Re: [PATCH 1/2] powerpc: iommu: Bring back table group release_ownership() call

[Jason Gunthorpe]

On Fri, Jan 26, 2024 at 09:29:55AM -0600, Timothy Pearson wrote:
> > On Fri, Jan 26, 2024 at 08:43:12PM +0530, Shivaprasad G Bhat wrote:
> >> > Also, is there any chance someone can work on actually fixing this to
> >> > be a proper iommu driver? I think that will become important for power
> >> > to use the common dma_iommu code in the next year...
> >> We are looking into it.
> > 
> > Okay, let me know, I can possibly help make parts of this happen
> > 
> > power is the last still-current architecture to be outside the modern
> > IOMMU and DMA API design and I'm going to start proposing things that
> > will not be efficient on power because of this.
> 
> I can get development resources on this fairly rapidly, including
> testing.  We should figure out the best way forward and how to deal
> with the VFIO side of things, even if that's a rewrite at the end of
> the day the machine-specific codebase isn't *that* large for our two
> target flavors (64-bit PowerNV and 64-bit pSeries).

I have a feeling the way forward is to just start a power driver under
drivers/iommu/ and use kconfig to make the user exclusively select
either the legacy arch or the modern iommu.

Get that working to a level where dma_iommu is happy.

Get iommufd support in the new driver good enough to run your
application.

Just forget about the weird KVM and SPAPR stuff, leave it under the
kconfig of the old code and nobody will run it. Almost nobody already
runs it, apparently.

Remove it in a few years

>From what I remember the code under the arch directory is already very
nearly almost an iommu driver. I think someone could fairly quickly
get to something working and using dma_iommu.c. If s390 is any
experience there is some benchmarking and tweaking to get performance
equal to the arch's tweaked dma_iommu copy.

Jason