Internet Engineering Task Force (IETF) J. Uttaro

Internet Engineering Task Force (IETF) J. Uttaro
Request for Comments: 9494 Independent Contributor
Updates: 6368 E. Chen
Category: Standards Track Palo Alto Networks
ISSN: 2070-1721 B. Decraene
Orange
J. Scudder
Juniper Networks
November 2023

Long-Lived Graceful Restart for BGP

Abstract

This document introduces a BGP capability called the "Long-Lived
Graceful Restart Capability" (or "LLGR Capability"). The benefit of
this capability is that stale routes can be retained for a longer
time upon session failure than is provided for by BGP Graceful
Restart (as described in RFC 4724). A well-known BGP community
called "LLGR_STALE" is introduced for marking stale routes retained
for a longer time. A second well-known BGP community called
"NO_LLGR" is introduced for marking routes for which these procedures
should not be applied. We also specify that such long-lived stale
routes be treated as the least preferred and that their
advertisements be limited to BGP speakers that have advertised the
capability. Use of this extension is not advisable in all cases, and
we provide guidelines to help determine if it is.

This memo updates RFC 6368 by specifying that the LLGR_STALE
community must be propagated into, or out of, the path attributes
exchanged between the Provider Edge (PE) and Customer Edge (CE)
routers.

Status of This Memo

This is an Internet Standards Track document.

This document is a product of the Internet Engineering Task Force
(IETF). It represents the consensus of the IETF community. It has
received public review and has been approved for publication by the
Internet Engineering Steering Group (IESG). Further information on
Internet Standards is available in Section 2 of RFC 7841.

Information about the current status of this document, any errata,
and how to provide feedback on it may be obtained at
https://www.rfc-editor.org/info/rfc9494.

Copyright Notice

Copyright (c) 2023 IETF Trust and the persons identified as the
document authors. All rights reserved.

This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(https://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Revised BSD License text as described in Section 4.e of the
Trust Legal Provisions and are provided without warranty as described
in the Revised BSD License.

Table of Contents

1. Introduction
2. Terminology
2.1. Definitions
2.2. Abbreviations
2.3. Requirements Language
3. Protocol Extensions
3.1. Long-Lived Graceful Restart Capability
3.2. LLGR_STALE Community
3.3. NO_LLGR Community
4. Theory of Operation
4.1. Use of the Graceful Restart Capability
4.2. Session Resets
4.3. Processing LLGR_STALE Routes
4.4. Route Selection
4.5. Errors
4.6. Optional Partial Deployment Procedure
4.7. Procedures When BGP Is the PE-CE Protocol in a VPN
4.7.1. Procedures When EBGP Is the PE-CE Protocol in a VPN
4.7.2. Procedures When IBGP Is the PE-CE Protocol in a VPN
5. Deployment Considerations
5.1. When BGP Is the PE-CE Protocol in a VPN
5.2. Risks of Depreferencing Routes
6. Security Considerations
7. Examples of Operation
8. IANA Considerations
9. References
9.1. Normative References
9.2. Informative References
Acknowledgements
Contributors
Authors' Addresses

1. Introduction

Routing protocols in general, and BGP in particular, have
historically been designed with a focus on "correctness", where a key
part of correctness is for each network element's forwarding state to
converge to the current state of the network as quickly as possible.
For this reason, the protocol was designed to remove state advertised
by routers that went down (from a BGP perspective) as quickly as
possible. Over time, this has been relaxed somewhat, notably by BGP
Graceful Restart (GR) [RFC4724]; however, the paradigm has remained
one of attempting to rapidly remove stale state from the network.

Over time, two phenomena have arisen that call into question the
underlying assumptions of this paradigm.

1. The widespread adoption of tunneled forwarding infrastructures
(for example, MPLS). Such infrastructures eliminate the risk of
some types of forwarding loops that can arise in hop-by-hop
forwarding; thus, they reduce one of the motivations for strong
consistency between forwarding elements.

2. The increasing use of BGP as a transport for data that is less
closely associated with packet forwarding than was originally the
case. Examples include the use of BGP for auto-discovery
(Virtual Private LAN Service (VPLS) [RFC4761]) and filter
programming (Flow Specification (FLOWSPEC) [RFC8955]). In these
cases, BGP data takes on a character more akin to configuration
than to conventional routing.

The observations above motivate a desire to offer network operators
the ability to choose to retain BGP data for a longer period than has
hitherto been possible when the BGP control plane fails for some
reason. Although the semantics of BGP Graceful Restart [RFC4724] are
close to those desired, several gaps exist, most notably in the
maximum time for which stale information can be retained: Graceful
Restart imposes a 4095-second upper bound.

In this document, we introduce a BGP capability called the "Long-
Lived Graceful Restart Capability". The goal of this capability is
that stale information can be retained for a longer time across a
session reset. We also introduce two BGP well-known communities:

* LLGR_STALE to mark such information, and

* NO_LLGR to indicate that these procedures should not be applied to
the marked route.

Long-lived stale information is to be treated as least preferred, and
its advertisement limited to BGP speakers that support the
capability. Where possible, we reference the semantics of BGP
Graceful Restart [RFC4724] rather than specifying similar semantics
in this document.

The expected deployment model for this extension is that it will only
be invoked for certain address families. This is discussed in more
detail in Section 5. The use of this extension may be combined with
that of conventional Graceful Restart; in such a case, it is invoked
after the conventional Graceful Restart interval has elapsed. When
not combined, LLGR is invoked immediately. Apart from the potential
to greatly extend the timer, the most obvious difference between LLGR
and conventional Graceful Restart is that in LLGR, routes are
"depreferenced"; that is, they are treated as least preferred.
Contrarily, in conventional GR, route preference is not affected.
The design choice to treat long-lived stale routes as least preferred
was informed by the expectation that they might be retained for
(potentially) an almost unbounded period of time; whereas, in the
conventional Graceful Restart case, stale routes are retained for
only a brief interval. In the case of Graceful Restart, the trade-
off between advertising new route status (at the cost of routing
churn) and not advertising it (at the cost of suboptimal or incorrect
route selection) is resolved in favor of not advertising. In the
case of LLGR, it is resolved in favor of advertising new state, using
stale information only as a last resort.

Section 7 provides some simple examples illustrating the operation of
this extension.

2. Terminology

2.1. Definitions

Depreference: A route is said to be depreferenced if it has its
route selection preference reduced in reaction to some event.

Helper: Sometimes referred to as "helper router". During Graceful
Restart or Long-Lived Graceful Restart, the router that detects a
session failure and applies the listed procedures. [RFC4724]
refers to this as the "receiving speaker".

Route: In this document, "route" means any information encoded as
BGP Network Layer Reachability Information (NLRI) and a set of path
attributes. As discussed above, the connection between such routes
and the installation of forwarding state may be quite remote.

Further note that, for brevity, in this document when we reference
conventional Graceful Restart, we cite its base specification,
[RFC4724]. That specification has been updated by [RFC8538]. The
citation to [RFC4724] is not intended to be limiting.

2.2. Abbreviations

CE: Customer Edge (See [RFC4364] for more information on Customer
Edge routers.)

EoR: End-of-RIB (See Section 2 of [RFC4724] for more information on
End-of-RIB markers.)

GR: Graceful Restart (See [RFC4724] for more information on GR.)
This term is also sometimes referred to herein as "conventional
Graceful Restart" or "conventional GR" to distinguish it from the
"Long-Lived Graceful Restart" or "LLGR" defined by this document.

LLGR: Long-Lived Graceful Restart

LLST: Long-Lived Stale Time

PE: Provider Edge (See [RFC4364] for more information on Provider
Edge routers.)

VRF: VPN Routing and Forwarding (See [RFC4364] for more information
on VRF tables.)

2.3. Requirements Language

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in
BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
capitals, as shown here.

3. Protocol Extensions

A BGP capability and two BGP communities are introduced in the
subsections that follow.

3.1. Long-Lived Graceful Restart Capability

The "Long-Lived Graceful Restart Capability", or "LLGR Capability",
(value: 71) is a BGP capability [RFC5492] that can be used by a BGP
speaker to indicate its ability to preserve its state according to
the procedures of this document. If the LLGR capability is
advertised, the Graceful Restart capability [RFC4724] MUST also be
advertised; see Section 4.1.

The capability value consists of zero or more tuples <AFI, SAFI,
Flags, LLST> as follows:

+--------------------------------------------------+
| Address Family Identifier (16 bits) |
+--------------------------------------------------+
| Subsequent Address Family Identifier (8 bits) |
+--------------------------------------------------+
| Flags for Address Family (8 bits) |
+--------------------------------------------------+
| Long-Lived Stale Time (24 bits) |
+--------------------------------------------------+
| ... |
+--------------------------------------------------+
| Address Family Identifier (16 bits) |
+--------------------------------------------------+
| Subsequent Address Family Identifier (8 bits) |
+--------------------------------------------------+
| Flags for Address Family (8 bits) |
+--------------------------------------------------+
| Long-Lived Stale Time (24 bits) |
+--------------------------------------------------+

The meaning of the fields are as follows:

Address Family Identifier (AFI), Subsequent Address Family
Identifier (SAFI):
The AFI and SAFI, taken in combination, indicate that the BGP
speaker has the ability to preserve its forwarding state for the
address family during a subsequent BGP restart. Routes may be
either:

* explicitly associated with a particular AFI and SAFI if using
the encoding described in [RFC4760], or

* implicitly associated with <AFI=IPv4, SAFI=Unicast> if using
the encoding described in [RFC4271].

Flags for Address Family:
This field contains bit flags relating to routes that were
advertised with the given AFI and SAFI.

0 1 2 3 4 5 6 7
+-+-+-+-+-+-+-+-+
|F| Reserved |
+-+-+-+-+-+-+-+-+

The most significant bit is used to indicate whether the state for
routes that were advertised with the given AFI and SAFI has indeed
been preserved during the previous BGP restart. When set (value
1), the bit indicates that the state has been preserved. This bit
is called the "F bit" since it was historically used to indicate
the preservation of forwarding state. Use of the F bit is
detailed in Section 4.2. The remaining bits are reserved and MUST
be set to zero by the sender and ignored by the receiver.

