This patch removes the (ugly and racy) packet receiving thread and the
kernel socket usage. Instead, packets are received directly by registering
the ethernet type and handling skbs instead of self-allocated buffers.

Some consequences and comments:

 * we don't copy the payload data when forwarding/sending/receiving data
   anymore. This should boost performance.
 * packets from/to different interfaces can be (theoretically) processed
   simultaneously. Only the big originator hash lock might be in the way.
 * no more polling or sleeping/wakeup/scheduling issues when receiving
   packets
 * this might introduce new race conditions.
 * aggregation and vis code still use packet buffers and are not (yet)
   converted.
 * all spinlocks were converted to irqsave/restore versions to solve
   some lifelock issues when preempted. This might be overkill, some
   of these locks might be reverted later.
 * skb copies are only done if neccesary to avoid overhead

performance differences:

 * we made some "benchmarks" with intel laptops.
 * bandwidth on Gigabit Ethernet increased from ~500 MBit/s to ~920 MBit/s
 * ping latency decresed from ~2ms to ~0.2 ms

I did some tests on my 9 node qemu environment and could confirm that
usual sending/receiving, forwarding, vis, batctl ping etc works.
Signed-off-by: Simon Wunderlich <siwu@hrz.tu-chemnitz.de>
Acked-by: Sven Eckelmann <sven.eckelmann@gmx.de>
Acked-by: Marek Lindner <lindner_marek@yahoo.de>
Acked-by: Linus Lüssing <linus.luessing@web.de>
Signed-off-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Greg Kroah-Hartman <gregkh@suse.de>