[PATCH v4] devicetree: Add generic IOMMU device tree bindings

[Will Deacon]

On Thu, Jul 10, 2014 at 11:57:38AM +0100, Thierry Reding wrote:
> On Thu, Jul 10, 2014 at 11:23:34AM +0100, Will Deacon wrote:
> > On Thu, Jul 10, 2014 at 10:49:10AM +0100, Thierry Reding wrote:
> > > On Wed, Jul 09, 2014 at 07:10:48PM +0100, Will Deacon wrote:
> > > > On Wed, Jul 09, 2014 at 03:21:27PM +0100, Thierry Reding wrote:
> > > > > Anything beyond that (e.g. logical grouping of masters) isn't directly
> > > > > within the scope of the binding (it doesn't describe hardware but some
> > > > > policy pertaining to some specific use-case).
> > > > 
> > > > This *is* for hardware. I can use PCI as an example, but this could equally
> > > > apply to other types of bus. If you have a bunch of PCI master devices
> > > > sitting being a non-transparent bridge, they can end up sharing the same
> > > > master device ID (requester ID). This means that there is no way in the
> > > > IOMMU to initialise a translation for one of these devices without also
> > > > affecting the others. We currently have iommu_groups to deal with this, but
> > > > it *is* a property of the hardware and we absolutely need a way to describe
> > > > it. I'm happy to add it later, but we need to think about it now to avoid
> > > > merging something that can't easily be extended.
> > > > 
> > > > For PCI, the topology is probable but even then, we need this information to
> > > > describe the resulting master device ID emitted by the bridge for the
> > > > upstream group. One way to do this with your binding would be to treat all
> > > > of the upstream masters as having the same device ID.
> > > 
> > > Yes, I think that makes most sense. After all from the IOMMU's point of
> > > view requests from all devices behind the bridge will originate from the
> > > same ID.
> > > 
> > > So technically it's not really correct to encode the master ID within
> > > each of the devices, but rather they should be inheriting the ID from
> > > the non-transparent bridge.
> > 
> > Indeed. Is that possible with your binding, or would we just duplicate the
> > IDs between the masters?
> 
> No, the binding only describes direct relationships between the IOMMU
> and masters. There's no way to translate them inbetween or inherit them.

[...]

> ? That way some code could walk up the IOMMU tree to resolve this. Or
> perhaps even easier:
> 
> 	iommu {
> 		#iommu-cells = <1>;
> 	};
> 
> 	bridge {
> 		iommus = <&/iommu 42>;
> 
> 		device at 0 {
> 			...
> 		};
> 
> 		device at 1 {
> 			...
> 		};
> 
> 		...
> 	};

Yes, I like that. Good thinking!

> And we could enhance the binding by defining that the iommus node is
> inherited by devices on a bus, which by what you're saying would be the
> sensible thing to do anyway.
> 
> In the second example above, the presence of an iommus property in the
> bridge would indicate that it's non-transparent regarding IOMMU
> translation and therefore the master ID should be inherited. Devices
> could still override by providing their own iommus property, though I'd
> be a little surprised if there ever was hardware like that.
> 
> > > > With virtualisation, we may want to assign a group of devices to a guest but
> > > > without emulating the bridge. This would need something the device-tree to
> > > > describe that they are grouped together.
> > > 
> > > But that's also a software decision, isn't it? Virtualization doesn't
> > > have anything to do with the hardware description. Or am I missing
> > > something? Of course I guess you could generate a DTB for the guest and
> > > group device together, in which case you're pretty much free to do what
> > > you want since you're essentially defining your own hardware.
> > 
> > If you're doing device passthrough and you want to allow the guest to
> > program the IOMMU, I think that virtualisation is directly related to the
> > hardware description, since the guest will be bound by physical properties
> > of the system.
> 
> Evidently you know much better what the requirements are here and what
> will actually be required. I guess we'll need to have more discussions
> along with examples of use-cases.

It's still early days for getting this stuff up and running on ARM, so I
agree that we'll have to come back to it a few times once we've got concrete
examples and code.

Will