| Message ID | 20190719130423.11195-1-chris@chris-wilson.co.uk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm/i915: Hide unshrinkable context objects from the shrinker | expand |
On 19/07/2019 14:04, Chris Wilson wrote: > The shrinker cannot touch objects used by the contexts (logical state > and ring). Currently we mark those as "pin_global" to let the shrinker > skip over them, however, if we remove them from the shrinker lists > entirely, we don't event have to include them in our shrink accounting. > > By keeping the unshrinkable objects in our shrinker tracking, we report > a large number of objects available to be shrunk, and leave the shrinker > deeply unsatisfied when we fail to reclaim those. The shrinker will > persist in trying to reclaim the unavailable objects, forcing the system > into a livelock (not even hitting the dread oomkiller). > > v2: Extend unshrinkable protection for perma-pinned scratch and guc > allocations (Tvrtko) > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> > --- > drivers/gpu/drm/i915/gem/i915_gem_object.c | 11 ++-- > drivers/gpu/drm/i915/gem/i915_gem_object.h | 4 ++ > drivers/gpu/drm/i915/gem/i915_gem_pages.c | 13 +---- > drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 57 ++++++++++++++++++++ > drivers/gpu/drm/i915/gt/intel_context.c | 4 +- > drivers/gpu/drm/i915/gt/intel_gt.c | 2 + > drivers/gpu/drm/i915/gt/intel_ringbuffer.c | 17 +++--- > drivers/gpu/drm/i915/gt/uc/intel_guc.c | 2 + > drivers/gpu/drm/i915/i915_debugfs.c | 3 +- > drivers/gpu/drm/i915/i915_vma.c | 15 ++++++ > drivers/gpu/drm/i915/i915_vma.h | 4 ++ > 11 files changed, 101 insertions(+), 31 deletions(-) > > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c > index d5197a2a106f..4ea97fca9c35 100644 > --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c > +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c > @@ -63,6 +63,8 @@ void i915_gem_object_init(struct drm_i915_gem_object *obj, > spin_lock_init(&obj->vma.lock); > INIT_LIST_HEAD(&obj->vma.list); > > + INIT_LIST_HEAD(&obj->mm.link); > + > INIT_LIST_HEAD(&obj->lut_list); > INIT_LIST_HEAD(&obj->batch_pool_link); > > @@ -273,14 +275,7 @@ void i915_gem_free_object(struct drm_gem_object *gem_obj) > * or else we may oom whilst there are plenty of deferred > * freed objects. > */ > - if (i915_gem_object_has_pages(obj) && > - i915_gem_object_is_shrinkable(obj)) { > - unsigned long flags; > - > - spin_lock_irqsave(&i915->mm.obj_lock, flags); > - list_del_init(&obj->mm.link); > - spin_unlock_irqrestore(&i915->mm.obj_lock, flags); > - } > + i915_gem_object_make_unshrinkable(obj); > > /* > * Since we require blocking on struct_mutex to unbind the freed > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h > index 67aea07ea019..3714cf234d64 100644 > --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h > +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h > @@ -394,6 +394,10 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, > unsigned int flags); > void i915_gem_object_unpin_from_display_plane(struct i915_vma *vma); > > +void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj); > +void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj); > +void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj); > + > static inline bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) > { > if (obj->cache_dirty) > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_pages.c b/drivers/gpu/drm/i915/gem/i915_gem_pages.c > index b36ad269f4ea..92ad3cc220e3 100644 > --- a/drivers/gpu/drm/i915/gem/i915_gem_pages.c > +++ b/drivers/gpu/drm/i915/gem/i915_gem_pages.c > @@ -153,24 +153,13 @@ static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj) > struct sg_table * > __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj) > { > - struct drm_i915_private *i915 = to_i915(obj->base.dev); > struct sg_table *pages; > > pages = fetch_and_zero(&obj->mm.pages); > if (IS_ERR_OR_NULL(pages)) > return pages; > > - if (i915_gem_object_is_shrinkable(obj)) { > - unsigned long flags; > - > - spin_lock_irqsave(&i915->mm.obj_lock, flags); > - > - list_del(&obj->mm.link); > - i915->mm.shrink_count--; > - i915->mm.shrink_memory -= obj->base.size; > - > - spin_unlock_irqrestore(&i915->mm.obj_lock, flags); > - } > + i915_gem_object_make_unshrinkable(obj); > > if (obj->mm.mapping) { > void *ptr; > diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c > index 3f4c6bdcc3c3..14abfd77365f 100644 > --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c > +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c > @@ -530,3 +530,60 @@ void i915_gem_shrinker_taints_mutex(struct drm_i915_private *i915, > if (unlock) > mutex_release(&i915->drm.struct_mutex.dep_map, 0, _RET_IP_); > } > + > +void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj) > +{ > + if (!list_empty(&obj->mm.link)) { Which lock protects this list head? > + struct drm_i915_private *i915 = to_i915(obj->base.dev); > + unsigned long flags; > + > + spin_lock_irqsave(&i915->mm.obj_lock, flags); > + GEM_BUG_ON(list_empty(&obj->mm.link)); > + > + list_del_init(&obj->mm.link); > + i915->mm.shrink_count--; > + i915->mm.shrink_memory -= obj->base.size; > + > + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); > + } > +} > + > +void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj) > +{ > + GEM_BUG_ON(!i915_gem_object_has_pages(obj)); > + GEM_BUG_ON(!list_empty(&obj->mm.link)); > + > + if (i915_gem_object_is_shrinkable(obj)) { > + struct drm_i915_private *i915 = to_i915(obj->base.dev); > + unsigned long flags; > + > + spin_lock_irqsave(&i915->mm.obj_lock, flags); > + GEM_BUG_ON(!kref_read(&obj->base.refcount)); > + > + list_add_tail(&obj->mm.link, &i915->mm.shrink_list); > + i915->mm.shrink_count++; > + i915->mm.shrink_memory += obj->base.size; > + > + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); > + } > +} > + > +void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj) > +{ > + GEM_BUG_ON(!i915_gem_object_has_pages(obj)); > + GEM_BUG_ON(!list_empty(&obj->mm.link)); > + > + if (i915_gem_object_is_shrinkable(obj)) { > + struct drm_i915_private *i915 = to_i915(obj->base.dev); > + unsigned long flags; > + > + spin_lock_irqsave(&i915->mm.obj_lock, flags); > + GEM_BUG_ON(!kref_read(&obj->base.refcount)); > + > + list_add_tail(&obj->mm.link, &i915->mm.purge_list); > + i915->mm.shrink_count++; > + i915->mm.shrink_memory += obj->base.size; > + > + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); > + } > +} Common helper for the above to passing in the correct list from each? > diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c > index 9e4f51ce52ff..9830edda1ade 100644 > --- a/drivers/gpu/drm/i915/gt/intel_context.c > +++ b/drivers/gpu/drm/i915/gt/intel_context.c > @@ -118,7 +118,7 @@ static int __context_pin_state(struct i915_vma *vma) > * And mark it as a globally pinned object to let the shrinker know > * it cannot reclaim the object until we release it. > */ > - vma->obj->pin_global++; > + i915_vma_make_unshrinkable(vma); > vma->obj->mm.dirty = true; > > return 0; > @@ -126,8 +126,8 @@ static int __context_pin_state(struct i915_vma *vma) > > static void __context_unpin_state(struct i915_vma *vma) > { > - vma->obj->pin_global--; > __i915_vma_unpin(vma); > + i915_vma_make_shrinkable(vma); > } > > static void __intel_context_retire(struct i915_active *active) > diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c > index f7e69db4019d..5b16b233a059 100644 > --- a/drivers/gpu/drm/i915/gt/intel_gt.c > +++ b/drivers/gpu/drm/i915/gt/intel_gt.c > @@ -231,6 +231,8 @@ int intel_gt_init_scratch(struct intel_gt *gt, unsigned int size) > if (ret) > goto err_unref; > > + i915_gem_object_make_unshrinkable(obj); > + > gt->scratch = vma; > return 0; > > diff --git a/drivers/gpu/drm/i915/gt/intel_ringbuffer.c b/drivers/gpu/drm/i915/gt/intel_ringbuffer.c > index 38ec11ae6ed7..d8efb88f33f3 100644 > --- a/drivers/gpu/drm/i915/gt/intel_ringbuffer.c > +++ b/drivers/gpu/drm/i915/gt/intel_ringbuffer.c > @@ -1238,7 +1238,7 @@ int intel_ring_pin(struct intel_ring *ring) > goto err_ring; > } > > - vma->obj->pin_global++; > + i915_vma_make_unshrinkable(vma); > > GEM_BUG_ON(ring->vaddr); > ring->vaddr = addr; > @@ -1267,6 +1267,8 @@ void intel_ring_reset(struct intel_ring *ring, u32 tail) > > void intel_ring_unpin(struct intel_ring *ring) > { > + struct i915_vma *vma = ring->vma; > + > if (!atomic_dec_and_test(&ring->pin_count)) > return; > > @@ -1275,18 +1277,17 @@ void intel_ring_unpin(struct intel_ring *ring) > /* Discard any unused bytes beyond that submitted to hw. */ > intel_ring_reset(ring, ring->tail); > > - GEM_BUG_ON(!ring->vma); > - i915_vma_unset_ggtt_write(ring->vma); > - if (i915_vma_is_map_and_fenceable(ring->vma)) > - i915_vma_unpin_iomap(ring->vma); > + i915_vma_unset_ggtt_write(vma); > + if (i915_vma_is_map_and_fenceable(vma)) > + i915_vma_unpin_iomap(vma); > else > - i915_gem_object_unpin_map(ring->vma->obj); > + i915_gem_object_unpin_map(vma->obj); > > GEM_BUG_ON(!ring->vaddr); > ring->vaddr = NULL; > > - ring->vma->obj->pin_global--; > - i915_vma_unpin(ring->vma); > + i915_vma_unpin(vma); > + i915_vma_make_purgeable(vma); Why is ring purgeable and scratch or context state shrinkable? > > intel_timeline_unpin(ring->timeline); > } > diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc.c b/drivers/gpu/drm/i915/gt/uc/intel_guc.c > index 83f2c197375f..a5ba0164fd14 100644 > --- a/drivers/gpu/drm/i915/gt/uc/intel_guc.c > +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc.c > @@ -597,6 +597,8 @@ struct i915_vma *intel_guc_allocate_vma(struct intel_guc *guc, u32 size) > goto err; > } > > + i915_gem_object_make_unshrinkable(obj); > + > return vma; > > err: > diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c > index 6b84d04a6a28..c43f270085f5 100644 > --- a/drivers/gpu/drm/i915/i915_debugfs.c > +++ b/drivers/gpu/drm/i915/i915_debugfs.c > @@ -363,8 +363,9 @@ static int i915_gem_object_info(struct seq_file *m, void *data) > struct drm_i915_private *i915 = node_to_i915(m->private); > int ret; > > - seq_printf(m, "%u shrinkable objects, %llu bytes\n", > + seq_printf(m, "%u shrinkable [%u free] objects, %llu bytes\n", > i915->mm.shrink_count, > + atomic_read(&i915->mm.free_count), > i915->mm.shrink_memory); > > seq_putc(m, '\n'); > diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c > index eb16a1a93bbc..e27ddbff487e 100644 > --- a/drivers/gpu/drm/i915/i915_vma.c > +++ b/drivers/gpu/drm/i915/i915_vma.c > @@ -1030,6 +1030,21 @@ int i915_vma_unbind(struct i915_vma *vma) > return 