On 2023/12/9 09:47, Jason Gunthorpe wrote:
On Fri, Nov 17, 2023 at 05:07:11AM -0800, Yi Liu wrote:
Take Intel VT-d as an example, the stage-1 translation table is I/O page
table. As the below diagram shows, guest I/O page table pointer in GPA
(guest physical address) is passed to host and be used to perform the stage-1
address translation. Along with it, modifications to present mappings in the
guest I/O page table should be followed with an IOTLB invalidation.
I've been looking at what the three HW's need for invalidation, it is
a bit messy.. Here is my thinking. Please let me know if I got it right
What is the starting point of the guest memory walks:
Intel: Single Scalable Mode PASID table entry indexed by a RID & PASID
AMD: GCR3 table (a table of PASIDs) indexed by RID
ARM: CD table (a table of PASIDs) indexed by RID
What key does the physical HW use for invalidation:
Intel: Domain-ID (stored in hypervisor, per PASID), PASID
AMD: Domain-ID (stored in hypervisor, per RID), PASID
ARM: VMID (stored in hypervisor, per RID), ASID (stored in guest)
Why key does the VM use for invalidation:
Intel: vDomain-ID (per PASID), PASID
AMD: vDomain-ID (per RID), PASID
ARM: ASID
What is in a Nested domain:
Intel: A single IO page table refereed to by a PASID entry
Each vDomain-ID,PASID allocates a unique nesting domain
AMD: A GCR3 table pointer
Nesting domains are created for every unique GCR3 pointer.
vDomain-ID can possibly refer to multiple Nesting domains :(
Per section '2.2.6.3 Guest CR3 Table' in below spec, DTE has domainID,
and it points to a guest CR3 Table (it should be a set of guest CRs3)
which is indexed by PASID. So it looks like the PASID is per-DommainID.
HW should use domainID+PASID to tag the cache, otherwise there would be
cache conflict...
https://www.amd.com/content/dam/amd/en/documents/processor-tech-docs/specifications/48882_IOMMU.pdf
ARM: A CD table pointer
Nesting domains are created for every unique CD table top pointer.
How does the hypervisor compute it's cache tag:
Intel: Each time a nesting domain is attached to a (RID,PASID) the
hypervisor driver will try to find a (DID,PASID) that is
already used by this domain, or allocate a new DID.
AMD: When creating the Nesting Domain the vDomain-ID should be passed
in. The hypervisor driver will allocate a unique pDomain-ID for
each vDomain-ID (hand wave). Several Nesting Domains will share
the same p/vDomain-ID.
ARM: The VMID is uniquely assigned to the Nesting Parent when it
is allocated, in some configurations it has to be shared with
KVM's VMID so allocating the Nesting Parent will require a KVM FD.
Will ATC be forwarded or synthesized:
Intel: The (vDomain-ID,PASID) is a unique nesting domain so
the hypervisor knows exactly which RIDs this nesting domain is
linked to and can generate an ATC invalidation. Plan is to
supress/discard the ATC invalidations from the VM and generate
them in the hypervisor.
AMD: (vDomain-ID,PASID) is ambiguous, it can refer to multiple GCR3
tables. We know which maximal set of RIDs it represents, but not
the actual set. I expect AMD will forward the ATC invalidation
to avoid over invalidation.
ARM: ASID is ambiguous. We have no idea which Nesting Domain/CD table
the ASID is contained in. ARM must forward the ATC invalidation
from the guest.
What iommufd object should receive the IOTLB invalidation command list:
Intel: The Nesting domain. The command list has to be broken up per
(vDomain-ID,PASID) and that batch delivered to the single
nesting domain. Kernel ignores vDomain-ID/PASID and just
invalidates whatever the nesting domain is actually attached to
this is what we are doing in current series[1]. is it? Guest needs to
issue invalidation request with vDomain-ID and PASID (if it is enabled),
and affected pages for sure. But in hypervisor side, use vDomainID and
PASID info to figure out the target HWPT, then invoke a cache invalidation
on the HWPT with the invalidation range is enough. Kernel can figure out
what device/pasid this HWPT has been attached and invalidate the caches.
