Re: [PATCH 1/3] ipuv3-crtc: add remove action for devres data

[Daniel Vetter]

On Tue, Apr 02, 2019 at 05:49:23PM +0200, Philipp Zabel wrote:
> Hi Michael,
> 
> On Tue, 2019-04-02 at 15:49 +0200, Michael Grzeschik wrote:
> > The destroy function in drm_mode_config_cleanup will remove the objects
> > in ipu-drm-core by calling its destroy functions if the bind function
> > fails. The drm_crtc is also part of the devres allocated ipu_crtc
> > object. The ipu_crtc object will already be cleaned up if the bind for
> > the crtc fails. This leads drm_crtc_cleanup try to clean already freed
> > memory.
> > 
> > We fix this issue by adding the devres action ipu_crtc_remove_head which
> > will remove its head from the objects in ipu-drm-core which then never
> > calls its destroy function anymore.
> > 
> > Signed-off-by: Michael Grzeschik <m.grzeschik@xxxxxxxxxxxxxx>
> > ---
> >  drivers/gpu/drm/imx/ipuv3-crtc.c | 12 ++++++++++++
> >  1 file changed, 12 insertions(+)
> > 
> > diff --git a/drivers/gpu/drm/imx/ipuv3-crtc.c b/drivers/gpu/drm/imx/ipuv3-crtc.c
> > index ec3602ebbc1cd..fa1ee33a43d77 100644
> > --- a/drivers/gpu/drm/imx/ipuv3-crtc.c
> > +++ b/drivers/gpu/drm/imx/ipuv3-crtc.c
> > @@ -429,6 +429,14 @@ static int ipu_crtc_init(struct ipu_crtc *ipu_crtc,
> >  	return ret;
> >  }
> >  
> > +static void ipu_crtc_remove_head(void *data)
> > +{
> > +	struct ipu_crtc *ipu_crtc = data;
> > +	struct drm_crtc *crtc = &ipu_crtc->base;
> > +
> > +	list_del(&crtc->head);
> 
> I don't think reaching into drm_crtc internals like this is going to be
> robust. Currently, this is either missing the rest of drm_crtc_cleanup,
> or it will crash if drm_crtc_init_with_planes hasn't been called yet.
> 
> I think you could call devm_add_action with a function that calls
> drm_crtc_cleanup after drm_crtc_init_with_planes in ipu_crtc_init.
> 
> Alternatively, the ipu_crtc allocation could be changed to kzalloc. It
> would then have to freed manually in the drm_crtc_funcs->destroy
> callback.

Dirty secret of devm: You can't use it for any drm_ structure, because the
lifetimes of those do not match the lifetimes of the underlying device.
We'd need to tie the lifetimes of drm objects to drm_device. Noralf has
some patches to move that forward. We'd need something like
drm_dev_kzalloc which releases memory when drm_device is freed.

Agreed that digging even deeper into the devm deadend is not a good idea.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch