Skip to contentSkip to conversation

media: v4l2-ioctl: zero the ext control built for VIDIOC_{G,S}_CTRL

PATCH Superseded Nick Rogers 3 participants 4 messages Sep 23 – Sep 25 Media (V4L/DVB)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 media: v4l2-ioctl: zero the ext control built for VIDIOC_{G,S}_CTRL

When a driver implements the extended control ioctls but has no control handler, v4l_g_ctrl() and v4l_s_ctrl() pass VIDIOC_G_CTRL and VIDIOC_S_CTRL on as a single struct v4l2_ext_control built on the stack. Only its id and value are set, and check_ext_ctrls() clears reserved[0] and reserved2[0]; the control's size and the rest of both structures are left uninitialized.

A driver that forwards the controls rather than handling them through the control framework sees that stack garbage. The virtio-media driver under review takes a nonzero size as a payload to copy from userspace, so VIDIOC_G_CTRL and VIDIOC_S_CTRL fail with -EINVAL through it whenever the stack is dirty. GStreamer's V4L2 encoders set their profile with VIDIOC_S_CTRL, and cannot negotiate against such a device.

Zero-initialize both structures.

Assisted-by Claude:claude-opus-5-5

---
Found running the virtio-media v9 series [1] in a VMM with a host-side stateful encoder: GStreamer's v4l2h264enc fails to negotiate because VIDIOC_S_CTRL returns -EINVAL. Tested on 6.18 with that series applied:
VIDIOC_G_CTRL and VIDIOC_S_CTRL now reach the device intact, and v4l2-compliance 1.30.1 reports the same results with and without this patch. Build-tested on media.git next (arm64, W=1, no new warnings).

[1] https://lore.kernel.org/all/[email protected]/

 drivers/media/v4l2-core/v4l2-ioctl.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/media/v4l2-core/v4l2-ioctl.c +4 -4
index 17ba1ae70..b7d248ab7 100644
@@ -2357,8 +2357,8 @@ static int v4l_g_ctrl(const struct v4l2_ioctl_ops *ops, struct file *file,
 	struct video_device *vfd = video_devdata(file);
 	struct v4l2_control *p = arg;
 	struct v4l2_fh *vfh = file_to_v4l2_fh(file);
-	struct v4l2_ext_controls ctrls;
-	struct v4l2_ext_control ctrl;
+	struct v4l2_ext_controls ctrls = {};
+	struct v4l2_ext_control ctrl = {};
 
 	if (vfh && vfh->ctrl_handler)
 		return v4l2_g_ctrl(vfh->ctrl_handler, p);
@@ -2388,8 +2388,8 @@ static int v4l_s_ctrl(const struct v4l2_ioctl_ops *ops, struct file *file,
 	struct video_device *vfd = video_devdata(file);
 	struct v4l2_control *p = arg;
 	struct v4l2_fh *vfh = file_to_v4l2_fh(file);
-	struct v4l2_ext_controls ctrls;
-	struct v4l2_ext_control ctrl;
+	struct v4l2_ext_controls ctrls = {};
+	struct v4l2_ext_control ctrl = {};
 	int ret;
 
 	if (vfh && vfh->ctrl_handler)
Signature · 1 line
2.54.0 (Apple Git-157)
Reviewed-by on v1 1/1
#1 lore raw

Hi,

Le mercredi 23 septembre 2026 à 17:09 +0100, Nick Rogers a écrit :

Quoted 14 lines

When a driver implements the extended control ioctls but has no control handler, v4l_g_ctrl() and v4l_s_ctrl() pass VIDIOC_G_CTRL and VIDIOC_S_CTRL on as a single struct v4l2_ext_control built on the stack. Only its id and value are set, and check_ext_ctrls() clears reserved[0] and reserved2[0]; the control's size and the rest of both structures are left uninitialized.

A driver that forwards the controls rather than handling them through the control framework sees that stack garbage. The virtio-media driver under review takes a nonzero size as a payload to copy from userspace, so VIDIOC_G_CTRL and VIDIOC_S_CTRL fail with -EINVAL through it whenever the stack is dirty. GStreamer's V4L2 encoders set their profile with VIDIOC_S_CTRL, and cannot negotiate against such a device.

Quoted 4 lines

Zero-initialize both structures.

Assisted-by Claude:claude-opus-5-5

One may argue that it could have been fixed in the unique driver not using the ctrl framework, but I personally like the simplicity of the fix.

