Re: [patch net-next-2.6] net: vlan: make non-hw-accel rx path similar to hw-accel

[Jiri Pirko <jpirko@redhat.com>]

Mon, Apr 04, 2011 at 09:51:51PM CEST, ebiederm@xmission.com wrote:
>Nicolas de Pesloüan <nicolas.2p.debian@gmail.com> writes:
>
>> Le 04/04/2011 09:14, Jiri Pirko a écrit :
>>> Mon, Apr 04, 2011 at 08:54:40AM CEST, nicolas.2p.debian@gmail.com wrote:
>>>> Le 03/04/2011 22:38, Jesse Gross a écrit :
>> <snip>
>>>>> It would be nice to merge all of this together.  One complication is
>>>>> the interaction of bridging and vlan on the same device.  Some people
>>>>> want to have a bridge for each vlan and a bridge for untagged packets.
>>>>>   On older kernels with vlan accelerated hardware this was possible
>>>>> because vlan devices would get packets before bridging and on current
>>>>> kernels it is possible with ebtables rules.  If we use rx_handler for
>>>>> both I believe we would need to extend it some to allow multiple
>>>>> handlers.
>>>>
>>>> I totally agree.
>>>
>>> I do not. The reason I do vlan_untag early is so actually emulates
>>> hw acceleration. The reason is to make rx path of hwaccel an
>>> nonhwaccel similar. If you move vlan untag to rx_handler, this goal
>>> wouldn't be achieved.
>>
>> Need to think more about that point.
>>
>>>> Remember that Jiri's original proposal (last summer) was to have
>>>> several rx_handlers per net_device. I still think we need several of
>>>> them, because the network stack need to be generic and allow for any
>>>> complex stacking setup. The rx_handler framework may need to be
>>>> enhanced for that, but I think it is the right tool to do all those
>>>> per net_device specific features.
>>>>
>>>>>> This would also cause protocol handlers to receive the untouched (tagged)
>>>>>> frame, if no setup required the frame to be untagged, which I think is the
>>>>>> right thing to do.
>>>>>
>>>>> At the very least we need to make sure that these packets are marked
>>>>> as PACKET_OTHERHOST because protocol handlers don't pay attention to
>>>>> the vlan field.
>>>>
>>>> Agreed.
>>>>
>>>>>>> @@ -3177,7 +3183,7 @@ ncls:
>>>>>>>                        ret = deliver_skb(skb, pt_prev, orig_dev);
>>>>>>>                        pt_prev = NULL;
>>>>>>>                }
>>>>>>> -             if (vlan_hwaccel_do_receive(&skb)) {
>>>>>>> +             if (vlan_do_receive(&skb)) {
>>>>>>>                        ret = __netif_receive_skb(skb);
>>>>>>>                        goto out;
>>>>>>>                } else if (unlikely(!skb))
>>>>>>
>>>>>> Why are you calling __netif_receive_skb here? Can't we simply goto
>>>>>> another_round?
>>>>>
>>>>> This code (other than the name change) predates the
>>>>> another_round/rx_handler changes.
>>>>
>>>> Yes, you are right. Let's keep this for a possible follow-up patch,
>>>> to avoid skb reinjection when it is not strictly necessary.
>>>
>>> To do another round here was my attention do do in follow up patch (I'm
>>> still figuring out how to move this effectively into rx_handlers)
>>
>> So you want to move vlan_do_receive into an rx_handler, but want untagging to
>> stay hard-coded at the beginning of __netif_receive_skb. I don't think I
>> understand the rational behind that.
>
>__netif_receive_skb is actually late for untagging.  eth_type_trans
>would be better but not path of control into __netif_receive_skb
>actually calls eth_type_trans.

Why __netif_receive_skb is late?

>
>Eric
>