Hi Nikos,

On 19.08.2014 07:52, Nikos Mavrogiannopoulos wrote:
> On Mon, 2014-08-11 at 10:09 +0300, Cristian Stoica wrote:
>> Yes, that kernel patch is used with COP_FLAG_AEAD_TLS_TYPE but there is
>> a difference from the cryptodev implementation (which I used for
>> inspiration) in that it registers an algorithm to do the same work. The
>> reason is that our HW supports AES-CBC-HMAC-SHA1 as a primitive
>> operation and needs that tls(...) algorithm to advertise the capability.
> That makes sense. What I realized after introducing this API, is when
> speeding that part, the bottleneck becomes the context switches in
> read()->decrypt() (and encrypt->write). An API that would combine these
> operations would have provided a significant boost.

The speed-up we get with the current API is pretty good but we know is
due only to hardware acceleration. We're doing some profiling and for
several reasons we wanted keep API changes to minimum. Yet if we see
improvements in this area we'll send patches for review.

Cristian S.

Cryptodev-linux-devel mailing list

Reply via email to