The MPLS WG Archive

[Eric Gray <eric.gray@sandburst.com>]

Paul,

    See below...

You wrote:

> Regarding the scenario discussed in Section 3.5.7.1.4. of RFC 3036
>
> "The combination of Downstream Unsolicited mode and conservative label
>    retention can lead to a situation where an LSR releases the label for
>    a FEC that it later needs.  For example, if LSR Rd advertises to LSR
>    Ru the label for a FEC for which it is not Ru's next hop, Ru will
>    release the label.  If Ru's next hop for the FEC later changes to Rd,
>    it needs the previously released label "
>
> The section says the following "To deal with this situation either Ru
> can explicitly request the label when it needs it, or Rd can periodically
> readvertise it to Ru ".
>
> Does this mean that even though during session initialization Ru stated it
> was configured for downstream unsolicited it is allowed to send a label
> request message. Have I understood this correctly ?

Yes.

>
>
> and does this mean the downstream will accept a request from an
> upstream router that has signaled that it is configured for Du ?

Again, yes.

>
> The latter mentioned solution i.e. periodically readvertise, is this
> inferring periodic label refresh?

There are a couple of things to realize here:

1   it does - in general - not make sense to combine the two modes
    that are discussed in this citation (i.e. - Downstream unsolicited
    and conservative retention) and
2  the upstream LSR should have sent a release message for any
    labels it has decided not to use.

The latter of these two things is a BIG part of the reason why the
former is true (i.e. - the two modes do not make sense together).
This would tend to support the idea that the combination is unlikely
in practice.  Plus the fact that the upstream LSR is supposed to
send release messages is a mitigation for the behavior you describe:
the behavior is not refresh, in the usual sense, since the labels being
readvertised have been explicitly released.

Whether or not it makes sense to litterally refresh label mappings
- either to support a bad upstream implementation (one that does
not explicitly release labels which does not intend to use) or to allow
a downstream LSR to avoid keeping state on labels it has issued - is
another matter.  I don't think it's a good idea.

>
>
> Thanks
>
> Paul

--
Eric Gray