daniel at ffwll.ch
Mon Dec 12 15:41:40 PST 2011
On Fri, Dec 9, 2011 at 18:44, Dave Airlie <airlied at gmail.com> wrote:
> On Fri, Dec 9, 2011 at 5:40 PM, Rob Clark <rob.clark at linaro.org> wrote:
>> On Wed, Oct 19, 2011 at 8:27 AM, Daniel Vetter <daniel at ffwll.ch> wrote:
>>>> +static void omap_framebuffer_destroy(struct drm_framebuffer *fb)
>>>> + struct drm_device *dev = fb->dev;
>>>> + struct omap_framebuffer *omap_fb = to_omap_framebuffer(fb);
>>>> + DBG("destroy: FB ID: %d (%p)", fb->base.id, fb);
>>>> + drm_framebuffer_cleanup(fb);
>>> This is a bit a general mess in kms. All drivers need to call this, and
>>> for added hilarity
>>> - drm_crtc.c drm_mode_rmfb has a TODO that this is missing
>>> - nouveau/radeon only do this _after_ unref'ing the backing storage
>>> - gma500 also tries to restore the kernel console here which should be
>>> done in lastclose (because you might disable another userspace fb on
>>> another output).
>>> Can I prod you to write the patches to clean this up and move
>>> drm_framebuffer_cleanup into common code?
>> fyi, I'm starting to look at this.. looks like crtc/encoder/connector
>> cleanup could also benefit from some similar treatment. I'll start w/
>> just fb, since that is a resource that is actually dynamically
>> created/destroyed, so seems a bit more critical..
> I'm not sure reversing the flow here is the right answer, this is
> mainly trying to avoid midlayering things, the driver controls when
> stuff happens and it controls init the framebuffer object it controls
> destroying it as well.
The above proposal was more in the light of eventually switching to
properly ref-counted framebuffers. The common kms code would then keep
track of the references it knows about from the current modeset
configuration and hence also responsible for dropping them (i.e. what
drm_framebuffer_cleanup currently does). The driver could then still
hold onto references it needs internally and is also still in control
of fb destruction - the driver's fb_destroy would be called when the
last reference disappears.
So I don't see an "inversion of control through stupid middle-layer"
issue here, more like "resource lifetime mess through lack of
refcounting". But I agree that we can keep the drm_framebuffer_cleanup
where it currently is until it completely disappears with proper fb
daniel.vetter at ffwll.ch - +41 (0) 79 364 57 48 - http://blog.ffwll.ch
More information about the dri-devel