[Intel-gfx] [PATCH] drm/i915: Let execlist_update_context() cover !FULL_PPGTT mode.

[Zhi Wang]

Hi Chris:
     Thanks for the explanation! :P Have you already sent the patch to 
keep PD structure under aliasing PPGTT mode? I tried drm-intel-nightly 
branch and still got kernel panic under aliasing PPGTT mode. T_T

I fixed it like this, is this acceptable as a hot fix? If it's 
acceptable as a hot fix, I can send a patch. :P

diff --git a/drivers/gpu/drm/i915/i915_gem_gtt.c 
b/drivers/gpu/drm/i915/i915_gem_gtt.c
index 22b3374..21e06c9 100644
--- a/drivers/gpu/drm/i915/i915_gem_gtt.c
+++ b/drivers/gpu/drm/i915/i915_gem_gtt.c
@@ -2658,6 +2658,19 @@ static int aliasing_gtt_bind_vma(struct i915_vma 
*vma,
  	if (vma->obj->gt_ro)
  		pte_flags |= PTE_READ_ONLY;

+	if (flags & I915_VMA_LOCAL_BIND) {
+		struct i915_hw_ppgtt *appgtt = i915->mm.aliasing_ppgtt;
+
+		ret = appgtt->base.allocate_va_range(&appgtt->base,
+						     vma->node.start,
+						     vma->node.size);
+		if (ret)
+			return ret;
+
+		appgtt->base.insert_entries(&appgtt->base,
+					    vma->pages, vma->node.start,
+					    cache_level, pte_flags);
+	}

  	if (flags & I915_VMA_GLOBAL_BIND) {
  		intel_runtime_pm_get(i915);
@@ -2666,14 +2679,6 @@ static int aliasing_gtt_bind_vma(struct i915_vma 
*vma,
  					cache_level, pte_flags);
  		intel_runtime_pm_put(i915);
  	}
-
-	if (flags & I915_VMA_LOCAL_BIND) {
-		struct i915_hw_ppgtt *appgtt = i915->mm.aliasing_ppgtt;
-		appgtt->base.insert_entries(&appgtt->base,
-					    vma->pages, vma->node.start,
-					    cache_level, pte_flags);
-	}
-
  	return 0;
  }



On 02/06/17 19:38, Chris Wilson wrote:
> On Mon, Feb 06, 2017 at 07:03:39PM +0800, Zhi Wang wrote:
>> Thanks for the merging. I guess that with your patches of keeping PD
>> structure under aliasing PPGTT mode in 32bit page table, the amount
>> of PDPs will not change anymore under aliasing PPGTT mode. :P
>
> Yes. That's the idea atm, we will preallocate the aliasing table
> and then keep the tree intact. The only disadvantage with gen8+ aliasing
> mode (compared to gen6) will be that we still have to walk the va range
> to pin the tree so that when we call clear_range afterwards we don't
> reap. The benefit from my pov, is that the special case for
> aliasing_ppgtt is within the aliasing_ppgtt code and not in the common
> code (which should be optimised for full-ppgtt as that is the
> default/typical use).
> -Chris
>