Re: [PATCH v1 4/5] ima: support fs-verity file digest based signatures

From: Stefan Berger
Date: Sat Jan 15 2022 - 22:32:30 EST


On 1/15/22 01:21, Eric Biggers wrote:
On Sat, Jan 15, 2022 at 08:31:01AM +0300, Vitaly Chikunov wrote:
Eric,

On Sun, Jan 09, 2022 at 01:07:18PM -0800, Eric Biggers wrote:
On Sun, Jan 09, 2022 at 11:45:37PM +0300, Vitaly Chikunov wrote:
On Wed, Jan 05, 2022 at 03:37:39PM -0800, Eric Biggers wrote:
On Fri, Dec 31, 2021 at 10:35:00AM -0500, Mimi Zohar wrote:
On Thu, 2021-12-02 at 14:07 -0800, Eric Biggers wrote:
On Thu, Dec 02, 2021 at 04:55:06PM -0500, Mimi Zohar wrote:
case IMA_VERITY_DIGSIG:
- fallthrough;
+ set_bit(IMA_DIGSIG, &iint->atomic_flags);
+
+ /*
+ * The IMA signature is based on a hash of IMA_VERITY_DIGSIG
+ * and the fs-verity file digest, not directly on the
+ * fs-verity file digest. Both digests should probably be
+ * included in the IMA measurement list, but for now this
+ * digest is only used for verifying the IMA signature.
+ */
+ verity_digest[0] = IMA_VERITY_DIGSIG;
+ memcpy(verity_digest + 1, iint->ima_hash->digest,
+ iint->ima_hash->length);
+
+ hash.hdr.algo = iint->ima_hash->algo;
+ hash.hdr.length = iint->ima_hash->length;
This is still wrong because the bytes being signed don't include the hash
algorithm. Unless you mean for it to be implicitly always SHA-256? fs-verity
supports SHA-512 too, and it may support other hash algorithms in the future.
IMA assumes that the file hash algorithm and the signature algorithm
are the same. If they're not the same, for whatever reason, the
signature verification would simply fail.

Based on the v2 signature header 'type' field, IMA can differentiate
between regular IMA file hash based signatures and fs-verity file
digest based signatures. The digest field (d-ng) in the IMA
meausrement list prefixes the digest with the hash algorithm. I'm
missing the reason for needing to hash fs-verity's file digest with
other metadata, and sign that hash rather than fs-verity's file digest
directly.
Because if someone signs a raw hash, then they also implicitly sign the same
hash value for all supported hash algorithms that produce the same length hash.
Unless there is broken hash algorithm allowing for preimage attacks this
is irrelevant. If there is two broken algorithms allowing for collisions,
colliding hashes could be prepared even if algo id is hashed too.

Only one algorithm needs to be broken. For example, SM3 has the same hash
length as SHA-256. If SM3 support were to be added to fs-verity, and if someone
were to find a way to find an input that has a specific SM3 digest, then they
could also make it match a specific SHA-256 digest. Someone might intend to
sign a SHA-256 digest, but if they are only signing the raw 32 bytes of the
digest, then they would also be signing the corresponding SM3 digest. That's
why the digest that is signed *must* also include the algorithm used in the
digest (not the algorithm(s) used in the signature, which is different).
I think it will be beneficial if we pass hash algo id to the
akcipher_alg::verify. In fact, ecrdsa should only be used with streebog.
And perhaps, sm2 with sm3, pkcs1 with md/sha/sm3, and ecdsa with sha family
hashes.

I was going to reply to this thread again, but I got a bit distracted by
everything else being broken. Yes, the kernel needs to be restricting which
hash algorithms can be used with each public key algorithm, along the lines of
what you said. I asked the BoringSSL maintainers for advice, and they confirmed
that ECDSA just signs/verifies a raw hash, and in fact it *must* be a raw hash
for it to be secure. This is a design flaw in ECDSA, which was fixed in newer
algorithms such as EdDSA and SM2 as those have a hash built-in to the signature
scheme. To mitigate it, the allowed hash algorithms must be restricted; in the
case of ECDSA, that means to the SHA family (preferably excluding SHA-1).

akcipher_alg::verify doesn't actually know which hash algorithm is used, except
in the case of rsa-pkcs1pad where it is built into the name of the algorithm.
So it can't check the hash algorithm. I believe it needs to happen in
public_key_verify_signature() (and I'm working on a patch for that).

Now, SM2 is different from ECDSA and ECRDSA in that it uses the modern design
that includes the hash into the signature algorithm. This means that it must be
used to sign/verify *data*, not a hash. (Well, you can sign/verify a hash, but
SM2 will hash it again internally.) Currently, public_key_verify_signature()
allows SM2 to be used to sign/verify a hash, skipping the SM2 internal hash, and
IMA uses this. This is broken and must be removed, since it isn't actually the
SM2 algorithm as specified anymore, but rather some homebrew thing with unknown
security properties. (Well, I'm not confident about SM2, but homebrew is worse.)

Adding fs-verity support to IMA also complicates things, as doing it naively
would introduce an ambiguity about what is signed. Naively, the *data* that is
signed (considering the hash as part of the signature algorithm) would be either
the whole file, in the case of traditional IMA, or the fsverity_descriptor
struct, in the case of IMA with fs-verity. However, a file could have contents
which match an fsverity_descriptor struct; that would create an ambiguity.

Assuming that it needs to be allowed that the same key can sign files for both
traditional and fs-verity hashing, solving this problem will require a second
hash. The easiest way to do this would be sign/verify the following struct:

struct ima_file_id {
u8 is_fsverity;
u8 hash_algorithm;
u8 hash[];
};


To calrify, I suppose that for ECDSA NIST P256 you would allow pairing with any of the SHA family hashes (also as defined by the existing OIDs) and as the standard allows today? And the same then applies for NIST p384 etc.?

Further, I suppose similar restriction would apply for ECRDSA to pair it with Streebog only, as Vitaly said.


What's happening now is that to verify a signature, IMA/integrity subsystem fills out the following structure:

struct public_key_signature pks;

pks.hash_algo = hash_algo_name[hdr->hash_algo];  // name of hash algo will go into this here, e.g., 'sha256'
pks.pkey_algo = pk->pkey_algo; // this is either 'rsa', 'ecdsa-', 'ecrdsa-' or 'sm2' string

It then calls:

    ret = verify_signature(key, &pks);

IMO, in the call path down this function the pairing of public key and hash algo would have to be enforced in order to enforce the standards. Would this not be sufficient to be able to stay with the standards ?

File hashes: IMA calculates the hash over a file itself by calling crypto functions, so at least the digest's bytes are trusted input in that respect and using the sha family type of hashes directly with ECDSA should work. Which algorithm IMA is supposed to use for the hashing is given in the xattr bytestream header. IMA could then take that type of hash, lookup the hash function, perform the hashing on the data, and let verify_signature enforce the pairing, rejecting file signatures with wrong pairing. This way the only thing that is needed is 'enforcement of pairing'.

Fsverity: How much control does a user have over the hash family fsverity is using? Can IMA ECDSA/RSA users tell it to use a sha family hash and ECRDSA users make it use a Streebog hash so that also the pairing of hash and key type can work 'naturally' and we don't need the level of indirection via your structure above?

    Stefan



This would be the *data* that is signed/verified -- meaning that it would be
hashed again as part of the signature algorithm (whether that hash is built-in
to the signature algorithm, as is the case for modern algorithms, or handled by
the caller as is the case for legacy algorithms). Note that both traditional
and fs-verity hashes would need to use this same method for it to be secure; the
kernel must not accept signatures using the old method at the same time.

- Eric