On Mon, Aug 19, 2024 at 02:36:15PM -0300, Jason Gunthorpe wrote: > On Thu, Aug 15, 2024 at 05:50:06PM -0700, Nicolin Chen wrote: > > > Though only driver would know whether it would eventually access > > the vdev_id list, I'd like to keep things in the way of having a > > core-managed VIOMMU object (IOMMU_VIOMMU_TYPE_DEFAULT), so the > > viommu invalidation handler could have a lock at its top level to > > protect any potential access to the vdev_id list. > > It is a bit tortured to keep the xarray hidden. It would be better to > find a way to expose the right struct to the driver. Yes. I will try adding set/unset_vdev_id to the default viommu ops. > > > > @@ -3249,6 +3266,19 @@ arm_smmu_convert_user_cmd(struct arm_smmu_domain *s2_parent, > > > > cmd->cmd[0] &= ~CMDQ_TLBI_0_VMID; > > > > cmd->cmd[0] |= FIELD_PREP(CMDQ_TLBI_0_VMID, vmid); > > > > break; > > > > + case CMDQ_OP_ATC_INV: > > > > + case CMDQ_OP_CFGI_CD: > > > > + case CMDQ_OP_CFGI_CD_ALL: > > > > > > Oh, I didn't catch on that CD was needing this too.. :\ > > > > Well, viommu cache has a very wide range :) > > > > > That makes the other op much more useless than I expected. I really > > > wanted to break these two series apart. > > > > HWPT invalidate and VIOMMU invalidate are somewhat duplicated in > > both concept and implementation for SMMUv3. It's not a problem to > > have both but practically I can't think of the reason why VMM not > > simply stick to the wider VIOMMU invalidate uAPI alone.. > > > > > Maybe we need to drop the hwpt invalidation from the other series and > > > > Yea, the hwpt invalidate is just one patch in your series, it's > > easy to move if we want to. > > > > aim to merge this all together through the iommufd tree. > > > > I have been hoping for that, as you can see those driver patches > > are included here :) > > Well, this series has to go through iommufd of course > > I was hoping will could take the nesting enablement and we'd do the > viommu next window. > > But nesting enablment with out viommu is alot less useful than I had > thought :( Actually, without viommu, the hwpt cache invalidate alone could still support non-SVA case? Though we still have the blocker at the msi mapping... It still requires a solution, even for viommu series. Thanks Nicolin
