Add an rmb() between checking the ownership bit of an event queue entry and
reading the contents of the EQE.  Without this barrier, the CPU could read
stale contents of the EQE before HW writes the EQE but have the read of the
ownership bit reordered until after HW finishes writing, which leads to the
driver processing an incorrect event.  This was actually observed to happen
when multiple completion queues are in heavy use on an IBM JS20 PowerPC 970
system.

Also explain the existing rmb() in completion queue poll (there for the same
reason) and slightly improve debugging output.
Signed-off-by: Roland Dreier <roland@topspin.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>