Re: [PATCH v5 4/4] mm: Defer ZONE_DEVICE page initialization to the point where we init pgmap

From: Dan Williams
Date: Wed Oct 10 2018 - 14:19:04 EST


On Wed, Oct 10, 2018 at 10:30 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
>
> On Wed 10-10-18 09:39:08, Alexander Duyck wrote:
> > On 10/10/2018 2:58 AM, Michal Hocko wrote:
> > > On Tue 09-10-18 13:26:41, Alexander Duyck wrote:
> > > [...]
> > > > I would think with that being the case we still probably need the call to
> > > > __SetPageReserved to set the bit with the expectation that it will not be
> > > > cleared for device-pages since the pages are not onlined. Removing the call
> > > > to __SetPageReserved would probably introduce a number of regressions as
> > > > there are multiple spots that use the reserved bit to determine if a page
> > > > can be swapped out to disk, mapped as system memory, or migrated.
> > >
> > > PageReserved is meant to tell any potential pfn walkers that might get
> > > to this struct page to back off and not touch it. Even though
> > > ZONE_DEVICE doesn't online pages in traditional sense it makes those
> > > pages available for further use so the page reserved bit should be
> > > cleared.
> >
> > So from what I can tell that isn't necessarily the case. Specifically if the
> > pagemap type is MEMORY_DEVICE_PRIVATE or MEMORY_DEVICE_PUBLIC both are
> > special cases where the memory may not be accessible to the CPU or cannot be
> > pinned in order to allow for eviction.
>
> Could you give me an example please?
>
> > The specific case that Dan and Yi are referring to is for the type
> > MEMORY_DEVICE_FS_DAX. For that type I could probably look at not setting the
> > reserved bit. Part of me wants to say that we should wait and clear the bit
> > later, but that would end up just adding time back to initialization. At
> > this point I would consider the change more of a follow-up optimization
> > rather than a fix though since this is tailoring things specifically for DAX
> > versus the other ZONE_DEVICE types.
>
> I thought I have already made it clear that these zone device hacks are
> not acceptable to the generic hotplug code. If the current reserve bit
> handling is not correct then give us a specific reason for that and we
> can start thinking about the proper fix.
>

Right, so we're in a situation where a hack is needed for KVM's
current interpretation of the Reserved flag relative to dax mapped
pages. I'm arguing to push that knowledge / handling as deep as
possible into the core rather than hack the leaf implementations like
KVM, i.e. disable the Reserved flag for all non-MEMORY_DEVICE_*
ZONE_DEVICE types.

Here is the KVM thread about why they need a change:

https://lkml.org/lkml/2018/9/7/552

...and where I pushed back on a KVM-local hack:

https://lkml.org/lkml/2018/9/19/154