relation between kernel FIB and hardware FIB

[Mark]

Hello

This is more a genric question on the implementation of routers. Does a
software in the control plane (let's consider RIP, OSPF or BGP protocols)
have to install routing entries in both kernel and underlying hardware
(given that the hardware is capable to maintain its FIB), or in such case
the kernel doesn't need to have a routing table and the routing table is
constructed and maintained entirely by the control plane routing prototols ?

Thanks.

Mark

[Barry Margolin]

In article <k3fes8$fo9$1...@speranza.aioe.org>,

Usually they each maintain separate tables. Each routing protocol
typically needs to maintain its own state, because they operate
independently -- a RIP route shouldn't normally be advertised via BGP.
The hardware FIB is created by merging all the information from the
different protocols.

Also, the memory for the hardware FIB is usually limited, so you just
want the minimal forwarding information. But routing protocols need to
maintain state, sometimes quite a bit of it (e.g. the AS path in BGP).

--
Barry Margolin, bar...@alum.mit.edu
Arlington, MA
*** PLEASE post questions in newsgroups, not directly to me ***

[Mark]

"Barry Margolin" <bar...@alum.mit.edu> wrote in message
news:barmar-2FD372....@news.eternal-september.org...

> Usually they each maintain separate tables. Each routing protocol
> typically needs to maintain its own state, because they operate
> independently -- a RIP route shouldn't normally be advertised via BGP.
> The hardware FIB is created by merging all the information from the
> different protocols.

Thanks for the reply. To be more specific -- a routing protocol maintains
its state and RIB table should reflect the RIB information down to the
kernel or/and hardware; so my question is -- should it be done on both
kernel and hardware, i.e. FIB in the hardware, or only FIB in hardware? (By
the hardware let's consider L3 switch, or network processor). Obviously if
we have a generic x86 machine with NICs and OS Linux, then the only FIB we
update is the kernel's FIB, but what if we have Linux and user-level
applications implementing routing protocols, and underlying network
processor capable of maintaining its own FIB tables?

> Also, the memory for the hardware FIB is usually limited, so you just
> want the minimal forwarding information. But routing protocols need to
> maintain state, sometimes quite a bit of it (e.g. the AS path in BGP).
>

Mark

[Barry Margolin]

In article <k3fpih$ch7$1...@speranza.aioe.org>,

"Mark" <mark_cruz...@hotmail.com> wrote:

> "Barry Margolin" <bar...@alum.mit.edu> wrote in message
> news:barmar-2FD372....@news.eternal-september.org...
> > Usually they each maintain separate tables. Each routing protocol
> > typically needs to maintain its own state, because they operate
> > independently -- a RIP route shouldn't normally be advertised via BGP.
> > The hardware FIB is created by merging all the information from the
> > different protocols.
>
> Thanks for the reply. To be more specific -- a routing protocol maintains
> its state and RIB table should reflect the RIB information down to the
> kernel or/and hardware; so my question is -- should it be done on both
> kernel and hardware, i.e. FIB in the hardware, or only FIB in hardware? (By
> the hardware let's consider L3 switch, or network processor). Obviously if
> we have a generic x86 machine with NICs and OS Linux, then the only FIB we
> update is the kernel's FIB, but what if we have Linux and user-level
> applications implementing routing protocols, and underlying network
> processor capable of maintaining its own FIB tables?

If you're implementing a router, you can do whatever you want. The main
benefit of keeping a copy of the FIB in the kernel is that it doesn't
have to interrupt the hardware if an application asks to view the FIB.