Re: [PATCH 0/2] crypto: x86/aes-ni-xts - recover and improve performance

[Eric Biggers]

On Tue, Dec 22, 2020 at 05:06:27PM +0100, Ard Biesheuvel wrote:
> The AES-NI implementation of XTS was impacted significantly by the retpoline
> changes, which is due to the fact that both its asm helper and the chaining
> mode glue library use indirect calls for processing small quantitities of
> data
> 
> So let's fix this, by:
> - creating a minimal, backportable fix that recovers most of the performance,
>   by reducing the number of indirect calls substantially;
> - for future releases, rewrite the XTS implementation completely, and replace
>   the glue helper with a core asm routine that is more flexible, making the C
>   code wrapper much more straight-forward.
> 
> This results in a substantial performance improvement: around ~2x for 1k and
> 4k blocks, and more than 3x for ~1k blocks that require ciphertext stealing
> (benchmarked using tcrypt using 1420 byte blocks - full results below)
> 
> It also allows us to enable the same driver for i386.
> 
> Cc: Megha Dey <megha.dey@xxxxxxxxx>
> Cc: Eric Biggers <ebiggers@xxxxxxxxxx>
> Cc: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
> 
> Ard Biesheuvel (2):
>   crypto: x86/aes-ni-xts - use direct calls to and 4-way stride
>   crypto: x86/aes-ni-xts - rewrite and drop indirections via glue helper
> 
>  arch/x86/crypto/aesni-intel_asm.S  | 353 ++++++++++++++++----
>  arch/x86/crypto/aesni-intel_glue.c | 230 +++++++------
>  2 files changed, 412 insertions(+), 171 deletions(-)
> 
> -- 
> 2.17.1
> 
> Benchmarked using tcrypt on a Intel(R) Core(TM) i7-8650U CPU @ 1.90GHz.

Thanks for doing this!  I didn't realize that there was such a big performance
regression here.  Getting rid of these indirect calls looks like the right
approach; this all seems to have been written for a world where indirect calls
are much faster...

I did some quick benchmarks on Zen ("AMD Ryzen Threadripper 1950X 16-Core
Processor") with CONFIG_RETPOLINE=y and confirmed the speedup on 4096-byte
blocks is around 2x there too.  (It's over 2x for AES-128-XTS and AES-192-XTS,
and a bit under 2x for AES-256-XTS.  And most of the speedup comes from the
first patch.)  Also, the extra self-tests are passing.

So feel free to add:

	Tested-by: Eric Biggers <ebiggers@xxxxxxxxxx> # x86_64

Note that this patch series didn't apply cleanly, as it seems to depend on some
other patches you've sent out recently.  So I actually tested your
"for-kernelci" branch instead of applying these directly.

- Eric