On Wed, May 24, 2017 at 07:01:38PM +0100, Jean-Philippe Brucker wrote: > Address Translation Service (ATS) is an extension to PCIe allowing > endpoints to manage their own IOTLB, called Address Translation Cache > (ATC). Instead of having every memory transaction processed by the IOMMU, > the endpoint can first send an Address Translation Requests for an IOVA, > obtain the corresponding Physical Address from the IOMMU and store it in > its ATC. Subsequent transactions to this memory region can be performed on > the PA, in which case they are marked 'translated' and (partially) bypass > the IOMMU. > > Since the extension uses fields that were previously reserved in the > PCIe Translation Layer Packet, it seems ill-advised to enabled it on a > system that doesn't fully support ATS. > > To "old" root complexes that simply ignored the new AT field, an Address > Translation Request will look exactly like a Memory Read Request, so the > root bridge will forward a memory read to the IOMMU instead of a > translation request. If the access succeeds, the RC will send a Read > Completion, which looks like a Translation Completion, back to the > endpoint. As a result the endpoint might end up storing the content of > memory instead of a physical address in its ATC. In reality, it's more > likely that the size fields will be invalid and either end will detect the > error, but in any case, it is undesirable. > > Add a way for firmware to tell the OS that ATS is supported by the PCI > root complex. Can't firmware have already enabled ATS? Often for things like this, not present means "use firmware setting". > > Signed-off-by: Jean-Philippe Brucker <jean-philippe.brucker@xxxxxxx> > --- > Documentation/devicetree/bindings/pci/pci-iommu.txt | 8 ++++++++ > 1 file changed, 8 insertions(+) > > diff --git a/Documentation/devicetree/bindings/pci/pci-iommu.txt b/Documentation/devicetree/bindings/pci/pci-iommu.txt > index 0def586fdcdf..f21a68ec471a 100644 > --- a/Documentation/devicetree/bindings/pci/pci-iommu.txt > +++ b/Documentation/devicetree/bindings/pci/pci-iommu.txt > @@ -44,6 +44,14 @@ Optional properties > - iommu-map-mask: A mask to be applied to each Requester ID prior to being > mapped to an IOMMU specifier per the iommu-map property. > > +- ats-supported: if present, the root complex supports the Address > + Translation Service (ATS). It is able to interpret the AT field in PCIe > + Transaction Layer Packets, and forward Translation Completions or > + Invalidation Requests to endpoints. Why can't this be based on the compatible strings? Rob
