Chas writes:

could you try this patch as well as the other?  it should remove the
race condition that exists in the current code.  in the old code, the
atm_may_send check is after the send.  this is racing with the clip_pop
which will occur right after the downcall to the driver's send (atleast
the he driver works this way, its trying to free those outstanding skb's
as soon as possible)  if we do the work before the down call this should
help some.  i also moved the xoff_lock to protect xoff and the
netif_queue calls.  you really want these to occur together w/o
clip_pop interrupting.