0; > } > > +void i915_vma_make_unshrinkable(struct i915_vma *vma) > +{ > + i915_gem_object_make_unshrinkable(vma->obj); > +} > + > +void i915_vma_make_shrinkable(struct i915_vma *vma) > +{ > + i915_gem_object_make_shrinkable(vma->obj); > +} > + > +void i915_vma_make_purgeable(struct i915_vma *vma) > +{ > + i915_gem_object_make_purgeable(vma->obj); > +} Would i915_vma_make_*object*_... be a better name? I am thinking the concept does not apply to vma's. > + > #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) > #include "selftests/i915_vma.c" > #endif > diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h > index 4b769db649bf..a24bd6787ef7 100644 > --- a/drivers/gpu/drm/i915/i915_vma.h > +++ b/drivers/gpu/drm/i915/i915_vma.h > @@ -459,4 +459,8 @@ void i915_vma_parked(struct drm_i915_private *i915); > struct i915_vma *i915_vma_alloc(void); > void i915_vma_free(struct i915_vma *vma); > > +void i915_vma_make_unshrinkable(struct i915_vma *vma); > +void i915_vma_make_shrinkable(struct i915_vma *vma); > +void i915_vma_make_purgeable(struct i915_vma *vma); > + > #endif > Regards, Tvrtko
Quoting Tvrtko Ursulin (2019-07-19 16:38:28) > > On 19/07/2019 14:04, Chris Wilson wrote: > > +void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj) > > +{ > > + if (!list_empty(&obj->mm.link)) { > > Which lock protects this list head? Hmm, I was thinking this would be nicely ordered by the caller. But no, not strongly protected against the shrinker... > > + struct drm_i915_private *i915 = to_i915(obj->base.dev); > > + unsigned long flags; > > + > > + spin_lock_irqsave(&i915->mm.obj_lock, flags); > > + GEM_BUG_ON(list_empty(&obj->mm.link)); ..and so this should be a regular if() not BUG_ON. > > + list_del_init(&obj->mm.link); > > + i915->mm.shrink_count--; > > + i915->mm.shrink_memory -= obj->base.size; > > + > > + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); > > + } > > +} > > + > > +void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj) > > +{ > > + GEM_BUG_ON(!i915_gem_object_has_pages(obj)); > > + GEM_BUG_ON(!list_empty(&obj->mm.link)); > > + > > + if (i915_gem_object_is_shrinkable(obj)) { > > + struct drm_i915_private *i915 = to_i915(obj->base.dev); > > + unsigned long flags; > > + > > + spin_lock_irqsave(&i915->mm.obj_lock, flags); > > + GEM_BUG_ON(!kref_read(&obj->base.refcount)); > > + > > + list_add_tail(&obj->mm.link, &i915->mm.shrink_list); > > + i915->mm.shrink_count++; > > + i915->mm.shrink_memory += obj->base.size; > > + > > + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); > > + } > > +} > > + > > +void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj) > > +{ > > + GEM_BUG_ON(!i915_gem_object_has_pages(obj)); > > + GEM_BUG_ON(!list_empty(&obj->mm.link)); > > + > > + if (i915_gem_object_is_shrinkable(obj)) { > > + struct drm_i915_private *i915 = to_i915(obj->base.dev); > > + unsigned long flags; > > + > > + spin_lock_irqsave(&i915->mm.obj_lock, flags); > > + GEM_BUG_ON(!kref_read(&obj->base.refcount)); > > + > > + list_add_tail(&obj->mm.link, &i915->mm.purge_list); > > + i915->mm.shrink_count++; > > + i915->mm.shrink_memory += obj->base.size; > > + > > + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); > > + } > > +} > > Common helper for the above to passing in the correct list from each? It's also worth making that has_pages into a has_pinned_pages. > > diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c > > index 9e4f51ce52ff..9830edda1ade 100644 > > --- a/drivers/gpu/drm/i915/gt/intel_context.c > > +++ b/drivers/gpu/drm/i915/gt/intel_context.c > > @@ -118,7 +118,7 @@ static int __context_pin_state(struct i915_vma *vma) > > * And mark it as a globally pinned object to let the shrinker know > > * it cannot reclaim the object until we release it. > > */ > > - vma->obj->pin_global++; > > + i915_vma_make_unshrinkable(vma); > > vma->obj->mm.dirty = true; > > > > return 0; > > @@ -126,8 +126,8 @@ static int __context_pin_state(struct i915_vma *vma) > > > > static void __context_unpin_state(struct i915_vma *vma) > > { > > - vma->obj->pin_global--; > > __i915_vma_unpin(vma); > > + i915_vma_make_shrinkable(vma); > > } > > > > static void __intel_context_retire(struct i915_active *active) > > diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c > > index f7e69db4019d..5b16b233a059 100644 > > --- a/drivers/gpu/drm/i915/gt/intel_gt.c > > +++ b/drivers/gpu/drm/i915/gt/intel_gt.c > > @@ -231,6 +231,8 @@ int intel_gt_init_scratch(struct intel_gt *gt, unsigned int size) > > if (ret) > > goto err_unref; > > > > + i915_gem_object_make_unshrinkable(obj); > > + > > gt->scratch = vma; > > return 0; > > > > diff --git a/drivers/gpu/drm/i915/gt/intel_ringbuffer.c b/drivers/gpu/drm/i915/gt/intel_ringbuffer.c > > index 38ec11ae6ed7..d8efb88f33f3 100644 > > --- a/drivers/gpu/drm/i915/gt/intel_ringbuffer.c > > +++ b/drivers/gpu/drm/i915/gt/intel_ringbuffer.c > > @@ -1238,7 +1238,7 @@ int intel_ring_pin(struct intel_ring *ring) > > goto err_ring; > > } > > > > - vma->obj->pin_global++; > > + i915_vma_make_unshrinkable(vma); > > > > GEM_BUG_ON(ring->vaddr); > > ring->vaddr = addr; > > @@ -1267,6 +1267,8 @@ void intel_ring_reset(struct intel_ring *ring, u32 tail) > > > > void intel_ring_unpin(struct intel_ring *ring) > > { > > + struct i915_vma *vma = ring->vma; > > + > > if (!atomic_dec_and_test(&ring->pin_count)) > > return; > > > > @@ -1275,18 +1277,17 @@ void intel_ring_unpin(struct intel_ring *ring) > > /* Discard any unused bytes beyond that submitted to hw. */ > > intel_ring_reset(ring, ring->tail); > > > > - GEM_BUG_ON(!ring->vma); > > - i915_vma_unset_ggtt_write(ring->vma); > > - if (i915_vma_is_map_and_fenceable(ring->vma)) > > - i915_vma_unpin_iomap(ring->vma); > > + i915_vma_unset_ggtt_write(vma); > > + if (i915_vma_is_map_and_fenceable(vma)) > > + i915_vma_unpin_iomap(vma); > > else > > - i915_gem_object_unpin_map(ring->vma->obj); > > + i915_gem_object_unpin_map(vma->obj); > > > > GEM_BUG_ON(!ring->vaddr); > > ring->vaddr = NULL; > > > > - ring->vma->obj->pin_global--; > > - i915_vma_unpin(ring->vma); > > + i915_vma_unpin(vma); > > + i915_vma_make_purgeable(vma); > > Why is ring purgeable and scratch or context state shrinkable? Because the ring contains nothing, but context state must be preserved. I was thinking the explicit selection would be clearer. scratch will be thrown away at the end of the driver. I don't see an instance where we make scratch shrinkable (except for a brief period it is an internal object but is pinned). > > +void i915_vma_make_unshrinkable(struct i915_vma *vma) > > +{ > > + i915_gem_object_make_unshrinkable(vma->obj); > > +} > > + > > +void i915_vma_make_shrinkable(struct i915_vma *vma) > > +{ > > + i915_gem_object_make_shrinkable(vma->obj); > > +} > > + > > +void i915_vma_make_purgeable(struct i915_vma *vma) > > +{ > > + i915_gem_object_make_purgeable(vma->obj); > > +} > > Would i915_vma_make_*object*_... be a better name? I am thinking the > concept does not apply to vma's. I'm planning ahead for a common backing store shared between objects and vma, where vma doesn't operate on the object per-se, and we have a first-class reference counted vma. -Chris
