- fd = xopenat(dir_fd, path, basic_flags|rdwr_flags, 0);
+ fd = xopenat(dir_fd, path, basic_flags|rdwr_flags, /* xopen_flags = */ 0, /* mode = */ 0);
else
direct = direct_flags != 0;
if (fd < 0) {
if (open_flags >= 0 || !(ERRNO_IS_PRIVILEGE(r) || r == -EROFS))
return r;
- fd = xopenat(dir_fd, path, basic_flags|direct_flags|O_RDONLY, 0);
+ fd = xopenat(dir_fd, path, basic_flags|direct_flags|O_RDONLY, /* xopen_flags = */ 0, /* mode = */ 0);
if (fd < 0 && direct_flags != 0) /* as above */
- fd = xopenat(dir_fd, path, basic_flags|O_RDONLY, 0);
+ fd = xopenat(dir_fd, path, basic_flags|O_RDONLY, /* xopen_flags = */ 0, /* mode = */ 0);
else
direct = direct_flags != 0;
if (fd < 0)
}
/* Now that the block device is released, let's also try to remove it */
- if (control >= 0)
- for (unsigned n_attempts = 0;;) {
+ if (control >= 0) {
+ useconds_t delay = 5 * USEC_PER_MSEC; /* A total delay of 5090 ms between 39 attempts,
+ * (4*5 + 5*10 + 5*20 + … + 3*640) = 5090. */
+
+ for (unsigned attempt = 1;; attempt++) {
if (ioctl(control, LOOP_CTL_REMOVE, d->nr) >= 0)
break;
- if (errno != EBUSY || ++n_attempts >= 64) {
+ if (errno != EBUSY || attempt > 38) {
log_debug_errno(errno, "Failed to remove device %s: %m", strna(d->node));
break;
}
- (void) usleep(50 * USEC_PER_MSEC);
+ if (attempt % 5 == 0) {
+ log_debug("Device is still busy after %u attempts…", attempt);
+ delay *= 2;
+ }
+
+ (void) usleep_safe(delay);
}
+ }
free(d->node);
sd_device_unref(d->dev);
if (r < 0)
return r;
+ uint64_t device_size;
+ r = blockdev_get_device_size(fd, &device_size);
+ if (r < 0)
+ return r;
+
r = sd_device_get_devnum(dev, &devnum);
if (r < 0)
return r;
.uevent_seqnum_not_before = UINT64_MAX,
.timestamp_not_before = USEC_INFINITY,
.sector_size = sector_size,
+ .device_size = device_size,
+ .created = false,
};
*ret = d;
return -EINVAL;
current_offset *= 512U;
- if (ioctl(partition_fd, BLKGETSIZE64, ¤t_size) < 0)
- return -EINVAL;
+ r = blockdev_get_device_size(partition_fd, ¤t_size);
+ if (r < 0)
+ return r;
if (size == UINT64_MAX && offset == UINT64_MAX)
return 0;
* which is a kernel generated string, subject to file system namespaces and such.
*
* .lo_file_name is useful since userspace can select it freely when creating a loopback block
- * device, and we can use it for /dev/loop/by-ref/ symlinks, and similar, so that apps can recognize
- * their own loopback files. */
+ * device, and we can use it for /dev/disk/by-loop-ref/ symlinks, and similar, so that apps can
+ * recognize their own loopback files. */
if (name && strlen(name) >= sizeof(info.lo_file_name))
return -ENOBUFS;