> -----Original Message-----
> From: Inki Dae <[email protected]>
> Sent: Monday, November 10, 2025 2:22 PM
> To: Hoyoung Lee <[email protected]>
> Cc: Seung-Woo Kim <[email protected]>; Kyungmin Park
> <[email protected]>; David Airlie <[email protected]>; Simona
> Vetter <[email protected]>; Krzysztof Kozlowski <[email protected]>; Alim
> Akhtar <[email protected]>; [email protected]; linux-
> [email protected]; [email protected]; linux-
> [email protected]
> Subject: Re: [PATCH 2/3] drm/exynos: Convert to drmm_mode_config_init()
> and drop manual cleanup
> 
> 2025년 9월 29일 (월) 오후 1:54, Hoyoung Lee <[email protected]>님이 작
> 성:
> >
> > Switch mode-config initialization to drmm_mode_config_init() so that
> > the lifetime is tied to drm_device. Remove explicit
> > drm_mode_config_cleanup() from error and unbind paths since cleanup is
> now managed by DRM.
> >
> > No functional change intended.
> >
> > Signed-off-by: Hoyoung Lee <[email protected]>
> > ---
> >  drivers/gpu/drm/exynos/exynos_drm_drv.c | 4 +---
> >  1 file changed, 1 insertion(+), 3 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/exynos/exynos_drm_drv.c
> > b/drivers/gpu/drm/exynos/exynos_drm_drv.c
> > index 6cc7bf77bcac..1aea71778ab1 100644
> > --- a/drivers/gpu/drm/exynos/exynos_drm_drv.c
> > +++ b/drivers/gpu/drm/exynos/exynos_drm_drv.c
> > @@ -257,7 +257,7 @@ static int exynos_drm_bind(struct device *dev)
> >         dev_set_drvdata(dev, drm);
> >         drm->dev_private = (void *)private;
> >
> > -       drm_mode_config_init(drm);
> > +       drmm_mode_config_init(drm);
> >
> >         exynos_drm_mode_config_init(drm);
> >
> > @@ -297,7 +297,6 @@ static int exynos_drm_bind(struct device *dev)
> >  err_unbind_all:
> >         component_unbind_all(drm->dev, drm);
> >  err_mode_config_cleanup:
> > -       drm_mode_config_cleanup(drm);
> 
> In the current implementation, there is a potential dereference issue
> because the private object may be freed before to_dma_dev(dev) is called.
> When drmm_mode_config_init() is invoked, it registers
> drm_mode_config_cleanup() as a managed action. This means that the cleanup
> function will be automatically executed later when
> drm_dev_put() is called.
> 
> The problem arises when drm_dev_put() is called without explicitly
> invoking drm_mode_config_cleanup() first, as in the original code. In that
> case, the managed cleanup is performed later, which allows
> to_dma_dev(dev) to be called after the private object has already been
> released.
> 
> For reference, the following sequence may occur internally when
> drm_mode_config_cleanup() is executed:
> 1. drm_mode_config_cleanup() is called.
> 2. During the cleanup of FBs, planes, CRTCs, encoders, and connectors,
> framebuffers or GEM objects may be released.
> 3. At this point, Exynos-specific code could invoke to_dma_dev(dev).
> 
> Therefore, the private object must remain valid until
> drm_mode_config_cleanup() completes.
> It would be safer to adjust the code so that kfree(private) is performed
> after drm_dev_put(drm) to ensure the private data remains available during
> cleanup.
> 
> Thanks,
> Inki Dae
> 
> >         exynos_drm_cleanup_dma(drm);
> >         kfree(private);
> >         dev_set_drvdata(dev, NULL);
> > @@ -317,7 +316,6 @@ static void exynos_drm_unbind(struct device *dev)
> >         drm_atomic_helper_shutdown(drm);
> >
> >         component_unbind_all(drm->dev, drm);
> > -       drm_mode_config_cleanup(drm);
> 
> Ditto.
> 
> >         exynos_drm_cleanup_dma(drm);
> >
> >         kfree(drm->dev_private);
> > --
> > 2.34.1
> >
> >

Hi, Inki
Thanks for the review and for pointing out the to_dma_dev() path

If I understand you correctly, fine with using DRMM, but kfree(priv) should 
occur after drm_dev_put(drm)
That would mean releasing the drm_device first and freeing dev_private 
afterwards.
Of course, we will also need to adjust the probe() error-unwind (err_free) 
order accordingly.
Do you anticipate any side effects from this ordering change? I’d appreciate 
your thoughts.

BRs,
Hoyoung Lee


Reply via email to