ipv6: a question about ECMP

ipv6: a question about ECMP

[Duan Jiong]

Hi,

  After reading the ip6_pol_route(), i have a question about ECMP. Why we call
the rt6_multipath_select() after calling rt6_select()?
  In my opinion, the route returned by rt6_select() has a highest score, but the route
returned by rt6_multipath_select() may has a lower score than the former, because the
ECMP don't take the route preference into consideration. That means that the kernel will
choose a less-desirable route.

Thanks,
  Duan
  

Re: ipv6: a question about ECMP

[Hannes Frederic Sowa]

Hi Duan!

On Thu, Nov 07, 2013 at 06:33:20PM +0800, Duan Jiong wrote:
>   After reading the ip6_pol_route(), i have a question about ECMP. Why we call
> the rt6_multipath_select() after calling rt6_select()?
>   In my opinion, the route returned by rt6_select() has a highest score, but the route
> returned by rt6_multipath_select() may has a lower score than the former, because the
> ECMP don't take the route preference into consideration. That means that the kernel will
> choose a less-desirable route.

ECMP routes only differ in the gateway the specify, so I doubt there will be
any change in the score they woud receive. rt6_multipath_select does merly
make sure we don't select the same route again and again.

Please note, the rt6_info's siblings fields were added for the solely purpose
of ECMP and the insertion only updates the siblings list if the above criteria
did hold. They make sure the routes lookup up do differ on each lookup, so it
does actually do multipath and does not depend on the order the routes where
inserted.

Hope that helps,

  Hannes

Re: ipv6: a question about ECMP

[sowmini varadhan]

On Thu, Nov 7, 2013 at 7:16 AM, Hannes Frederic Sowa
<hannes@stressinduktion.org> wrote:
> Hi Duan!
>
> On Thu, Nov 07, 2013 at 06:33:20PM +0800, Duan Jiong wrote:
>>   After reading the ip6_pol_route(), i have a question about ECMP. Why we call
>> the rt6_multipath_select() after calling rt6_select()?
>>   In my opinion, the route returned by rt6_select() has a highest score, but the route
>> returned by rt6_multipath_select() may has a lower score than the former, because the
>> ECMP don't take the route preference into consideration. That means that the kernel will
>> choose a less-desirable route.
>
> ECMP routes only differ in the gateway the specify, so I doubt there will be
> any change in the score they woud receive. rt6_multipath_select does merly
> make sure we don't select the same route again and again.

 rt6_multipath_select() -> rt6_socre_route() seems to require that the
interface *must* matchi, which is consistent with your assertion above that
"ECMP routes differ in gw only".

But for IPv6, the gw addr is a a link-local, which is only required to be
unique on the link. Thus, e.g.,  you can have fe80::1 as the gw on both eth0 and
eth1.

What is the assumption around "cost" for ECMP here- are we assuming some
form of link bundling (Section 6 of rfc 2991) here? or is the "multiple parallel
links" case handled somewhere else, that I am missing?

--Sowmini

>
> Please note, the rt6_info's siblings fields were added for the solely purpose
> of ECMP and the insertion only updates the siblings list if the above criteria
> did hold. They make sure the routes lookup up do differ on each lookup, so it
> does actually do multipath and does not depend on the order the routes where
> inserted.
>
> Hope that helps,
>
>   Hannes
>
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: ipv6: a question about ECMP

[Nicolas Dichtel]

Le 07/11/2013 19:32, sowmini varadhan a écrit :
> On Thu, Nov 7, 2013 at 7:16 AM, Hannes Frederic Sowa
> <hannes@stressinduktion.org> wrote:
>> Hi Duan!
>>
>> On Thu, Nov 07, 2013 at 06:33:20PM +0800, Duan Jiong wrote:
>>>    After reading the ip6_pol_route(), i have a question about ECMP. Why we call
>>> the rt6_multipath_select() after calling rt6_select()?
>>>    In my opinion, the route returned by rt6_select() has a highest score, but the route
>>> returned by rt6_multipath_select() may has a lower score than the former, because the
>>> ECMP don't take the route preference into consideration. That means that the kernel will
>>> choose a less-desirable route.
>>
>> ECMP routes only differ in the gateway the specify, so I doubt there will be
>> any change in the score they woud receive. rt6_multipath_select does merly
>> make sure we don't select the same route again and again.
>
>   rt6_multipath_select() -> rt6_socre_route() seems to require that the
> interface *must* matchi, which is consistent with your assertion above that
> "ECMP routes differ in gw only".
In fact, ECMP routes have the same metric/weight and destination but not the
same next hop (ie gw + oif).

>
> But for IPv6, the gw addr is a a link-local, which is only required to be
> unique on the link. Thus, e.g.,  you can have fe80::1 as the gw on both eth0 and
> eth1.
Yes, oif can be different.
Note that gw can also be a global address.

>
> What is the assumption around "cost" for ECMP here- are we assuming some
> form of link bundling (Section 6 of rfc 2991) here? or is the "multiple parallel
> links" case handled somewhere else, that I am missing?
rt6_score_route() is called to check requested oif (see 52bd4c0c1551 "ipv6: fix 
ecmp lookup when oif is specified").

Regards,
Nicolas

>
> --Sowmini
>
>>
>> Please note, the rt6_info's siblings fields were added for the solely purpose
>> of ECMP and the insertion only updates the siblings list if the above criteria
>> did hold. They make sure the routes lookup up do differ on each lookup, so it
>> does actually do multipath and does not depend on the order the routes where
>> inserted.
>>
>> Hope that helps,
>>
>>    Hannes
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe netdev" in
>> the body of a message to majordomo@vger.kernel.org
>> More majordomo info at  http://vger.kernel.org/majordomo-info.html