Re: [GIT PULL] iommu: Kill off pgsize_bitmap field from struct iommu_ops

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

On Wed, Apr 01, 2015 at 04:38:54PM +0100, Joerg Roedel wrote:
> Hi Will,

Hi Joerg,

> On Tue, Mar 31, 2015 at 03:49:56PM +0100, Will Deacon wrote:
> > But isn't this restriction already the case in practice? For example, if
> > I have a domain with some mapping already configured, then that mapping
> > will be using some fixed set of page sizes. Attaching a device behind
> > another IOMMU that doesn't support that page size would effectively require
> > the domain page tables to be freed and re-allocated from scratch.
> 
> The problem is that this restriction depends on the IOMMU driver in use.
> From the beginning of the IOMMU-API the backend drivers have supported
> sharing a domain by multiple devices. In fact, the domain is just an
> abstraction for an address space that can be attached to devices just
> like cpu page-tables are assigned to threads of a process.

I think the domain is a useful abstraction; I just don't think it works
too well when you attach devices upstream of different IOMMUs to the same
domain. If all the devices attached to a domain share an IOMMU, everything
works nicely.

> We can discuss whether this fundamental concept of the IOMMU-API needs
> to be changed (moving into the direction this patch-set proposes). I
> could imaging removing the domain concept entirely and just have
> map/unmap functions like this:
> 
> 	iommu_map(struct device *dev, dma_addr_t iova,
> 		  phys_addr_t phys, size_t size)
> 	iommu_unmap(struct device *dev, dma_addr_t iova,
> 		    phys_addr_t phys, size_t size)
> 
> This would require some changes for the users of the IOMMU-API, but it
> shouldn't be too hard.
> 
> Or we keep the desired semantics as they are now and do everything
> needed for that in the IOMMU drivers. For the arm-smmu this would mean
> exposing a common set of supported pgsizes between IOMMUs, or to build
> multiple page-tables to match the different IOMMU capabilities.
> 
> I am open for disussions for either way, but I like the current
> semantics a bit more, as it allows us to share page-tables between
> devices and we can move all of the nasty code in VFIO that already
> creates multiple domains to get different page-tables into the
> IOMMU core or the drivers (were it belongs).

I agree that sharing page tables is desirable and I certainly wouldn't
suggest removing the domain abstraction.

> What I definitly don't want is a mixture of both concepts depending on
> the IOMMU driver in use. We should settle on one way and force the
> drivers to behave accordingly. We don't need a common API when every
> driver behaves differently.

Agreed. How would you feel about restricting domains to be per-IOMMU
instance? VFIO already copes with this, so I think we'd just need
something to keep legacy KVM device passthrough working on x86. Maybe we
could add a new domain type using your new series (DOMAIN_X86_KVM_LEGACY
or something) and avoid the cross-IOMMU domain checks for that.

Will