Per spec, the Consumer/Producer bit is defined only for Extended Address Space descriptors and should be ignored for QWord/DWord/Word Address Space descriptors. My understanding is that this is because x86 BIOSes didn't use the bit consistently, so it couldn't be relied upon. The Extended descriptors were added later and are presumably more reliable.
Linux currently looks at Consumer/Producer for all Address Space descriptors, although we don't use the result very much. The x86 and ia64 host bridge driver code doesn't use Consumer/Producer to identify windows; it assumes all descriptors are windows.
We do currently use Consumer/Producer to decide whether to apply _TRA. The change in these patches is to ignore Consumer/Producer for QWord/DWord/Word, so we'll apply _TRA from Consumer descriptors where we didn't before. Per spec, that should be safe because _TRA is required to be zero for Consumer resources.
If we want to describe host bridge register space and ECAM space directly in the PNP0A03 device on ARM64 (and I guess potentially even on x86 & ia64), I think we need changes like this so the arch code can use IORESOURCE_WINDOW to tell which descriptors are windows and which are registers.
This is a subtle area, so please take a look and let me know what you think.
These patches are based on v4.9-rc1. It's getting pretty late in the cycle, but I think we'd really like to get the ARM64 ACPI PCI story sorted out for the v4.10 merge window.
---
Bjorn Helgaas (2): ACPI: Combine acpi_dev_resource_address_space() and acpi_dev_resource_ext_address_space() ACPI: Ignore Consumer/Producer for QWord/DWord/Word Address Space
drivers/acpi/ioapic.c | 3 - drivers/acpi/resource.c | 130 ++++++++++++++++++---------------------- drivers/pnp/pnpacpi/rsparser.c | 3 - include/linux/acpi.h | 2 - 4 files changed, 59 insertions(+), 79 deletions(-)