On 25 September 2015 at 23:01, Ingo Molnar [off-list ref] wrote:

* Matt Fleming [off-list ref] wrote:

quoted

From: Ard Biesheuvel <redacted>

The new Properties Table feature introduced in UEFIv2.5 may split
memory regions that cover PE/COFF memory images into separate code
and data regions. Since these regions only differ in the type (runtime
code vs runtime data) and the permission bits, but not in the memory
type attributes (UC/WC/WT/WB), the spec does not require them to be
aligned to 64 KB.

Since the relative offset of PE/COFF .text and .data segments cannot
be changed on the fly, this means that we can no longer pad out those
regions to be mappable using 64 KB pages.
Unfortunately, there is no annotation in the UEFI memory map that
identifies data regions that were split off from a code region, so we
must apply this logic to all adjacent runtime regions whose attributes
only differ in the permission bits.

So instead of rounding each memory region to 64 KB alignment at both
ends, only round down regions that are not directly preceded by another
runtime region with the same type attributes. Since the UEFI spec does
not mandate that the memory map be sorted, this means we also need to
sort it first.

So I think this is fundamentally wrong as well, similarly to the related x86 fix.

I think for compatibility reasons the whole 'EFI runtime image' should be mapped
in a single go, as closely matching the EFI layouts and offsets as possible. We
are not talking about gigabytes here, right?

As I explained in the other thread, this is really not necessary, and
never has been until the firmware started splitting up PE/COFF images
into several sections each. As long as we keep those PE/COFF images
together, everything will work as before, and the only complication is
that the memory map does not contain any clues about which regions
belong to a single PE/COFF image, so we need to keep all adjacent
EFI_MEMORY_RUNTIME regions adjacent in the VA mapping.

Even if technically they are 'separate sections', the x86 bug shows that they
aren't. So we should not pretend so on the Linux side either and we should not
tear them apart (and then work hard to preserve the interdependencies, some
declared, some hidden!).

This is about relocations and interdependencies at the symbol level,
and such interdependencies only exist internally inside PE/COFF
images.

If we allocate the EFI runtime as a single virtual memory block then issues like
rounding between sections does not even come up as a problem: we map the original
offsets and sizes byte by byte.

Well, by that reasoning, we should not call SetVirtualAddressMap() in
the first place, and just use the 1:1 mapping UEFI uses natively. This
is more than feasible on arm64, and I actually fought hard against
using SetVirtualAddressMap() at all, but I was overruled by others. I
think this is also trivially possible on X64, since the 1:1 mapping is
already active alongside the VA mapping.

-- 
Ard.