Re: [PATCH 1/2] arm: msm: Add System MMU support.

[Arnd Bergmann <arnd@arndb.de>]

On Thursday 29 July 2010, Roedel, Joerg wrote:
> On Thu, Jul 29, 2010 at 07:25:47AM -0400, Arnd Bergmann wrote:
> > On Thursday 29 July 2010, Roedel, Joerg wrote:
> > 
> > > > You designed it for what you need at the time. It should have been
> > > > named appropriately to avoid confusion. Later, when we actually
> > > > understand what other IOMMUs need, we can evolve the specific API for
> > > > generic purposes. Then we can rename the API to more generic.
> > > 
> > > At the time the iommu-api was written is was generic enough for what we
> > > had. So it was designed as an generic API. At this point in time nobody
> > > knew what the future requirements would we. So today it turns out that
> > > it is not generic enough anymore for latest hardware. The logical
> > > consequence is to fix this in the iommu-api.
> > 
> > Well, I think the real question is why we have two APIs that both claim
> > to work with IOMMUs in a generic way and how we could unify the two.
> 
> The DMA-API itself does not claim to be an iommu-frontend. The purpose
> of the DMA-API is to convert physical memory addresses into dma handles
> and do all the management of these handles. Backend implementations can
> use hardware iommus for this task. But depending on the hardware in the
> system the DMA-API can very well be implemented without any hardware
> support. This is an important difference to the IOMMU-API which needs
> hardware because it exposes hardware iommu features to software.

Well, you could call that a limitation in the IOMMU API ;-)

The idea behind the DMA mapping API is to allow a device driver
to work without knowing if the hardware can, cannot or must use
an IOMMU.

> > I really think we should not extend the (KVM) IOMMU API further but
> > just use the generic DMA mapping api for KVM and extend it as necessary.
> > It already has the concept of cache coherency and mapping/unmapping
> > that are in the IOMMU API and could be extended to support domains
> > as well, through the use of dma_attrs.
> 
> If we find a nice and clean way to expose lower-level iommu
> functionality through the DMA-API, thats fine. We could certainly
> discuss ideas in this direction.  I think this is going to be hard
> because the DMA-API today does not provide enough flexibility to let the
> user choose both sides of a io-virtual<->cpu-physical address mapping.
> Thats fine for most drivers because it makes sense for them to use the
> generic io-address-allocator the DMA-API provides but not for KVM which
> needs this flexibility.

One way to do this would be to add a new attribute, e.g.

enum dma_attr {
        DMA_ATTR_WRITE_BARRIER,
        DMA_ATTR_WEAK_ORDERING,
	DMA_ATTR_FIXED_MAPPING, /* this one is new */
        DMA_ATTR_MAX,
};

struct dma_attrs {
        unsigned long flags[__DMA_ATTRS_LONGS];
	dma_add_t dest;
};

Nothing except for KVM would need to use that attribute, and KVM would
obviously need a way to check if this is supported by the underlying
implementation.

	Arnd