Re: [PATCH 5/5] overlayfs: Use vfs_getxattr_noperm() for real inode

[Miklos Szeredi]

On Thu, Jul 7, 2016 at 8:35 PM, Vivek Goyal <vgoyal@xxxxxxxxxx> wrote:
> On Wed, Jul 06, 2016 at 04:58:37PM +0200, Miklos Szeredi wrote:
>> On Wed, Jul 6, 2016 at 12:54 PM, Vivek Goyal <vgoyal@xxxxxxxxxx> wrote:
>> > On Wed, Jul 06, 2016 at 06:36:49AM +0200, Miklos Szeredi wrote:
>> >> On Tue, Jul 5, 2016 at 11:16 PM, Vivek Goyal <vgoyal@xxxxxxxxxx> wrote:
>> >> > On Tue, Jul 05, 2016 at 01:29:39PM -0700, Casey Schaufler wrote:
>> >> >> On 7/5/2016 8:50 AM, Vivek Goyal wrote:
>> >> >> > ovl_getxattr() currently uses vfs_getxattr() on realinode. This fails
>> >> >> > if mounter does not have DAC/MAC permission to access getxattr.
>> >> >> >
>> >> >> > Specifically this becomes a problem when selinux is trying to initialize
>> >> >> > overlay inode and does ->getxattr(overlay_inode). A task might trigger
>> >> >> > initialization of overlay inode and we will access real inode xattr in the
>> >> >> > context of mounter and if mounter does not have permissions, then inode
>> >> >> > selinux context initialization fails and inode is labeled as unlabeled_t.
>> >> >> >
>> >> >> > One way to deal with it is to let SELinux do getxattr checks both on
>> >> >> > overlay inode and underlying inode and overlay can call vfs_getxattr_noperm()
>> >> >> > to make sure when selinux is trying to initialize label on inode, it does
>> >> >> > not go through checks on lower levels and initialization is successful.
>> >> >> > And after inode initialization, SELinux will make sure task has getatttr
>> >> >> > permission.
>> >> >> >
>> >> >> > One issue with this approach is that it does not work for directories as
>> >> >> > d_real() returns the overlay dentry for directories and not the underlying
>> >> >> > directory dentry.
>> >> >> >
>> >> >> > Another way to deal with it to introduce another function pointer in
>> >> >> > inode_operations, say getxattr_noperm(), which is responsible to get
>> >> >> > xattr without any checks. SELinux initialization code will call this
>> >> >> > first if it is available on inode. So user space code path will call
>> >> >> > ->getxattr() and that will go through checks and SELinux internal
>> >> >> > initialization will call ->getxattr_noperm() and that will not
>> >> >> > go through checks.
>> >> >> >
>> >> >> > For now, I am just converting ovl_getxattr() to get xattr without
>> >> >> > any checks on underlying inode. That means it is possible for
>> >> >> > a task to get xattr of a file/dir on lower/upper through overlay mount
>> >> >> > while it is not possible outside overlay mount.
>> >> >> >
>> >> >> > If this is a major concern, I can look into implementing getxattr_noperm().
>> >> >>
>> >> >> This is a major concern.
>> >> >
>> >> > Hmm.., In that case I will write patch to provide another inode operation
>> >> > getxattr_noperm() and a wrapper which falls back to getxattr() if noperm
>> >> > variant is not defined. That should take care of this issue.
>> >>
>> >> That's not going to fly.  A slighly better, but still quite ugly
>> >> solution would be to add a "flags" arg to the current ->getxattr()
>> >> callback indicating whether the caller wants permission checking
>> >> inside the call or not.
>> >>
>> >
>> > Ok, will try that.
>> >
>> >> But we already have the current->creds.  Can't that be used to control
>> >> the permission checking done by the callback?
>> >
>> > Sorry, did not get how to use current->creds to control permission
>> > checking.
>>
>> I'm not sure about the details either.  But current->creds *is* the
>> context provided for the VFS and filesystems to check permissions.  It
>> might make sense to use that to indicate to overlayfs that permission
>> should not be checked.
>
> That sounds like raising capabilities of task temporarily to do
> getxattr(). But AFAIK, there is no cap which will override SELinux checks.

So a new capability can be invented for this purpose?

> I am taking a step back re-thinking about the problem.
>
> - For context mounts this is not a problem at all as overlay inode will
>   get its label from context= mount option and we will not call into
>   ovl_getxattr().
>
> - For non-context mounts this is a problem only if mounter is not
>   privileged enough to do getattr. And that's not going to be a common
>   case either.
>
> IOW, this does not look like a common case. And if getxattr() fails,
> SELinux already seems to mark inode as unlabeled_t. And my understanding
> is that task can't access unlabeled_t anyway, so there is no information
> leak.
>
> So for now, why not leave it as it is. Only side affect I seem to see
> is following warnings on console.
>
> SELinux: inode_doinit_with_dentry:  getxattr returned 13 for dev=overlay ino=29147
>
> This is for information purposes only and given getxattr() can fail in
> stacked configuration, I think we can change this to KERN_DEBUG instead
> of KERN_WARNING.

I'm fine with this as well.

Thanks,
Miklos