Skip to contentSkip to conversation

drm/i915: Restore all GGTT VMAs on resume

PATCH In review Tvrtko Ursulin 4 participants 8 messages Jul 6, 2015 – Jul 8, 2015

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

PATCH 1/1 drm/i915: Restore all GGTT VMAs on resume

From: Tvrtko Ursulin <[email protected]>

When rotated and partial views were added no one spotted the resume path which assumes only one GGTT VMA per object and hence is now skipping rebind of alternative views.

Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>

---
Similarly to my recent debugfs patch, it would seem quicker path could be to walk GGTT active & inactive lists, but I assume we want to call i915_gem_clflush_object only once per object which would make that problematic.

---
 drivers/gpu/drm/i915/i915_gem_gtt.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/i915_gem_gtt.c +16 -7
index 8cfa3906f440..5742a2a74cf8 100644
@@ -2543,6 +2543,8 @@ void i915_gem_restore_gtt_mappings(struct drm_device *dev)
 	struct drm_i915_private *dev_priv = dev->dev_private;
 	struct drm_i915_gem_object *obj;
 	struct i915_address_space *vm;
+	struct i915_vma *vma;
+	bool flush;
 
 	i915_check_and_clear_faults(dev);
 
@@ -2552,16 +2554,23 @@ void i915_gem_restore_gtt_mappings(struct drm_device *dev)
 				       dev_priv->gtt.base.total,
 				       true);
 
+	/* Cache flush objects bound into GGTT and rebind them. */
+	vm = &dev_priv->gtt.base;
 	list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
-		struct i915_vma *vma = i915_gem_obj_to_vma(obj,
-							   &dev_priv->gtt.base);
-		if (!vma)
-			continue;
+		flush = false;
+		list_for_each_entry(vma, &obj->vma_list, vma_link) {
+			if (vma->vm != vm)
+				continue;
 
-		i915_gem_clflush_object(obj, obj->pin_display);
-		WARN_ON(i915_vma_bind(vma, obj->cache_level, PIN_UPDATE));
-	}
+			WARN_ON(i915_vma_bind(vma, obj->cache_level,
+					      PIN_UPDATE));
 
+			flush = true;
+		}
+
+		if (flush)
+			i915_gem_clflush_object(obj, obj->pin_display);
+	}
 
 	if (INTEL_INFO(dev)->gen >= 8) {
 		if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev))
Signature · 6 lines
2.4.2

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Reviewed-by on v1 1/1
#1 lore raw
Quoted 10 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

When rotated and partial views were added no one spotted the resume path which assumes only one GGTT VMA per object and hence is now skipping rebind of alternative views.

Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>
Quoted 5 lines

---
Similarly to my recent debugfs patch, it would seem quicker path could be to walk GGTT active & inactive lists, but I assume we want to call i915_gem_clflush_object only once per object which would make that problematic.

It's a bit easier because there are no active entries after suspend, and we may not care too much for over clflushing.

We should actually clear the GGTT of all but pinned entries on hibernate (which would save a bunch of work on early resume), and there are a bunch of non-objects inside the GGTT that should be evicted/restored properly. (Would have been trivial had we made them all objects in the first place.)

The patch is in the right direction and the other issues are present irrespective of this patch.

I guess a begrudingly

-Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw
Quoted 17 lines · Chris Wilson
Quoted 10 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

When rotated and partial views were added no one spotted the resume path which assumes only one GGTT VMA per object and hence is now skipping rebind of alternative views.

Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>
Quoted 5 lines

---
Similarly to my recent debugfs patch, it would seem quicker path could be to walk GGTT active & inactive lists, but I assume we want to call i915_gem_clflush_object only once per object which would make that problematic.

Quoted 8 lines

It's a bit easier because there are no active entries after suspend, and we may not care too much for over clflushing.

We should actually clear the GGTT of all but pinned entries on hibernate (which would save a bunch of work on early resume), and there are a bunch of non-objects inside the GGTT that should be evicted/restored properly. (Would have been trivial had we made them all objects in the first place.)

What are these non-objects and where do I find the code with restores them improperly?

Quoted 5 lines

The patch is in the right direction and the other issues are present irrespective of this patch.

I guess a begrudingly

Thanks!

Regards,

Tvrtko
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#3 lore raw
Quoted 22 lines · Tvrtko Ursulin
Quoted 20 lines · Chris Wilson
Quoted 10 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

When rotated and partial views were added no one spotted the resume path which assumes only one GGTT VMA per object and hence is now skipping rebind of alternative views.

Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>
Quoted 5 lines

---
Similarly to my recent debugfs patch, it would seem quicker path could be to walk GGTT active & inactive lists, but I assume we want to call i915_gem_clflush_object only once per object which would make that problematic.

It's a bit easier because there are no active entries after suspend, and we may not care too much for over clflushing.

Quoted 8 lines
Quoted 5 lines

We should actually clear the GGTT of all but pinned entries on hibernate (which would save a bunch of work on early resume), and there are a bunch of non-objects inside the GGTT that should be evicted/restored properly. (Would have been trivial had we made them all objects in the first place.)

What are these non-objects and where do I find the code with restores them improperly?

The PDEs are allocated directly in the GGTT and ignored completely across hibernation. All we ideally need to do is evict them on freeze. -Chris

Signature · 5 lines
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#4 lore raw
In reply to Chris Wilson
Quoted 7 lines · Chris Wilson
Quoted 6 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

When rotated and partial views were added no one spotted the resume path which assumes only one GGTT VMA per object and hence is now skipping rebind of alternative views.

Quoted 9 lines
Quoted 9 lines
Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>

---
Similarly to my recent debugfs patch, it would seem quicker path could be to walk GGTT active & inactive lists, but I assume we want to call i915_gem_clflush_object only once per object which would make that problematic.

Patch is missing offending commit that introduce this issue and which platforms are affected (I guess all due to partial view?). Also should be
cc: stable I presume?

Can you please supply this so I can ammend the commit message? Applied to -fixes meanwhile, thanks.
-Daniel

Quoted 19 lines

It's a bit easier because there are no active entries after suspend, and we may not care too much for over clflushing.

We should actually clear the GGTT of all but pinned entries on hibernate (which would save a bunch of work on early resume), and there are a bunch of non-objects inside the GGTT that should be evicted/restored properly. (Would have been trivial had we made them all objects in the first place.)

The patch is in the right direction and the other issues are present irrespective of this patch.

I guess a begrudingly

-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#5 lore raw
Quoted 18 lines · Daniel Vetter
Quoted 7 lines · Chris Wilson
Quoted 6 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

When rotated and partial views were added no one spotted the resume path which assumes only one GGTT VMA per object and hence is now skipping rebind of alternative views.

Quoted 9 lines
Quoted 9 lines
Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>

---
Similarly to my recent debugfs patch, it would seem quicker path could be to walk GGTT active & inactive lists, but I assume we want to call i915_gem_clflush_object only once per object which would make that problematic.

Quoted 6 lines

Patch is missing offending commit that introduce this issue and which platforms are affected (I guess all due to partial view?). Also should be
cc: stable I presume?

Can you please supply this so I can ammend the commit message? Applied to -fixes meanwhile, thanks.

Ha, "offending" commits. In the sense of that the bug only manifests when they are present they are either of these two:

commit c5ad54cf7dd8922bd1cee2d5871aebf73dc9638e

Author: Joonas Lahtinen <[email protected]>
Date:   Wed May 6 14:36:09 2015 +0300

drm/i915: Use partial view in mmap fault handler

commit 3b7a5119b5d2def1161226a4c6a643db537dff7e

Author: Sonika Jindal <[email protected]>
Date:   Fri Apr 10 14:37:29 2015 +0530
    drm/i915/skl: Support for 90/270 rotation

But they are just the tips of two longish streams of developments
so they should not necessarily be viewed as offending towards the respective authors. They are just final feature enablements.

All platforms are affected due to partial views I suppose, although I am not sure if they are actually used/triggered in the field. That's why I was not sure should stable be copied.

Regards,

Tvrtko
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#6 lore raw
Quoted 42 lines · Tvrtko Ursulin
Quoted 18 lines · Daniel Vetter
Quoted 7 lines · Chris Wilson
Quoted 6 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

When rotated and partial views were added no one spotted the resume path which assumes only one GGTT VMA per object and hence is now skipping rebind of alternative views.

Quoted 9 lines
Quoted 9 lines
Cc Daniel Vetter <[email protected]>
Cc Chris Wilson <[email protected]>
Cc Joonas Lahtinen <[email protected]>

---
Similarly to my recent debugfs patch, it would seem quicker path could be to walk GGTT active & inactive lists, but I assume we want to call i915_gem_clflush_object only once per object which would make that problematic.

Quoted 6 lines

Patch is missing offending commit that introduce this issue and which platforms are affected (I guess all due to partial view?). Also should be
cc: stable I presume?

Can you please supply this so I can ammend the commit message? Applied to -fixes meanwhile, thanks.

Ha, "offending" commits. In the sense of that the bug only manifests when they are present they are either of these two:

commit c5ad54cf7dd8922bd1cee2d5871aebf73dc9638e

Author: Joonas Lahtinen <[email protected]>
Date:   Wed May 6 14:36:09 2015 +0300

drm/i915: Use partial view in mmap fault handler

commit 3b7a5119b5d2def1161226a4c6a643db537dff7e

Author: Sonika Jindal <[email protected]>
Date:   Fri Apr 10 14:37:29 2015 +0530
    drm/i915/skl: Support for 90/270 rotation
Quoted 7 lines

But they are just the tips of two longish streams of developments so they should not necessarily be viewed as offending towards the respective authors. They are just final feature enablements.

All platforms are affected due to partial views I suppose, although I am not sure if they are actually used/triggered in the field. That's why I was not sure should stable be copied.

Well commits are only in 4.2, so -fixes is ok.

Thanks, Daniel

Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Tested-by on v1 1/1
#7 lore raw
In reply to Tvrtko Ursulin

Task id: 6731 -------------------------------------Summary-------------------------------------
Platform Delta drm-intel-nightly Series Applied

ILK                                  302/302              302/302
SNB                                  312/316              312/316
IVB                                  345/345              345/345

BYT -2 289/289 287/289
HSW 382/382 382/382
-------------------------------------Detailed-------------------------------------
Platform Test drm-intel-nightly Series Applied
*BYT igt@gem_partial_pwrite_pread@reads PASS(1) FAIL(1)
*BYT igt@gem_partial_pwrite_pread@reads-uncached PASS(1) FAIL(1)
Note: You need to pay more attention to line start with '*'
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx