Skip to contentSkip to conversation

drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

PATCH Superseded [email protected] 4 participants 18 messages Apr 2, 2014 – May 16, 2014

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

PATCH 1/1 drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

Issue: VIZ-3772

---
 drivers/gpu/drm/i915/i915_gem.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_gem.c +5 -1
index c70121d..1d161c7 100644
@@ -3619,6 +3619,10 @@ unlock:
 
 static bool is_pin_display(struct drm_i915_gem_object *obj)
 {
+	struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
+	if (!vma)
+		return false;
+
 	/* There are 3 sources that pin objects:
 	 *   1. The display engine (scanouts, sprites, cursors);
 	 *   2. Reservations for execbuffer;
@@ -3630,7 +3634,7 @@ static bool is_pin_display(struct drm_i915_gem_object *obj)
 	 * subtracting the potential reference by the user, any pin_count
 	 * remains, it must be due to another use by the display engine.
 	 */
-	return i915_gem_obj_to_ggtt(obj)->pin_count - !!obj->user_pin_count;
+	return vma->pin_count - !!obj->user_pin_count;
 }
 
 /*
Signature · 1 line
1.9.0
#1 lore raw
Quoted 3 lines · On Wed, Apr 02, 2014 at 07:21:01PM +0100, [email protected] wrote:

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

Or we could just do
was_pin_display = obj->pin_display;
obj->pin_display = true;

err_unpin_display:
obj->pin_display = was_pin_display;

And then we wouldn't even need the essay... However, it is almost a theory of operation and quite useful...
-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#2 lore raw
In reply to [email protected]
Quoted 7 lines · On Wed, Apr 02, 2014 at 07:21:01PM +0100, [email protected] wrote:

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

Quoted 7 lines

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

Issue: VIZ-3772

Have you looked into provoking this with an igt testcase? On a hunch a busy load (to extend the race window) plus the usual interruptor trick to jump out of wait_seqno calls should be able to make this go kaboom on command. But I haven't analyzed the bug in detail.

Thanks, Daniel

Quoted 35 lines
---
 drivers/gpu/drm/i915/i915_gem.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/gpu/drm/i915/i915_gem.c +5 -1
index c70121d..1d161c7 100644
@@ -3619,6 +3619,10 @@ unlock:
 
 static bool is_pin_display(struct drm_i915_gem_object *obj)
 {
+	struct i915_vma *vma = i915_gem_obj_to_ggtt(obj);
+	if (!vma)
+		return false;
+
 	/* There are 3 sources that pin objects:
 	 *   1. The display engine (scanouts, sprites, cursors);
 	 *   2. Reservations for execbuffer;
@@ -3630,7 +3634,7 @@ static bool is_pin_display(struct drm_i915_gem_object *obj)
 	 * subtracting the potential reference by the user, any pin_count
 	 * remains, it must be due to another use by the display engine.
 	 */
-	return i915_gem_obj_to_ggtt(obj)->pin_count - !!obj->user_pin_count;
+	return vma->pin_count - !!obj->user_pin_count;
 }
 
 /*
Signature · 6 lines
1.9.0

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#3 lore raw

Hi Daniel,

Sorry, this fell through the cracks:

Quoted 16 lines

Subject: Re: [Intel-gfx] [PATCH] drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

Quoted 12 lines · On Wed, Apr 02, 2014 at 07:21:01PM +0100, [email protected] wrote:

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

Quoted 7 lines
Quoted 2 lines

Issue: VIZ-3772

Have you looked into provoking this with an igt testcase? On a hunch a busy load (to extend the race window) plus the usual interruptor trick to jump out of wait_seqno calls should be able to make this go kaboom on command. But I haven't analyzed the bug in detail.

AFAICT, the only sequence where this likely to happen (because we are handling a recently created object) is:

intelfb_alloc -> intel_pin_and_fence_fb_obj -> i915_gem_object_pin_to_display_plane -> i915_gem_object_set_cache_level -> is_pin_display

Now, I think intelfb_alloc only happens during bootup, so reproducing it on IGT would be difficult. I still feel the fix is valid though :)

-- Oscar

#4 lore raw
Quoted 28 lines · Oscar

Hi Daniel,

Sorry, this fell through the cracks:

Quoted 16 lines

Subject: Re: [Intel-gfx] [PATCH] drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

Quoted 12 lines · On Wed, Apr 02, 2014 at 07:21:01PM +0100, [email protected] wrote:

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

Quoted 7 lines
Quoted 2 lines

Issue: VIZ-3772

Have you looked into provoking this with an igt testcase? On a hunch a busy load (to extend the race window) plus the usual interruptor trick to jump out of wait_seqno calls should be able to make this go kaboom on command. But I haven't analyzed the bug in detail.

Quoted 6 lines

AFAICT, the only sequence where this likely to happen (because we are handling a recently created object) is:

intelfb_alloc -> intel_pin_and_fence_fb_obj -> i915_gem_object_pin_to_display_plane -> i915_gem_object_set_cache_level -> is_pin_display

Now, I think intelfb_alloc only happens during bootup, so reproducing it on IGT would be difficult. I still feel the fix is valid though :)

Did you consider my alternative fix of restoring the old value in the error path?
-Chris

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

---------------------------------------------------------------------
Intel Corporation (UK) Limited
Registered No. 1134945 (England)
Registered Office: Pipers Way, Swindon SN3 1RJ
VAT No: 860 2173 47

Quoted 45 lines

-----Original Message-----
From: Chris Wilson [mailto:[email protected]]
Sent: Monday, May 12, 2014 11:09 AM
To: Mateo Lozano, Oscar
Cc: Daniel Vetter; [email protected]
Subject: Re: [Intel-gfx] [PATCH] drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

Quoted 8 lines · Oscar

Hi Daniel,

Sorry, this fell through the cracks:

Quoted 4 lines

Subject: Re: [Intel-gfx] [PATCH] drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

On Wed, Apr 02, 2014 at 07:21:01PM +0100, [email protected]

Quoted 15 lines · wrote:
Quoted 15 lines
Quoted 14 lines

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

Issue: VIZ-3772

Quoted 6 lines
Quoted 4 lines

Have you looked into provoking this with an igt testcase? On a hunch a busy load (to extend the race window) plus the usual interruptor trick to jump out of wait_seqno calls should be able to make this go kaboom on command. But I haven't analyzed the bug in detail.

AFAICT, the only sequence where this likely to happen (because we are

handling a recently created object) is:

Quoted 5 lines

intelfb_alloc -> intel_pin_and_fence_fb_obj ->
i915_gem_object_pin_to_display_plane ->
i915_gem_object_set_cache_level -> is_pin_display

Quoted 6 lines
Quoted 1 line

Now, I think intelfb_alloc only happens during bootup, so reproducing it on

IGT would be difficult.

Quoted 1 line

I still feel the fix is valid though :)

Did you consider my alternative fix of restoring the old value in the error path? -Chris

Is that directed to Daniel or me? Restoring the old value is way easier, but I thought you wanted to keep is_pin_display as a theory of operation? And Daniel might still want an IGT test, I suppose...

#6 lore raw
Quoted 4 lines · Oscar
Quoted 1 line

Did you consider my alternative fix of restoring the old value in the error path?

Is that directed to Daniel or me? Restoring the old value is way easier, but I thought you wanted to keep is_pin_display as a theory of operation? And Daniel might still want an IGT test, I suppose...

I'd like to be able to kill is_pin_display(), it's very fragile but I don't have a better alternative yet.
-Chris

Signature · 1 line
Chris Wilson, Intel Open Source Technology Centre
#7 lore raw
In reply to Mateo Lozano, Oscar
Quoted 23 lines · Oscar

Hi Daniel,

Sorry, this fell through the cracks:

Quoted 19 lines

Subject: Re: [Intel-gfx] [PATCH] drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

Quoted 14 lines · On Wed, Apr 02, 2014 at 07:21:01PM +0100, [email protected] wrote:

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

Issue: VIZ-3772

Quoted 8 lines
Quoted 4 lines

Have you looked into provoking this with an igt testcase? On a hunch a busy load (to extend the race window) plus the usual interruptor trick to jump out of wait_seqno calls should be able to make this go kaboom on command. But I haven't analyzed the bug in detail.

AFAICT, the only sequence where this likely to happen (because we are handling a recently created object) is:

intelfb_alloc -> intel_pin_and_fence_fb_obj -> i915_gem_object_pin_to_display_plane -> i915_gem_object_set_cache_level -> is_pin_display

Pageflipping to a freshly allocated BO without ever touching it beforehand should be able to achive the same. If this is really all that's needed.

But looking at the code a better way should be:
1. Create new bo, wrap it in a kms fb.
2. Slap busy load onto that bo, e.g. reapeatedly fill it with the blitter.
3. Enable evil interruptor (igt_fork_signal_helper).
4. Submit pageflip

-> Boom since the set_cache_level will block, get interrupted and exit early with -EINTR.

Given sufficient overkill in 2. this should be 100% reliable to reproduce.

Cheers, Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#8 lore raw
Quoted 42 lines · Daniel Vetter
Quoted 23 lines · Oscar

Hi Daniel,

Sorry, this fell through the cracks:

Quoted 19 lines

Subject: Re: [Intel-gfx] [PATCH] drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

Quoted 14 lines · On Wed, Apr 02, 2014 at 07:21:01PM +0100, [email protected] wrote:

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

Issue: VIZ-3772

Quoted 8 lines
Quoted 4 lines

Have you looked into provoking this with an igt testcase? On a hunch a busy load (to extend the race window) plus the usual interruptor trick to jump out of wait_seqno calls should be able to make this go kaboom on command. But I haven't analyzed the bug in detail.

AFAICT, the only sequence where this likely to happen (because we are handling a recently created object) is:

intelfb_alloc -> intel_pin_and_fence_fb_obj -> i915_gem_object_pin_to_display_plane -> i915_gem_object_set_cache_level -> is_pin_display

Pageflipping to a freshly allocated BO without ever touching it beforehand should be able to achive the same. If this is really all that's needed.

But looking at the code a better way should be:
1. Create new bo, wrap it in a kms fb.
2. Slap busy load onto that bo, e.g. reapeatedly fill it with the blitter.
3. Enable evil interruptor (igt_fork_signal_helper).
4. Submit pageflip

Quoted 4 lines

-> Boom since the set_cache_level will block, get interrupted and exit early with -EINTR.

Given sufficient overkill in 2. this should be 100% reliable to reproduce.

Aside: Those kinds of tricks are a big reason why I think igts aren't just useful as testcases, but also to really understand how a bug comes about. At least ime finally figuring out the last ingredient to make an igt fully reliably often resulted in a suddenly much clearer understanding of the bug at hand.

I call this "review by asking for an igt" ;-)
-Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#9 lore raw
Quoted 1 line

I call this "review by asking for an igt" ;-) -Daniel

Ok, I´ll give it a try. At least I will learn something about the kms code, a.k.a. "learning by igt" :D

#10 lore raw
In reply to Daniel Vetter
Quoted 10 lines
Quoted 10 lines

But looking at the code a better way should be:
1. Create new bo, wrap it in a kms fb.
2. Slap busy load onto that bo, e.g. reapeatedly fill it with the blitter.
3. Enable evil interruptor (igt_fork_signal_helper).
4. Submit pageflip

-> Boom since the set_cache_level will block, get interrupted and exit early with -EINTR.

Given sufficient overkill in 2. this should be 100% reliable to reproduce.

As soon as I execbuffer to the bo, it gets a vma for the GGTT vm:

	vm = ctx->vm;
	if (!USES_FULL_PPGTT(dev))
		vm = &dev_priv->gtt.base;

	...

	/* Look up object handles */
	ret = eb_lookup_vmas(eb, exec, args, vm, file);
	if (ret)
		goto err;

And then it becomes impossible to reproduce the problem :( Is there any other trick to make set_cache_level fail?

-- Oscar

#11 lore raw
Quoted 24 lines · Oscar
Quoted 10 lines
Quoted 10 lines

But looking at the code a better way should be:
1. Create new bo, wrap it in a kms fb.
2. Slap busy load onto that bo, e.g. reapeatedly fill it with the blitter.
3. Enable evil interruptor (igt_fork_signal_helper).
4. Submit pageflip

-> Boom since the set_cache_level will block, get interrupted and exit early with -EINTR.

Given sufficient overkill in 2. this should be 100% reliable to reproduce.

As soon as I execbuffer to the bo, it gets a vma for the GGTT vm:

	vm = ctx->vm;
	if (!USES_FULL_PPGTT(dev))
		vm = &dev_priv->gtt.base;

	...

	/* Look up object handles */
	ret = eb_lookup_vmas(eb, exec, args, vm, file);
	if (ret)
		goto err;
Quoted 2 lines

And then it becomes impossible to reproduce the problem :( Is there any other trick to make set_cache_level fail?

What if you make the pin_to_ggtt fail instead? Can you create an object that's too big for the ggtt?

Signature · 2 lines
Ville Syrjälä
Intel OTC
#12 lore raw
Quoted 34 lines

-----Original Message-----
From: Ville Syrjälä [mailto:[email protected]]
Sent: Thursday, May 15, 2014 2:34 PM
To: Mateo Lozano, Oscar
Cc: Daniel Vetter; [email protected]
Subject: Re: [Intel-gfx] [PATCH] drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

Quoted 25 lines · Oscar
Quoted 11 lines
Quoted 11 lines

But looking at the code a better way should be:
1. Create new bo, wrap it in a kms fb.
2. Slap busy load onto that bo, e.g. reapeatedly fill it with the blitter.
3. Enable evil interruptor (igt_fork_signal_helper).
4. Submit pageflip

-> Boom since the set_cache_level will block, get interrupted and -> exit
early with -EINTR.

Given sufficient overkill in 2. this should be 100% reliable to reproduce.

As soon as I execbuffer to the bo, it gets a vma for the GGTT vm:

	vm = ctx->vm;
	if (!USES_FULL_PPGTT(dev))
		vm = &dev_priv->gtt.base;

	...

	/* Look up object handles */
	ret = eb_lookup_vmas(eb, exec, args, vm, file);
	if (ret)
		goto err;
Quoted 5 lines
Quoted 2 lines

And then it becomes impossible to reproduce the problem :( Is there any other trick to make set_cache_level fail?

What if you make the pin_to_ggtt fail instead? Can you create an object that's too big for the ggtt?

Thanks Ville: that worked like a charm ;)

#13 lore raw
In reply to Mateo Lozano, Oscar
Quoted 24 lines · Oscar
Quoted 10 lines
Quoted 10 lines

But looking at the code a better way should be:
1. Create new bo, wrap it in a kms fb.
2. Slap busy load onto that bo, e.g. reapeatedly fill it with the blitter.
3. Enable evil interruptor (igt_fork_signal_helper).
4. Submit pageflip

-> Boom since the set_cache_level will block, get interrupted and exit early with -EINTR.

Given sufficient overkill in 2. this should be 100% reliable to reproduce.

As soon as I execbuffer to the bo, it gets a vma for the GGTT vm:

	vm = ctx->vm;
	if (!USES_FULL_PPGTT(dev))
		vm = &dev_priv->gtt.base;

	...

	/* Look up object handles */
	ret = eb_lookup_vmas(eb, exec, args, vm, file);
	if (ret)
		goto err;
Quoted 2 lines

And then it becomes impossible to reproduce the problem :( Is there any other trick to make set_cache_level fail?

i915.ppgtt=2 should still make this blow up. The bug kinda doesn't exist without full ppgtt I think ...
-Daniel

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
#14 lore raw

PATCH v2 1/1 drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

In reply to [email protected]

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

v2: Chris Wilson says restoring the old value is easier, but that is_pin_display is useful as a theory of operation. Take the solomonic
decision: at least this way is_pin_display is a little more robust (until Chris can kill it off).

Issue: VIZ-3772

---
 drivers/gpu/drm/i915/i915_gem.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +14 -2
index 034ba2c..211b778 100644
@@ -3641,6 +3641,15 @@ unlock:
 
 static bool is_pin_display(struct drm_i915_gem_object *obj)
 {
+	struct i915_vma *vma;
+
+	if (list_empty(&obj->vma_list))
+		return false;
+
+	vma = i915_gem_obj_to_ggtt(obj);
+	if (!vma)
+		return false;
+
 	/* There are 3 sources that pin objects:
 	 *   1. The display engine (scanouts, sprites, cursors);
 	 *   2. Reservations for execbuffer;
@@ -3652,7 +3661,7 @@ static bool is_pin_display(struct drm_i915_gem_object *obj)
 	 * subtracting the potential reference by the user, any pin_count
 	 * remains, it must be due to another use by the display engine.
 	 */
-	return i915_gem_obj_to_ggtt(obj)->pin_count - !!obj->user_pin_count;
+	return vma->pin_count - !!obj->user_pin_count;
 }
 
 /*
@@ -3666,6 +3675,7 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 				     struct intel_ring_buffer *pipelined)
 {
 	u32 old_read_domains, old_write_domain;
+	bool was_pin_display;
 	int ret;
 
 	if (pipelined != obj->ring) {
@@ -3677,6 +3687,7 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	/* Mark the pin_display early so that we account for the
 	 * display coherency whilst setting up the cache domains.
 	 */
+	was_pin_display = obj->pin_display;
 	obj->pin_display = true;
 
 	/* The display engine is not coherent with the LLC cache on gen6.  As
@@ -3719,7 +3730,8 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	return 0;
 
 err_unpin_display:
-	obj->pin_display = is_pin_display(obj);
+	WARN_ON(was_pin_display != is_pin_display(obj));
+	obj->pin_display = was_pin_display;
 	return ret;
 }
 
Signature · 1 line
1.9.0
Reviewed-by on v2 1/1
#15 lore raw
Quoted 12 lines · On Fri, May 16, 2014 at 12:08:22PM +0100, [email protected] wrote:

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

Quoted 6 lines

v2: Chris Wilson says restoring the old value is easier, but that is_pin_display is useful as a theory of operation. Take the solomonic
decision: at least this way is_pin_display is a little more robust (until Chris can kill it off).

Issue: VIZ-3772

I heard you wrote a testcase?

Quoted 10 lines
---
 drivers/gpu/drm/i915/i915_gem.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +0 -0
index 034ba2c..211b778 100644

drivers/gpu/drm/i915/i915_gem.c unlock: · line 3647

Quoted 7 lines
drivers/gpu/drm/i915/i915_gem.c +4 -0
@@ -3641,6 +3641,15 @@ unlock:
 
 static bool is_pin_display(struct drm_i915_gem_object *obj)
 {
+	struct i915_vma *vma;
+
+	if (list_empty(&obj->vma_list))
+		return false;

Hmm, this is so that we don't trigger the WARN from inside i915_gem_obj_to_ggtt(). I would say that means the WARN in the callee has outlived its usefulness. Other callers WARN if they fail to find the ggtt_vma they expect, so I think we can just drop the WARN and save the duplication here.

Quoted 34 lines
drivers/gpu/drm/i915/i915_gem.c +8 -1
+
+	vma = i915_gem_obj_to_ggtt(obj);
+	if (!vma)
+		return false;
+
 	/* There are 3 sources that pin objects:
 	 *   1. The display engine (scanouts, sprites, cursors);
 	 *   2. Reservations for execbuffer;
@@ -3652,7 +3661,7 @@ static bool is_pin_display(struct drm_i915_gem_object *obj)
 	 * subtracting the potential reference by the user, any pin_count
 	 * remains, it must be due to another use by the display engine.
 	 */
-	return i915_gem_obj_to_ggtt(obj)->pin_count - !!obj->user_pin_count;
+	return vma->pin_count - !!obj->user_pin_count;
 }
 
 /*
@@ -3666,6 +3675,7 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 				     struct intel_ring_buffer *pipelined)
 {
 	u32 old_read_domains, old_write_domain;
+	bool was_pin_display;
 	int ret;
 
 	if (pipelined != obj->ring) {
@@ -3677,6 +3687,7 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	/* Mark the pin_display early so that we account for the
 	 * display coherency whilst setting up the cache domains.
 	 */
+	was_pin_display = obj->pin_display;
 	obj->pin_display = true;
 
 	/* The display engine is not coherent with the LLC cache on gen6.  As

drivers/gpu/drm/i915/i915_gem.c i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj, · line 3736

Quoted 8 lines
drivers/gpu/drm/i915/i915_gem.c +2 -1
@@ -3719,7 +3730,8 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	return 0;
 
 err_unpin_display:
-	obj->pin_display = is_pin_display(obj);
+	WARN_ON(was_pin_display != is_pin_display(obj));
+	obj->pin_display = was_pin_display;
 	return ret;
 }

Ok, this looks like a useful check.

Other than the debate over the placement of the WARN() in i915_gem_obj_to_ggtt() (maybe leave a comment here to remind us to drop the WARN and the check later?),

-Chris

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

PATCH v3 1/1 drm/i915: Gracefully handle obj not bound to GGTT in is_pin_display

In reply to [email protected]

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

The IGT kms_flip/bo-too-big tests for this bug.

v2: Chris Wilson says restoring the old value is easier, but that is_pin_display is useful as a theory of operation. Take the solomonic
decision: at least this way is_pin_display is a little more robust (until Chris can kill it off).

v3: Chris suggests the WARN in i915_gem_obj_to_ggtt has outlived its
usefulness: add a reminder to remove it.

Issue: VIZ-3772

---
 drivers/gpu/drm/i915/i915_gem.c | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +17 -2
index 034ba2c..f6c0351 100644
@@ -3641,6 +3641,15 @@ unlock:
 
 static bool is_pin_display(struct drm_i915_gem_object *obj)
 {
+	struct i915_vma *vma;
+
+	if (list_empty(&obj->vma_list))
+		return false;
+
+	vma = i915_gem_obj_to_ggtt(obj);
+	if (!vma)
+		return false;
+
 	/* There are 3 sources that pin objects:
 	 *   1. The display engine (scanouts, sprites, cursors);
 	 *   2. Reservations for execbuffer;
@@ -3652,7 +3661,7 @@ static bool is_pin_display(struct drm_i915_gem_object *obj)
 	 * subtracting the potential reference by the user, any pin_count
 	 * remains, it must be due to another use by the display engine.
 	 */
-	return i915_gem_obj_to_ggtt(obj)->pin_count - !!obj->user_pin_count;
+	return vma->pin_count - !!obj->user_pin_count;
 }
 
 /*
@@ -3666,6 +3675,7 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 				     struct intel_ring_buffer *pipelined)
 {
 	u32 old_read_domains, old_write_domain;
+	bool was_pin_display;
 	int ret;
 
 	if (pipelined != obj->ring) {
@@ -3677,6 +3687,7 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	/* Mark the pin_display early so that we account for the
 	 * display coherency whilst setting up the cache domains.
 	 */
+	was_pin_display = obj->pin_display;
 	obj->pin_display = true;
 
 	/* The display engine is not coherent with the LLC cache on gen6.  As
@@ -3719,7 +3730,8 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	return 0;
 
 err_unpin_display:
-	obj->pin_display = is_pin_display(obj);
+	WARN_ON(was_pin_display != is_pin_display(obj));
+	obj->pin_display = was_pin_display;
 	return ret;
 }
 
@@ -5115,6 +5127,9 @@ struct i915_vma *i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj)
 {
 	struct i915_vma *vma;
 
