[neutron][queens] dhcp issue

Brian Haley haleyb.dev at gmail.com
Wed Dec 11 20:13:26 UTC 2019

On 12/11/19 12:36 PM, Ignazio Cassano wrote:
> Hello Jeremy, thanks for your help.
> Indeed when more subnet of a provider vlan are created, instance on one 
> of the subnet receives the static routes of all subnets.
> Is it correct?

Yes, this behavior is correct, the dhcp-agent will add host routes for 
all subnets in a network.  I looks like in your case you had a network 
with 12 subnets?  I wonder whether it's worth checking either the number 
of subnets is below some threshold (10?) and just stop supplying the 
routes, or if there is some other way to determine the packet size.

The only issue with limiting the packet size in the reply is the admin 
wouldn't know we did it, since adding the subnet is asynchronous from 
the dhcp-agent adding it's option logic, and things might fail in random 
ways, but I guess having a running dnsmasq is better than not.

I guess at worst we should document that adding too many subnets, or 
extra dhcp options for that matter, could cause issues.


> Il Mer 11 Dic 2019, 16:29 Jeremy Stanley <fungi at yuggoth.org 
> <mailto:fungi at yuggoth.org>> ha scritto:
>     On 2019-12-11 10:10:17 +0100 (+0100), Ignazio Cassano wrote:
>     [...]
>      > Dec 11 09:49:06 podto2-osctrl01 dnsmasq-dhcp[70093]: cannot send
>      >     DHCP/BOOTP option 121: no space left in packet
>     [...]
>     This error is fairly straightforward. DHCP replies are limited to
>     what you can fit in a single UDP packet and how large of a datagram
>     the client agrees to accept. The number of routes you're trying to
>     include in via option 121 could be the problem (is it a bunch?) but
>     also the other fields and their lengths can also contribute, as can
>     the MTU on that network segment:
>     https://tools.ietf.org/html/rfc2131
>     According to IETF RFC 2131 you should be able to assume clients will
>     accept a datagram of up to 576 octets in length (which implies your
>     entire options field can use at least 312 octets). Anything over
>     this requires negotiation of a maximum DHCP message size with the
>     client.
>     -- 
>     Jeremy Stanley

More information about the openstack-discuss mailing list