xfs: fix buffer lookup vs release race

Since commit 298f3422 ("xfs: lockless buffer lookup") the buffer
lookup fastpath is done without a hash-wide lock (then pag_buf_lock, now
bc_lock) and only under RCU protection.  But this means that nothing
serializes lookups against the temporary 0 reference count for buffers
that are added to the LRU after dropping the last regular reference,
and a concurrent lookup would fail to find them.

Fix this by doing all b_hold modifications under b_lock.  We're already
doing this for release so this "only" ~ doubles the b_lock round trips.
We'll later look into the lockref infrastructure to optimize the number
of lock round trips again.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Carlos Maiolino <cem@kernel.org>