Re: [PATCH 1/7] fs: Add user namesapace member to struct super_block

[Stephen Smalley]

On 08/06/2015 11:44 AM, Seth Forshee wrote:
> On Thu, Aug 06, 2015 at 10:51:16AM -0400, Stephen Smalley wrote:
>> On 08/06/2015 10:20 AM, Seth Forshee wrote:
>>> On Wed, Aug 05, 2015 at 04:19:03PM -0500, Eric W. Biederman wrote:
>>>> Seth Forshee <seth.forshee@xxxxxxxxxxxxx> writes:
>>>>
>>>>> On Wed, Jul 15, 2015 at 09:47:11PM -0500, Eric W. Biederman wrote:
>>>>>> Seth Forshee <seth.forshee@xxxxxxxxxxxxx> writes:
>>>>>>
>>>>>>> Initially this will be used to eliminate the implicit MNT_NODEV
>>>>>>> flag for mounts from user namespaces. In the future it will also
>>>>>>> be used for translating ids and checking capabilities for
>>>>>>> filesystems mounted from user namespaces.
>>>>>>>
>>>>>>> s_user_ns is initialized in alloc_super() and is generally set to
>>>>>>> current_user_ns(). To avoid security and corruption issues, two
>>>>>>> additional mount checks are also added:
>>>>>>>
>>>>>>>  - do_new_mount() gains a check that the user has CAP_SYS_ADMIN
>>>>>>>    in current_user_ns().
>>>>>>>
>>>>>>>  - sget() will fail with EBUSY when the filesystem it's looking
>>>>>>>    for is already mounted from another user namespace.
>>>>>>>
>>>>>>> proc needs some special handling here. The user namespace of
>>>>>>> current isn't appropriate when forking as a result of clone (2)
>>>>>>> with CLONE_NEWPID|CLONE_NEWUSER, as it will make proc unmountable
>>>>>>> from within the new user namespace. Instead, the user namespace
>>>>>>> which owns the new pid namespace should be used. sget_userns() is
>>>>>>> added to allow passing of a user namespace other than that of
>>>>>>> current, and this is used by proc_mount(). sget() becomes a
>>>>>>> wrapper around sget_userns() which passes current_user_ns().
>>>>>>
>>>>>> From bits of the previous conversation.
>>>>>>
>>>>>> We need sget_userns(..., &init_user_ns) for sysfs.  The sysfs
>>>>>> xattrs can travel from one mount of sysfs to another via the sysfs
>>>>>> backing store.
>>>>>>
>>>>>> For tmpfs and any other filesystems we support mounting without
>>>>>> privilige that support xattrs.  We need to identify them and
>>>>>> see if userspace is taking advantage of the ability to set
>>>>>> xattrs and file caps (unlikely).  If they are we need to call
>>>>>> sget_userns(..., &init_user_ns) on those filesystems as well.
>>>>>>
>>>>>> Possibly/Probably we should just do that for all of the interesting
>>>>>> filesystems to start with and then change back to an ordinary old sget
>>>>>> after we have done the testing and confirmed we will not be introducing
>>>>>> userspace regressions.
>>>>>
>>>>> I was reviewing everything in preparation for sending v2 patches, and I
>>>>> realized that doing this has an undesirable side effect. In patch 2 the
>>>>> implicit nodev is removed for unprivileged mounts, and instead s_user_ns
>>>>> is used to block opening devices in these mounts. When we set s_user_ns
>>>>> to &init_user_ns, it becomes possible to open device nodes from
>>>>> unprivileged mounts of these filesystems.
>>>>>
>>>>> This doesn't pose a real problem today. The only filesystems it will
>>>>> affect is sysfs, tmpfs, and ramfs (no others need s_user_ns =
>>>>> &init_user_ns for user namespace mounts), and all of these aren't
>>>>> problems. sysfs is okay because kernfs doesn't (currently?) allow device
>>>>> nodes, and a user would require CAP_MKNOD to create any device nodes in
>>>>> a tmpfs or ramfs mount.
>>>>>
>>>>> But for sysfs in particular it does mean that we will need to make sure
>>>>> that there's no way that device nodes could start appearing in an
>>>>> unprivileged mount.
>>>>
>>>> Good point about nodev.  
>>>>
>>>> For tmpfs and ramfs and security labels the smack policy of allowing but
>>>> filtering security labels mean smack once it has those bits will not
>>>> care which user namespace ramfs and tmpfs live in.  The labels should
>>>> pretty much stay the same in any case.
>>>
>>> Smack does care which namespace ramfs and tmpfs are in. With the patch
>>> I've got right now, if s_user_ns != &init_user_ns and the label of an
>>> inode does not match that of the root inode then
>>> security_inode_permission() will return EACCES.
>>>
>>> So if something with CAP_MAC_ADMIN is changing security labels in such a
>>> mount, suddenly those inodes might become inaccessible. And while it may
>>> be unlikely that anyone is doing this it's impossible for me to prove
>>> that's the case.
>>>
>>>> If the same class of handling will also apply to selinux and those are
>>>> the only two security modules that apply labels than we can leave tmpfs
>>>> and ramfs with the security labels of whomever mounted them.
>>>
>>> For SELinux I now have a patch which applies mountpoint labeling to
>>> mounts for which s_user_ns != &init_user_ns. I'm less sure then with
>>> Smack how this behavior will differ from what happens today, but my
>>> understanding is that this means that the label of the mountpoint is
>>> used for all objects from that superblock. Afaik it does not have the
>>> Smack behavior of denying access to filesystem objects which have a
>>> different label in the backing store.
>>>
>>>> For sysfs things get a little more interesting.  Assuming tmpfs and
>>>> ramfs don't need s_user_ns == &init_user_ns, sysfs may be fine operating
>>>> with possibly invalid securitly labels set on a different mount of
>>>> selinux.  (I am wondering now how all of these labels work in the
>>>> context of nfs).
>>>
>>> If someone was using Smack to label sysfs then a mount with s_user_ns !=
>>> &init_user_ns is going to leave inaccessible anything without the same
>>> label as the process which performed the mount.
>>>
>>> Again with SELinux I'm less certain, but I think you could end up with a
>>> sysfs superblock that has mountpoint labeling, and thus any labels set
>>> in the mount in the init namespace would be ignored.
>>
>> If you're using the logic I suggested for SELinux, then SELinux will
>> only use mountpoint labeling if SELinux would otherwise fetch the
>> extended attribute value from the filesystem via ->getxattr (this is the
>> SECURITY_FS_USE_XATTR test in the code).  As this is not the case for
>> purely in-memory filesystems like tmpfs, ramfs, or sysfs, SELinux will
>> still label those filesystems in the usual manner, i.e. it initially
>> computes a default label for new inodes, and if userspace later performs
>> a setxattr(), then it updates its internal state at that time from the
>> relevant hooks (inode_post_setxattr or inode_setsecurity).
>> So nothing should change for SELinux wrt labeling of tmpfs, ramfs, or
>> sysfs in userns mounts aside from not allowing the use of the additional
>> mount options (e.g. context=).
> 
> This is the patch I have currently:
> 
> http://kernel.ubuntu.com/git/sforshee/linux.git/commit/?h=userns-mounts&id=080e5f5ee58143a56cfc57b4e51dff58b7a3cb1a
> 
> I haven't been able to figure out which labeling behavior sysfs would
> end up with normally from just inspecting the code. kernfs does support
> xattrs, but now that I look at the implementation it handles security
> xattrs differently and calls security_inode_setsecurity whenever one is
> written. I'm not sure how all of that is going to work out in practice
> with SELinux.

sysfs would have a labeling behavior of SECURITY_FS_USE_GENFS
(policy-driven).  It wouldn't make sense to configure sysfs with
SECURITY_FS_USE_XATTR, because that would cause SELinux to ask the
filesystem via ->getxattr for the initial value for the label when the
inode is first instantiated, and sysfs would have no answer there.  So,
in practice, sysfs will still get labeled exactly as before, and there
would be no change in behavior.  Similarly for tmpfs
(SECURITY_FS_USE_TRANS) or ramfs.  The only filesystem types that get
SECURITY_FS_USE_XATTR are the ones that actually support storing SELinux
attributes persistently and therefore could provide an initial value
from backing store.

>> Also, a superblock can only have a single labeling behavior, so you
>> can't have different mounts of sysfs, one using mountpoint labeling and
>> one not.  An inode can only have one label, no matter how you reach it.
> 
> There are multiple sysfs superblocks though, see sysfs_mount(). It calls
> kernfs_mount_ns(), passing a kobject for the current net ns.
> kernfs_test_super() only matches if the net ns matches an existing
> superblock, so you end up with a different superblock per net ns.
> 
> For kobjects which aren't namespaced, the same path within two different
> sysfs superblocks will be backed by the same kernfs node. kernfs stashes
> the security context inside the kernfs node, so inodes in different
> superblocks backed by the same kernfs node will have the same security
> context.
> 
> So, with sysfs you can have different superblocks with (partially) the
> same backing store, and it would be possible for those superblocks to
> end up with different labeling behavior. I think we want to avoid having
> security labels applied to sysfs files in the init namespace and have
> those get lost in a mount from another namespace.

As long as we prohibit context= mounts on the userns mounts (which your
patch does), then this shouldn't be possible.  Maybe we should just do
that for sysfs always.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/