Several reports about leaf corruption has been floating on the list, one of them
points to __btrfs_drop_extents(), and we find that the leaf becomes corrupted
after __btrfs_drop_extents(), it's really a rare case but it does exist.

The problem turns out to be btrfs_next_leaf() called in __btrfs_drop_extents().

So in btrfs_next_leaf(), we release the current path to re-search the last key of
the leaf for locating next leaf, and we've taken it into account that there might
be balance operations between leafs during this 'unlock and re-lock' dance, so
we check the path again and advance it if there are now more items available.
But things are a bit different if that last key happens to be removed and balance
gets a bigger key as the last one, and btrfs_search_slot will return it with
ret > 0, IOW, nothing change in this leaf except the new last key, then we think
we're okay because there is no more item balanced in, fine, we thinks we can
go to the next leaf.

However, we should return that bigger key, otherwise we deserve leaf corruption,
for example, in endio, skipping that key means that __btrfs_drop_extents() thinks
it has dropped all extent matched the required range and finish_ordered_io can
safely insert a new extent, but it actually doesn't and ends up a leaf
corruption.

One may be asking that why our locking on extent io tree doesn't work as
expected, ie. it should avoid this kind of race situation.  But in
__btrfs_drop_extents(), we don't always find extents which are included within
our locking range, IOW, extents can start before our searching start, in this
case locking on extent io tree doesn't protect us from the race.

This takes the special case into account.

Reviewed-by: Filipe Manana <fdmanana@gmail.com>
Signed-off-by: Liu Bo <bo.li.liu@oracle.com>
Signed-off-by: Chris Mason <clm@fb.com>