On Thu, Jan 8, 2026 at 4:54 PM Jason Gunthorpe <[email protected]> wrote: > > On Fri, Jan 09, 2026 at 12:43:32AM +0000, David Matlack wrote: > > On 2026-01-08 08:36 PM, Jason Gunthorpe wrote: > > > On Thu, Jan 08, 2026 at 09:29:29PM +0000, David Matlack wrote: > > > > On 2026-01-08 02:33 PM, Jason Gunthorpe wrote: > > > > > On Thu, Jan 08, 2026 at 10:24:19AM -0800, David Matlack wrote: > > > > > > > > Oh, I was thinking about a compatability only flow only in the > > > > > > > > type 1 > > > > > > > > emulation that internally magically converts a VMA to a dmabuf, > > > > > > > > but I > > > > > > > > haven't written anything.. It is a bit tricky and the type 1 > > > > > > > > emulation > > > > > > > > has not been as popular as I expected?? > > > > > > > > > > > > > > In part because of this gap, I'd guess. Thanks, > > > > > > > > > > > > Lack of huge mappings in the IOMMU when using VFIO_TYPE1_IOMMU is > > > > > > another gap I'm aware of. > > > > > > vfio_dma_mapping_test.vfio_type1_iommu_anonymous_hugetlb_1gb.dma_map_unmap > > > > > > fails when IOMMUFD_VFIO_CONTAINER is enabled. > > > > > > > > > > What is this? I'm not aware of it.. > > > > > > > > It's one of the test cases within > > > > tools/testing/selftests/vfio/vfio_dma_mapping_test.c. > > > > > > > > Here's the output when running with CONFIG_IOMMUFD_VFIO_CONTAINER=y: > > > > > > > > # RUN > > > > vfio_dma_mapping_test.vfio_type1_iommu_anonymous_hugetlb_1gb.dma_map_unmap > > > > ... > > > > Mapped HVA 0x7f0480000000 (size 0x40000000) at IOVA 0x0 > > > > Searching for IOVA 0x0 in > > > > /sys/kernel/debug/iommu/intel/0000:6a:01.0/domain_translation_struct > > > > Found IOMMU mappings for IOVA 0x0: > > > > PGD: 0x0000000203475027 > > > > P4D: 0x0000000203476027 > > > > PUD: 0x0000000203477027 > > > > PMD: 0x00000001e7562027 > > > > PTE: 0x00000041c0000067 > > > > # > > > > tools/testing/selftests/vfio/vfio_dma_mapping_test.c:188:dma_map_unmap:Expected > > > > 0 (0) == mapping.pte (282394099815) > > > > # dma_map_unmap: Test terminated by assertion > > > > # FAIL > > > > vfio_dma_mapping_test.vfio_type1_iommu_anonymous_hugetlb_1gb.dma_map_unmap > > > > > > I can't think of any reason this would fail, I think your tests have > > > found a real bug?? Can you check into it, what kernel call fails and > > > where does the kernel code come from? > > > > Oh I thought it was by design. This code in iommufd_vfio_set_iommu(): > > > > /* > > * The difference between TYPE1 and TYPE1v2 is the ability to unmap in > > * the middle of mapped ranges. This is complicated by huge page > > support > > * which creates single large IOPTEs that cannot be split by the iommu > > * driver. TYPE1 is very old at this point and likely nothing uses it, > > * however it is simple enough to emulate by simply disabling the > > * problematic large IOPTEs. Then we can safely unmap within any > > range. > > */ > > if (type == VFIO_TYPE1_IOMMU) > > rc = iopt_disable_large_pages(&ioas->iopt); > > > > git-blame says some guy named Jason Gunthorpe wrote it :P > > Er, maybe I mis understood the output then? > > This is not a "failure" though, the map succeeded and gave a small > page mapping. > > This is not reflecting a bug in iommufd but a bug in the TYPE1 support > in VFIO itself because it definitely cannot maintain the required > unmap anywhere semantic if it mapped in a 1G huge page like this. > > Basically, if you are mapping with TYPE1 mode then this should be triggered: > > if (!strcmp(variant->iommu_mode, "iommufd_compat_type1")) > mapping_size = SZ_4K; > > And VFIO should be the one to fail, not iommufd. > > If you really want to test TYPE1 you need to test what makes it > unique, which is that you can map any VMA and then unmap any slice of > it. Including within what should otherwise be a 1G page. > > But I doubt anyone cares enough to fix this, so just exclude > VFIO_TYPE1_IOMMU from this test?
Ah, ok, thanks for the explanation. So VFIO_TYPE1_IOMMU should always use 4K mappings regardless of backend (VFIO or iommufd) so that unmap can work as intended. I think excluding VFIO_TYPE1_IOMMU from this assertion makes sense if we don't care about fixing it.
