On Thu, Mar 07, 2024 at 09:54:53AM +0800, Baolu Lu wrote:

On 2024/3/7 0:01, Jason Gunthorpe wrote:

quoted

On Wed, Mar 06, 2024 at 11:15:50PM +0800, Zhangfei Gao wrote:

quoted

Double checked, this does not send flags, 0 is OK,
Only domain_alloc_user in iommufd_hwpt_paging_alloc requires flags.

In my debug, I need this patch, otherwise NULL pointer errors happen
since SVA is not set.

This is some driver bug, we need to get rid of these
iommu_dev_enable_feature() requirements.

Yes. Especially iopf should be independent of SVA.

The problem in the arm smmu v3 driver is that enabling iopf is actually
done in the enabling SVA path, while the enabling iopf path does nothing
except for some checks. It doesn't matter if iopf is tied with SVA, but
when it comes to delivering iopf to user space, we need to decouple it.

Yes. Each driver is going to need to get a to a NOP for the feature
things then we can remove them.

I did half the ARM driver in my part 2, the iopf bit still needs
doing there.

Jason