Hi Elwyn,
My apologies for the late response on this.
Please find responses inline.
Regards,
Arun
==============================================================
Elwyn Davies wrote:
<eliminating agreed items.
Arun Satyanarayana wrote:
Hi Elwyn,
Thanks much for the review. Please find responses inline.
Regards,
Arun (and co-authors)
==============================================================
Elwyn Davies wrote:
I am the assigned Gen-ART reviewer for
draft-ietf-ccamp-rsvp-restart-ext-05.txt. For background on Gen-ART,
please see the FAQ at
<http://www.alvestrand.no/ietf/gen/art/gen-art-FAQ.html>.
Please consider these comments along with any other last call
comments you may receive.
Summary
=======
Generally an excellent document. I am not sufficiently familiar with
Thanks much.
the details of GMPLS RSVP to be sure that the technical specification
of state recovery are correct but nothing stuck out as missing or
broken.
The one significant issue that I would suggest might need additional
justification/consideration is the implication that it should be used
after the control plane has suffered a security attack (s6, para 3).
RFC3473 specifies the original restart mechanisms for use after
faults and says nothing about security attacks. I would think that if
a particular node has suffered an attack that requires RSVP-TE
restart, there would be concerns about the whole chain of RSVP-TE
nodes. I would particularly worried about restarting an ingress node
in these circumstances. Would it be possible for a compromised node
further down the chain to force a restart and load inappropriate
state? I don't know if this is feasible but I'd want to be sure this
couldn't happen.
Agree that the text is a bit open to interpretation. The basic surmise
here is that the RSVP agent was already protected against a security
breach (via means detailed in RFC2747). So, RSVP signaling state itself
is not compromised. However, due to some other part of the control plane
suffering a more penetrative attack, it is possible that the entire
control plane is restarted to clear such compromised state. The text
refers to RSVP state recovery under these conditions. Will re-work the
text to reflect this, if you feel this is more explicit in indicating
the case in question.
I do think this needs more explanation.
Will add text to this effect.
I have the impression that a graceful restart is going to create a
fair bit of traffic on a router with many paths set up. I didn't see
any suggestions for values of the Recovery Period and its
relationship to number of paths to recover - summarizing the total
messages per path involved in recovery might help give a handle on this.
Recovery times and actual number of messages are very dependant on the
following factors at least:
- The CPU speed on which the restarting as well as neighboring RSVP
agents reside on.
- The CPU availability during the Recovery Period (there may be routing
and/or other control plane being recovered during this period).
- Actual implementation specifics that affect the time taken to
communicate with the hardware to read current connection state:
one implementation may be able read the current connection and label
state via an intermediate software agent quickly; there could be
another implementation may choose to superimpose recovered state
on existing data-plane state and depend on errors if any to validate
state being recovered.
- Forwarding plane flush timers: ingress agents that reside on
packet-originating nodes would possibly have other timers on the
forwarding plane that dictate the upper bound on the Recovery Period.
As an example, it is critical to know that a GMPLS LSP used as an
IP nexthop interface is still available for forwarding - the
keepalive mechanism could be RSVP signaling state combined with
RSVP hellos.
These factors make such text far too open-ended to introduce in this
I-D. It is preferable to suggest values in a companion Informational/BCP
document (hopefully based on real numbers gathered from operational
networks).
Clearly there are difficulties in expressing the full problem but your
own last point would be a good one to make since it is not performance
dependent. Just a sentence reminding people how much time is needed per
LSP (one round trip time or whatever) would be a good idea.
Ok, will do.
Otherwise there are a few minor comments, mostly editorial.
Details
=======
s3, para 8: The name of the Capability object seems a bit general for
something which is very specific to the RecoveryPath mechanism. It
might be better to use something like RECOVERYPATH_CAP (to match with
RESTART_CAP.
The nomenclature was chosen so since the object is carried in Hellos and
there are bits available for carrying other capability information in
the future. Will change the name if you feel strongly about this.
I don't feel terribly strongly, but it is not clear to me that the
semantics of CAPABILITY (tied to RESTART_CAP) make it useful for
anything except other restart associated capabilities, so the generic
name might be confusing.
Hmmm... not the idea the authors intended to convey. We have explicitly
allowed the Capability object to be carried in all Hellos. Also, there
is no text tying it's presence to one of the bits defined in this I-D. I
feel this is sufficient to allow its use to convey other Capabilities as
well in the future.
s4.4.2, last para: Should something be said about any partial state
that has already been recovered when the transition happens? I don't
know enough about the details of the two processes to know if this
could cause problems with interactions between the two processes.
This is covered in s4.5.2.3 (Procedures on Expiration of Recovery
Period). The restarting node cannot make a decision reliably on partial
state until the expiration of the Recovery Period.
Hmm! I am not sure that this statement doesn't actually make matters
worse! So what does it do between the time the bit switches and the end
of the recovery period?
>
Clearly we don't want very elaborate statements about all of this
because it is going to be a rare event (I assume). However, I think you
need to be absolutely clear whether the restarting node should accept
the partial state that it has got before the bit switches or discard it
and behave as if it never got any refreshed state. I am unclear as to
whether this last (discard process) is possible and whether the state
could be in an inconsistent state (too many states) as a result of the
partial interchange.
Would the following text added to the end of s4.4.2 be sufficient to
address your concern?
Specifically, the restarting node MUST continue processing of
all Path messages containing a Recovery_Label object, that have been
received and saved per procedures defined in Section 4.5.2.1. Such
processing MUST follow procedures for handling a Path message with a
Recovery_Label object as defined in Section 9.5.2 of RFC3473.
s5.2.1, last para: As with s4.4.2, does anything need to be said
about partial state already recovered?
Handling of partially recovered state is covered in s4.5.2.3 (Procedures
on Expiration of Recovery Period). Since the SRefresh procedures are
only an optimization to the RecoveryPath procedures defined earlier in
the document, no explicit text has been included to handle partial
state.
Will include text if you feel strongly about this.
Depends a bit on what you say about the other case. Make it clear what
it ought to do to continue.
Will add text referring to the new text suggested above.
/Elwyn
_______________________________________________
Gen-art mailing list
[email protected]
https://www1.ietf.org/mailman/listinfo/gen-art
