Re: [manet] Tsvart early review of draft-ietf-manet-dlep-credit-flow-control-09

[Lou Berger <[e-mail]>]

Hi David
Thanks for the response, please see in line below...

----------
On March 22, 2022 8:13:47 PM "Black, David" <[e-mail]> wrote:

> Lou,
>
> Thanks for the comprehensive response.  I'm copying and replying to only the items that merit further attention - any item not responded to is ok with me.
>
>>> -- [1] -- Number of Documents
> [...]
>>> I strongly suggest folding the DA credit extension draft into the traffic
>>> classification draft, as the resulting two-draft structure will be easier to
>>>  explain, and more importantly, should be easier for implementers to understand.
>>
>> This was the decision of the WG made a long time ago. I defer to the
>> chairs on this.
>
> I would ask the WG to revisit this decision based on the surprisingly short length
> of the DA credit extension draft (about a page of actual protocol specification
> plus one addition to an IANA registry).
>

The working group has been discussing this since ietf100. Each time it comes up it seems the group is on the fence but generally agrees that it is good from a conformance standpoint to be unambiguous in which functions are supported in an RFC. If we put everything in one document, then vendors will have to say they only support part of the document, and users won't be able to tell from a spec sheet what function is possible with the device.

We discussed it again in this meeting and came to the same conclusion. For my standpoint this topic is again closed, but of course it is up to the iesg to decide.

>>> -- [2] -- Flow Match Criteria
>>>
>>> The Traffic Classification draft does not explain how traffic classification is
>>> actually performed, i.e., how the applicable FID is determined for a packet,
>>> particularly when more than one FID is possible.
> [...]
>>
>> It took me a bit to decide how to address this (valid) comment. I
>> propose the following change to
>> draft-ietf-manet-dlep-traffic-classification-06:
>>
>> At the end of section 2.3.1, I propose adding the following text:
>>
>>          If a packet matches both a DSCP Field Value,see Section 2.2.
>>          and a Priority Field value, the DSCP
>>          associated TID MUST take precedence.
>
> That's headed in the right direction, but more work is needed on a
> couple of aspects.
>
> The above text does not cover situations where 0 DSCPs and/or 0 PCPs
> are used in a FID, and does not mention the VLAN identifier.  Suggestion:
>
> If a packet matches both a DiffServ Traffic Classification
> Sub-Data Item (see Section 2.2), e.g., DSCP match, and
> an Ethernet Traffic Classification Sub-Data Item (see Section
> 2.3), e.g., PCP match, then the TID with which the DiffServ
> Traffic Classification Sub-Data Item is associated MUST
> take precedence.
>

This works for me. Thank you for the text!

> In addition, I think there's an implicit assumption that TIDs do not
> overlap, perhaps because they are associated with non-overlapping
> destinations.  Under that assumption, the same packet could not match
> FIDs under more than one TID.  If that's correct, that assumption needs
> to be explicitly stated (or referred to in the base DLEP spec?).  If that's
> not correct, then an additional precedence rule is needed, e.g., lowest
> numbered TID takes precedence.
>

Sure we can use that - lowest tid.

>>> -- [3] -- Router MAY ignore flow control
>>>
>>> The Management Considerations section in both the traffic classification and DA
>>> credit extension drafts contains this text:
>>>
>>>     When modem-provided credit window information exceeds the capabilities of a
>>>     router, the router MAY use a subset of the provided credit windows.
>>>
>>> In other words, the router MAY ignore any credit windows that exceed the
>>> routers capabilities ... which in practice could amount to "MAY ignore any
>>> credit windows that the router implementer views as inconvenient" ... with the
>>> router proceeding to overrun the corresponding modem queues.  That doesn't seem
>> right, so if this is intended, some serious rationale/explanation ought to be
>>> provided.
>>
>> I personally have no objection to MUST, but perhaps a SHOULD is the
>> right answer given that there is an alternative defined.
>
> This isn't about MUST/SHOULD/MAY - it's about queue overrun which isn't supposed to
> happen and can't be addressed via just a keyword change.  Suppose a router ignores the
> 17th modem credit window, *and* proceeds to overrun the corresponding modem
> queue.  That overrun is an undesirable failure mode.
>

I'm not sure I'd call it a failure. It will just result in unexpected packet loss, which can happen at any time in a radio network anyway...

