Previous work on dev-iotlb message broke vhost on either SMMU or virtio-iommu since dev-iotlb (or PCIe ATS) is not yet supported for those archs.
An initial idea is that we can let IOMMU to export this information to vhost so that vhost would know whether the vIOMMU would support dev-iotlb, then vhost can conditionally register to dev-iotlb or the old iotlb way. We can work based on some previous patch to introduce PCIIOMMUOps as Yi Liu proposed [1]. However it's not as easy as I thought since vhost_iommu_region_add() does not have a PCIDevice context at all since it's completely a backend. It seems non-trivial to pass over a PCI device to the backend during init. E.g. when the IOMMU notifier registered hdev->vdev is still NULL. To make the fix smaller and easier, this patch goes the other way to leverage the flag_changed() hook of vIOMMUs so that SMMU and virtio-iommu can trap the dev-iotlb registration and fail it. Then vhost could try the fallback solution as using UNMAP invalidation for it's translations. [1] https://lore.kernel.org/qemu-devel/1599735398-6829-4-git-send-email-yi.l....@intel.com/ Reported-by: Eric Auger <eric.au...@redhat.com> Fixes: b68ba1ca57677acf870d5ab10579e6105c1f5338 Reviewed-by: Eric Auger <eric.au...@redhat.com> Tested-by: Eric Auger <eric.au...@redhat.com> Signed-off-by: Peter Xu <pet...@redhat.com> --- hw/arm/smmuv3.c | 5 +++++ hw/virtio/vhost.c | 13 +++++++++++-- hw/virtio/virtio-iommu.c | 5 +++++ 3 files changed, 21 insertions(+), 2 deletions(-) diff --git a/hw/arm/smmuv3.c b/hw/arm/smmuv3.c index 98b99d4fe8e..bd1f97000d9 100644 --- a/hw/arm/smmuv3.c +++ b/hw/arm/smmuv3.c @@ -1497,6 +1497,11 @@ static int smmuv3_notify_flag_changed(IOMMUMemoryRegion *iommu, SMMUv3State *s3 = sdev->smmu; SMMUState *s = &(s3->smmu_state); + if (new & IOMMU_NOTIFIER_DEVIOTLB_UNMAP) { + error_setg(errp, "SMMUv3 does not support dev-iotlb yet"); + return -EINVAL; + } + if (new & IOMMU_NOTIFIER_MAP) { error_setg(errp, "device %02x.%02x.%x requires iommu MAP notifier which is " diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c index 28c7d781721..6e17d631f77 100644 --- a/hw/virtio/vhost.c +++ b/hw/virtio/vhost.c @@ -704,6 +704,7 @@ static void vhost_iommu_region_add(MemoryListener *listener, Int128 end; int iommu_idx; IOMMUMemoryRegion *iommu_mr; + int ret; if (!memory_region_is_iommu(section->mr)) { return; @@ -726,8 +727,16 @@ static void vhost_iommu_region_add(MemoryListener *listener, iommu->iommu_offset = section->offset_within_address_space - section->offset_within_region; iommu->hdev = dev; - memory_region_register_iommu_notifier(section->mr, &iommu->n, - &error_fatal); + ret = memory_region_register_iommu_notifier(section->mr, &iommu->n, NULL); + if (ret) { + /* + * Some vIOMMUs do not support dev-iotlb yet. If so, try to use the + * UNMAP legacy message + */ + iommu->n.notifier_flags = IOMMU_NOTIFIER_UNMAP; + memory_region_register_iommu_notifier(section->mr, &iommu->n, + &error_fatal); + } QLIST_INSERT_HEAD(&dev->iommu_list, iommu, iommu_next); /* TODO: can replay help performance here? */ } diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c index 6b9ef7f6b2b..c2883a2f6c8 100644 --- a/hw/virtio/virtio-iommu.c +++ b/hw/virtio/virtio-iommu.c @@ -893,6 +893,11 @@ static int virtio_iommu_notify_flag_changed(IOMMUMemoryRegion *iommu_mr, IOMMUNotifierFlag new, Error **errp) { + if (new & IOMMU_NOTIFIER_DEVIOTLB_UNMAP) { + error_setg(errp, "Virtio-iommu does not support dev-iotlb yet"); + return -EINVAL; + } + if (old == IOMMU_NOTIFIER_NONE) { trace_virtio_iommu_notify_flag_add(iommu_mr->parent_obj.name); } else if (new == IOMMU_NOTIFIER_NONE) { -- 2.26.2
