Re: [Qemu-devel] [PATCH] Add a memory barrier to DMA functions

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Tue, 2012-05-22 at 14:34 +1000, Benjamin Herrenschmidt wrote:
> So here's the latest try :-) I've kept is simple, I don't
> add anything to map/unmap at this stage, so we might still
> have a problem with drivers who do that a lot without any
> explicit barrier (ahci ?). My preference is to also add
> barriers to map/unmap by default but we can discuss it.
> 
> Note that I've put the barrier in an inline "helper" so
> we can nag about the type of barriers, or try to be smart,
> or use flags in the DMAContext or whatever we want reasonably
> easily as we have a single spot to modify.
> 
> I'm now going to see if I can measure a performance hit on
> my x86 laptop, but if somebody who has existing x86 guest setups
> wants to help, that would be much welcome :-) 

Also an idea I had to make it easier & avoid a clutter of
_relaxed variants of everything:

Can we change DMADirection to be a DMAAttributes or DMAFlags,
and basically have more than just the direction in there ?

That way we can easily have "relaxed" be a flag, and thus
apply to most of the accessors without adding a bunch of
different variant (especially since we already have that
"cancel" variant for map, I don't want to make 2 more
for _relaxed).

If I do that change I'd also like to change the enum so that
read and write are distinct bits, and so we can set both for
bidirectional. This makes more sense if we ever have to play
with cache issues etc... and might help using more optimal
barriers. It might also fit better to IOMMUs who provide
distinct read vs. write permissions.

Cheers,
Ben.