Re: [mpls] Updated rev -08 posted [Re: Benjamin Kaduk's Discuss on draft-ietf-mpls-ldp-yang-07: (with DISCUSS and COMMENT)]

[Benjamin Kaduk <[e-mail]>]

On Fri, Feb 28, 2020 at 04:39:14AM +0000, Kamran Raza (skraza) wrote:
> Thanks for your detailed review, Ben. 
> We have posted a new rev -08 that takes care almost all of your comments and comments from other IESG reviewers. 
> On behalf of authors, Please see inline [skraza] for each of your comment/Q:
> 
>  
>     ----------------------------------------------------------------------
>     DISCUSS:
>     ----------------------------------------------------------------------
>     
>     The YANG doctor last call review of the -06 notes that these modules are
>     in violation of the expectations of RFC 8349 (at least w.r.t. defining a
>     new identity for the control-plane protocol and possibly more).  This
>     document should either be brought into compliance or explain why it
>     diverges.
> 
> [skraza]: Fixed.  
>     
>     The YANG doctor's comments about default values for "local"
>     configuration (e.g., graceful-restart configuration) causing the global
>     configuration to never take effect should also be addressed.
> 
> [skraza]: Fixed. 

These both look addressed, to my eye, but I am very much not a YANG expert.
I'd recommend trying to get a dedicated YANG review, perhaps from the
incoming Management AD Rob Wilton.

>     Please include some justification for why LDP IPv6 is considered an
>     "extended feature" (which is particularly surprising given that Section
>     2 classifies "IP" to refer to both IPv4 and IPv6 together).
>     
> [skraza]:  This was an early decision taken and presented to WG and WG had agreed.
> Moreover, Sec 1.1 does points out why LDP IPv6 is considered as an extended feature. The "base" features are the one most commonly implemented and deployed features.
> The idea was that base must not contain any if-feature - given that LDPv6 is not widely implemented and was branded as if-feature, it is classified in the extended category then.
> BTW, note that LDP has separate RFC for LDP IPv6. 

It's not really clear to me that just because the protocol is described in
separate RFCs for IPv4 and IPv6 we cannot present a unified management
interface on top of it.  I note that several other ADs expressed concern
about this topic, so I've moved this document onto the agenda for this
week's IESG telechat to get a better sense on what the right thing to do is
here.

> 
>     We need to define the format/semantics of the md5-key string (e.g., is
>     it hex? base64{url,}?) either directly or by reference (as the YANG
>     doctor notes).  Using a crypto-type would probably be appropriate, as
>     would adding a note that tcp-md5 is obsoleted by TCP-AO.
>     
> [skraza]: It is beyond our control as we are using it as-is from the RFC. Made the change anyways to add crypto-algo to widen the scope/type of key beyond MD5.

I'm still concerned that there isn't enough text here for me to be able to
implement something and have confidence it will interoperate.  For example,
the recommendations in RFC 3562 include the ability to use a direct binary
key for TCP-md5.  Does this YANG stanza allow for doing so?  If yes, what
mechanism is used for converting input into "raw" binary, if needed?

That is, even if the actual protocol semantics for how TCP-MD5 uses a given
key/password are out of your control, there is still ambiguity as to what
mapping (if any) occurs between the YANG-configured value and what is
passed to the TCP-MD5 implementation.

Furthermore, the authentication-type still only has a "string" leaf for the
key, with (if I am reading the YANG correctly) the recommended key-chain
functionality for protected key storage being left as an extension.  It's
not clear to me why we should be leaving the best-practice formulation as
an extension.

>     In the peer-af-policy-container grouping's
>     label-policy/advertise/prefix-list, we need to say if the filter is for
>     inclusion or exclusion of outgoing label advertisements.
>     Similarly for the incoming label advertisements in the 'accept'
>     container (and most of the <foo>-list-ref usages?).
>     
>     In the policy-container grouping's
>     label-policy/assign/independent-mode/prefix-list, the description
>     suggests that the contents will provide not just a list of prefixes that
>     act as a filter, but also a map from prefix to label.  This is a
>     qualitatively different usage than the previous occurrences of
>     prefix-list-ref, and it seems like a different type may be needed for
>     it.  Similarly for ordered-mode.
>     
> [skraza]: Fixed as per latest changes that use rtgwg policy model constructs.

Ah, this took me a little more work to find, since the changes were in the
prefix-list-ref typedef.  Thanks.

>     (pro-forma) This document has 6 authors, and per RFC 7322, as this is
>     more than five, we are supposed to consider whether that's appropriate.
>     Seeing nothing in the shepherd writeup or similar, I mention it here.
>     
> [skraza]: We had taken an advise from Loa (MPLS WG chair) and would like to keep it to 6.

Sure; this was presumed to just be "pro forma" but I put it in since the
shepherd writeup didn't mention it specifically.

>     ----------------------------------------------------------------------
>     COMMENT:
>     ----------------------------------------------------------------------
>     
>     Please also respond to the other YANG doctor, rtg-dir, and other
>     directorate reviewers' comments.
>     
>     I support Roman's Discuss.
>     
>     I suggest making a reference to RFC 5920 at some point, perhaps from the
>     security considerations.
> [skraza]: Done.
>     
>     We have several instances of a "container ipv4" that's a "container with
>     presence" in the jargon of RFC 7950.  As SEC AD one thing I look for is
>     edge cases, such as when a subset of these containers are present but
>     not all (and not none).  While for this specific case, it seems fairly
>     natural to assume that IPv4 is enabled if any are present, it does make
>     me wonder if the modeling is at an optimal state if there's even the
>     possibility for such skewed signals.  (I didn't specifically audit for
>     any other cases of "redundant" containers-with-presence.)
>     
> [skraza]: This is the same structure as RFC 8344. The edge cases are OK. The ipv4/ipv6 container may contain none, one, or more nodes in the container. ipv4/ipv6 address family is enabled as long as the container is “present”. Your assumption is correct, and the modeling (and the protocol as well) is in a good state, even with such possibilities. The signals should be clear enough.
> 
>     It's not entirely clear to me why we need (all of) the specialization of
>     structures into using inet:ipv4-address and inet:ipv6-address as opposed
>     to just inet:ip-address.  I do see that the latter is used in several
>     places already, and so assume that this was already considered; perhaps
>     some of the discussion/motivation could be included in the document,
>     though.
>     
> [skraza]: When the type of an attribute is for a specific address family (ipv4 or ipv6), we use the address family specific type (inet:ipv4-address or inet:ipv6-address), so that user cannot provide a value of the wrong format.  inet:ip-address is simply a union of net:ipv4-address and inet:ipv6-address, and is used when the address family is not restricted. This is particular convenient when the user makes a query on a particular peer when the user does not have a particular address family in mind. The system will provide the address in a proper format.
> 
>     Section 1
>     
>        The data model is defined for following constructs that are used for
>        managing the protocol:
>     
>     nit: "the following".
> [skraza]: Fixed.    
> 
>     I share the YANG doctor's concern that separating the configuration and
>     operational state constructs is not consistent with the NMDA model.
> [skraza]: Fixed.
>     
>     Section 1.1
>     
>        The "base" category contains the basic and fundamental features that
>        are covered in LDP base specification [RFC5036] and constitute the
>        minumum requirements for a typical base LDP deployment.  Whereas, the
>        "extended" category contains all other non-base features.  All the
>     
>     This statement ("all other") is not future-proof.
> [skraza]:  Removed the "all".
>     
>     Section 3
>     
>        o  This module aims to address only the core LDP parameters as per
>           RFC specification, as well as some widely deployed non-RFC
>           features (such as label policies, session authentication etc).
>           Any vendor specific feature should be defined in a vendor-specific
>           augmentation of this model.
>     
>     [Even for widely-deployed non-RFC features we still need to provide a
>     clear description of what semantics are being covered, whether
>     in-document or via an external reference.]
>     
> [skraza]: The only non-RFC config was related to label policies - Reworked/enhanced the text.
> 
>        o  This model is a VPN Forwarding and Routing (VRF)-centric model.
>           It is important to note that [RFC4364] defines VRF tables and
>     
>     side note: I would like the IETF to someday get to a point where "VPN"
>     (the 'P' is for "private") is used only for protocols/deployments that
>     provide confidentiality protection ("privacy") for the data being
>     conveyed.  We are, of course, not there now, and I cannot insist on any
>     change here.  But please consider whether an alternate phrasing would
>     suffice, that includes just the "virtual" part but not "VPN".  Since we
>     are mostly using "VRF" in the rest of the document, this is not as
>     daunting as it sometimes is...
>     
>     Section 4
>     
>     It might be nice to have the tree show the ipv4 and ipv6 children in
>     as similar an order as is possible (with the understanding that the
>     semantics are not affected by the order of appearance, of course).
>     
>            |  +--rw address-families
>            |  |  +--rw ipv4!
>            |  |  |  +--rw enable?                           boolean
>            |  |  |  +--ro label-distribution-controlmode?   enumeration
>            |  |  |  +--ro bindings
>            |  |  |  |  +--ro address* [address]
>            |  |  |  |  |  +--ro address               inet:ipv4-address
>            |  |  |  |  |  +--ro advertisement-type?   advertised-received
>            |  |  |  |  |  +--ro peer
>            |  |  |  |  |     +--ro lsr-id?           leafref
>            |  |  |  |  |     +--ro label-space-id?   leafref
>     
>     Does this imply that there is only one peer per address?  Is that going
>     to be a limitation for any deployment cases?
> 
> [skraza]: Not an issue. Note that a peer may have one or more bound addresses. 
>     
>               +--rw ldp-ext:session-downstream-on-demand
>               |       {session-downstream-on-demand-config}?
>               |  +--rw ldp-ext:enable?      boolean
>     
>     nit(?): missing '|' to connect down to ldp-ext:enable?
>     Similarly for ldp-ext:max-wait a few lines later, and some other
>     occurrences later in the document.
> 
> 
> [skraza]: Some of these instances/typos were due to hand picked sub-tree. 
> Not an issue anymore as we are printing the entire tree as generated by the tool.
> 
>     
>     Section 5.1.2
>     
>     Am I reading this right that (e.g.) per-interface hello-holdtime is
>     configurable as an extension and only the global hello-holdtime is
>     present in the "base" module?
> 
> [skraza]: Yes.
>     
>     Section 7
>     
>     "NHLFE" is used only once in this document (and is not listed as
>     "well-known" at
>     https://www.rfc-editor.org/materials/abbrev.expansion.txt); please
>     expand it out.
>     
> [skraza]: It is well known and listed in the above url. Please x-chk.

To be clear, the linked content uses the convention of an "*" marking to
indicate that the expansion is "well-known" and does not need to be
provided in a given document.  The entry for "Next Hop Label Forwarding
Entry" does not include such a marking; perhaps the responsible AD should
request it to be added, if it is indeed well-known?

>     Section 9
>     
>     Is 2018 still the right copyright year for the modules themselves?
> 
> 
> [skraza]: 2020-02-25 rev now.
> 
>     
>     Section 9.1
>     
>          typedef ldp-address-family {
>            type identityref {
>              base rt:address-family;
>            }
>     
>     Why can't we use rt:address-family directly in the one container that
>     uses this?
>     
> [skraza]: Fixed now.
> 
>            description
>              "Duration represented as 32 bit seconds with infinite.";
>     
>     nits: "32-bit" (hyphenated), "with infinity".
> [skraza]:  Fixed.    
> 
>          typedef downstream-upstream {
>            type enumeration {
>              enum downstream {
>                description "Downstream information.";
>              }
>              enum upstream {
>                description "Upstream information.";
>              }
>            }
>            description
>              "Received or advertised.";
>          }
>     
>     nit: copy/paste on the description?
> [skraza]:  Fixed.        
> 
>              leaf used-in-forwarding {
>                type boolean;
>                description
>                  "'true' if the lable is used in forwarding.";
>     
>     nit: s/lable/label/
> [skraza]:  Fixed.        
>     
>          grouping graceful-restart-attributes-per-peer {
>            description
>              "Per peer graceful restart attributes.
>               On the local side, these attributes are configuration and
>               operational state data. One the peer side, these attributes
>               are operational state data reveiced from the peer.";
>     
>     nit: s/reveiced/received/
> [skraza]:  Fixed.        
>     
>          grouping peer-state-derived {
>            description
>              "The peer state information derived from the LDP protocol
>               operatoins.";
>     
>     nit: s/operatoins/operations/
> [skraza]:  Fixed.            
> 
>          grouping peer-state-derived {
>          [...]
>            leaf next-keep-alive {
>              type uint16;
>              units seconds;
>              config false;
>              description "Time to send the next KeepAlive message.";
>            }
>     
>     nit: this description text sounds like it's giving an absolute time, but
>     given that it's a uint16 I presume it's "time [from now] until sending".
> [skraza]:  Clarified.       
> 
>            container received-peer-state {
>              config false;
>              description
>                "Operational state information learned from the peer.";
>     
>              uses graceful-restart-attributes-per-peer;
>     
>              container capability {
>                description "Configure capability.";
>     
>     This is config-false; "Configure" in the description cannot be correct,
>     though I suppose "Configured" could be.
>     (Also applies to several of the child containers' descriptions.)
>     
> [skraza]:  Fixed all these.
>   
>            leaf up-time {
>              type string;
>              config false;
>              description "Up time. The interval format in ISO 8601.";
>     
>     We probably should have some sort of reference for ISO 8601 (though of
>     course it cannot actually *be* a <xref> in the YANG module itself)
>     
> [skraza]:  Changed to use timeticks64 now.
> 
>            leaf notification {
>              type yang:counter64;
>              description
>                "The number of messages sent or received.";
>     
>     Is this the same as or different than "leaf total-messages"?
> [skraza]:  Fixed.    
> 
>                    leaf label-distribution-controlmode {
>     
>     nit: any reason to not hyphenate control-mode as well?
> [skraza]:  Fixed.        
> 
>                container interfaces {
>                  description
>                    "A list of interfaces for LDP Basic Descovery.";
>     
>     nit: s/Descovery/Discovery/
> [skraza]:  Fixed.    
>     
>              container discovery {
>                description
>                  "Neibgbor discovery configuration and operational state.";
>     
>     nit: s/Neibgbor/Neighbor/
> [skraza]:  Fixed.    
>     
>                        // ipv4
>                        container hello-adjacencies {
>                          config false;
>                          description
>                            "Containing a list of hello adjacencies.";
>     
>                          list hello-adjacency {
>                            key "adjacent-address";
>                            config false;
>                            description "List of hello adjacencies.";
>     
>                            leaf adjacent-address {
>                              type inet:ipv4-address;
>                              description
>                                "Neighbor address of the hello adjacency.";
>                            }
>     
>     I'm not an expert here, but I'm not really seeing why we need to
>     specialize this container for v4 vs v6 as opposed to just using
>     inet:ip-address and having something that's applicable to both.
> 
> [skraza]: We could. There are pros and cons either way. The reasons to use separate containers in this model are:
> 1) Operators usually plan and configure IPv4 and IPv6 separately, rather than mixing the address family together. The separate containers provide more focus and clear separations.
> 2) Often times, there are attributes with different types for different address families. Mixing two together brings up some modeling difficulties to specifies differently for different address families.
> 3) Sometimes, there are attributes that are applicable to one address family but not the other. Mixing the two containers together would require some difficult conditions, which would complicate the model and make the model less usable.
> 4) Mixing the two containers can make the YANG file more compact, but not the configuration data and operational state data.
> 5) The existing RFCs use the style of separate containers, such as RFC 8344 and RFC 8349.

Thanks for writing this out; it helps me understand the situation better.

>                        leaf adjacent-address {
>                          type inet:ipv4-address;
>                          description
>                            "Configures a remote LDP neighbor and enables
>                             extended LDP discovery of the specified
>                             neighbor.";
>     
>     Given that there's a sibling leaf "enable", do we still want "and
>     enables" here?
> [skraza]: We don’t need “enables” here. Removed it.
> 
>     
>     Should peer/capability have a comment in its description that the
>     container exists to be augmented by other (extension) modules?
>     
> [skraza]: Yes. It is better. Added as suggested.
> 
>          notification mpls-ldp-fec-event {
>     
>     Am I reading this right that we only generate notifications on 0->1 and
>     1->0 transitions of "number of prefixes active for a given FEC"?  Would
>     we ever want to know about addition/removal of prefixes that do leave
>     the FEC up?
>     
> [skraza]: Here, the prefix is the FEC. Addition/removal of a prefix is equivalent to the addition/removal of the FEC. Changed the leaf name from prefix to fec to avoid the confusion.
> 
>     Section 9.2
>     
>     I can't find a pattern into which if-feature statements go into
>     containers used to augment things vs. the augment directives themselves.
>     Can we be more consistent (or am I missing the pattern)?
>     
> [skraza]: It is not intended to set up a strict pattern between feature name and containers, while the purpose is to provide more descriptive feature names. That said, the parent container names are often used to provide the contexts for the feature name. e.g. In the case of “dual-stack-transport-pereference”, it seems that the parent container name “peers” is also beneficial, so we have the feature name “dual-stack-transport-preference” to “peers-dual-stack-transport-preference” to make the feature name fit better with the pattern and more descriptive.

Thanks for the explanation

>          feature dual-stack-transport-pereference {
>            description
>              "This feature indicates that the system allows to configure
>               the transport connection pereference in a dual-stack setup.";
>     
>     nit: s/pereference/preference/ (twice)
> [skraza]: Fixed.
>     
>          feature policy-targeted-discovery-config {
>            description
>              "This feature indicates that the system allows to configure
>               policies to control the acceptance of targeted neighbor
>               discovery hello messages.";
>     
>     nit: s/hello/Hello/
>     [skraza]: Fixed.
> 
>          typedef neighbor-list-ref {
>            type string;
>            description
>              "A type for a reference to a neighbor address list.
>               The string value is the name identifier for uniquely
>               identifying the referenced address list, which contains a list
>               of addresses that a routing policy can applied. The definition
>               of such an address list is outside the scope of this
>               document.";
>     
>     If I'm reading correctly between here and where this is used, the
>     semantics of this are better described as "implementation-specific" --
>     that is, this is a label that refers to some local configured state,
>     specifically, a list of neighbor addresses used as an ACL.  (Emphasis on
>     *local* -- it seems like it does not need to interact with anything
>     else.)
>     I do sympathize with the YANG doctor, though -- as-is, this feels
>     underspecified.
>     (BTW, IP ACLs are not considered to be a security measure by the broader
>     security community.)
>     Similarly for the subsequent <foo>-list-refs.
> 
> [skraza]: We do understand the concerns, and have fixed these two types by using the https://tools.ietf.org/html/draft-ietf-rtgwg-policy-model, which was not available when we initially started to work on this model.

Great!

>              container interfaces {
>                description
>                  "A list of interfaces on which forwarding is disabled.";
>     
>     nit: given the separate 'ldp-disable' leaf, perhaps s/is/can be/ is in
>     order.
> [skraza]: Understand the confusion. Changed as suggested.
> 
>     Also, the description of the list itself implies additional
>     functionality, of listing the interfaces used for basic discovery.
>     
> [skraza]: The description was wrong. We don’t intend to use this list for discovery. Fixed.
> 
>          grouping graceful-restart-augment {
>            description "Augmentation to graceful restart.";
>     
>            leaf helper-enable {
>              if-feature graceful-restart-helper-mode;
>              type boolean;
>              default false;
>              description
>                "Enable or disable graceful restart helper mode.";
>     
>     Where is this "helper mode" documented?
> [skraza]: This is not an RFC defined term but commonly known and used when GR timer is set to 0.
>   Fixed the description in yang to highlight this.

Is this just staged for the -09?  I am not seeing it in the -08.

>          augment "/rt:routing/rt:control-plane-protocols/ldp:mpls-ldp/"
>            + "ldp:global" {
>            description "Graceful forwarding nexthop augmentation.";
>            uses global-forwarding-nexthop-augment;
>          }
>     
>     nit: I don't think "Graceful" is correct, in the description.
>     [skraza]: Removed it.
> 
>     Why is (the augmented)
>     /rt:routing/rt:control-plane-protocols/ldp:mpls-ldp/ldp:global/ldp:address-families/ipv6/label-distribution-controlmode
>     'config false'?  Doesn't RFC 5036 say that an LSR might support either
>     as a configuration option?
>     
> [skraza]: Yes, but no implementation supports such a cfg as change of such a mode is a massive disruptive operation - all implementation pick one mode as the default and only supported mode and stick to it.
>   The LDP YANG design team had discussed this early enough and agreed to not provide this a cfg option.

Ah, understood.

>                leaf adjacent-address {
>                  type inet:ipv6-address;
>                  description
>                    "Configures a remote LDP neighbor and enables
>                     extended LDP discovery of the specified
>                     neighbor.";
>     
>     As for the ipv4 case, do we still want "and enables" here?
> [skraza]: Fixed by removing it.    
> 
>          augment "/rt:routing/rt:control-plane-protocols/ldp:mpls-ldp/"
>            + "ldp:discovery/ldp:interfaces/ldp:interface/"
>            + "ldp:address-families/ldp-ext:ipv6" {
>            description "Interface address family IPv6 augmentation.";
>            uses interface-address-family-ipv6-augment;
>     
>     Didn't we just create .../ldp:address-families/ldp-ext:ipv6 with an
>     earlier augmentation in this module?  (Similarly for
>     /rt:routing/rt:control-plane-protocols/ldp:mpls-ldp/ldp:peers/ldp:peer/ldp:address-families/ldp-ext:ipv6
>     which we augment with "uses peer-af-policy-container".)
>     
> [skraza]: Yes. It is ok to augment a node which is part of another augmentation. The result is to add them all together. Doing so is to avoid too much irregular patterns, so that we have more similar construction between IPv4 and IPv6.
> 
>     Appendix A
>     
>                      "is-router": [null],
>     
>     This looks weird to me ("array containing one element with value literal
>     null") -- why does the array come into play?
>     
>     [Xufeng]: Here is a little explanation to the seemingly confusing notation: the pair of square brackets does not indicate a normal array. Instead, it is a special notation to indicate that the value of a “empty” type leaf is “present” (Section 6.9. of RFC 7951). Semantically the line above says that “(the neighbor) is a router.”

Ah, thanks for the explanations.

-Ben