Re: [PATCH v2 4/4] coco/tdx-host: Initialize trusted DMA on all capable IOMMUs

[Jason Gunthorpe]

On Fri, Oct 16, 2026 at 07:17:31AM +0000, Tian, Kevin wrote:
> +Aneesh.
> 
> > From: Baolu Lu <baolu.lu@xxxxxxxxxxxxxxx>
> > Sent: Friday, October 16, 2026 3:03 PM
> > 
> > On 10/16/2026 1:37 AM, Jason Gunthorpe wrote:
> > > On Thu, Oct 15, 2026 at 04:31:31PM +0800, Xu Yilun wrote:
> > >
> > >> An alternative could be, bringing up one IOMMU when an associated
> > >> device requests link encryption. Then a no-tdx-support iommu fails the
> > >> encryption request naturally.
> > > That is what ARM is planning to do. I think it is reasonable to
> > > trigger creation of the secure iommu when the TSM is connected to TDX.
> > >
> > > Fail connect if the iommu could not be created.
> > 
> > Yes, agreed. The Intel iommu driver could use per-device interfaces,
> > just as ARM is planning. There are some differences, though.
> > 
> > ARM's interaction with the iommu driver in viommu_init() appears to
> > consist of two steps:
> > 
> >   - arm_smmu_get_realm_params(dev, &params): get and validate the
> >     physical iommu information from the arm-smmu-v3 driver;
> >   - cca_psmmu_get(&params): activate the physical smmu for Realm use.
> > 
> > Intel differs in two ways:
> > 
> >   - ARM queries the physical iommu information from the iommu driver
> >     in the viommu_init() path. On Intel, the secure iommu must be set
> >     up during TSM connect, because it is part of establishing PCIe
> >     link encryption.
> 
> My understanding is that the psmmu is required for link encryption, so
> it's similar ordering here i.e. activated before tsm connect?

No, AFAIK? The psmmu has to be linked before creating the vdev, and in
the current spec the only way to get the measurements is through the
vdev, but I am expecting that to change.

I have no idea what brilliant idea caused the VT-d iommu HW to become
involved in DOE and link IDE configuration in the root port. That
makes very little sense from an architecture perspective.

> >   - The TDX Connect architecture places requirements on the host
> >     iommu driver while the secure IOMMU is being set up. For example,
> >     the host iommu driver must block the enhanced command (ECMD)
> >     interface while the TDH.IOMMU.SETUP SEAMCALL runs, and it must
> >     reserve the upper half of the domain ID space. These steps have
> >     to happen atomically under the iommu driver's own locks. As a
> >     result, the Intel iommu driver needs to merge the two steps above
> >     into one.
> 
> Didn't get this. Both ECMD and DID reservation are TDX specific things.
> How are they related to the two steps in the ARM sides and why calls
> for merging?
> 
> and why are they must be done atomically? they are conditions to
> execute TDH.IOMMU.SETUP, but it is perfectly fine to reserve DID
> at an early point then block ECMD later...

That's going to be a bit of a PITA, since the DIDs can be in use when
the tdx module is start potentially.

I suppose you can start out by just failing if this has happened.

But as soon as reasonably possible once the system opts into a TDX
operation the VT-d drivers should stop using the upper did space.

Otherwise you will have to implement live DID renaming, which is a
PITA in an iommu driver.

Jason