Re: [PATCH] AES i586 optimized (oops)

[Felipe Alfaro Solana <felipe_alfaro@linuxmail.org>]

On Thu, 2004-05-13 at 17:29, Felipe Alfaro Solana wrote:
> On Thu, 2004-05-13 at 16:05, Felipe Alfaro Solana wrote:
> > On Thu, 2004-05-13 at 13:42, Fruhwirth Clemens wrote:
> > > > 
> > > > Some benchmark figures would be useful.  Cache-cold and cache-hot.
> > > 
> > > I posted this patch for the first time about 3/4 year ago. The first
> > > response has been the same. Please have a look at
> > 
> > I'll give this a spin, since I'm very interested in AES. Currently, I'm
> > using IPSec with AES ESP all over my 100Mbps LAN. I'll apply this one to
> > 2.6.6-mm2 and will compare with vanillas 2.6.6 and 2.6.6-mm2.
> 
> I got an oops on 2.6.6-mm2 plus i586 AES patch when invoking setkey to
> setup my SPD database with preshared keys entries. Attached dmesg and
> config.

As pointed before, it was the CONFIG_REGPARM config option. Reverting
the force-config_regparm-to-y.patch and setting CONFIG_REGPARM to no
solves the problems.

Having tried the new i586-optimized AES, I must say that throughput
increases from ~7MB/s with vanilla AES to more than ~9.4MB/s with the
i586-optimized AES (when FTPing a large file over the AES/ESP encrypted
link).

That's like saying that I'm getting more than 80% of the maximum
sustained, unencrypted throughput for my network link (which is at
~11MB/s). That's really impressive, since that's a Security Association
between a 700Mhz PIII and a 2GHz P4.

Nice work, indeed.