[Intel-gfx] [PATCH 2/3] drm/i915/fbdev: Limit the global async-domain synchronization

[Chris Wilson]

On Tue, Jun 21, 2016 at 09:35:44AM +0200, Daniel Vetter wrote:
> On Fri, Jun 17, 2016 at 06:54:48PM +0100, Chris Wilson wrote:
> > During cleanup we have to synchronise with the async task we are using
> > to initialise and register our fbdev. Currently, we are using a full
> > synchronisation on the global domain, but we can restrict this to just
> > synchronising up to our task if we remember our cookie.
> > 
> > v2: async_synchronize_cookie() takes an exclusive upper bound, to
> > synchronize with our task we have to pass in the next cookie.
> > v3: Drop premature disregarding of the active cookie (we need to wait
> > until the task is complete before continuing in the teardown).
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Lukas Wunner <lukas at wunner.de>
> > ---
> >  drivers/gpu/drm/i915/intel_drv.h   |  1 +
> >  drivers/gpu/drm/i915/intel_fbdev.c | 29 ++++++++++++++++-------------
> >  2 files changed, 17 insertions(+), 13 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/intel_drv.h b/drivers/gpu/drm/i915/intel_drv.h
> > index 0c1dc9bae170..b657ddd2d078 100644
> > --- a/drivers/gpu/drm/i915/intel_drv.h
> > +++ b/drivers/gpu/drm/i915/intel_drv.h
> > @@ -159,6 +159,7 @@ struct intel_framebuffer {
> >  struct intel_fbdev {
> >  	struct drm_fb_helper helper;
> >  	struct intel_framebuffer *fb;
> > +	async_cookie_t cookie;
> >  	int preferred_bpp;
> >  };
> >  
> > diff --git a/drivers/gpu/drm/i915/intel_fbdev.c b/drivers/gpu/drm/i915/intel_fbdev.c
> > index 4babefc51eb2..638e420a59cb 100644
> > --- a/drivers/gpu/drm/i915/intel_fbdev.c
> > +++ b/drivers/gpu/drm/i915/intel_fbdev.c
> > @@ -538,8 +538,7 @@ static const struct drm_fb_helper_funcs intel_fb_helper_funcs = {
> >  	.fb_probe = intelfb_create,
> >  };
> >  
> > -static void intel_fbdev_destroy(struct drm_device *dev,
> > -				struct intel_fbdev *ifbdev)
> > +static void intel_fbdev_destroy(struct intel_fbdev *ifbdev)
> >  {
> >  	/* We rely on the object-free to release the VMA pinning for
> >  	 * the info->screen_base mmaping. Leaking the VMA is simpler than
> > @@ -552,12 +551,14 @@ static void intel_fbdev_destroy(struct drm_device *dev,
> >  	drm_fb_helper_fini(&ifbdev->helper);
> >  
> >  	if (ifbdev->fb) {
> > -		mutex_lock(&dev->struct_mutex);
> > +		mutex_lock(&ifbdev->helper.dev->struct_mutex);
> >  		intel_unpin_fb_obj(&ifbdev->fb->base, BIT(DRM_ROTATE_0));
> > -		mutex_unlock(&dev->struct_mutex);
> > +		mutex_unlock(&ifbdev->helper.dev->struct_mutex);
> >  
> >  		drm_framebuffer_remove(&ifbdev->fb->base);
> >  	}
> > +
> > +	kfree(ifbdev);
> >  }
> >  
> >  /*
> > @@ -732,32 +733,34 @@ int intel_fbdev_init(struct drm_device *dev)
> >  
> >  static void intel_fbdev_initial_config(void *data, async_cookie_t cookie)
> >  {
> > -	struct drm_i915_private *dev_priv = data;
> > -	struct intel_fbdev *ifbdev = dev_priv->fbdev;
> > +	struct intel_fbdev *ifbdev = data;
> >  
> >  	/* Due to peculiar init order wrt to hpd handling this is separate. */
> >  	if (drm_fb_helper_initial_config(&ifbdev->helper,
> >  					 ifbdev->preferred_bpp))
> > -		intel_fbdev_fini(dev_priv->dev);
> > +		intel_fbdev_fini(ifbdev->helper.dev);
> >  }
> >  
> >  void intel_fbdev_initial_config_async(struct drm_device *dev)
> >  {
> > -	async_schedule(intel_fbdev_initial_config, to_i915(dev));
> > +	struct intel_fbdev *ifbdev = to_i915(dev)->fbdev;
> > +
> > +	ifbdev->cookie = async_schedule(intel_fbdev_initial_config, ifbdev);
> >  }
> >  
> >  void intel_fbdev_fini(struct drm_device *dev)
> >  {
> >  	struct drm_i915_private *dev_priv = dev->dev_private;
> > -	if (!dev_priv->fbdev)
> > +	struct intel_fbdev *ifbdev = dev_priv->fbdev;
> > +
> > +	if (!ifbdev)
> >  		return;
> >  
> >  	flush_work(&dev_priv->fbdev_suspend_work);
> > +	if (ifbdev->cookie && !current_is_async())
> > +		async_synchronize_cookie(ifbdev->cookie + 1);
> 
> First I went like wtf about the cookie+1, but the main use case for this
> function (or intended use-case at least) is to synchronize with everything
> before your own async task when you register. To uphold deterministic dev
> node ordering ...

Yup, it's a total wtf. Definitely scores high on Rusty's how to screw
with your API consumers. The whole async-vs-sync kernel is the same. If
only the kernel had fences as a completion variable...

> Needs a comment in the code imo, this is too suprising:
> 
> 	/* Only synchronizes with all _preceeding_ async tasks, hence + 1 */
> 
> Or whatever you feel like.

Ok.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre