On 10/13/2010 8:21 AM, Daniel P. Berrange wrote:
On Mon, Oct 11, 2010 at 12:21:19PM +0530, Harsh Prateek Bora wrote:
This patch introduces new attribute to filesystem element to support customizable security for mount type. Valid mount_security are: passthrough and mapped.
Usage: <filesystem type='mount' mount_security='passthrough'> <source dir='/export/to/guest'/> <target dir='mount_tag'/> </filesystem>
Here is the detailed explanation on these security models:
Security model: mapped ----------------------
Fileserver intercepts and maps all the file object create requests. Files on the fileserver will be created with Fileserver's user credentials and the client-user's credentials are stored in extended attributes. During getattr() server extracts the client-user's credentials from extended attributes and sends to the client.
This adds a great deal of security in the cloud environments where the guest's(client) user space is kept completely isolated from host's user space.
Security model : passthrough ----------------------------
In this security model, Fileserver passes down all requests to the underlying filesystem. File system objects on the fileserver will be created with client-user's credentials. This is done by setting setuid()/setgid() during creation or chmod/chown after file creation. At the end of create protocol request, files on the fileserver will be owned by cleint-user's uid/gid. This model mimic's current NFSv3 level of security.
In your first patch you had a 3rd option 'none', whose semantics I had asked about, because they appeared to be the same as passthrough. Looking at the QEMU virtio-9p.c file comments though, it appears that there is in fact a difference.
- In 'passthrough' the user/group ownership is preserved from the guest requests. - In 'mapped' the guest user/group ownership is stored in xtended attrs with files on host keeping same uid/gid as the QEMU process - In 'none' the user/group ownership from guest is completely ignored, and all files just have uid/gid matching the QEMU process.
If this interpretation of QEMU code is correct, then I think we should support all 3 options in libvirt after all.
The 'passthrough' option should be the default in libvirt, because that matches the semantics of the <filesystem> element usage in LXC / OpenVZ drivers in libvirt.
I think I'd give 'none' a different name in the XML , perhaps call it 'squash', inspired by NFS root squash, which squashes uid/gid onto a single user.
Finally I'm thinking that the security attribute should be called 'accessmode' rather than 'mount_security', mostly because I don't like underscores in XML attribute/element names.
I am complete in agreement with all the observations above. Harsh can quickly make these changes and repost the patch. Thanks, JV
Regards, Daniel