[PATCH] dma-buf/sync_file: only enable fence signalling during wait

[Daniel Vetter]

On Tue, Jul 12, 2016 at 10:46:45AM +0200, Maarten Lankhorst wrote:
> Op 11-07-16 om 22:27 schreef Gustavo Padovan:
> > 2016-07-10 Maarten Lankhorst <maarten.lankhorst at linux.intel.com>:
> >
> >> Op 08-07-16 om 17:44 schreef Gustavo Padovan:
> >>> From: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
> >>>
> >>> Signalling doesn't need to be enabled at sync_file creation, it is only
> >>> required if userspace waiting the fence to signal through poll().
> >>>
> >>> Thus we delay fence_add_callback() until poll is called. It only adds the
> >>> callback the first time poll() is called. This avoid re-adding the same
> >>> callback multiple times.
> >>>
> >>> v2: rebase and update to work with new fence support for sync_file
> >>>
> >>> Signed-off-by: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
> >>> ---
> >>> This patch applies on top of my latest sync_file changes to support
> >>> fence_array: https://lkml.org/lkml/2016/7/4/534
> >>>
> >>>  drivers/dma-buf/sync_file.c | 23 ++++++++++++++---------
> >>>  include/linux/sync_file.h   |  2 ++
> >>>  2 files changed, 16 insertions(+), 9 deletions(-)
> >>>
> >>> diff --git a/drivers/dma-buf/sync_file.c b/drivers/dma-buf/sync_file.c
> >>> index 61a687c..1db4a64 100644
> >>> --- a/drivers/dma-buf/sync_file.c
> >>> +++ b/drivers/dma-buf/sync_file.c
> >>> @@ -86,8 +86,6 @@ struct sync_file *sync_file_create(struct fence *fence)
> >>>  		 fence->ops->get_timeline_name(fence), fence->context,
> >>>  		 fence->seqno);
> >>>  
> >>> -	fence_add_callback(fence, &sync_file->cb, fence_check_cb_func);
> >>> -
> >>>  	return sync_file;
> >>>  }
> >>>  EXPORT_SYMBOL(sync_file_create);
> >>> @@ -269,9 +267,6 @@ static struct sync_file *sync_file_merge(const char *name, struct sync_file *a,
> >>>  		goto err;
> >>>  	}
> >>>  
> >>> -	fence_add_callback(sync_file->fence, &sync_file->cb,
> >>> -			   fence_check_cb_func);
> >>> -
> >>>  	strlcpy(sync_file->name, name, sizeof(sync_file->name));
> >>>  	return sync_file;
> >>>  
> >>> @@ -286,7 +281,6 @@ static void sync_file_free(struct kref *kref)
> >>>  	struct sync_file *sync_file = container_of(kref, struct sync_file,
> >>>  						     kref);
> >>>  
> >>> -	fence_remove_callback(sync_file->fence, &sync_file->cb);
> >>>  	fence_put(sync_file->fence);
> >>>  	kfree(sync_file);
> >>>  }
> >>> @@ -306,13 +300,24 @@ static unsigned int sync_file_poll(struct file *file, poll_table *wait)
> >>>  
> >>>  	poll_wait(file, &sync_file->wq, wait);
> >>>  
> >>> +	if (!sync_file->enabled) {
> >>> +		fence_add_callback(sync_file->fence, &sync_file->cb,
> >>> +				   fence_check_cb_func);
> >>> +		sync_file->enabled = true;
> >>> +	}
> >> Won't this blow up completely with 2 threads polling at the same time?
> > Indeed, using atomic operations on enabled should fix this.
> No, it still would blow up without locking around fence_remove/add_callback too..
> 
> Personally I would just add the callback once, then remove it in destructor.
> 
> Something like:
> 
> poll:
> if (!atomic_xchg(&sync_file->enabled, 1)) {
> if (fence_add_callback(...) < 0)
> wake up sync_file->wq, fence is signaled
> }
> 
> sync_file_free:
> if (atomic_read(&sync_file->enabled))
> fence_remove_callback(...);
> 
> fence_put()
> 
> It's not like fence can disable hw signaling when all callbacks are removed anyway,
> it's harmless to keep it on the list.

+1 from me on this plan, removing the callbacks doesn't provide any value
(before the sync_file is destroyed). Also, when someone polls, you can
expect to get repolled a lot until the fences finally all signal.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch