USB core sanity check

  
  Periodically folk have run into problems where usb-ohci oopses
  due to device refcount bugs ...
    
  This is a minor patch to move the sanity check out of usb-ohci
  into the generic bits of usbcore.    There are comments that
  suggest a path for a more comprehensive approach too.
    
  Applies cleanly against 2.5.7 and I've been testing with it
  for a while.  I can't think of any reason it shouldn't also go
  into 2.4, beyond the patch not applying cleanly there ... :)

drivers/usb/usb.c

@@ -791,9 +791,22 @@ struct usb_device *usb_alloc_dev(struct usb_device *parent, struct usb_bus *bus)
 
 // usbcore-internal ...
 // but usb_dec_dev_use() is #defined to this, and that's public!!
+// FIXME the public call should BUG() whenever count goes to zero,
+// the usbcore-internal one should do so _unless_ it does so...
 void usb_free_dev(struct usb_device *dev)
 {
 	if (atomic_dec_and_test(&dev->refcnt)) {
+		/* Normally only goes to zero in usb_disconnect(), from
+		 * khubd or from roothub shutdown (rmmod/apmd/... thread).
+		 * Abnormally, roothub init errors can happen, so HCDs
+		 * call this directly.
+		 *
+		 * Otherwise this is a nasty device driver bug, often in
+		 * disconnect processing.
+		 */
+		if (in_interrupt ())
+			BUG ();
+
 		dev->bus->op->deallocate(dev);
 		usb_destroy_configuration(dev);