> > I think the crypto framework still
quoted text
> > does too many context switches for small operations.  IIRC It also
> > doesn't do much load balancing when you have multiple accelerators in
> > the system.
> 
> I'm not too interested in accelerator cards at this point, just
> software implementations (i.e., cryptosoft.c), and my current
> understanding is that these are handled without context switches.  Is
> that mistaken?

No, those operations are immediate done and then return.  In 10 years
I've heard a lot of people talk about writing custom back ends for
this which can still operate on mbufs, but they never saw the light of
day.

quoted text
> (I am somewhat interested in also experimenting with using high-end
> graphics cards with programmable GPUs to handle offloading crypto work
> too, but that's a way off project, if ever; the plan I already
> outlined feels much more reasonable at this point.)

Same here; I've heard chatter about this for more than 10 years now,
and never seen anything which was more than a research project.
Sorry, but it sounds like a massive waste of time and electricity.

I suppose it depends what you are trying to do.  Userland crypto?
Then just do it in userland.  Or in the kernel, you want to do IPSEC?
You want to move every single IPSEC packet twice more over the PCI
bus?  Sorry, but that sounds like a rather significant
de-optimization.  That is why the hifn cards suck.