> Would it be reasonable for a router that ignores at least one credit window to be
> prohibited from sending traffic that does not match any credit window (or that
> could match an ignored credit window)?  That would address the overrun problem.
>
> Suggestion - at the end of Section 2.1 in the flow control draft, add a sentence that
> begins with:
>
> If a router is unable to associate a credit window with traffic to be
> sent a modem, then ...
>
>>> -- [4] -- Association Not Well-Specified
> [...]
>> Okay I'll try to improve the language in section 2 to cover (I) and (II)
>
> Better, but still hard to follow.    Suggestion:
>
> OLD
>   The TID provides the information to
>    support router traffic classification, based on the FIDs contained in
>    the TID, see [I-D.ietf-manet-dlep-traffic-classification].  As
>    defined, each credit window has a corresponding FID.
> NEW
>   The TID provides the information to
>    support router traffic classification, based on the FIDs contained in
>    the TID, see [I-D.ietf-manet-dlep-traffic-classification].  As
>    defined, each credit window has a corresponding FID, so traffic is
>    mapped to a credit window by locating a matching FID that
>    is contained in the TID that is associated with the traffic's destination.
>


Wfm. Thank you again for the text!

>> > -- [5] -- Initialization vs. in-flight traffic
>> >
>> > The Credit Window Initialization data item appears to be intended to establish
>> > common state for a Credit Window (e.g., size, number of credits) across the
>> > router and modem ... but ... that data item appears to be allowed to be sent
>> > while there's traffic in flight.  The result would be that the modem would
>> > count in-flight traffic against the initialized Credit Window, but the router
>> > would not.   The resulting inconsistency deserves discussion - it may be
>> > acceptable if the amount of traffic in flight is miniscule by comparison to
>> > both the Credit Window size and initial credit balance.
>>
>> This is the purpose of the credit window status DI.
>
> That does not solve the problem if the amount of traffic in flight between router
> and the modem is a significant fraction of the credit window size.  E.g., suppose we
> have a very slow modem with a 1kbyte queue and there could be 512 bytes in flight
> between router and modem.  If the Credit Window Status arrives with 512 bytes in
> flight behind it, and the modem immediately does a Credit Window Initialization to 1k,
> the router can then send another 1k for a total of 1.5k which overruns the 1k modem
> queue by 50% (oops).
>
> I expect the reaction to that example to be along the lines of: That can't happen
> because of <A>, <B> and <C>.  If so, then <A>, <B> and <C> need to be clearly stated
> in the draft (e.g., <A> might be " the amount of traffic in flight is miniscule by
> comparison to both the Credit Window size and initial credit balance").
>

Since you've been so helpful :-) do you have any texts to propose to address this point?

>> -- [6] -- Security Considerations
> [...]
>> How about:
>>
>>      This document introduces credit window control and flow mechanisms
>>      to DLEP.  These mechanisms expose vulnerabilities similar to
>>      existing DLEP messages, e.g., Destination UP or Down message
>>      injection attacks. The security mechanisms documented in <xref
>>      target="RFC8175"/> can be applied equally to the mechanism defined
>>      in this document.
>
> Second sentence would be better if the example were specific to this draft, perhaps:
>

Umm, I was trying to show relationship to existing exposures.

>         These mechanisms expose vulnerabilities analogous to vulnerabilities
>         exposed by existing DLEP messages, e.g., an injected message resizes
>         a credit window to a tiny value causing denial of service.
>

How about

an injected message resizes    a credit window to a tiny value causing denial of service, which is fundamentally no different than existing exposure of injecting a destination down message.

>>> [C] Data item usage:  It's not clear which data items are required, allowed or
>>> prohibited in which DLEP messages.
> [...]
>> Unexpected DIs in messages is defined by the base spec -- and yes, it's
>> an error.
>
> I suggest adding a version of that sentence to both drafts that define DIs -
> the flow control draft and the traffic classification draft.
>

Sure, although this is so fundamental to the protocol I would think any implementer would see this as obvious.

>> > [D] Uninitialized window: The second sentence of Section 2.3.1 in the flow
>> > control draft begins with:
>> >
>> >     This Data Item SHOULD be included in any Session Initialization Response
>> >     Message that also ...
>> >
>> > What happens if that SHOULD is not adhered to?  Please explain.
>>
>> There's a MAY immediately following.  Other places in the document
>> describe that use of an FID/credit window in a message before it is
>> defined/initialized are treated as errors.
>
> That MAY doesn't appear to help.  I suggest starting the sentence with:
>
>    In order to avoid errors caused by use of undefined FIDs or uninitialized
>    credit windows, this Data Item SHOULD be included ...
>

Okay

>>> [E] Credit Window Size: In Section 2.3.1 of the flow control draft (Credit
>>> Window Initialization), what happens if the Credit Value exceeds the Credit
>>> Window Size?
>> Okay will add a sentence.
>
> Nit to fix in new sentence ...
>
>    If the resulting Credit Value results in the credit
>    window exceeding the represented Credit Window Max Size, the Credit
>    Window Size is used as the new credit window size.
>
> In last line - s/Window Size/Window Max Size/
>

Good catch.

Thank you again! I'll do the update in the next few days and upload

Lou

> Thanks, --David
>
> -----Original Message-----
> From: Lou Berger <[e-mail]>
> Sent: Tuesday, March 22, 2022 10:06 AM
> To: Black, David; [e-mail]
> Cc: MANET IETF
> Subject: Fwd: Tsvart early review of draft-ietf-manet-dlep-credit-flow-control-09
>
>
> [EXTERNAL EMAIL]
>
>
> This is a resend as it seems the first message didn't make it to David
> or the archive.
>
> -------- Forwarded Message --------
> Subject: Re: Tsvart early review of
> draft-ietf-manet-dlep-credit-flow-control-09
> Date: Thu, 24 Feb 2022 07:18:55 -0500
> From: Lou Berger <[e-mail]>
> To: David Black <[e-mail]>, [e-mail]
> CC: [e-mail], [e-mail]
>
> David,
>
> Please see (overdue) responses in-line below.
>
> On 11/19/2021 12:02 PM, David Black via Datatracker wrote:
>> Reviewer: David Black
>> Review result: On the Right Track
>>
>> TSVART Review of:
>>
>>          DLEP Credit-Based Flow Control Messages and Data Items
>>                  draft-ietf-manet-dlep-credit-flow-control-09
>>          DLEP DiffServ Aware Credit Window Extension
>>                  draft-ietf-manet-dlep-da-credit-extension-12
>>          DLEP Traffic Classification Data Item
>>                  draft-ietf-manet-dlep-traffic-classification-06
>>
>> Reviewer: David L. Black <[e-mail]>
>> Date: November 19, 2021
>>
>> This is a combined review of all three drafts.
>>
>> These three drafts specify a flow control mechanism between a "router" and a
>> "modem" in the sending direction (router to modem) to prevent overrun of
>> buffers in the "modem".  The mechanism is designed to provide flow control for
>> multiple queues in the "modem" that are independently protected, with the
>> "modem" in control of how much data the router is allowed to send - to a first
>> approximation, there is a separate instance of flow control, called a Credit
>> Window, for each queue.  Both Diffserv (DSCP) and Ethernet Priority (PCP)
>> values can be used to classify traffic to determine which flow control instance
>> (Credit Window) applies to each packet being sent to the "modem" by the
>> "router".
>>
>> Each of the drafts is reasonably well written, but there are some difficulties
>> in understanding the combination of the three drafts, which have to be used
>> together in order to implement this flow control mechanism. That's one of a
>> number of issues that deserve attention.
>>
>> ****** Major Issues:
>>
>> -- [1] -- Number of Documents
>>
>> I understand and see the merit in specifying the flow control mechanism
>> (-credit-flow-control draft) independently of traffic classification
>> (-traffic-classification draft).  In contrast, I do not see a strong rationale
>> for the separate DA credit extension draft (-da-credit-extension), which has
>> very little content (about a page of actual protocol specification plus one
>> addition to an IANA registry).
>>
>> I strongly suggest folding the DA credit extension draft into the traffic
>> classification draft, as the resulting two-draft structure will be easier to
>> explain, and more importantly, should be easier for implementers to understand.
>
> This was the decision of the WG made a long time ago. I defer to the
> chairs on this.
>
>> -- [2] -- Flow Match Criteria
>>
>> The Traffic Classification draft does not explain how traffic classification is
>> actually performed, i.e., how the applicable FID is determined for a packet,
>> particularly when more than one FID is possible. For example, suppose a packet
>> carries a DSCP that matches FID 5 and a PCP that matches FID 7 - does that
>> packet consume FID 5 credits or FID 7 credits?  If the router and modem make
>> different decisions, the result may be undesirable.  This example is within the
>> same TID - the situation appears to be worse across TIDs for the same
>> destination, because the same match criteria could appear in multiple TIDs,
>> e.g., same DSCP could match different FIDs under different TIDs.  Clarity is
>> needed here so that the router and modem agree on which FID's credits are
>> consumed by each packet, and on a related note, the DA credit extension draft
>> does not prohibit use of Ethernet traffic classification - perhaps it should.
>
> It took me a bit to decide how to address this (valid) comment. I
> propose the following change to
> draft-ietf-manet-dlep-traffic-classification-06:
>
> At the end of section 2.3.1, I propose adding the following text:
>
>           If a packet matches both a DSCP Field Value,see Section 2.2.
>           and a Priority Field value, the DSCP
>           associated TID MUST take precedence.
>
>> -- [3] -- Router MAY ignore flow control
>>
>> The Management Considerations section in both the traffic classification and DA
>> credit extension drafts contains this text:
>>
>>     When modem-provided credit window information exceeds the capabilities of a
>>     router, the router MAY use a subset of the provided credit windows.
>>
>> In other words, the router MAY ignore any credit windows that exceed the
>> routers capabilities ... which in practice could amount to "MAY ignore any
>> credit windows that the router implementer views as inconvenient" ... with the
>> router proceeding to overrun the corresponding modem queues.  That doesn't seem
>> right, so if this is intended, some serious rationale/explanation ought to be
>> provided.
>
> I personally have no objection to MUST, but perhaps a SHOULD is the
> right answer given that there is an alternative defined.
>
>
>> -- [4] -- Association Not Well-Specified
>>
>> The overall association structure is that flow control instances are identified
>> by FIDs which are associated with TIDs which are associated with destinations.
>> That structure is described towards the end of Section 2 of the flow control
>> draft, but the details of how it is done are mostly missing - the flow control
>> draft ought to state that:
>>
>>          i) FIDs are associated with TIDs via traffic classification information
>>          specified in the traffic classification draft. ii) TIDs (and their
>>          associated FIDs) are associated with destinations via use of the Credit
>>          Window Association data item in DLEP Destination Up and Destination
>>          Update messages.
>>
>> The first item (i) can be mostly inferred from the last few paragraphs in
>> section 2 of the flow control draft (and needs to be more explicit), but the
>> second item (ii) is missing because the flow control draft does not specify how
>> to determine the "destination" with which a TID is associated by a Credit
>> Window Association data item (in a DLEP Destination Up or Destination Update)
>> message.  I would expect that the destination is a relatively obvious field in
>> those messages, but the draft needs to specify that field.  This should be easy
>> to fix, although it made the drafts much more difficult to understand.
>
> Okay I'll try to improve the language in section 2 to cover (I) and (II)
>
>
>>
>> -- [5] -- Initialization vs. in-flight traffic
>>
>> The Credit Window Initialization data item appears to be intended to establish
>> common state for a Credit Window (e.g., size, number of credits) across the
>> router and modem ... but ... that data item appears to be allowed to be sent
>> while there's traffic in flight.  The result would be that the modem would
>> count in-flight traffic against the initialized Credit Window, but the router
>> would not.   The resulting inconsistency deserves discussion - it may be
>> acceptable if the amount of traffic in flight is miniscule by comparison to
>> both the Credit Window size and initial credit balance.
>
> This is the purpose of the credit window status DI.  The current text says:
>
>           Once the credit window is identified, the modem SHOULD check the
>           received Current Credit Window Size field value against the outstanding credit
>           window's available credits at the time the most recent Credit Window
>           Initialization or Grant Data Item associated with the indicated FID
>           was sent.  If the values significantly differ, i.e., greater than can
>           be accounted for based on observed data frames, then the modem SHOULD
>           send a Credit Window Initialization Data Item to reset the associated
>           credit window size to the modem's current view of the available
>           credits.  As defined above, Credit Window Initialization Data Items are
>           sent in Session Update Messages.  When multiple Data Items need to be
>           sent, they SHOULD be combined into a single message when possible.
>           Alternatively, and also in cases where there are small differences,
>           the modem MAY adjust the values sent in Credit Window Grant Data Items
>           to account for the reported Credit Window.
>
> is this insufficient? if so, can you suggest some alternate wording.
>
>
>> -- [6] -- Security Considerations
>>
>> Although this is a Transport review, I found a security issue that would be
>> better dealt with now before the security directorate points it out ;-) - the
>> security considerations sections in each of the 3 drafts claims that adding
>> credit window control and flow functionality to DLEP does not introduce any new
>> security considerations (vulnerabilities). That's a nice try, but it's
>> incorrect.
>>
>> These drafts specify a new resource (credits in a credit window) that is
>> subject to resource exhaustion attacks that could cause denial-of-service.  For
>> example, suppose an attacker injects a Credit Window Initialization data item
>> that contains almost no credits and/or specifies a ridiculously tiny Window
>> (Max) Size.  I expect that the protocol contains mechanisms to counter this and
>> related attacks on credit resources (e.g., if something looks wrong, the modem
>> reinitializes the Credit Window), but the current text incorrectly asserts the
>> non-existence of such attacks.  These sorts of attacks definitely exist - I am
>> aware of a (subsequently fixed) resource exhaustion problem in another
>> credit-based flow control mechanism caused by an unanticipated environmental
>> "attack" on signal integrity of credit exchange messages, resulting in message
>> discard and credit loss.
>
> How about:
>
>        This document introduces credit window control and flow mechanisms
>       to DLEP.  These mechanisms expose vulnerabilities similar to
>       existing DLEP messages, e.g., Destination UP or Down message
>       injection attacks. The security mechanisms documented in <xref
>       target="RFC8175"/> can be applied equally to the mechanism defined
>       in this document.
>
>> ***** Minor Issues:
>>
>> [A] Packets consume credits: Section 2 of the flow control draft needs to say
>> somewhere early in the section that each packet that a router sends consumes a
>> credit for each octet in the packet.  This is to be found at the end of the
>> second paragraph in Section 2.1, but ought to be stated much earlier, e.g., at
>> the end of second paragraph in Section 2.
> okay
>>
>> [B] Bidirectional messages: Both the Credit Control Message and its response
>> can be sent in both directions (i.e., by both modems and routers).  Has the
>> possibility of using different message types in each direction been considered?
>>   That might help out people who have to read packet traces/dumps.  NOTE: "Yes,
>> that was carefully considered and the WG decided not to do that." is a fine
>> response.
> It was discussed, and the current approach was selected by the WG.
>> [C] Data item usage:  It's not clear which data items are required, allowed or
>> prohibited in which DLEP messages.  It appears that any data item MAY be used
>> in any DLEP message, which is surely not what was intended.  Both the flow
>> control and traffic classification drafts ought to spell out these details,
>> including what happens when  the rules are violated, e.g., data item shows up
>> in a message where it's not supposed to be used - is the data item ignored or
>> does it cause an error?
>
> Unexpected DIs in messages is defined by the base spec -- and yes, it's
> an error.  These drafts don't modify that behavior.
>
>
>>
>> [D] Uninitialized window: The second sentence of Section 2.3.1 in the flow
>> control draft begins with:
>>
>>     This Data Item SHOULD be included in any Session Initialization Response
>>     Message that also ...
>>
>> What happens if that SHOULD is not adhered to?  Please explain.
>
> There's a may immediately following.  Other places in the document
> describe that use of an FID/credit window in a message before it is
> defined/initialized are treated as errors.
>
>> [E] Credit Window Size: In Section 2.3.1 of the flow control draft (Credit
>> Window Initialization), what happens if the Credit Value exceeds the Credit
>> Window Size?
> Okay will add a sentence.
>> I also strongly suggest renaming Credit Window Size to Credit Window Max Size
>> or something similar to avoid possible confusion of this concept with current
>> credit balance (Credit Value)
> Sure
>>
