Re: [PATCH v8 05/14] drm: add DRM_CLIENT_CAP_LUMINANCE

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:
- [High] TOCTOU race between `drm_backlight_link()` and device-wide backlight inhibit/uninhibit iterations causes permanent count corruption.
- [High] The LUMINANCE property is unconditionally exposed to all atomic clients, violating the DRM_CLIENT_CAP_LUMINANCE capability contract.
- [Low] Clients cannot cleanly unset DRM_CLIENT_CAP_LUMINANCE if they have previously unset DRM_CLIENT_CAP_ATOMIC.
--

commit 3e9d95a8d4fc1bbb26ab6124a33c012ee7b1f848
Author: Mario Limonciello <mario.limonciello@amd.com>

drm: add DRM_CLIENT_CAP_LUMINANCE

This patch adds a DRIVER_CONNECTOR_LUMINANCE driver feature and a
DRM_CLIENT_CAP_LUMINANCE client capability. When a client sets this
capability, every DRM-connected backlight is marked as taken over, and
legacy sysfs writes return -EBUSY.

> diff --git a/drivers/gpu/drm/drm_ioctl.c b/drivers/gpu/drm/drm_ioctl.c
> index 9039a39c43243..de2c9d9ab5285 100644
> --- a/drivers/gpu/drm/drm_ioctl.c
> +++ b/drivers/gpu/drm/drm_ioctl.c

[ ... ]

> @@ -392,6 +394,28 @@ drm_setclientcap(struct drm_device *dev, void *data, struct drm_file *file_priv)
>  		file_priv->plane_color_pipeline = req->value;
>  		break;
>  	}
> +	case DRM_CLIENT_CAP_LUMINANCE:
> +		if (!drm_core_check_feature(dev, DRIVER_CONNECTOR_LUMINANCE))
> +			return -EOPNOTSUPP;
> +		if (!file_priv->atomic)
> +			return -EINVAL;

[Severity: Low]
Can this prevent clients from cleanly unsetting DRM_CLIENT_CAP_LUMINANCE if
they have previously unset DRM_CLIENT_CAP_ATOMIC?

If a client sets ATOMIC=1, LUMINANCE=1, ATOMIC=0, and then tries to unset
LUMINANCE=0, this unconditional atomic check will cause the last call to
return -EINVAL.

> +		if (req->value > 1)
> +			return -EINVAL;
> +
> +		/*
> +		 * Serialize the compare-and-act so concurrent ioctls on a shared
> +		 * fd cannot unbalance the device-wide inhibit count.
> +		 */
> +		mutex_lock(&dev->mode_config.mutex);
> +		if (req->value != file_priv->supports_luminance_control) {
> +			if (req->value)
> +				drm_backlight_inhibit_legacy_all(dev);

[Severity: High]
Does this call introduce a TOCTOU race between drm_backlight_link() and the
device-wide backlight inhibit iteration?

Looking at drm_backlight_inhibit_legacy_all() in
drivers/gpu/drm/drm_backlight.c, it drops drm_backlight_lock before
iterating connectors:

	scoped_guard(spinlock, &drm_backlight_lock)
		dev->mode_config.luminance_clients++;

	drm_connector_list_iter_begin(dev, &iter);

If a concurrent hotplug/probe thread calls drm_backlight_link(), it reads
the already-incremented dev->mode_config.luminance_clients and assigns it to
the new connector's b->luminance_clients.

Then the ioctl thread resumes, iterates the connector list, reaches this
newly linked connector, and calls drm_backlight_inhibit_legacy(), which
increments b->luminance_clients again. This causes the connector's count to
get permanently stuck at a non-zero value, locking out legacy sysfs backlight
control for the rest of the system uptime.

> +			else
> +				drm_backlight_uninhibit_legacy_all(dev);
> +			file_priv->supports_luminance_control = req->value;

[Severity: High]
Is the LUMINANCE property being unconditionally exposed to all atomic clients,
violating the DRM_CLIENT_CAP_LUMINANCE capability contract?

The UAPI behavior documented in include/uapi/drm/drm.h states: "The LUMINANCE
property becomes visible on supported connectors".

However, the capability tracked here (file_priv->supports_luminance_control)
is never utilized by the DRM core (e.g., in drm_mode_object_get_properties)
to filter property visibility. drm_mode_object_get_properties() only checks
if the client is atomic, but fails to check if the property is luminance and
if the client supports it:

	if ((prop->flags & DRM_MODE_PROP_ATOMIC) && !atomic)
		continue;

As a result, any generic atomic client (like a compositor restoring state)
will see and restore the LUMINANCE property without setting
DRM_CLIENT_CAP_LUMINANCE, bypassing the sysfs inhibition.

> +		}
> +		mutex_unlock(&dev->mode_config.mutex);
> +		break;
>  	default:
>  		return -EINVAL;
>  	}

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260908044035.62093-1-mario.limonciello@amd.com?part=5