Skip to contentSkip to conversation

drm/gpuvm: Do not prepare NULL objects

PATCH In review Jonathan Cavitt 4 participants 6 messages Jan 30 – Apr 17

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/gpuvm: Do not prepare NULL objects

Statis analysis issue:

drm_gpuvm_prepare_range issues an exec_object_prepare call to all drm_gem_objects mapped between addr and addr + range. However, it is possible (albeit very unlikely) that the objects found through drm_gpuvm_for_each_va_range (as connected to va->gem) are NULL, as seen in other functions such as drm_gpuva_link and drm_gpuva_unlink_defer.

Do not prepare NULL objects.

Fixes 50c1a36f594b ("drm/gpuvm: track/lock/validate external/evicted objects")
Cc Matthew Brost <[email protected]>
Cc Thomas Hellström <[email protected]>
---
 drivers/gpu/drm/drm_gpuvm.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/drm_gpuvm.c +3 -0
index 14469765a780..76f70686f0e6 100644
@@ -1322,6 +1322,9 @@ drm_gpuvm_prepare_range(struct drm_gpuvm *gpuvm, struct drm_exec *exec,
 	drm_gpuvm_for_each_va_range(va, gpuvm, addr, end) {
 		struct drm_gem_object *obj = va->gem.obj;
 
+		if (unlikely(!obj))
+			continue;
+
 		ret = exec_prepare_obj(exec, obj, num_fences);
 		if (ret)
 			return ret;
Signature · 1 line
2.43.0
Reviewed-by on v1 1/1
#1 lore raw

Hi Jonathan,

Quoted 35 lines · Jonathan Cavitt

Statis analysis issue:

drm_gpuvm_prepare_range issues an exec_object_prepare call to all drm_gem_objects mapped between addr and addr + range. However, it is possible (albeit very unlikely) that the objects found through drm_gpuvm_for_each_va_range (as connected to va->gem) are NULL, as seen in other functions such as drm_gpuva_link and drm_gpuva_unlink_defer.

Do not prepare NULL objects.

Fixes 50c1a36f594b ("drm/gpuvm: track/lock/validate external/evicted objects")
Cc Matthew Brost <[email protected]>
Cc Thomas Hellström <[email protected]>
---
 drivers/gpu/drm/drm_gpuvm.c | 3 +++
 1 file changed, 3 insertions(+)
drivers/gpu/drm/drm_gpuvm.c +3 -0
index 14469765a780..76f70686f0e6 100644
@@ -1322,6 +1322,9 @@ drm_gpuvm_prepare_range(struct drm_gpuvm *gpuvm, struct drm_exec *exec,
 	drm_gpuvm_for_each_va_range(va, gpuvm, addr, end) {
 		struct drm_gem_object *obj = va->gem.obj;
 
+		if (unlikely(!obj))
+			continue;
+
 		ret = exec_prepare_obj(exec, obj, num_fences);
 		if (ret)
 			return ret;
Signature · 1 line
2.43.0

Since there alerady such checks in other functions, as you pointed out above and the change looks sane:

Signature · 2 lines
Best Regards,
Krzysztof
maintainer
#2 lore raw
In reply to Jonathan Cavitt

(Cc: Alice)

Quoted 5 lines · Jonathan Cavitt

Statis analysis issue:

drm_gpuvm_prepare_range issues an exec_object_prepare call to all drm_gem_objects mapped between addr and addr + range. However, it is possible (albeit very unlikely) that the objects found through

It is not as unlikely as you might think. The documentation of drm_gpuvm_for_each_va_range() explicitly states:

"This iterator does not skip over the &drm_gpuvm's @kernel_alloc_node."

And the kernel_alloc_node does not have a GEM object set. Also, drivers are free to insert VAs that do not have a GEM object set, e.g. when they have the DRM_GPUVA_SPARSE flag set.

Quoted 2 lines

drm_gpuvm_for_each_va_range (as connected to va->gem) are NULL, as seen in other functions such as drm_gpuva_link and drm_gpuva_unlink_defer.

That said, there is a "real" justification, other than "we do the same thing in other functions" that would have been good to put this into the commit message.

The reason I say "would have been" is because this patch has already been applied -- unfortunately without going through the corresponding maintainer (and non-Intel reviewers). I only noticed it by accident as I saw it pop up in drm-misc-next.

No action needed as the code itself is obviously correct, but for the next time, please make sure to send patches to the corresponding maintainers and reviewers.

- Danilo

#3 lore raw

-----Original Message-----
From: Danilo Krummrich <[email protected]>
Sent: Thursday, April 16, 2026 1:01 PM
To: Cavitt, Jonathan <[email protected]>
Cc: [email protected]; Gupta, Saurabhg <[email protected]>; Zuo, Alex <[email protected]>; Brost, Matthew <[email protected]>; [email protected]; [email protected]; dri-devel <[email protected]>
Subject: Re: [PATCH] drm/gpuvm: Do not prepare NULL objects

Quoted 25 lines

(Cc: Alice)

Quoted 5 lines · Jonathan Cavitt

Statis analysis issue:

drm_gpuvm_prepare_range issues an exec_object_prepare call to all drm_gem_objects mapped between addr and addr + range. However, it is possible (albeit very unlikely) that the objects found through

It is not as unlikely as you might think. The documentation of drm_gpuvm_for_each_va_range() explicitly states:

"This iterator does not skip over the &drm_gpuvm's @kernel_alloc_node."

And the kernel_alloc_node does not have a GEM object set. Also, drivers are free to insert VAs that do not have a GEM object set, e.g. when they have the DRM_GPUVA_SPARSE flag set.

Quoted 2 lines

drm_gpuvm_for_each_va_range (as connected to va->gem) are NULL, as seen in other functions such as drm_gpuva_link and drm_gpuva_unlink_defer.

That said, there is a "real" justification, other than "we do the same thing in other functions" that would have been good to put this into the commit message.

Quoted 7 lines

The reason I say "would have been" is because this patch has already been applied -- unfortunately without going through the corresponding maintainer (and non-Intel reviewers). I only noticed it by accident as I saw it pop up in drm-misc-next.

No action needed as the code itself is obviously correct, but for the next time, please make sure to send patches to the corresponding maintainers and reviewers.

Question for the future:
Is Lankhorst, Maarten a proper maintainer for this component?
If not, is there a list I can reference to determine who the proper maintainers I should be contacting are?
-Jonathan Cavitt

Quoted 3 lines

- Danilo

maintainer
#4 lore raw
Quoted 2 lines · Jonathan Cavitt

Question for the future:
Is Lankhorst, Maarten a proper maintainer for this component?

Maarten is one of the drm-misc tree maintainers.

DRM is a bit special in this regard compared to other kernel subsystems, as DRM also allows non-maintainers to push to development trees. Yet, those development trees still need to be run by someone and those people are the tree maintainers.

Additionally, there are people maintaining the individual components, drivers, etc. and those are usually the people with the strongest domain knowledge.

Patches should be sent to both the tree maintainers and the component maintainers.

Quoted 2 lines

If not, is there a list I can reference to determine who the proper maintainers I should be contacting are?

Absolutely, everything is documented in the MAINTAINERS file in the kernel tree.

The best way to figure out who to send patches to is to use ./scripts/get_maintainer.pl. The script should provide you with the full list of all relevant maintainers and reviewers for your patch.

Additionally, it also outputs people who have been involved in the corresponding component and may be worth Cc'ing as well.

If you run it for your patch you should get the output below.

I hope this helps!

- Danilo

--

	$ ./scripts/get_maintainer.pl ./20260130_jonathan_cavitt_drm_gpuvm_do_not_prepare_null_objects.mbx
	Danilo Krummrich <[email protected]> (maintainer:DRM GPUVM,blamed_fixes:1/1=100%)
	Matthew Brost <[email protected]> (reviewer:DRM GPUVM)
	"Thomas Hellström" <[email protected]> (reviewer:DRM GPUVM,blamed_fixes:1/1=100%)
	Alice Ryhl <[email protected]> (reviewer:DRM GPUVM)
	Maarten Lankhorst <[email protected]> (maintainer:DRM DRIVERS AND MISC GPU PATCHES)
	Maxime Ripard <[email protected]> (maintainer:DRM DRIVERS AND MISC GPU PATCHES)
	Thomas Zimmermann <[email protected]> (maintainer:DRM DRIVERS AND MISC GPU PATCHES)
	David Airlie <[email protected]> (maintainer:DRM DRIVERS)
	Simona Vetter <[email protected]> (maintainer:DRM DRIVERS)
	"Christian König" <[email protected]> (blamed_fixes:1/1=100%)
	Boris Brezillon <[email protected]> (blamed_fixes:1/1=100%)
	[email protected] (open list:DRM GPUVM)
	[email protected] (open list)
	DRM GPUVM status: Supported
#5 lore raw
Quoted 25 lines · Danilo Krummrich
Quoted 2 lines · Jonathan Cavitt

Question for the future:
Is Lankhorst, Maarten a proper maintainer for this component?

Maarten is one of the drm-misc tree maintainers.

DRM is a bit special in this regard compared to other kernel subsystems, as DRM also allows non-maintainers to push to development trees. Yet, those development trees still need to be run by someone and those people are the tree maintainers.

Additionally, there are people maintaining the individual components, drivers, etc. and those are usually the people with the strongest domain knowledge.

Patches should be sent to both the tree maintainers and the component maintainers.

Quoted 2 lines

If not, is there a list I can reference to determine who the proper maintainers I should be contacting are?

Absolutely, everything is documented in the MAINTAINERS file in the kernel tree.

The best way to figure out who to send patches to is to use ./scripts/get_maintainer.pl. The script should provide you with the full list of all relevant maintainers and reviewers for your patch.

Quoted 6 lines

Additionally, it also outputs people who have been involved in the corresponding component and may be worth Cc'ing as well.

If you run it for your patch you should get the output below.

I hope this helps!

That's a great response! Indeed, the get_maintainer.pl is a good tool to use.

Signature · 2 lines
Regards,
Luben