FAT has a somewhat obnoxious rule: byte E5 in the first position is
encoded as 05.  If it was *always* that case we could simply fold it
into the codepage tables, but that rule applies only to the first
character.

Note that in the common codepages (incl. 437), E5 is lower-case sigma
(σ), which we currently upper case to upper case sigma (Σ, E4)
preventing a match -- which DOS *doesn't* do (it thinks of the Greek
characters as symbols, not subject to case distinctions.)  Putting
that special case in the codepage-generation scripts seems too ugly,
though.

"Bytes left" should be "bytes read"

GNU ld 2.17, which is still included with Debian 4.0 (etch),
mishandles PC-relative relocations to an absolute address.  Hack
around it by making pm_entry a segment-relative symbol, completely
unnecessarily so.

Actually close the file properly on EOF, even if not aligned to a
sector boundary.  Sigh.

Print the unknown keyword if we run into one.

We would incorrectly not set SI=0 when returning from getfssec upon
end of file.  Fix this.

Decimal, not hex...

Add routine to write decimal numbers, currently unused, but likely to
become useful at some time or another.

UHCI supposedly wants a null command (0FFh) sent to the KBC as part of
the A20-flipping sequence.  See if that works for us.

We have made it this far without it, but it seems like a disaster
waiting to happen.  We can afford the extra four bytes in sector 1.