> Subject: Re: [PATCH 1/2] sched/wait: Break up long wake list walk > > On Tue, Aug 22, 2017 at 2:24 PM, Andi Kleen <ak@xxxxxxxxxxxxxxx> wrote: > > > > I believe in this case it's used by threads, so a reference count > > limit wouldn't help. > > For the first migration try, yes. But if it's some kind of "try and try again" > pattern, the second time you try and there are people waiting for the page, > the page count (not the map count) would be elevanted. > > So it's possible that depending on exactly what the deeper problem is, the > "this page is very busy, don't migrate" case might be discoverable, and the > page count might be part of it. > > However, after PeterZ made that comment that page migration should have > that should_numa_migrate_memory() filter, I am looking at that > mpol_misplaced() code. > > And honestly, that MPOL_PREFERRED / MPOL_F_LOCAL case really looks like > complete garbage to me. > > It looks like garbage exactly because it says "always migrate to the current > node", but that's crazy - if it's a group of threads all running together on the > same VM, that obviously will just bounce the page around for absolute zero > good ewason. > > The *other* memory policies look fairly sane. They basically have a fairly > well-defined preferred node for the policy (although the > "MPOL_INTERLEAVE" looks wrong for a hugepage). But > MPOL_PREFERRED/MPOL_F_LOCAL really looks completely broken. > > Maybe people expected that anybody who uses MPOL_F_LOCAL will also > bind all threads to one single node? > > Could we perhaps make that "MPOL_PREFERRED / MPOL_F_LOCAL" case just > do the MPOL_F_MORON policy, which *does* use that "should I migrate to > the local node" filter? > > IOW, we've been looking at the waiters (because the problem shows up due > to the excessive wait queues), but maybe the source of the problem comes > from the numa balancing code just insanely bouncing pages back-and-forth if > you use that "always balance to local node" thing. > > Untested (as always) patch attached. The patch doesn’t work. Thanks, Kan ��.n������g��� �a��� �&ޖ)���)��h���&������梷���� �Ǟ�m������)������^�����������v���O��zf������
