Re: [PATCH 2/2] dma-buf/sync_file: rework fence storage in struct file

[Chris Wilson]

On Thu, Jun 30, 2016 at 11:09:48AM -0300, Gustavo Padovan wrote:
> From: Gustavo Padovan <gustavo.padovan@xxxxxxxxxxxxxxx>
> 
> Create sync_file->fence to abstract the type of fence we are using for
> each sync_file. If only one fence is present we use a normal struct fence
> but if there is more fences to be added to the sync_file a fence_array
> is created.
> 
> This change cleans up sync_file a bit. We don't need to have sync_file_cb
> array anymore. Instead, as we always have  one fence, only one fence
> callback is registered per sync_file.
> 
> v3: Comments from Chris Wilson and Christian König
> 	- struct sync_file lost status member in favor of fence_is_signaled()
> 	- drop use of fence_array_teardown()
> 	- use sizeof(*fence) to allocate only an array on fence pointers
> 
> v2: Comments from Chris Wilson and Christian König
> 	- Not using fence_ops anymore
> 	- fence_is_array() was created to differentiate fence from fence_array
> 	- fence_array_teardown() is now exported and used under fence_is_array()
> 	- struct sync_file lost num_fences member
> 
> Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
> Cc: Christian König <christian.koenig@xxxxxxx>
> Signed-off-by: Gustavo Padovan <gustavo.padovan@xxxxxxxxxxxxxxx>
> ---
>  drivers/dma-buf/sync_file.c          | 163 ++++++++++++++++++++++-------------
>  drivers/staging/android/sync_debug.c |  13 ++-
>  include/linux/sync_file.h            |  17 ++--
>  3 files changed, 118 insertions(+), 75 deletions(-)
> 
> diff --git a/drivers/dma-buf/sync_file.c b/drivers/dma-buf/sync_file.c
> index 9aaa608..7a5aaae 100644
> --- a/drivers/dma-buf/sync_file.c
> +++ b/drivers/dma-buf/sync_file.c
> @@ -28,11 +28,11 @@
>  
>  static const struct file_operations sync_file_fops;
>  
> -static struct sync_file *sync_file_alloc(int size)
> +static struct sync_file *sync_file_alloc(void)
>  {
>  	struct sync_file *sync_file;
>  
> -	sync_file = kzalloc(size, GFP_KERNEL);
> +	sync_file = kzalloc(sizeof(*sync_file), GFP_KERNEL);
>  	if (!sync_file)
>  		return NULL;
>  
> @@ -45,6 +45,8 @@ static struct sync_file *sync_file_alloc(int size)
>  
>  	init_waitqueue_head(&sync_file->wq);
>  
> +	INIT_LIST_HEAD(&sync_file->cb.node);
> +
>  	return sync_file;
>  
>  err:
> @@ -54,14 +56,11 @@ err:
>  
>  static void fence_check_cb_func(struct fence *f, struct fence_cb *cb)
>  {
> -	struct sync_file_cb *check;
>  	struct sync_file *sync_file;
>  
> -	check = container_of(cb, struct sync_file_cb, cb);
> -	sync_file = check->sync_file;
> +	sync_file = container_of(cb, struct sync_file, cb);
>  
> -	if (atomic_dec_and_test(&sync_file->status))
> -		wake_up_all(&sync_file->wq);
> +	wake_up_all(&sync_file->wq);
>  }
>  
>  /**
> @@ -76,22 +75,18 @@ struct sync_file *sync_file_create(struct fence *fence)
>  {
>  	struct sync_file *sync_file;
>  
> -	sync_file = sync_file_alloc(offsetof(struct sync_file, cbs[1]));
> +	sync_file = sync_file_alloc();
>  	if (!sync_file)
>  		return NULL;
>  
> -	sync_file->num_fences = 1;
> -	atomic_set(&sync_file->status, 1);
> +	sync_file->fence = fence;
> +
>  	snprintf(sync_file->name, sizeof(sync_file->name), "%s-%s%llu-%d",
>  		 fence->ops->get_driver_name(fence),
>  		 fence->ops->get_timeline_name(fence), fence->context,
>  		 fence->seqno);
>  
> -	sync_file->cbs[0].fence = fence;
> -	sync_file->cbs[0].sync_file = sync_file;
> -	if (fence_add_callback(fence, &sync_file->cbs[0].cb,
> -			       fence_check_cb_func))
> -		atomic_dec(&sync_file->status);
> +	fence_add_callback(fence, &sync_file->cb, fence_check_cb_func);
>  
>  	return sync_file;
>  }
> @@ -121,14 +116,48 @@ err:
>  	return NULL;
>  }
>  
> -static void sync_file_add_pt(struct sync_file *sync_file, int *i,
> -			     struct fence *fence)
> +static int sync_file_set_fence(struct sync_file *sync_file,
> +			       struct fence **fences, int num_fences)
> +{
> +	struct fence_array *array;
> +
> +	/*
> +	 * The reference for the fences in the new sync_file and holded

s/holded/held/

> +	 * in add_fence() during the merge procedure, so for num_fences == 1
> +	 * we already own a new reference to the fence. For num_fence > 1
> +	 * we own the reference of the fence_array creation.
> +	 */

Thanks, that really does help understanding the reference handling.

> +	if (num_fences == 1) {
> +		sync_file->fence = fences[0];
> +	} else {
> +		array = fence_array_create(num_fences, fences,
> +					   fence_context_alloc(1), 1, false);
> +		if (!array)
> +			return -ENOMEM;
> +
> +		sync_file->fence = (struct fence *)array;

I'd prefer sync_file->fence = &array->base;

> +	}
> +
> +	return 0;
> +}
> +
> +static struct fence **get_fences(struct sync_file *sync_file, int *num_fences)
> +{
> +	if (fence_is_array(sync_file->fence)) {
> +		struct fence_array *array = to_fence_array(sync_file->fence);
> +		*num_fences = array->num_fences;
> +		return array->fences;
> +	} else {
> +		*num_fences = 1;
> +		return &sync_file->fence;
> +	}
> +}
> +
> +static void add_fence(struct fence **fences, int *i, struct fence *fence)
>  {
> -	sync_file->cbs[*i].fence = fence;
> -	sync_file->cbs[*i].sync_file = sync_file;
> +	fences[*i] = fence;
>  
> -	if (!fence_add_callback(fence, &sync_file->cbs[*i].cb,
> -				fence_check_cb_func)) {
> +	if (!fence_is_signaled(fence)) {
>  		fence_get(fence);
>  		(*i)++;
>  	}
> @@ -147,16 +176,21 @@ static void sync_file_add_pt(struct sync_file *sync_file, int *i,
>  static struct sync_file *sync_file_merge(const char *name, struct sync_file *a,
>  					 struct sync_file *b)
>  {
> -	int num_fences = a->num_fences + b->num_fences;
>  	struct sync_file *sync_file;
> -	int i, i_a, i_b;
> -	unsigned long size = offsetof(struct sync_file, cbs[num_fences]);
> +	struct fence **fences, **a_fences, **b_fences;
> +	int i, i_a, i_b, num_fences, a_num_fences, b_num_fences;
>  
> -	sync_file = sync_file_alloc(size);
> +	sync_file = sync_file_alloc();
>  	if (!sync_file)
>  		return NULL;
>  
> -	atomic_set(&sync_file->status, num_fences);
> +	a_fences = get_fences(a, &a_num_fences);
> +	b_fences = get_fences(b, &b_num_fences);
> +	num_fences = a_num_fences + b_num_fences;

Do we need overflow paranoia here? Probably, as this is on a user ioctl
path.

> +
> +	fences = kcalloc(num_fences, sizeof(*fences), GFP_KERNEL);
> +	if (!fences)
> +		goto err;
>  
>  	/*
>  	 * Assume sync_file a and b are both ordered and have no
> @@ -165,55 +199,66 @@ static struct sync_file *sync_file_merge(const char *name, struct sync_file *a,
>  	 * If a sync_file can only be created with sync_file_merge
>  	 * and sync_file_create, this is a reasonable assumption.
>  	 */
> -	for (i = i_a = i_b = 0; i_a < a->num_fences && i_b < b->num_fences; ) {
> -		struct fence *pt_a = a->cbs[i_a].fence;
> -		struct fence *pt_b = b->cbs[i_b].fence;
> +	for (i = i_a = i_b = 0; i_a < a_num_fences && i_b < b_num_fences; ) {
> +		struct fence *pt_a = a_fences[i_a];
> +		struct fence *pt_b = b_fences[i_b];
>  
>  		if (pt_a->context < pt_b->context) {
> -			sync_file_add_pt(sync_file, &i, pt_a);
> +			add_fence(fences, &i, pt_a);
>  
>  			i_a++;
>  		} else if (pt_a->context > pt_b->context) {
> -			sync_file_add_pt(sync_file, &i, pt_b);
> +			add_fence(fences, &i, pt_b);
>  
>  			i_b++;
>  		} else {
>  			if (pt_a->seqno - pt_b->seqno <= INT_MAX)
> -				sync_file_add_pt(sync_file, &i, pt_a);
> +				add_fence(fences, &i, pt_a);
>  			else
> -				sync_file_add_pt(sync_file, &i, pt_b);
> +				add_fence(fences, &i, pt_b);
>  
>  			i_a++;
>  			i_b++;
>  		}
>  	}
>  
> -	for (; i_a < a->num_fences; i_a++)
> -		sync_file_add_pt(sync_file, &i, a->cbs[i_a].fence);
> +	for (; i_a < a_num_fences; i_a++)
> +		add_fence(fences, &i, a_fences[i_a]);
> +
> +	for (; i_b < b_num_fences; i_b++)
> +		add_fence(fences, &i, b_fences[i_b]);
> +
> +	if (num_fences > i) {
> +		fences = krealloc(fences, i * sizeof(**fences),

Hmm, still too large, sizeof(*fences);

Looking pretty good. Putting fence-array to good use.

With those minor tweaks,
Reviewed-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre