The current writer throttling in balance_dirty_pages() assumes that the
writer will be effectively throttled on request queues.

That works fine when the amount of data which can be placed into a
queue is "much less than" total memory.

But if the machine has a small amount of memory, or many disks, or has
large request queues, or large requests, it can go wrong.

For example, with mem=96m and dirty_async_ratio=15, we want to be able
to clamp dirty+writeback memory at 15 megabytes.  But it doesn't work,
because a single SCSI request queue can hold 40 megs or more.  The
heavy writer keeps on dirtying memory until that queue fills up.

So add a test for that - if we did some writeback, and we're *still*
over the dirty+writeback threshold then make the caller take an
explicit nap on some writes terminating.  And keep on doing that until
the dirty+writeback memory subsides.