Re: [PATCH v4 28/41] vfio/iommufd: Implement the iommufd backend

[Jason Gunthorpe <jgg@nvidia.com>]

On Thu, Nov 09, 2023 at 01:21:59PM +0000, Joao Martins wrote:
> On 09/11/2023 13:09, Jason Gunthorpe wrote:
> > On Thu, Nov 09, 2023 at 01:03:02PM +0000, Joao Martins wrote:
> > 
> >>> I am not talking about mdevs; but rather the regular (non mdev) case not being
> >>> able to use dirty tracking with autodomains hwpt allocation.
> >>
> >> ... without any vIOMMU.
> > 
> > Ah, well, that is troublesome isn't it..
> > 
> > So do we teach autodomains to be more featured in the kernel or do we
> > teach the generic qemu code to effectively implement autodomains in
> > userspace?
> 
> The latter is actually what we have been doing. Well I wouldn't call autodomains
> in qemu, but rather just allocate a hwpt, instead of attaching the IOAS
> directly. But well mdevs don't have domains and we overlooked that. I would turn
> the exception into an exception rather than making the norm, doesn't look to be
> much complexity added?

Autodomains are complex because of things like mdev and iommu
non-uniformity's. Qemu can't just allocate a single HWPT, it needs to
be annoyingly managed.

> What I last re-collect is that autodomains represents the 'simple users' that
> don't care much beyond the basics of IOMMU features (I recall the example was
> DPDK apps and the like). You could say that for current needs IOMMU autodomains
> suffices for qemu.

Yes, that was my intention. Aside from that it primarily exists to
support vfio compatibility

> Connecting autodomains to this enforcing on the hwpt is relatively simple btw,
> it just needs to connect the dirty tracking flag with same semantic of
> hwpt-alloc equivalent and pass the hwpt flags into the domain allocation.

Yes

> It's more of what of a question should be the expectations to the user when
> using ATTACH_HWPT with an IOAS_ID versus direct manipulation of HWPT. I am
> wondering if dirty tracking is alone here or whether there's more features that
> start to mud the simplicity of autodomains that would approximate of hwpt-alloc.

This is why I had been thinking of a pure HWPT based scheme

So it seems we cannot have a simple model where the generic qmeu layer
just works in IOAS :( It might as well always work in HWPT and
understand all the auto domains complexity itself.

Jason