Message-ID: <15d129d8e601a98d5b447da6a6032c8f984d2638.camel@huaweicloud.com>
Date: Mon, 06 May 2024 14:34:03 +0200
From: Roberto Sassu <roberto.sassu@...weicloud.com>
To: Amir Goldstein <amir73il@...il.com>
Cc: Stefan Berger <stefanb@...ux.ibm.com>, linux-integrity@...r.kernel.org, 
 linux-unionfs@...r.kernel.org, linux-kernel@...r.kernel.org,
 zohar@...ux.ibm.com,  roberto.sassu@...wei.com, miklos@...redi.hu,
 brauner@...nel.org, Vivek Goyal <vgoyal@...hat.com>
Subject: Re: [RFC PATCH v2 0/2] ima: Fix detection of read/write violations
 on stacked filesystems

On Sat, 2024-04-27 at 12:03 +0300, Amir Goldstein wrote:
> On Fri, Apr 26, 2024 at 10:34 AM Roberto Sassu
> <roberto.sassu@...weicloud.com> wrote:
> > 
> > On Thu, 2024-04-25 at 15:37 +0300, Amir Goldstein wrote:
> > > > On Thu, Apr 25, 2024 at 2:30 PM Roberto Sassu
> > > > <roberto.sassu@...weicloud.com> wrote:
> > > > > > 
> > > > > > On Tue, 2024-04-23 at 09:02 +0300, Amir Goldstein wrote:
> > > > > > > > On Mon, Apr 22, 2024 at 6:07 PM Stefan Berger <stefanb@...ux.ibm.com> wrote:
> > > > > > > > > > 
> > > > > > > > > > This series fixes the detection of read/write violations on stacked
> > > > > > > > > > filesystems. To be able to access the relevant dentries necessary to
> > > > > > > > > > detect files opened for writing on a stacked filesystem a new d_real_type
> > > > > > > > > > D_REAL_FILEDATA is introduced that allows callers to access all relevant
> > > > > > > > > > files involved in a stacked filesystem while traversing the layers.
> > > > > > > > > > 
> > > > > > > > 
> > > > > > > > Stefan,
> > > > > > > > 
> > > > > > > > Both Miklos and myself objected to this solution:
> > > > > > > > https://lore.kernel.org/linux-unionfs/CAJfpeguctirEYECoigcAsJwpGPCX2NyfMZ8H8GHGW-0UyKfjgg@mail.gmail.com/
> > > > > > > > 
> > > > > > > > Not sure what you are hoping to achieve from re-posting the same solution.
> > > > > > > > 
> > > > > > > > I stopped counting how many times I already argued that *all* IMA/EVM
> > > > > > > > assertions,
> > > > > > > > including rw-ro violations should be enforced only on the real inode.
> > > > > > > > I know this does not work - so you should find out why it does not work and fix
> > > > > > > > the problem.
> > > > > > > > 
> > > > > > > > Enforcing IMA/EVM on the overlayfs inode layer is just the wrong way IMO.
> > > > > > > > Not once have I heard an argument from IMA/EVM developers why it is really
> > > > > > > > needed to enforce IMA/EVM on the overlayfs inode layer and not on the
> > > > > > > > real inode.
> > > > > > 
> > > > > > Ok, I try to provide an example regarding this, and we see if it makes
> > > > > > sense.
> > > > > > 
> > > > > > # echo test > test-file
> > > > > > # chown 2000 d/test-file
> > > > > > # ls -l a/test-file
> > > > > > -rw-r--r--. 1 2000 root 25 Apr 25 10:50 a/test-file
> > > > > > 
> > > > > > Initially there is a file in the lower layer with UID 2000.
> > > > > > 
> > > > > > 
> > > > > > # mount -t overlay -olowerdir=a,upperdir=b,workdir=c,metacopy=on overlay d
> > > > > > # chown 3000 d/test-file
> > > > > > # ls -l d/test-file
> > > > > > -rw-r--r--. 1 3000 root 25 Apr 25 10:50 d/test-file
> > > > > > # ls -l a/test-file
> > > > > > -rw-r--r--. 1 2000 root 25 Apr 25 10:50 a/test-file
> > > > > > # ls -l b/test-file
> > > > > > -rw-r--r--. 1 3000 root 25 Apr 25 10:50 b/test-file
> > > > > > 
> > > > > > If I have a policy like this:
> > > > > > 
> > > > > > # echo "measure fsname=overlay fowner=3000" > /sys/kernel/security/ima/policy
> > > > > > 
> > > > > > there won't be any match on the real file which still has UID 2000. But
> > > > > > what is observable by the processes through overlayfs is UID 3000.
> > > > > > 
> > > > 
> > > > ok, it is simple to write an ima policy that is not respected by overlayfs.
> > > > 
> > > > My question is: under what circumstances is a policy like the above
> > > > useful in the real world?
> > > > 
> > > > Correct me if I am wrong, but AFAIK, the purpose of IMA/EVM is to
> > > > mitigate attack vectors of tampering with files offline or after the
> > > > file's data/metadata were measured. Is that a correct description?
> > 
> > (For now I would talk about IMA, EVM can be considered separately).
> > 
> > The main purpose of IMA is to evaluate files being accessed, and record
> > the access together with a file digest in a measurement list,
> > allow/deny access to the file (appraisal), or add a new event to audit
> > logs.
> > 
> > How files are selected depends on the IMA policy. A rule can be
> > subject-based or object-based, depending on whether respectively
> > process or file attributes are matched. It can also be both.
> > 
> > A subject-based rule means that you identify a process/set of
> > processes, and you evaluate everything it/they read.
> > 
> > An object-based rule means that you identify a file/set of files, and
> > you evaluate any process accessing them.
> > 
> > Since processes normally would access the top most layer (overlayfs),
> > the IMA policy should be written in terms of metadata seen in that
> > layer (but not necessarily).
> > 
> > This is just for identifying the set of files to
> > measure/appraise/audit, not which file is going to be evaluated, which
> > will be always the persistent one.
> > 
> > I have to admit, things are not very clear also to me.
> > 
> > Suppose you have a file in the lower filesystem with SELinux label
> > user_t, and then on overlayfs with metadata copy, you change the label
> > of this file to unconfined_t.
> > 
> > What will happen exactly? On the overlayfs layer, you will have a
> > permission request with the new label unconfined_t, but when overlayfs
> > calls vfs_open(), there will be another permission request with the old
> > label.
> 
> CC Vivek who was involved with ovl+selinux, but I think the answer is
> that ovl sepolicy is expected to be associated with the mount ctx and
> not the objects and there was a need to implement the security hook
> security_inode_copy_up() to be able to compose a safe sepolicy for
> overlayfs.
> 
> > 
> > It is kind of the same challenge we are facing with IMA, we can observe
> > the file operations at different layers. That is why I think having
> > stacked IMA calls is a good idea (other than really fixing the
> > violations).
> > 
> > The current problem, that is very difficult to solve, is that the
> > policy should cover all layers, or some events will be missed. Now we
> > have overlayfs-specific code to detect changes in the backing inode,
> > while with stacked IMA calls, we can detect the change at the layer
> > where it happened (and we can remove the overlayfs-specific code).
> > 
> > Ideally, what I would do to cover all layers is that if there is a
> > match at one layer, the lower layers should automatically match too,
> > but it is not that easy since after the vfs_open() recursive calls we
> > start calling IMA in the botton most layer first.
> > 
> > (What I did with the stacked IMA calls is just an experiment to see how
> > far we can go, but still we didn't make any decision with Mimi).
> > 
> > > > AFAIK, IMA/EVM policy is system-wide and not namespace aware
> > > > so the policy has to be set on the container's host and not inside
> > > > containers. Is that correct?
> > 
> > I know that overlayfs is primarily aiming at containers, but I would
> > suggest to not add that complexity yet and just consider the host.
> > 
> > > > If those statements are true then please try to explain to me what is
> > > > the thread model for tampering with overlayfs files, without tampering
> > > > with the real upper and/or lower files.
> > 
> > I hope at this point is clear that what we care about is that, or the
> > process is reading the content of the file whose digest is recorded in
> > the measurement list, or we must signal to remote verifiers concurrent
> > accesses that make the statement above false.
> > 
> > > > My thesis is that if an IMA/EVM policy is good enough to prevent
> > > > tampering with the real lower/upper files, then no extra measures
> > > > are needed to protect the virtual overlayfs files against tampering.
> > 
> > What you say is correct, but the way you identify files to
> > measure/appraise/audit can be different.
> > 
> 
> IIUC, the problem as you described it, is similar to the problem
> of how to display the overlayfs path of mmaped files in /proc/self/maps
> when the actual inode mapped to memory is the real inode.
> 
> The solution for this problem was the somewhat awkward
> "file with fake path" - it was recently changed to use the new
> file_user_path() helper.
> 
> I'm not sure how this can help IMA, but in theory, you could
> always attach iint state to the real inode and only the the real inode
> but the rules that filter by path/object in the IMA hooks that take a file
> argument could take file_user_path() into account.

