[PATCH 04/20] iommu/fsl_pamu: Replace set_platform_dma_ops() with IOMMU_DOMAIN_PLATFORM

Jason Gunthorpe jgg at nvidia.com
Wed May 3 05:54:07 PDT 2023


On Wed, May 03, 2023 at 11:57:59AM +0100, Robin Murphy wrote:
> On 2023-05-01 19:02, Jason Gunthorpe wrote:
> > It is not clear what this is actually doing, most likely this is IDENTITY
> > behavior, but I think there is a chance it is BLOCKING given how the PAMU
> > stuff is oddly used.
> 
> Logically it has to be identity, since there are no DMA ops interacting with
> this driver (it's not the TCE IOMMU of arch/powerpc/kernel/iommu.c), so any
> device using a kernel driver rather than VFIO must be using dma-direct and
> thus require an identity mapping.

Yes, that is the usual argument.. FSL just seems so odd because it has
drivers manipulating the iommu so it wasn't clear to me if those
drivers were the only drivers hooked here or not.

> At this point I finally got sufficiently fed up of this driver always being
> the mystery weirdo and tracked down an old QorIQ reference manual, and now I
> have about half an hours' worth of understanding of how the PAMU actually
> works.

I'm going to post my other FSL patches for the group setup then before
you forget ;)

> Based on that, what setup_liodns() is doing is indeed setting up identity
> for everything initially. It also becomes apparent that it's never supported
> giving a PCI device back to its regular driver after using vfio-pci, since
> an attach/detach cycle will then leave the PPAACE invalid and thus DMA
> blocked. Oh well, that's been broken for 10 years; nobody cares. Call it an
> identity domain and move on.

Okay, easy to do thanks for checking the manual

Jason



More information about the Linux-mediatek mailing list