On Mon, 25 Jan 2021 17:03:58 +0800 Shenming Lu <lushenming@xxxxxxxxxx> wrote: > Hi, > > The static pinning and mapping problem in VFIO and possible solutions > have been discussed a lot [1, 2]. One of the solutions is to add I/O > page fault support for VFIO devices. Different from those relatively > complicated software approaches such as presenting a vIOMMU that provides > the DMA buffer information (might include para-virtualized optimizations), > IOPF mainly depends on the hardware faulting capability, such as the PCIe > PRI extension or Arm SMMU stall model. What's more, the IOPF support in > the IOMMU driver is being implemented in SVA [3]. So do we consider to > add IOPF support for VFIO passthrough based on the IOPF part of SVA at > present? > > We have implemented a basic demo only for one stage of translation (GPA > -> HPA in virtualization, note that it can be configured at either stage), > and tested on Hisilicon Kunpeng920 board. The nested mode is more complicated > since VFIO only handles the second stage page faults (same as the non-nested > case), while the first stage page faults need to be further delivered to > the guest, which is being implemented in [4] on ARM. My thought on this > is to report the page faults to VFIO regardless of the occured stage (try > to carry the stage information), and handle respectively according to the > configured mode in VFIO. Or the IOMMU driver might evolve to support more... > > Might TODO: > - Optimize the faulting path, and measure the performance (it might still > be a big issue). > - Add support for PRI. > - Add a MMU notifier to avoid pinning. > - Add support for the nested mode. > ... > > Any comments and suggestions are very welcome. :-) I expect performance to be pretty bad here, the lookup involved per fault is excessive. There are cases where a user is not going to be willing to have a slow ramp up of performance for their devices as they fault in pages, so we might need to considering making this configurable through the vfio interface. Our page mapping also only grows here, should mappings expire or do we need a least recently mapped tracker to avoid exceeding the user's locked memory limit? How does a user know what to set for a locked memory limit? The behavior here would lead to cases where an idle system might be ok, but as soon as load increases with more inflight DMA, we start seeing "unpredictable" I/O faults from the user perspective. Seems like there are lots of outstanding considerations and I'd also like to hear from the SVA folks about how this meshes with their work. Thanks, Alex