Long-Lived Stale Time:
This time (in seconds) specifies how long stale information (for
this AFI/SAFI) may be retained by the receiver (in addition to the
period specified by the "Restart Time" in the Graceful Restart
Capability). Because the potential use cases for this extension
vary widely, there is no suggested default value for the LLST.

3.2. LLGR_STALE Community

The well-known BGP community LLGR_STALE (value: 0xFFFF0006) can be
used to mark stale routes retained for a longer period of time (see
[RFC1997] for more information on BGP communities). Such long-lived
stale routes are to be handled according to the procedures specified
in Section 4.

An implementation MAY allow users to configure policies that accept,
reject, or modify routes based on the presence or absence of this
community.

3.3. NO_LLGR Community

The well-known BGP community NO_LLGR (value: 0xFFFF0007) can be used
to mark routes that a BGP speaker does not want to be treated
according to these procedures, as detailed in Section 4.

An implementation MAY allow users to configure policies that accept,
reject, or modify routes based on the presence or absence of this
community.

4. Theory of Operation

If a BGP speaker is configured to support the procedures of this
document, it MUST use BGP Capabilities Advertisement [RFC5492] to
advertise the Long-Lived Graceful Restart Capability. The setting of
the parameters for an AFI/SAFI depends on the properties of the BGP
speaker, network scale, and local configuration.

In the presence of the Long-Lived Graceful Restart Capability, the
procedures specified in [RFC4724] continue to apply unless explicitly
revised by this document.

4.1. Use of the Graceful Restart Capability

If the LLGR Capability is advertised, the Graceful Restart capability
MUST also be advertised. If it is not so advertised, the LLGR
Capability MUST be disregarded. The purpose for mandating this is to
enable the reuse of certain base mechanisms that are common to both
"flavors" notably: origination, collection, and processing of EoR as
well as the finite-state-machine modifications and connection-reset
logic introduced by GR.

We observe that, if support for conventional Graceful Restart is not
desired for the session, the conventional GR phase can be skipped by
omitting all AFIs/SAFIs from the GR Capability, advertising a Restart
Time of zero, or both. Section 4.2 discusses the interaction of
conventional and LLGR.

4.2. Session Resets

BGP Graceful Restart [RFC4724] defines conditions under which a BGP
session can reset and have its associated routes retained. If such a
reset occurs for a session in which the LLGR Capability has also been
exchanged, the following procedures apply:

* If the Graceful Restart Capability that was received does not list
all AFIs/SAFIs supported by the session, then the GR Restart Time
shall be deemed zero for those AFIs/SAFIs that are not listed.

* Similarly, if the received LLGR Capability does not list all AFIs/
SAFIs supported by the session, then the Long-Lived Stale Time
shall be deemed zero for those AFIs/SAFIs that are not listed.

The following text in Section 4.2 of [RFC4724] no longer applies:

| If the session does not get re-established within the "Restart
| Time" that the peer advertised previously, the Receiving Speaker
| MUST delete all the stale routes from the peer that it is
| retaining.

and the following procedures are specified instead:

After the session goes down, and before the session is re-
established, the stale routes for an AFI/SAFI MUST be retained. The
interval for which they are retained is limited by the sum of the
Restart Time in the received Graceful Restart Capability and the
Long-Lived Stale Time in the received Long-Lived Graceful Restart
Capability. The timers received in the Long-Lived Graceful Restart
Capability SHOULD be modifiable by local configuration, which may
impose an upper bound, a lower bound, or both on their respective
values.

If the value of the Restart Time or the Long-Lived Stale Time is
zero, the duration of the corresponding period would be zero seconds.
For example, if the Restart Time is zero and the Long-Lived Stale
Time is nonzero, only the procedures particular to LLGR would apply.
Conversely, if the Long-Lived Stale Time is zero and the Restart Time
is nonzero, only the procedures of GR would apply. If both are zero,
none of these procedures would apply, only those of the base BGP
specification [RFC4271] (although EoR would still be used as detailed
in [RFC4724]). And finally, if both are nonzero, then the procedures
would be applied serially: first those of GR and then those of LLGR.
During the first interval, we observe that, while the procedures of
GR are in effect, route preference would not be affected. During the
second interval, while LLGR procedures are in effect, routes would be
treated as least preferred as specified elsewhere in this document.

Once the Restart Time period ends (including the case in which the
Restart Time is zero), the LLGR period is said to have begun and the
following procedures MUST be performed:

* For each AFI/SAFI for which it has received a nonzero Long-Lived
Stale Time, the helper router MUST start a timer for that Long-
Lived Stale Time. If the timer for the Long-Lived Stale Time for
a given AFI/SAFI expires before the session is re-established, the
helper MUST delete all stale routes of that AFI/SAFI from the
neighbor that it is retaining.

* The helper router MUST attach the LLGR_STALE community to the
stale routes being retained. Note that this requirement implies
that the routes would need to be readvertised in order to
disseminate the modified community.

* If any of the routes from the peer have been marked with the
NO_LLGR community, either as sent by the peer or as the result of
a configured policy, they MUST NOT be retained and MUST be removed
as per the normal operation of [RFC4271].

* The helper router MUST perform the procedures listed in
Section 4.3.

Once the session is re-established, the procedures specified in
[RFC4724] apply for the stale routes irrespective of whether the
stale routes are retained during the Restart Time period or the Long-
Lived Stale Time period. However, in the case of consecutive
restarts, the previously marked stale routes MUST NOT be deleted
before the timer for the Long-Lived Stale Time expires.

Similar to [RFC4724], once the LLGR Period begins, the Helper MUST
immediately remove all the stale routes from the peer that it is
retaining for that address family if any of the following occur:

* the F bit for a specific address family is not set in the newly
received LLGR Capability, or

* a specific address family is not included in the newly received
LLGR Capability, or

* the LLGR and accompanying GR Capability are not received in the
re-established session at all.

If a Long-Lived Stale Time timer is running for routes with a given
AFI/SAFI received from a peer, it MUST NOT be updated (other than by
manual operator intervention) until the peer has established and
synchronized a new session. The session is termed "synchronized" for
a given AFI/SAFI once the EoR for that AFI/SAFI has been received
from the peer or once the Selection_Deferral_Timer discussed in
[RFC4724] expires.

The value of a Long-Lived Stale Time in the capability received from
a neighbor MAY be reduced by local configuration.

While the session is down, the expiration of a Long-Lived Stale Time
timer is treated analogously to the expiration of the Restart Time
timer in [RFC4724], other than applying only to the AFI/SAFI it
accompanies. However, the timer continues to run once the session
has re-established. The timer is neither stopped nor updated until
the EoR marker is received for the relevant AFI/SAFI from the peer.
If the timer expires during synchronization with the peer, any stale
routes that the peer has not refreshed are removed. If the session
subsequently resets prior to becoming synchronized, any remaining
routes (for the AFI/SAFI whose LLST timer expired) MUST be removed
immediately.

4.3. Processing LLGR_STALE Routes

A BGP speaker that has advertised the Long-Lived Graceful Restart
Capability to a neighbor MUST perform the following upon receiving a
route from that neighbor with the LLGR_STALE community or upon
attaching the LLGR_STALE community itself per Section 4.2:

* Treat the route as the least preferred in route selection (see
below). See Section 5.2 for a discussion of potential risks
inherent in doing this.

* The route SHOULD NOT be advertised to any neighbor from which the
Long-Lived Graceful Restart Capability has not been received. The
exception is described in Section 4.6. Note that this requirement
implies that such routes should be withdrawn from any such
neighbor.

* The LLGR_STALE community MUST NOT be removed when the route is
further advertised.

4.4. Route Selection

A least preferred route MUST be treated as less preferred than any
other route that is not also least preferred. When performing route
selection between two routes when both are least preferred, normal
tiebreaking applies. Note that this would only be expected to happen
if the only routes available for selection were least preferred; in
all other cases, such routes would have been eliminated from
consideration.

4.5. Errors

If the LLGR Capability is received without an accompanying GR
Capability, the LLGR Capability MUST be ignored, that is, the
implementation MUST behave as though no LLGR Capability has been
received.

4.6. Optional Partial Deployment Procedure

Ideally, all routers in an Autonomous System (AS) would support this
specification before it were enabled. However, to facilitate
incremental deployment, stale routes MAY be advertised to neighbors
that have not advertised the Long-Lived Graceful Restart Capability
under the following conditions:

* The neighbors MUST be internal (Internal BGP (IBGP) or
Confederation) neighbors.

* The NO_EXPORT community [RFC1997] MUST be attached to the stale
routes.

* The stale routes MUST have their LOCAL_PREF set to zero. See
Section 5.2 for a discussion of potential risks inherent in doing
this.

If this strategy for partial deployment is used, the network operator
should set the LOCAL_PREF to zero for all long-lived stale routes
throughout the Autonomous System. This trades off a small reduction
in flexibility (ordering may not be preserved between competing long-
lived stale routes) for consistency between routers that do, and do
not, support this specification. Since the consistency of route
selection can be important for preventing forwarding loops, the
latter consideration dominates.

4.7. Procedures When BGP Is the PE-CE Protocol in a VPN

4.7.1. Procedures When EBGP Is the PE-CE Protocol in a VPN

In VPN deployments (for example, [RFC4364]), External BGP (EBGP) is
often used as a PE-CE protocol. It may be a practical necessity in
such deployments to accommodate interoperation with peer routers that
cannot easily be upgraded to support specifications such as this one.
This leads to a problem: the procedures defined elsewhere in this
document generally prevent LLGR stale routes from being sent across
EBGP sessions that don't support LLGR, but this could prevent the VPN
routes from being used for their intended purpose.

We observe that the principal motivation for restricting the
propagation of "stale" routing information is the desire to prevent
it from spreading without limit once it exits the "safe" perimeter.
We further observe that VPN deployments are typically topologically
constrained, making this concern moot. For this reason, an
implementation MAY advertise stale routes over a PE-CE session, when
explicitly configured to do so. That is, the second rule listed in
Section 4.3 MAY be disregarded in such cases. All other rules
continue to apply. Finally, if this exception is used, the
implementation SHOULD, by default, attach the NO_EXPORT community to
the routes in question, as an additional protection against stale
routes spreading without limit. Attachment of the NO_EXPORT
community MAY be disabled by explicit configuration in order to
accommodate exceptional cases.

See further discussion of using an explicitly configured policy to
mitigate this issue in Section 5.1.

4.7.2. Procedures When IBGP Is the PE-CE Protocol in a VPN

If IBGP is used as the PE-CE protocol, following the procedures of
[RFC6368], then when a PE router imports a VPN route that contains
the ATTR_SET attribute into a destination VRF and subsequently
advertises that route to a CE router:

* If the CE router supports the procedures of this document (in
other words, if the CE router has advertised the LLGR Capability):

In addition to including the path attributes derived from the
ATTR_SET attribute in the advertised route as per [RFC6368],
the PE router MUST also include the LLGR_STALE community if it
is present in the path attributes of the imported route, even
if it is not present in the ATTR_SET attribute.

* If the CE router does not support the procedures of this document:

Then the optional procedures of Section 4.6 MAY be followed,
attaching the NO_EXPORT community and setting the value of
LOCAL_PREF to zero, overriding the value found in the ATTR_SET.

Similarly, when a PE router receives a route from a CE into its VRF
and subsequently exports that route to a VPN address family:

* If the PE router supports the procedures of this document (in
other words, if the PE router has advertised the LLGR Capability):

In addition to including in the VPN route the ATTR_SET derived
from the path attributes as per [RFC6368], the PE router MUST
also include the LLGR_STALE community in the VPN route if it is
present in the path attributes of the route as received from
the CE.

* If the PE router does not support the procedures of this document:

There exists no ideal solution. The CE could advertise a route
with LLGR_STALE, with the understanding that the LLGR_STALE
marking will only be honored by the provider network if
appropriate policy configuration exists on the PE (see
Section 5.1). It is at least guaranteed that LLGR_STALE will
be propagated when the route is propagated beyond the provider
network, or the CE could refrain from advertising the
LLGR_STALE route to the incapable PE.

5. Deployment Considerations

The deployment considerations discussed in [RFC4724] apply to this
document. In addition, network operators are cautioned to carefully
consider the potential disadvantages of deploying these procedures
for a given AFI/SAFI. Most notably, if used for an AFI/SAFI that
conveys conventional reachability information, the use of a long-
lived stale route could result in a loss of connectivity for the
covered prefix. This specification takes pains to mitigate this risk
where possible by making such routes least preferred and by
restricting the scope of such routes to routers that support these
procedures (or, optionally, a single Autonomous System, see
Section 4.6). However, if a stale route is chosen as best for a
given prefix, then according to the normal rules of IP forwarding,
that route will be used for matching destinations, even if a non-
stale less specific matching route is also available. Networks in
which the deployment of these procedures would be especially
concerning include those that do not use "tunneled" forwarding (in
other words, those using conventional hop-by-hop forwarding).

Implementations MUST NOT enable these procedures by default. They
MUST require affirmative configuration per AFI/SAFI in order to
enable them.

The procedures of this document do not alter the route resolvability
requirement of Section 9.1.2.1 of [RFC4271]. Because of this, it
will commonly be the case that "stale" IBGP routes will only continue
to be used if the router depicted in the next hop remains resolvable,
even if its BGP component is down. Details of IGP fault-tolerance
strategies are beyond the scope of this document. In addition to the
foregoing, it may be advisable to check the viability of the next hop
through other means, for example, Bidirectional Forwarding Detection
(BFD) [RFC5880]. This may be especially useful in cases where the
next hop is known directly at the network layer, notably EBGP.

As discussed in this document, after a BGP session goes down and
before the session is re-established, stale routes may be retained
for up to two consecutive periods, controlled by the Restart Time and
the Long-Lived Stale Time, respectively:

* During the first period, routing churn would be prevented, but
with potential persistent packet loss.

* During the second period, potential persistent packet loss may be
reduced, but routing churn would be visible throughout the
network.

The setting of the relevant parameters for a particular application
should take into account trade-offs, network dynamics, and potential
failure scenarios. If needed, the first period can be bypassed
either by local configuration or by setting the Restart Time in the
Graceful Restart Capability to zero and/or not listing the AFI/SAFI
in that capability.

The setting of the F bit (and the Forwarding State bit of the
accompanying GR Capability) depends, in part, on deployment
considerations. The F bit can be understood as an indication that
the Helper should flush associated routes (if the bit is left clear).
As discussed in Section 1, an important use case for LLGR is for
routes that are more akin to configuration than to conventional
routing. For such routes, it may make sense to always set the F bit,
regardless of other considerations. Likewise, for control-plane-only
entities, such as dedicated route reflectors that do not participate
in the forwarding plane, it makes sense to always set the F bit.
Overall, the rule of thumb is that if loss of state on the restarting
router can reasonably be expected to cause a forwarding loop or
persistent packet loss, the F bit should be set scrupulously
according to whether state has been retained. Specifics of whether
or not the F bit is set are implementation dependent and may also be
controlled by configuration. Also, for every AFI/SAFI represented in
the LLGR Capability that is also represented in the GR Capability,
there will be two corresponding F bits: the LLGR F bit and the GR F
bit. If the LLGR F bit is set, the corresponding GR F bit should
also be set, since to do otherwise would cause the state to be
cleared on the Receiving Router per the normal rules of GR, violating
the intent of the set LLGR bit.

5.1. When BGP Is the PE-CE Protocol in a VPN

As discussed in Section 4.7, it may be necessary for a PE to
advertise stale routes to a CE in some VPN deployments, even if the
CE does not support this specification. In that case, the operator
configuring their PE to advertise such routes should notify the
operator of the CE receiving the routes, and the CE should be
configured to depreference the routes.

Similarly, it may be necessary for a CE to advertise stale routes to
a PE, even if the PE does not support this specification. In that
case, the operator configuring their CE to advertise such routes
should notify the operator of the PE receiving the routes, and the PE
should be configured to depreference the routes.

Typical BGP implementations will be able to be configured to
depreference routes by matching on the LLGR_STALE community and
setting the LOCAL_PREF for matching routes to zero, similar to the
procedure described in Section 4.6.

5.2. Risks of Depreferencing Routes

Depreferencing EBGP routes is considered safe, no different from the
common practice of applying a routing policy to an EBGP session.
However, the same is not always true of IBGP.

Consistent route selection is a fundamental tenet of IBGP correctness
and safe operation in hop-by-hop routed networks. When routers
within an AS apply different criteria in selecting routes, they can
arrive at inconsistent route selections. This can lead to the
formation of forwarding loops unless some form of tunneled forwarding
is used to prevent "core" routers from making a (potentially
inconsistent) forwarding decision based on the IP header.

This specification uses the state of a peering session as an input to
the selection criteria, depreferencing routes that are associated
with a session that has gone down but that have not yet aged out.
Since different routers within an AS might have different notions as
to whether their respective sessions with a given peer are up or
down, they might apply different selection criteria to routes from
that peer. This could result in a forwarding loop forming between
such routers.

For an example of such a forwarding loop, consider the following
simple topology:

A ---- B ---- C ------------------------- D
^ ^
| |
R1 R2

Figure 1

In this example, A - D are routers with a full mesh of IBGP sessions
between them (the sessions are not shown). The short links have unit
cost, the long link has cost 5. Routers A and D are AS border
routers, each advertising some route, R, with the same LOCAL_PREF
into the AS: denoted R1 and R2 in the diagram. In ordinary
operation, it can be seen that routers B and C will select R1 for
forwarding and will forward toward A.

Suppose that the session between A and B goes down for some reason,
and it stays down long enough for LLGR processing to be invoked on B.
Then, on B, route R1 will be depreferenced, leading to the selection
of R2 by B. However, C will continue to prefer R1. In this case, it
can be seen that a forwarding loop for packets destined to R would
form between B and C. (We note that other forwarding loop scenarios
can be constructed for conventional GR, but these are generally
considered less severe since GR can remain in effect for a much more
limited interval.)

The potential benefits of this specification can outweigh the risks
discussed above, as long as care is exercised in deployment. The
cardinal rule to be followed is that if a given set of routes is
being used within an AS for hop-by-hop forwarding, enabling LLGR
procedures is not recommended. If tunneled forwarding (such as MPLS)
is used within the AS, or if routes are being used for purposes other
than hop-by-hop forwarding, less caution is needed; however, the
operator should still carefully consider the consequences of enabling
LLGR.

6. Security Considerations

The security implications of the LLGR mechanism defined in this
document are akin to those incurred by the maintenance of stale
routing information within a network. However, since the retention
time may be much longer, the window during which certain attacks are
feasible may substantially increase. This is particularly relevant
when considering the maintenance of routing information that is used
for service segregation, such as MPLS label entries.

For MPLS VPN services, the effectiveness of the traffic isolation
between VPNs relies on the correctness of the MPLS labels between
ingress and egress PEs. In particular, when an egress PE withdraws a
label L1 allocated to a VPN1 route, this label must not be assigned
to a VPN route of a different VPN until all ingress PEs stop using
the old VPN1 route using L1.

Such a corner case may happen today if the propagation of VPN routes
by BGP messages between PEs takes more time than the label
reallocation delay on a PE. Given that we can generally bound the
worst-case BGP propagation time to a few minutes (for example, 2-5
minutes), the security breach will not occur if PEs are designed to
not reallocate a previously used and withdrawn label before a few
minutes.

The problem is made worse with BGP GR between PEs because VPN routes
can be stalled for a longer period of time (for example, 20 minutes).

This is further aggravated by the LLGR extension specified in this
document because VPN routes can be stalled for a much longer period
of time (for example, 2 hours, 1 day).

In order to exploit the vulnerability described above, an attacker
needs to engineer a specific LLGR state between two PE devices and
also cause the label reallocation to occur such that the two
topologies overlap. To avoid the potential for a VPN breach, the
operator should ensure that the lower bound for label reuse is
greater than the upper bound on the LLST before enabling LLGR for a
VPN address family. Section 4.2 discusses the provision of an upper
bound on LLST. Details of features for setting a lower bound on
label reuse time are beyond the scope of this document; however,
factors that might need to be taken into account when setting this
value include:

* The load of the BGP route churn on a PE (in terms of the number of
VPN labels advertised and the churn rate).

