Ugh sorry I just dug this patch out from last week and forgot I had just
picked an arbitrary number to make sure it was working. You are correct,
what I _meant_ to do (and will do after I respond) was calculate how
much we wanted to flush and then divide that by how much the delayed
inodes reserve and then multiply it by 2 for good measure.

I guess that's as good a place as any to start :)

It'd be nice to see some data on how often this actually stops the full
run.  If it doesn't, just pitch the whole _NR path entirely?

- z