Hi Alvaro
I'm sorry that it has taken longer than I thought to reply to your comments!
Please find our replies below. I will post an updated version of the document
as soon as I can.
Many thanks
Jon
<snip>
----------------------------------------------------------------------
DISCUSS:
----------------------------------------------------------------------
I am balloting DISCUSS because I think that there are some technical and
clarity issues that makes understanding, and potentially implementing this
document hard. I also want to discuss the "backwards compatibility" and the
use of TLVs as sub-TLVs in PCEP as introduced in this document.
(1) MSD Definition. The MSD may be learned from a variety of sources,
including the SR-PCE-CAPABILITY sub-TLV defined in this document. It is
important then for the MSD to be defined consistently everywhere. Please use
the BMI-MSD definition from rfc8491.
[Jon] Happy to change this. This draft actually pre-dates RFC 8491, but now
the definition has moved there, we can point to it.
(2) Ability to signal the MSD per link, not just per node. Clearly the
calculation of paths through specific links (using an Adjacency SID, for
example) would require that information (if different/lower from what the node
may support).
Note that §6.1 seems to assume that the MSD will normally be advertised through
different mechanisms, and it uses that to work around the fact that there's no
link-specific information: "Furthermore, whenever a PCE learns the MSD for a
link via different means, it MUST use that value for that link regardless of
the MSD value exchanged in the SR-PCE-CAPABILITY sub-TLV." However, the text
doesn't mandate the IGP/BGP-LS information to be available to the PCE. IOW, as
written, the specification can't guarantee the proper calculation of paths that
require the PCE to consider link MSDs.
[Jon] In many deployments we anticipate that link MSDs are homogeneous. In
those cases, link state routing might not distribute per-link MSDs (e.g.
routers might not even support RFC 8491). In such cases, the per-node MSD on
the PCEP session is sufficient. All the draft says is that MSD information
available from link state routing, if any, must take priority over that defined
on the PCEP session. I don't see a problem with that.
(3) Extensibility to advertise other MSD-Types. [This point is not
DISCUSS-worthy, but I'm including it here since I'm already talking about the
MSD.]
rfc8491 (aka I-D.ietf-isis-segment-routing-msd) and
I-D.ietf-ospf-segment-routing-msd encode the MSD advertisement as a pair:
MSD-Type and MSD-Value, with the expectation that "new MSD-Types will be
defined to signal additional capabilities, e.g., entropy labels, SIDs that can
be imposed through recirculation, or SIDs associated with another data plane
such as IPv6." IOW, the encoding is reusable with other dataplanes. I peeked
into draft-negi-pce-segment-routing-ipv6 [*] and i don't see anything in there
that couldn't be signaled using the SR-PCE-CAPABILITY sub-TLV defined here (+
the MSD_Value). I think this is important for consistency.
[*] I realize that draft-negi-pce-segment-routing-ipv6 is not even a WG
document, but it is the only potential reference I found to what a different
dataplane might look line.
[Jon] This document (and the SR-PCE-CAPABILITY) is scoped to MPLS only. I
believe that draft-negi defines its own SRV6-PCE-CAPABILITY TLV and I don't see
any reference to MSD in it, but if a new MSD type is needed for other dataplane
types, I think it's expected that a new SR capability TLV will be defined to
convey it. I don't expect to generalize the SR-PCE-CAPABILITY TLV.
Note that the MSD in the SR-PCE-CAPABILITY is a BMI-MSD. Although RFC 8491
defines a generic MSD container, the MSD in this document is specifically a
BMI-MSD.
(4) §6.2.2 (Interpreting the SR-ERO):
o If the subobjects contain NAI only, then the PCC first converts
each NAI into a SID index value by looking it up in its local
database, and then proceeds as above.
I believe that this step in the interpretation of the SR-ERO is not properly
specified.
Which "local database" are you referring to? §6.2.2.1 mentions the SR-DB (when
talking about errors)...but the specification should be clear about which
database and what the specific procedure is.
[Jon] You are right, this should be more explicit. How about this.
NEW
If the subobjects contain NAI only, the PCC first converts
each NAI into a SID index value and then proceeds as above.
To convert an NAI to a SID index, the PCC looks for a fully-specified
prefix or adjacency matching the fields in the NAI. If the PCC finds
a matching prefix/adjacency, and the matching prefix/adjacency has a SID
associated
with it, then the PCC uses that SID. If the PCC cannot find a
matching prefix/adjacency, or if the matching prefix/adjacency has no SID
associated
with it, the PCC behaves as specified in section 6.2.2.1.
END NEW
For example, what is the specific process that the PCC needs to follow to
convert a Node ID/IP address to the SID/MPLS label? What if the SR-DB doesn't
contain an SID associated to the specific Node ID/IP address? How should the
router react to that? This case is not covered in the Error Handling section
(6.2.2.1) either.
[Jon] This is specified in 6.2.2.1. First bullet - if the prefix is found in
the SR-DB but has no SID, send error TBD3. Second bullet - if the NAI is not
found in the SR-DB, send error TBD4.
A pointer to the SR-DB (as defined in I-D.ietf-spring-segment-routing-policy)
is not enough because it is composed of optional information (according to the
description in §3 (Segment Routing Database)). This document should be
specific about what information must be contained in the SR-DB for the
conversion to be successful.
[Jon] Hopefully the new text proposed above will do that.
The requirement of the information to be contained in the SR-DB makes
I-D.ietf-spring-segment-routing-policy a Normative reference.
[Jon] Rather than add an extra normative dependency, we would prefer to remove
the dependency on the definition of the SR-DB and instead explicitly define in
our document what information is to searched.
(5) §7 (Backward Compatibility) "Some implementations, which are compliant
with an earlier version of this specification..." <sigh>
I understand that there may be implementations that are non-compliant with this
specification out in the field. However, why is this document making
accommodations for them? Not only are these implementations not compliant with
this document, but are also non compliant with rfc8408, which implies the use
of a new sub-TLV per PST.
I didn't find a discussion on the mailing list related to this issue.
Specifying alternate behavior to accommodate non-compliant implementations is
not the best way to define new functionality. If the support for those old
implementations was an imperative then the new functionality should have been
fully specified to seamlessly interoperate with what is already deployed. The
current result is two ways to do the same thing...
While I would prefer for this "backwards compatibility" not to be built into
the specification, I am looking for discussion in the WG and a better
justification that the current one (which can be reduced to "non-compliant
implementations exist, so we need to fit them in here somehow").
[Jon] Yes, this section was painful to write, and was done after an on-list
consultation with the working group. I can provide some references. The
relevant thread starts here.
https://www.ietf.org/mail-archive/web/pce/current/msg05397.html
We got firm feedback from one vendor that the old behaviour needed to be
accommodated.
https://www.ietf.org/mail-archive/web/pce/current/msg05415.html
There were good reasons for us changing the spec at a late stage, but this
unfortunately left us needing to make a special case of the SR-CAPABILITY-TLV
or else break a fielded implementation. So we collectively held our noses and
did it. Hopefully, this plus the thread above gives you the background to the
decision.
(6) sub-TLV Space for the PATH-SETUP-TYPE-CAPABILITY TLV
rfc8408 failed to set up a sub-TLV registry for the PATH-SETUP-TYPE-CAPABILITY
TLV. The bigger issue is that it also doesn't say that other PCE TLVs can be
used as sub-TLVs (nor does it prohibit that). The Type for the
SR-PCE-CAPABILITY sub-TLV is allocated from the PCEP TLV Type Indicators
registry, making it a TLV. I also couldn't find any mention of sub-TLVs in
rfc5440, or the potential intent to have a single space from which both TLVs
and sub-TLVs could come.
The question is: are all TLVs (defined in the PCEP TLV Type Indicators
registry) able to be used as sub-TLVs? This question is general, but also
specific to the PATH-SETUP-TYPE-CAPABILITY TLV. At a minimum, it should be
made clear which can be used with the PATH-SETUP-TYPE-CAPABILITY TLV -- because
this is the first document to define a new PST and sub-TLV, it seems
appropriate to Update rfc8408 here...but rfc5440 may also need an Update.
[Jon] I don't think there is an intent that any TLV can be used as a sub-TLV.
This argues for making a new registry for the sub-TLVs of the
PATH-SETUP-TYPE-CAPABILITY TLV (although we will have a vestigial code point
allocation for SR-PCE-CAPABILITY as a top-level TLV because of section 7). I
think therefore that this draft should create the registry and update RFC 8408.
----------------------------------------------------------------------
COMMENT:
----------------------------------------------------------------------
These comments don't raise to the level of a DISCUSS, but I would like to also
see them addressed.
(1) [nit] "Both node segments and adjacency segments can be used for SR Traffic
Engineering (SR-TE)." I find the use of SR-TE (instead of simply SR) gratuitous
and potentially confusing; it introduces a new term which doesn't represent new
functionality as compared to exiting segment routing documents.
[Jon] OK
(2) "This document is relevant to the MPLS forwarding plane only." I believe
that I-D.ietf-spring-segment-routing-mpls should be a Normative reference.
[Jon] OK
(3) In §3, the first two paragraphs have redundant text:
"In an SR network, the ingress node of an SR path prepends an SR header to all
outgoing packets. The SR header consists of a list of SIDs (or MPLS labels in
the context of this document)....In SR networks, an ingress node of an SR path
prepends an SR header to all outgoing packets. The SR header consists of a
list of SIDs (or MPLS labels in the context of this document)."
[Jon] Oops - I will delete the duplicate text from the 2nd para.
(4) §3: "...the PCEP messages (e.g., Path Computation Request, Path Computation
Reply, Path Computation Report, Path Computation Update, Path Computation
Initiate, etc.,) MUST be formatted according to [RFC5440], [RFC8231],
[RFC8281], and any other applicable PCEP specifications." I'm not sure what
behavior is being specified here -- IOW, why do we need Normative language?
This document defines the extensions referred to here, so the format should
already be compliant with the RFCs mentioned. s/MUST/must
[Jon] OK: "...the PCEP messages [...] are formatted according to..."
(5) Following up from the last point... §4 seems to address that MUST by
saying that there's no requirement for "changes in the format of the PCReq and
PCRep messages specified in [RFC5440], PCInitiate message specified in
[RFC8281], and PCRpt and PCUpd messages specified in [RFC8231]." I find this
section unnecessary.
[Jon] Agreed - will remove it.
(6) [nit] §5.3.1 defines the "L Flag"... §6.1, for example, uses "L flag" to
refer to the L bit (§5.1.1). Please try to be consistent to avoid
confusion...or even better, use a different letter.
[Jon] Yes, unfortunate that they have the same name. I'll clarify it.
(7) §5.1.1 says that a "PCEP speaker SHOULD indicate its support of the
function described in this document by sending a PATH-SETUP-TYPE-CAPABILITY
TLV...[and]...MUST also include the SR-PCE-CAPABILITY sub-TLV"...but §6.1 then
says that "if a PCE receives an SR-PCE-CAPABILITY sub-TLV with the L flag and
MSD both set to zero then it MUST assume that the PCC is not capable of
imposing a SID stack of any depth and hence is not SR-TE capable". Wait, the
sub-TLV is included because the TLV says that it supports SR. Isn't this then
a contradiction?? What good is it to signal support if the node is "not
capable of imposing a SID stack of any depth"? Shouldn't this combination
result in an error message?
[Jon] I can't think of any legitimate reason to signal that, so yes, this
should probably be an error case instead. I'll update the draft.
(8) §6.2.2 "According to [I-D.ietf-spring-segment-routing-policy], each SR-ERO
subobject in the sequence identifies a segment that the traffic will be
directed to, in the order given."
The SR-ERO subobject is defined in this document, so its interpretation is of
obvious importance. Because of that, I think that the text above makes the
reference Normative.
However, I looked in I-D.ietf-spring-segment-routing-policy and I find no
mention of the SR-ERO, ERO, or sequence. The only related text (that I could
find) is the generic one about SR being an "ordered list of segments"...so I
think that the reference to I-D.ietf-spring-segment-routing-policy is out of
place.
Suggestion: replace the reference to I-D.ietf-spring-segment-routing-policy
with a reference to rfc8402.
[Jon] On reflection I'm not sure we need any reference for this fairly
straightforward statement. I would amend it to:
NEW
The SR-ERO contains a sequence of subobjects. Each SR-ERO subobject in
the sequence identifies a segment that the traffic will be directed
to, in the order given. That is, the first subobject identifies the
first segment the traffic will be directed to, the second
subobject represents the second segment, and so on.
END NEW
(9) §6.2.2 "If the subobjects contain SID index values, then the PCC converts
them into the corresponding MPLS labels by following the procedure defined in
[I-D.ietf-spring-segment-routing-mpls]." I think this statement requires
I-D.ietf-spring-segment-routing-mpls to be a Normative reference.
[Jon] OK
(10) §6.2.2 Only the third procedure ends with "...and then directs the packet
to the segment identified by the first SID", which is the obvious next step,
but the text is only talking about the conversion. Either make sure that it is
clear that all the processes continue with sending, or take this piece of text
out. Be consistent.
[Jon] I'll remove it from the final bullet and then add this after the bullets:
"For all cases above, after the PCC has imposed the label stack on the packet,
it sends the packet to the segment identified by the first SID. "
(11) §5.5 "...the PCE MUST minimize the SID depth of the computed path." If
the B bit is not set (meaning not bound), what behavior is this phrase
standardizing? Given that we're not standardizing the computation done by the
PCE, how can it be enforced?
[Jon] I think Martin pointed this out too. I think the normative language is
inappropriate here. Instead: "... the PCE minimizes the SID depth..." The
MUST can be used only in conjunction with B=1 since only then can the PCE's
behaviour be enforced.
(12) §8.1 (Controlling the Path Setup Type) I find this section out of place
in this document. rfc8408 is the document that specifies the support for
multiple path setup methods...while this document adds the SR-related type. If
kept, then I think this document should be tagged to Update rfc8408.
[Jon] I agree that most of this text is generic and could have been written in
RFC 8408. I think we have already agreed to update RFC 8408, above, so there
is presumably nothing more to do here.
_______________________________________________
Pce mailing list
Pce@ietf.org
https://www.ietf.org/mailman/listinfo/pce