Quoted 39 lines

---
Found running the virtio-media v9 series [1] in a VMM with a host-side stateful encoder: GStreamer's v4l2h264enc fails to negotiate because VIDIOC_S_CTRL returns -EINVAL. Tested on 6.18 with that series applied:
VIDIOC_G_CTRL and VIDIOC_S_CTRL now reach the device intact, and v4l2-compliance 1.30.1 reports the same results with and without this patch. Build-tested on media.git next (arm64, W=1, no new warnings).

[1] https://lore.kernel.org/all/[email protected]/

 drivers/media/v4l2-core/v4l2-ioctl.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
drivers/media/v4l2-core/v4l2-ioctl.c +4 -4
index 17ba1ae70..b7d248ab7 100644
@@ -2357,8 +2357,8 @@ static int v4l_g_ctrl(const struct v4l2_ioctl_ops *ops, struct file *file,
 	struct video_device *vfd = video_devdata(file);
 	struct v4l2_control *p = arg;
 	struct v4l2_fh *vfh = file_to_v4l2_fh(file);
-	struct v4l2_ext_controls ctrls;
-	struct v4l2_ext_control ctrl;
+	struct v4l2_ext_controls ctrls = {};
+	struct v4l2_ext_control ctrl = {};
 
 	if (vfh && vfh->ctrl_handler)
 		return v4l2_g_ctrl(vfh->ctrl_handler, p);
@@ -2388,8 +2388,8 @@ static int v4l_s_ctrl(const struct v4l2_ioctl_ops *ops, struct file *file,
 	struct video_device *vfd = video_devdata(file);
 	struct v4l2_control *p = arg;
 	struct v4l2_fh *vfh = file_to_v4l2_fh(file);
-	struct v4l2_ext_controls ctrls;
-	struct v4l2_ext_control ctrl;
+	struct v4l2_ext_controls ctrls = {};
+	struct v4l2_ext_control ctrl = {};
 	int ret;
 
 	if (vfh && vfh->ctrl_handler)
#2 lore raw
In reply to Nick Rogers
Quoted 26 lines · Nick Rogers

When a driver implements the extended control ioctls but has no control handler, v4l_g_ctrl() and v4l_s_ctrl() pass VIDIOC_G_CTRL and VIDIOC_S_CTRL on as a single struct v4l2_ext_control built on the stack. Only its id and value are set, and check_ext_ctrls() clears reserved[0] and reserved2[0]; the control's size and the rest of both structures are left uninitialized.

A driver that forwards the controls rather than handling them through the control framework sees that stack garbage. The virtio-media driver under review takes a nonzero size as a payload to copy from userspace, so VIDIOC_G_CTRL and VIDIOC_S_CTRL fail with -EINVAL through it whenever the stack is dirty. GStreamer's V4L2 encoders set their profile with VIDIOC_S_CTRL, and cannot negotiate against such a device.

Zero-initialize both structures.

Assisted-by Claude:claude-opus-5-5

---
Found running the virtio-media v9 series [1] in a VMM with a host-side stateful encoder: GStreamer's v4l2h264enc fails to negotiate because VIDIOC_S_CTRL returns -EINVAL. Tested on 6.18 with that series applied:
VIDIOC_G_CTRL and VIDIOC_S_CTRL now reach the device intact, and v4l2-compliance 1.30.1 reports the same results with and without this patch. Build-tested on media.git next (arm64, W=1, no new warnings).

Quoted 1 line

[1] https://lore.kernel.org/all/[email protected]/

I'm wondering what the API contract intent is here. If it is that drivers are supposed to fill the control structures themselves, then virtio-media should also be fixed, regardless of this safeguard.

Indeed it is likely that virtio-media will be ported to some older kernels, which may not have this patch, in which case the same bug will arise again.

#3 lore raw
In reply to Nick Rogers
Quoted 14 lines · Nick Rogers

When a driver implements the extended control ioctls but has no control handler, v4l_g_ctrl() and v4l_s_ctrl() pass VIDIOC_G_CTRL and VIDIOC_S_CTRL on as a single struct v4l2_ext_control built on the stack. Only its id and value are set, and check_ext_ctrls() clears reserved[0] and reserved2[0]; the control's size and the rest of both structures are left uninitialized.

A driver that forwards the controls rather than handling them through the control framework sees that stack garbage. The virtio-media driver under review takes a nonzero size as a payload to copy from userspace, so VIDIOC_G_CTRL and VIDIOC_S_CTRL fail with -EINVAL through it whenever the stack is dirty. GStreamer's V4L2 encoders set their profile with VIDIOC_S_CTRL, and cannot negotiate against such a device.

Quoted 3 lines

Zero-initialize both structures.

Assisted-by Claude:claude-opus-5-5

nit: per the kernel guidelines [1], this should be

Assisted-by LLM

[1] https://docs.kernel.org/process/coding-assistants.html#attribution