Re: [Last-Call] Secdir last call review of draft-ietf-bess-mvpn-fast-failover-11

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi Daniel,
thank you for spotting it, fixed.
 I enjoyed our discussion and, again, thank you for the review and your thoughtful comments.

Best regards,
Greg

On Thu, Nov 12, 2020 at 1:32 PM Daniel Migault <daniel.migault@xxxxxxxxxxxx> wrote:
Thanks for the response and explanation. I am fine with the text you proposed and I consider all my concerns being addressed.

I am reading your text as implicitly suggesting the following lines of your response, which seems reasonable.   
 
"""
 it is difficult to make any assumptions on how the convergence in the control plane may impact the forwarding plane and what effect that will have on a multicast flow. I think that very much depends on the implementation and the HW capabilities of the platform used.
"""

There  one additional nit s/sectionSection/Section/ which I think comes from the conversion. 

Thanks for all your clarifications!

Yours, 
Daniel


From: Greg Mirsky <gregimirsky@xxxxxxxxx>
Sent: Thursday, November 12, 2020 3:14 PM
To: Daniel Migault <daniel.migault@xxxxxxxxxxxx>
Cc: secdir@xxxxxxxx <secdir@xxxxxxxx>; BESS <bess@xxxxxxxx>; last-call@xxxxxxxx <last-call@xxxxxxxx>; draft-ietf-bess-mvpn-fast-failover.all@xxxxxxxx <draft-ietf-bess-mvpn-fast-failover.all@xxxxxxxx>
Subject: Re: Secdir last call review of draft-ietf-bess-mvpn-fast-failover-11
 
Hi Daniel,
thank you for the additional information. I understand your concerns and agree that it is helpful to provide implementors and operators with useful information about the potential impact the new functionality may demonstrate in the network and how to mitigate the risks. I believe it is important to recognize that this draft proposes mechanisms that expedite the failure detection in a P-tunnel from the perspective of a downstream PE. And the detection directly impacts the control plane, not the data plane. I believe that it is difficult to make any assumptions on how the convergence in the control plane may impact the forwarding plane and what effect that will have on a multicast flow. I think that very much depends on the implementation and the HW capabilities of the platform used. I've moved the new text from Section 3.1 into the Security Considerations section. Please let me know if you think that is a more appropriate place for that paragraph.
Also, I've realized that the text I've proposed earlier that refers to 1:N and N:M protection might be the source of questions and arguments and would like to withdraw it. I hope you'll agree.

I've attached the new diff reflecting the changes in the working version.

Regards,
Greg


On Thu, Nov 12, 2020 at 9:02 AM Daniel Migault <daniel.migault@xxxxxxxxxxxx> wrote:
Hi Greg, 

Thanks for the response Greg. This seems to go in the right direction, but I think it would be nice to detail a bit on the negative impact that may result from the fast-fail over. 

"""
unnecessary failover negatively impacting the multicast service
"""

I apology to appear being maybe a bit picky, but, at least to me, the security consideration section is the place to point on specific impacts that an operator may not have thought of and the text appears to me a bit too vague on what can impact negatively the multicast service. 

Let me dig a bit on what I mean and probably what information I would have expected to find. Maybe that would have been useful I provided those earlier. Again, not being an expert in this area, please take my following recommendations with a pitch of salt. 

What I would like, for example, to understand is whether having a fast-failover between nodes that work properly results in a packet lost or not. 
I also envision that in some cases, this will result in packet re-ordering which might result in packet being rejected by the end node. 
In IPsec vpns, we have specific counters, keys that make fail-over relatively complex as a context has to be maintained between the old and the new node to pass anti replay protection and enable appropriated encryption/decryption. It would be good to clarify if any parameters need - or not - to be synchronized between the two nodes as its transfer represents a risk of disrupting the traffic, and thus may be mentioned. 
There probably other points I am missing due to my lack of expertise - especially those due to operational practices. 
I believe that any information that you could think of that would encourage you to double check/validate a network outage is present over performing the fast failover might be useful information.    
Similarly, it would be good to mention cases where an operator may choose not to deploy such mechanism.

Yours, 
Daniel


From: Greg Mirsky <gregimirsky@xxxxxxxxx>
Sent: Thursday, November 12, 2020 10:47 AM
To: Daniel Migault <daniel.migault@xxxxxxxxxxxx>
Cc: secdir@xxxxxxxx <secdir@xxxxxxxx>; BESS <bess@xxxxxxxx>; last-call@xxxxxxxx <last-call@xxxxxxxx>; draft-ietf-bess-mvpn-fast-failover.all@xxxxxxxx <draft-ietf-bess-mvpn-fast-failover.all@xxxxxxxx>
Subject: Re: Secdir last call review of draft-ietf-bess-mvpn-fast-failover-11
 
Hi Daniel,
thank you for your kind consideration of my notes. I've top-copied what appeared to me as the remaining open issues. I hope I've not missed any of your questions. Please find my notes in-line below tagged GIM>>. Attached are the updated working version and the new diff.

Regards,
Greg

<mglt>
sure. If you know the network is down, then fast fail-over is definitively a plus. What I think could be useful is to evaluate the cost associated to a fast-fail-over without any network failure.  This would be useful for an operator to evaluate whether it should spend more time in diagnosing a network failure versus performing a fast-fail-over.
Typically, if a fast failover comes a no cost at all, one operator would maybe use one exchange to test the liveness of a node rather than 3.      

At that point, it seems to me that additional text coudl be added to characterize the impact. These could be high level and indicative, but it seems to me that knowing these impacts presents some value to the operators.      
</mglt>
GIM>> I would like to add a new paragraph in Section 3.1:
NEW TEXT:
   All methods described in this section may produce false-negative
   state changes that can be the trigger for an unnecessary failover
   negatively impacting the multicast service provided by the VPN.  An
   operator expected to consider the network environment and use
   available controls of the mechanism used to determine the status of a
   P-tunnel.

Would the new text be helpful?

<mglt>
Thanks for the feed back, It seems to me important to mention it is not recommended these two mechanism co-exist.
How to avoid false negative transition might be out of scope of the draft I agree, but it seems to me worth being mentioned especially in relation to the impacts associated to a fail-over.  In case the fast-failover comes with no impact this becomes less of a problem for operator deploying it.
 
</mglt>
GIM>> I hope that the new text presented above addresses this concern.

<mglt>
I understand the document is addressing a 1:N scenario. That said, if M:N scenario leverage from 1:N protection it seems to me worth raising the issue.
</mglt>
GIM>> I propose adding the clarification of the use of the Sandby PE in Section 4:
OLD TEXT:
   The procedures described below are limited to the case where the site
   that contains C-S is connected to two or more PEs, though, to
   simplify the description, the case of dual-homing is described.
NEW TEXT:
   The procedures described below are limited to the case where the site
   that contains C-S is connected to two or more PEs, though, to
   simplify the description, the case of dual-homing is described.  Such
   a redundancy protection scheme, referred to as 1:N protection, is the
   special case of M:N protection, where M working instances are sharing
   protection of the N standby instances.  In addition to a network
   failure detection mechanism, the latter scheme requires using a
   mechanism to coordinate the failover among working instances.  For
   that reason, M:N protection is outside the scope of this
   specification. 

On Wed, Nov 11, 2020 at 8:48 AM Daniel Migault <daniel.migault@xxxxxxxxxxxx> wrote:
Hi Greg, 

Thanks for the response and clarifications. Most of my comments have been addressed/answered. However, it seems to me that some additional text might be added to the security consideration section document the impact on the network of a fast-failover operation. The knowledge of these impact might be useful for an operator to determine when the trigger can be done. 

Please see more comments inline. 

Yours, 
Daniel 


From: Greg Mirsky <gregimirsky@xxxxxxxxx>
Sent: Tuesday, November 10, 2020 9:13 PM
To: Daniel Migault <daniel.migault@xxxxxxxxxxxx>
Cc: secdir@xxxxxxxx <secdir@xxxxxxxx>; BESS <bess@xxxxxxxx>; last-call@xxxxxxxx <last-call@xxxxxxxx>; draft-ietf-bess-mvpn-fast-failover.all@xxxxxxxx <draft-ietf-bess-mvpn-fast-failover.all@xxxxxxxx>
Subject: Re: Secdir last call review of draft-ietf-bess-mvpn-fast-failover-11
 
Hi Daniel,
many thanks for the review, thoughtful comments, and questions, all are much appreciated. Also, my apologies for the long delay to respond to your comments. Please find my answers and notes in-line below tagged by GIM>>. Attached are the new working version and the diff to -12.

Regards,
Greg

On Fri, Oct 23, 2020 at 5:36 AM Daniel Migault via Datatracker <noreply@xxxxxxxx> wrote:
Reviewer: Daniel Migault
Review result: Has Nits

Hi,


I reviewed this document as part of the Security Directorate's ongoing effort to
review all IETF documents being processed by the IESG.  These comments were
written primarily for the benefit of the Security Area Directors.  Document
authors, document editors, and WG chairs should treat these comments just like
any other IETF Last Call comments.  Please note also that my expertise in BGP is
limited, so feel free to take these comments with a pitch of salt. 

Review Results: Has Nits

Please find my comments below.

Yours,
Daniel


                  Multicast VPN Fast Upstream Failover
                 draft-ietf-bess-mvpn-fast-failover-11

Abstract

   This document defines multicast VPN extensions and procedures that
   allow fast failover for upstream failures, by allowing downstream PEs
   to take into account the status of Provider-Tunnels (P-tunnels) when
   selecting the Upstream PE for a VPN multicast flow, and extending BGP
   MVPN routing so that a C-multicast route can be advertised toward a
   Standby Upstream PE.

<mglt>
Though it might be just a nit, if MVPN
designates multicast VPN, it might be
clarifying to specify the acronym in the
first sentence. This would later make
the correlation with BGP MVPN clearer.

</mglt>
GIM>> I've updated s/BGP MVPN/BGP multicast VPN/. Also, s/mVPN/MVPN/ throughout the document.


1.  Introduction

   In the context of multicast in BGP/MPLS VPNs, it is desirable to
   provide mechanisms allowing fast recovery of connectivity on
   different types of failures.  This document addresses failures of
   elements in the provider network that are upstream of PEs connected
   to VPN sites with receivers.

<mglt>
Well I am not familiar with neither BGP
nor MPLS. It seems that BGP/MLPS IP VPNS
and MPLS/BGP IP VPNs are both used. I am
wondering if there is a distinction
between the two and a preferred way to
designate these VPNs.  My understanding
is that the VPN-IPv4 characterizes the
VPN while MPLS is used by the backbone
for the transport.  Since the PE are
connected to the backbone the VPN-IPv4
needs to be labeled.

</mglt>
GIM>> I understand that this document often sends the reader to check RFC 6513 and/or RFC 6514. BGP/MPLS MVPN identifies the case of providing a multicast service over an IP VPN that is overlayed on the MPLS data plane using the BGP control plane.

   Section 3 describes local procedures allowing an egress PE (a PE
   connected to a receiver site) to take into account the status of
   P-tunnels to determine the Upstream Multicast Hop (UMH) for a given
   (C-S, C-G).  This method does not provide a "fast failover" solution
<mglt>
I understand the limitation is due to
BGP convergence.

</mglt>
GIM>> Yes, a dynamic routing protocol, BGP in this case, provides the service restoration functionality but the restoration time is significant and affects the experience of a client.

   when used alone, but can be used together with the mechanism
   described in Section 4 for a "fast failover" solution.

   Section 4 describes protocol extensions that can speed up failover by
   not requiring any multicast VPN routing message exchange at recovery
   time.

   Moreover, section 5 describes a "hot leaf standby" mechanism, that
   uses a combination of these two mechanisms.  This approach has
   similarities with the solution described in [RFC7431] to improve
   failover times when PIM routing is used in a network given some
   topology and metric constraints.


[...]

3.1.1.  mVPN Tunnel Root Tracking

   A condition to consider that the status of a P-tunnel is up is that
   the root of the tunnel, as determined in the x-PMSI Tunnel attribute,
   is reachable through unicast routing tables.  In this case, the
   downstream PE can immediately update its UMH when the reachability
   condition changes.

   That is similar to BGP next-hop tracking for VPN routes, except that
   the address considered is not the BGP next-hop address, but the root
   address in the x-PMSI Tunnel attribute.

   If BGP next-hop tracking is done for VPN routes and the root address
   of a given tunnel happens to be the same as the next-hop address in
   the BGP A-D Route advertising the tunnel, then checking, in unicast
   routing tables, whether the tunnel root is reachable, will be
   unnecessary duplication and thus will not bring any specific benefit.

<mglt>
It seems to me that x-PMSI address
designates a different interface than
the one used by the Tunnel itself. If
that is correct, such mechanisms seems
to assume that one equipment up on one
interface will be up on the other
interfaces. I have the impression that a
configuration change in a PE may end up
in the P-tunnel being down, while the PE
still being reachable though the x-PMSI
Tunnel attribute. If that is a possible
scenario, the current mechanisms may not
provide more efficient mechanism than
then those of the standard BGP.
GIM>> That is a very interesting angle, thank you. Yes, in OAM, and in the Fault Management (FM) OAM in particular, we have to make some assumptions about the state of the remote system based on a single event or change of state. Usually, AFAIK, operators use not a physical interface but a loopback to associate with a tunnel. With a fast IGP convergence, a loopback interface is reachable as long as there's a path through the network between two nodes.
<mglt>
Thanks for the clarification
</mglt>

Similarly, it is assumed the tunnel is
either up or down and the determination
of not being up if being down.  I am not
convinced that the two only states.
Typically services under DDoS may be
down for a small amount of time. While
this affects the network, there is not
always a clear cut between the PE being
up or down.
</mglt>
GIM>>  In defect detection a system often has some hysteresis, i.e., time that the system has to wait to change its state. For example, BFD changes state from Up to Down after the system does not receive N consecutive packets (usually 3). As a result, in some cases, the system can be tuned to detect relatively short outages while in others be slower and miss short-lived outages.


[...]

3.1.6.  BFD Discriminator Attribute

   P-tunnel status may be derived from the status of a multipoint BFD
   session [RFC8562] whose discriminator is advertised along with an
   x-PMSI A-D Route.

   This document defines the format and ways of using a new BGP
   attribute called the "BFD Discriminator".  It is an optional
   transitive BGP attribute.  In Section 7.2, IANA is requested to
   allocate the codepoint value (TBA2).  The format of this attribute is
   shown in Figure 1.

<mglt>
I feel that the sentence "In Section ...
TBA2)." should be removed.

</mglt>
GIM>> We use this to mark where to note the allocated value. Usually, this text is replaced by the RFC Editor to read 
In Section 7.2 IANA allocated codepoint XXX.



       0                   1                   2                   3
       0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |    BFD Mode   |                  Reserved                     |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      |                       BFD Discriminator                       |
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
      ~                         Optional TLVs                         ~
      +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+


            Figure 1: Format of the BFD Discriminator Attribute

   Where:

      BFD Mode field is the one octet long.  This specification defines
      the P2MP BFD Session as value 1 Section 7.2.

      Reserved field is three octets long, and the value MUST be zeroed
      on transmission and ignored on receipt.

      BFD Discriminator field is four octets long.





Morin, et al.             Expires April 5, 2021                 [Page 7]

Internet-Draft         mVPN Fast Upstream Failover          October 2020


      Optional TLVs is the optional variable-length field that MAY be
      used in the BFD Discriminator attribute for future extensions.
      TLVs MAY be included in a sequential or nested manner.  To allow
      for TLV nesting, it is advised to define a new TLV as a variable-
      length object.  Figure 2 presents the Optional TLV format TLV that
      consists of:

      *  one octet-long field of TLV 's Type value (Section 7.3)

      *  one octet-long field of the length of the Value field in octets

      *  variable length Value field.

      The length of a TLV MUST be multiple of four octets.
<mglt>
I am wondering why the constraint on the
length is not mentioned in the paragraph
associated to the field - as opposed to
a  separate paragraph.

</mglt>
GIM>> There might be a slight confusion due to the use of Length and length. Capitalized - the name of the field which value is the length of the Value field. The last sentence refers to the overall length of a TLV, including lengths of Type, Length and Value fields.

<mglt>
you are correct that might have confused me. 
</mglt>

[..]

8.  Security Considerations

   This document describes procedures based on [RFC6513] and [RFC6514]
   and hence shares the security considerations respectively represented
   in these specifications.

   This document uses p2mp BFD, as defined in [RFC8562], which, in turn,
   is based on [RFC5880].  Security considerations relevant to each
   protocol are discussed in the respective protocol specifications.  An
   implementation that supports this specification MUST use a mechanism
   to control the maximum number of p2mp BFD sessions that can be active
   at the same time.

<mglt>
At a high level view - or at least my
interpretation of it - the document
proposes a mechanism based on BFD to
detect fault in the path.  Upon a fault
detection a fail-over operation is
instructed using BGP. This rocedure is
expected to perform a faster fail-over
than traditional BGP convergence on
maintaining routing tables. Once the
fail over has been performed, BFD is
confirms the new path is "legitimate"
and works.

It seems correct to me that the current
protocol relies on BGP / BFD security.
That said, having BFD authentication
based on MD5 or SHA1 may suggest that
stronger primitives be recommended.
While this does not concerns the current
document, it seems to me that the
information might be relayed to routing
ADs.

What remains unclear to me - and I
assume this might be due to my lake or
expertise in routing area - is the impact
associated to performing a fail-over
both on 1) the data plane and 2) the
standard BGP way to establish routing
tables.

