Re: [RFC PATCH v2 2/2] dm-inlinecrypt: add target for inline block device encryption

[Milan Broz <gmazyland@gmail.com>]

On 10/17/24 10:28 PM, Eric Biggers wrote:
> On Thu, Oct 17, 2024 at 10:17:04PM +0200, Milan Broz wrote:
>> On 10/17/24 9:44 PM, Eric Biggers wrote:
>>> On Wed, Oct 16, 2024 at 04:27:48PM -0700, Eric Biggers wrote:
>>>> Add a new device-mapper target "dm-inlinecrypt" that is similar to
>>>> dm-crypt but uses the blk-crypto API instead of the regular crypto API.
>>>> This allows it to take advantage of inline encryption hardware such as
>>>> that commonly built into UFS host controllers.
>>>
>>> A slight difference in behavior vs. dm-crypt that I just became aware of:
>>> dm-crypt allows XTS keys whose first half equals the second half, i.e.
>>> cipher key == tweak key.  dm-inlinecrypt typically will not allow this.  Inline
>>> encryption hardware typically rejects such keys, and blk-crypto-fallback rejects
>>> them too because it uses CRYPTO_TFM_REQ_FORBID_WEAK_KEYS.
>>>
>>> IMO, rejecting these weak keys is desirable, and the fact that dm-inlinecrypt
>>> fixes this issue with dm-crypt will just need to be documented.
>>
>> Hm, I thought this is already rejected in crypto API (at least in FIPS mode)...
>>
>> It should be rejected exactly as you described even for dm-crypt,
>> just the check should be (IMO) part of crypto API (set keys), not dm-crypt itself.
>>
>> And here I think we should not be backward "compatible" as it is security issue,
>> both XTS keys just must not be the same.
>>
> 
> In "FIPS mode" such keys are always rejected, but otherwise it is opt-in via the
> flag CRYPTO_TFM_REQ_FORBID_WEAK_KEYS.  dm-crypt doesn't use that flag.
> 
> We could certainly try to fix that in dm-crypt, though I expect that some
> dm-crypt users have started relying on such keys.  It is a common misconception
> that XTS is secure when the two halves of the key are the same.

Ah, ok, I missed that weak keys flag.

We never did that in cryptsetup (including LUKS and plain with hashed passwords),
with the exception for benchmark (where it it was not a real key, just all zeroes,
and was fixed years ago as it did not work in FIPS) -- and obviously the case when
user set the key explicitly.

The same check is now in VeraCrypt (that uses dm-crypt on Linux).
I know about several broken HW crypto, but actually no dm-crypt user.

IMO we should set CRYPTO_TFM_REQ_FORBID_WEAK_KEYS by default. We can always introduce
flag to disable it, but I would really like to know if there are any real users....

Milan