On Jun 9, 2005, at 7:34 AM, Manuel Bouyer wrote:

> Hi,
> at which spl level should ether_input() be called ?
> When called from a real device driver it's at splnet, but when it  
> is from
> a pseudo-device it's not (and I got a panic because of this).

It should be called at splnet().

> Parts of ether_input are protected with splnet(), but there is one  
> missing:
> the call to bridge_input(). In my case, a write to /dev/tap0 ended up
> calling ex_start() at a lower priority than IPL_NET (probably IPL_BIO,
> I didn't check). bridge_input() itself also needs to run at splnet().
>
> I see 2 ways to fix this:
> 1) declare that ether_input() needs to always be called at splnet()  
> (as it's
>    already this way for the common case), add splnet()/splx()  
> around the
>    (*ifp->if_input) call in tap_dev_write(), and remove the unecessary
>    splnet()/splx() in ether_input().

That seems the most appropriate.

> 2) add the missing splnet()/splx() in ether_input().
>
> I think the most efficient way it 1). But if the convention is that
> (*ifp->if_input) doesn't need to be called at splnet() (I didn't find
> infos about this) then we need to do 2).

if_input() should always be called at splnet(), IMO.

>
> -- 
> Manuel Bouyer <bouyer@antioche.eu.org>
>      NetBSD: 26 ans d'experience feront toujours la difference
> --
>

-- thorpej