Re: [PATCH v3 1/1] mm/migrate: Add migrate_device_pfns

[Matthew Brost <matthew.brost@intel.com>]

On Wed, Oct 23, 2024 at 06:22:17PM -0700, Andrew Morton wrote:
> On Wed, 23 Oct 2024 16:39:43 -0700 Matthew Brost <matthew.brost@intel.com> wrote:
> 
> > Part of series [1]. Sending as individual patch ahead of that series as
> > this is a prerequisite for merging.
> 
> That's news to me - singleton patches are perfectly OK?
> 

I've merged a couple of other patches outside of the DRM subsystem for
pending series which we then have picked up in a following kernel
release. If I have this flow wrong, my mistake.

> On Wed, 23 Oct 2024 16:39:44 -0700 Matthew Brost <matthew.brost@intel.com> wrote:
> 
> > Implement migrate_device_pfns to prepare an array of PFNs for migration.
> > Handles non-contiguous ranges of device pages that require migration.
> 
> OK, that's "what".  We're more interested in "why".
> 

Sure can add. The 'why' is:

A non-contiguous allocation of device pages can occur if a device is
under memory pressure within a single driver allocation of device
memory. Additionally, a driver allocation of memory can also be evicted
under memory pressure. Therefore, an interface for migrating a set of
non-contiguous device pages is required.

Matt

> > +EXPORT_SYMBOL(migrate_device_pfns);
> 
> And it's exported to modules, which adds to the significance.
> 
> Please fully describe the reasons for proposing this change.