Instead of opencoding the fsync calling sequence use vfs_fsync.  This also
gets rid of the useless i_mutex over the data writeout.

Consolidate the remaining special code for syncing directories and document
it's quirks.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: J. Bruce Fields <bfields@citi.umich.edu>