Message-ID: <154208f2a248482e8664903f734cf0b4@huawei.com>
Date:   Wed, 27 Oct 2021 09:34:48 +0000
From:   Roberto Sassu <roberto.sassu@...wei.com>
To:     Deven Bowers <deven.desai@...ux.microsoft.com>,
        Eric Biggers <ebiggers@...nel.org>
CC:     "corbet@....net" <corbet@....net>,
        "axboe@...nel.dk" <axboe@...nel.dk>,
        "agk@...hat.com" <agk@...hat.com>,
        "snitzer@...hat.com" <snitzer@...hat.com>,
        "tytso@....edu" <tytso@....edu>,
        "paul@...l-moore.com" <paul@...l-moore.com>,
        "eparis@...hat.com" <eparis@...hat.com>,
        "jmorris@...ei.org" <jmorris@...ei.org>,
        "serge@...lyn.com" <serge@...lyn.com>,
        "jannh@...gle.com" <jannh@...gle.com>,
        "dm-devel@...hat.com" <dm-devel@...hat.com>,
        "linux-doc@...r.kernel.org" <linux-doc@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "linux-block@...r.kernel.org" <linux-block@...r.kernel.org>,
        "linux-fscrypt@...r.kernel.org" <linux-fscrypt@...r.kernel.org>,
        "linux-audit@...hat.com" <linux-audit@...hat.com>,
        "linux-security-module@...r.kernel.org" 
        <linux-security-module@...r.kernel.org>
Subject: RE: [RFC PATCH v7 12/16] fsverity|security: add security hooks to
 fsverity digest and signature

> From: Deven Bowers [mailto:deven.desai@...ux.microsoft.com]
> Sent: Tuesday, October 26, 2021 9:04 PM
> On 10/15/2021 1:11 PM, Eric Biggers wrote:
> 
> > On Fri, Oct 15, 2021 at 12:25:53PM -0700, Deven Bowers wrote:
> >> On 10/13/2021 12:24 PM, Eric Biggers wrote:
> >>> On Wed, Oct 13, 2021 at 12:06:31PM -
> 0700,deven.desai@...ux.microsoft.com  wrote:
> >>>> From: Fan Wu<wufan@...ux.microsoft.com>
> >>>>
> >>>> Add security_inode_setsecurity to fsverity signature verification.
> >>>> This can let LSMs save the signature data and digest hashes provided
> >>>> by fsverity.
> >>> Can you elaborate on why LSMs need this information?
> >> The proposed LSM (IPE) of this series will be the only one to need
> >> this information at the  moment. IPE’s goal is to have provide
> >> trust-based access control. Trust and Integrity are tied together,
> >> as you cannot prove trust without proving integrity.
> > I think you mean authenticity, not integrity?
> I’ve heard a lot of people use these terms in overloaded ways.
> 
> If we’re working with the definition of authenticity being
> “the property that a resource was _actually_ sent/created by a
> party”, and integrity being “the property that a resource was not
> modified from a point of time”, then yes. Though the statement isn’t
> false, though, because you’d need to prove integrity in the process of
> proving authenticity.
> 
> If not, could you clarify what you mean by authenticity and integrity,
> so that we can use consistent definitions?
> > Also how does this differ from IMA?  I know that IMA doesn't support fs-verity
> > file hashes, but that could be changed.  Why not extend IMA to cover your use
> > case(s)?
> We looked at extending IMA to cover our requirements extensively the
> past year
> based on feedback the last time I posted these patches. We implemented a
> prototype that had half of our requirements, but found it resulted in a
> large change list that would result in a large amount of pain in respect
> to maintenance, in addition to other more architectural concerns about the
> implementation. We weren’t convinced it was the correct direction, for our
> needs.
> 
> There was a presentation done at LSS 2021 around this prototype done by my
> colleague, Fan, who authored this patch and implemented the aforementioned
> prototype.
> 
> In general, IMA provides a whole suite of amazing functionality when it
> comes to everything integrity, as the fs-verity documentation states
> itself:
> 
>     IMA specifies a system-wide policy that specifies which
>     files are hashed and what to do with those hashes, such
>     as log them, authenticate them, or add them to a
>     measurement list.
> 
> Instead, IPE provides a fine-tuned way to _only_ enforce an access control
> policy to these files based on the defined trust requirements in the policy,
> under various contexts, (you might have different requirements for what
> executes in a general purpose, versus loadable kernel modules, for example).
> It will never provide bother to log, measure, or revalidate these hashes
> because
> that’s not its purpose. This is why it belongs at the LSM layer instead
> of the
> integrity subsystem layer, as it is providing access control based on a
> policy,
> versus providing deep integrations with the actual integrity claim.
> 
> IPE is trying to be agnostic to how precisely “trust” is provided, as
> opposed to be deeply integrated into the mechanism that provides
> “trust”.
> >> IPE needs the digest information to be able to compare a digest
> >> provided by the policy author, against the digest calculated by
> >> fsverity to make a decision on whether that specific file, represented
> >> by the digest is authorized for the actions specified in the policy.
> >>
> >> A more concrete example, if an IPE policy author writes:
> >>
> >>      op=EXECUTE fsverity_digest=<HexDigest > action=DENY
> >>
> >> IPE takes the digest provided by this security hook, stores it
> >> in IPE's security blob on the inode. If this file is later
> >> executed, IPE compares the digest stored in the LSM blob,
> >> provided by this hook, against <HexDigest> in the policy, if
> >> it matches, it denies the access, performing a revocation
> >> of that file.
> > Do you have a better example?  This one is pretty useless since one can get
> > around it just by executing a file that doesn't have fs-verity enabled.
> Here’s a more complete example:
> 
>     policy_name=”fs-exec-only” policy_version=0.0.1
>     DEFAULT action=ALLOW
> 
>     DEFAULT op=EXECUTE action=DENY
>     op=EXECUTE fsverity_digest=<Digest> action=DENY
>     op=EXECUTE fsverity_signature=TRUE action=ALLOW
> 
> Execution is prohibited unless it is a signed fs-verity file;
> However, after one of those executables was signed and published,
> an exploitable vulnerability in said executable was found, a new
> version was published without that vulnerability. We need to
> revoke trust for that executable since it could be used to exploit
> the system, so the first rule prevents it from matching the second.

