On Wed, Jun 15, 2016 at 6:33 PM, Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> wrote:
> On Wed, Jun 15, 2016 at 06:01:41PM +0200, Daniel Vetter wrote:
>> 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.
>
> Something is very fishy then. The behaviour of drm_new_master() would
> appear to be to create a drm_master owned by the device, but really it is
> owned by file_priv?
>
> * checks back on drm_master
>
> So drm_master is the authentication structure that needs to be unique to
> a hierachy. So drm_new_set_master() and here really do appear backwards.
>
> The old drm_new_set_master() makes more sense, it assigns to the
> file_priv, and then performs a setmaster operation. In that ordering,
> you could even do the file_priv local operation of creating the new
> master structure before taking the lock to perform setmaster.
I didn't change the logic, the old new_set_master didn't first create
the master for file_priv, it first created it for file_priv->minor,
and that drm_minor is the device-unique drm_minor for the
primary/legacy node. The change only moves from that drm_minor for the
legacy node to the drm_device.
Wrt refences, every file_priv holds a ref for it's master. On top of
that the current master structure (after this patch stored in
dev->master, before in dev->minors[legacy]->master. The only tricky
bit is that when the current master closes, we do an implicit
dropmaster first.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
