[Intel-gfx] [PATCH 12/14] drm: Move master pointer from drm_minor to drm_device

[Daniel Vetter]

On Wed, Jun 15, 2016 at 01:10:35PM +0100, Chris Wilson wrote:
> On Tue, Jun 14, 2016 at 08:51:07PM +0200, Daniel Vetter wrote:
> > There can only be one current master, and it's for the overall device.
> > Render/control minors don't support master-based auth at all.
> > 
> > This simplifies the master logic a lot, at least in my eyes: All these
> > additional pointer chases are just confusing.
> 
> One master for the device, on the struct drm_device, as opposed to hidden
> behind the first of three minors, makes sense.
> 
> > @@ -128,13 +128,13 @@ static int drm_new_set_master(struct drm_device *dev, struct drm_file *fpriv)
> >  	lockdep_assert_held_once(&dev->master_mutex);
> >  
> >  	/* create a new master */
> > -	fpriv->minor->master = drm_master_create(fpriv->minor->dev);
> > -	if (!fpriv->minor->master)
> > +	dev->master = drm_master_create(dev);
> > +	if (!dev->master)
> >  		return -ENOMEM;
> >  
> >  	/* take another reference for the copy in the local file priv */
> >  	old_master = fpriv->master;
> > -	fpriv->master = drm_master_get(fpriv->minor->master);
> > +	fpriv->master = drm_master_get(dev->master);
> >  
> >  	if (dev->driver->master_create) {
> >  		ret = dev->driver->master_create(dev, fpriv->master);
> 
> > @@ -234,10 +234,10 @@ int drm_master_open(struct drm_file *file_priv)
> >  	/* if there is no current master make this fd it, but do not create
> >  	 * any master object for render clients */
> >  	mutex_lock(&dev->master_mutex);
> > -	if (!file_priv->minor->master)
> > +	if (!dev->master)
> >  		ret = drm_new_set_master(dev, file_priv);
> >  	else
> > -		file_priv->master = drm_master_get(file_priv->minor->master);
> > +		file_priv->master = drm_master_get(dev->master);
> >  	mutex_unlock(&dev->master_mutex);
> 
> You could take the opportunity to make this a bit simpler:
> 
> 	if (!READ_ONCE(dev->master)) {
> 		int ret;
> 
> 		ret = 0;
> 		mutex_lock(&dev->master_mutex);
> 		if (!dev->master)
> 			ret = drm_new_master(dev);
> 		mutex_unlock(&dev->master_mutex);
> 		if (ret)
> 			return ret;
> 	}
> 
> 	file_priv->master = drm_master_get(dev->master);

drm_master_get(dev->master) must be under the master_mutex, without it we
could race with a drm_master_put(&dev->master) and end up doing a kref_get
when the refcount already reached 0.

> 	return 0;
> 
> Just to straighten out the kref dance.
> 
> >  
> >  	return ret;
> > @@ -271,11 +271,11 @@ void drm_master_release(struct drm_file *file_priv)
> >  		mutex_unlock(&dev->struct_mutex);
> >  	}
> >  
> > -	if (file_priv->minor->master == file_priv->master) {
> > +	if (dev->master == file_priv->master) {
> >  		/* drop the reference held my the minor */
> >  		if (dev->driver->master_drop)
> >  			dev->driver->master_drop(dev, file_priv, true);
> > -		drm_master_put(&file_priv->minor->master);
> > +		drm_master_put(&dev->master);
> 
> This still makes me uneasy. This is not equivalent to dropmaster_ioctl
> and subsequent setmaster_ioctl will fail as dev->master is still
> assigned (but the owner has gone).

drm_master_put clears the pointer passed to it, so dev->master will be set
to NULL. And it does the same as drop_master (wrt dev->master at least,
master_release also needs to clean up file_priv->master on top). Not sure
it's worth it to extract those 5 lines into a __drm_drop_master() helper
function? I can respin with that if you want. On the master_open/setmaster
side the shared code is already extracted in drm_new_set_master().
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch