[PATCH v6 6/8] dma-mapping: detect and configure IOMMU in of_dma_configure

[Laurent Pinchart]

Hi Will,

On Monday 15 December 2014 18:09:33 Will Deacon wrote:
> On Mon, Dec 15, 2014 at 05:16:50PM +0000, Laurent Pinchart wrote:
> > On Monday 15 December 2014 16:40:41 Will Deacon wrote:
> >> On Sun, Dec 14, 2014 at 03:49:34PM +0000, Laurent Pinchart wrote:
> >>> On Wednesday 10 December 2014 15:08:53 Will Deacon wrote:
> >>>> On Wed, Dec 10, 2014 at 02:52:56PM +0000, Rob Clark wrote:
> >>>>> so, what is the way for a driver that explicitly wants to manage
> >>>>> it's own device virtual address space to opt out of this?  I
> >>>>> suspect that won't be the common case, but for a gpu, if dma layer
> >>>>> all of a sudden thinks it is in control of the gpu's virtual
> >>>>> address space, things are going to end in tears..
> >>>> 
> >>>> I think you'll need to detach from the DMA domain, then have the
> >>>> driver manage everything itself. As you say, it's not the common
> >>>> case, so you may need to add some hooks for detaching from the
> >>>> default domain and swizzling your DMA ops.
> >>> 
> >>> I'm wondering if it's such an exotic case after all. I can see two
> >>> reasons not to use the default domain. In addition to special
> >>> requirements coming from the bus master side, the IOMMU itself might
> >>> not support one domain per bus master (I'm of course raising the issue
> >>> from a very selfish Renesas IPMMU point of view).
> >> 
> >> Do you mean that certain masters must be grouped into the same domain,
> >> or that the IOMMU can fail with -ENOSPC?
> > 
> > My IOMMU has hardware supports for 4 domains, and serves N masters (where
> > N is dependent on the SoC but is > 4). In its current form the driver
> > supports a single domain and thus detaches devices from the default
> > domain in the add_device callback:
>
> Hmm, ok. Ideally, you wouldn't need to do any of that in the driver, but I
> can understand why you decided to go down that route.

I'm of course open to alternative suggestions :-)

> > 	/*
> > 	 * Detach the device from the default ARM VA mapping and attach it to
> > 	 * our private mapping.
> > 	 */
> > 	arm_iommu_detach_device(dev);
> > 	ret = arm_iommu_attach_device(dev, mmu->mapping);
> > 	if (ret < 0) {
> > 	
> > 		dev_err(dev, "Failed to attach device to VA mapping\n");
> > 		return ret;
> > 	
> > 	}
> > 
> > I would have implemented that in the of_xlate callback, but that's too
> > early as the ARM default domain isn't created yet at that point.
> 
> Yup, the mythical ->get_default_domain might be the right place instead.
> 
> > Using a single domain is a bit of a waste of resources in my case, so an
> > evolution would be to create four domains and assign devices to them based
> > on a policy. The policy could be fixed (round-robin for instance), or
> > configurable (possibly through DT, although it's really a policy, not a
> > hardware description).
> 
> I think having one default domain, which is home to all of the masters that
> don't have any DMA restrictions is a good use of the hardware. That then
> leaves you with three domains to cover VFIO, devices with DMA limitations
> and potentially device isolation (if we had a way to describe that).
> 
> >> For the former, we need a way to represent IOMMU groups for the platform
> >> bus.
> > 
> > To be honest I'm not entirely sure how IOMMU groups are supposed to be
> > used. I understand they can be used by VFIO to group several masters that
> > will be able to see each other's memory through the same page table, and
> > also that a page table could be shared between multiple groups. When it
> > comes to group creation, though, things get fuzzy. I started with
> > creating one group per master in my driver, which is probably not the
> > thing to do. The Exynos IOMMU driver used to do the same, until Marek's
> > patch series converting it to DT- based instantiation (on top of your
> > patch set) has removed groups altogether. Groups seem to be more or less
> > optional, except in a couple of places (for instance the remove_device
> > callback will not be called by the BUS_NOTIFY_DEL_DEVICE notifier if the
> > device isn't part of an iommu group).
> > 
> > I'd appreciate if someone could clarify this to help me make an informed
> > opinion on the topic.
> 
> Ok, an iommu_group is the minimum granularity for which a specific IOMMU
> can offer address translation. So, if your IPMMU can isolate an arbitrary
> device (assuming there is a domain available), then each device is in its
> own iommu_group. This isn't always the case, for example if two masters are
> behind some sort of bridge that makes them indistinguishable to the IOMMU
> (perhaps they appear to have the same master ID), then they would have to
> be in the same iommu_group. Essentially, iommu_groups are a property of
> the hardware and should be instantiated by the bus. PCI does this, but
> we don't yet have anything for the platform bus.
> 
> VFIO puts multiple groups (now called vfio_groups) into a container. The
> container is synoymous to an iommu_domain (i.e. a shared address space).

Thank you for the explanation, that was very informative. It would be nice to 
capture that somewhere in Documentation/ ;-)

I'll thus create one group per bus master in the of_xlate function, as my 
platforms don't have masters indistinguishable to the IOMMUs (at least not to 
my knowledge).

-- 
Regards,

Laurent Pinchart