Re: [mpls] I-D Action: draft-rosen-mpls-rfc3107bis-01.txt

[Lucy yong <[e-mail]>]

I thought about this example :) and think the followings may result different control plane and data plane operations.
Add <a,b>  to NLRI, ->  <a,b,x ...>
Replace <x> by  <a,b,x ...>

The first one works regardless if x exists or not, especially when x means multiple labels.

Lucy

From: Eric C Rosen [mailto:[e-mail]]
Sent: Tuesday, June 21, 2016 9:56 AM
To: Lucy yong
Cc: [e-mail]
Subject: Re: [mpls] I-D Action: draft-rosen-mpls-rfc3107bis-01.txt


*         Sn 3.2.2, the listed local policies seem all related to label swapping actions.
 it is valuable to add a new local policy as follow:

o   Add a single label or a sequence of labels to the NLRI before propagating the route.
     This policy results that, when a node receives a MPLS packet, it will pop out the label(s) and forward the packet to next hop.
I think this is a sub-case of "replacing one label with multiple labels".
[Lucy] IMO, two are different. The case here does not replace any label in NLRI, just add one or multiple labels to the NLRI.
That's the same as replacing <x> by <a,b,x>.





*         Sn 4 data plane description is not sufficient. When applying mutli labels on data plane, we need to specify the rules to fill the label stack entries beside pushing labels; e.g. TTL, EXP. To backward comparable (RFC3032), need to clarify TTL, EXP processing applying to the top label before and after label process action.
3107bis does not convey TTL or TC values, so I think all we need to say is that "the setting of the TTL and TC fields in the label stack of a data packet is determined by local policy".
[Lucy] True. But 3107 does describes data plane operation accordingly. So it should at least mention these fields setting must be compatible to RFC3032 because no intention to change here.
I'll add the following text:

   When pushing labels onto a packet's label stack, the Time-to-Live
   (TTL) field ([RFC3032], [RFC3443]) and the Traffic Class (TC) field
   ([RFC3032], [RFC5462]) of each label stack entry must of course be
   set.  This document does not specify any set of rules for setting
   these fields; that is a matter of local policy.






*         This feature gives each next hop flexibility to determine how many labels to bind a prefix, which may impact Path MTU. We SHOULD avoid each path segment to fragment labeled packets.
MPLS does not have fragmentation, so it is not possible to fragment labeled packets ;-)
[Lucy] pls check RFC3032.
I'd forgotten that RFC3032 contains an elaborate discussion about applying IP fragmentation in the case where the MPLS payload is an IP datagram and pushing on labels might cause the MTU to be exceeded.   In 1997, it seemed like that might be important, but I don't think those procedures are actually used much (if at all).

I don't think 3107bis needs to mention MTU issues, any more than any other draft that provides a way of distributing labels.

You're right that a router should not be instructed to push on so many labels that the MTU is exceeded.  But this is always an issue with MPLS, and is not specifically related to 3107bis.  It is outside the scope of this draft to specify procedures for determining the maximum number of labels that can be safely pushed.
[Lucy]  This mechanism adds more complex for Path MTU computation. It is good to point out it when describing data plane.
The "data plane" section of 3107bis is about forming a packet's label stack based on the information in SAFI-128 or SAFI-4 UPDATES.  It is not a general discussion of MPLS data plane issues.





*         This example in Sn 4 is not correct.



   In this case, if S1 receives an MPLS data packet whose top label is
   L21 and whose second label is L22, S1 will remove both L21 and L22
  from the label stack, and replace them with <L11,L12,...L1k>.  Note
   that the fact that L21 is a context label is known only to S1; other
   BGP speakers do not know how S1 will interpret L21 (or L22).

   The ability to replace one or more labels by one or more labels can
   provide great flexibility, but must be done carefully.  Let's suppose
   again that S1 receives an UPDATE that specifies prefix P, label stack
   <L11,L12,...,L1k>, and next hop N1.  And suppose that S1 propagates
   this UPDATE to BGP speaker S2 after setting next hop self and after
   replacing the label field with <L21,L22,...L2k>.  Finally, suppose
   that S1 programs its data plane so that when it processes a received
   MPLS packet whose top label is L21, it replaces L21 with
   <L11,L12,...,L1k>, and then tunnels the packet to N1.

   In this case, BGP speaker S2 will have received a route with prefix
   P, label field <L21,L22,...L2k>, and next hop S1.  If S2 decides to
   forward an IP packet according to this route, it will push
   <L21,L22,...L2k> onto the packet's label stack, and tunnel the packet
   to S1.  S1 will replace L21 with <L11,L12,...,L1k>, and will tunnel
   the packet to N1.  N1 will receive the packet with the following
   label stack: <L11,L12,...L1k,L22,...L2k>.  While this may be useful
   in certain scenarios, it may provide unintended results in other
   scenarios. -end

   Lucy: Label <L21,L22,L2k> is advertised by S1, it does not make a sense that S1
   programs its data plane so that when it processes a received MPLS packet whose top label
   is L21, it replaces L21 with <L11,L12,..L1k>, and tunnel the packet to N1, i.e.
   N1 will receive the packet with the following
   label stack: <L11,L12,...L1k,L22,...L2k>.
Whether this example corresponds to an actual use case is debatable.  The example merely shows something that can be done with the specified mechanisms.

This scenario would only be useful if S1 knows somehow that L22 will rise to the top of the packet's label stack at a node to which L22 is meaningful.
[Lucy] IMO: this is a fault case.
As I said, this is just an example of something you could do with the mechanisms of 3107bis.  If there's no use case for it, no one will do it.





S1 should replace <L21, L22, ..L2k> with
   <L11,L12,...,L1k> in this case.
No, 3107bis does not (and should not) modify any of the rules for processing the label stack of an incoming packet.
[Lucy] ??
If S1 really wanted to replace <L21,L22,...,L2k> with <L11,L12,...,L1k> it would program each of <L21,L22,...,L2(k-1)> as a POP, and it would program L2k as a "POP and then PUSH <L11,L12,...,L2k>".  But that's a different example than the one given in the draft.