Re: [PATCH v1 02/20] iommu: Introduce a test_dev domain op and an internal helper

[Jason Gunthorpe <jgg@nvidia.com>]

On Wed, Nov 05, 2025 at 10:18:01AM -0800, Nicolin Chen wrote:
> On Wed, Nov 05, 2025 at 06:57:31AM +0000, Tian, Kevin wrote:
> > > From: Jason Gunthorpe <jgg@nvidia.com>
> > > Sent: Tuesday, November 4, 2025 2:54 AM
> > > 
> > > On Thu, Oct 30, 2025 at 12:43:59PM -0700, Nicolin Chen wrote:
> > > 
> > > > FWIW, I am thinking of another design based on Jason's remarks:
> > > > https://lore.kernel.org/linux-iommu/aQBopHFub8wyQh5C@Asurada-
> > > Nvidia/
> > > >
> > > > So, instead of core initiating the round trip between the blocking
> > > > domain and group->domain, it forwards dev_reset_prepare/done to the
> > > > driver where it does a low-level attachment that wouldn't fail:
> > > >   For SMMUv3, it's an STE update.
> > > >   For intel_iommu, it seems to be the context table update?
> > > 
> > > Kevin, how bad do you think the UAPI issue is if we ignore it?
> > > 
> > 
> > yeah probably better to leave it. I didn't see a clean way and the
> > value didn't justify the complexity.
> > 
> > Regarding to PF reset, it's a devastating operation while the vf user
> > is operating the vf w/o any awareness. there must be certain
> > coordination in userspace. otherwise nobody can recover the
> > registers. Comparing to that, solving the domain attach problem
> > is less important...
> 
> If I capture these correctly, we should go with a -EBUSY version:
>  - Reject concurrent attachments during a device reset
>  - Skip reset for devices having sibling group devices
>  - Allow PF to stop IOMMU, ignoring VFs
> ?
> 
> That sounds pretty much like this v4:
> https://lore.kernel.org/linux-iommu/0f6021b500c74db33af8118210dd7a2b2fd31b3c.1756682135.git.nicolinc@nvidia.com/
> by dropping the SRIOV concern.

It seems like the simplest answer..

I'd ignore the VFs, I think it is already really weird/dangerous to be
resetting the PF while VFs have drivers bound.. Not sure there is
anything we can do to make this work better.

Jason