Re: [PATCH] security: selinux: allow per-file labeling for bpffs

From: Stephen Smalley
Date: Wed Feb 12 2020 - 13:08:30 EST


On 2/12/20 12:46 PM, Steven Moreland wrote:
And I strongly encourage our downstream in the same way :) I try, I
try. However, I am a n00b here (thanks for merging "my" first linux
patch).

Looking at this code, I was wondering, why isn't SELinux labelling
completely orthogonal to the fs type? Is this a measurable
memory/performance thing?

If you just mean why don't we turn on SE_SBGENFS for all filesystem types, that's discussed in
https://github.com/SELinuxProject/selinux-kernel/issues/2

It isn't always safe so we have been whitelisting the filesystem types that are supported.

More generally, labeling in SELinux goes beyond just GENFS; there are the SECURITY_FS_USE_* filesystem labeling behaviors defined by policy and those are also based on fstype, just not hardcoded in the kernel.



On Tue, Feb 11, 2020 at 7:17 PM Paul Moore <paul@xxxxxxxxxxxxxx> wrote:

On Thu, Feb 6, 2020 at 1:12 PM Stephen Smalley <sds@xxxxxxxxxxxxx> wrote:
On 2/6/20 12:41 PM, Steven Moreland wrote:
On Thu, Feb 6, 2020 at 9:35 AM Stephen Smalley <sds@xxxxxxxxxxxxx> wrote:

On 2/6/20 12:21 PM, Stephen Smalley wrote:
On 2/6/20 11:55 AM, Steven Moreland wrote:
From: Connor O'Brien <connoro@xxxxxxxxxx>

Add support for genfscon per-file labeling of bpffs files. This allows
for separate permissions for different pinned bpf objects, which may
be completely unrelated to each other.

Do you want bpf fs to also support userspace labeling of files via
setxattr()? If so, you'll want to also add it to
selinux_is_genfs_special_handling() as well.


Android doesn't currently have this use case.

The only caveat I would note here is that it appears that bpf fs
supports rename, link, unlink, rmdir etc by userspace, which means that
name-based labeling via genfscon isn't necessarily safe/stable. See
https://github.com/SELinuxProject/selinux-kernel/issues/2


Android restricts ownership of these files to a single process (bpfloader) and
so this isn't a concern in our architecture. Is it a concern in general?

I guess if the inodes are pinned in memory, then only the original name
under which the file is created will be relevant to determining the
label and subsequent rename/link operations won't have any effect. So as
long as the bpfloader creates the files with the same names being
specified in policy, that should line up and be stable for the lifecycle
of the inode.

The alternative model is to have bpfloader look up a context from the
userspace file_contexts configuration via selabel_lookup(3) and friends,
and set it on the file explicitly. That's what e.g. ueventd does for
device nodes. However, one difference here is that you could currently
only do this via setxattr()/setfilecon() after creating the file so that
the file would temporarily exist in the default label for bpf fs, if
that matters. ueventd can instead use setfscreatecon(3) before creating
the file so that it is originally created in the right label but that
requires the filesystem to call security_inode_init_security() from its
function that originally creates the inode, which tmpfs/devtmpfs does
but bpf does not. So you'd have to add that to the bpf filesystem code
if you wanted to support setfscreatecon(3) on it.

Considering the relative maturity of bpf, and bpffs, I think it's okay
to take this small step right now, with the understanding that more
work may need to be done, depending on how this is generally adopted
by distros and users (for those of you not following the other thread,
I've merged the v3 draft of this patch).

However, I've been noticing a trend from the Android folks of tossing
patches over the wall without much thought beyond the Android use
case. I understand the Android devs have a job to do, and products to
focus on, but I would strongly encourage them to think a bit longer
about more general use cases before submitting patches upstream.

--
paul moore
www.paul-moore.com