We used to try to request 8 times more vram than needed, which would
fail if the card has a too small BAR (observed with qemu & kvm).
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
CC: stable@kernel.org