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@xxxxxxxxxxxxxxxxxx>
> > Cc: Lukas Wunner <lukas@xxxxxxxxx>
> > ---
> > 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
_______________________________________________
Intel-gfx mailing list
Intel-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
