s390 PCI dma ops

s390 PCI dma ops

[Christoph Hellwig]

Hi S/390 PCI and IOMMU maintainers,

We've recently tried to convert all implementations of the IOMMU API to
use drivers/iommu/dma-iommu.c instead of their local IOMMU ops.  I've
looked at S/390 and while some comments mention things about the iommu
API the implementation looks quite separate.  Do arch/s390/pci/pci_dma.c
and drivers/iommu/s390-iommu.c driver the same hardware?  If so would a
switch to drivers/iommu/dma-iommu.c be feasibel?

Re: s390 PCI dma ops

[Matthew Rosato]

On 9/11/20 3:27 AM, Christoph Hellwig wrote:
> Hi S/390 PCI and IOMMU maintainers,
> 
> We've recently tried to convert all implementations of the IOMMU API to
> use drivers/iommu/dma-iommu.c instead of their local IOMMU ops.  I've
> looked at S/390 and while some comments mention things about the iommu
> API the implementation looks quite separate.  Do arch/s390/pci/pci_dma.c
> and drivers/iommu/s390-iommu.c driver the same hardware?  If so would a

Yes, these drivers are dealing with the same hardware.

> switch to drivers/iommu/dma-iommu.c be feasibel?
> 

Hmm...  s390 does some particular things that are arch-specific, largely 
as forms of optimization.  One such example is a case where we avoid 
reusing IO TLB Page Table Entries until we have to do a global flush.

I'd need to take a closer look at what AMD did / what intel is doing. 
Fundamentally we are open to the idea as the maintenance benefit is 
clear, but I don't think it's a trivial swap - I think we would need 
extensions to dma-iommu to accommodate some of our special cases / 
optimizations.