On Wed, Mar 03, 2021 at 01:27:34PM +0800, Lu Baolu wrote: > I have tested this framework with the Intel VT-d implementation. It > works as expected. Hence, > > Reviewed-by: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> > Tested-by: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx> Thanks! > One possible future optimization is that we could allow the system > administrators to choose between handle PRQs in a workqueue or handle > them synchronously. One research discovered that most of the software > latency of handling a single page fault exists in the schedule part. > Hence, synchronous processing will get shorter software latency if PRQs > are rare and limited. Yes, the risk is that processing a fault synchronously will take too much time, leading to PRI queue overflow if the IOMMU keeps receiving page faults. That's why I opted for the workqueue initially, but it's definitely something we can tweak Thanks, Jean
