On Mon, Aug 31, 2026 at 1:50 AM Jesse Zhang <[email protected]> wrote:
>
> Set prop->modify and prop->user_wptr (read from the wptr poll BO) and
> call init_mqd, so a re-enabled compute queue keeps its context-saved
> rptr and resumes at the first un-consumed packet.
>
> Also set modify=false on the create and kernel-ring
> (amdgpu_ring_to_mqd_prop) paths, so every init_mqd caller states its own
> value rather than relying on the zeroed prop.
>
> v3: set the mqd_prop modify flag and call init_mqd instead of a separate
> update_mqd callback, per review.
>
> Signed-off-by: Jesse Zhang <[email protected]>
> Suggested-by: Alexander Deucher <[email protected]>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c | 2 ++
> drivers/gpu/drm/amd/amdgpu/mes_userqueue.c | 20 +++++++++++++++++++-
> 2 files changed, 21 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> index 686c92e96025..9fd0ab0be40e 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
> @@ -840,6 +840,8 @@ static void amdgpu_ring_to_mqd_prop(struct amdgpu_ring
> *ring,
> prop->use_doorbell = ring->use_doorbell;
> prop->doorbell_index = ring->doorbell_index;
> prop->kernel_queue = true;
> + /* kernel rings always build the MQD from scratch, never a MODIFY */
> + prop->modify = false;
>
> /* map_queues packet doesn't need activate the queue,
> * so only kiq need set this field.
> diff --git a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
> b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
> index 82bb4369451e..c6ac3194e4da 100644
> --- a/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
> +++ b/drivers/gpu/drm/amd/amdgpu/mes_userqueue.c
> @@ -479,6 +479,7 @@ static int mes_userq_mqd_create(struct
> amdgpu_usermode_queue *queue,
> userq_props->use_doorbell = true;
> userq_props->doorbell_index = queue->doorbell_index;
> userq_props->fence_address = queue->fence_drv->gpu_addr;
> + userq_props->modify = false;
>
> if (queue->queue_type == AMDGPU_HW_IP_COMPUTE) {
> struct drm_amdgpu_userq_mqd_compute_gfx11 *compute_mqd;
> @@ -681,6 +682,20 @@ static int mes_userq_mqd_create(struct
> amdgpu_usermode_queue *queue,
> return r;
> }
>
> +/* Read the queue's live user wptr from its poll BO; 0 if not mapped. */
> +static u64 mes_userq_read_user_wptr(struct amdgpu_usermode_queue *queue)
> +{
> + u64 user_wptr = 0;
> + void *wc = NULL;
> +
> + if (queue->wptr_obj.obj &&
> + !amdgpu_bo_kmap(queue->wptr_obj.obj, &wc) && wc) {
> + user_wptr = *(volatile u64 *)wc;
> + amdgpu_bo_kunmap(queue->wptr_obj.obj);
> + }
> + return user_wptr;
> +}
Can't we just skip this? The firmware should update the MQD and store
the current rptr/wptr when we unmap it. So when we modify the queue,
all we need to do is skip any changes to the rptr/wrptr when we modify
the MQD. No need to save the wptr. Other than that, the patches look
good to me.
Alex
> +
> static int mes_userq_compute_mqd_update(struct amdgpu_usermode_queue *queue,
> struct drm_amdgpu_userq_in *args_in)
> {
> @@ -707,8 +722,11 @@ static int mes_userq_compute_mqd_update(struct
> amdgpu_usermode_queue *queue,
>
> userq_props->queue_size = args_in->queue_size;
> userq_props->hqd_base_gpu_addr = args_in->queue_va;
> + userq_props->modify = true;
> + userq_props->user_wptr = mes_userq_read_user_wptr(queue);
>
> - retval = mqd_hw_default->init_mqd(adev, (void *)queue->mqd.cpu_ptr,
> userq_props);
> + retval = mqd_hw_default->init_mqd(adev, (void *)queue->mqd.cpu_ptr,
> + userq_props);
>
> free:
> kfree(compute_mqd_v11);
> --
> 2.49.0
>
