badly broken on big-endian

* passing little-endian to pci_unmap_single() et.al.
* cpu_to_le32() before passing value to writel()
* worse, cpu_to_le64() and shifting/masking result before the same
* 			hmp->tx_ring[i].status_n_length = cpu_to_le32(
				DescEndRing |
				(hmp->tx_ring[i].status_n_length & 0x0000FFFF));
  is obviously bogus on big-endian.  Not hard to untangle, fortunately...
* poisoning addresses in rx_ring is better done after we'd done
pci_unmap_single() on them, not before that.  [this one affects little-endian
as well, obviously, provided that pci_unmap_single() is not a no-op on target
in question]
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>
Signed-off-by: NJeff Garzik <jeff@garzik.org>