Re: [PATCH 0/6] prohibit pinning pages in ZONE_MOVABLE

[Jason Gunthorpe]

On Fri, Dec 04, 2020 at 10:55:30AM -0500, Pavel Tatashin wrote:
> On Thu, Dec 3, 2020 at 11:03 PM Joonsoo Kim <js1304@xxxxxxxxx> wrote:
> >
> > Hello,
> >
> > On Wed, Dec 02, 2020 at 12:23:24AM -0500, Pavel Tatashin wrote:
> > > When page is pinned it cannot be moved and its physical address stays
> > > the same until pages is unpinned.
> > >
> > > This is useful functionality to allows userland to implementation DMA
> > > access. For example, it is used by vfio in vfio_pin_pages().
> > >
> > > However, this functionality breaks memory hotplug/hotremove assumptions
> > > that pages in ZONE_MOVABLE can always be migrated.
> > >
> > > This patch series fixes this issue by forcing new allocations during
> > > page pinning to omit ZONE_MOVABLE, and also to migrate any existing
> > > pages from ZONE_MOVABLE during pinning.
> >
> > I love what this patchset does, but, at least, it's better to consider
> > the side-effect of this patchset and inform it in somewhere. IIUC,
> > ZONE_MOVABLE exists for two purposes.
> >
> > 1) increasing availability of THP
> > 2) memory hot-unplug
> >
> > Potential issue would come from the case 1). They uses ZONE_MOVABLE
> > for THP availability and hard guarantee for migration isn't required
> > until now. So, there would be a system with following congifuration.
> >
> > - memory layout: ZONE_NORMAL-512MB, ZONE_MOVABLE-512MB
> > - memory usage: unmovable-256MB, movable pinned-256MB, movable
> >   unpinned-512MB
> >
> > With this patchset, movable pinned should be placed in ZONE_NORMAL so
> > 512MB is required for ZONE_NORMAL. ZONE_NORMAL would be exhausted and
> > system performance would be highly afftect according to memory usage
> > pattern.
> >
> > I'm not sure whether such configuration exists or not, but, at least,
> > it's better to write down this risk on commit message or something
> > else.
> 
> Yes, this indeed could be a problem for some configurations. I will
> add your comment to the commit log of one of the patches.

It sounds like there is some inherent tension here, breaking THP's
when doing pin_user_pages() is a really nasty thing to do. DMA
benefits greatly from THP.

I know nothing about ZONE_MOVABLE, is this auto-setup or an admin
option? If the result of this patch is standard systems can no longer
pin > 80% of their memory I have some regression concerns..

Jason