arm64: improved memory map handling for /dev/mem, ACPI etc

Quoted 70 lines · Dave Young

Quoted 57 lines · Ard Biesheuvel

Quoted 47 lines · Dave Young

Quoted 36 lines · Ard Biesheuvel

This series was split off from the UEFI virtmap for kexec series that I posted earlier today. The main purpose is to deal with the need to classify memory ranges as RAM or non-RAM in a consistent and comprehensive manner. This series applies on top of the other series.

Patch #1 avoids an early panic if the UEFI memory map is available but UEFI support itself fails to initialize. In this case, there is no need to panic early, and we have a better chance of being able to inform the user if we deal with this error condition at a later time.

Patch #2 adds iomem resource registration of UEFI memory regions. This is necessary because otherwise, drivers could potentially claim regions that are in active use by the firmware. This applies to both MMIO (NOR flash, RTC) and RAM ranges (runtime services code and data).

Patch #3-6 adds support to UEFI and non-UEFI code paths to record all memory known to the system in the 'physmem' memblock table (if enabled). This fulfils a need in the /dev/mem and (upcoming) ACPI layers to be able to classify ranges as being backed by normal RAM even if they are not covered by the 'memory' memblock table, and are hence not covered by the linear direct mapping. The physmem code is pre-existing code that only needs minor tweaking to be made suitable for this purpose.

Patch #7 enables the 'physmem' memblock table for arm64, and wires it into the handling of /dev/mem mappings, both to decide whether it should be mapped as MT_NORMAL, and whether read-write access can be allowed. (Non-RAM regions can be mapped read-write as long as they are not claimed by a driver in the iomem resource table. RAM regions can only be mapped read-only, and only if they are not covered by the 'memory' memblock table, and hence not covered by the linear mapping)

Finally, patch #8 changes the way the virtual memory map is handled by the early UEFI code. Specifically, it memblock_remove()s rather than _reserves() UEFI reserved RAM regions, so that they are removed entirely from the linear mapping.

Quoted 9 lines

Ard Biesheuvel (8):
arm64/efi: use UEFI memory map unconditionally if available arm64/efi: register UEFI reserved regions as iomem resources
memblock: add physmem to memblock_dump_all() output
memblock: introduce memblock_add_phys() and memblock_is_physmem()
of: fdt: register physmem in early_init_dt_scan_memory() arm64/efi: register physmem in reserve_regions()
arm64: use 'physmem' memblock to improve CONFIG_STRICT_DEVMEM handling arm64/efi: memblock_remove rather than _reserve UEFI reserved RAM

Quoted 2 lines

Ard, It is much cleaner for this splitting.

Thanks for having a look.

Quoted 3 lines

I wonder if some of them can become general code such as register reserved regions as iomem resources?

Quoted 7 lines

AFAICT, the x86 code adds reservations for such regions to the E820 memory map, which in turn is used to memblock_reserve() the actual memory. Also, sharing of the RTC is handled with a dedicated mutex in the runtime services wrapper code (and x86 does not even use the time related runtime services as they are broken on many firmware implementations) so blindly applying the same logic to x86 or ia64 would likely break stuff.

Ok, I think bootloader will pass the E820 ranges though it can not tell which region is for what.

Quoted 1 line

Quoted 7 lines

Quoted 5 lines

Do you (or Mark) have any feedback on the utility of this series in the ACPI context? I failed to mention in the cover letter that memblock_is_physmem() now serves the purpose of page_is_ram(), i.e., whether a physical region is backed by a slice of RAM that was left out of the kernel's linear mapping.

I will do some test later along with the stable mapping stuff.

Quoted 1 line

Cheers.

Quoted 6 lines

BTW, with previous stable mapping patches on modified 3.18 tree kexec kernel panics because of damaged efi mempry map arrays. It contains some random value instead of right addresses so that virt_to_phys return same value for vendor field. Will do more debugging maybe I missed something. Any clue for debugging above issue will be appreciated.

No clues, unfortunately. The memory map is not used after early boot, so it may be getting clobbered at any point between the first boot and the second boot.