Re: [RFC] [PATCH 4/4] esp: add the pcrypt hooks to esp

[Steffen Klassert <steffen.klassert@secunet.com>]

On Fri, Mar 27, 2009 at 04:44:57PM +0800, Herbert Xu wrote:
> On Mon, Mar 16, 2009 at 12:55:26PM +0100, Steffen Klassert wrote:
> >
> > @@ -447,7 +448,7 @@ static int esp_init_aead(struct xfrm_state *x)
> >  	struct crypto_aead *aead;
> >  	int err;
> >  
> > -	aead = crypto_alloc_aead(x->aead->alg_name, 0, 0);
> > +	aead = crypto_alloc_aead_tfm(x->aead->alg_name, 0, 0);
> 
> I'd like this to be configurable.  After all, there's not much
> point in doing this if your crypto is done through PCI.
> 

Indeed, it should be configurable. I was not sure what's the best way to     
make it configurable, so I decided to just make it the default to be
able to use it in this first version.

> The easiest way is to let the user specify pcrypt through the
> algorithm name.  In fact, as long as pcrypt has a higher priority
> than the default software algorithms, the user simply has to
> instantiate it in order for it to be the default for all uses
> of that algorithm.
> 

For IPsec I thought about something like 'pcrypt(authenc(...,...))' to
be able to process each crypto request with just one
parallelization/serialization call. Perhaps I'm missing something, but
actually I don't see how to choose for this from userspace through the
name without adding an entry for each possible combination of hmac and
blkcipher algorithm to the xfrm_aead_list. I'd appreciate any hint here.
My idea to make it configurable was to do it similar to the async
algorithms by

aead = crypto_alloc_aead_tfm(x->aead->alg_name, flags, CRYPTO_ALG_PCRYPT);

if CRYPTO_ALG_PCRYPT is set in flags, the crypto manager tries to choose
a parallel algorithm, if it is not set parallel algorithms will be ignored.
~