Re: [PATCH v2 1/9] drm: atmel-hlcdc: add a ->cleanup_fb() operation

From: Daniel Vetter
Date: Fri Mar 18 2016 - 13:58:16 EST


On Thu, Mar 17, 2016 at 09:49:42AM +0100, Boris Brezillon wrote:
> Hi Daniel,
>
> On Wed, 16 Mar 2016 16:17:38 +0100
> Daniel Vetter <daniel@xxxxxxxx> wrote:
>
> > On Wed, Mar 16, 2016 at 02:57:35PM +0100, Boris Brezillon wrote:
> > > Add a ->cleanup_fb() operation to avoid memory leaks when the atomic
> > > operation is interrupted after the ->prepare_fb() call.
> > >
> > > Signed-off-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxxxxxxxxx>
> > > Fixes 2389fc1 ("drm: atmel-hlcdc: Atomic mode-setting conversion")
> > > Reviewed-by: Nicolas Ferre <nicolas.ferre@xxxxxxxxx>
> > > Tested-by: Nicolas Ferre <nicolas.ferre@xxxxxxxxx>
> > > ---
> > > drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_dc.h | 2 ++
> > > drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c | 22 +++++++++++++++++++---
> > > 2 files changed, 21 insertions(+), 3 deletions(-)
> > >
> > > diff --git a/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_dc.h b/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_dc.h
> > > index fed517f..ec64140 100644
> > > --- a/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_dc.h
> > > +++ b/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_dc.h
> > > @@ -81,11 +81,13 @@ struct atmel_hlcdc_plane_properties {
> > > * @layer: HLCDC layer structure
> > > * @properties: pointer to the property definitions structure
> > > * @rotation: current rotation status
> > > + * @prepared: flagging the plane has prepared for an atomic update
> > > */
> > > struct atmel_hlcdc_plane {
> > > struct drm_plane base;
> > > struct atmel_hlcdc_layer layer;
> > > struct atmel_hlcdc_plane_properties *properties;
> > > + bool prepared;
> > > };
> > >
> > > static inline struct atmel_hlcdc_plane *
> > > diff --git a/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c b/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c
> > > index 1ffe9c3..35027d0 100644
> > > --- a/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c
> > > +++ b/drivers/gpu/drm/atmel-hlcdc/atmel_hlcdc_plane.c
> > > @@ -715,11 +715,25 @@ static int atmel_hlcdc_plane_prepare_fb(struct drm_plane *p,
> > > const struct drm_plane_state *new_state)
> > > {
> > > struct atmel_hlcdc_plane *plane = drm_plane_to_atmel_hlcdc_plane(p);
> > > + int ret;
> > >
> > > - if (!new_state->fb)
> > > - return 0;
> > > + ret = atmel_hlcdc_layer_update_start(&plane->layer);
> > > + if (!ret)
> > > + plane->prepared = true;
> >
> > Atomic helpers will keep track of this for you, and only call ->cleanup_fb
> > on a plane combo where it did (successfully) call ->prepare_fb.
>
> Hm, it's not exactly encoding the same thing. What I want to do here is
> call atmel_hlcdc_layer_update_rollback() only if the atomic update has
> failed (see below, I set ->prepared back to false in the
> ->atomic_update() method). AFAICT, ->cleanup_fb() is also called
> when the whole operation succeed (and in this case I don't want to
> call atmel_hlcdc_layer_update_rollback()).
>
> Let me know if you see a better approach to do that.

Ah makes sense. And yeah I guess this is the only approach really. Note
that you should put plane_prepared into plane_state though, in case we
start to pipeline updates.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch