Skip to contentSkip to conversation

drm/amdgpu/userq: lock and validate wptr BOs before reading their GPU offset on restore

PATCH In review Jesse Zhang 2 participants 3 messages Aug 14 – Aug 17

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

PATCH 1/1 drm/amdgpu/userq: lock and validate wptr BOs before reading their GPU offset on restore

amdgpu_userq_vm_validate_and_restore_queue() reads each queue's wptr BO GPU offset via amdgpu_bo_gpu_offset() after only calling amdgpu_ttm_alloc_gart() on it. But the wptr BOs are not part of this VM (their reservation object is their own, not vm->root), so neither amdgpu_vm_validate() nor amdgpu_userq_bo_validate() (which only handles the VM's evicted list) covers them. As a result a wptr BO can still be in TTM_PL_SYSTEM and unreserved when its offset is read, tripping the amdgpu_bo_gpu_offset() sanity checks from the restore worker:

  ------------[ cut here ]------------
  WARNING: amdgpu_object.c:1486 at amdgpu_bo_gpu_offset+0x75/0xa0 [amdgpu], CPU#3: kworker/3:1/116
  Workqueue: events amdgpu_userq_restore_worker [amdgpu]
  RIP: 0010:amdgpu_bo_gpu_offset+0x75/0xa0 [amdgpu]
  Call Trace:
   <TASK>
   amdgpu_userq_vm_validate_and_restore_queue+0x629/0x960 [amdgpu]
   amdgpu_userq_restore_worker+0xa6/0x180 [amdgpu]
   process_scheduled_works+0xa6/0x460
   worker_thread+0x13c/0x290
   kthread+0xfb/0x140
   ret_from_fork+0x1b6/0x2b0
   ret_from_fork_asm+0x1a/0x30
   </TASK>
  ---[ end trace 0000000000000000 ]---
  ------------[ cut here ]------------
  WARNING: amdgpu_object.c:1485 at amdgpu_bo_gpu_offset+0x9a/0xa0 [amdgpu], CPU#2: kworker/2:1/127
  Workqueue: events amdgpu_userq_restore_worker [amdgpu]
  RIP: 0010:amdgpu_bo_gpu_offset+0x9a/0xa0 [amdgpu]

amdgpu_ttm_alloc_gart() only creates the GART mapping; it does not migrate the BO out of system memory, so the offset read is bogus (the queue would resume with a wrong wptr address).

Lock each wptr BO into the drm_exec context and validate it into GTT inside the drm_exec_until_all_locked() block, mirroring what the create path (mes_userq_create_wptr_mapping) already does, so that the offset read later is safe and correct.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 26 +++++++++++++++++++++++
 1 file changed, 26 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +26 -0
index 17cc48d87c4d..59aa4802c111 100644
@@ -1054,6 +1054,32 @@ amdgpu_userq_vm_validate_and_restore_queue(struct amdgpu_userq_mgr *uq_mgr)
 		drm_exec_retry_on_contention(&exec);
 		if (unlikely(ret))
 			goto unlock_all;
+
+		/*
+		 * The per-queue wptr BOs are not part of this VM (their resv is
+		 * their own, not vm->root), so the validation above does not
+		 * cover them. Lock and validate each into GTT here so that
+		 * reading its GPU offset below is safe - matching what the
+		 * create path (mes_userq_create_wptr_mapping) does.
+		 */
+		xa_for_each(&uq_mgr->userq_xa, tmp_key, queue) {
+			struct ttm_operation_ctx wptr_ctx = { false, false };
+
+			bo = queue->wptr_obj.obj;
+			if (!bo)
+				continue;
+
+			ret = drm_exec_prepare_obj(&exec, &bo->tbo.base,
+						   TTM_NUM_MOVE_FENCES + 1);
+			drm_exec_retry_on_contention(&exec);
+			if (unlikely(ret))
+				goto unlock_all;
+
+			amdgpu_bo_placement_from_domain(bo, bo->allowed_domains);
+			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &wptr_ctx);
+			if (unlikely(ret))
+				goto unlock_all;
+		}
 	}
 
 	if (invalidated) {
Signature · 1 line
2.49.0
maintainer
#1 lore raw
Quoted 48 lines · Jesse Zhang

amdgpu_userq_vm_validate_and_restore_queue() reads each queue's wptr BO GPU offset via amdgpu_bo_gpu_offset() after only calling amdgpu_ttm_alloc_gart() on it. But the wptr BOs are not part of this VM (their reservation object is their own, not vm->root), so neither amdgpu_vm_validate() nor amdgpu_userq_bo_validate() (which only handles the VM's evicted list) covers them. As a result a wptr BO can still be in TTM_PL_SYSTEM and unreserved when its offset is read, tripping the amdgpu_bo_gpu_offset() sanity checks from the restore worker:

  ------------[ cut here ]------------
  WARNING: amdgpu_object.c:1486 at amdgpu_bo_gpu_offset+0x75/0xa0 [amdgpu], CPU#3: kworker/3:1/116
  Workqueue: events amdgpu_userq_restore_worker [amdgpu]
  RIP: 0010:amdgpu_bo_gpu_offset+0x75/0xa0 [amdgpu]
  Call Trace:
   <TASK>
   amdgpu_userq_vm_validate_and_restore_queue+0x629/0x960 [amdgpu]
   amdgpu_userq_restore_worker+0xa6/0x180 [amdgpu]
   process_scheduled_works+0xa6/0x460
   worker_thread+0x13c/0x290
   kthread+0xfb/0x140
   ret_from_fork+0x1b6/0x2b0
   ret_from_fork_asm+0x1a/0x30
   </TASK>
  ---[ end trace 0000000000000000 ]---
  ------------[ cut here ]------------
  WARNING: amdgpu_object.c:1485 at amdgpu_bo_gpu_offset+0x9a/0xa0 [amdgpu], CPU#2: kworker/2:1/127
  Workqueue: events amdgpu_userq_restore_worker [amdgpu]
  RIP: 0010:amdgpu_bo_gpu_offset+0x9a/0xa0 [amdgpu]

amdgpu_ttm_alloc_gart() only creates the GART mapping; it does not migrate the BO out of system memory, so the offset read is bogus (the queue would resume with a wrong wptr address).

Lock each wptr BO into the drm_exec context and validate it into GTT inside the drm_exec_until_all_locked() block, mirroring what the create path (mes_userq_create_wptr_mapping) already does, so that the offset read later is safe and correct.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 26 +++++++++++++++++++++++
 1 file changed, 26 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +0 -0
index 17cc48d87c4d..59aa4802c111 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c amdgpu_userq_vm_validate_and_restore_queue(struct amdgpu_userq_mgr *uq_mgr) · line 1061

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +5 -0
@@ -1054,6 +1054,32 @@ amdgpu_userq_vm_validate_and_restore_queue(struct amdgpu_userq_mgr *uq_mgr)
 		drm_exec_retry_on_contention(&exec);
 		if (unlikely(ret))
 			goto unlock_all;
+
+		/*
+		 * The per-queue wptr BOs are not part of this VM (their resv is
+		 * their own, not vm->root), so the validation above does not
+		 * cover them.

That's not correct.

The WPTR BOs absolutely must be part of the VM or otherwise the MES/CP/SDMA FW wouldn't be able to read it.

Quoted 12 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +12 -0
+		Lock and validate each into GTT here so that
+		 * reading its GPU offset below is safe - matching what the
+		 * create path (mes_userq_create_wptr_mapping) does.
+		 */
+		xa_for_each(&uq_mgr->userq_xa, tmp_key, queue) {
+			struct ttm_operation_ctx wptr_ctx = { false, false };
+
+			bo = queue->wptr_obj.obj;
+			if (!bo)
+				continue;
+
+			ret = drm_exec_prepare_obj(&exec, &bo->tbo.base,

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c amdgpu_userq_vm_validate_and_restore_queue(struct amdgpu_userq_mgr *uq_mgr) · line 1082

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +10 -0
+						   TTM_NUM_MOVE_FENCES + 1);
+			drm_exec_retry_on_contention(&exec);
+			if (unlikely(ret))
+				goto unlock_all;
+
+			amdgpu_bo_placement_from_domain(bo, bo->allowed_domains);
+			ret = ttm_bo_validate(&bo->tbo, &bo->placement, &wptr_ctx);
+			if (unlikely(ret))
+				goto unlock_all;
+		}

Clear NAK to that, this is clearly not correct.

Regards,
Christian.

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +0 -0
 	}
 
 	if (invalidated) {
#2 lore raw

AMD General

Quoted 76 lines

-----Original Message-----
From: Koenig, Christian <[email protected]>
Sent: Friday, August 14, 2026 7:30 PM
To: Zhang, Jesse(Jie) <[email protected]>; [email protected]
Cc: Deucher, Alexander <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: lock and validate wptr BOs before reading their GPU offset on restore

Quoted 10 lines · Jesse Zhang

amdgpu_userq_vm_validate_and_restore_queue() reads each queue's wptr BO GPU offset via amdgpu_bo_gpu_offset() after only calling amdgpu_ttm_alloc_gart() on it. But the wptr BOs are not part of this VM (their reservation object is their own, not vm->root), so neither amdgpu_vm_validate() nor amdgpu_userq_bo_validate() (which only handles the VM's evicted list) covers them. As a result a wptr BO can still be in TTM_PL_SYSTEM and unreserved when its offset is read, tripping the
amdgpu_bo_gpu_offset() sanity checks from the restore worker:

Quoted 2 lines
  ------------[ cut here ]------------
  WARNING: amdgpu_object.c:1486 at amdgpu_bo_gpu_offset+0x75/0xa0

[amdgpu], CPU#3: kworker/3:1/116

Quoted 7 lines
  Workqueue: events amdgpu_userq_restore_worker [amdgpu]
  RIP: 0010:amdgpu_bo_gpu_offset+0x75/0xa0 [amdgpu]
  Call Trace:
   <TASK>
   amdgpu_userq_vm_validate_and_restore_queue+0x629/0x960 [amdgpu]
   amdgpu_userq_restore_worker+0xa6/0x180 [amdgpu]
   process_scheduled_works+0xa6/0x460
Quoted 8 lines
   worker_thread+0x13c/0x290
   kthread+0xfb/0x140
   ret_from_fork+0x1b6/0x2b0
   ret_from_fork_asm+0x1a/0x30
   </TASK>
  ---[ end trace 0000000000000000 ]---
  ------------[ cut here ]------------
  WARNING: amdgpu_object.c:1485 at amdgpu_bo_gpu_offset+0x9a/0xa0

[amdgpu], CPU#2: kworker/2:1/127

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 24 lines
  Workqueue: events amdgpu_userq_restore_worker [amdgpu]
  RIP: 0010:amdgpu_bo_gpu_offset+0x9a/0xa0 [amdgpu]

amdgpu_ttm_alloc_gart() only creates the GART mapping; it does not migrate the BO out of system memory, so the offset read is bogus (the queue would resume with a wrong wptr address).

Lock each wptr BO into the drm_exec context and validate it into GTT inside the drm_exec_until_all_locked() block, mirroring what the create path (mes_userq_create_wptr_mapping) already does, so that the offset read later is safe and correct.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 26
+++++++++++++++++++++++
 1 file changed, 26 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +0 -0
index 17cc48d87c4d..59aa4802c111 100644
@@ -1054,6 +1054,32 @@

amdgpu_userq_vm_validate_and_restore_queue(struct amdgpu_userq_mgr *uq_mgr)

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c amdgpu_userq_vm_validate_and_restore_queue(struct amdgpu_userq_mgr *uq_mgr)

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +5 -0
            drm_exec_retry_on_contention(&exec);
            if (unlikely(ret))
                    goto unlock_all;
+
+           /*
+            * The per-queue wptr BOs are not part of this VM (their resv is
+            * their own, not vm->root), so the validation above does not
+            * cover them.
Quoted 4 lines

That's not correct.

The WPTR BOs absolutely must be part of the VM or otherwise the MES/CP/SDMA FW wouldn't be able to read it.

Thanks Christian, you're right that “not part of this VM” is incorrect wording — the WPTR BO is VM-mapped.
The issue I’m hitting is different: in restore, amdgpu_bo_gpu_offset() can be called without the WPTR BO reservation lock held, which triggers WARN_ON and races with BO state changes. I’ll send v2 with corrected description and a narrower fix that only adds each WPTR BO to the drm_exec lock set before reading/updating its GPU offset.

Thanks
Jesse

Quoted 33 lines
Quoted 14 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +13 -0
+           Lock and validate each into GTT here so that
+            * reading its GPU offset below is safe - matching what the
+            * create path (mes_userq_create_wptr_mapping) does.
+            */
+           xa_for_each(&uq_mgr->userq_xa, tmp_key, queue) {
+                   struct ttm_operation_ctx wptr_ctx = { false, false };
+
+                   bo = queue->wptr_obj.obj;
+                   if (!bo)
+                           continue;
+
+                   ret = drm_exec_prepare_obj(&exec, &bo->tbo.base,
+                                              TTM_NUM_MOVE_FENCES + 1);

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c amdgpu_userq_vm_validate_and_restore_queue(struct amdgpu_userq_mgr *uq_mgr) · line 1082

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +9 -0
+                   drm_exec_retry_on_contention(&exec);
+                   if (unlikely(ret))
+                           goto unlock_all;
+
+                   amdgpu_bo_placement_from_domain(bo, bo-
 allowed_domains);
+                   ret = ttm_bo_validate(&bo->tbo, &bo->placement, &wptr_ctx);
+                   if (unlikely(ret))
+                           goto unlock_all;
+           }

Clear NAK to that, this is clearly not correct.

Regards,
Christian.

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +0 -0
    }
 
    if (invalidated) {