[PATCH 1/3] drm/amdgpu: add user queue vm identifier

[PATCH 1/3] drm/amdgpu: add user queue vm identifier

[Prike Liang]

Add a user queue vm identifier for each userqueue
kms opt accessing.

Signed-off-by: Prike Liang <Prike.Liang@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 5 +++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c    | 2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h    | 2 +-
 3 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index 8888be4e758c..a8bb3f1b8649 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
@@ -1249,10 +1249,13 @@ amdgpu_userq_evict(struct amdgpu_userq_mgr *uq_mgr,
 int amdgpu_userq_mgr_init(struct amdgpu_userq_mgr *userq_mgr, struct drm_file *file_priv,
 			  struct amdgpu_device *adev)
 {
+	struct amdgpu_fpriv *fpriv = uq_mgr_to_fpriv(userq_mgr);
+
 	mutex_init(&userq_mgr->userq_mutex);
 	xa_init_flags(&userq_mgr->userq_mgr_xa, XA_FLAGS_ALLOC);
 	userq_mgr->adev = adev;
 	userq_mgr->file = file_priv;
+	fpriv->vm.is_userq_context = true;
 
 	INIT_DELAYED_WORK(&userq_mgr->resume_work, amdgpu_userq_restore_worker);
 	return 0;
@@ -1262,6 +1265,7 @@ void amdgpu_userq_mgr_fini(struct amdgpu_userq_mgr *userq_mgr)
 {
 	struct amdgpu_usermode_queue *queue;
 	unsigned long queue_id;
+	struct amdgpu_fpriv *fpriv = uq_mgr_to_fpriv(userq_mgr);
 
 	cancel_delayed_work_sync(&userq_mgr->resume_work);
 
@@ -1276,6 +1280,7 @@ void amdgpu_userq_mgr_fini(struct amdgpu_userq_mgr *userq_mgr)
 	xa_destroy(&userq_mgr->userq_mgr_xa);
 	mutex_unlock(&userq_mgr->userq_mutex);
 	mutex_destroy(&userq_mgr->userq_mutex);
+	fpriv->vm.is_userq_context = false;
 }
 
 int amdgpu_userq_suspend(struct amdgpu_device *adev)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index db66b4232de0..48a64d828eb8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -2591,7 +2591,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 	ttm_lru_bulk_move_init(&vm->lru_bulk_move);
 
 	vm->is_compute_context = false;
-
+	vm->is_userq_context = false;
 	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
 				    AMDGPU_VM_USE_CPU_FOR_GFX);
 
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
index 77207f4e448e..cbe7dfa4ffcb 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
@@ -439,7 +439,7 @@ struct amdgpu_vm {
 	struct ttm_lru_bulk_move lru_bulk_move;
 	/* Flag to indicate if VM is used for compute */
 	bool			is_compute_context;
-
+	bool			is_userq_context;
 	/* Memory partition number, -1 means any partition */
 	int8_t			mem_id;
 
-- 
2.34.1

[PATCH 2/3] drm/amdgpu: rework the userq tlb flush

[Prike Liang]

The userq TLB flush should be performed after
the vm update complete and input fence signal.

Signed-off-by: Prike Liang <Prike.Liang@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c    | 5 +++--
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index a8bb3f1b8649..01707d72d982 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
@@ -384,6 +384,7 @@ amdgpu_userq_wait_for_last_fence(struct amdgpu_userq_mgr *uq_mgr,
 			return -ETIME;
 		}
 	}
+	amdgpu_gmc_flush_gpu_tlb_pasid(uq_mgr->adev, queue->vm->pasid, 2, true, 0);
 
 	return ret;
 }
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 48a64d828eb8..1537f0b6b712 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -1062,7 +1062,8 @@ amdgpu_vm_tlb_flush(struct amdgpu_vm_update_params *params,
 	}
 
 	/* Prepare a TLB flush fence to be attached to PTs */
-	if (!params->unlocked && vm->is_compute_context) {
+	if (!params->unlocked && (vm->is_compute_context ||
+	     vm->is_userq_context)) {
 		amdgpu_vm_tlb_fence_create(params->adev, vm, fence);
 
 		/* Makes sure no PD/PT is freed before the flush */
@@ -2792,7 +2793,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	amdgpu_bo_unreserve(root);
 	amdgpu_bo_unref(&root);
 	WARN_ON(vm->root.bo);
-
+	amdgpu_gmc_flush_gpu_tlb_pasid(adev, vm->pasid, 2, true, 0);
 	amdgpu_vm_fini_entities(vm);
 
 	if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {
-- 
2.34.1

Re: [PATCH 2/3] drm/amdgpu: rework the userq tlb flush

[Christian König]

On 11/5/25 07:16, Prike Liang wrote:
> The userq TLB flush should be performed after
> the vm update complete and input fence signal.
> 
> Signed-off-by: Prike Liang <Prike.Liang@amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 1 +
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c    | 5 +++--
>  2 files changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> index a8bb3f1b8649..01707d72d982 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> @@ -384,6 +384,7 @@ amdgpu_userq_wait_for_last_fence(struct amdgpu_userq_mgr *uq_mgr,
>  			return -ETIME;
>  		}
>  	}
> +	amdgpu_gmc_flush_gpu_tlb_pasid(uq_mgr->adev, queue->vm->pasid, 2, true, 0);

Please drop that one, it doesn't seem to make any sense at all.

>  
>  	return ret;
>  }
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index 48a64d828eb8..1537f0b6b712 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -1062,7 +1062,8 @@ amdgpu_vm_tlb_flush(struct amdgpu_vm_update_params *params,
>  	}
>  
>  	/* Prepare a TLB flush fence to be attached to PTs */
> -	if (!params->unlocked && vm->is_compute_context) {
> +	if (!params->unlocked && (vm->is_compute_context ||
> +	     vm->is_userq_context)) {

Just drop the is_compute_context check here, don't add any new checks.

>  		amdgpu_vm_tlb_fence_create(params->adev, vm, fence);
>  
>  		/* Makes sure no PD/PT is freed before the flush */
> @@ -2792,7 +2793,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
>  	amdgpu_bo_unreserve(root);
>  	amdgpu_bo_unref(&root);
>  	WARN_ON(vm->root.bo);
> -
> +	amdgpu_gmc_flush_gpu_tlb_pasid(adev, vm->pasid, 2, true, 0);

Please drop that one as well, it doesn't seem to make sense.

Regards,
Christian.

>  	amdgpu_vm_fini_entities(vm);
>  
>  	if (!RB_EMPTY_ROOT(&vm->va.rb_root)) {

[PATCH 3/3] drm/amdgpu: fix the vm update tlb sequence

[Prike Liang]

The VM TLB sequence is incremented by the VM update
commit invoker, so it should not be increased again
in the commit callback.

Signed-off-by: Prike Liang <Prike.Liang@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c  | 2 --
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c   | 2 ++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c | 3 ---
 3 files changed, 2 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
index 0c1ef5850a5e..f7a5e71346df 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
@@ -106,8 +106,6 @@ static int amdgpu_vm_cpu_update(struct amdgpu_vm_update_params *p,
 static int amdgpu_vm_cpu_commit(struct amdgpu_vm_update_params *p,
 				struct dma_fence **fence)
 {
-	if (p->needs_flush)
-		atomic64_inc(&p->vm->tlb_seq);
 
 	mb();
 	amdgpu_device_flush_hdp(p->adev, NULL);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
index 30022123b0bf..869f9cb28d94 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
@@ -420,6 +420,8 @@ int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		goto exit;
 
 	r = vm->update_funcs->commit(&params, NULL);
+	if (params.needs_flush)
+		atomic64_inc(&vm->tlb_seq);
 exit:
 	drm_dev_exit(idx);
 	return r;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
index 46d9fb433ab2..0ceb2474a721 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
@@ -115,9 +115,6 @@ static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
 	WARN_ON(ib->length_dw == 0);
 	amdgpu_ring_pad_ib(ring, ib);
 
-	if (p->needs_flush)
-		atomic64_inc(&p->vm->tlb_seq);
-
 	WARN_ON(ib->length_dw > p->num_dw_left);
 	f = amdgpu_job_submit(p->job);
 
-- 
2.34.1

Re: [PATCH 3/3] drm/amdgpu: fix the vm update tlb sequence

[Christian König]

On 11/5/25 07:16, Prike Liang wrote:
> The VM TLB sequence is incremented by the VM update
> commit invoker, so it should not be increased again
> in the commit callback.

Clear NAK to that. It must be incremented by *both*!

Regards,
Christian.

> 
> Signed-off-by: Prike Liang <Prike.Liang@amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c  | 2 --
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c   | 2 ++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c | 3 ---
>  3 files changed, 2 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
> index 0c1ef5850a5e..f7a5e71346df 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
> @@ -106,8 +106,6 @@ static int amdgpu_vm_cpu_update(struct amdgpu_vm_update_params *p,
>  static int amdgpu_vm_cpu_commit(struct amdgpu_vm_update_params *p,
>  				struct dma_fence **fence)
>  {
> -	if (p->needs_flush)
> -		atomic64_inc(&p->vm->tlb_seq);
>  
>  	mb();
>  	amdgpu_device_flush_hdp(p->adev, NULL);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
> index 30022123b0bf..869f9cb28d94 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c
> @@ -420,6 +420,8 @@ int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>  		goto exit;
>  
>  	r = vm->update_funcs->commit(&params, NULL);
> +	if (params.needs_flush)
> +		atomic64_inc(&vm->tlb_seq);
>  exit:
>  	drm_dev_exit(idx);
>  	return r;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
> index 46d9fb433ab2..0ceb2474a721 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c
> @@ -115,9 +115,6 @@ static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
>  	WARN_ON(ib->length_dw == 0);
>  	amdgpu_ring_pad_ib(ring, ib);
>  
> -	if (p->needs_flush)
> -		atomic64_inc(&p->vm->tlb_seq);
> -
>  	WARN_ON(ib->length_dw > p->num_dw_left);
>  	f = amdgpu_job_submit(p->job);
>  

Re: [PATCH 1/3] drm/amdgpu: add user queue vm identifier

[Christian König]

On 11/5/25 07:16, Prike Liang wrote:
> Add a user queue vm identifier for each userqueue
> kms opt accessing.

Clear NAK to that approach, we don't want any specialized handling in the VM if userqueues are enabled.

Christian.

> 
> Signed-off-by: Prike Liang <Prike.Liang@amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 5 +++++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c    | 2 +-
>  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h    | 2 +-
>  3 files changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> index 8888be4e758c..a8bb3f1b8649 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> @@ -1249,10 +1249,13 @@ amdgpu_userq_evict(struct amdgpu_userq_mgr *uq_mgr,
>  int amdgpu_userq_mgr_init(struct amdgpu_userq_mgr *userq_mgr, struct drm_file *file_priv,
>  			  struct amdgpu_device *adev)
>  {
> +	struct amdgpu_fpriv *fpriv = uq_mgr_to_fpriv(userq_mgr);
> +
>  	mutex_init(&userq_mgr->userq_mutex);
>  	xa_init_flags(&userq_mgr->userq_mgr_xa, XA_FLAGS_ALLOC);
>  	userq_mgr->adev = adev;
>  	userq_mgr->file = file_priv;
> +	fpriv->vm.is_userq_context = true;
>  
>  	INIT_DELAYED_WORK(&userq_mgr->resume_work, amdgpu_userq_restore_worker);
>  	return 0;
> @@ -1262,6 +1265,7 @@ void amdgpu_userq_mgr_fini(struct amdgpu_userq_mgr *userq_mgr)
>  {
>  	struct amdgpu_usermode_queue *queue;
>  	unsigned long queue_id;
> +	struct amdgpu_fpriv *fpriv = uq_mgr_to_fpriv(userq_mgr);
>  
>  	cancel_delayed_work_sync(&userq_mgr->resume_work);
>  
> @@ -1276,6 +1280,7 @@ void amdgpu_userq_mgr_fini(struct amdgpu_userq_mgr *userq_mgr)
>  	xa_destroy(&userq_mgr->userq_mgr_xa);
>  	mutex_unlock(&userq_mgr->userq_mutex);
>  	mutex_destroy(&userq_mgr->userq_mutex);
> +	fpriv->vm.is_userq_context = false;
>  }
>  
>  int amdgpu_userq_suspend(struct amdgpu_device *adev)
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index db66b4232de0..48a64d828eb8 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -2591,7 +2591,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
>  	ttm_lru_bulk_move_init(&vm->lru_bulk_move);
>  
>  	vm->is_compute_context = false;
> -
> +	vm->is_userq_context = false;
>  	vm->use_cpu_for_update = !!(adev->vm_manager.vm_update_mode &
>  				    AMDGPU_VM_USE_CPU_FOR_GFX);
>  
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> index 77207f4e448e..cbe7dfa4ffcb 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
> @@ -439,7 +439,7 @@ struct amdgpu_vm {
>  	struct ttm_lru_bulk_move lru_bulk_move;
>  	/* Flag to indicate if VM is used for compute */
>  	bool			is_compute_context;
> -
> +	bool			is_userq_context;
>  	/* Memory partition number, -1 means any partition */
>  	int8_t			mem_id;
>