[Intel-gfx] [RFC] mm, drm/i915: Mark pinned shmemfs pages as unevictable

[Jerome Glisse]

On Tue, Jun 06, 2017 at 02:30:15PM +0200, Vlastimil Babka wrote:
> On 06/06/2017 02:14 PM, Michal Hocko wrote:
> > On Tue 06-06-17 13:04:36, Chris Wilson wrote:
> >> Similar in principle to the treatment of get_user_pages, pages that
> >> i915.ko acquires from shmemfs are not immediately reclaimable and so
> >> should be excluded from the mm accounting and vmscan until they have
> >> been returned to the system via shrink_slab/i915_gem_shrink. By moving
> >> the unreclaimable pages off the inactive anon lru, not only should
> >> vmscan be improved by avoiding walking unreclaimable pages, but the
> >> system should also have a better idea of how much memory it can reclaim
> >> at that moment in time.
> > 
> > That is certainly desirable. Peter has proposed a generic pin_page (or
> > similar) API. What happened with it? I think it would be a better
> > approach than (ab)using mlock API. I am also not familiar with the i915
> > code to be sure that using lock_page is really safe here. I think that
> > all we need is to simply move those pages in/out to/from unevictable LRU
> > list on pin/unpining.
> 
> Hmm even when on unevictable list, the pages were still allocated as
> MOVABLE, while pinning prevents them from being migrated, so it doesn't
> play well with compaction/grouping by mobility/CMA etc. Addressing that
> would be more useful IMHO, and e.g. one of the features envisioned for
> the pinning API was to first migrate the pinned pages out of movable
> zones and CMA/MOVABLE pageblocks.

Cost would be high, GPU dataset can be big (giga byte range) so i don't
see copying out of MOVABLE as something sane to do here. Maybe we can
reuse the lru pointer to store a pointer to function and metadata so that
who ever pin a page provide a way to unpin it through this function.

Issue then is how to handle double pin (ie when 2 different driver want
to pin same page).

Cheers,
Jérôme