[Intel-gfx] [PATCH] drm/i915: Remove guard page insertion around unevictable nodes

[Chris Wilson]

Quoting Matthew Auld (2021-01-25 11:16:13)
> On Sun, 24 Jan 2021 at 13:57, Chris Wilson <chris at chris-wilson.co.uk> wrote:
> >
> > Assume that unevictable nodes are not in the GTT and so we can ignore
> > page boundary concerns, and so allow regular nodes to abutt against
> > irregular unevictable nodes.
> >
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  drivers/gpu/drm/i915/i915_drv.h       |  2 --
> >  drivers/gpu/drm/i915/i915_gem_evict.c |  6 ++++--
> >  drivers/gpu/drm/i915/i915_vma.h       | 10 +++++++++-
> >  drivers/gpu/drm/i915/i915_vma_types.h |  2 ++
> >  4 files changed, 15 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > index 99cf861df92d..69c5a185ecff 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.h
> > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > @@ -357,8 +357,6 @@ enum i915_cache_level {
> >         I915_CACHE_WT, /* hsw:gt3e WriteThrough for scanouts */
> >  };
> >
> > -#define I915_COLOR_UNEVICTABLE (-1) /* a non-vma sharing the address space */
> > -
> >  struct intel_fbc {
> >         /* This is always the inner lock when overlapping with struct_mutex and
> >          * it's the outer lock when overlapping with stolen_lock. */
> > diff --git a/drivers/gpu/drm/i915/i915_gem_evict.c b/drivers/gpu/drm/i915/i915_gem_evict.c
> > index 4d2d59a9942b..aef88fdb9f66 100644
> > --- a/drivers/gpu/drm/i915/i915_gem_evict.c
> > +++ b/drivers/gpu/drm/i915/i915_gem_evict.c
> > @@ -313,11 +313,13 @@ int i915_gem_evict_for_node(struct i915_address_space *vm,
> >                  */
> >                 if (i915_vm_has_cache_coloring(vm)) {
> >                         if (node->start + node->size == target->start) {
> > -                               if (node->color == target->color)
> > +                               if (i915_node_color_matches(node,
> > +                                                           target->color))
> >                                         continue;
> >                         }
> >                         if (node->start == target->start + target->size) {
> > -                               if (node->color == target->color)
> > +                               if (i915_node_color_matches(node,
> > +                                                           target->color))
> >                                         continue;
> >                         }
> >                 }
> 
> Since we bail early on seeing COLOR_UNEVICTABLE, and since we have to
> enlarge our search space by a page on both ends, do we need something
> like:

Are we not doing the opposite here, and skipping the evict if either
node/target is unevictable?

So we always expand the search by a page if the vm has coloring enabled,
but then skip the guard page eviction if either side says no.
-Chris