[PATCHv4 2/7] driver/core: Populate IOMMU'able devices in order

[Will Deacon]

On Tue, Nov 12, 2013 at 11:34:20PM +0000, Stephen Warren wrote:
> On 11/11/2013 01:31 AM, Hiroshi Doyu wrote:
> > An "IOMMU device" on the bus is poplulated first, "IOMMU'able devices"
> > are done later.
> > 
> > With CONFIG_OF_IOMMU, "#stream-id-cells" DT binding would be used to
> > identify whether a device is IOMMU'able or not. If a device is
> > IOMMU'able, we'll defer to populate that device till an iommu device
> > is populated. Once an iommu device is populated, "dev->bus->iommu_ops"
> > is set in the bus. Then, those defered IOMMU'able devices are
> > populated and configured as IOMMU'abled with help of the already
> > populated iommu device via iommu_ops->add_device().
> 
> This looks fairly neat and clean.
> 
> I'm still worried about using #stream-id-cells in DT nodes though. While
> I do understand that the *Linux* device model currently only allows each
> struct device to be affected by a single IOMMU, I worry that encoding
> that same restriction into DT is a mistake. I'd far rather see a
> property like:
> 
> SMMU:
>     smmu: smmu at xxxxxx {
>         #smmu-cells = <1>;
>     }
> 
> Affected device:
>     smmus = <&smmu 1>;
>     (perhaps with smmu-names too)
> 
> That would allow the DT to represent basically arbitrary HW configurations.
> 
> The implementation of this patch would then be almost as trivial; you'd
> just need to walk the smmus property to find each phandle in turn, just
> like any other phandle+specifier property, and validate that the SMMU
> driver was already probe()d for each.

There are a few problems with that:

  1.) It assumes all devices sharing an SMMU have the same number of
      "smmu cells"

  2.) It moves SMMU-specific data out to the device, which makes it
      impossible to describe more complicated topologies where IDs can be
      remapped/remastered, potentially by multiple SMMUs and/or bus bridges.

When writing the binding for the ARM SMMU driver, I originally started with
something similar to what you're suggesting, but was later forced down a
different route when I realised what sort of systems were being built.

We will have similar problems for PCIe RIDs and GIC DIDs (I spoke about this
at the ARM mini-summit). They are not fixed across the system: they
originate from a device, but can change as they traverse the system
topology.

Will