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c index d5197a2a106f..4ea97fca9c35 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c @@ -63,6 +63,8 @@ void i915_gem_object_init(struct drm_i915_gem_object *obj, spin_lock_init(&obj->vma.lock); INIT_LIST_HEAD(&obj->vma.list); + INIT_LIST_HEAD(&obj->mm.link); + INIT_LIST_HEAD(&obj->lut_list); INIT_LIST_HEAD(&obj->batch_pool_link); @@ -273,14 +275,7 @@ void i915_gem_free_object(struct drm_gem_object *gem_obj) * or else we may oom whilst there are plenty of deferred * freed objects. */ - if (i915_gem_object_has_pages(obj) && - i915_gem_object_is_shrinkable(obj)) { - unsigned long flags; - - spin_lock_irqsave(&i915->mm.obj_lock, flags); - list_del_init(&obj->mm.link); - spin_unlock_irqrestore(&i915->mm.obj_lock, flags); - } + i915_gem_object_make_unshrinkable(obj); /* * Since we require blocking on struct_mutex to unbind the freed diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.h b/drivers/gpu/drm/i915/gem/i915_gem_object.h index 67aea07ea019..3714cf234d64 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_object.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_object.h @@ -394,6 +394,10 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, unsigned int flags); void i915_gem_object_unpin_from_display_plane(struct i915_vma *vma); +void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj); +void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj); +void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj); + static inline bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) { if (obj->cache_dirty) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_pages.c b/drivers/gpu/drm/i915/gem/i915_gem_pages.c index b36ad269f4ea..92ad3cc220e3 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_pages.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_pages.c @@ -153,24 +153,13 @@ static void __i915_gem_object_reset_page_iter(struct drm_i915_gem_object *obj) struct sg_table * __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj) { - struct drm_i915_private *i915 = to_i915(obj->base.dev); struct sg_table *pages; pages = fetch_and_zero(&obj->mm.pages); if (IS_ERR_OR_NULL(pages)) return pages; - if (i915_gem_object_is_shrinkable(obj)) { - unsigned long flags; - - spin_lock_irqsave(&i915->mm.obj_lock, flags); - - list_del(&obj->mm.link); - i915->mm.shrink_count--; - i915->mm.shrink_memory -= obj->base.size; - - spin_unlock_irqrestore(&i915->mm.obj_lock, flags); - } + i915_gem_object_make_unshrinkable(obj); if (obj->mm.mapping) { void *ptr; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c index 3f4c6bdcc3c3..14abfd77365f 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_shrinker.c @@ -530,3 +530,60 @@ void i915_gem_shrinker_taints_mutex(struct drm_i915_private *i915, if (unlock) mutex_release(&i915->drm.struct_mutex.dep_map, 0, _RET_IP_); } + +void i915_gem_object_make_unshrinkable(struct drm_i915_gem_object *obj) +{ + if (!list_empty(&obj->mm.link)) { + struct drm_i915_private *i915 = to_i915(obj->base.dev); + unsigned long flags; + + spin_lock_irqsave(&i915->mm.obj_lock, flags); + GEM_BUG_ON(list_empty(&obj->mm.link)); + + list_del_init(&obj->mm.link); + i915->mm.shrink_count--; + i915->mm.shrink_memory -= obj->base.size; + + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); + } +} + +void i915_gem_object_make_shrinkable(struct drm_i915_gem_object *obj) +{ + GEM_BUG_ON(!i915_gem_object_has_pages(obj)); + GEM_BUG_ON(!list_empty(&obj->mm.link)); + + if (i915_gem_object_is_shrinkable(obj)) { + struct drm_i915_private *i915 = to_i915(obj->base.dev); + unsigned long flags; + + spin_lock_irqsave(&i915->mm.obj_lock, flags); + GEM_BUG_ON(!kref_read(&obj->base.refcount)); + + list_add_tail(&obj->mm.link, &i915->mm.shrink_list); + i915->mm.shrink_count++; + i915->mm.shrink_memory += obj->base.size; + + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); + } +} + +void i915_gem_object_make_purgeable(struct drm_i915_gem_object *obj) +{ + GEM_BUG_ON(!i915_gem_object_has_pages(obj)); + GEM_BUG_ON(!list_empty(&obj->mm.link)); + + if (i915_gem_object_is_shrinkable(obj)) { + struct drm_i915_private *i915 = to_i915(obj->base.dev); + unsigned long flags; + + spin_lock_irqsave(&i915->mm.obj_lock, flags); + GEM_BUG_ON(!kref_read(&obj->base.refcount)); + + list_add_tail(&obj->mm.link, &i915->mm.purge_list); + i915->mm.shrink_count++; + i915->mm.shrink_memory += obj->base.size; + + spin_unlock_irqrestore(&i915->mm.obj_lock, flags); + } +} diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c index 9e4f51ce52ff..9830edda1ade 100644 --- a/drivers/gpu/drm/i915/gt/intel_context.c +++ b/drivers/gpu/drm/i915/gt/intel_context.c @@ -118,7 +118,7 @@ static int __context_pin_state(struct i915_vma *vma) * And mark it as a globally pinned object to let the shrinker know * it cannot reclaim the object until we release it. */ - vma->obj->pin_global++; + i915_vma_make_unshrinkable(vma); vma->obj->mm.dirty = true; return 0; @@ -126,8 +126,8 @@ static int __context_pin_state(struct i915_vma *vma) static void __context_unpin_state(struct i915_vma *vma) { - vma->obj->pin_global--; __i915_vma_unpin(vma); + i915_vma_make_shrinkable(vma); } static void __intel_context_retire(struct i915_active *active) diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c index f7e69db4019d..5b16b233a059 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt.c +++ b/drivers/gpu/drm/i915/gt/intel_gt.c @@ -231,6 +231,8 @@ int intel_gt_init_scratch(struct intel_gt *gt, unsigned int size) if (ret) goto err_unref; + i915_gem_object_make_unshrinkable(obj); + gt->scratch = vma; return 0; diff --git a/drivers/gpu/drm/i915/gt/intel_ringbuffer.c b/drivers/gpu/drm/i915/gt/intel_ringbuffer.c index 38ec11ae6ed7..d8efb88f33f3 100644 --- a/drivers/gpu/drm/i915/gt/intel_ringbuffer.c +++ b/drivers/gpu/drm/i915/gt/intel_ringbuffer.c @@ -1238,7 +1238,7 @@ int intel_ring_pin(struct intel_ring *ring) goto err_ring; } - vma->obj->pin_global++; + i915_vma_make_unshrinkable(vma); GEM_BUG_ON(ring->vaddr); ring->vaddr = addr; @@ -1267,6 +1267,8 @@ void intel_ring_reset(struct intel_ring *ring, u32 tail) void intel_ring_unpin(struct intel_ring *ring) { + struct i915_vma *vma = ring->vma; + if (!atomic_dec_and_test(&ring->pin_count)) return; @@ -1275,18 +1277,17 @@ void intel_ring_unpin(struct intel_ring *ring) /* Discard any unused bytes beyond that submitted to hw. */ intel_ring_reset(ring, ring->tail); - GEM_BUG_ON(!ring->vma); - i915_vma_unset_ggtt_write(ring->vma); - if (i915_vma_is_map_and_fenceable(ring->vma)) - i915_vma_unpin_iomap(ring->vma); + i915_vma_unset_ggtt_write(vma); + if (i915_vma_is_map_and_fenceable(vma)) + i915_vma_unpin_iomap(vma); else - i915_gem_object_unpin_map(ring->vma->obj); + i915_gem_object_unpin_map(vma->obj); GEM_BUG_ON(!ring->vaddr); ring->vaddr = NULL; - ring->vma->obj->pin_global--; - i915_vma_unpin(ring->vma); + i915_vma_unpin(vma); + i915_vma_make_purgeable(vma); intel_timeline_unpin(ring->timeline); } diff --git a/drivers/gpu/drm/i915/gt/uc/intel_guc.c b/drivers/gpu/drm/i915/gt/uc/intel_guc.c index 83f2c197375f..a5ba0164fd14 100644 --- a/drivers/gpu/drm/i915/gt/uc/intel_guc.c +++ b/drivers/gpu/drm/i915/gt/uc/intel_guc.c @@ -597,6 +597,8 @@ struct i915_vma *intel_guc_allocate_vma(struct intel_guc *guc, u32 size) goto err; } + i915_gem_object_make_unshrinkable(obj); + return vma; err: diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c index 6b84d04a6a28..c43f270085f5 100644 --- a/drivers/gpu/drm/i915/i915_debugfs.c +++ b/drivers/gpu/drm/i915/i915_debugfs.c @@ -363,8 +363,9 @@ static int i915_gem_object_info(struct seq_file *m, void *data) struct drm_i915_private *i915 = node_to_i915(m->private); int ret; - seq_printf(m, "%u shrinkable objects, %llu bytes\n", + seq_printf(m, "%u shrinkable [%u free] objects, %llu bytes\n", i915->mm.shrink_count, + atomic_read(&i915->mm.free_count), i915->mm.shrink_memory); seq_putc(m, '\n'); diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c index eb16a1a93bbc..e27ddbff487e 100644 --- a/drivers/gpu/drm/i915/i915_vma.c +++ b/drivers/gpu/drm/i915/i915_vma.c @@ -1030,6 +1030,21 @@ int i915_vma_unbind(struct i915_vma *vma) return 0; } +void i915_vma_make_unshrinkable(struct i915_vma *vma) +{ + i915_gem_object_make_unshrinkable(vma->obj); +} + +void i915_vma_make_shrinkable(struct i915_vma *vma) +{ + i915_gem_object_make_shrinkable(vma->obj); +} + +void i915_vma_make_purgeable(struct i915_vma *vma) +{ + i915_gem_object_make_purgeable(vma->obj); +} + #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) #include "selftests/i915_vma.c" #endif diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h index 4b769db649bf..a24bd6787ef7 100644 --- a/drivers/gpu/drm/i915/i915_vma.h +++ b/drivers/gpu/drm/i915/i915_vma.h @@ -459,4 +459,8 @@ void i915_vma_parked(struct drm_i915_private *i915); struct i915_vma *i915_vma_alloc(void); void i915_vma_free(struct i915_vma *vma); +void i915_vma_make_unshrinkable(struct i915_vma *vma); +void i915_vma_make_shrinkable(struct i915_vma *vma); +void i915_vma_make_purgeable(struct i915_vma *vma); + #endif
The shrinker cannot touch objects used by the contexts (logical state and ring). Currently we mark those as "pin_global" to let the shrinker skip over them, however, if we remove them from the shrinker lists entirely, we don't event have to include them in our shrink accounting. By keeping the unshrinkable objects in our shrinker tracking, we report a large number of objects available to be shrunk, and leave the shrinker deeply unsatisfied when we fail to reclaim those. The shrinker will persist in trying to reclaim the unavailable objects, forcing the system into a livelock (not even hitting the dread oomkiller). v2: Extend unshrinkable protection for perma-pinned scratch and guc allocations (Tvrtko) Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_object.c | 11 ++-- drivers/gpu/drm/i915/gem/i915_gem_object.h | 4 ++ drivers/gpu/drm/i915/gem/i915_gem_pages.c | 13 +---- drivers/gpu/drm/i915/gem/i915_gem_shrinker.c | 57 ++++++++++++++++++++ drivers/gpu/drm/i915/gt/intel_context.c | 4 +- drivers/gpu/drm/i915/gt/intel_gt.c | 2 + drivers/gpu/drm/i915/gt/intel_ringbuffer.c | 17 +++--- drivers/gpu/drm/i915/gt/uc/intel_guc.c | 2 + drivers/gpu/drm/i915/i915_debugfs.c | 3 +- drivers/gpu/drm/i915/i915_vma.c | 15 ++++++ drivers/gpu/drm/i915/i915_vma.h | 4 ++ 11 files changed, 101 insertions(+), 31 deletions(-)