Greg,
On 26/05/18 20:49, Greg Mirsky wrote:
Hi Bob,
thank you for the thorough review, detailed questions and helpful
comments. Please find my answers in-line and tagged GIM>>.
I've updated the working version of the draft based on your comments
and suggestions. Appreciate your feedback whether all questions have
been addressed.
Attached please find the diff of -16 and the working version and the
copy of the working version of the draft.
Regards,
Greg
On Mon, May 21, 2018 at 5:20 PM, Bob Briscoe <i...@bobbriscoe.net
<mailto:i...@bobbriscoe.net>> wrote:
Reviewer: Bob Briscoe
Review result: Not Ready
Altho this is a TSV-ART review, I did not find many
transport-related issues to
focus on, except a need to justify why lack of information for
adapting the
transmit interval is not an issue.
Nonetheless, I did find a few other non-trivial technical issues,
including 2
security issues, enumerated below (I mis-spent some of my early
research career
working on a multicast session control and security, for which we used
beaconing control channels). However, I only have passing prior
knowledge of
BFD, so my critique might be off-beam.
==Main Technical Concerns===
1/ Mandatory return path?
RFC5880 is the base RFC that this draft updates. RFC5880 says that
"unidirectional links" are in scope, but only as long as there is
a return path.
The introduction of this bfd-multipoint draft seems to contradict
that, making
a return path optional: "
As an option, the tail may notify the head of the lack of
multipoint
connectivity. Details of tail notification to the head are outside
the scope of this document.
"
It's allowable for irrelevant details to be outside the scope, but
surely it
needs to be clear whether at least the existence of a return path
is mandatory.
GIM>> Thank you for highlighting this issue. I think that the second
paragraph of Introduction is the appropriate place to note that this
mechanism does not require existence of a return path from tails to
the head. Would the following be acceptable:
NEW TEXT:
Use of BFD in
Demand mode enables a tail monitor availability of a multipoint path
even without the existence of some kind of a return path to the head.
2/ Mechanism for verifying connectivity, or not?
The introduction seems to contradict itself:
"
As multipoint transmissions are inherently unidirectional, this
mechanism purports only to verify this unidirectional connectivity.
"
"
Term "connectivity" in this document is not being used in the
context
of connectivity verification in transport network but as an
alternative to "continuity", i.e. existence of a forwarding path
between the sender and the receiver.
"
How can this mechanism verify connectivity, but not be used in the
context of
connectivity verification in the transport network?
GIM>> This draft defines the base specification for multipoint BFD. In
order for multipoint BFD to support the transport-like connectivity
verification we need to do work similar to described in RFC 6428.
[BB]: Caveat: I am having to talk in generalizations, cos I don't
actually know how you are going to get this protocol to work in a wide
range of circumstances, given inherent problems like multipoint feedback
implosion {Note 1}.
My point was that, having broken up the drafts in this way, this draft
on its own no longer defined a workable protocol. Therefore, it needed
some references to other drafts (even if they are placeholders), so that
the extent of the pre-requisite collection of work is clear. The refs
you give later go a long way to fixing this issue.
If each pre-requisite protocol is intended to only represent one
example, the citation can say that and the ref can be informative. But
with zero examples for all the prerequisite parts, all the reader sees
is a dismembered octopus, not a protocol.
3/ Use case
The introduction seems to be written rather academically. Surely,
in cases
where there is never a return path, only the tails will ever be
able to verify
connectivity. The head could continue transmitting BFD packets
(and data
packets) for years without ever knowing whether it is connected to
anything.
Knowledge of connectivity is surely of little use if it excludes
the link
sender, which is the node that always controls routing.
If there are scenarios where it is useful for tails but not the
head to be able
to verify connectivity, can you please give a concrete example?
GIM>> One example could be a multicast system with 1+1 protection.
Without multipoint BFD tails would not be able to detect the failure
of the muticast path from the head. Other examples discussed in
several drafts:
* BESS WG draft MVPN fast upstream failover
<https://tools.ietf.org/html/draft-ietf-bess-mvpn-fast-failover-03>
* Individual draft BFD for Multipoint Networks and VRRP Use Case
<https://tools.ietf.org/html/draft-mirsky-bfd-p2mp-vrrp-use-case-01>
* Individual draft BFD for Multipoint Networks and PIM-SM Use Case
<https://tools.ietf.org/html/draft-mirsky-pim-bfd-p2mp-use-case-00>
I am not sure how references to non-WG drafts affect the progress of
this document. Appreciate your suggestion.
[BB]: In my experience, informative refs to non-WG drafts as use-cases
would be OK during doc development. However, if a non-WG draft fails to
proceed, its citation has to be removed later. So choose those that are
most likely to proceed.
Nonetheless, if you cite some specs that turn this into a workable
protocol (see previous issue) use-cases might not be necessary.
4/ Interval adaptation
Text is needed to describe why it is not an issue for the head to
be unaware
whether it needs to adapt its transmit interval. Otherwise, this seems
potentially problematic.
GIM>> Very interesting, thank you. I wouldn't say that the case when a
tail cannot process incoming mpBFD control packets at the offered rate
is entirely non-issue. Such scenario must be handled by the
implementation and may be controlled by local policy, e.g., close the
MultipointTail session.
[BB]: Fair enough.
In some scenarios, this issue will not necessarily be so unlikely tho:
* If asymmetric crypto is used to solve the group message authentication
problem (see later), the processing burden on any lightweight endpoints
might lead to message verification leaving less available processor
resource than needed for the host's other tasks.
* Each tail might be joined to a very large number of multipoint sessions.
Where would you suggest to add the text?
I would suggest a new section listing potential issues when there is no
back channel.
5/ Inability to authenticate the sender with symmetric keys
In unicast scenarios, symmetric keys can be used for message
authentication,
because each end knows there is only one other node with the
shared key. But,
in multipoint scenarios, all the tails would share the key, so a
shared key
does not authenticate who sent the message - any tail can spoof
the head from
the viewpoint of the other tails.
Therefore text is needed to say that:
* multipoint message authentication is limited to cases where all
tails are
trusted not to spoof the head, if shared keys are used. *
otherwise asymmetric
message authentication would be needed, e.g. TESLA [RFC4082]
GIM>> Thank you for the suggested text. Would the Security
Considerations section be appropriate place:
[BB]: Well, the point limits the applicability of the assumption about
security in 5. 'Assumptions', so this would fit well there.
Then "Security Considerations" should point to everywhere in the doc
that discusses security, such as this (to save time for security reviewers).
NEW TEXT:
Use of shared keys to authenticate BFD Control packet in multipoint
scenarios is limited because tail can spoof the head from the
viewpoint of the other tails. Thus, if shared keys are used, all
tails MUST be trusted not to spoof the head.
[BB]: Normally a MUST is applied to implementations. It would be rather
odd to require users/operators to satisfy a spec requirement,
particularly requiring them to trust each other. I think this should be
written as an applicability statement not a normative requirement.
Otherwise, asymmetric
message authentication would be needed, e.g., Timed Efficient Stream
Loss-Tolerant Authentication (TESLA) as described in [RFC4082].
[BB]: If you are going to allow for cases where all tails are trusted
not to spoof the head, then the assumption written in section 5 is no
longer correct.
[FYI, RFC4082 is only a generic description. Many RFCs have been written
to authenticate specific protocols along TESLA lines.]
A related nit: Section 5 says all tails are assumed to have a common
authentication key. In cases with symmetric message
authentication, surely the
head also needs the same key.
GIM>> Thank you. Please check the updated text:
NEW TEXT:
If authentication is in use, the head and all tails must be
configured to have a common authentication key in order for the tails
to validate received the multipoint BFD Control packets.
[BB]: Yup. Except delete "received the".
Also see above about whether this is now a correct assumption.
6/ Source address spoofing
A 3-way handshake makes a protocol robust against simple source
address
spoofing. Without a 3WHS, surely the spec. needs to highlight this
vulnerability or discuss ways to address it or why it is not an issue.
GIM>> Because mpBFD control packets cannot be demultiplexed by tail
based on the value of Your Discriminator field as per RFC 5880,
the new procedure outlined in Section 4.7:
IP and MPLS multipoint tails MUST demultiplex BFD packets based on a
combination of the source address, My Discriminator and the identity
of the multipoint tree which the Multipoint BFD Control packet was
received from. Together they uniquely identify the head of the
multipoint path.
and described in details in Section 4.13.2:
If the Multipoint (M) bit is set
If the Your Discriminator field is nonzero, the packet MUST be
discarded.
Select a session as based on source address, My Discriminator
and the identity of the multipoint tree which the Multipoint
BFD Control packet was received. If a session is found, and
bfd.SessionType is not MultipointTail, the packet MUST be
discarded. If a session is not found, a new session of type
MultipointTail MAY be created, or the packet MAY be discarded.
This choice is outside the scope of this specification.
Would you suggest additional text to a use case where the new
demultiplexing is not sufficent to protect from source address spoofing?
[BB]: I seem to have become co-opted into redesigning this protocol. I'd
prefer to limit my involvement to reviewing :)
7/ Scope
On eight occasions an issue is raised, but resolution is stated as
outside the
scope of this document. It is OK to limit the scope of a spec, for
example to
allow for multiple solutions to each issue. But at least one
solution must
already exist for each of these eight issues. So, at least one
example solution
ought to be cited in each case. If any issues are open, then this
should not be
on the standards track.
It would be more useful to state why each issue is out of scope.
This would be
helped by stating from the start what the scope of the document is.
GIM>> I've listed all eight occasions with the explanation for each one:
1. Details of tail notification to the head are outside the scope of
this document. - Notifications by tails addressed in
draft-ietf-bfd-multipoint-active-tail
<https://datatracker.ietf.org/doc/draft-ietf-bfd-multipoint-active-tail/>.
Will add as informational reference.
[BB]: Good.
Nonetheless, given you have confirmed that a reverse path is optional,
the doc still needs to address the case where there is no reverse path.
{Note 1} For the active tail draft, you might find the following ideas
for scaling multipoint feedback useful:
*Statistical feedback:*
Nonnenmacher, Jö. & Biersack, E.W., "Scalable Feedback for Large Groups
<https://dl.acm.org/citation.cfm?id=312251>," IEEE/ACM Transactions on
Networking 7(3):375--386 (June 1999)
FUHRMANN, T., AND WIDMER, J. "On the scaling of feedback algorithms for
very large multicast groups
<https://dl.acm.org/citation.cfm?id=2294709>," Computer Communications
24, 5-6 (March 2001), 539 547;
WIDMER, J., AND FUHRMANN, T. Extremum feedback for very large multicast
groups. Tech. Rep. TR 12-2001, Prakfische Informatik IV, University of
Mannheim, Germany, May 2001.
Also, anycast can be used to select different representative feedback
tails, e.g. for a certain time, which might overlap with the periods for
which a few other tails are selected using subsequent anycasts.
*Logical 'AND' feedback:*
Burbridge, T., Soppera, A., Briscoe, R. and Jacquet, A. "Method and
device for co-ordinating networked group members
<https://worldwide.espacenet.com/publicationDetails/biblio?II=0&ND=3&adjacent=true&locale=en_EP&FT=D&date=20060406&CC=US&NR=2006075022A1&KC=A1#>"
Patent WO2004059479, (Jul 2004; Priority 24 Dec 2002)
[AFAICT this patent is still being maintained, so use of it in a
protocol would require an IPR declaration.]
1. Details of how the head keeps track of tails and how tails alert
their connectivity to the head are outside scope of this document.
- Same as #1.
[BB]: And my response is same as #1.
1. Bootstrapping BFD session to multipoint MPLS LSP in case of
penultimate hop popping is outside the scope of this document. -
It may use control plane as in MVPN draft. Will add as
informational reference.
[BB]: Good.
1. Use of other types of encapsulation of the BFD control message
over multipoint LSP is outside the scope of this document. - This
in reference to ACH encapsulation that is discussed in
draft-mirsky-mpls-p2mp-bfd
<https://tools.ietf.org/html/draft-mirsky-mpls-p2mp-bfd-03>.
Should it be added as informational reference? What would be the
imacpt of progressing this specification?
[BB]: See earlier comment about citing individual drafts (I don't have
enough BFD knowledge to give a BFD-specific answer).
Also, in my review I should also have said: when creating new
encapsulations, pls see the common transport issues related to
encapsulation:
https://trac.ietf.org/trac/tsv/wiki/tsvdir-common-issues#TunnelingprotocolsandTransportRelatedIssues
1. Change in the value of bfd.RequiredMinRxInterval is outside the
scope of this document. - Same as #1.
[BB]: And my response is same as #1.
1. If a session is not found, a new session of type MultipointTail
MAY be created, or the packet MAY be discarded. This choice is
outside the scope of this specification. - I propose to add "based
on local policy" to the last sentence.
[BB]: On what basis will local policy decide? It's my job as a reviewer
to ensure that this spec does not contain any loose ends (open issues).
1. The exact method of selection is application-specific and is thus
outside the scope of this specification. - This is copied from RFC
5880: "The exact method of selection is application specific and
is thus outside the scope of this specification." as the section
is to replace Section 6.8.6.
[BB]: OK.
1. If a matching session is not found, a new session of type
PointToPoint MAY be created, or the packet MAY be discarded. This
choice is outside the scope of this specification. - Same as #6.
[BB]: And my response is same as #6.
[Sry, my embedded comments have broken your numbered list.]
There is also one issue that is "for further discussion". Does
this mean the
document is not ready yet?
GIM>> I think that the question left for further discussion is
non-technical:
The semantic difference between Down and AdminDown state is for
further discussion.
I propose to remove the sentence altogether.
[BB]: OK.
8/ Incremental deployment
Section 4.4.1. "New State Variable Values" defines bfd.SessionType =
PointToPoint as well as a couple of Multipoint alternatives.
Presumably this
spec does not require all existing PointToPoint systems to support
this state
value. Is the implication that only Multipoint systems that happen
to be in
PointToPoint mode should use this state?
GIM>> You're aboultely right, existing implementations of BFD don't
need to support bfd.SessionType variable. Only implementations that
support the base BFD, single-hop or multi-hop, and this specification,
mpBFD, should support bfd.SessionType and set it to PointToPoint value
when BFD is in single-hop or multi-hop mode. When in mpBFD mode,
bfd.SessionType will be set to either MultipointHead or MultipointClient.
[BB]: Doesn't something need to be written (or referenced) to clarify
all this? AFAIR, this spec. never made clear that multipoint is a fork
in implementations.
==Nits==
* Sometimes 'tree' is used to mean a multipoint path in general. I
suspect
'path' was intended.
GIM>> I've found six occasions of "tree" and s/tree/path/
4.8. Packet consumption on tails
s/Node/Nodes/
s/packet/packets/
s/demultiplex/demultiplexed/
GIM>> Accepted all three.
4.9. Bringing Up and Shutting Down Multipoint BFD Service
"
a newly
started head (that does not have any previous state information
available) SHOULD start with...
"
...
"
... (so long as the restarted head
is using the same or a larger value of bfd.DesiredMinTxInterval
than
it did previously).
"
If it has no state available, how can it know whether a value is
larger than
previously?
GIM>> You are right, the BFD system at the head would not know the
previous value of bfd.DesiredMinTxInterval. This text is to caution
operator from decreasing bfd.DesiredMinTxInterval upon restart of the
BFD system.
4.9. Bringing Up and Shutting Down Multipoint BFD Service
There are a number of "SHOULD"s and "SHOULD NOT"s that do not
state or give
examples of circumstances in which the "SHOULD" would not be
appropriate. If
there are none, "MUST" would be more appropriate.
GIM>> In the first paragraph SHOULD may not be followed if the
implementation can differentiate between the very first start and
restarts of BFD system. If it is the first start of BFD system, the
head MAY directly progress to Up state skipping Down state.
The last paragraph describes graceful shuttdown. The head MAY shut the
BFD mp session abruptly by just stopping transmission of BFD Control
packets.
[BB]: I assume you will say all this in the next rev, not just in this
email.
4.10. Timer Manipulation
"
Because of the one-to-many mapping, a session of type
MultipointHead
SHOULD NOT initiate a Poll Sequence in conjunction with timer value
changes. However, to indicate a change in the packets,
MultipointHead session MUST send packets with the P bit set.
MultipointTail session MUST NOT reply if the packet has M and P
bits
set and bfd.RequiredMinRxInterval set to 0.
"
The initial "SHOULD NOT" needs to be written another way. Perhaps
"
...a session of type MultipointHead
does not initiate a Poll Sequence
"
The head's normative action is defined by the following "MUST",
then the tail's
"MUST NOT reply" is what prevents the poll sequence happening.
GIM>> A Poll Sequence starts with the initiator setting Poll bit.
Unless the peer sends BFD Control packet with Finl bit set the poll
sequence would continue indefinetely. The initial SHOULD NOT, in my
opinion, correctly points that the mechanism of Poll Sequence not to
be used by MultipointHead when changing transmission interval. I think
that MUST in the first paragraph can be downgraded to MAY because the
MultipointHead doesn't need to use transition period when changing the
transmission interval to lower level, i.e., decreasing frequency. May
I propose the following:
OLD TEXT:
Because of the one-to-many mapping, a session of type MultipointHead
SHOULD NOT initiate a Poll Sequence in conjunction with timer value
changes. However, to indicate a change in the packets,
MultipointHead session MUST send packets with the P bit set.
NEW TEXT:
Because of the one-to-many mapping, a session of type MultipointHead
SHOULD NOT initiate a Poll Sequence in conjunction with timer value
changes. However, to indicate a change in the packets,
MultipointHead session MAY send packets with the P bit set during
transition period.
[BB]: If I were an implementer, I would not know what this is saying I
ought to implement. The spec needs to answer this question: If the head
changes the packets what happens differently if it sets the P bit vs. if
it doesn't?
4.11. Detection Times
Delete "in the calculation" (repetition).
GIM>> Done.
4.13.1. Reception of BFD Control Packets
Some actions seem to be only relevant to PointToPoint sessions,
but they are
stated for all session types. Specifically: "the transmission of
Echo packets,
if any, MUST cease." "the Poll Sequence MUST be terminated." "MUST
cease the
periodic transmission of BFD Control packets" "MUST send periodic
BFD Control
packets"
"
If bfd.SessionType is PointToPoint, update the Detection Time as
described in section 6.8.4 of [RFC5880]. If bfd.SessionType is
MultipointTail,
"
The second sentence above ought to start on a new line as an Else if.
GIM>> Hope I got it right:
If bfd.SessionType is PointToPoint, update the Detection Time as
described in section 6.8.4 of [RFC5880].
Else
If bfd.SessionType is MultipointTail, then update the Detection
Time as the product of the last received values of Desired Min
TX Interval and Detect Mult, as described in Section 5.11 of
this specification.
4.13.2. Demultiplexing BFD Control Packets
"
This section is part of the replacement for [RFC5880] section
6.8.6,
separated for clarity.
"
Do you mean "This section replaces the sentence: "If the
Multipoint (M) bit is
nonzero, the packet MUST be discarded." in [RFC5880] section 6.8.6?
The statements under "If the Multipoint (M) bit is set" are not
formatted like
the rest of the if-else logic, and I think an Else is missing at
the start of
the statement after the nested "If".
GIM>> Agree, the paragraph is not structured properly. How about this
formating:
If the Multipoint (M) bit is set
If the Your Discriminator field is nonzero, the packet MUST be
discarded.
Select a session as based on source address, My Discriminator
and the identity of the multipoint path which the Multipoint
BFD Control packet was received.
If a session is found, and bfd.SessionType is not
MultipointTail, the packet MUST be discarded.
Else
If a session is not found, a new session of type
MultipointTail MAY be created, or the packet MAY be
discarded. This choice is outside the scope of this
specification.
[BB]: As long as this represents the logic you want, fine. The point is
that the indentation is the only clue to whether one 'if' is conditional
on a previous 'if', or not.
HTH
Bob
--
________________________________________________________________
Bob Briscoe http://bobbriscoe.net/
