On Fri 07-07-23 09:39:05, Christian Brauner wrote:

On Thu, Jul 06, 2023 at 06:12:55PM +0200, Jan Kara wrote:

quoted

On Thu 06-07-23 08:55:18, Christoph Hellwig wrote:

quoted

On Tue, Jul 04, 2023 at 02:56:54PM +0200, Jan Kara wrote:

quoted

When we don't allow opening of mounted block devices for writing, bind
mounting is broken because the bind mount tries to open the block device
before finding the superblock for it already exists. Reorganize the
mounting code to first look whether the superblock for a particular
device is already mounted and open the block device only if it is not.

Warning: this might be a rathole.

I really hate how mount_bdev / get_tree_bdev try to deal with multiple
mounts.

The idea to just open the device and work from there just feels very
bogus.

There is really no good reason to have the bdev to find a superblock,
the dev_t does just fine (and in fact I have a patch to remove
the bdev based get_super and just use the dev_t based one all the
time).  So I'd really like to actually turn this around and only
open when we need to allocate a new super block.  That probably
means tearning sget_fc apart a bit, so it will turn into a fair
amount of work, but I think it's the right thing to do.

Well, this is exactly what this patch does - we use dev_t to lookup the
superblock in sget_fc() and we open the block device only if we cannot find
matching superblock and need to create a new one...

Can you do this rework independent of the bdev_handle work that you're
doing so this series doesn't depend on the other work and we can get
the VFS bits merged for this?

Yeah, it should be doable. I'll have a look into it.

								Honza
-- 
Jan Kara [off-list ref]
SUSE Labs, CR