On Fri, 2011-11-18 at 16:27 +0100, Joerg Roedel wrote: > On Fri, Nov 18, 2011 at 07:56:30AM -0700, Alex Williamson wrote: > > As you say, the interface is to tell userspace what devices the IOMMU > > can safely distinguish between. On x86, the IOMMU cannot distinguish > > DMA to the interrupt region between devices without interrupt remapping. > > Therefore, to only expose devices we can safely distinguish between, we > > shouldn't expose any devices by default if we don't provide this > > isolation. The interrupt remapping capability of the IOMMU isn't > > exposed to userspace, nor should userspace need to make the leap in > > understanding how MSIs are generated on a specific platform and whether > > interrupt remapping is required to provide isolation. > > But the isolation domains are different. The iommu-group interface is > about device isolation and interrupt remapping is about DMA address > space isolation (MSI range vs. the rest). The iommu can still > distinguish between devices when it can not distinguish between DMA and > MSI interrupts. I guess I fail to see the difference. We group devices behind certain bridges together because we can't distinguish DMA from those devices. MSI presents an address window across all devices for which we potentially can't distinguish between any of them. If instead of MSI the window was VGA and the IOMMU could differentiate all devices, so long as they don't do DMA to 0xa0000-0xc0000, what would we do? I think we'd have to exclude such an IOMMU from providing grouping information. > So what you want to do belongs to the VFIO driver or (as it is today) > into the KVM device assignment code and not into the iommu-group > interface. The trouble is that interrupt remapping closing a hole in DMA isolation is a platform issue. Is vfio supposed to know that on architecture foo we don't have such a hole and we don't need to look for interrupt remapping. Or maybe that platform bar solved it differently and we need to instead check flag MSI_OK. Current KVM doesn't care about this because it only does device assignment on x86. Another way to solve this might be to expose a more generic capability for the IOMMU. Instead of exposing "interrupt remapping support" it could expose "full DMA isolation", then amd_iommu/intel-iommu could set it based on interrupt remapping capability. The trouble is that then we expose grouping information for devices, but can't figure out the domain capabilities until we actually create a domain and attach those devices, which is kind of a crappy experience for a vfio user to find out they don't actually get to use those device. Instead, it seems like the device-group interface should only provide grouping information for devices with full DMA isolation, and hide the rest. Then we know that if iommu_device_group() returns an error, we can't safely use that device. Thanks, Alex -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
