Re: [PATCH] crypto: vmx - CTR: always increment IV as quadword

[Eric Biggers <ebiggers@kernel.org>]

On Mon, May 20, 2019 at 11:59:05AM +1000, Daniel Axtens wrote:
> Daniel Axtens <dja@axtens.net> writes:
> 
> > The kernel self-tests picked up an issue with CTR mode:
> > alg: skcipher: p8_aes_ctr encryption test failed (wrong result) on test vector 3, cfg="uneven misaligned splits, may sleep"
> >
> > Test vector 3 has an IV of FFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFD, so
> > after 3 increments it should wrap around to 0.
> >
> > In the aesp8-ppc code from OpenSSL, there are two paths that
> > increment IVs: the bulk (8 at a time) path, and the individual
> > path which is used when there are fewer than 8 AES blocks to
> > process.
> >
> > In the bulk path, the IV is incremented with vadduqm: "Vector
> > Add Unsigned Quadword Modulo", which does 128-bit addition.
> >
> > In the individual path, however, the IV is incremented with
> > vadduwm: "Vector Add Unsigned Word Modulo", which instead
> > does 4 32-bit additions. Thus the IV would instead become
> > FFFFFFFFFFFFFFFFFFFFFFFF00000000, throwing off the result.
> >
> > Use vadduqm.
> >
> > This was probably a typo originally, what with q and w being
> > adjacent. It is a pretty narrow edge case: I am really
> > impressed by the quality of the kernel self-tests!
> >
> > Fixes: 5c380d623ed3 ("crypto: vmx - Add support for VMS instructions by ASM")
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Daniel Axtens <dja@axtens.net>
> >
> > ---
> >
> > I'll pass this along internally to get it into OpenSSL as well.
> 
> I passed this along to OpenSSL and got pretty comprehensively schooled:
> https://github.com/openssl/openssl/pull/8942
> 
> It seems we tweak the openssl code to use a 128-bit counter, whereas
> the original code was in fact designed for a 32-bit counter. We must
> have changed the vaddu instruction in the bulk path but not in the
> individual path, as they're both vadduwm (4x32-bit) upstream.
> 
> I think this change is still correct with regards to the kernel,
> but I guess it's probably something where I should have done a more
> thorough read of the documentation before diving in to the code, and
> perhaps we should note it in the code somewhere too. Ah well.
> 
> Regards,
> Daniel
> 

Ah, I didn't realize there are multiple conventions for CTR.  Yes, all CTR
implementations in the kernel follow the 128-bit convention, and evidently the
test vectors test for that.  Apparently the VMX OpenSSL code got incompletely
changed from the 32-bit convention by this commit, so that's what you're fixing:

	commit 1d4aa0b4c1816e8ca92a6aadb0d8f6b43c56c0d0
	Author: Leonidas Da Silva Barbosa <leosilva@linux.vnet.ibm.com>
	Date:   Fri Aug 14 10:12:22 2015 -0300

	    crypto: vmx - Fixing AES-CTR counter bug
	    
	    AES-CTR is using a counter 8bytes-8bytes what miss match with
	    kernel specs.
	    
	    In the previous code a vadduwm was done to increment counter.
	    Replacing this for a vadduqm now considering both cases counter
	    8-8 bytes and full 16bytes.

A comment in the code would indeed be helpful.

Note that the kernel doesn't currently need a 32-bit CTR implementation for GCM
like OpenSSL does, because the kernel currently only supports 12-byte IVs with
GCM.  So the low 32 bits of the counter start at 1 and don't overflow,
regardless of whether the counter is incremented mod 2^32 or mod 2^128.

- Eric