[PATCH v5 7/9] drm/amdgpu: add userq va unmap validated helper

[Christian König]

On 04.07.25 12:33, Prike Liang wrote:
> This helper can validate the userq whether can be
> unmapped prior to the userq VA GEM unmap.
> 
> Signed-off-by: Prike Liang <Prike.Liang at amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 78 +++++++++++++++++++++++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h |  3 +
>  2 files changed, 81 insertions(+)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> index 25a35ab7395b..30838e5279bd 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
> @@ -1180,3 +1180,81 @@ int amdgpu_userq_start_sched_for_enforce_isolation(struct amdgpu_device *adev,
>  	mutex_unlock(&adev->userq_mutex);
>  	return ret;
>  }
> +
> +static bool amdgpu_userq_gem_va_unmap_queue_retrieve(struct amdgpu_usermode_queue *queue,
> +							uint64_t va)
> +{
> +	va = va << AMDGPU_GPU_PAGE_SHIFT | AMDGPU_GMC_HOLE_END;

Still clear NAK to doing that here.

> +
> +	switch (queue->queue_type) {
> +	case AMDGPU_HW_IP_GFX:
> +		if (queue->queue_va == va ||
> +		    queue->wptr_va  == va ||
> +		    queue->rptr_va  == va ||
> +		    queue->shadow_va == va ||
> +		    queue->csa_va  == va)
> +			return true;
> +		break;
> +	case AMDGPU_HW_IP_COMPUTE:
> +		  if (queue->queue_va == va ||
> +		      queue->wptr_va == va ||
> +		      queue->rptr_va  == va ||
> +		      queue->eop_va  == va)
> +			return true;
> +		break;
> +	case AMDGPU_HW_IP_DMA:
> +		if (queue->queue_va == va ||
> +		    queue->wptr_va == va ||
> +		    queue->rptr_va == va ||
> +		    queue->csa_va == va)
> +			return true;
> +		break;
> +	default:
> +		break;
> +	}
> +
> +	return false;
> +}
> +
> +int amdgpu_userq_gem_va_unmap_validate(struct amdgpu_vm *vm,
> +				uint64_t va)

I don't see an user of this function?

> +{
> +	struct amdgpu_fpriv *fpriv = vm_to_fpriv(vm);
> +	struct amdgpu_userq_mgr *uq_mgr = &fpriv->userq_mgr;
> +
> +	if (&fpriv->vm == vm) {

This check is complete nosense.

> +		struct amdgpu_usermode_queue *queue;
> +		int queue_id, r = 0;
> +
> +		if (mutex_trylock(&uq_mgr->userq_mutex)) {
> +			/* If here the userq bo is busy and needs to deactivate and prevent reusing it.*/
> +			idr_for_each_entry(&uq_mgr->userq_idr, queue, queue_id) {
> +				struct dma_fence *f = queue->last_fence;
> +
> +
> +				if (!amdgpu_userq_gem_va_unmap_queue_retrieve(queue, va)) {
> +					dev_dbg(uq_mgr->adev->dev, "queue(id:%d) not belond to vm:%p\n",
> +						queue_id,vm);
> +					continue;
> +				}
> +
> +				if (f && !dma_fence_is_signaled(f)) {
> +
> +					dev_warn(uq_mgr->adev->dev, "try to unmap the busy queue(id:%d):%p under vm:%p\n",
> +						queue_id, queue, vm);
> +					/* Need to set a resonable state for avoiding reusing this queue*/
> +					queue->state = AMDGPU_USERQ_STATE_HUNG;
> +					r++;
> +				}
> +			}
> +			mutex_unlock(&uq_mgr->userq_mutex);
> +			return r;
> +		} else {
> +			/* do we need a try lock again before return*/
> +			return -EBUSY;
> +		}
> +
> +	}
> +
> +	return 0;
> +}
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h
> index 194ec7a6b3b2..08c49d738ec1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h
> @@ -31,6 +31,7 @@
>  #define to_ev_fence(f) container_of(f, struct amdgpu_eviction_fence, base)
>  #define uq_mgr_to_fpriv(u) container_of(u, struct amdgpu_fpriv, userq_mgr)
>  #define work_to_uq_mgr(w, name) container_of(w, struct amdgpu_userq_mgr, name)
> +#define vm_to_fpriv(v)  container_of(v, struct amdgpu_fpriv, vm)

NAK to that, the VM should not be casted to fpriv.

>  
>  enum amdgpu_userq_state {
>  	AMDGPU_USERQ_STATE_UNMAPPED = 0,
> @@ -148,4 +149,6 @@ bool amdgpu_userq_buffer_vas_mapped(struct amdgpu_vm *vm,
>  int amdgpu_userq_buffer_va_put(struct amdgpu_vm *vm, u64 addr);
>  int amdgpu_userq_buffer_vas_put(struct amdgpu_vm *vm,
>  			struct amdgpu_usermode_queue *queue);
> +int amdgpu_userq_gem_va_unmap_validate(struct amdgpu_vm *vm,
> +				uint64_t va);
>  #endif