There are a couple of different solutions to this problem...

1) Compile all LKM code as quasi-PIC on platforms with limited
   jmp/call range -- meaning that references are always stored in and
   loaded from the data segment.  This would cause a slight
   performance hit, but it would be no worse than regular PIC code.

2) Reserve VA space around the loaded kernel image to store LKM code.

The latter is probably easier and more expedient to implement, though
the former isn't really hard for a GCC g00r00.  Perhaps you'd like to
work on it, David?  B-)