* jamal [off-list ref] 2006-06-30 22:23

I am  not certain i understood then: Are we in the mode where the
refcount is not needed because chances are small that a device will
disappear? It seems to me after all this trouble that it may not be so
bad to refcount (I guess i meant refcount the device on input to the ifb
and decrement on the output).

Based on two principles:

 a) All netdevices a packet is processed through is properly
    refcounted until the packet is queued for the first time.
    Given that iif is strictly set to the previous device, a
    refcnt is certainly taken up to the point where the skb is
    put into rq. There is one exception to this: The fact that
    you don't set skb->dev = ifb when reinjecting at ingress
    will not take a refcnt on the ifb. This sounds like a broken
    architecture to me but it doesn't matter as you want to
    prohibit ifb -> ifb anyways.
 
 b) The netdevice used for xmit via dev_queue_xmit() is already
    protected by the initial xmit attempt.

I can't see reason why to hurt performance by introducing more
atomic operations in your fast path.

If you have more concerns regarding these patches, feel free to
fix your own architecture or propose to drop the patches, I think
I've spend enough time on this.