[PATCH 09/35] drm/gem: support BO freeing without dev->struct_mutex

[Daniel Vetter]

On Tue, Apr 26, 2016 at 08:47:32PM +0100, Chris Wilson wrote:
> On Tue, Apr 26, 2016 at 07:29:42PM +0200, Daniel Vetter wrote:
> > Finally all the core gem and a lot of drivers are entirely free of
> > dev->struct_mutex depencies, and we can start to have an entirely
> > lockless unref path.
> > 
> > To make sure that no one who touches the core code accidentally breaks
> > existing drivers which still require dev->struct_mutex I've made the
> > might_lock check unconditional.
> > 
> > While at it de-inline the ref/unref functions, they've become a bit
> > too big.
> > 
> > v2: Make it not leak like a sieve.
> > 
> > Signed-off-by: Daniel Vetter <daniel.vetter at intel.com>
> > ---
> >  drivers/gpu/drm/drm_gem.c | 64 ++++++++++++++++++++++++++++++++++++++---------
> >  include/drm/drmP.h        | 12 ++++++++-
> >  include/drm/drm_gem.h     | 45 ++-------------------------------
> >  3 files changed, 65 insertions(+), 56 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/drm_gem.c b/drivers/gpu/drm/drm_gem.c
> > index 25dac31eef37..8f2eff448bb5 100644
> > --- a/drivers/gpu/drm/drm_gem.c
> > +++ b/drivers/gpu/drm/drm_gem.c
> > @@ -788,16 +788,7 @@ drm_gem_object_release(struct drm_gem_object *obj)
> >  }
> >  EXPORT_SYMBOL(drm_gem_object_release);
> >  
> > -/**
> > - * drm_gem_object_free - free a GEM object
> > - * @kref: kref of the object to free
> > - *
> > - * Called after the last reference to the object has been lost.
> > - * Must be called holding struct_ mutex
> > - *
> > - * Frees the object
> > - */
> > -void
> > +static void
> >  drm_gem_object_free(struct kref *kref)
> >  {
> >  	struct drm_gem_object *obj =
> > @@ -806,10 +797,59 @@ drm_gem_object_free(struct kref *kref)
> >  
> >  	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
> >  
> > -	if (dev->driver->gem_free_object != NULL)
> > +	if (dev->driver->gem_free_object_unlocked != NULL)
> > +		dev->driver->gem_free_object_unlocked(obj);
> > +	else if (dev->driver->gem_free_object != NULL)
> >  		dev->driver->gem_free_object(obj);
> >  }
> > -EXPORT_SYMBOL(drm_gem_object_free);
> > +
> > +/**
> > + * drm_gem_object_unreference_unlocked - release a GEM BO reference
> > + * @obj: GEM buffer object
> > + *
> > + * This releases a reference to @obj. Callers must not hold the
> > + * dev->struct_mutex lock when calling this function.
> > + */
> > +void
> > +drm_gem_object_unreference_unlocked(struct drm_gem_object *obj)
> > +{
> 
> I have i915.ko in a state where we could use this as well. I would like
> to have our inlines back though. I would add
> 
> static inline void
> i915_gem_object_put(struct drm_i915_gem_object *obj)
> {
> 	kref_put(&obj->base.refcount, drm_gem_object_free);
> }
> 
> Hmm, considering how simple that is, maybe I won't ask for
> 
> static inline void
> __drm_gem_object_unreference_unlocked(struct drm_gem_object *obj)
> {
> 	BUG_ON(obj->dev->driver->gem_free_object == NULL);
> 	kref_put(&obj->refcount, drm_gem_object_free);
> }

Yeah, something like that makes sense. Imo the core versions really need
all the silly locking checks and all that to make sure no one abuses them
by accident, or breaks a driver still using struct_mutex. Adding a
fastpath for drivers using gem_free_object_unlocked just wrapping the
kref_put certainly looks like a good idea.

And if we document the exact use case I think we could even nuke the
BUG_ON too. This can be audited with a simple

$ git grep __drm_gem_object_unreference_unlocked -- drivers/gpu/drm/$driver

so imo nothing too dangerous.

Ack/r-b on the patch itself?
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch