Re: [PATCH v4] Introduce v3 namespaced file capabilities

From: Stefan Berger
Date: Tue Jun 20 2017 - 13:34:08 EST

On 06/20/2017 08:19 AM, Stefan Berger wrote:
On 06/20/2017 01:42 AM, Amir Goldstein wrote:
On Tue, Jun 20, 2017 at 12:34 AM, Eric W. Biederman
<ebiederm@xxxxxxxxxxxx> wrote:
"Serge E. Hallyn" <serge@xxxxxxxxxx> writes:

Quoting Stefan Berger (stefanb@xxxxxxxxxxxxxxxxxx):
On 06/14/2017 11:05 PM, Serge E. Hallyn wrote:
On Wed, Jun 14, 2017 at 08:27:40AM -0400, Stefan Berger wrote:
On 06/13/2017 07:55 PM, Serge E. Hallyn wrote:
Quoting Stefan Berger (stefanb@xxxxxxxxxxxxxxxxxx):
If all extended
attributes were to support this model, maybe the 'uid' could be
associated with the 'name' of the xattr rather than its 'value' (not
sure whether that's possible).
Right, I missed that in your original email when I saw it this morning.
It's not what my patch does, but it's an interesting idea. Do you have
a patch to that effect? We might even be able to generalize that to
No, I don't have a patch. It may not be possible to implement it.
The xattr_handler's take the name of the xattr as input to get().
That may be ok though. Assume the host created a container with
100000 as the uid for root, which created a container with 130000 as
uid for root. If root in the nested container tries to read the
xattr, the kernel can check for[130000] first, then[100000], then Or, it can do a listxattr
and look for those. Am I overlooking one?

So one could try to encode the mapped uid in the name. However, that
I thought that's exactly what you were suggesting in your original
email? "security.capability[uid=2000]"

could lead to problems with stale xattrs in a shared filesystem over
time unless one could limit the number of xattrs with the same
prefix, e.g., security.capability*. So I doubt that it would work.
Hm. Yeah. But really how many setups are there like that? I.e. if
you launch a regular docker or lxd container, the image doesn't do a
bind mount of a shared image, it layers something above it or does a
copy. What setups do you know of where multiple containers in different
user namespaces mount the same filesystem shared and writeable?
I think I have something now that accomodates userns access to

Encoding of uid is in the attribute name now as follows:<uid>

1) The 'plain' security.capability is only r/w accessible from the
host (init_user_ns).
2) When userns reads/writes 'security.capability' it will read/write
security.capability@uid=<uid> instead, with uid being the uid of
root , e.g. 1000.
3) When listing xattrs for userns the host's security.capability is
filtered out to avoid read failures iof 'security.capability' if
security.capability@uid=<uid> is read but not there. (see 1) and 2))
4) security.capability* may all be read from anywhere
5) security.capability@uid=<uid> may be read or written directly
from a userns if <uid> matches the uid of root (current_uid())
This looks very close to what we want. One exception - we do want
to support root in a user namespace being able to write
security.capability@uid=<x> where <x> is a valid uid mapped in its
namespace. In that case the name should be rewritten to be
security.capability@uid=<y> where y is the unmapped kuid.val.


so far my patch hasn't yet hit Linus' tree. Given that, would you
mind taking a look and seeing what you think of this approach? If
we may decide to go this route, we probably should stop my patch
from hitting Linus' tree before we have to continue supporting it.
Agreed. I will take a look. I also want to see how all of this works
in the context of stackable filesystems. As that is the one case that
looked like it could be a problem case in your current patchset.

Apropos stackable filesystems [cc some overlayfs folks], is there any
way that parts of this work could be generalized towards ns aware
trusted@uid.* xattr?

I am at least removing all string comparison with xattr names from the core code and move the enabled xattr names into a list. For the security.* extended attribute names we would enumerated the enabled ones in that list, only security.capability for now. I am not sure how the trusted.* space works.

I extended 'the infrastructure' now to support prefix matching for trusted.* and probably others as well. It's fairly easy to do that but would not write the code like that for exact string matching to support security.capability. The patch lets me write from within the userns if uid=100 exists, rejects it otherwise. It may be written out as for root mapping to uid 1000. I can list this entry on the host. For some reason trusted.* is not listed at all inside the userns. So something else needs to be enabled as well. For now it looks like this: