Yes, that could be a problem. Using test_and_set_bit() can
guard the freezing sequence itself, but it won't handle
letting two threads of control freeze and unfreeze safely
without a reference count.
We want this thing to be able to be used flexbly, which means
we can't just assume that this is a short code sequence and
the unfreeze will come quickly. That pretty much rules
out using a new lock around the operation or anything
like that.
So I guess we could replace the state bit with a reference
count. It doesn't even need to be atomic since it is set
and tested under dev_queue->_xmit_lock
--
