[PATCH 0/4] ACPI: DMA ranges management

[Lorenzo Pieralisi]

On Fri, Jul 28, 2017 at 04:55:36PM +0100, Robin Murphy wrote:
> On 28/07/17 15:09, Lorenzo Pieralisi wrote:
> > On Fri, Jul 28, 2017 at 02:08:01PM +0100, Robin Murphy wrote:
> > 
> > [...]
> > 
> >>>>> To ensure that dma_set_mask() and friends actually respect _DMA, would
> >>>>> you consider introducing a dma_supported() callback to check the input
> >>>>> dma_mask against the FW defined limits? This would end up aggressively
> >>>>> clipping the dma_mask to 32-bits for devices like the above if the _DMA
> >>>>> limit was less than 64-bits, but that is probably preferable to the
> >>>>> controller accessing unintended addresses.
> >>>>>
> >>>>> Also, how would you feel about adding support for the IORT named_node
> >>>>> memory_address_limit field?
> >>>>
> >>>> We will certainly need that for some platform devices, so if you fancy
> >>>> giving it a go before Lorenzo or I get there, feel free!
> >>>
> >>> I can do it for v2 but I would like to understand why using _DMA is
> >>> not good enough for named components - having two bindings describing
> >>> the same thing is not ideal and I'd rather avoid it - if there is
> >>> a reason I am happy to add the necessary code.
> >>
> >> My interpretation of "_DMA is only defined under devices that represent
> >> buses." (ACPI 6.0, section 6.2.4) is that "devices that represent buses"
> >> are those that have other device objects as children.
> > 
> > Well if that was the case we would not be able to use _DMA for
> > eg PNP0A03 PCI host bridges that have no child ACPI devices, which
> > defeats the whole purpose of what I am doing.
> > 
> > The question here is what the _DMA object binding exactly means when
> > it refers to a "bus" and that's something I will figure out (and possibly
> > change) ASAP.
> > 
> >> In other words (excuse my novice pseudo-ASL), this would be valid:
> >>
> >> Scope(_SB)
> >> {
> >> 	Device (Bus)
> >> 	{
> >> 		...
> >> 		Method (_DMA ... )
> >> 		Device (Dev1)
> >> 		{
> >> 			...
> >> 		}
> >> 	}
> >> }
> >>
> >> but this should be invalid:
> >>
> >> Scope(_SB)
> >> {
> >> 	Device (Dev2)
> >> 	{
> >> 		...
> >> 		Method (_DMA ... )
> >> 	}
> >> }
> > 
> > Not sure about that (see above) and I agree that's what needs
> > clarification.
> > 
> >> Thus in the case where Dev2 is wired directly to an SMMU input, but
> >> fewer address bits are wired up between the two than both the device and
> >> SMMU interfaces are capable of, memory address limit is enough to
> >> describe that without having to insert a fake "bus" object above it just
> >> to hold the _DMA method.
> > 
> > BTW, how would you describe that in DT ? A "dma-ranges" property in the
> > device DT node right ? Arguably "dma-ranges" was not meant to be used
> > like that either ;-)
> 
> I believe that in real Open Firmware, the full PCI hierarchy is
> described in the device tree - I had assumed that ACPI expected the
> equivalent (i.e. the firmware probes PCI and assigns resources, so
> bridges/endpoints/etc. would be represented in the namespace with
> appropriate _CRS), thus the "bus with invisible children" case would
> only need to apply to DT. In terms of DTspec, it does not say that
> "dma-ranges" cannot be present on nodes without children, but that *is*
> implied of #address-cells and #size cells, so there does exist a similar
> ambiguity about what exactly counts as "a memory-mapped bus whose
> devicetree parent can be accessed from DMA operations originating from
> the bus". Certainly in the current Linux code, of_dma_configure()
> *doesn't* parse "dma-ranges" on leaf nodes (which is an open problem for
> some PCI host bridges in extant FDTs).
> 
> As for the case of straightforward interconnect widths/offsets (rather
> than potentially arbitrary windows), the 'fake bus' notion is already
> alive and well:
> 
> $ git grep 'soc {' arch/arm*/boot/dts
> 
> and the current "dma-ranges" users thankfully have consistent-enough
> topologies that they don't need to get much crazier than that.
> 
> (side note: up at the other end, I'm not entirely convinced that what I
> did for Juno is actually legal either)
> 
> > Long and short of it is: I do not like having two ways of describing
> > the same thing. I agree that the _DMA object usage requires
> > clarifications from a spec point of view but I want to do that before
> > plugging in code that may use bindings inconsistently.
> 
> I'd still argue that they are describing different things, just that one
> (the number of address bits wired up between a device and an SMMU)
> happens to be possible to describe as a subset of the other (an
> arbitrary mapping between two address spaces). The use-cases don't
> entirely overlap either - the information in _DMA is also likely to be
> wanted by non-ECAM PCI host controller drivers to configure their
> inbound windows, irrespective of anything to do with IOMMUs, whereas
> IORT code in hypervisors or other situations without a full ACPI
> namespace available may need to make decisions that the device memory
> address size limit is necessary for (well, that's the argument I've
> heard anyway).

Ok, I thought about it a bit more and I think the points you raised
should be tackled, certainly using _DMA objects on devices that
are not PCI host bridges (probably the only devices _DMA object was
devised to apply to) is moot, to say the least, whereas IORT address
limits are well defined for named components.

I will rework the code to use _DMA object (and its ranges) for PCI
devices, IORT named components address capabilities for anything else.

I will seek ACPI specs clarification anyway to make sure that the _DMA
usage is solid for PCI devices, if you have more comments just let me
know please.

Thanks,
Lorenzo