Message ID | 20230913080423.523953-1-eric.auger@redhat.com (mailing list archive) |
---|---|
Headers | show |
Series | VIRTIO-IOMMU/VFIO: Don't assume 64b IOVA space | expand |
The original issue I found : After starting a VM which has two ice PFs and a virtio-iommu device, qemu-kvm and VM guest dmesg throw lots of duplicate VFIO_MAP_DMA errors After testing with Eric's build, the original issue is gone and the Tier1 regression test against ice PF and virtio iommu device gets PASS as well. Tested-by: Yanghang Liu <yanghliu@redhat.com> On Wed, Sep 13, 2023 at 4:06 PM Eric Auger <eric.auger@redhat.com> wrote: > > On x86, when assigning VFIO-PCI devices protected with virtio-iommu > we encounter the case where the guest tries to map IOVAs beyond 48b > whereas the physical VTD IOMMU only supports 48b. This ends up with > VFIO_MAP_DMA failures at qemu level because at kernel level, > vfio_iommu_iova_dma_valid() check returns false on vfio_map_do_map(). > > This is due to the fact the virtio-iommu currently unconditionally > exposes an IOVA range of 64b through its config input range fields. > > This series removes this assumption by retrieving the usable IOVA > regions through the VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE UAPI when > a VFIO device is attached. This info is communicated to the > virtio-iommu memory region, transformed into the inversed info, ie. > the host reserved IOVA regions. Then those latter are combined with the > reserved IOVA regions set though the virtio-iommu reserved-regions > property. That way, the guest virtio-iommu driver, unchanged, is > able to probe the whole set of reserved regions and prevent any IOVA > belonging to those ranges from beeing used, achieving the original goal. > > Best Regards > > Eric > > This series can be found at: > https://github.com/eauger/qemu/tree/virtio-iommu_geometry_v2 > > History: > v1 -> v2: > - Remove "[PATCH 12/13] virtio-iommu: Resize memory region according > to the max iova info" which causes way too much trouble: trigger > a coredump in vhost, causes duplication of IOMMU notifiers causing > EEXIST vfio_dma_map errors, ... This looks like a bad usage of the > memory API so I prefer removing this from this series. So I was > also obliged to remove the vfio_find_hostwin() check in the case > of an IOMMU MR. > - Let range_inverse_array() take low/high args instead of hardcoding > 0, UINT64_MAX which both complexifies the algo and the tests. > - Move range function description in header. > - Check that if set_iova_ranges is called several times, new resv > regions are included in previous ones > > Eric Auger (12): > memory: Let ReservedRegion use Range > memory: Introduce memory_region_iommu_set_iova_ranges > vfio: Collect container iova range info > virtio-iommu: Rename reserved_regions into prop_resv_regions > virtio-iommu: Introduce per IOMMUDevice reserved regions > range: Introduce range_inverse_array() > virtio-iommu: Implement set_iova_ranges() callback > range: Make range_compare() public > util/reserved-region: Add new ReservedRegion helpers > virtio-iommu: Consolidate host reserved regions and property set ones > test: Add some tests for range and resv-mem helpers > vfio: Remove 64-bit IOVA address space assumption > > include/exec/memory.h | 30 +++- > include/hw/vfio/vfio-common.h | 2 + > include/hw/virtio/virtio-iommu.h | 7 +- > include/qemu/range.h | 14 ++ > include/qemu/reserved-region.h | 32 ++++ > hw/core/qdev-properties-system.c | 9 +- > hw/vfio/common.c | 70 +++++++-- > hw/virtio/virtio-iommu-pci.c | 8 +- > hw/virtio/virtio-iommu.c | 110 ++++++++++++-- > softmmu/memory.c | 15 ++ > tests/unit/test-resv-mem.c | 251 +++++++++++++++++++++++++++++++ > util/range.c | 51 ++++++- > util/reserved-region.c | 94 ++++++++++++ > hw/virtio/trace-events | 1 + > tests/unit/meson.build | 1 + > util/meson.build | 1 + > 16 files changed, 655 insertions(+), 41 deletions(-) > create mode 100644 include/qemu/reserved-region.h > create mode 100644 tests/unit/test-resv-mem.c > create mode 100644 util/reserved-region.c > > -- > 2.41.0 > >