On Thu, May 29, 2025 at 06:46:20AM +0000, Duan, Zhenzhong wrote: > >Looking at the kernel iommufd_hwpt_invalidate() routine and > >intel_nested_cache_invalidate_user(), it doesn't seem possible to > >return a different number of cache entries. Are you anticipating > >other implementations (sMMU) ? > > Yes, same for sMMU's arm_vsmmu_cache_invalidate() and selftest's > mock_viommu_cache_invalidate() and mock_domain_cache_invalidate_user(). > > I'm not sure if this should apply to all types of IOMMUs, uAPI doc doesn't > talk about it.
It should. The uAPI defines that at entry_num: * struct iommu_hwpt_invalidate - ioctl(IOMMU_HWPT_INVALIDATE) ... * @entry_num: Input the number of cache invalidation requests in the array. * Output the number of requests successfully handled by kernel. This applies to either ret != 0 case too. > @Liu, Yi L, @nicol...@nvidia.com, @Jason Gunthorpe, should I treat ret = 0 > and total_entries != cache.entry_num as a kernel bug or not? Selftest has that coverage, so it would be a kernel bug that will unlikely occur. That being said, it doesn't hurt to do that IMHO. Thanks Nicolin
