Re: [PATCH ipvs,v3 00/20] Support v6 real servers in v4 pools and vice versa

[Alex Gartrell <agartrell@fb.com>]

On 8/28/14 8:15 AM, Julian Anastasov wrote:
>
> 	Hello,
>
> On Wed, 27 Aug 2014, Alex Gartrell wrote:
>
>> At Facebook we use ipip forwarding to deliver packets from our layer 4 ipvs
>> load balancers to our layer 7 proxies.  Today these layer 7 proxies are all
>> dual stacked, so we can simply send v4 over v4 and v6 over v6.  In the
>> future, we're going to have v6-only layer 7 load balancers (no internal v4
>> address).  To deal with this, we'll forward v4 packets in v6 tunnels.  This
>> patchset introduces the necessary functionality into ipvs.
>>
>> The noteworthy limitation of this is that it is not compatible with state
>> synchronization, so great care is taken to keep these things mutually
>> exclusive.
>>
>> This patchset includes changes that add an additional netlink attribute to
>> destinations and changes that plumb the destination address family through
>> parts of the code where it was assumed that it was the same as the service.
>> Finally, there's a change that updates the transmit functions for tunneling
>> to share common code and support v4 in v6 and vice versa.
>>
>> Changes for v2:
>>
>> Introduced crosses_local_route_boundary and update_pmtu functions and
>> conditionally do the ip_hdr operations.  The latter means that df will be
>> effectively zero when we forward an ipv6 packet over an ipv4 tunnel.
>>
>> Additionally, I addressed Julian's other statements.
>>
>> Changes for v3:
>>
>> added back the first two patches
>> checkpatch.pl all of the things
>> pull out the mtu changes
>> other previously detailed fixes
>
> 	I think, we are almost at the final:

Awesome :)

>
> Patch 5:
> 	- we can mix patch 5 and 6?

Yeah sure


>
> Patch 7:
> 	- I guess addr_type var should be in the
> 	'if (skb_af == AF_INET6) {' block. So, do not
> 	forget to compile next time with CONFIG_IP_VS_IPV6=n

I'll make sure I compile each patch with this option.

>
> Patch 9:
> 	- for ensure_mtu_is_adequate():
> 		- move 'struct netns_ipvs *ipvs' below, where ipvs is
> 		assigned
> 		- move 'struct net *net' below, where net is assigned

no problem

>
> Patch 10:
> 	- 'is'->'if' in
> 	/* We only care about the df field is sysctl_pmtu_disc(ipvs) is set */
>

Yeah

>
> 	After checking the cp->daddr usage I prepared
> a patch that we should insert somewhere before the last one.
> Attached.

I'm just going to make one small change: dbuf isn't declared without 
CONFIG_IP_VS_IPV6.

> Later we have to wait
> "ipvs: properly declare tunnel encapsulation" to appear in net-next.
>

No problem, it'll give me time to test it in our environment.