Re: [PATCH v9 04/16] drm/imx: dc-kms: Use driver specific .atomic_commit_tail callback

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] drm/imx: dc: plane disable skipped on modeset in dc_drm_atomic_commit_tail()
--

commit ff022f683849ca82589cd00028c8fb7c412423dc
Author: Liu Ying <victor.liu@nxp.com>

drm/imx: dc-kms: Use driver specific .atomic_commit_tail callback

This commit introduces a custom atomic_commit_tail callback for the i.MX8qxp
Display Controller KMS driver that passes DRM_PLANE_COMMIT_ACTIVE_ONLY and
DRM_PLANE_COMMIT_NO_DISABLE_AFTER_MODESET to drm_atomic_helper_commit_planes()
to avoid accessing plane registers while clocks are disabled.

> diff --git a/drivers/gpu/drm/imx/dc/dc-kms.c b/drivers/gpu/drm/imx/dc/dc-kms.c
> index 0e2ce6a425487..9ffa15ac43a25 100644
> --- a/drivers/gpu/drm/imx/dc/dc-kms.c
> +++ b/drivers/gpu/drm/imx/dc/dc-kms.c
[ ... ]
> @@ -30,6 +31,29 @@ static const struct drm_mode_config_funcs dc_drm_mode_config_funcs = {
>  	.atomic_commit = drm_atomic_helper_commit,
>  };
>  
> +static void dc_drm_atomic_commit_tail(struct drm_atomic_commit *state)
> +{
> +	struct drm_device *dev = state->dev;
> +
> +	drm_atomic_helper_commit_modeset_disables(dev, state);
> +
> +	drm_atomic_helper_commit_planes(dev, state,
> +					DRM_PLANE_COMMIT_ACTIVE_ONLY |
> +					DRM_PLANE_COMMIT_NO_DISABLE_AFTER_MODESET);

[Severity: High]
Could passing DRM_PLANE_COMMIT_NO_DISABLE_AFTER_MODESET (and
DRM_PLANE_COMMIT_ACTIVE_ONLY) here leave the plane's FetchUnit source buffer
and LayerBlend enabled in hardware after the old framebuffer is cleaned up?

In drm_atomic_helper_commit_planes(), DRM_PLANE_COMMIT_NO_DISABLE_AFTER_MODESET
skips calling funcs->atomic_disable() when the CRTC needs a modeset:

    if (disabling && funcs->atomic_disable) {
        struct drm_crtc_state *crtc_state;

        crtc_state = old_plane_state->crtc->state;

        if (drm_atomic_crtc_needs_modeset(crtc_state) &&
            no_disable)
            continue;

        funcs->atomic_disable(plane, state);
    }

That flag is meant for drivers where the CRTC atomic_disable callback also
disables the CRTC's planes. However, in drivers/gpu/drm/imx/dc/dc-crtc.c,
dc_crtc_atomic_disable() only disables FrameGen and calls pm_runtime_put() on
the Display Engine and the shared Pixel Engine without disabling the planes:

    enable_irq(dc_crtc->irq_dec_seqcomplete);
    dc_fg_disable(dc_crtc->fg);
    DC_CRTC_WAIT_FOR_COMPLETION_TIMEOUT(dec_seqcomplete_done);
    disable_irq(dc_crtc->irq_dec_seqcomplete);

    dc_fg_disable_clock(dc_crtc->fg);

    /* request pixel engine power-off as plane is off too */
    pm_runtime_put(dc_drm->pe->dev);

And in drivers/gpu/drm/imx/dc/dc-plane.c, __dc_plane_atomic_disable() is what
actually disables the FetchUnit source buffer and LayerBlend:

    /* disable fetchunit in shadow */
    fu_ops = dc_fu_get_ops(dplane->fu);
    fu_ops->disable_src_buf(dplane->fu, DC_FETCHUNIT_FRAC0);

    /* set ExtDst's source to ConstFrame */
    dc_ed_pec_src_sel(dplane->ed, dc_cf_get_link_id(dplane->cf));

When dc_crtc_atomic_enable() runs with new_crtc_state->plane_mask == 0, it
switches ed_cont's source selection to cf_cont, but it does not call
fu_ops->disable_src_buf():

    /* show CRTC background if no plane is enabled */
    if (new_crtc_state->plane_mask == 0) {
        /* constframe in content stream shows black frame */
        dc_cf_constantcolor_black(dc_crtc->cf_cont);

        cf_link = dc_cf_get_link_id(dc_crtc->cf_cont);
        dc_ed_pec_src_sel(dc_crtc->ed_cont, cf_link);
    }

If an active CRTC with an enabled primary plane undergoes a modeset while
remaining active and disables its plane in the same commit,
dc_crtc_atomic_disable() calls asynchronous pm_runtime_put(dc_drm->pe->dev)
and dc_crtc_atomic_begin() immediately calls
pm_runtime_resume_and_get(dc_drm->pe->dev) before runtime suspend runs (or the
second CRTC on the shared Pixel Engine keeps dc_drm->pe->dev active), so
dc_pe_runtime_resume() does not reset the Pixel Engine hardware. Because
drm_atomic_crtc_needs_modeset() is true, DRM_PLANE_COMMIT_NO_DISABLE_AFTER_MODESET
skips dc_plane_atomic_disable(), leaving SOURCEBUFFERENABLE and LayerBlend
enabled after drm_atomic_helper_cleanup_planes() cleans up the framebuffer.

Similarly, because drm_atomic_helper_swap_state() swaps crtc->state before
dc_drm_atomic_commit_tail() runs, plane_crtc_active(old_plane_state) evaluates
to false whenever a CRTC is being disabled (new_crtc_state->active == false).
With DRM_PLANE_COMMIT_ACTIVE_ONLY, disabling a CRTC and its plane while the
second CRTC keeps the shared dc_drm->pe->dev powered on, followed by
re-enabling the first CRTC with plane_mask == 0, also skips
dc_plane_atomic_disable() without ever calling fu_ops->disable_src_buf().

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261103-imx8-dc-prefetch-v9-0-3b985e7636e2@nxp.com?part=4