On Fri, 7 Jul 2023 at 18:21, Dong Wei <Dong.Wei@xxxxxxx> wrote: > > On Arm systems today, the ACPI RSDP is found using the UEFI Configuration Table. This is true for all Arm SystemReady compliant systems: 1) SystemReady LS: LBBRv1 is using a minimal UEFI FW to load LinuxBoot, that minimal UEFI FW is producing the UEFI Configuration Table. We are working on LBBRv2. LBBRv2 is based on Coreboot loading LinuxBoot. But we do not have a way today to get CoreBoot to produce this pointer to ACPI RSDP. Arm does not support x86 E820 BIOS interface. 2) SystemReady IR: this solution uses DT rather than ACPI. 3) SystemReady ES: this solution can use UBoot or EDK2, and it requires ACPI. Since both UBoot and EDK2 support UEFI now, so ACPI RSDP can be found using the UEFI Configuration Table. 4) SystemReady SR: this solution typically uses EDK2 and requires ACPI, so no issue finding RSDP via UEFI Configuration Table. > > > > So the ACPI RSDP issue only exist if we want to remove the minimum UEFI FW and go to CoreBoot completely to load LinuxBoot. We are currently exploring how to solve that issue… > Hello Dong, This fixes the RSDP issue perhaps, but that is not the only problem. I have mentioned this many times already, but let me mention it again for completeness: ACPI does not have a memory map, and ARM is much more finicky about mapping memory regions with the right attributes, given that uncached accesses don't snoop the caches like they do on x86. This means it is essential that memory mappings constructed from AML code (which doesn't provide any context pertaining to the memory type either) are created with the right memory type. Currently, the Linux/arm64 glue code for the ACPI core cross-references every memory mapping created from a SystemMemory OpRegion by AML code against the EFI memory map, and uses the EFI memory type and attributes to infer the memory type to program into the page tables. So simply providing the RSDP is *not* sufficient: on arm64, more work is needed and currently, booting ACPI without a EFI memory map results in SystemMemory OpRegions not working at all. Of course, we might be able to work around that by providing a 'coreboot' memory map for doing ACPI on arm64, but that results in more fragmentation and an inflated validation matrix, which puts the burden on the Linux subsystem maintainers to make sure that all these different combinations remain supported. AIUI, this memory type issue does not exist for RISC-V today, but I'd suggest to the RISC-V maintainers to take a careful look at arm64's acpi_os_ioremap() implementation and decide whether or not RISC-V needs similar logic.
