Re: [patch v2, kernel version 3.2.1] net/ipv4/ip_gre: Ethernet multipoint GRE over IP

["Štefan Gula" <steweg@ynet.sk>]

Dňa 17. januára 2012 11:11, David Lamparter <equinox@diac24.net> napísal/a:
> On Tue, Jan 17, 2012 at 10:56:51AM +0100, Štefan Gula wrote:
>> Dňa 17. januára 2012 10:50, David Lamparter <equinox@diac24.net> napísal/a:
>> > On Mon, Jan 16, 2012 at 11:52:25PM +0100, Štefan Gula wrote:
>> >> Dňa 16. januára 2012 21:29, David Lamparter <equinox@diac24.net> napísal/a:
>> >> > At the risk of repeating myself, Linux GRE support already has
>> >> > provisions for multipoint tunnels. And unlike your code, those reuse the
>> >> > existing neighbor table infrastructure, including all of its user
>> >> > interface and introspection capabilities.
>> >> >
>> >> > It's actually slightly visible in your patch:
>> >> >
>> >> > On Mon, Jan 16, 2012 at 08:45:14PM +0100, Štefan Gula wrote:
>> >> >> +++ linux-3.2.1-my/net/ipv4/ip_gre.c  2012-01-16 20:42:03.000000000 +0100
>> >> >> @@ -716,7 +942,19 @@ static netdev_tx_t ipgre_tunnel_xmit(str
>> >> > [...]
>> >> >>               /* NBMA tunnel */
>> >> >>
>> >> >>               if (skb_dst(skb) == NULL) {
>> >> >
>> >> >
>> >> > -David
>> >>
>> >> That code you are referring to is used only for routed traffic inside
>> >> GRE - L3 traffic over L3 routed infrastructure. My patch is dealing
>> >> with L2 traffic over L3 routed infrastructure - so the decision here
>> >> is based on destination MAC addresses and not based on IPv4/IPv6
>> >> addresses.
>> >
>> > Yes, it currently only does IPv4/IPv6 -> IPv4 through the neighbor
>> > table. That doesn't mean it can't be extended to handle ethernet
>> > addresses the same way.
>>
>> Routing mechanisms and switching mechanisms works completely
>> different, in switching you simply don't have anything like next-hop
>> from routing, which can be resolved by utilizing modified ARP message
>> and there is also absolutely no hierarchy in MAC address (like you
>> have in routing table), so I seriously doubt that it can be done the
>> same way, but I am opened to ideas here. So how would you do like to
>> do that?
>
> NBMA GRE does not use routing mechanisms, it uses the neighbor table.
> The neighbor table does not use any hierarchy in its lookups.
>
> So you have the existing neighbor table use cases:
>  - IPv4 -> MAC with in-kernel ARP
>  - IPv6 -> MAC with in-kernel ND
>  - IPv4 -> IPv4 on NBMA GRE devices (without in-kernel filling)
>  - IPv6 -> IPv4 on NBMA GRE devices (without in-kernel filling)
> You could add to that:
>  - MAC -> IPv4 on NBMA GRETAP devices
>
>
> -David
Current status:
 - no general neighbor table exists (currently used build-in tables
for forwarding mechanism ARP, ND, RT, RT6 or combinations to get the
right information)

Your proposal:
 - creating another kernel build-in table -  MAC -> IPv4 on NBMA
GRETAP device table doesn't currently exists at all in kernel
 - adjust/implement learning mechanism to populate the table
 - adjust/implement forwarding mechanisms to forward frames based on
information from given table

My patch:
- creates table for holding MAC -> IPv4 on NBMA GRETAP device
- adjust/implement learning mechanism (using bridge code logic)
- adjust/implement forwarding mechanism (using bridge code logic)

So I think we are talking about the same thing here, only you are
talking in more general way, I am talking more about the actual
implementation.