Re: [RFC PATCH crypto 4/4] AES-NI: Add support to Intel AES-NIinstructions for x86_64 platform

[Huang Ying]

On Fri, 2009-01-09 at 15:01 +0800, Herbert Xu wrote:
> > - ablkcipher asynchronous machanism is used to delay a crypto request
> >   to work queue context upon FPU state is using by other kernel
> >   context.
> 
> Actually I was thinking of something as simple as just using
> cryptd as is.  That is, register a blkcipher algorithm for your
> version of cbc-aes, but don't register under the name cbc(aes),
> e.g., register it under the name of __cbc-aes-aesni for both alg
> name and driver name.  You can register it under cbc(aes) too
> if you add the fallback in there.
> 
> Then for the real thing, just allocate the blkcipher __cbc_*,
> plus the ablkcipher cryptd(__cbc_*), and invoke the right one
> depending on caller context.

I have ever considered this method too. This one is simpler, but the
drawbacks are as follow:

- cryptd thread is not per-CPU, so I think there will be some
unnecessary cache inter-CPU migration. Why not use a dedicate workqueue
or just system events workqueue?

- with cryptd(__*-aes-aesni), we need 4 internal tfms for each external
tfm allocation request. For example, for one external cbc(aes) tfm
allocation request, we need one cbc(aes) ablkcipher tfm, one
cryptd(cbc-aes-aesni) tfm, and two cbc-aes-aesni tfm. Do we use too much
memory? And we need to call aesni_set_key() twice.

> This is even simpler for the modes which you don't implement.
> In that case you would just allocate FOO(aes-aesni) in conjunction
> with cryptd(FOO(aes-aesni)).

Yes. This is really a simple method.

Best Regards,
Huang Ying

Attachment: signature.asc
Description: This is a digitally signed message part