On Fri, 06 Oct 2006 09:48:27 +0100
Alex Zeffertt <ajz@xxxxxxxxxxxxxxxxxxxxxx> wrote:
> Hi list,
>
> I've been reading through the 8021q.o module source recently and
> I am confused by the choice of spin lock used in the following
> code:
>
>
> int vlan_skb_recv(struct sk_buff *skb, struct net_device *dev,
> struct packet_type* ptype)
> {
> ...
> spin_lock_bh(&vlan_group_lock);
> skb->dev = __find_vlan_dev(dev, vid);
> if (!skb->dev) {
> spin_unlock_bh(&vlan_group_lock);
> ...
>
>
>
> Now, I may have this wrong but as I understand it spin_lock_bh()
> will disable softIRQs and spin_unlock_bh() will enable softIRQs (*).
> But vlan_skb_recv() is called from within a softIRQ, and so it
> should execute with softIRQs disabled. Calling spin_unlock_bh()
> will re-enable softIRQs, possibly having undesired consequences.
The _bh is unnecessary and can be removed. It is only called in
RX path and yes softIRQ's are already disabled. The bh disable/enable
is a counter so it does the right thing when calls are nested like
this. It just wastes a few instructions.
> Should we be using spin_lock_irqsave/spin_lock_irqrestore instead?
> I know that this is more heavy handed in that it stops *all* IRQs,
> but there doesn't seem to be a spin_lock_bhsave/spin_lock_bhrestore!
That is overkill.
--
Stephen Hemminger <shemminger@xxxxxxxx>
