On Sun, Dec 29, 2024 at 09:36:41PM -0800, David Rientjes wrote: > On Thu, 26 Dec 2024, Karim Manaouil wrote: > > memory node will have a standard top-tier node to promote to *or* that we > stash that promotion node information at the time of demotion Not sure stashing this information generalizes the way you want, and may require space in the folio/page struct that may not be available. Things also get more complicated the further away from virtual memory mappings you get, because associations become weaker (physical pages backing file cache may have weak, ephemeral associations, for example). We should consider the implications of limiting promotion to one or more levels in a single proximity domain - and whether the "two step" migration process is ultimately harmful or not. Just trying to save ourselves from crawling down a rabbit hole. ~Gregory