Sorry, it took some time to elaborate this proposal, but I think it is
mostly correct. I even tried it (as an experiment, we didn't make any
decision yet).

So, the basic idea is to use the stacked IMA calls I proposed (by
adding security_file_post_open() in backing-file.c, and evaluate each
layer independently.

The policy match on each IMA call will be on the overlayfs inode, so
that we see the same metadata as the process opening the file, but we
attach the state to the real inode (regardless of the layer).

Violations are going to work because there is no layer mismatch (write
on one layer, read on a different layer), they are checked always on
the real inode. Since overlayfs has to always open the real inode,
i_write/readcount will be always incremented.

If there is a match at one layer, IMA is not going to process the file
again, since the result is cached from the previous IMA call.

i_version comparison will also be on the real inode, no need to provide
one on overlayfs.

Invalidation (like by setting security.ima) will happen on the real
inode.

I went even further, to think that when we check the authenticity of
security.ima with EVM we can pass the real inode, but we are not sure
yet.

We will continue to validate this proposal.

Thanks

Roberto

> For example, if you have security_file_post_open(f) in
> backing_file_open(), then you can use d_real_inode()
> in process_measurement to get to the IMA state and check the
> rules referring to the real inode and you can use file_user_path(file)
> to check the rules referring to "front object".
> 
> In case of a nested overlayfs, you should get two post_open
> hook, both will refer to the same IMA state on the real inode, but
> each hook with have a different file_user_path(), so you will have
> an opportunity to apply the path/object based rules on every layer
> in the nested overlayfs.
> 
> I hope what I tried to explain is clear and I hope there are not
> many traps down this road, but you won't know unless you try...
> 
> Thanks,
> Amir.

Powered by blists - more mailing lists