@@ -4247,6 +4247,10 @@ i915_gem_object_pin_view(struct drm_i915_gem_object *obj,
bound = vma ? vma->bound : 0;
if (vma == NULL || !drm_mm_node_allocated(&vma->node)) {
+ /* In true PPGTT, bind has possibly changed PDEs, which
+ * means we must do a context switch before the GPU can
+ * accurately read some of the VMAs.
+ */
vma = i915_gem_object_bind_to_vm(obj, vm, alignment,
flags, view);
if (IS_ERR(vma))
@@ -573,8 +573,20 @@ static inline bool should_skip_switch(struct intel_engine_cs *ring,
struct intel_context *from,
struct intel_context *to)
{
- if (from == to && !to->remap_slice)
- return true;
+ struct drm_i915_private *dev_priv = ring->dev->dev_private;
+
+ if (to->remap_slice)
+ return false;
+
+ if (to->ppgtt) {
+ if (from == to && !test_bit(ring->id,
+ &to->ppgtt->pd_dirty_rings))
+ return true;
+ } else if (dev_priv->mm.aliasing_ppgtt) {
+ if (from == to && !test_bit(ring->id,
+ &dev_priv->mm.aliasing_ppgtt->pd_dirty_rings))
+ return true;
+ }
return false;
}
@@ -610,10 +622,7 @@ needs_pd_load_post(struct intel_engine_cs *ring, struct intel_context *to)
if (ring != &dev_priv->ring[RCS])
return false;
- if (!to->legacy_hw_ctx.initialized)
- return true;
-
- if (i915_gem_context_is_default(to))
+ if (to->ppgtt->pd_dirty_rings)
return true;
return false;
@@ -664,6 +673,9 @@ static int do_switch(struct intel_engine_cs *ring,
ret = to->ppgtt->switch_mm(to->ppgtt, ring);
if (ret)
goto unpin_out;
+
+ /* Doing a PD load always reloads the page dirs */
+ clear_bit(ring->id, &to->ppgtt->pd_dirty_rings);
}
if (ring != &dev_priv->ring[RCS]) {
@@ -696,6 +708,8 @@ static int do_switch(struct intel_engine_cs *ring,
if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to))
hw_flags |= MI_RESTORE_INHIBIT;
+ else if (to->ppgtt && test_and_clear_bit(ring->id, &to->ppgtt->pd_dirty_rings))
+ hw_flags |= MI_FORCE_RESTORE;
ret = mi_set_context(ring, to, hw_flags);
if (ret)
@@ -1187,6 +1187,13 @@ i915_gem_ringbuffer_submission(struct drm_device *dev, struct drm_file *file,
if (ret)
goto error;
+ if (ctx->ppgtt)
+ WARN(ctx->ppgtt->pd_dirty_rings & (1<<ring->id),
+ "%s didn't clear reload\n", ring->name);
+ else if (dev_priv->mm.aliasing_ppgtt)
+ WARN(dev_priv->mm.aliasing_ppgtt->pd_dirty_rings &
+ (1<<ring->id), "%s didn't clear reload\n", ring->name);
+
instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
instp_mask = I915_EXEC_CONSTANTS_MASK;
switch (instp_mode) {
@@ -1161,6 +1161,16 @@ static void gen6_ppgtt_unmap_pages(struct i915_hw_ppgtt *ppgtt)
4096, PCI_DMA_BIDIRECTIONAL);
}
+/* PDE TLBs are a pain invalidate pre GEN8. It requires a context reload. If we
+ * are switching between contexts with the same LRCA, we also must do a force
+ * restore.
+ */
+static inline void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
+{
+ /* If current vm != vm, */
+ ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask;
+}
+
static int gen6_alloc_va_range(struct i915_address_space *vm,
uint64_t start, uint64_t length)
{
@@ -1180,6 +1190,7 @@ static int gen6_alloc_va_range(struct i915_address_space *vm,
GEN6_PTES);
}
+ mark_tlbs_dirty(ppgtt);
return 0;
}
@@ -300,6 +300,7 @@ struct i915_hw_ppgtt {
struct i915_address_space base;
struct kref ref;
struct drm_mm_node node;
+ unsigned long pd_dirty_rings;
unsigned num_pd_entries;
unsigned num_pd_pages; /* gen8+ */
union {