(In reply to Mike Lothian from comment #2)
> I'll revert that on the 4.9-next-wip branch and see if it makes a difference
> 
> I'm bisecting on linus's tree at the moment to see where the regression
> crept in

It's probably due to the d3cold support that when upstream in 4.8.

> 
> I've also been runnig with these local patches to use Dave's D3cold work (I
> reverted those before reporting) I'll attach them to the bug though

You shouldn't need those (or the workaround from my tree referenced in comment
1) now that the d3cold support has gone upstream via the pci tree.