On 6/17/24 3:25 AM, Duan, Zhenzhong wrote:
Hi Cédric,
-----Original Message-----
From: Cédric Le Goater <c...@redhat.com>
Sent: Friday, June 14, 2024 6:05 PM
To: eric.au...@redhat.com; eric.auger....@gmail.com; qemu-
de...@nongnu.org; qemu-...@nongnu.org; m...@redhat.com; jean-
phili...@linaro.org; peter.mayd...@linaro.org; yangh...@redhat.com; Duan,
Zhenzhong <zhenzhong.d...@intel.com>
Cc: alex.william...@redhat.com; jasow...@redhat.com;
pbonz...@redhat.com; berra...@redhat.com
Subject: Re: [PATCH v3 1/7] HostIOMMUDevice: Store the VFIO/VDPA agent
Talking of which, why are we passing a 'VFIODevice *' parameter to
HostIOMMUDeviceClass::realize ? I don't see a good reason
I think a 'VFIOContainerBase *' would be more appropriate since
'HostIOMMUDevice' represents a device on the host which is common
to all VFIO devices.
In that case, HostIOMMUDevice::agent wouldn't need to be opaque
anymore. It could simply be a 'VFIOContainerBase *' and
hiod_legacy_vfio_get_iova_ranges() in patch 3 would grab the
'iova_ranges' from the 'VFIOContainerBase *' directly.
This means some rework :
* vfio_device_get_aw_bits() would use a 'VFIOContainerBase *' instead.
* HostIOMMUDevice::name would be removed. This is just for error
messages.
* hiod_iommufd_vfio_realize() would use VFIOIOMMUFDContainer::be.
That said, I think we need the QOMification changes first.
OK I need to review your series first. At the moment I have just
addressed Zhenzhong's comment in v4, just sent.
Yep. Just take a look at mine. If both of you agree with above
proposal, I can care of it and resend all 3. It's a small change.
I would suggest using opaque pointer and VFIODevice for two reasons,
1. in nesting series vIOMMU needs to attach/detaching hwpt which is VFIODevice
operations.
See
https://github.com/yiliu1765/qemu/commit/3ca559d35adc9840555e361a56708af4c6338b3d
OK. I realized later on that we needed the device id anyhow.
2. If we plan to support VDPA Device in future, the opaque pointer can also
point
to an VDPADevice structure.
Thanks,
C.