* The label allocation policy on the PE, which possibly depends upon
the size of the pool of the VPN labels (which can be restricted by
hardware considerations or other MPLS usages), the label
allocation scheme (for example, per route or per VRF/CE), and the
reallocation policy (for example, least recently used label).

Note that [RFC4781], which defines the Graceful Restart Mechanism for
BGP with MPLS, is also applicable to LLGR.

7. Examples of Operation

For illustrative purposes, we present a few examples of how this
specification might be used in practice. These examples are neither
exhaustive nor normative.

Consider the following scenario: A border router, ASBR1, has an IBGP
peering with a route reflector, RR1, from which it learns routes. It
has an EBGP peering with an external peer, EXT, to which it
advertises those routes. The external peer has advertised the GR and
LLGR Capabilities to ASBR1. ASBR1 is configured to support GR and
LLGR on its sessions with RR1 and EXT. RR1 advertises a GR Restart
Time of 1 (second) and an LLST of 3600 (seconds):

+==========+=====================================================+
| Time | Event |
+==========+=====================================================+
| t | ASBR1's IBGP session with RR fails. ASBR1 retains |
| | RR's routes according to the rules of GR [RFC4724]. |
+----------+-----------------------------------------------------+
| t+1 | GR Restart Time expires. ASBR1 transitions RR's |
| | routes to long-lived stale routes by attaching the |
| | LLGR_STALE community and depreferencing them. |
| | However, since it has no backup routes, it |
| | continues to make use of them. It re-announces |
| | them to EXT with the LLGR_STALE community attached. |
+----------+-----------------------------------------------------+
| t+1+3600 | LLST expires. ASBR1 removes RR's stale routes from |
| | its own RIB and sends BGP updates to withdraw them |
| | from EXT. |
+----------+-----------------------------------------------------+

Table 1

Next, imagine the same scenario, but suppose RR1 advertised a GR
Restart Time of zero, effectively disabling GR. Equally, ASBR1 could
have used a local configuration to override RR1's offered Restart
Time, setting it to a locally configured value of zero:

+==========+=======================================================+
| Time | Event |
+==========+=======================================================+
| t | ASBR1's IBGP session with RR fails. ASBR1 |
| | transitions RR's routes to long-lived stale routes by |
| | attaching the LLGR_STALE community and depreferencing |
| | them. However, since it has no backup routes, it |
| | continues to make use of them. It re-announces them |
| | to EXT with the LLGR_STALE community attached. |
+----------+-------------------------------------------------------+
| t+0+3600 | LLST expires. ASBR1 removes RR's stale routes from |
| | its own RIB and sends BGP updates to withdraw them |
| | from EXT. |
+----------+-------------------------------------------------------+

Table 2

Next, imagine the original scenario, but consider that the ASBR1-RR1
session comes back up and becomes synchronized 180 seconds after the
failure was detected:

+=========+=====================================================+
| Time | Event |
+=========+=====================================================+
| t | ASBR1's IBGP session with RR fails. ASBR1 retains |
| | RR's routes according to the rules of GR [RFC4724]. |
+---------+-----------------------------------------------------+
| t+1 | GR Restart Time expires. ASBR1 transitions RR's |
| | routes to long-lived stale routes by attaching the |
| | LLGR_STALE community and depreferencing them. |
| | However, since it has no backup routes, it |
| | continues to make use of them. It re-announces |
| | them to EXT with the LLGR_STALE community attached. |
+---------+-----------------------------------------------------+
| t+1+179 | Session is re-established and resynchronized. |
| | ASBR1 removes the LLGR_STALE community from RR1's |
| | routes and re-announces them to EXT with the |
| | LLGR_STALE community removed. |
+---------+-----------------------------------------------------+

Table 3

Finally, imagine the original scenario, but consider that EXT has not
advertised the LLGR Capability to ASBR1:

+==========+======================================================+
| Time | Event |
+==========+======================================================+
| t | ASBR1's IBGP session with RR fails. ASBR1 retains |
| | RR's routes according to the rules of GR [RFC4724]. |
+----------+------------------------------------------------------+
| t+1 | GR Restart Time expires. ASBR1 transitions RR's |
| | routes to long-lived stale routes by attaching the |
| | LLGR_STALE community and depreferencing them. |
| | However, since it has no backup routes, it continues |
| | to make use of them. It withdraws them from EXT. |
+----------+------------------------------------------------------+
| t+1+3600 | LLST expires. ASBR1 removes RR's stale routes from |
| | its own RIB. |
+----------+------------------------------------------------------+

Table 4

8. IANA Considerations

This document defines a BGP capability called the "Long-Lived
Graceful Restart Capability". IANA has assigned a value of 71 from
the "Capability Codes" registry.

This document introduces two BGP well-known communities:

* the first called "LLGR_STALE" for marking long-lived stale routes,
and

* the second called "NO_LLGR" for marking routes that should not be
retained if stale.

IANA has assigned these well-known community values 0xFFFF0006 and
0xFFFF0007, respectively, from the "BGP Well-known Communities"
registry.

