- The balance_dirty_pages() logic is simply wrong.  It goes:

	if (value > threshold)
		go_and_write(value - threshold);

  which is just fine for a single process writing data.  But
  for many processes, they *all* go and bring things back into
  balance, and too much data gets written out.

- The

	go_and_write(this much)

  logic is inoperative, because I turned off the ->writeback_mapping()
  function in ext2.  So a call to writeback_unlocked_inodes(this_much)
  doesn't actually decrement and test *this_much.  It will walk every
  inode, all the time.  Silly.

So quickly fixing the above things, the amount of dirty+writeback
memory in the machine nicely stabilises at 500 megabytes across
the run.