+	/* This WARN has probably outlived its usefulness (callers already
+	 * WARN if they don't find the GGTT vma they expect). When removing,
+	 * remember to remove the pre-check in is_pin_display() as well */
 	if (WARN_ON(list_empty(&obj->vma_list)))
 		return NULL;
 
Signature · 1 line
1.9.0
#17 lore raw
Quoted 19 lines · On Fri, May 16, 2014 at 02:20:43PM +0100, [email protected] wrote:

From: Oscar Mateo <[email protected]>

Otherwise, we do a NULL pointer dereference.

I've seen this happen while handling an error in
i915_gem_object_pin_to_display_plane():

If i915_gem_object_set_cache_level() fails, we call is_pin_display() to handle the error. At this point, the object is still not pinned to GGTT and maybe not even bound, so we have to check before we dereference its GGTT vma.

The IGT kms_flip/bo-too-big tests for this bug.

v2: Chris Wilson says restoring the old value is easier, but that is_pin_display is useful as a theory of operation. Take the solomonic
decision: at least this way is_pin_display is a little more robust (until Chris can kill it off).

Quoted 5 lines

v3: Chris suggests the WARN in i915_gem_obj_to_ggtt has outlived its
usefulness: add a reminder to remove it.

Issue: VIZ-3772

Queued for -next, thanks for the patch.
-Daniel

Quoted 76 lines
---
 drivers/gpu/drm/i915/i915_gem.c | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)
drivers/gpu/drm/i915/i915_gem.c +17 -2
index 034ba2c..f6c0351 100644
@@ -3641,6 +3641,15 @@ unlock:
 
 static bool is_pin_display(struct drm_i915_gem_object *obj)
 {
+	struct i915_vma *vma;
+
+	if (list_empty(&obj->vma_list))
+		return false;
+
+	vma = i915_gem_obj_to_ggtt(obj);
+	if (!vma)
+		return false;
+
 	/* There are 3 sources that pin objects:
 	 *   1. The display engine (scanouts, sprites, cursors);
 	 *   2. Reservations for execbuffer;
@@ -3652,7 +3661,7 @@ static bool is_pin_display(struct drm_i915_gem_object *obj)
 	 * subtracting the potential reference by the user, any pin_count
 	 * remains, it must be due to another use by the display engine.
 	 */
-	return i915_gem_obj_to_ggtt(obj)->pin_count - !!obj->user_pin_count;
+	return vma->pin_count - !!obj->user_pin_count;
 }
 
 /*
@@ -3666,6 +3675,7 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 				     struct intel_ring_buffer *pipelined)
 {
 	u32 old_read_domains, old_write_domain;
+	bool was_pin_display;
 	int ret;
 
 	if (pipelined != obj->ring) {
@@ -3677,6 +3687,7 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	/* Mark the pin_display early so that we account for the
 	 * display coherency whilst setting up the cache domains.
 	 */
+	was_pin_display = obj->pin_display;
 	obj->pin_display = true;
 
 	/* The display engine is not coherent with the LLC cache on gen6.  As
@@ -3719,7 +3730,8 @@ i915_gem_object_pin_to_display_plane(struct drm_i915_gem_object *obj,
 	return 0;
 
 err_unpin_display:
-	obj->pin_display = is_pin_display(obj);
+	WARN_ON(was_pin_display != is_pin_display(obj));
+	obj->pin_display = was_pin_display;
 	return ret;
 }
 
@@ -5115,6 +5127,9 @@ struct i915_vma *i915_gem_obj_to_ggtt(struct drm_i915_gem_object *obj)
 {
 	struct i915_vma *vma;
 
+	/* This WARN has probably outlived its usefulness (callers already
+	 * WARN if they don't find the GGTT vma they expect). When removing,
+	 * remember to remove the pre-check in is_pin_display() as well */
 	if (WARN_ON(list_empty(&obj->vma_list)))
 		return NULL;
 
Signature · 6 lines
1.9.0

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch