Re: [PATCH 3/3] [CRYPTO] Add optimized SHA-1 implementation for x86_64

[Benjamin Gilbert <bgilbert@cs.cmu.edu>]

Andi Kleen wrote:
> Benjamin Gilbert <bgilbert@cs.cmu.edu> writes:
>> +#define EXPAND(i)						\
>> +	movl	OFFSET(i % 16)(DATA), TMP;			\
>> +	xorl	OFFSET((i + 2) % 16)(DATA), TMP;		\
> 
> Such overlapping memory accesses are somewhat dangerous as they tend
> to stall some CPUs.  Better probably to do a quad load and then extract.

OFFSET(i) is defined as 4*(i), so they don't actually overlap. 
(Arguably that macro should go away.)

> I haven't checked in detail if it's possible but it's suspicious you
> never use quad operations for anything. You keep at least half
> the CPU's bits idle all the time.

SHA-1 fundamentally wants to work with 32-bit quantities.  It might be 
possible to use quad operations for some things, with sufficient 
cleverness, but I doubt it'd be worth the effort.

> Gut feeling is that the unroll factor is far too large.
> Have you tried a smaller one? That would save icache
> which is very important in the kernel.

That seems to be the consensus.  I'll see if I can find some time to try 
linux@horizon.com's suggestion and report back.

I don't think, though, that cache footprint is the *only* thing that 
matters.  Leaving aside /dev/urandom, there are cases where throughput 
matters a lot.  This patch set came out of some work on a hashing block 
device driver in which SHA is, by far, the biggest CPU user.  One could 
imagine content-addressable filesystems, or even IPsec under the right 
workloads, being in a similar situation.

Would it be more palatable to roll the patch as an optimized CryptoAPI 
module rather than as a lib/sha1.c replacement?  That wouldn't help 
/dev/urandom, of course, but for other cases it would allow the user to 
ask for the optimized version if needed, and not pay the footprint costs 
otherwise.

--Benjamin Gilbert