Re: [RFC v2] /dev/iommu uAPI proposal

[Jason Gunthorpe]

On Fri, Aug 06, 2021 at 02:45:26PM +1000, David Gibson wrote:

> Well, that's kind of what I'm doing.  PCI currently has the notion of
> "default" address space for a RID, but there's no guarantee that other
> buses (or even future PCI extensions) will.  The idea is that
> "endpoint" means exactly the (RID, PASID) or (SID, SSID) or whatever
> future variations are on that.

This is already happening in this proposal, it is why I insisted that
the driver facing API has to be very explicit. That API specifies
exactly what the device silicon is doing.

However, that is placed at the IOASID level. There is no reason to
create endpoint objects that are 1:1 with IOASID objects, eg for
PASID.

We need to have clear software layers and responsibilities, I think
this is where the VFIO container design has fallen behind.

The device driver is responsible to delcare what TLPs the device it
controls will issue

The system layer is responsible to determine how those TLPs can be
matched to IO page tables, if at all

The IO page table layer is responsible to map the TLPs to physical
memory.

Each must stay in its box and we should not create objects that smush
together, say, the device and system layers because it will only make
a mess of the software design.

Since the system layer doesn't have any concrete objects in our
environment (which is based on devices and IO page tables) it has to
exist as metadata attached to the other two objects.

Jason