blkgs are chained from both blkcgs and request_queues and thus
subjected to two locks - blkcg->lock and q->queue_lock.  As both blkcg
and q can go away anytime, locking during removal is tricky.  It's
currently solved by wrapping removal inside RCU, which makes the
synchronization complex.  There are three locks to worry about - the
outer RCU, q lock and blkcg lock, and it leads to nasty subtle
complications like conditional synchronize_rcu() on queue exit paths.

For all other paths, blkcg lock is naturally nested inside q lock and
the only exception is blkcg removal path, which is a very cold path
and can be implemented as clumsy but conceptually-simple reverse
double lock dancing.

This patch updates blkg removal path such that blkgs are removed while
holding both q and blkcg locks, which is trivial for request queue
exit path - blkg_destroy_all().  The blkcg removal path,
blkiocg_pre_destroy(), implements reverse double lock dancing
essentially identical to ioc_release_fn().

This simplifies blkg locking - no half-dead blkgs to worry about.  Now
unnecessary RCU annotations will be removed by the next patch.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Vivek Goyal <vgoyal@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>