Re: Use of usb_find_interface in open is racy

[Russ Dill <russ.dill@gmail.com>]

On Wed, Nov 18, 2009 at 8:31 AM, Alan Stern <stern@rowland.harvard.edu> wrote:
> On Wed, 18 Nov 2009, Jiri Kosina wrote:
>
>> On Tue, 17 Nov 2009, Russ Dill wrote:
>>
>> > Many usb drivers that create character devices use "struct
>> > usb_class_driver", a set of fops, and a usb_find_interface in their
>> > open call. A prime example is drivers/usb/usb-skeleton.c. A race
>> > occurs when userspace receives a hotplug event for the addition for
>> > the interface and then opens the associated device file before the
>> > device is added to the driver's klist_devices.
>> >
>> > The usb core senses a new usb device (usb_new_device) and calls
>> > device_add. This eventually gets down to really_probe and the
>> > usb-skeleton probe function, skel_probe. skel_probe calls
>> > usb_register_dev() which registers the associated character device for
>> > skel_class. The hotplug events for the class device get emitted.
>> >
>> > User space receives the hotplug event for the class device, makes the
>> > device node and notifies another program that opens the device node.
>> > The program opens the device node which calls into usb_open and then
>> > skel_open. skel_open calls usb_find_interface. usb_find_interfaces
>> > searches the klist_devices of skel_driver, finds no device associated
>> > with the minor number and returns NULL. skel_open returns -ENODEV.
>> >
>> > Control returns to really_probe and really_probe calls driver_bound
>> > which adds the device to the list of devices associated with
>> > skel_driver (klist_devices).
>> >
>> > I'm not sure what the right way to solve this is. A call to
>> > wait_for_device_probe() in the skel_open call before calling
>> > usb_find_interface fixes the problem, but it is a rather large hammer.
>
> I think the proper answer is for usb_find_interface() to use
> bus_for_each_dev() instead of driver_for_each_device().
>

This would work, bus_find_device would probably be a better match.