Re: [RFC PATCH v2 11/12] crypto: adiantum - add Adiantum support

[Eric Biggers <ebiggers@kernel.org>]

On Tue, Oct 23, 2018 at 07:40:34AM -0300, Ard Biesheuvel wrote:
> On 20 October 2018 at 15:12, Eric Biggers <ebiggers@kernel.org> wrote:
> > Hi Ard,
> >
> > On Sat, Oct 20, 2018 at 12:17:58PM +0800, Ard Biesheuvel wrote:
> >> On 16 October 2018 at 01:54, Eric Biggers <ebiggers@kernel.org> wrote:
> >> > From: Eric Biggers <ebiggers@google.com>
> >> >
> >> > Add support for the Adiantum encryption mode.  Adiantum was designed by
> >> > Paul Crowley and is specified by our paper:
> >> >
> >> >     Adiantum: length-preserving encryption for entry-level processors
> >> >     (https://eprint.iacr.org/2018/720.pdf)
> >> >
> >> > See our paper for full details; this patch only provides an overview.
> >> >
> >> > Adiantum is a tweakable, length-preserving encryption mode designed for
> >> > fast and secure disk encryption, especially on CPUs without dedicated
> >> > crypto instructions.  Adiantum encrypts each sector using the XChaCha12
> >> > stream cipher, two passes of an ε-almost-∆-universal (εA∆U) hash
> >> > function, and an invocation of the AES-256 block cipher on a single
> >> > 16-byte block.  On CPUs without AES instructions, Adiantum is much
> >> > faster than AES-XTS; for example, on ARM Cortex-A7, on 4096-byte sectors
> >> > Adiantum encryption is about 4 times faster than AES-256-XTS encryption,
> >> > and decryption about 5 times faster.
> >> >
> >> > Adiantum is a specialization of the more general HBSH construction.  Our
> >> > earlier proposal, HPolyC, was also a HBSH specialization, but it used a
> >> > different εA∆U hash function, one based on Poly1305 only.  Adiantum's
> >> > εA∆U hash function, which is based primarily on the "NH" hash function
> >> > like that used in UMAC (RFC4418), is about twice as fast as HPolyC's;
> >> > consequently, Adiantum is about 20% faster than HPolyC.
> >> >
> >> > This speed comes with no loss of security: Adiantum is provably just as
> >> > secure as HPolyC, in fact slightly *more* secure.  Like HPolyC,
> >> > Adiantum's security is reducible to that of XChaCha12 and AES-256,
> >> > subject to a security bound.  XChaCha12 itself has a security reduction
> >> > to ChaCha12.  Therefore, one need not "trust" Adiantum; one need only
> >> > trust ChaCha12 and AES-256.  Note that the εA∆U hash function is only
> >> > used for its proven combinatorical properties so cannot be "broken".
> >> >
> >>
> >> So what happens if the part of the input covered by the block cipher
> >> is identical between different generations of the same disk block
> >> (whose sector count is used as the 'outer' IV). How are we not in the
> >> same boat as before when using stream ciphers for disk encryption?
> >>
> >
> > This is the point of the hash step.  The value encrypted with the block cipher
> > to produce the intermediate value C_M (used as the stream cipher nonce) is
> > H(T, P_L) + P_R.  (T is the tweak a.k.a the IV, P_L is the plaintext except the
> > last 16 bytes, P_R is the last 16 bytes.)  A collision in this value occurs iff:
> >
> >         H(T1, P1_L) + P1_R = H(T2, P2_L) + P2_R
> > i.e.
> >         H(T1, P1_L) - H(T2, P2_L) = P2_R - P1_R
> >
> > If (T1, P1_L) = (T2, P2_L) then P1_R != P2_R so the equation has no solutions
> > (since we don't consider queries where the whole input is the same; those
> > unavoidably produce the same ciphertext).  Otherwise (T1, P1_L) != (T2, P2_L),
> > and since the hash function H is ε-almost-∆-universal over integers mod 2^128,
> > the equation is true for at most a very small proportion 'ε' of hash keys.
> > But, the hash key is chosen at random and is unknown to the attacker.
> >
> > The same applies in the other direction, for chosen ciphertext attacks.
> >
> > Basically, it's very difficult for an attacker to cause the intermediate value
> > C_M to be reused, and the outputs will appear random until they do.
> >
> > Of course, all this is explained much more precisely and comprehensively in our
> > paper.  See section 5, "Security reduction".
> >
> 
> Thanks for the explanation. I saw that the result of the AES
> encryption was used as the XChaCha nonce, but I failed to spot that
> the result of the nhpoly1305 pass is added/subtracted to/from that
> particular block first.
> 
> In any case, this looks good to me: as far as I can tell, the code
> implements the algorithm as described in the paper, and the plumbing
> into the crypto API looks correct to me as well.
> 
> Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
> 
> Whether the paper is correct is a different matter: it looks
> convincing to me but IANAC.
> 
> The only request I have is to add a speed test to tcrypt as well so we
> can easily benchmark it.

I'll add an Adiantum testing mode to tcrypt, though I have to admit that I
really dislike tcrypt, so I've actually been using a custom patch instead
(https://git.kernel.org/pub/scm/linux/kernel/git/ebiggers/linux.git/commit/?h=cryptobench).

FWIW, I mostly dislike tcrypt's lack of flexibility.  Every benchmark setting
has to be coded explicitly in tcrypt, with a mode number to select it.  It's
often missing what I want, or what I want is included but comes with a bunch of
unwanted tests too.  Also tcrypt has to be built as a loadable module, so it
can't be used with CONFIG_MODULES=n.  And the benchmark results only go directly
to the kernel log, where they aren't easily retrievable from a script.

The patch I've been using just exposes a file /proc/cryptobench (note: it maybe
should be made into a char device instead) to which you can write commands like:

	algtype=skcipher algname=adiantum(xchacha12,aes) keysize=32 bufsize=4096 niter=1000

Then userspace reads back the result:

	SUCCESS algname=adiantum(xchacha12,aes) driver_name=adiantum(xchacha12-software,aes-aesni,nhpoly1305-generic) measurement=0x30a5abd5776e0af enc_time=44831104 dec_time=38303077

It's then pretty straightforward to wrap this API in a userspace script that
does all the benchmarks you need, and formats the results nicely.

For correctness testing there's also an option 'sgl_fuzz' that randomizes the
scatterlist division.

Currently the patch is missing some features such as AEAD support, but at some
point I'd like to get it into a state where it can be included upstream.

- Eric