converting sync driver to async

converting sync driver to async

[avital sela]

Hello,

I have written sha and aes HW accelerator drivers for our custom
drivers (thanks for your patient help) and they work great.
To exploit the full potential of the HW I want to port them to the
ASYNC interface and have some questions.
1. Using my current setup I see the following results (measured with iperf)
no ipsec  -85 mbits
ESP in tunnel mode with null algs  - 45 mbits
ESP in tunnel mode with aes and sha1 - 32 mbits
Is it reasonable to have the throughput drop so much even without any
actual crypto going on? Can this mostly be attributed to the bigger
packets ?
My understanding of the results is that the maximum throughput I can
hope to achieve on this systen by only improving the crypto drivers is
45 mbits?
2. Is there any comparison somewhere between OCF performance and Acrypto?
3. Using drivers/crypto/talitos as a reference I am setting the alg
name to "authenc(hmac(sha1),cbc(aes))",cra_flags to
CRYPTO_ALG_TYPE_AEAD | CRYPTO_ALG_ASYNC;
   and cra_type to &crypto_aead_type. My existing driver had a
cra_list field. Is that required for this type of driver as well?

Thanks,
Avital

Re: converting sync driver to async

[Herbert Xu]

On Tue, Feb 16, 2010 at 06:06:37PM +0200, avital sela wrote:
> Hello,
> 
> I have written sha and aes HW accelerator drivers for our custom
> drivers (thanks for your patient help) and they work great.
> To exploit the full potential of the HW I want to port them to the
> ASYNC interface and have some questions.
> 1. Using my current setup I see the following results (measured with iperf)
> no ipsec  -85 mbits
> ESP in tunnel mode with null algs  - 45 mbits
> ESP in tunnel mode with aes and sha1 - 32 mbits
> Is it reasonable to have the throughput drop so much even without any
> actual crypto going on? Can this mostly be attributed to the bigger
> packets ?

That should be easy to tell.  Just disable GSO/SG/checksum offload
on the physical interface and retest with no ipsec.

> My understanding of the results is that the maximum throughput I can
> hope to achieve on this systen by only improving the crypto drivers is
> 45 mbits?

Yes.

> 2. Is there any comparison somewhere between OCF performance and Acrypto?

Not that I'm aware unfortunately.

> 3. Using drivers/crypto/talitos as a reference I am setting the alg
> name to "authenc(hmac(sha1),cbc(aes))",cra_flags to
> CRYPTO_ALG_TYPE_AEAD | CRYPTO_ALG_ASYNC;
>    and cra_type to &crypto_aead_type. My existing driver had a
> cra_list field. Is that required for this type of driver as well?

The cra_list field does not have to be initialised so no you
don't need to make it explicit.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt