Re: [PATCH 1/9] drm/arcpgu: replace struct drm_simple_display_pipe with regular atomic helpers

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

New issues:
- [High] NULL pointer dereference when enabling CRTC without a primary plane framebuffer.

Pre-existing issues:
- [High] Unsafe direct access of global plane and CRTC state pointers in atomic commit callbacks.
--

commit 6344532446cab68117e44660f110ef0c8aec15d7
Author: Ze Huang <ze.huang@oss.qualcomm.com>

drm/arcpgu: replace struct drm_simple_display_pipe with regular atomic helpers

This commit replaces drm_simple_display_pipe with regular atomic helpers.
It instantiates the plane, CRTC, and encoder directly, wiring them up with
standard atomic helpers to remove the dependency on the simple-KMS interface.

> diff --git a/drivers/gpu/drm/tiny/arcpgu.c b/drivers/gpu/drm/tiny/arcpgu.c
> index c93d61ac0bb75..375cdb79e4e89 100644
> --- a/drivers/gpu/drm/tiny/arcpgu.c
> +++ b/drivers/gpu/drm/tiny/arcpgu.c
[ ... ]
> @@ -117,7 +117,7 @@ static const u32 arc_pgu_supported_formats[] = {
>  
>  static void arc_pgu_set_pxl_fmt(struct arcpgu_drm_private *arcpgu)
>  {
> -	const struct drm_framebuffer *fb = arcpgu->pipe.plane.state->fb;
> +	const struct drm_framebuffer *fb = arcpgu->plane.state->fb;

[Severity: High]
This isn't a bug introduced by this patch, but does reading the global 
plane->state pointer here instead of using the state passed via the 
drm_atomic_commit parameter cause issues during non-blocking atomic commits?

If multiple non-blocking commits are queued rapidly, the global pointers 
advance to the newest state. When the first commit's commit_tail finally 
executes, reading plane->state directly instead of extracting the specific 
commit's state could cause the older commit to program the hardware using a 
future commit's state.

[Severity: High]
Can this dereference a NULL pointer if the CRTC is enabled without a primary 
plane framebuffer?

The transition from drm_simple_display_pipe to generic atomic helpers removed 
the underlying protection that previously prevented the CRTC from enabling 
without a valid framebuffer. If userspace performs an atomic commit that 
enables the CRTC but sets the primary plane's framebuffer to 0 (NULL), 
drm_atomic_helper_check_crtc_primary_plane verifies the plane is attached but 
does not enforce plane_state->fb != NULL.

This would allow the commit to proceed to arcpgu_crtc_helper_atomic_enable(), 
which calls arc_pgu_mode_set(), and then arc_pgu_set_pxl_fmt(), blindly 
dereferencing fb->format->format here.

>  	uint32_t pixel_format = fb->format->format;
>  	u32 format = DRM_FORMAT_INVALID;
>  	int i;
[ ... ]
> @@ -155,7 +155,7 @@ static enum drm_mode_status arc_pgu_mode_valid(struct drm_simple_display_pipe *p
>  
>  static void arc_pgu_mode_set(struct arcpgu_drm_private *arcpgu)
>  {
> -	struct drm_display_mode *m = &arcpgu->pipe.crtc.state->adjusted_mode;
> +	struct drm_display_mode *m = &arcpgu->crtc.state->adjusted_mode;

[Severity: High]
This is a pre-existing issue, but could reading the global crtc->state pointer 
directly cause hardware state corruption or frame tearing?

Similar to the plane state issue above, if rapid non-blocking atomic commits 
are generated, arc_pgu_mode_set() might read from the newest global state 
rather than the specific commit's state object.

>  	u32 val;
>  
>  	arc_pgu_write(arcpgu, ARCPGU_REG_FMT,
[ ... ]
> @@ -218,35 +218,106 @@ static void arc_pgu_disable(struct drm_simple_display_pipe *pipe)
>  			      ~ARCPGU_CTRL_ENABLE_MASK);
>  }
>  
> -static void arc_pgu_update(struct drm_simple_display_pipe *pipe,
> -			   struct drm_plane_state *state)
> +static void arcpgu_plane_helper_atomic_update(struct drm_plane *plane,
> +					      struct drm_atomic_commit *state)
>  {
>  	struct arcpgu_drm_private *arcpgu;
>  	struct drm_gem_dma_object *gem;
>  
> -	if (!pipe->plane.state->fb)
> +	if (!plane->state->fb)

[Severity: High]
This is a pre-existing issue, but does arcpgu_plane_helper_atomic_update() 
safely handle non-blocking atomic commits when reading plane->state->fb 
directly?

Reading the global plane->state pointer instead of extracting the new plane 
state from the drm_atomic_commit parameter could result in the same race 
condition during asynchronous commit_tail execution.

>  		return;
>  
> -	arcpgu = pipe_to_arcpgu_priv(pipe);
> -	gem = drm_fb_dma_get_gem_obj(pipe->plane.state->fb, 0);
> +	arcpgu = dev_to_arcpgu(plane->dev);
> +	gem = drm_fb_dma_get_gem_obj(plane->state->fb, 0);
>  	arc_pgu_write(arcpgu, ARCPGU_REG_BUF0_ADDR, gem->dma_addr);
>  }

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260705-drm-simple-kms-removal-v1-0-b4e1ca053623@oss.qualcomm.com?part=1