Re: [PATCH 0/2] virtio-fs: Add 'file' mount option

[Hanna Czenczek <hreitz@redhat.com>]

On 10.07.24 20:42, Josef Bacik wrote:
> On Wed, Jul 10, 2024 at 09:28:08AM +0200, Hanna Czenczek wrote:
>> On 09.07.24 19:56, Josef Bacik wrote:
>>> On Tue, Jul 09, 2024 at 01:19:16PM +0200, Hanna Czenczek wrote:
>>>> Hi,
>>>>
>>>> We want to be able to mount filesystems that just consist of one regular
>>>> file via virtio-fs, i.e. no root directory, just a file as the root
>>>> node.
>>>>
>>>> While that is possible via FUSE itself (through the 'rootmode' mount
>>>> option, which is automatically set by the fusermount help program to
>>>> match the mount point's inode mode), there is no virtio-fs option yet
>>>> that would allow changing the rootmode from S_IFDIR to S_IFREG.
>>>>
>>>> To do that, this series introduces a new 'file' mount option that does
>>>> precisely that.  Alternatively, we could provide the same 'rootmode'
>>>> option that FUSE has, but as laid out in patch 1's commit description,
>>>> that option is a bit cumbersome for virtio-fs (in a way that it is not
>>>> for FUSE), and its usefulness as a more general option is limited.
>>>>
>>> All this does is make file an alias for something a little easier for users to
>>> read, which can easily be done in libfuse.  Add the code to lib/mount.c to alias
>>> 'file' to turn it into rootmode=S_IFREG when it sends it to the kernel, it's not
>>> necessary to do this in the kernel.  Thanks,
>> This series is not about normal FUSE filesystems (file_system_type
>> fuse_fs_type, “fuse”), but about virtio-fs (file_system_type virtio_fs_type,
>> “virtiofs”), i.e. a case where libfuse and fusermount are not involved at
>> all.  As far as I’m aware, mounting a virtio-fs filesystem with a
>> non-directory root inode is currently not possible at all.
> Ok so I think I had it backwards in my head, my apologies.
>
> That being said I still don't understand why this requires a change to virtiofs
> at all.
>
> I have a virtiofs thing attached to my VM.  Inside the vm I do
>
> mount -t virtiofs <name of thing I've attached to the vm> /directory
>
> and then on the host machine, virtiofsd is a "normal" FUSE driver, except it's
> talking over the socket you setup between the guest and the host.  I assume this
> is all correct?
>
> So then the question is, why does it matter what virtiofsd is exposing?  I guess
> that's the better question.  The guest shouldn't have to care if it's a
> directory or a file right?  The mountpoint is going to be a directory, whatever
> is backing it shouldn't matter.  Could you describe the exact thing you're
> trying to accomplish?  Thanks,

The mount point needs to be of the same mode as the root node of the 
mounted filesystem, or it’ll be inaccessible after mounting[1].  In this 
case, I want to export a regular file as the root node, so the root node 
must be a regular file, too:

host$ echo foo > /tmp/bar

host$ virtiofsd --shared-dir /tmp/bar --socket-path /tmp/viofsd.sock 
--sandbox none


guest# mkdir /tmp/mnt-dir

guest# mount -t virtiofs virtiofs-tag /tmp/mnt-dir

guest# stat /tmp/mnt-dir
stat: cannot statx '/tmp/mnt-dir': Input/output error

guest# cat /tmp/mnt-dir
cat: /tmp/mnt-dir: Input/output error

guest# ls /tmp/mnt-dir
ls: cannot access '/tmp/mnt-dir': Input/output error

guest# umount /tmp/mnt-dir

(following with this series applied)

guest# touch /tmp/mnt-file

guest# mount -t virtiofs virtiofs-tag /tmp/mnt-file -o file

guest# stat /tmp/mnt-file
   File: /tmp/mnt-file
   Size: 4               Blocks: 8          IO Block: 4096   regular file
[...]

guest# cat /tmp/mnt-file
foo

guest# ls --file-type /tmp/mnt-file
/tmp/mnt-file

guest# ls --file-type /tmp
mnt-dir/
mnt-file
[...]


[1] As far as I remember, FUSE/virtio-fs will present the root node’s 
mode as 'rootmode' during mounting, and so the d_is_dir() equality 
checks in do_move_mount() and graft_tree() just check whether that 
matches the mount point’s mode.  So, like in the example above, mounting 
a filesystem whose root node is a regular file to a directory mount 
point without '-o file' succeeds.  But accessing it then fails, probably 
because the mismatch is then noticed somewhere (virtiofsd receives a 
GETATTR request, that’s it), i.e. the root node is supposed to be a 
directory, but it turns out not to be after all.

Hanna