Skip to contentSkip to conversation

drm/nouveau: nouveau_sched_fini() should check for init failure

PATCH In review Timur Tabi 2 participants 5 messages Jan 31, 2024 – Feb 1, 2024

Subsystems (3)
lore
Share & tools
lore

Timur Tabi

author
#0 lore raw

PATCH 1/1 drm/nouveau: nouveau_sched_fini() should check for init failure

If initialization fails, Nouveau can still call the _fini() function to clean up, with the expectation that the function can handle if its corresponding _init() function was never called or exited with error.

Such is not the case with nouveau_sched_fini(), which still attempts to wait for jobs to finish even if the underlying data structures were never initialized.

Fixes 5f03a507b29e ("drm/nouveau: implement 1:1 scheduler - entity relationship")
---
 drivers/gpu/drm/nouveau/nouveau_sched.c | 5 +++++
 drivers/gpu/drm/nouveau/nouveau_sched.h | 2 ++
 2 files changed, 7 insertions(+)
drivers/gpu/drm/nouveau/nouveau_sched.c +5 -0
index dd98f6910f9c..9c771bc0e332 100644
@@ -443,6 +443,8 @@ nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm,
 	INIT_LIST_HEAD(&sched->job.list.head);
 	init_waitqueue_head(&sched->job.wq);
 
+	sched->initialized = true;
+
 	return 0;
 
 fail_sched:
@@ -459,6 +461,9 @@ nouveau_sched_fini(struct nouveau_sched *sched)
 	struct drm_gpu_scheduler *drm_sched = &sched->base;
 	struct drm_sched_entity *entity = &sched->entity;
 
+	if (!sched->initialized)
+		return;
+
 	rmb(); /* for list_empty to work without lock */
 	wait_event(sched->job.wq, list_empty(&sched->job.list.head));
 
drivers/gpu/drm/nouveau/nouveau_sched.h +2 -0
index a6528f5981e6..351931c706aa 100644
@@ -109,6 +109,8 @@ struct nouveau_sched {
 		} list;
 		struct wait_queue_head wq;
 	} job;
+
+	bool initialized;
 };
 
 int nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm,
Signature · 1 line
2.34.1
#1 lore raw
Quoted 7 lines · Timur Tabi

If initialization fails, Nouveau can still call the _fini() function to clean up, with the expectation that the function can handle if its corresponding _init() function was never called or exited with error.

Such is not the case with nouveau_sched_fini(), which still attempts to wait for jobs to finish even if the underlying data structures were never initialized.

Good catch!

Quoted 13 lines
Fixes 5f03a507b29e ("drm/nouveau: implement 1:1 scheduler - entity relationship")
---
  drivers/gpu/drm/nouveau/nouveau_sched.c | 5 +++++
  drivers/gpu/drm/nouveau/nouveau_sched.h | 2 ++
  2 files changed, 7 insertions(+)
drivers/gpu/drm/nouveau/nouveau_sched.c +0 -0
index dd98f6910f9c..9c771bc0e332 100644

drivers/gpu/drm/nouveau/nouveau_sched.c nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm, · line 446

Quoted 4 lines
drivers/gpu/drm/nouveau/nouveau_sched.c +1 -0
@@ -443,6 +443,8 @@ nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm,
  	INIT_LIST_HEAD(&sched->job.list.head);
  	init_waitqueue_head(&sched->job.wq);
 
+	sched->initialized = true;

I wonder if we should allocate struct nouveau_sched dynamically instead and just check for NULL in the corresponding *_fini() functions.

Actually, in nouveau_abi16_ioctl_channel_alloc() we can omit creating a scheduler instance entirely if !nouveau_cli_uvmm().

- Danilo

Quoted 27 lines
drivers/gpu/drm/nouveau/nouveau_sched.c +4 -0
+
  	return 0;
 
  fail_sched:
@@ -459,6 +461,9 @@ nouveau_sched_fini(struct nouveau_sched *sched)
  	struct drm_gpu_scheduler *drm_sched = &sched->base;
  	struct drm_sched_entity *entity = &sched->entity;
 
+	if (!sched->initialized)
+		return;
+
  	rmb(); /* for list_empty to work without lock */
  	wait_event(sched->job.wq, list_empty(&sched->job.list.head));
 
drivers/gpu/drm/nouveau/nouveau_sched.h +2 -0
index a6528f5981e6..351931c706aa 100644
@@ -109,6 +109,8 @@ struct nouveau_sched {
  		} list;
  		struct wait_queue_head wq;
  	} job;
+
+	bool initialized;
  };
 
  int nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm,
author
#2 lore raw
Quoted 7 lines · Danilo Krummrich

I wonder if we should allocate struct nouveau_sched dynamically instead and just
check for NULL in the corresponding *_fini() functions.

Actually, in nouveau_abi16_ioctl_channel_alloc() we can omit creating a scheduler
instance entirely if !nouveau_cli_uvmm().

I am all for a better solution if you can think of one. I don't really like the idea of "bool initialized", but I couldn't think of a more reliable test.

#3 lore raw
Quoted 11 lines · Timur Tabi
Quoted 7 lines · Danilo Krummrich

I wonder if we should allocate struct nouveau_sched dynamically instead and just
check for NULL in the corresponding *_fini() functions.

Actually, in nouveau_abi16_ioctl_channel_alloc() we can omit creating a scheduler
instance entirely if !nouveau_cli_uvmm().

I am all for a better solution if you can think of one. I don't really like the idea of "bool initialized", but I couldn't think of a more reliable test.

Yeah, it's not really ideal that there is a single error path calling the corresponding *_fini() in nouveau_abi16_ioctl_channel_alloc() and nouveau_cli_init() only. Maybe this can be re-worked to only reverse what has actually been done already.

However, I'd probably not do that for -fixes. Maybe we should go with the dynamic allocation and a NULL check for now.

Do you want to send another patch? Otherwise I can also come up with something and add your Reported-by.

- Danilo

Quoted 2 lines
author
#4 lore raw

On Thu, 2024-02-01 at 17:35 +0100, Danilo Krummrich wrote:
Do you want to send another patch? Otherwise I can also come up with something and add your Reported-by.

I think it's best if you do it, since I'm not at all familiar with the abi16 stuff.

I'm able to repro the problem easily by just having r535_gsp_wpr_meta_init() return an error code instead of 0.