Re: [PATCH v2] drm/radeon: restore hardware polling in fence_is_signaled to fix performance regression

[Christian König]

On 7/30/26 03:56, 2564278112@xxxxxx wrote:
> From: Wang Jiang <jiangwang@xxxxxxxxxx>
> 
> Commit 9eb00b5f5697b ("drm/radeon: delete radeon_fence_process in
> is_signaled, no deadlock") removed the hardware status check from
> radeon_fence_is_signaled() to fix a self-deadlock caused by
> wake_up_all(&rdev->fence_queue) being called with the fence queue
> lock held.
> 
> However, removing the hardware poll entirely causes significant
> performance regression (e.g. glxgears FPS drop) because the signaled
> check becomes purely passive — it only reads the cached last_seq
> without probing the GPU, so completed GPU work is not detected in
> time, causing unnecessary CPU stalls in sync-heavy workloads.
> 
> Fix this by calling radeon_fence_read() directly in
> radeon_fence_is_signaled() to poll the hardware fence counter
> without updating last_seq or calling wake_up_all(). This restores
> timely fence detection while avoiding the original deadlock, since
> last_seq update and waiter wakeup remain solely in the interrupt
> handler's responsibility.
> 
> Fixes: 9eb00b5f5697b ("drm/radeon: delete radeon_fence_process in is_signaled, no deadlock")
> Signed-off-by: Wang Jiang <jiangwang@xxxxxxxxxx>
> ---
>  drivers/gpu/drm/radeon/radeon_fence.c | 12 ++++++++++++
>  1 file changed, 12 insertions(+)
> 
> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
> index 02a40e4750c7..047ec5a1d194 100644
> --- a/drivers/gpu/drm/radeon/radeon_fence.c
> +++ b/drivers/gpu/drm/radeon/radeon_fence.c
> @@ -360,6 +360,18 @@ static bool radeon_fence_is_signaled(struct dma_fence *f)
>  	if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq)
>  		return true;
>  
> +	/*
> +	* Poll hardware directly without updating last_seq.
> +	* This allows immediate detection of signaled fences (performance)
> +	* while leaving last_seq update and wake_up_all() to the interrupt
> +	* handler, avoiding:
> +	* 1. Deadlock from calling wake_up_all() with fence lock held
> +	* 2. Wake event stealing by advancing last_seq without waking waiters
> +	* 3. rw_semaphore usage in atomic/irq context (unsafe on PREEMPT_RT)
> +	*/
> +	if (radeon_fence_read(rdev, ring) >= (u32)seq)
> +			return true;
> +

That check is not correct, you need to take wrap around into account.

Try using __dma_fence_is_later() instead.

Regards,
Christian.

>  	return false;
>  }
>