On 4/29/24 08:50, Zhenzhong Duan wrote:
Hi,
The most important change in this version is instroducing a common
HostIOMMUDeviceCaps structure in HostIOMMUDevice and a new interface
between vIOMMU and HostIOMMUDevice.
HostIOMMUDeviceClass::realize() is introduced to initialize
HostIOMMUDeviceCaps and other fields of HostIOMMUDevice variants.
HostIOMMUDeviceClass::check_cap() is introduced to query host IOMMU
device capabilities.
After the change, part2 is only 3 patches, so merge it with part1 to be
a single prerequisite series, same for changelog. If anyone doesn't like
that, I can split again.
The class tree is as below:
HostIOMMUDevice
| .caps
| .realize()
| .check_cap()
|
.-----------------------------------------------.
| | |
HostIOMMUDeviceLegacyVFIO {HostIOMMUDeviceLegacyVDPA} HostIOMMUDeviceIOMMUFD
| .vdev | {.vdev} | .iommufd
| .devid
| [.ioas_id]
| [.attach_hwpt()]
| [.detach_hwpt()]
|
.----------------------.
| |
HostIOMMUDeviceIOMMUFDVFIO {HostIOMMUDeviceIOMMUFDVDPA}
| .vdev | {.vdev}
* The attributes in [] will be implemented in nesting series.
* The classes in {} will be implemented in future.
* .vdev in different class points to different agent device,
* i.e., for VFIO it points to VFIODevice.
PATCH1-4: Introduce HostIOMMUDevice and its sub classes
PATCH5-11: Introduce HostIOMMUDeviceCaps, implement .realize() and .check_cap()
handler
PATCH12-16: Create HostIOMMUDevice instance and pass to vIOMMU
PATCH17-19: Implement compatibility check between host IOMMU and
vIOMMU(intel_iommu)
Qemu code can be found at:
https://github.com/yiliu1765/qemu/tree/zhenzhong/iommufd_nesting_preq_v3
Besides the compatibility check in this series, in nesting series, this
host IOMMU device is extended for much wider usage. For anyone interested
on the nesting series, here is the link:
https://github.com/yiliu1765/qemu/tree/zhenzhong/iommufd_nesting_rfcv2
v4 should be a good candidate, we will need feedback from the vIOMMU
maintainers though.
However, have you considered another/complementary approach which
would be to create an host IOMMU (iommufd) backend object and a vIOMMU
device object together for each vfio-pci device being plugged in the
machine ?
Something like,
-device pcie-root-port,port=23,chassis=8,id=pci.8,bus=pcie.0 \
-object iommufd,id=iommufd1 \
-device
intel-iommu,intremap=on,device-iotlb=on,caching-mode=on,iommufd=iommufd1 \
-device vfio-pci,host=0000:08:10.0,bus=pci.1,iommufd=iommufd0
The vIOMMU device would be linked to the host IOMMU (iommufd) backend
object at realize time and it would simplify the discovery of the host
IOMMU properties. The implementation would be more straight forward.
That said, I didn't study deeply what needs to be done. The vIOMMU
implementation is not ready yet to support multiple instances and some
massaging is needed to change that first.
Thanks,
C.
