On Mon, Dec 20, 2021 at 09:03:12PM -0400, Jason Gunthorpe wrote:

That just leave the THP splitting.. I suppose we get the PTL, then
compute the current value of the new bit based on refcount and diffuse
it to all tail pages, then update the PMD and release the PTL. Safe
against concurrent WP - don't need DoubleMap horrors because it isn't
a counter.

One of the things I've been trying to figure out is how we do
can_split_huge_page().  Maybe an rmap walk to figure out how many
refcounts we would subtract if we did unmap it from everywhere it's
currently mapped?  (just to be clear, we call unmap_page() as the
next thing, so I don't mind warming up the rbtree cachelines
if it's mapped anywhere)