Re: [PATCH 2/2] drm/vkms: Use drm_dev_unplug() for ConfigFS device removal

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [High] drm/vkms: broken CRC debugfs unplug handling in vkms_destroy()
--

commit 58c6203197ea5aa86a82e01783503c18d44dd2a1
Author: Sebastian Wick <sebastian.wick@redhat.com>

drm/vkms: Use drm_dev_unplug() for ConfigFS device removal

This commit switches vkms_destroy() from drm_dev_unregister() to
drm_dev_unplug() so that ioctls on open file descriptors are rejected with
-ENODEV and in-flight operations guarded by drm_dev_enter()/drm_dev_exit()
complete before ConfigFS device removal proceeds.

> diff --git a/drivers/gpu/drm/vkms/vkms_drv.c b/drivers/gpu/drm/vkms/vkms_drv.c
> index 5a640b531d88f..716f9c966c211 100644
> --- a/drivers/gpu/drm/vkms/vkms_drv.c
> +++ b/drivers/gpu/drm/vkms/vkms_drv.c
> @@ -259,7 +259,7 @@ void vkms_destroy(struct vkms_config *config)
>  
>  	fdev = config->dev->faux_dev;
>  
> -	drm_dev_unregister(&config->dev->drm);
> +	drm_dev_unplug(&config->dev->drm);
>  	drm_atomic_helper_shutdown(&config->dev->drm);
>  	devres_release_group(&fdev->dev, NULL);

[Severity: High]
How does this interact with an open CRTC CRC debugfs file descriptor
(/sys/kernel/debug/dri/<N>/crtc-<M>/crc/data) when the VKMS device is
disabled via ConfigFS?

In drivers/gpu/drm/vkms/vkms_composer.c, vkms_set_crc_source() guards the
callback with drm_dev_enter():

drivers/gpu/drm/vkms/vkms_composer.c:vkms_set_crc_source() {
	...
	if (!drm_dev_enter(crtc->dev, &idx))
		return -ENODEV;

	ret = vkms_crc_parse_source(src_name, &enabled);

	vkms_set_composer(out, enabled);

	drm_dev_exit(idx);
	return ret;
}

When the crc/data file descriptor is closed, full_proxy_release() invokes
crtc_crc_release() in drivers/gpu/drm/drm_debugfs_crc.c, which calls
crtc->funcs->set_crc_source(crtc, NULL).

Two issues seem possible here:

1. Unlike /dev/dri/card<N> file descriptors, opening crc/data via
crtc_crc_open() does not acquire a drm_dev_get() reference on crtc->dev,
and debugfs_remove_recursive() during drm_dev_unregister() does not wait
for open file descriptors to close.

In vkms_destroy(), drm_dev_unplug() runs synchronize_srcu() before
devres_release_group(&fdev->dev, NULL) drops the devm reference via
devm_drm_dev_init_release() -> drm_dev_put(). If no /dev/dri/card<N> fd is
open (or after it is closed first, such as during process exit when card<N>
has a lower fd number than crc/data), drm_dev_release() frees struct
vkms_output (embedding struct drm_crtc) and struct vkms_device (embedding
struct drm_device). When the crc/data fd is subsequently closed, wouldn't
crtc_crc_release() and drm_dev_enter(crtc->dev, &idx) in
vkms_set_crc_source() dereference the freed struct drm_crtc and
crtc->dev?

2. Conversely, if struct drm_device is still kept alive by an open
/dev/dri/card<N> fd when crc/data is closed after drm_dev_unplug() sets
dev->unplugged = true, drm_dev_enter(crtc->dev, &idx) in
vkms_set_crc_source(crtc, NULL) will return false and exit early with
-ENODEV. Doesn't that skip vkms_set_composer(out, false), leaving
out->composer_enabled set to true and leaking the vblank reference from
drm_crtc_vblank_put(&out->crtc)?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261031225901.3200167-1-sebastian.wick@redhat.com?part=2