Re: [RFC PATCH v3 5/7] dma-mapping: detect and configure IOMMU in of_dma_configure

[Will Deacon <will.deacon-5wv7dgnIgG8@public.gmane.org>]

On Mon, Oct 27, 2014 at 11:30:33AM +0000, Laurent Pinchart wrote:
> Hi Will,

Hey Laurent,

> On Monday 27 October 2014 10:51:59 Will Deacon wrote:
> > On Tue, Oct 14, 2014 at 01:53:59PM +0100, Laurent Pinchart wrote:
> > > On Monday 22 September 2014 18:50:27 Will Deacon wrote:
> > > > On Mon, Sep 22, 2014 at 10:29:10AM +0100, Thierry Reding wrote:
> > > > > Agreed. I wonder how useful it is to know the set of IOMMU instances
> > > > > that each device can master through. Wouldn't it be more useful to
> > > > > keep a list of master interfaces for each device? The set of IOMMU
> > > > > instances can trivially be derived from that.
> > > > 
> > > > I'm struggling to think how that would look. What do you mean by `master
> > > > interfaces' in terms of the code we have in Linux? At the end of the
> > > > day, the list of IOMMU instances (i.e. iommu_dma_mapping) exists because
> > > > you and Laurent have use-cases involving devices mastering through
> > > > multiple IOMMUs. If it doesn't work for you, it might be best for you to
> > > > send me the patch ;)
> > > 
> > > Just for the record, I've brought up the topic of masters being served by
> > > multiple IOMMUs, but don't have a use case for it (yet at least). I do
> > > have masters served through multiple streams with separate stream IDs, but
> > > all by the same IOMMU.
> > 
> > Ok. I spoke to Arnd, David and Joerg at LPC and the consensus was that the
> > DMA-mapping API should *not* be exposed to the details of masters that
> > master through multiple IOMMUs. Instead, that should be abstracted by the
> > device API by exposing that device as a single struct device.
> 
> I'm not sure to follow you here. Aren't we already exposing masters that 
> master through multiple IOMMUs as single instances of struct device ?

Hmm, yes, now you've confused me too! The conclusion was certainly that
dma-mapping should not be the one dealing with the I/O topology. Domain
allocation would then be an iommu callback (something like
->get_default_domain), but the rest of the details weren't fleshed out.

Joerg?

> > So, that's certainly an area that needs more work and I'll drop the limited
> > support I'd cooked up from this patch set in the next version.
> 
> How about masters connected to multiple stream IDs of the same IOMMU ?

That should still be handled, as I believe that will be a common case.

Will