>> [F] TID association with destination:
>>
>> Section 2.3.2 of the flow control draft (Credit Window Association) says that a
>> the traffic classification information associated with a TID "MUST be
>> associated with the destination" - what does that mean?
>>
>> Specifically:
>> - Does that traffic classification information replace any existing     traffic
>> classification information, e.g., if the info associated with that TID has been
>> updated since that TID was most recently associated with that destination? - Is
>> there any limit on the number of TIDs or amount of traffic classification
>> information that can be associated with a destination?  If so, what happens if
>> that limit is exceeded? Also note that "the destination" is not currently
>> defined, see [4] above.
>
> I'll drop:
>
> That clause, the next MUST is sufficient.
>
>> [G] Bad Credit Window math: The next to last paragraph in Section 2.3.3 (Credit
>> Window Grant) of the flow control draft contains this text about adding credit
>> to a window:
>>
>>     If the increase results in a window overflow, i.e., the size of the credit
>>     window after the increase is smaller than the original credit window size,
>>     the Credit Window must be set to its maximum (0xFFFFFFFFFFFFFFFF).
>>
>> That's doubly wrong:
>> - consider a window max size of 10, a current window size of 3 and an addition
>> of 15 credits.  Overflow occurs, but the window size increases, as the new
>> window size is 8. - The maximum size of the credit window may be considerably
>> smaller than 0xFFFFFFFFFFFFFFFF.
>>
>> For the first item, just use the concept of overflow in modular arithmetic.
>> For the second item, remove that all-F's constant and refer back to the Credit
>> Window Initialization material for max size.
> WFM
>>
>> This text is also an example of why [E] strongly suggests renaming Credit
>> Window Size to Credit Window Max Size or somthing similar.
> done
>> [H] Amount of imprecision: Section 2.3.4 of the flow control draft discusses
>> modem comparison of internal values with values received from the router, and
>> contains this text:
>>
>>     If the values significantly differ, i.e., greater than can be  accounted for
>>     based on observed data frames,
>>
>> That needs some guidance on what a significant difference is and how much of a
>> "greater than" difference ought to trigger the consequences, accompanied by a
>> warning against precise (== vs. !=) comparisons (e.g., courtesy of [5] above).
> I believe we (once) discussed this ad concluded it could be
> implementation dependent.  If you feel strongly on this, can you propose
> some language that we can discuss in the WG?
>> [I] Update means what?: The last paragraph of Section 2.1 of the traffic
>> classification draft (Traffic Classification Data Item) contains:
>>
>>     the router MUST update the information using the values carried in the Data
>>     Item.
>>
>> What does "update" mean, e.g., "replace", "merge"?
> MUST replace.
>>
>> [J] Error behavior: Section 2.1.1 of the traffic classification draft leaves
>> error behavior as a "go fish" exercise for the reader:
>>
>>     Any errors or inconsistencies encountered in parsing
>>     Sub-Data Items are handled in the same fashion as any other Data
>>     Item parsing error encountered in DLEP.
>>
>> That doesn't tell an implementer what to do - this needs to cite a reference
>> that specifies the applicable error behavior.
> I disagree, DLEP has defined error processing in the base spec. There
> has been discussion about changing that behavior in an update.  I don't
> think repeating what is in the base spec is a good idea, particularly in
> light of that potential update.
>> [K] RFC 2474 CU field: Section 2.2 of the traffic classification draft defines
>> the DS Field as an octet and reproduces the definition from RFC 2474 that
>> includes the DSCP and CU fields.  Unfortunately, this does not reflect updates
>> to RFC 2474 - those two bits are now the ECN Field, which may be non-zero.
>>
>> The fix for this is simple - use the 6 bit DSCP field from RFC 2474 and replace
>> the CU field with two zero bits to produce an octet.
> okay, changed to stat that DSCP comes from 2474 and the bits "MUST be zero"
>>
>> [L] 802.1Q DEI field: This has a similar problem to the CU field ... and a
>> similar solution - replace the DEI field with a zero bit.
> Done!
>>
>> ****** Nits:
>>
>> FID paragraph in section 2.3.5 of the flow control draft:
>>          The FID also uniquely identifies a credit window
>> s/identifies/indicates/ for consistency.
> sure
>>
>> End of Section 2 of the traffic classification draft:
>>          TID and FID values is a modem-local scope.
>> s/is a/have/
>
> Thank you!
>
> I'll publish a matching update shortly.
>
> Lou
>
>
>>
>>