| Message ID | 20231011175516.541374-1-eric.auger@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | VIRTIO-IOMMU/VFIO: Don't assume 64b IOVA space | expand |
On Wed, Oct 11, 2023 at 07:52:16PM +0200, Eric Auger wrote: > This applies on top of vfio-next: > https://github.com/legoater/qemu/, vfio-next branch virtio things make sense Reviewed-by: Michael S. Tsirkin <mst@redhat.com> let me know how you want to merge all this. > 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/vfio-next-iommu_geometry-v3 > > History: > v2 -> v3: > - rebase on top of vfio-next (including iommufd prereq) > - take into account IOVA range info capability may not be offered by > old kernel and use nr_iovas = -1 to encode that [Alex] > - use GList * everywhere instead of arrays (in the range_inverse_array) > with the benefice it sorts ranges retrieved from the kernel which are > not garanteed to be sorted. Rework the tests accordingly [Alex] > - Make sure resv_regions GList is build before the probe() [Jean] > per device list is first populated with prop resv regions on > IOMMUDevice creation and then rebuilt on set_iova() > - Add a warning if set_iova builds a valid list after probe was > called [Jean] > - Build host windows on top of IOVA valid ranges if this info can > be retrieved from the kernel. As many windows are created as > valid ranges > 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. > - 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 (13): > 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 > range: Make range_compare() public > util/reserved-region: Add new ReservedRegion helpers > virtio-iommu: Introduce per IOMMUDevice reserved regions > range: Introduce range_inverse_array() > virtio-iommu: Record whether a probe request has been issued > virtio-iommu: Implement set_iova_ranges() callback > 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 | 34 +++- > 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 | 23 ++- > hw/vfio/container.c | 67 ++++++- > hw/virtio/virtio-iommu-pci.c | 8 +- > hw/virtio/virtio-iommu.c | 155 +++++++++++++-- > system/memory.c | 13 ++ > tests/unit/test-resv-mem.c | 318 +++++++++++++++++++++++++++++++ > util/range.c | 61 +++++- > util/reserved-region.c | 91 +++++++++ > hw/virtio/trace-events | 1 + > tests/unit/meson.build | 1 + > util/meson.build | 1 + > 17 files changed, 791 insertions(+), 46 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
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, Oct 18, 2023 at 9:45 PM Michael S. Tsirkin <mst@redhat.com> wrote: > > On Wed, Oct 11, 2023 at 07:52:16PM +0200, Eric Auger wrote: > > This applies on top of vfio-next: > > https://github.com/legoater/qemu/, vfio-next branch > > virtio things make sense > > Reviewed-by: Michael S. Tsirkin <mst@redhat.com> > > let me know how you want to merge all this. > > > > > 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/vfio-next-iommu_geometry-v3 > > > > History: > > v2 -> v3: > > - rebase on top of vfio-next (including iommufd prereq) > > - take into account IOVA range info capability may not be offered by > > old kernel and use nr_iovas = -1 to encode that [Alex] > > - use GList * everywhere instead of arrays (in the range_inverse_array) > > with the benefice it sorts ranges retrieved from the kernel which are > > not garanteed to be sorted. Rework the tests accordingly [Alex] > > - Make sure resv_regions GList is build before the probe() [Jean] > > per device list is first populated with prop resv regions on > > IOMMUDevice creation and then rebuilt on set_iova() > > - Add a warning if set_iova builds a valid list after probe was > > called [Jean] > > - Build host windows on top of IOVA valid ranges if this info can > > be retrieved from the kernel. As many windows are created as > > valid ranges > > 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. > > - 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 (13): > > 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 > > range: Make range_compare() public > > util/reserved-region: Add new ReservedRegion helpers > > virtio-iommu: Introduce per IOMMUDevice reserved regions > > range: Introduce range_inverse_array() > > virtio-iommu: Record whether a probe request has been issued > > virtio-iommu: Implement set_iova_ranges() callback > > 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 | 34 +++- > > 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 | 23 ++- > > hw/vfio/container.c | 67 ++++++- > > hw/virtio/virtio-iommu-pci.c | 8 +- > > hw/virtio/virtio-iommu.c | 155 +++++++++++++-- > > system/memory.c | 13 ++ > > tests/unit/test-resv-mem.c | 318 +++++++++++++++++++++++++++++++ > > util/range.c | 61 +++++- > > util/reserved-region.c | 91 +++++++++ > > hw/virtio/trace-events | 1 + > > tests/unit/meson.build | 1 + > > util/meson.build | 1 + > > 17 files changed, 791 insertions(+), 46 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 > >
Hi Yanghang, On 10/19/23 11:07, YangHang Liu wrote: > 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> Thank you for testing! Eric > > > On Wed, Oct 18, 2023 at 9:45 PM Michael S. Tsirkin <mst@redhat.com> wrote: >> On Wed, Oct 11, 2023 at 07:52:16PM +0200, Eric Auger wrote: >>> This applies on top of vfio-next: >>> https://github.com/legoater/qemu/, vfio-next branch >> virtio things make sense >> >> Reviewed-by: Michael S. Tsirkin <mst@redhat.com> >> >> let me know how you want to merge all this. >> >> >> >>> 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/vfio-next-iommu_geometry-v3 >>> >>> History: >>> v2 -> v3: >>> - rebase on top of vfio-next (including iommufd prereq) >>> - take into account IOVA range info capability may not be offered by >>> old kernel and use nr_iovas = -1 to encode that [Alex] >>> - use GList * everywhere instead of arrays (in the range_inverse_array) >>> with the benefice it sorts ranges retrieved from the kernel which are >>> not garanteed to be sorted. Rework the tests accordingly [Alex] >>> - Make sure resv_regions GList is build before the probe() [Jean] >>> per device list is first populated with prop resv regions on >>> IOMMUDevice creation and then rebuilt on set_iova() >>> - Add a warning if set_iova builds a valid list after probe was >>> called [Jean] >>> - Build host windows on top of IOVA valid ranges if this info can >>> be retrieved from the kernel. As many windows are created as >>> valid ranges >>> 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. >>> - 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 (13): >>> 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 >>> range: Make range_compare() public >>> util/reserved-region: Add new ReservedRegion helpers >>> virtio-iommu: Introduce per IOMMUDevice reserved regions >>> range: Introduce range_inverse_array() >>> virtio-iommu: Record whether a probe request has been issued >>> virtio-iommu: Implement set_iova_ranges() callback >>> 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 | 34 +++- >>> 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 | 23 ++- >>> hw/vfio/container.c | 67 ++++++- >>> hw/virtio/virtio-iommu-pci.c | 8 +- >>> hw/virtio/virtio-iommu.c | 155 +++++++++++++-- >>> system/memory.c | 13 ++ >>> tests/unit/test-resv-mem.c | 318 +++++++++++++++++++++++++++++++ >>> util/range.c | 61 +++++- >>> util/reserved-region.c | 91 +++++++++ >>> hw/virtio/trace-events | 1 + >>> tests/unit/meson.build | 1 + >>> util/meson.build | 1 + >>> 17 files changed, 791 insertions(+), 46 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 >>
On 10/18/23 15:37, Michael S. Tsirkin wrote: > On Wed, Oct 11, 2023 at 07:52:16PM +0200, Eric Auger wrote: >> This applies on top of vfio-next: >> https://github.com/legoater/qemu/, vfio-next branch > > virtio things make sense > > Reviewed-by: Michael S. Tsirkin <mst@redhat.com> > > let me know how you want to merge all this. Michael, I will grab the series if that's OK. Thanks, C.
On Thu, Oct 19, 2023 at 01:07:41PM +0200, Cédric Le Goater wrote: > On 10/18/23 15:37, Michael S. Tsirkin wrote: > > On Wed, Oct 11, 2023 at 07:52:16PM +0200, Eric Auger wrote: > > > This applies on top of vfio-next: > > > https://github.com/legoater/qemu/, vfio-next branch > > > > virtio things make sense > > > > Reviewed-by: Michael S. Tsirkin <mst@redhat.com> > > > > let me know how you want to merge all this. > > Michael, > > I will grab the series if that's OK. fine by me > Thanks, > > C.
Hi Cédric, On 10/19/23 13:07, Cédric Le Goater wrote: > On 10/18/23 15:37, Michael S. Tsirkin wrote: >> On Wed, Oct 11, 2023 at 07:52:16PM +0200, Eric Auger wrote: >>> This applies on top of vfio-next: >>> https://github.com/legoater/qemu/, vfio-next branch >> >> virtio things make sense >> >> Reviewed-by: Michael S. Tsirkin <mst@redhat.com> >> >> let me know how you want to merge all this. > > Michael, > > I will grab the series if that's OK. I have just sent a v4 taking into account Alex' suggestions, collecting Michael's and Alex' R-b, and YangHang's T-b. This should be ready to go if you don't have any other comments and if this survives your non regression tests ;-) Eric > > Thanks, > > C. >
Hello Eric, On 10/19/23 15:51, Eric Auger wrote: > Hi Cédric, > > On 10/19/23 13:07, Cédric Le Goater wrote: >> On 10/18/23 15:37, Michael S. Tsirkin wrote: >>> On Wed, Oct 11, 2023 at 07:52:16PM +0200, Eric Auger wrote: >>>> This applies on top of vfio-next: >>>> https://github.com/legoater/qemu/, vfio-next branch >>> >>> virtio things make sense >>> >>> Reviewed-by: Michael S. Tsirkin <mst@redhat.com> >>> >>> let me know how you want to merge all this. >> >> Michael, >> >> I will grab the series if that's OK. > > I have just sent a v4 taking into account Alex' suggestions, collecting > Michael's and Alex' R-b, and YangHang's T-b. > This should be ready to go if you don't have any other comments and if > this survives your non regression tests ;-) Sure. I did a simple fix in patch 2 for a documentation breakage. No need to resend. I should apply in the morning. Cheers, C.