RE: [PATCH 2/3] crypto: inside-secure - Reduce stack usage

[Pascal Van Leeuwen]

> -----Original Message-----
> From: Arnd Bergmann <arnd@xxxxxxxx>
> Sent: Monday, September 30, 2019 10:12 PM
> To: Pascal Van Leeuwen <pvanleeuwen@xxxxxxxxxxxxxx>
> Cc: Antoine Tenart <antoine.tenart@xxxxxxxxxxx>; Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>;
> David S. Miller <davem@xxxxxxxxxxxxx>; Pascal van Leeuwen <pascalvanl@xxxxxxxxx>; Ard
> Biesheuvel <ard.biesheuvel@xxxxxxxxxx>; Eric Biggers <ebiggers@xxxxxxxxxx>; linux-
> crypto@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH 2/3] crypto: inside-secure - Reduce stack usage
> 
> On Mon, Sep 30, 2019 at 9:04 PM Pascal Van Leeuwen
> <pvanleeuwen@xxxxxxxxxxxxxx> wrote:
> 
> > > Alternatively, it should be possible to shrink these allocations
> > > as the extra buffers appear to be largely unnecessary, but doing
> > > this would be a much more invasive change.
> > >
> > Actually, for HMAC-SHA512 you DO need all that buffer space.
> > You could shrink it to 2 * ctx->state_sz but then your simple indexing
> > is no longer going to fly. Not sure if that would be worth the effort.
> 
> Stack allocations can no longer be dynamically sized in the kernel,
> so that would not work.
> 
I was actually referring to your kzalloc, not to the original stack
based implementation ...

> What I noticed though is that the largest part of safexcel_ahash_export_state
> is used in the 'cache' member, and this is apparently only referenced inside of
> safexcel_hmac_init_iv, which you call twice. If that cache can be allocated
> only once, you save SHA512_BLOCK_SIZE bytes in one of the two paths.
> 
Well ... hmmm ... "my"... This is not originally "my" code. And until you
brought this up, I did not fully realise it was using this export_state
struct to store those digests. Seems to have something to do with directly
taking the crypto_ahash_export state output, which is much more than that,
in case you need to continue the hash (which you don't here).

I guess you need to "catch" that output somewhere, so probably you could
save a bit of memory but I doubt it would be a significant improvement.

> > I don't like the part where you dynamically allocate the cryto_aes_ctx
> > though, I think that was not necessary considering its a lot smaller.
> 
> I count 484 bytes for it, which is really large.
> 
Maybe I should've counted myself ... totally unexpectedly huge. Why??
Anyway, glad I got rid of it already then.

> > And it conflicts with another change I have waiting that gets rid of
> > aes_expandkey and that struct alltogether (since it was really just
> > abused to do a key size check, which was very wasteful since the
> > function actually generates all roundkeys we don't need at all ...)
> 
> Right, this is what I noticed there. With 480 of the 484 bytes gone,
> you are well below the warning limit even without the other change.
> 
And by "other change" you mean the safexcel_ahash_export_state?
Ok, good to known, although I do like to improve that one as well,
but preferably by not exporting the cache so I don't need the full
struct.

Regards,
Pascal van Leeuwen
Silicon IP Architect, Multi-Protocol Engines @ Verimatrix
www.insidesecure.com