Opening a block device may require a long operation such as waiting for
the cdrom tray to close. Performing this operation with locks held locks
out other attempts to open the device. These processes waiting to open
the device are not killable.
To avoid this issue and still be able to perform time-consuming checks
at open() time the block device driver can provide open_finish(). If it
does opening the device proceeds even when an error is returned from
open(), bd_mutex is released and open_finish() is called. If
open_finish() succeeds the device is now open, if it fails release() is
called.
When -ERESTARTSYS is returned from open() blkdev_get may loop without
calling open_finish(). On -ERESTARTSYS open_finish() is not called.
Move a ret = 0 assignment up in the if/else branching to avoid returning
-ENXIO. Previously the return value was ignored on the unhandled branch.
Signed-off-by: Michal Suchanek <[email protected]>
---
Documentation/filesystems/locking.rst | 2 ++
fs/block_dev.c | 21 +++++++++++++++++----
include/linux/blkdev.h | 1 +
3 files changed, 20 insertions(+), 4 deletions(-)
diff --git a/Documentation/filesystems/locking.rst b/Documentation/filesystems/locking.rst
index fc3a0704553c..2471ced5a8cf 100644
--- a/Documentation/filesystems/locking.rst
+++ b/Documentation/filesystems/locking.rst
@@ -456,6 +456,7 @@ block_device_operations
prototypes::
int (*open) (struct block_device *, fmode_t);
+ int (*open_finish) (struct block_device *, fmode_t, int);
int (*release) (struct gendisk *, fmode_t);
int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
int (*compat_ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
@@ -473,6 +474,7 @@ locking rules:
ops bd_mutex
======================= ===================
open: yes
+open_finish: no
release: yes
ioctl: no
compat_ioctl: no
diff --git a/fs/block_dev.c b/fs/block_dev.c
index 9c073dbdc1b0..009b5dedb1f7 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -1526,6 +1526,7 @@ static int __blkdev_get(struct block_device *bdev, fmode_t mode, int for_part)
int partno;
int perm = 0;
bool first_open = false;
+ bool need_finish = false;
if (mode & FMODE_READ)
perm |= MAY_READ;
@@ -1581,6 +1582,8 @@ static int __blkdev_get(struct block_device *bdev, fmode_t mode, int for_part)
put_disk_and_module(disk);
goto restart;
}
+ if (bdev->bd_disk->fops->open_finish)
+ need_finish = true;
}
if (!ret) {
@@ -1601,7 +1604,7 @@ static int __blkdev_get(struct block_device *bdev, fmode_t mode, int for_part)
invalidate_partitions(disk, bdev);
}
- if (ret)
+ if (ret && !need_finish)
goto out_clear;
} else {
struct block_device *whole;
@@ -1627,10 +1630,14 @@ static int __blkdev_get(struct block_device *bdev, fmode_t mode, int for_part)
if (bdev->bd_bdi == &noop_backing_dev_info)
bdev->bd_bdi = bdi_get(disk->queue->backing_dev_info);
} else {
+ ret = 0;
if (bdev->bd_contains == bdev) {
- ret = 0;
- if (bdev->bd_disk->fops->open)
+ if (bdev->bd_disk->fops->open) {
ret = bdev->bd_disk->fops->open(bdev, mode);
+ if ((ret != -ERESTARTSYS) &&
+ bdev->bd_disk->fops->open_finish)
+ need_finish = true;
+ }
/* the same as first opener case, read comment there */
if (bdev->bd_invalidated) {
if (!ret)
@@ -1638,7 +1645,7 @@ static int __blkdev_get(struct block_device *bdev, fmode_t mode, int for_part)
else if (ret == -ENOMEDIUM)
invalidate_partitions(bdev->bd_disk, bdev);
}
- if (ret)
+ if (ret && !need_finish)
goto out_unlock_bdev;
}
}
@@ -1650,6 +1657,12 @@ static int __blkdev_get(struct block_device *bdev, fmode_t mode, int for_part)
/* only one opener holds refs to the module and disk */
if (!first_open)
put_disk_and_module(disk);
+ if (ret && need_finish)
+ ret = bdev->bd_disk->fops->open_finish(bdev, mode, ret);
+ if (ret) {
+ __blkdev_put(bdev, mode, for_part);
+ return ret;
+ }
return 0;
out_clear:
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index f3ea78b0c91c..b67e93c6afb7 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -1695,6 +1695,7 @@ static inline struct bio_vec *rq_integrity_vec(struct request *rq)
struct block_device_operations {
int (*open) (struct block_device *, fmode_t);
+ int (*open_finish)(struct block_device *bdev, fmode_t mode, int ret);
void (*release) (struct gendisk *, fmode_t);
int (*rw_page)(struct block_device *, sector_t, struct page *, unsigned int);
int (*ioctl) (struct block_device *, fmode_t, unsigned, unsigned long);
--
2.23.0
Please make sure you CC linux-block if you add block device ops.
On Fri, Oct 25, 2019 at 01:21:42PM +0200, Michal Suchanek wrote:
> Opening a block device may require a long operation such as waiting for
> the cdrom tray to close. Performing this operation with locks held locks
> out other attempts to open the device. These processes waiting to open
> the device are not killable.
>
> To avoid this issue and still be able to perform time-consuming checks
> at open() time the block device driver can provide open_finish(). If it
> does opening the device proceeds even when an error is returned from
> open(), bd_mutex is released and open_finish() is called. If
> open_finish() succeeds the device is now open, if it fails release() is
> called.
>
> When -ERESTARTSYS is returned from open() blkdev_get may loop without
> calling open_finish(). On -ERESTARTSYS open_finish() is not called.
>
> Move a ret = 0 assignment up in the if/else branching to avoid returning
> -ENXIO. Previously the return value was ignored on the unhandled branch.
Still a complete nack for splitting a fundamental operation over two
ops, especially just for working around a piece of buggy software.
On Mon, Nov 04, 2019 at 04:17:27PM -0800, Christoph Hellwig wrote:
> Please make sure you CC linux-block if you add block device ops.
>
> On Fri, Oct 25, 2019 at 01:21:42PM +0200, Michal Suchanek wrote:
> > Opening a block device may require a long operation such as waiting for
> > the cdrom tray to close. Performing this operation with locks held locks
> > out other attempts to open the device. These processes waiting to open
> > the device are not killable.
> >
> > To avoid this issue and still be able to perform time-consuming checks
> > at open() time the block device driver can provide open_finish(). If it
> > does opening the device proceeds even when an error is returned from
> > open(), bd_mutex is released and open_finish() is called. If
> > open_finish() succeeds the device is now open, if it fails release() is
> > called.
> >
> > When -ERESTARTSYS is returned from open() blkdev_get may loop without
> > calling open_finish(). On -ERESTARTSYS open_finish() is not called.
> >
> > Move a ret = 0 assignment up in the if/else branching to avoid returning
> > -ENXIO. Previously the return value was ignored on the unhandled branch.
>
> Still a complete nack for splitting a fundamental operation over two
> ops, especially just for working around a piece of buggy software.
Still did not provide an awesome alternative that does not sneed
splitting the operation.
What is it, specifically?
Thanks
Michal