> > I don't believe this is accurate.  Calling skb_linearize() (on a kernel
 > > with CONFIG_HIGHMEM set) can end up calling local_bh_enable() in
 > > kunmap_skb_frag(), which can obviously cause problems if the initial
 > > context relies on having BHs disabled (as hard_start_xmit does).

 > local_bh_{enable,disable}() nests, so this is not a problem

Duh.  OK, then the only bugs seem to be that iw_nes does skb_linearize
with irqs off (due to being an LLTX driver), and mv643xx_eth leaks an
skb on its error path if skb_linearize fails.

 - R.