Re: [PATCH] media: ipu-bridge: hold a reference on the IVSC software node

[Andy Shevchenko <andriy.shevchenko@linux.intel.com>]

+Cc: Bart for the double check of the approach. This sounds like it might need
to be fixed in driver core (software node implementation).

On Sun, Oct 11, 2026 at 03:15:18PM +0200, Eduardo Ruiz Duarte wrote:
> ipu_bridge_instantiate_ivsc() attaches the IVSC software node to the mei-csi
> device with set_secondary_fwnode(), but never takes a reference on it, while
> the driver core unconditionally drops one when that device goes away.
> 
> device_add() calls software_node_notify(), which does a kobject_get() on the
> swnode behind dev_fwnode(dev). device_del() calls
> software_node_notify_remove(), which does the matching kobject_put() -- and
> that put sits *outside* the "if (swnode->managed)" block, so it runs for every
> device whose fwnode is a software node, managed or not:
> 
> 	sysfs_remove_link(&swnode->kobj, dev_name(dev));
> 	sysfs_remove_link(&dev->kobj, "software_node");
> 	kobject_put(&swnode->kobj);
> 
> 	if (swnode->managed) {
> 		set_secondary_fwnode(dev, NULL);
> 		kobject_put(&swnode->kobj);
> 	}
> 
> csi_dev is the mei-csi device, which the MEI client bus has already added by
> the time ipu_bridge_instantiate_ivsc() runs. The add-side get therefore never
> happens for this node, but the del-side put always does.
> 
> On hardware where the MEI bus resets -- which it does on every s2idle and S4
> resume on a Dell XPS 13 9340 -- mei-csi is destroyed and recreated each cycle,
> so the node loses one reference per resume. INTC10CF-N starts at 3: one for its
> own registration plus one for each of its two port children, since
> kobject_add() of a child takes a reference on the parent. The third resume
> therefore frees it, software_node_find_by_name() starts returning NULL, and
> isys walks the fwnode graph into freed memory:
> 
>   refcount_t: addition on 0; use-after-free
>   WARNING: lib/refcount.c:25 at refcount_warn_saturate+0x7e/0xb0
>    kobject_get+0x94/0xb0
>    software_node_graph_get_port_parent+0x89/0xd0
>    fwnode_graph_get_port_parent+0x70/0xd0
>    fwnode_graph_get_endpoint_by_id+0x93/0x280
>    isys_notifier_init+0x92/0x180 [intel_ipu6_isys]
>    isys_register_devices+0x3c9/0x4e0 [intel_ipu6_isys]
>    isys_probe+0x3df/0x500 [intel_ipu6_isys]
>    auxiliary_bus_probe+0x4c/0x70
>    ...
>    deferred_probe_work_func+0xa5/0x110
> 
> followed by an underflow and a saturate on the same kobject, and the camera
> never comes back.
> 
> Take a reference so the attachment owns one, balancing the put the core will do
> when csi_dev is destroyed. Where csi_dev is never destroyed this leaks a single
> kobject reference, which is what the lifetime of the node already assumes.
> 
> Found while making the camera survive hibernation on a Dell XPS 13 9340
> (ov02c10 behind an INTC10CF IVSC); with suspend-then-hibernate the first cycle
> burns two resumes, so the failure consistently showed up on the second
> hibernation. Tested on 7.0.0-1015-oem: two consecutive hibernations with the
> node alive throughout and no refcount warnings, where before the second one
> hung the machine.
> 
> Signed-off-by: Eduardo Ruiz Duarte <toorandom@gmail.com>
> ---
>  drivers/media/pci/intel/ipu-bridge.c | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> --- a/drivers/media/pci/intel/ipu-bridge.c
> +++ b/drivers/media/pci/intel/ipu-bridge.c
> @@ -765,6 +765,15 @@
>  	if (!fwnode)
>  		return -ENODEV;
>  
> +	/*
> +	 * Balance the kobject_put() that software_node_notify_remove() will do
> +	 * on this node when csi_dev is destroyed. That put is unconditional,
> +	 * while the matching get in software_node_notify() only happens at
> +	 * device_add() time -- and csi_dev was added before we got here, so it
> +	 * never ran for this node.
> +	 */
> +	fwnode_handle_get(fwnode);
> +
>  	set_secondary_fwnode(sensor->csi_dev, fwnode);
>  
>  	return 0;

-- 
With Best Regards,
Andy Shevchenko