Re: [PATCH v8 4/5] iommufd: Extend IOMMU_GET_HW_INFO to report PASID capability

[Nicolin Chen <nicolinc@nvidia.com>]

On Thu, Mar 20, 2025 at 08:48:49PM +0800, Yi Liu wrote:
> On 2025/3/20 01:58, Nicolin Chen wrote:
> > On Thu, Mar 13, 2025 at 05:47:52AM -0700, Yi Liu wrote:
> > > PASID usage requires PASID support in both device and IOMMU. Since the
> > > iommu drivers always enable the PASID capability for the device if it
> > > is supported, this extends the IOMMU_GET_HW_INFO to report the PASID
> > > capability to userspace. Also, enhances the selftest accordingly.
> > 
> > Overall, I am a bit confused by the out_capabilities field in the
> > IOMMU_GET_HW_INFO. Why these capabilities cannot be reported via
> > the driver specific data structure?
> > 
> > E.g. we don't report the "NESTING" capability in out_capabilities
> > at all and that works fine since the iommu driver would reject if
> > it doesn't support.
> 
> NESTING is a bit different. Userspace needs to know underlying PASID
> cap and further expose it to guest if it wants. While NESTING is not
> from this angle. It's just for the use of userspace. So a try and fail
> is ok.

Hmm, would you please elaborate the difference more?

Also, what's that "further expose"?

> > Mind elaborate the reason for these two new capabilities? What is
> > the benefit from keeping them in the core v.s. driver level?
> 
> I view the PASID cap is generic just like the DIRTY_TRACKING cap.

Well, I actually don't get the DIRTY_TRACKING cap either. Based on
what I see from Zhenzhong's implementation in the QEMU, it totally
could be a vendor-specific cap: we can just add another PCIIOMMUOps
op for QEMU core to ask the IOMMU device model to get hw_info and
allocate a DIRTY_TRACKING-enabled HWPT to return that to the core,
instead of core detecting the out_capabilities in the common path.

In that regard, honestly, I don't quite get this out_capabilities.

> Reporting them in the driver-specific part is fine, but I doubt
> if it is better since PASID cap is generic to all vendors.

The argument could be that NESTING is generic to all vendors too,
yet we don't have that in out_capabilities :-/

Thanks
Nicolin