[PATCH 1/4] sync_file: add a mutex to protect fence and callback members. (v3)

Chris Wilson chris at chris-wilson.co.uk
Mon Mar 20 08:21:15 UTC 2017

On Mon, Mar 20, 2017 at 05:03:04PM +1000, Dave Airlie wrote:
> From: Dave Airlie <airlied at redhat.com>
> This patch allows the underlying fence in a sync_file to be changed
> or set to NULL. This isn't currently required but for Vulkan
> semaphores we need to be able to swap and reset the fence.
> In order to faciliate this, it uses rcu to protect the fence,
> along with a new mutex. The mutex also protects the callback.
> It also checks for NULL when retrieving the rcu protected
> fence in case it has been reset.
> v1.1: fix the locking (Julia Lawall).
> v2: use rcu try one
> v3: fix poll to use proper rcu, fixup merge/fill ioctls
> to not crash on NULL fence cases.
> Signed-off-by: Dave Airlie <airlied at redhat.com>
> ---
> @@ -124,13 +133,26 @@ struct dma_fence *sync_file_get_fence(int fd)
>  	if (!sync_file)
>  		return NULL;
> -	fence = dma_fence_get(sync_file->fence);
> +	if (!rcu_access_pointer(sync_file->fence))
> +		return NULL;
> +
> +	rcu_read_lock();
> +	fence = dma_fence_get_rcu_safe(&sync_file->fence);
> +	rcu_read_unlock();
> +
>  	fput(sync_file->file);

So poll will wait until the fence is set before the sync_file is
signaled, but here we return NULL. At the moment this is interpretted by
the callers as an error (since we can't distinguish between the lookup
error and the empty sync_file). However, if it is empty we also want to
delay the dependent execution until the fence is set to match the poll

Chris Wilson, Intel Open Source Technology Centre

More information about the amd-gfx mailing list