[1]
https://lore.kernel.org/linux-iommu/20231117131816.24359-1-yi.l.liu@xxxxxxxxx/
AMD: Any Nesting Domain in the vDomain-ID group. The command list has
to be broken up per (vDomain-ID). Kernel replaces
vDomain-ID with pDomain-ID from the nesting domain and executes
the invalidation.
ARM: The Nesting Parent domain. Kernel forces the VMID from the
Nesting Parent and executes the invalidation.
In all cases the VM issues an ATC invalidation with (vRID, PASID) as
the tag. The VMM must translate vRID -> dev_id -> pRID
For a pure SW flow the vRID can be mapped to the dev_id and the ATC
invalidation delivered to the device object (eg IOMMUFD_DEV_INVALIDATE)
Finally, we have the HW driven invalidation DMA queues that can be
directly assigned to the guest. AMD and SMMUv3+vCMDQ support this. In
this case the HW is directly processing invalidation commands without
a hypervisor trap.
To make this work the iommu needs to be programmed with:
AMD: A vDomain-ID -> pDomain-ID table
A vRID -> pRID table
This is all bound to some "virtual function"
ARM: A vRID -> pRID table
The vCMDQ is bound to a VM_ID, so to the Nesting Parent
For AMD, as above, I suggest the vDomain-ID be passed when creating
the nesting domain.
The AMD "virtual function".. It is probably best to create a new iommufd
object for this and it can be passed in to a few places
The vRID->pRID table should be some mostly common
IOMMUFD_DEV_ASSIGN_VIRTUAL_ID. AMD will need to pass in the virtual
function ID and ARM will need to pass in the Nesting Parent ID.
For the HW path some function will create the command queue and
DMA/mmap it. Taking in the virtual function/nesting parent as the
handle to associate it with.
For a SW path:
AMD: All invalidations can be delivered to the virtual function
and the kernel can use the vDomainID/vRID tables to translate
them fully
ARM: All invalidations can be delivered to the nesting parent
In many ways the nesting parent/virtual function are very similar
things. Perhaps ARM should also create a virtual function object which
is just welded to the nesting parent for API consistency.
So.. In short.. Invalidation is a PITA. The idea is the same but
annoying little details interfere with actually having a compltely
common API here. IMHO the uAPI in this series is fine. It will support
Intel invalidation and non-ATC invalidation on AMD/ARM. It should be
setup to allow that the target domain object can be any HWPT.
This HWPT is still nested domain. Is it? But it can represent a guest I/O
page table (VT-d), guest CD table (ARM), guest CR3 Table (AMD, it seems to
be a set of guest CR3 table pointers). May ARM and AMD guys keep me honest
here.
The Intel guest I/O page table case may be the simplest as userspace only
needs to provide the HWPT ID and the affected ranges for invalidating. As
mentioned above, kernel will find out the attached device/pasid and
invalidating cache with the device/pasid. For ARM and AMD case, extra
information is needed. Am I getting you correct?
ARM will be able to do IOTLB invalidation using this API.
IOMMUFD_DEV_INVALIDATE should be introduced with the same design as
HWPT invalidate. This would be used for AMD/ARM's ATC invalidation
(and just force the stream ID, userspace must direct the vRID to the
correct dev_id).
Then in yet another series we can tackle the entire "virtual function"
vRID/pRID translation stuff when the mmapable queue thing is
introduced.
Thus next steps:
- Respin this and lets focus on Intel only (this will be tough for
the holidays, but if it is available I will try)
I've respinned the iommufd cache invalidation part with the change to
report error_code/error_data per invalidation entry. yet still busy on
making Intel VTd part to report the error_code. Besides, I didn't see
other respin needed for Intel VT-d invalidation. If I missed thing, please
do let me know.:)
- Get an ARM patch that just does IOTLB invalidation and add it to my
part 3
- Start working on IOMMUFD_DEV_INVALIDATE along with an ARM
implementation of it
- Reorganize the AMD RFC broadly along these lines and lets see it
freshened up in the next months as well. I would like to see the
AMD support structured to implement the SW paths in first steps and
later add in the "virtual function" acceleration stuff. The latter
is going to be complex.
Jason
--
Regards,
Yi Liu
