This bug took forever to debug (just ask Ben :-).

When we move the completion event from the failed request to the sense
request, we risk either the initial complete and then later complete on
a long gone ->waiting.  I think this business of moving the completion
structure to the request sense is a bit bogus and always has been, and
the bug is fixed nicely by just rewriting this logic a bit.  So instead
we simply unconditionally dequeue the failed request (regardless of
whether it was REQ_PC or REQ_BLOCK_PC), and pass a reference to it in
the sense request.  When the sense completes, we call end io on the
originally failed request (which does the complete() etc).
Signed-off-by: Jens Axboe <axboe@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>