Regarding the data plane, I am wondering
if fail-over results in a lost of
packets for example - I suppose for
example that at least the packets in the
process of being forwarded might be
lost. I believe that providing details
on this may be good.
GIM>> You bring up a very topic for the discussion, thank you. With network failure detection in place, the fail-over can be viewed as the reaction to a network failure.  If that is the case, then packet loss experienced by service due to the fail-over is the result of the network failure. Would you agree with that view? A shorter failure detection interval and faster fail-over should minimize the packet loss and, as a result, the negative impact on the service itself.

<mglt>
sure. If you know the network is down, then fast fail-over is definitively a plus. What I think could be useful is to evaluate the cost associated to a fast-fail-over without any network failure.  This would be useful for an operator to evaluate whether it should spend more time in diagnosing a network failure versus performing a fast-fail-over. 
Typically, if a fast failover comes a no cost at all, one operator would maybe use one exchange to test the liveness of a node rather than 3.      

At that point, it seems to me that additional text coudl be added to characterize the impact. These could be high level and indicative, but it seems to me that knowing these impacts presents some value to the operators.      
</mglt>

If there are any impacts I would like to
understand also in which cases the
decision to perform a failover operation
may result in more harm than the event
that has been over-interpreted. An
hypothetical scenario could be that the
non reception of a BFD packet is
interpreted as a PE being down while it
may not be correct and the PE might have
been simply under stress. A "too fast" fail-over
may over interpreted it and perform a
fail-over. If such things could happen,
an attacker could leverage a micro event
to perform network operation that are
not negligible. Another way to see that
is that an attacker might not have
direct access to the control plan, but
could use the data plan to generate a
stress and sort of control the fail
over. It seems to me that some text
might be welcome to prevent such cases
to happen. This could be guidance for
declaring a tunnel down for example.
GIM>> I agree with your scenario. Over-short detection interval may produce a false-negative transition to the Down state in BFD and thus triggering the fail-over. I think that that is more an operational issue, something that an operator will consider when deploying the mechanism specified in this draft. Resulting from addressing RtgDir review the draft was updated to provide more guidance:
   In many cases, it is not practical to use both protection
   methods at the same time because uncorrelated timers might cause
   unnecessary switchovers and destabilize the network.
<mglt>
Thanks for the feed back, It seems to me important to mention it is not recommended these two mechanism co-exist. 
How to avoid false negative transition might be out of scope of the draft I agree, but it seems to me worth being mentioned especially in relation to the impacts associated to a fail-over.  In case the fast-failover comes with no impact this becomes less of a problem for operator deploying it.
 
</mglt>
Though the text above might not be general, I think that it also applies to the scenario you've presented.

Similarly, it would be good to add some
text regarding the interferences with
the non-fast forwarding fail over when
performed by the standard BGP.
Typically, my impression is that the
fast fail-over mechanism is a local
decision versus the BGP convergence that
is more global. As a result, even with
more time this two mechanisms may come
with different outcomes. One such
example to illustrate my purpose could
be the following. Note that this is only
illustrative of my purpose, and I let
you find and pick on ethat is more
appropriated.   I am thinking of a case
where a standby PE is be shared among
multiple PEs - supposing this situation
could occur.  Typically, if PE_1, PE_2
are shared by PE_a, ..., PE_z. In case
PE_a and PE_b are down, we expect PE_a
to switch to PE_1 and PE_b to switch to
PE_2. It seems to me that BGP would end
up in such situation while a local
decision may end up in PE_a and PE_a to
switch to PE_1.

</mglt>
GIM>> Thank you for the scenario that is very common in deploying protection based on the shared redundant resources. Such schemes, referred to as M:N protection, in addition to using mechanism detecting a network failure, e.g., BFD, require a protocol to coordinate the switchover. This specification applies to a more special deployment scenario where one working PE is protected by one or more standby PEs, i.e., 1:N protection.

<mglt>
I understand the document is addressing a 1:N scenario. That said, if M:N scenario leverage from 1:N protection it seems to me worth raising the issue. 
</mglt>
-- 
last-call mailing list
last-call@xxxxxxxx
https://www.ietf.org/mailman/listinfo/last-call

[Index of Archives]     [IETF Annoucements]     [IETF]     [IP Storage]     [Yosemite News]     [Linux SCTP]     [Linux Newbies]     [Mhonarc]     [Fedora Users]

  Powered by Linux