On Tue, Oct 30, 2018 at 1:18 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > On Mon 29-10-18 23:55:12, Dan Williams wrote: > > On Mon, Oct 29, 2018 at 11:29 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote: [..] > > That testing identified this initialization performance problem and > > thankfully got it addressed in time for the current merge window. > > And I still cannot see a word about that in changelogs. True, I would have preferred that commit 966cf44f637e "mm: defer ZONE_DEVICE page initialization to the point where we init pgmap" include some notes about the scaling advantages afforded by not serializing memmap_init_zone() work. I think this information got distributed across several patch series because removing the lock was not sufficient by itself, Alex went further to also rework the physical socket affinity of the nvdimm sub-system's async initialization threads. As the code gets refactored further it's a chance to add commentary on the scaling expectations of the design.
