[PATCH 5/8] sync_file: add support for a semaphore object

[Chris Wilson]

On Tue, Apr 04, 2017 at 02:27:30PM +1000, Dave Airlie wrote:
> From: Dave Airlie <airlied at redhat.com>
> 
> This object can be used to implement the Vulkan semaphores.
> 
> The object behaviour differs from fence, in that you can
> replace the underlying fence, and you cannot merge semaphores.
> 
> Signed-off-by: Dave Airlie <airlied at redhat.com>
> ---
>  drivers/dma-buf/sync_file.c    | 36 +++++++++++++++++++++++++++++++++++-
>  include/linux/sync_file.h      |  2 ++
>  include/uapi/linux/sync_file.h | 14 ++++++++++++++
>  3 files changed, 51 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/dma-buf/sync_file.c b/drivers/dma-buf/sync_file.c
> index 6376f6f..a82f6d8 100644
> --- a/drivers/dma-buf/sync_file.c
> +++ b/drivers/dma-buf/sync_file.c
> @@ -44,7 +44,7 @@ int sync_file_validate_type_flags(uint32_t type, uint32_t flags)
>  {
>  	if (flags)
>  		return -EINVAL;
> -	if (type != SYNC_FILE_TYPE_FENCE)
> +	if (type != SYNC_FILE_TYPE_FENCE && type != SYNC_FILE_TYPE_SEMAPHORE)
>  		return -EINVAL;
>  	return 0;
>  }
> @@ -200,6 +200,38 @@ sync_file_get_fence_locked(struct sync_file *sync_file)
>  					 sync_file_held(sync_file));
>  }
>  
> +/**
> + * sync_file_replace_fence - replace the fence related to the sync_file
> + * @sync_file:	 sync file to replace fence in
> + * @fence: fence to replace with (or NULL for no fence).
> + * Returns previous fence.
> + */
> +struct dma_fence *sync_file_replace_fence(struct sync_file *sync_file,
> +					  struct dma_fence *fence)
> +{
> +	struct dma_fence *ret_fence = NULL;
> +
> +	if (sync_file->type != SYNC_FILE_TYPE_SEMAPHORE)
> +		return NULL;
> +
> +	if (fence)
> +		dma_fence_get(fence);
> +
> +	mutex_lock(&sync_file->lock);
> +
> +	ret_fence = sync_file_get_fence_locked(sync_file);
> +	if (ret_fence) {
> +		if (test_bit(POLL_ENABLED, &ret_fence->flags))
> +			dma_fence_remove_callback(ret_fence, &sync_file->cb);

This is racy with sync_file_poll. And sync_file_poll now needs rcu
protection (as does all access to sync_file->fence), I need to check
whether the previous patches are complete. Also needs to handle, or
forbid, the caller passing in the same fence.

> +	}
> +
> +	RCU_INIT_POINTER(sync_file->fence, fence);
> +
> +	mutex_unlock(&sync_file->lock);
> +	return ret_fence;
> +}
> +EXPORT_SYMBOL(sync_file_replace_fence);
> +
>  static int sync_file_set_fence(struct sync_file *sync_file,
>  			       struct dma_fence **fences, int num_fences)
>  {
> @@ -278,6 +310,8 @@ static struct sync_file *sync_file_merge(const char *name, struct sync_file *a,
>  
>  	if (a->type != b->type)
>  		return NULL;
> +	if (a->type != SYNC_FILE_TYPE_FENCE)
> +		return NULL;

No rules for containerisation? Being able to pass in an array of current
semaphore in-fences through a single fd? It will have a bigger window
for change than doing everything inside the kernel, but since the kernel
cannot take a simultaneous snaphot of all in-semaphores either (could
use ww_mutex?) it seems harsh to exclude current ABI interop.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre