> I was looking at 2.6.9.  it looked to me that the radix tree tags
> were how dirty pages were identified, but it had to look at the tag
> for every page to find the dirty ones.  granted, this is a lot faster
> than what we do, but it doesn't let you find the dirty pages without
> examining the state for every page.

my understanding is, because tags are aggregated into intermediate nodes,
it doesn't need to visit each untagged leafs in many cases.

YAMAMOTO Takashi