Re: loop subsystem corrupted after mounting multiple btrfs sub-volumes

From: Al Viro
Date: Fri Feb 26 2016 - 15:30:19 EST

On Fri, Feb 26, 2016 at 03:05:27PM -0500, Austin S. Hemmelgarn wrote:
> >Where is /mnt/2?
> It's kind of interesting, but I can't reproduce _any_ of this
> behavior with either ext4 or BTRFS when I manually set up the loop
> devices and point mount(8) at those instead of using -o loop on a
> file. That really seems to indicate that this is caused by something
> mount(8) is doing when it's calling losetup. I'm running a mostly
> unmodified version of 4.4.2 (the only modification that would come
> even remotely close to this is that I changed the default mount
> options for everything from relatime to noatime), and util-linux
> 2.27.1 from Gentoo.

Sigh... sys_mount() (mount_bdev(), actually) has no way to tell if two
loop devices refer to the same underlying object. As far as it's
concerned, you are asking to mount a completely unrelated block device.
Which just happens to see the data (living in separate pagecache, even)
modified behind its back (with some delay) after it gets written to another
device. Filesystem drivers generally don't like when something is screwing
the underlying data, to put it mildly...

When you ask to mount the _same_ device, mount_bdev(), as well as btrfs
counterpart, makes sure that you get a reference to the same struct
super_block, which avoids all coherency problems - all mounted instances
refer to the same in-core objects (dentries, inodes, page cache, etc.).
They get separate struct vfsmount instances, but that only matters for
mountpoint crossing.

As soon as you've set the second /dev/loop alias for the same underlying
file, you are asking for all kinds of trouble. If you use the same one
consistently, you are OK. BTW, even
losetup /dev/loop0 /dev/sda1
mount -t ext2 /dev/sda1 /mnt/1
mount -t ext2 /dev/loop0 /mnt/2
is enough for trouble - you get (as far as ext2 knows) unrelated devices
screwing each other, with no good way to predict that. And you need to
check propagation through more than one layer - loop over loop over block
is also possible.

IMO on-demand losetup a-la -o loop is simply a bad idea...