2024年5月16日(木) 21:59 Manivannan Sadhasivam <mani@xxxxxxxxxx>: > > On Thu, May 16, 2024 at 01:38:40PM +0900, Shunsuke Mie wrote: > > Hi virtio folks, > > > > You forgot to CC the actual Virtio folks. I've CCed them now. Oops. thank you. > > I'm writing to discuss finding a workaround with Virtio drivers and legacy > > devices with limited memory access. > > > > # Background > > The Virtio specification defines a feature (VIRTIO_F_ACCESS_PLATFORM) to > > indicate devices requiring restricted memory access or IOMMU translation. This > > feature bit resides at position 33 in the 64-bit Features register on modern > > interfaces. When the linux virtio driver finds the flag, the driver uses DMA > > API that handles to use of appropriate memory. > > > > # Problem > > However, legacy devices only have a 32-bit register for the features bits. > > Consequently, these devices cannot represent the ACCESS_PLATFORM bit. As a > > result, legacy devices with restricted memory access cannot function > > properly[1]. This is a legacy spec issue, but I'd like to find a workaround. > > > > # Proposed Solutions > > I know these are not ideal, but I propose the following solution. > > Driver-side: > > - Implement special handling similar to xen_domain. > > In xen_domain, linux virtio driver enables to use the DMA API. > > - Introduce a CONFIG option to adjust the DMA API behavior. > > Device-side: > > Due to indistinguishability from the guest's perspective, a device-side > > solution might be difficult. > > > > I'm open to any comments or suggestions you may have on this issue or > > alternative approaches. > > > > [1] virtio-net PCI endpoint function using PCIe Endpoint Framework, > > https://lore.kernel.org/lkml/54ee46c3-c845-3df3-8ba0-0ee79a2acab1@xxxxxxxxxx/t/ > > The Linux PCIe endpoint framework is used to implement the virtio-net device on > > a legacy interface. This is necessary because of the framework and hardware > > limitation. > > > > We can fix the endpoint framework limitation, but the problem lies with some > platforms where we cannot write to vendor capability registers and still have > IOMMU. I agree, this is a problem caused by the inability to set the capability. I'm not sure, but are there any chips that support this? Also, I wasn't aware of the IOMMU issue. I thought that if the Linux DMA subsystem could handle IOMMU properly, there wouldn't be any problems. Is that incorrect? Shunsuke, Best > - Mani > > -- > மணிவண்ணன் சதாசிவம்
