->dma can be a 64bit variable on 64bit boxen; its value will fit into 32 bits
just fine (due to dma mask).  However, cpu_to_le32p(&...) will break if we
are on a 64bit big-endian; we'll end up up passing it the address of upper
32 bits and get 0 instead of correct value.  Fix is trivial...
Signed-off-by: Al Viro <viro@parcelfarce.linux.org.uk>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>