IANA has established a registry called the "Long-Lived Graceful
Restart Flags for Address Family" registry under the "Border Gateway
Protocol (BGP) Parameters" group. The registration procedures are
Standards Action (see [RFC8126]). The registry is initially
populated as follows:

+==============+=======================+============+===========+
| Bit Position | Name | Short Name | Reference |
+==============+=======================+============+===========+
| 0 | Preservation of state | F | RFC 9494 |
+--------------+-----------------------+------------+-----------+
| 1-7 | Unassigned | | |
+--------------+-----------------------+------------+-----------+

Table 5

9. References

9.1. Normative References

[RFC1997] Chandra, R., Traina, P., and T. Li, "BGP Communities
Attribute", RFC 1997, DOI 10.17487/RFC1997, August 1996,
<https://www.rfc-editor.org/info/rfc1997>.

[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/info/rfc2119>.

[RFC4271] Rekhter, Y., Ed., Li, T., Ed., and S. Hares, Ed., "A
Border Gateway Protocol 4 (BGP-4)", RFC 4271,
DOI 10.17487/RFC4271, January 2006,
<https://www.rfc-editor.org/info/rfc4271>.

[RFC4724] Sangli, S., Chen, E., Fernando, R., Scudder, J., and Y.
Rekhter, "Graceful Restart Mechanism for BGP", RFC 4724,
DOI 10.17487/RFC4724, January 2007,
<https://www.rfc-editor.org/info/rfc4724>.

[RFC4760] Bates, T., Chandra, R., Katz, D., and Y. Rekhter,
"Multiprotocol Extensions for BGP-4", RFC 4760,
DOI 10.17487/RFC4760, January 2007,
<https://www.rfc-editor.org/info/rfc4760>.

[RFC5492] Scudder, J. and R. Chandra, "Capabilities Advertisement
with BGP-4", RFC 5492, DOI 10.17487/RFC5492, February
2009, <https://www.rfc-editor.org/info/rfc5492>.

[RFC6368] Marques, P., Raszuk, R., Patel, K., Kumaki, K., and T.
Yamagata, "Internal BGP as the Provider/Customer Edge
Protocol for BGP/MPLS IP Virtual Private Networks (VPNs)",
RFC 6368, DOI 10.17487/RFC6368, September 2011,
<https://www.rfc-editor.org/info/rfc6368>.

[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
May 2017, <https://www.rfc-editor.org/info/rfc8174>.

[RFC8538] Patel, K., Fernando, R., Scudder, J., and J. Haas,
"Notification Message Support for BGP Graceful Restart",
RFC 8538, DOI 10.17487/RFC8538, March 2019,
<https://www.rfc-editor.org/info/rfc8538>.

9.2. Informative References

[RFC4364] Rosen, E. and Y. Rekhter, "BGP/MPLS IP Virtual Private
Networks (VPNs)", RFC 4364, DOI 10.17487/RFC4364, February
2006, <https://www.rfc-editor.org/info/rfc4364>.

[RFC4761] Kompella, K., Ed. and Y. Rekhter, Ed., "Virtual Private
LAN Service (VPLS) Using BGP for Auto-Discovery and
Signaling", RFC 4761, DOI 10.17487/RFC4761, January 2007,
<https://www.rfc-editor.org/info/rfc4761>.

[RFC4781] Rekhter, Y. and R. Aggarwal, "Graceful Restart Mechanism
for BGP with MPLS", RFC 4781, DOI 10.17487/RFC4781,
January 2007, <https://www.rfc-editor.org/info/rfc4781>.

[RFC5880] Katz, D. and D. Ward, "Bidirectional Forwarding Detection
(BFD)", RFC 5880, DOI 10.17487/RFC5880, June 2010,
<https://www.rfc-editor.org/info/rfc5880>.

[RFC8126] Cotton, M., Leiba, B., and T. Narten, "Guidelines for
Writing an IANA Considerations Section in RFCs", BCP 26,
RFC 8126, DOI 10.17487/RFC8126, June 2017,
<https://www.rfc-editor.org/info/rfc8126>.

[RFC8955] Loibl, C., Hares, S., Raszuk, R., McPherson, D., and M.
Bacher, "Dissemination of Flow Specification Rules",
RFC 8955, DOI 10.17487/RFC8955, December 2020,
<https://www.rfc-editor.org/info/rfc8955>.

Acknowledgements

We would like to thank Nabil Bitar, Martin Djernaes, Roberto
Fragassi, Jeffrey Haas, Jakob Heitz, Daniam Henriques, Nicolai
Leymann, Mike McBride, Paul Mattes, John Medamana, Pranav Mehta, Han
Nguyen, Saikat Ray, Valery Smyslov, and Bo Wu for their valuable
input and contributions to the discussion and solution.

Contributors

Clarence Filsfils
Cisco Systems
1150 Brussels
Belgium
Email: [email protected]

Pradosh Mohapatra
Sproute Networks
Email: [email protected]

Yakov Rekhter

Eric Rosen
Email: [email protected]

Rob Shakir
Google, Inc.
1600 Amphitheatre Parkway
Mountain View, CA 94043
United States of America
Email: [email protected]

Adam Simpson
Nokia
Email: [email protected]

Authors' Addresses

James Uttaro
Independent Contributor
Email: [email protected]

Enke Chen
Palo Alto Networks
Email: [email protected]

Bruno Decraene
Orange
Email: [email protected]

John G. Scudder
Juniper Networks
Email: [email protected]