In the Linux kernel, the following vulnerability has been resolved:

HID: universal-pidff: stop the device when force-feedback init fails

universal_pidff_probe() starts the device with hid_hw_start() and then, if
force-feedback initialisation fails, returns the error through a label that
only does "return error".  The device is left started.

The HID core does not unwind on the driver's behalf.  __hid_device_probe()
releases the devres group, closes the report and clears hdev->driver:

	if (ret) {
		devres_release_group(&hdev->dev, hdev->devres_group_id);
		hid_close_report(hdev);
		hdev->driver = NULL;
	}

The hidraw character device that hid_hw_start() registered through
hid_connect() is allocated with kzalloc() and added with cdev_device_add(),
so it is not devres-managed and survives that.  With hdev->driver NULL,
hid_device_remove() skips hid_hw_stop() as well, because it only unwinds
while a driver is still attached.  The registration therefore outlives the
device on both paths.

Opening the surviving /dev/hidrawX writes into freed memory.  KASAN reports
a use-after-free write from hidraw_open() -> hid_hw_open() -> the
transport's open callback, which takes a spinlock inside the freed object.
A descriptor that carries a PID usage page and no input reports is enough:
hidraw claims the device so hid_hw_start() succeeds, while hid->inputs
stays empty so force-feedback init fails.  The other failure returns in
hid_pidff_init_with_quirks() - no output reports, an allocation failure,
pidff_init_fields(), pidff_check_autocenter(), an unusable effect count,
input_ff_create() - all reach the same label.

Stop the device on that path.  hid-dr.c and hid-emsff.c, which start the
device with the same HID_CONNECT_DEFAULT & ~HID_CONNECT_FF mask, already do
this.  The two earlier gotos must keep returning without hid_hw_stop(),
since neither has a started device, so give the path that fails after the
start its own label.

Discovered by XBOW, triaged by Baul Lee <baul.lee>