On Fri, Sep 07, 2018 at 10:47:11AM +0800, Lu Baolu wrote: > > >>+ > >>+ intel_pasid_clear_entry(dev, pasid); > >>+ > >>+ if (!ecap_coherent(iommu->ecap)) { > >>+ pte = intel_pasid_get_entry(dev, pasid); > >>+ clflush_cache_range(pte, sizeof(*pte)); > >>+ } > >>+ > >>+ pasid_based_pasid_cache_invalidation(iommu, did, pasid); > >>+ pasid_based_iotlb_cache_invalidation(iommu, did, pasid); > >>+ > >>+ /* Device IOTLB doesn't need to be flushed in caching mode. */ > >>+ if (!cap_caching_mode(iommu->cap)) > >>+ pasid_based_dev_iotlb_cache_invalidation(iommu, dev, > >>pasid); > > > >can you elaborate, or point to any spec reference? > > > > In the driver, device iotlb doesn't get flushed in caching mode. I just > follow what have been done there. > > It also makes sense to me since only the bare metal host needs to > consider whether and how to flush the device iotlb. >
DavidW might remember, i think the idea was to help with cost of virtualization, we can avoid taking 2 exits vs handling it directly when we do iotlb flushing instead. the other optimization was to only do devtlb flushing when you unmap since when establish not-present to present there is no need to flush devtlb at that point. _______________________________________________ iommu mailing list iommu@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/iommu