hypothetical vlan rx path question

hypothetical vlan rx path question

[Jiri Pirko]

Hi guys.

Consider following code taken from 8139cp.c


static inline void cp_rx_skb (struct cp_private *cp, struct sk_buff *skb,
			      struct cp_desc *desc)
{
	skb->protocol = eth_type_trans (skb, cp->dev);

	cp->dev->stats.rx_packets++;
	cp->dev->stats.rx_bytes += skb->len;

#if CP_VLAN_TAG_USED
	if (cp->vlgrp && (desc->opts2 & cpu_to_le32(RxVlanTagged))) {
		vlan_hwaccel_receive_skb(skb, cp->vlgrp,
					 swab16(le32_to_cpu(desc->opts2) & 0xffff));
	} else
#endif
		netif_receive_skb(skb);
}


Now my question is why the check for cp->vlgrp is needed here. Because
in hypothetical case it might be possible to receive vlan packet as
non-vlan packet (vlan tag would be lost).

This is present in many drivers.

How about to kill this check entirely and let the later code to deside
what to do with the packet?

Thanks.

Jirka

Re: hypothetical vlan rx path question

[Stephen Hemminger]

On Wed, 13 Jul 2011 22:49:46 +0200
Jiri Pirko <jpirko@redhat.com> wrote:

> Hi guys.
> 
> Consider following code taken from 8139cp.c
> 
> 
> static inline void cp_rx_skb (struct cp_private *cp, struct sk_buff *skb,
> 			      struct cp_desc *desc)
> {
> 	skb->protocol = eth_type_trans (skb, cp->dev);
> 
> 	cp->dev->stats.rx_packets++;
> 	cp->dev->stats.rx_bytes += skb->len;
> 
> #if CP_VLAN_TAG_USED
> 	if (cp->vlgrp && (desc->opts2 & cpu_to_le32(RxVlanTagged))) {
> 		vlan_hwaccel_receive_skb(skb, cp->vlgrp,
> 					 swab16(le32_to_cpu(desc->opts2) & 0xffff));
> 	} else
> #endif
> 		netif_receive_skb(skb);
> }
> 
> 
> Now my question is why the check for cp->vlgrp is needed here. Because
> in hypothetical case it might be possible to receive vlan packet as
> non-vlan packet (vlan tag would be lost).
> 
> This is present in many drivers.
> 
> How about to kill this check entirely and let the later code to deside
> what to do with the packet?
> 
> Thanks.
> 
> Jirka

All this is moot with new vlan model. It should always just put
tagged packet up.

I think all drivers with .ndo_vlan_rx_register need to be still converted.

Re: hypothetical vlan rx path question

[Jiri Pirko]

Wed, Jul 13, 2011 at 11:11:27PM CEST, shemminger@linux-foundation.org wrote:
>On Wed, 13 Jul 2011 22:49:46 +0200
>Jiri Pirko <jpirko@redhat.com> wrote:
>
>> Hi guys.
>> 
>> Consider following code taken from 8139cp.c
>> 
>> 
>> static inline void cp_rx_skb (struct cp_private *cp, struct sk_buff *skb,
>> 			      struct cp_desc *desc)
>> {
>> 	skb->protocol = eth_type_trans (skb, cp->dev);
>> 
>> 	cp->dev->stats.rx_packets++;
>> 	cp->dev->stats.rx_bytes += skb->len;
>> 
>> #if CP_VLAN_TAG_USED
>> 	if (cp->vlgrp && (desc->opts2 & cpu_to_le32(RxVlanTagged))) {
>> 		vlan_hwaccel_receive_skb(skb, cp->vlgrp,
>> 					 swab16(le32_to_cpu(desc->opts2) & 0xffff));
>> 	} else
>> #endif
>> 		netif_receive_skb(skb);
>> }
>> 
>> 
>> Now my question is why the check for cp->vlgrp is needed here. Because
>> in hypothetical case it might be possible to receive vlan packet as
>> non-vlan packet (vlan tag would be lost).
>> 
>> This is present in many drivers.
>> 
>> How about to kill this check entirely and let the later code to deside
>> what to do with the packet?
>> 
>> Thanks.
>> 
>> Jirka
>
>All this is moot with new vlan model. It should always just put
>tagged packet up.

That's what I thought.
>
>I think all drivers with .ndo_vlan_rx_register need to be still converted.

Yes, I'm kinda on the way to do that.

Thanks Stephen.