[dm-devel] [PATCH] DM-CRYPT: Scale to multiple CPUs

[huang ying]

On Tue, Jun 1, 2010 at 6:07 AM, Herbert Xu <herbert at gondor.apana.org.au> wrote:
> On Mon, May 31, 2010 at 09:04:00PM +0200, Andi Kleen wrote:
>> > I mean how it is implemented now in crypto API, and I was almost
>> > sure that aes-ni acceleration code uses cryptd (iow real asynchronous processing)
>> > and also that not all CPU cores can run these instruction in parallel.
>>
>> I think you can configure it to use cryptd (or pcrypt), but it's not default
>> and usually higher overhead.
>
> Right.  The only reason aes-ni uses the async interface is to
> work around the fact that you have to save FPU state when using
> it.  For dm-crypt you can consider it to be synchronous.

The reason for aes-ni to use async interface is that the FPU/SSE
instructions enclosed by kernel_fpu_begin and kernel_fpu_end is not
reentranceable. That is, it is not safe to use FPU/SSE instructions in
interrupt context, if the instructions are used in interrupted
context. So irq_fpu_usable is used in the async interface of aes-ni,
and go asynchronous implementation only if !irq_fpu_usable. In most
instances, the async interface of aes-ni will run in synchronous mode.

Best Regards,
Huang Ying