Currently there are several minor problems with randomization base
generation code:

1. It might misbehave in low memory conditions. In particular there
might be enough space for the kernel on [0, block_sum] but after
	if (base < safe_addr)
		base = safe_addr;
it might not be enough anymore.

2. It does not correctly handle minimal address constraint. In condition
	if (base < safe_addr)
		base = safe_addr;
a synthetic value is compared with an address. If we have a memory
setup with memory holes due to offline memory regions, and safe_addr is
close to the end of the first online memory block - we might position
the kernel in invalid memory.

3. block_sum calculation logic contains off-by-one error. Let's say we
have a memory block in which the kernel fits perfectly
(end - start == kernel_size). In this case:
		if (end - start < kernel_size)
			continue;
		block_sum += end - start - kernel_size;
block_sum is not increased, while it is a valid kernel position.

So, address problems listed and explain algorithm used. Besides that
restructuring the code makes it possible to extend kernel positioning
algorithm further. Currently we pick position in between single
[min, max] range (min = safe_addr, max = memory_limit). In future we
can do that for multiple ranges as well (by calling
count_valid_kernel_positions for each range).
Reviewed-by: Philipp Rudo <prudo@linux.ibm.com>
Reviewed-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>