Skip to contentSkip to conversation

drm/i915: Don't forget to apply SNB PIPE_CONTROL GTT workaround.

PATCH In review Eric Anholt 4 participants 5 messages Jul 31, 2012 – Aug 5, 2012

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/i915: Don't forget to apply SNB PIPE_CONTROL GTT workaround.

If a buffer that was the target of a PIPE_CONTROL from userland was a reused one that hadn't been evicted which had not previously had this workaround applied, then we would not bind it into the GTT and the write would land somewhere else.

Based on a doubting-my-sanity debugging session with cworth, I'm pretty sure this will fix his reproducible GL_EXT_timer_query failures, and hopefully the intermittent OQ issues on snb that danvet's been working on.

I have not tested it yet, but hopefully when cworth gets home he will.

---
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |   20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_gem_execbuffer.c +10 -10
index 25b2c54..afb312e 100644
@@ -117,6 +117,16 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 	target_i915_obj = to_intel_bo(target_obj);
 	target_offset = target_i915_obj->gtt_offset;
 
+	/* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
+	 * pipe_control writes because the gpu doesn't properly redirect them
+	 * through the ppgtt for non_secure batchbuffers. */
+	if (unlikely(IS_GEN6(dev) &&
+	    reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
+	    !target_i915_obj->has_global_gtt_mapping)) {
+		i915_gem_gtt_bind_object(target_i915_obj,
+					 target_i915_obj->cache_level);
+	}
+
 	/* The target buffer should have appeared before us in the
 	 * exec_object list, so it should have a GTT space bound by now.
 	 */
@@ -225,16 +235,6 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 		io_mapping_unmap_atomic(reloc_page);
 	}
 
-	/* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
-	 * pipe_control writes because the gpu doesn't properly redirect them
-	 * through the ppgtt for non_secure batchbuffers. */
-	if (unlikely(IS_GEN6(dev) &&
-	    reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
-	    !target_i915_obj->has_global_gtt_mapping)) {
-		i915_gem_gtt_bind_object(target_i915_obj,
-					 target_i915_obj->cache_level);
-	}
-
 	/* and update the user's relocation entry */
 	reloc->presumed_offset = target_offset;
 
Signature · 1 line
1.7.10.4
Reviewed-by on v1 1/1
#1 lore raw
Quoted 11 lines · Eric Anholt

If a buffer that was the target of a PIPE_CONTROL from userland was a reused one that hadn't been evicted which had not previously had this workaround applied, then we would not bind it into the GTT and the write would land somewhere else.

Based on a doubting-my-sanity debugging session with cworth, I'm pretty sure this will fix his reproducible GL_EXT_timer_query failures, and hopefully the intermittent OQ issues on snb that danvet's been working on.

I have not tested it yet, but hopefully when cworth gets home he will.

Had to look hard to find the secret ingredient. The reason why this patch will work is that that in the middle of the error checking we have an early success return after applying the new reloc domains iff the object still holds the same GTT offset as presumed by the reloc.

Probably deserves a little more clarification in the changelog as to why the patch works.

I think this should also fix
https://bugs.freedesktop.org/show_bug.cgi?id=48019
but my snb has decided today is the day to hard hang instead. -Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#3 lore raw

PATCH 1/1 drm/i915: Don't forget to apply SNB PIPE_CONTROL GTT workaround.

In reply to Eric Anholt

If a buffer that was the target of a PIPE_CONTROL from userland was a reused one that hadn't been evicted which had not previously had this workaround applied, then the early return for a correct presumed_offset in this function meant we would not bind it into the GTT and the write would land somewhere else.

Fixes reproducible failures with GL_EXT_timer_query usage in apitrace, and I also expect it to fix the intermittent OQ issues on snb that danvet's been working on.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=48019

---
 drivers/gpu/drm/i915/i915_gem_execbuffer.c |   20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)
drivers/gpu/drm/i915/i915_gem_execbuffer.c +10 -10
index 25b2c54..afb312e 100644
@@ -117,6 +117,16 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 	target_i915_obj = to_intel_bo(target_obj);
 	target_offset = target_i915_obj->gtt_offset;
 
+	/* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
+	 * pipe_control writes because the gpu doesn't properly redirect them
+	 * through the ppgtt for non_secure batchbuffers. */
+	if (unlikely(IS_GEN6(dev) &&
+	    reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
+	    !target_i915_obj->has_global_gtt_mapping)) {
+		i915_gem_gtt_bind_object(target_i915_obj,
+					 target_i915_obj->cache_level);
+	}
+
 	/* The target buffer should have appeared before us in the
 	 * exec_object list, so it should have a GTT space bound by now.
 	 */
@@ -225,16 +235,6 @@ i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 		io_mapping_unmap_atomic(reloc_page);
 	}
 
-	/* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
-	 * pipe_control writes because the gpu doesn't properly redirect them
-	 * through the ppgtt for non_secure batchbuffers. */
-	if (unlikely(IS_GEN6(dev) &&
-	    reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
-	    !target_i915_obj->has_global_gtt_mapping)) {
-		i915_gem_gtt_bind_object(target_i915_obj,
-					 target_i915_obj->cache_level);
-	}
-
 	/* and update the user's relocation entry */
 	reloc->presumed_offset = target_offset;
 
Signature · 1 line
1.7.10.4
#4 lore raw
Quoted 10 lines · Eric Anholt

If a buffer that was the target of a PIPE_CONTROL from userland was a reused one that hadn't been evicted which had not previously had this workaround applied, then the early return for a correct presumed_offset in this function meant we would not bind it into the GTT and the write would land somewhere else.

Fixes reproducible failures with GL_EXT_timer_query usage in apitrace, and I also expect it to fix the intermittent OQ issues on snb that danvet's been working on.

Quoted 5 lines

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=48019

Picked up for -fixes, thanks for the patch. I've also added a bz line for #52932 to the commit.

I should have noticed this while banging against this particular wall, I guess I owe you a few beers for tracking it down ;-)

Cheers, Daniel

Signature · 3 lines
Daniel Vetter
Mail: [email protected]
Mobile: +41 (0)79 365 57 48