quoted

Also you are at the point the tty is closing so the net device may be
active. Don't you need to netif_stop_queue() or defer the buffer
kfrees until after the network device is unregistered so you don't pee
into free memory if you have a transmit occurring ?

I'm pretty sure that's what the semaphore down above this sequence is
accomplishing.  But if we do need the netif_stop_queue() let's do that
as a separate patch.

Follow the code path for sp_xmit(). If sp_xmit is called it digs out sp
from the ndetdev, locks sp->lock and stops the queue then calls sp_encaps
which touches sp->xbuff.

So if one thread of execution hits sp_xmit and another closes the ldisc
at just the wrong moment then we have no protection.

Alan