On Wed, 31 Jul 2013 10:21:19 +1000
Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> The only thing I think we still need to address is whether
> xfs_ioc_setattr() should allow users within a namespace to change
> the project ID of a file they otherwise own. That function is
> currently changed to use a inode_owner_or_capable() check and so if
> the uids match inside the namespace the modification is allowed.
Right, so before this change the caller has to own the file or be
CAP_FOWNER in init_user_ns. Changing to using inode_owner_or_capable()
means the caller has to own the file (and the inode's uid must be valid
in current_user_ns) or be CAP_FOWNER in current_user_ns. I don't
see how this lets the user the user do something inside the userns that
they can't do outside.
Basically I think we want to treat projids as a non namespace aware
identifier, but it sounds like maybe we don't want them manipulated
from userns context at all.
> However, right now for project IDs I think we have decided to limit
> manipulations to the init user namespace and not expose project IDs
> inside user namespaces at all. Hence I think that xfs_ioc_setattr()
> needs a further check for this...
If we don't want to allow setting the projid from a userns, the check
I would propose inside the if (mask & FSX_EXTSIZE) block is:
if (current_user_ns() != &init_user_ns)
Is the appropriate error return for this EINVAL? (thats what a similar
check in kernel/taskstats.c returns)