OSFP LSA_MAXSEQNO problems

[Alexander V. Chernikov]

On 14.08.2013 14:18, Ondrej Zajicek wrote:
> On Wed, Aug 14, 2013 at 03:35:07AM +0400, Alexander V. Chernikov wrote:
>> Hello list!
>>
>> Currently bird performs LSA premature aging in very strange way which
>> sometimes upsets quagga (up to SIGSEGV) and Cisco.
>> Every time aging is done via setting seq number to LSA_MAXSEQNO.
>>
>> RFC 2328, on the opposite, shows us only 2 cases where MaxSequenceNumber
>> is used in outgoing LSAs:
>> 1) Self LSA originating with given seq num (very rare event)
>> 2) [Premature] Aging of LSA from 1) because a) disappeared net and b)
>> seq number increase (12.1.6)
>>
>> For all other cases seq number should be either increased or left intact.
>>
>>
>> Bird, however, insist on setting LSA_MAXSEQNO while:
>> 1) Receiving self-originating LSA which is not in local LSDB (
>> ospf_lsupd_receive ) which is not a special case according to 14.1
>> 2) In all other places like handling if-down, external route disappear,
>> etc... ( ospf_lsupd_flush_nlsa )
> You are right that BIRD uses LSA_MAXSEQNO in more cases than prescribed
> by RFC 2328 (essentially in most cases when a local LSA is flushed), but
> these changes were introduced mostly as a reaction to problems with the
> usual flushing (just premature aging), where BIRD sends LSA with MaxAge,
> then forgets the LSA and when reintroduced the LSA later, it starts with
> InitSeqNum. In cases like external route flaps (and similar events for
> other kinds of LSAs) that caused that newer LSA meets MaxAge LSA during
> flood, is considered as older and removed (and combined with other
> minor inconsistencies in router behaviors this would lead to all kinds
> of strange events).
Let me return to this old, but still relevant topic :)
Two patches, unifying work with sn/age LSA fields with sn story-keeping 
stuff attached.

First one replaces most initial LSA header setup with single 
fill_lsa_header() function.
It is also responsible for (unique) LSA id generation.

Second one handles sequence number keeping for all LSA types using 
per-are FIB for most LSAs and
per-proto FIB for AS-boundary (type5/7/0x2007/0x4005).
I'd prefer to use fib2_init() for lookups, but fib_init() seems to be 
sufficient, too.

We definitely have to convert ipa_to_rid() to some stateful stuff at 
least in IPv6 case, but that's the next step.
We also need to deal more accurately with LSA_MAXSEQ.
>
> The idea behind the behavior is that you should either keep the old
> seqnum for your LSA when it is flushed (and continue in it when you
> reintroduce it), or you should use LSA_MAXSEQNO to ensure that old seqnum
> is eliminated from the OSPF domain so you could forget seqnum and later
> start from InitSeqNum. The behavior is not according to RFC 2328, but it
> is compatible with it. And i have a work-in-progress code that fixes several
> issues in the flooding code together with this.
>