[Intel-gfx] [PATCH] drm/fbdev: Update legacy plane->fb refcounting for atomic restore

[David Herrmann]

Hi

On Tue, Sep 22, 2015 at 11:55 AM, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> From: Matt Roper <matthew.d.roper at intel.com>
>
> Starting with commit
>
>         commit 28cc504e8d52248962f5b485bdc65f539e3fe21d
>         Author: Rob Clark <robdclark at gmail.com>
>         Date:   Tue Aug 25 15:36:00 2015 -0400
>
>             drm/i915: enable atomic fb-helper
>
> I've been seeing some panics on i915 when the DRM master shuts down that appear
> to be caused by using an already-freed framebuffer (i.e., we're unexpectedly
> dropping our initial FB's reference count to 0 and freeing it, which causes a
> crash when we try to restore it later).  Digging deeper, the state FB
> refcounting is working as expected, but we seem to be missing proper
> refcounting on the legacy plane->fb pointers in the new atomic fbdev code.
>
> Tracking plane->old_fb and then doing a ref/unref at the end of the
> fbdev restore like we do in the legacy ioctl's ensures we don't miscount
> references on plane->fb and avoids the panics.
>
> v2 from Daniel:
>
> Really do what the atomic ioctl does:
> - Also update plane->fb and plane->crtc.
> - Clear out plane->old_fb on failures too.
>
> Cc: Rob Clark <robdclark at gmail.com>
> Cc: intel-gfx at lists.freedesktop.org
> Signed-off-by: Matt Roper <matthew.d.roper at intel.com> (v1)
> Signed-off-by: Daniel Vetter <daniel.vetter at ffwll.ch>
> ---
>  drivers/gpu/drm/drm_fb_helper.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
>
> diff --git a/drivers/gpu/drm/drm_fb_helper.c b/drivers/gpu/drm/drm_fb_helper.c
> index 64fc5ca8fda2..1b564a69f561 100644
> --- a/drivers/gpu/drm/drm_fb_helper.c
> +++ b/drivers/gpu/drm/drm_fb_helper.c
> @@ -352,6 +352,8 @@ retry:
>         drm_for_each_plane(plane, dev) {
>                 struct drm_plane_state *plane_state;
>
> +               plane->old_fb = plane->fb;
> +
>                 plane_state = drm_atomic_get_plane_state(state, plane);
>                 if (IS_ERR(plane_state)) {
>                         ret = PTR_ERR(plane_state);
> @@ -382,6 +384,20 @@ retry:
>         }
>
>         ret = drm_atomic_commit(state);
> +
> +       drm_for_each_plane(plane, dev) {
> +               if (ret == 0) {
> +                       struct drm_framebuffer *new_fb = plane->state->fb;
> +                       if (new_fb)
> +                               drm_framebuffer_reference(new_fb);
> +                       plane->fb = new_fb;
> +                       plane->crtc = plane->state->crtc;
> +
> +                       if (plane->old_fb)
> +                               drm_framebuffer_unreference(plane->old_fb);
> +               }
> +       }
> +
>         if (ret != 0)
>                 goto fail;

Why move the fixup _before_ the error-check? drm_atomic_ioctl() does
this to make sure it never leaves stale "old_fb" pointers, but this is
clearly not what you do here. So either move it after the error-check
and drop the then redundant "if (ret ==0)" from the loop, or move it
into the "fail:" condition (maybe rename it) and properly take care of
"old_fb" and not-freeing "state".

Thanks
David