[RFC v2 3/3] dma-buf/sync_file: rework fence storage in struct file

[Gustavo Padovan]

2016-06-28 Chris Wilson <chris at chris-wilson.co.uk>:

> On Tue, Jun 28, 2016 at 11:25:00AM -0300, Gustavo Padovan wrote:
> > 2016-06-28 Chris Wilson <chris at chris-wilson.co.uk>:
> > 
> > > On Mon, Jun 27, 2016 at 04:29:22PM -0300, Gustavo Padovan wrote:
> > > > From: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
> > > > 
> > > > 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.
> > > > 
> > > > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > > > Cc: Christian König <christian.koenig at amd.com>
> > > > Signed-off-by: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
> > > > ---
> > > > @@ -76,21 +76,19 @@ 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;
> > > > +	sync_file->fence = fence;
> > > > +
> > > >  	atomic_set(&sync_file->status, 1);
> > > 
> > > sync_file->status => fence_is_signaled(sync_file->fence);
> > > 
> > > Both should just be an atomic read, except fence_is_signaled() will then
> > > do a secondary poll.
> > 
> > Not sure I follow. I set it to 1 here, but below when we call
> > fence_add_callback() and the fence is already signalled atomic_dec sets
> > sync_file->status to 0.
> 
> I'm just saying that usage sync_file->status is equivalent to
> fence_is_signaled(), i.e. we reduce the amount of bookkeeping local to
> sync_file.

Indeed, that makes sense, I'll remove status from sync_file.

> 
> > > >  	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))
> > > > +	if (fence_add_callback(fence, &sync_file->cb, fence_check_cb_func))
> > > >  		atomic_dec(&sync_file->status);
> > > >  
> > > >  	return sync_file;
> > > > @@ -121,14 +119,42 @@ 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)
> > > >  {
> > > > -	sync_file->cbs[*i].fence = fence;
> > > > -	sync_file->cbs[*i].sync_file = sync_file;
> > > > +	struct fence_array *array;
> > > > +
> > > > +	if (num_fences == 1) {
> > > > +		sync_file->fence = fences[0];
> > > 
> > > This steals the references.
> > > 
> > > > +	} else {
> > > > +		array = fence_array_create(num_fences, fences,
> > > > +					   fence_context_alloc(1), 1, false);
> > > 
> > > This creates a reference.
> > > 
> > > When we call fence_put(sync_fence->fence) we release a reference we
> > > never owned if num_fences == 1.
> > 
> > No, sync_file_merge() gets a new reference for each fence it is going to
> > add to the new fence. So for num_fences == 1 when sync_file->fence is
> > set we already hold a reference to it, so no matter if it is a fence or
> > a array we own a reference.
> 
> Ugh. Root cause appears to be that fence_array_create() does not behave
> how I would expect, in that it borrows references to the fences and
> not own a reference to the fences in its array. I beg for a comment as
> this function is very counter-intuitive for me.

There is this in fence_array_create():

 * The caller should allocte the fences array with num_fences size              
 * and fill it with the fences it wants to add to the object. Ownership of this 
 * array is take and fence_put() is used on each fence on release.     

	Gustavo