On 12/10/2018 17:35, Michael S. Tsirkin wrote: >> + list_del(&req->list); >> + kfree(req); > > So with DEBUG set, this will actually free memory that device still > DMA's into. Hardly pretty. I think you want to mark device broken, > queue the request and then wait for device to be reset. Ok, let's remove DEBUG for the moment >> +static int viommu_probe(struct virtio_device *vdev) >> +{ >> + struct device *parent_dev = vdev->dev.parent; >> + struct viommu_dev *viommu = NULL; >> + struct device *dev = &vdev->dev; >> + u64 input_start = 0; >> + u64 input_end = -1UL; >> + int ret; >> + >> + if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) >> + return -ENODEV; > > I'm a bit confused about what will happen if this device > happens to be behind an iommu itself. > > If we can't handle that, should we clear PLATFORM_IOMMU > e.g. like the balloon does? I think the DMA API can handle this device doing DMA through another IOMMU. I haven't tested this case because it is very unusual (IOMMUs themselves generally access the physical address space) but I don't see anything preventing it. What we can't handle is a device performing DMA through two daisy-chained IOMMUs, but clearing PLATFORM_IOMMU on the first one wouldn't make things work in that case, we'd need some core changes. >> +struct virtio_iommu_config { >> + /* Supported page sizes */ >> + __u64 page_size_mask; >> + /* Supported IOVA range */ >> + struct virtio_iommu_range { > > I'd rather we moved the definition outside even though gcc allows it - > some old userspace compilers might not. > >> + __u64 start; >> + __u64 end; >> + } input_range; >> + /* Max domain ID size */ >> + __u8 domain_bits; > > Let's add explicit padding here as well? Ok Thanks, Jean