Re: BPF hash algo (Re: [kernel-hardening] Re: [PATCH v7 3/6] random: use SipHash in place of MD5)

[Andy Lutomirski <luto@amacapital.net>]

On Fri, Dec 23, 2016 at 3:59 AM, Daniel Borkmann <daniel@iogearbox.net> wrote:
> On 12/23/2016 11:59 AM, Hannes Frederic Sowa wrote:
>>
>> On Fri, 2016-12-23 at 11:04 +0100, Daniel Borkmann wrote:
>>>
>>> On 12/22/2016 05:59 PM, Hannes Frederic Sowa wrote:
>>>>
>>>> On Thu, 2016-12-22 at 08:07 -0800, Andy Lutomirski wrote:
>
> [...]
>
>>>> The hashing is not a proper sha1 neither, unfortunately. I think that
>>>> is why it will have a custom implementation in iproute2?
>>>
>>>
>>> Still trying to catch up on this admittedly bit confusing thread. I
>>> did run automated tests over couple of days comparing the data I got
>>> from fdinfo with the one from af_alg and found no mismatch on the test
>>> cases varying from min to max possible program sizes. In the process
>>> of testing, as you might have seen on netdev, I found couple of other
>>> bugs in bpf code along the way and fixed them up as well. So my question,
>>> do you or Andy or anyone participating in claiming this have any
>>> concrete data or test cases that suggests something different? If yes,
>>> I'm very curious to hear about it and willing fix it up, of course.
>>> When I'm back from pto I'll prep and cook up my test suite to be
>>> included into the selftests/bpf/, should have done this initially,
>>> sorry about that. I'll also post something to expose the alg, that
>>> sounds fine to me.
>>
>>
>> Looking into your code closer, I noticed that you indeed seem to do the
>> finalization of sha-1 by hand by aligning and padding the buffer
>> accordingly and also patching in the necessary payload length.
>>
>> Apologies for my side for claiming that this is not correct sha1
>> output, I was only looking at sha_transform and its implementation and
>> couldn't see the padding and finalization round with embedding the data
>> length in there and hadn't thought of it being done manually.
>>
>> Anyway, is it difficult to get the sha finalization into some common
>> code library? It is not very bpf specific and crypto code reviewers
>> won't find it there at all.
>
>
> Yes, sure, I'll rework it that way (early next year when I'm back if
> that's fine with you).

Can we make it SHA-256 before 4.10 comes out, though?  This really
looks like it will be used in situations where collisions matter and
it will be exposed to malicious programs, and SHA-1 should not be used
for new designs for this purpose because it simply isn't long enough.

Also, a SHA-1 digest isn't a pile of u32s, so u32 digest[...] is very
misleading.  That should be u8 or, at the very least, __be32.

I realize that there isn't a sha-256 implementation in lib, but would
it really be so bad to make the bpf digest only work (for now) when
crypto is enabled?  I would *love* to see the crypto core learn how to
export simple primitives for direct use without needing the whole
crypto core, and this doesn't seem particularly hard to do, but I
don't think that's 4.10 material.

--Andy