With DIGLIM, revocation will be completely transparent since
digests of files in the packages being removed will be also removed
from the kernel. The next time IPE will evaluate an old version of the
file (assuming that there was a copy somewhere) it will realize that
the digest is not known anymore and will deny access.

Roberto

HUAWEI TECHNOLOGIES Duesseldorf GmbH, HRB 56063
Managing Director: Li Peng, Zhong Ronghua

> >> This brings me to your next comment:
> >>
> >>> The digest isn't meaningful without knowing the hash algorithm it uses.
> >> It's available here, but you aren't passing it to this function.
> >>
> >> The digest is meaningful without the algorithm in this case.
> > No, it's not.
> >
> > Digests are meaningless without knowing what algorithm they were created
> with.
> >
> > If your security policy is something like "Trust the file with digest $foo" and
> > multiple hash algorithms are possible, then the alorithm intended to be used
> > needs to be explicitly specified.  Otherwise any algorithm with the same length
> > digest will be accepted.  That's a fatal flaw if any of these algorithms is
> > cryptographically broken or was never intended to be a cryptographic
> algorithm
> > in the first place (e.g., a non-cryptographic checksum).
> >
> > Cryptosystems always need to specify the crypto algorithm(s) used; the
> adversary
> > must not be allowed to choose the algorithms.
> Oof. You’re completely right. The part I was missing is that as time
> goes on,
> the secure status of these cryptographic algorithms will change, and
> then we’ll
> need a way to migrate between algorithms. Additionally, tooling and the
> like will
> likely need a way to identify this from the policy text without
> consulting anything
> else. This is a major oversight for general use, the system that this
> was originally
> designed for only had support for a subset of the sha2-family (all
> separate lengths)
> so I hadn’t even considered it.
> 
> It's trivial to correct in a minimal amount of code, making the policy
> express the
> digest like so:
> 
>     fsverity_digest=<algo>:<digest>
> 
> and change the argument passed to the LSM hook to accept a structure
> containing these
> two fields.
> 
> > I'm not sure how these patches can be taken seriously when they're getting
> this
> > sort of thing wrong.
> That said, I, personally, hope that an honest mistake, in a series
> submitted as
> an RFC submitted in good faith, is not a reason to discount an entire patch
> series.
> 
> I hope you continue to provide feedback, as it is invaluable to making this
> system better, and making me, personally, a better developer.
> >>>> +					FS_VERITY_SIGNATURE_SEC_NAME,
> >>>> +					signature, sig_size, 0);
> >>> This is only for fs-verity built-in signatures which aren't the only way to do
> >>> signatures with fs-verity.  Are you sure this is what you're looking for?
> >> Could you elaborate on the other signature types that can be used
> >> with fs-verity? I’m 99% sure this is what I’m looking for as this
> >> is a signature validated in the kernel against the fs-verity keyring
> >> as part of the “fsverity enable” utility.
> >>
> >> It's important that the signature is validated in the kernel, as
> >> userspace is considered untrusted until the signature is validated
> >> for this case.
> >>
> >>> Can you elaborate on your use case for fs-verity built-in signatures,
> >> Sure, signatures, like digests, also provide a way to prove integrity,
> >> and the trust component comes from the validation against the keyring,
> >> as opposed to a fixed value in IPE’s policy. The use case for fs-verity
> >> built-in signatures is that we have a rw ext4 filesystem that has some
> >> executable files, and we want to have a execution policy (through IPE)
> >> that only _trusted_ executables can run. Perf is important here, hence
> >> fs-verity.
> > Most users of fs-verity built-in signatures have actually been enforcing their
> > security policy in userspace, by checking whether specific files have the
> > fs-verity bit set or not.  Such users could just store and verify signatures in
> > userspace instead, without any kernel involvement.  So that's what I've been
> > recommending (with limited success, unfortunately).
> I believe the difference in security models comes from this line
> (emphasis, mine):
> 
>  > by checking whether _specific files_ have the fs-verity bit set or not.
> 
> IPE policy is written by a system author who owns the system, but may
> not have 100% control over all of the application code running on the
> system.  In the case of applications which are not aware of IPE, the policy
> can still enforce that all of the code running on the system is trusted.
> 
> An example attack of what we're trying to mitigate:  A hostile actor
> could downloads a binary off the internet with all required
> dependencies into tmpfs and runs their malicious executable.
> 
> With us validating this information in the kernel, even if the attacker
> downloaded their malicious executable to /tmp and executed it, it would
> still fail to pass policy and be denied, as the kernel is the common
> entrypoint across all executables.
> 
> Operationally, this _could_ be done by digest, but the policies would
> quickly become gigantic on a cartoonish proportion, as you'll have to
> authorize every single executable and dependency by digest - and
> there would be a complicated update story as the policy would have to
> be updated to onboard new digests.
> 
> By using signatures, we can prevent the policy update, and keep the
> policy size small.
> 
> > If you really do need in-kernel signature verification, then that may be a
> > legitimate use case for the fs-verity built-in signatures, although I do wonder
> > why you aren't using IMA and its signature mechanism instead.
> >
> > - Eric

Powered by blists - more mailing lists