This fixes two issues in the CFQ IO-scheduler:

o Still a bad allocation under the queue lock
o We need to clear crq after mempool_alloc(), otherwise the rbtree
  pointers can contain garbage if slab poisoning is enabled. This causes
  crashes on front/back merges because rb_next() returns crap.
Signed-off-by: Jens Axboe <axboe@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>