On Fri, 2013-03-29 at 09:17 -0400, Steven Rostedt wrote:

I've thought about this too, but I wasn't sure we wanted two
synchronize_*() functions, as the caller does a synchronize as well.
That said, I think this is the more robust solution and it lets all
rx_handler() functions assume that their rx_handler_data is set. And it
removes the check from the fast path which outweighs an added
synchronization in the slow path.

Note that I used synchronize_net(), which does a
synchronize_rcu_expedited() when RTNL is locked, so its normally quite
fast.

Acked-by: Steven Rostedt <rostedt@goodmis.org>

Thanks!

Thanks a lot for your very detailed report and analysis !