Re: [PATCH v7 1/3] iommufd: Add data structure for Intel VT-d stage-1 cache invalidation

[Nicolin Chen <nicolinc@nvidia.com>]

On Wed, Jan 03, 2024 at 12:01:08PM -0400, Jason Gunthorpe wrote:
> On Wed, Jan 03, 2024 at 10:24:42AM +0800, Yi Liu wrote:
> > On 2024/1/3 07:38, Jason Gunthorpe wrote:
> > > On Fri, Dec 15, 2023 at 12:01:19PM +0800, Yi Liu wrote:
> > > > > I think I misread Yi's narrative: dev_id is a working approach
> > > > > for VMM to convert to a vRID, while he is asking for a better
> > > > > alternative :)
> > > > 
> > > > In concept, dev_id works, but in reality we have problem to get a dev_id
> > > > for a given device in intel iommu driver, hence I'm asking for help here. :)
> > > 
> > > I think we just need to solve this one way or another.. Even if you
> > > use a viommu object you still end up having difficult coupling to
> > > iommufd
> > > 
> > > Some:
> > >    iommufd_get_dev_id(struct iommufd_ctx *ictx, struct device *dev)
> > > 
> > > Callable by a driver (using the driver-callable function
> > > infrastructure we made for dirty tracking) Is really all that is
> > > needed here.
> > 
> > yep, I noticed IOMMUFD_DRIVER was selected by intel iommu driver when
> > IOMMUFD is configed. Maybe such an API could be compiled when
> > IOMMUFD_DRIVER is enabled as well. This does address my concern on making
> > intel iommu driver depending on iommufd. But still need a way to pass in
> > the iommufd_ctx pointer to intel iommu driver, and store it. Hence intel
> > iommu driver could call the above iommufd_get_dev_id(). One possible way is
> > passing it when attaching device to domain and clear it in detach. However,
> > this seems not ideal as iommufd_ctx information should be static between
> > bind_iommufd and unbind. But we don't call into intel iommu driver in the
> > bind and unbind operations. May need to add new iommu op. Any suggestion
> > here?
> 
> You can pass the ctx to the invalidate op, it is already implied
> because the passed iommu_domain is linked to a single iommufd ctx.

The device virtual id lookup API needs something similar, yet it
likely needs a viommu pointer (or its id) instead? As the table
is attached to a viommu while an ictx can have multiple viommus,
right?

Thanks
Nic