On Tue, Jan 24, 2012 at 12:36:14PM +0100, Hiroshi Doyu wrote:

quoted

A domain is, as you said, a virtual address space for IO devices. But
the important point is, an arbitrary number of devices can be part of a
domain. This also means that the devices can be behind different
hardware SMMUs. In this case your driver needs to program the page-table
pointer into more than one SMMU to give devices behind different SMMUs
the same address space.

Thank you for explaining.

Does the above mean that a buffer can be shared with different devices
which belong to different IOMMU devices(virtual address spaces)?

For example, assuming the following:

- We have "struct iommu_domain *domain1".
- "domain1" has iommu device "iommu_dev1" and "iommu_dev2".
- "iommu_dev1" has "client_dev1" and "client_dev2".
- "iommu_dev2" has "client_dev3" and "client_dev4".

"iommu_map(domain1, iova, pa, ...)" will create the following mapping
___at once___:

- (iova)-(pa) mapping in iommu_dev1(iommmu_dev1's virtual address space)
- (iova)-(pa) mapping in iommu_dev2(iommmu_dev2's virtual address space)

Is the above correct?

Yes, this is correct.

It seems that the same (iova) is used for different virtual address
spaces. What kind of case is this beneficial most in?

It is actually the _same_ virtual address space which is used by
iommu_dev1 and iommu_dev2. Think of it like multiple threads of a single
process. They also share the address space.
This is a requirement of the iommu-api which is beneficial for
virtualization and simplifies the usage of the api in general. The user
does not need to care which devices can be assigned to which domain
because of underlying hardware constraints. The goal of the iommu-api is
to hide such contraints.


	Joerg

-- 
AMD Operating System Research Center

Advanced Micro Devices GmbH Einsteinring 24 85609 Dornach
General Managers: Alberto Bozzo
Registration: Dornach, Landkr. Muenchen; Registerger. Muenchen, HRB Nr. 43632