[PATCH] drm: Don't assign fbs for universal cursor support to files

[Daniel Vetter]

On Thu, Feb 26, 2015 at 06:50:19PM -0800, Matt Roper wrote:
> On Wed, Feb 25, 2015 at 01:45:26PM +0000, Chris Wilson wrote:
> > The internal framebuffers we create to remap legacy cursor ioctls to
> > plane operations for the universal plane support shouldn't be linke to
> > the file like normal userspace framebuffers. This bug goes back to the
> > original universal cursor plane support introduced in
> > 
> > commit 161d0dc1dccb17ff7a38f462c7c0d4ef8bcc5662
> > Author: Matt Roper <matthew.d.roper at intel.com>
> > Date:   Tue Jun 10 08:28:10 2014 -0700
> > 
> >     drm: Support legacy cursor ioctls via universal planes when possible (v4)
> > 
> > The isn't too disastrous since fbs are small, we only create one when the
> > cursor bo gets changed and ultimately they'll be reaped when the window
> > server restarts.
> > 
> > Conceptually we'd want to just pass NULL for file_priv when creating it,
> > but the driver needs the file to lookup the underlying buffer object for
> > cursor id. Instead let's move the file_priv linking out of
> > add_framebuffer_internal() into the addfb ioctl implementation, which is
> > the only place it is needed. And also rename the function for a more
> > accurate since it only creates the fb, but doesn't add it anywhere.
> > 
> > Signed-off-by: Daniel Vetter <daniel.vetter at intel.com> (fix & commit msg)
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk> (provider of lipstick)
> > Cc: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > Cc: Matt Roper <matthew.d.roper at intel.com>
> > Cc: Rob Clark <robdclark at gmail.com>
> > Cc: Dave Airlie <airlied at redhat.com>
> > Cc: stable at vger.kernel.org
> > ---
> >  drivers/gpu/drm/drm_crtc.c | 35 +++++++++++++++++++----------------
> >  1 file changed, 19 insertions(+), 16 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/drm_crtc.c b/drivers/gpu/drm/drm_crtc.c
> > index 927f3445ff38..4c78d12c5418 100644
> > --- a/drivers/gpu/drm/drm_crtc.c
> > +++ b/drivers/gpu/drm/drm_crtc.c
> > @@ -43,9 +43,10 @@
> >  #include "drm_crtc_internal.h"
> >  #include "drm_internal.h"
> >  
> > -static struct drm_framebuffer *add_framebuffer_internal(struct drm_device *dev,
> > -							struct drm_mode_fb_cmd2 *r,
> > -							struct drm_file *file_priv);
> > +static struct drm_framebuffer *
> > +internal_framebuffer_create(struct drm_device *dev,
> > +			    struct drm_mode_fb_cmd2 *r,
> > +			    struct drm_file *file_priv);
> >  
> >  /* Avoid boilerplate.  I'm tired of typing. */
> >  #define DRM_ENUM_NAME_FN(fnname, list)				\
> > @@ -2919,13 +2920,11 @@ static int drm_mode_cursor_universal(struct drm_crtc *crtc,
> >  	 */
> >  	if (req->flags & DRM_MODE_CURSOR_BO) {
> >  		if (req->handle) {
> > -			fb = add_framebuffer_internal(dev, &fbreq, file_priv);
> > +			fb = internal_framebuffer_create(dev, &fbreq, file_priv);
> >  			if (IS_ERR(fb)) {
> >  				DRM_DEBUG_KMS("failed to wrap cursor buffer in drm framebuffer\n");
> >  				return PTR_ERR(fb);
> >  			}
> > -
> > -			drm_framebuffer_reference(fb);
> 
> Sorry for the delay reviewing this.  I'll provide an i-g-t test that
> checks for these memory leaks shortly.
> 
> If I'm not mistaken, this patch will work properly for normal operation,
> but I think we might run into problems if your display server gets
> killed while a wrapped cursor is onscreen and we need to restore the
> fbdev mode.  
> 
> From what I can see, we'll wind up in drm_plane_force_disable() which
> does:
> 
>         plane->old_fb = plane->fb;
>         ret = plane->funcs->disable_plane(plane);
>         if (ret) {
>                 DRM_ERROR("failed to disable plane with busy fb\n");
>                 plane->old_fb = NULL;
>                 return;
>         }
>         /* disconnect the plane from the fb and crtc: */
>         __drm_framebuffer_unreference(plane->old_fb);
> 
> Note the internal __drm_framebuffer_unreference() here rather than a
> traditional drm_framebuffer_unreference().  The internal version is only
> supposed to be used when we know we're not releasing the last reference
> and BUG()'s out if we actually take the reference count down to zero
> (which is exactly what we do in this case).
> 
> I guess we need to just do away with __drm_framebuffer_unreference() now since
> its only call-site is no longer guaranteed to be working on framebuffers that
> still have a remaining reference.

Nice issue you spotted, but this actually goes back to making the idr
reference a weak one in

commit 83f45fc360c8e16a330474860ebda872d1384c8c
Author: Daniel Vetter <daniel.vetter at ffwll.ch>
Date:   Wed Aug 6 09:10:18 2014 +0200

    drm: Don't grab an fb reference for the idr

The reference which was guaranteed to be around was from the idr. The race
is really small though since we still remove fbs right away once they go
away from the idr.

I'll prep a patch for this.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch