On 27 October 2015 at 06:02, Matt Fleming <matt@xxxxxxxxxxxxxxxxxxx> wrote: > On Fri, 23 Oct, at 10:37:46AM, Ard Biesheuvel wrote: >> >> After looking at the original (already merged) patch 11/11 again, I >> realize this is still not right: the problem is that efi_memory_map's >> phys_map member uses a void* type to hold a physical address, which >> happens to be correct in the normal case even when phys_addr_t is >> larger than void* (like on ARM with LPAE enabled) since the address it >> holds is the address of an allocation performed by the firmware, which >> only uses 1:1 addressable memory. >> >> However, overwriting memmap.phys_map with a value produced my >> memblock_alloc() is problematic, since the allocation may be above 4 >> GB on 32-bit (L)PAE platforms. So the correct way to do this would be >> to set the memblock limit to 4GB before memblock_alloc() on 32-bit >> platforms, and restore it afterwards. This is a bit of a kludge, >> though, and it would be more correct to change the type of >> efi_memory_map::phys_map to phys_addr_t, although I don't know what >> the potential fallout of that change is. Matt? > > I think that should be fine. The only potentially tricky situation we > could encounter is where 32-bit x86 firmware uses PAE but the kernel > is built without support. > > But that's not something I've ever seen enabled in the firmware and > there's a bunch of assumptions in the kernel already that would break > in that case. > Does UEFI even allow that? Even if it can describe memory over 4 GB, it uses a flat mapping so allocations done by the stub (which retrieves the memory map) should never reside at addresses over 4 GB. -- Ard. -- To unsubscribe from this list: send the line "unsubscribe linux-efi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
