[j-nsp] strange behavior of Junos LDP (wtf!!)
Piotr Marecki
p.marecki at swiat.pl
Wed Oct 13 17:31:47 EDT 2004
Ehlo ,
lets consider following topology :
Ra<--->Rb<--->Rc<--->Rd[x.x.x.x/32] , where :
Ra , Rd are "other vendor" routers , Rb and Rc are Juniper M series . There
is common IGP and LDP
Rd is advertising prefix fec x.x.x.x/32 to MPLS domain as it is egress
router for x.x.x.x/32.
All label bindings are propagated , in another words LSP for FEC x.x.x.x is
working . Consider following label values :
Ra ( x.x.x.x/32 , Push X ) ----> Rb ( X , swap Y ) -----> Rc ( Y , label
pop ) ------> Rd
1) Time T0
Now , for whatever reason Rd issues LabelWithdraw for FEC x.x.x.x/32 ( with
label = 3 , eg implict null ) while IGP topology stays the same.Now horror
begins
In this very moment Rc release label and issue LabelRelease message for
x.x.x.x/32
and label = 3 to Rd. It also update it's LFIB by deleting
ILM entry for label Y , eg it no longer recognize Y as legitimate label for
x.x.x.x.Problem is that Rc delays (?) sending LabelWithdraw
(X.X.X.X/32,Y) to Rb for 60 seconds.In effect tagged traffic for x.x.x.x is
dropped on Rc .That's it - Rb is still sending Y tagged traffic while Rc
doesn't hold
ILM entry for that particular label ( it's true for 6.2R2.4 , 6.2R3 and
6.4R2.4)
2) Time T0+60 seconds
Rc is so nice that ( after 60 seconds of traffic going to null ) it at last
sends LabelWithdraw message for
(x.x.x.x/32,Y).Rb responds with LabelRelease for x.x.x.x/32 and label=Y , it
delete LFIB for label X , however once
again Rb doesn't send LabelWithdraw for (x.x.x.x/32,X) to Ra.In effect Ra
sends labelled traffic to Rb , and Rb is
of course discarding it.
3) Time T0+120
Rb sends LabelWithdraw to Ra for (x.x.x.x/32,label X) -> now traffic can
reach destination.
Summary :
For n ( where n = amount of Juniper routers in path ) x 60 seconds traffic
is going to bitbucket.
This problem is of course limited only to case where IGP topology stays
unchanged - but nevertheless
it is f****d up.Question to J folks - what is going on ?.
regards
Piotr Marecki
Netia S.A
More information about the juniper-nsp
mailing list