BGPsec Operational Considerations
Internet Initiative Japan
5147 Crystal Springs
Bainbridge Island
Washington
98110
United States of America
randy@psg.com
Deployment of the BGPsec architecture and protocols has many
operational considerations. This document attempts to collect and
present the most critical and universal. Operational practices are expected to evolve as BGPsec is formalized and initially deployed.
Origin validation based on the Resource Public Key Infrastructure
(RPKI) is in its early phases. As
BGPsec may require
larger memory and/or more modern CPUs, it expected to be deployed
incrementally over a longer time span. BGPsec is a new protocol
with many operational considerations that this document attempts to
describe. As with most operational practices, they will likely change over time.
BGPsec relies on widespread propagation of the RPKI . How the RPKI is distributed and maintained
globally and within an operator's infrastructure may be different
for BGPsec than for origin validation.
BGPsec needs to be spoken only by an Autonomous System's (AS's) eBGP-speaking border
routers. It is designed so that it can be used to protect
announcements that are originated by resource-constrained edge
routers. This has special operational considerations, see .
Different prefixes may have different timing and replay
protection considerations.
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
when, and only when, they appear in all capitals, as shown here.
It is assumed that the reader understands BGP ,
BGPsec , the RPKI , the RPKI Repository Structure
, and Route Origin Authorizations (ROAs)
.
The considerations for RPKI objects (Certificates, Certificate
Revocation Lists (CRLs), manifests , and Ghostbusters Records ),
Trust Anchor Locators (TALs) , cache
behaviors of synchronization, and validation from the section on
RPKI Distribution and Maintenance of apply.
Specific considerations relating to ROA objects do not apply to this
document.
As described in ,
BGPsec-speaking routers are capable of generating their own
public/private key-pairs and having their certificates signed and
published in the RPKI by the RPKI Certification Authority (CA) system, and/or are given
public/private key-pairs by the operator.
A site/operator may use a single certificate/key in all their
routers, one certificate/key per router, or any granularity in
between.
A large operator, concerned that a compromise of one router's key
would make other routers vulnerable, may deploy a more complex
certificate/key distribution burden to reduce this exposure.
At the other end of the spectrum, an edge site with one or two
routers may choose to use a single certificate/key.
In anticipation of possible key compromise, a prudent operator
SHOULD pre-provision each router's 'next' key in the RPKI so that there
is no propagation delay for provisioning the new key.
BGPsec is spoken by edge routers in a network, specifically those that border
other networks/ASes.
In an AS where edge routers speak BGPsec and, therefore, inject
BGPsec paths into the iBGP (internal BGP), Route Reflectors (RRs) MUST have BGPsec
enabled if and only if there are eBGP (external BGP) speakers in their client cone,
i.e., an RR client or the transitive closure of a client's
customers.
A BGPsec-capable router MAY use the data it receives to influence
local policy within its network, see . In
deployment, this policy should fit into the AS's existing policy,
preferences, etc. This allows a network to incrementally deploy
BGPsec-enabled border routers.
eBGP speakers that face more critical peers or upstreams or downstreams
would be candidates for early deployment. Both securing one's own
announcements and validating received announcements should be
considered in partial deployment.
An operator should be aware that BGPsec, as any other policy
change, can cause traffic shifts in their network. And, as with
normal policy shift practice, a prudent operator has the tools and
methods to predict, measure, modify, etc.
On the other hand, an operator wanting to monitor router loading,
shifts in traffic, etc., might deploy incrementally while watching
those and similar effects.
BGPsec does not sign over communities, so they are not formally
trustable. Additionally, outsourcing verification is not a prudent
security practice. Therefore, an eBGP listener SHOULD NOT strongly
trust unsigned security signaling, such as communities, received
across a trust boundary.
An edge site that does not provide transit and trusts its
upstream(s) may only originate a signed prefix announcement and
not validate received announcements.
An operator might need to use hardware with limited resources. In
such cases, BGPsec protocol capability negotiation allows for a
resource-constrained edge router to hold only its own signing
key(s) and sign its announcements, but not receive signed
announcements. Therefore, the router would not have to deal with
the majority of the RPKI, potentially saving the need for
additional hardware.
As the vast majority of ASes are stubs, and they announce the
majority of prefixes, this allows for simpler and less expensive
incremental deployment. It may also mean that edge sites concerned
with routing security will be attracted to upstreams that support
BGPsec.
As BGPsec-signed paths cannot traverse non-BGPsec topology,
partial BGPsec deployment forms islands of assured paths. As
islands grow to touch each other, they become larger islands.
Unlike origin validation based on the RPKI, BGPsec marks a
received announcement as Valid or Not Valid, there is no explicit
NotFound state. In some sense, an unsigned BGP4 path is the
equivalent of NotFound. How this is used in routing is up to the
operator's local policy, similar to origin validation as in
.
As BGPsec will be rolled out over years and does not allow for
intermediate non-signing edge routers, coverage will be spotty for
a long time. This presents a dilemma; should a router evaluating
an inbound BGPsec_PATH as Not Valid be very strict and discard it?
On the other hand, it might be the only path to that prefix, and a
very low local-preference would cause it to be used and propagated
only if there was no alternative. Either choice is reasonable,
but we recommend dropping because of the next point.
Operators should be aware that accepting Not Valid announcements,
no matter the local preference, will often be the equivalent of
treating them as fully Valid. Local preference affects only
routes to the same set of destinations. Consider having a Valid
announcement from neighbor V for prefix 10.0.0.0/16 and a Not
Valid announcement for 10.0.666.0/24 from neighbor I. If local
policy on the router is not configured to discard the Not Valid
announcement from I, then the longest match forwarding will send
packets to neighbor I no matter the value of local preference.
Validation of signed paths is usually deployed at the eBGP
edge.
Local policy on the eBGP edge MAY convey the validation state of
a BGP-signed path through normal local policy mechanisms, e.g.,
setting a BGP community for internal use, or modifying a metric
value such as local-preference or Multi-Exit Discriminator (MED).
Some may choose to use the large Local-Pref hammer. Others may
choose to let AS path rule and set their internal metric, which
comes after AS path in the BGP decision process.
As the mildly stochastic timing of RPKI propagation may cause
version skew across routers, an AS Path that does not validate at
router R0 might validate at R1. Therefore, signed paths that are
Not Valid and yet propagated (because they are chosen as best path)
MUST NOT have signatures stripped and MUST be signed if sent to
external BGPsec speakers.
This implies that updates which a speaker judges to be Not Valid
MAY be propagated to iBGP peers. Therefore, unless local policy
ensures otherwise, a signed path learned via iBGP may be Not Valid.
If needed, the validation state should be signaled by normal local
policy mechanisms such as communities or metrics.
On the other hand, local policy on the eBGP edge might preclude
iBGP or eBGP announcement of signed AS Paths that are Not Valid.
A BGPsec speaker receiving a path SHOULD perform origin
validation per and
.
A route server is usually 'transparent', i.e., does not insert an AS
into the path so as not to increase the AS hop count, and thereby
affect downstream path choices. But, with BGPsec, a client router
R needs to be able to validate paths that are forward signed to
R. But the sending router cannot generate signatures to all the
possible clients. Therefore, a BGPsec-aware route server needs to
validate the incoming BGPsec_PATH and to forward updates that
can be validated by clients that must, therefore, know the route
server's AS. This implies that the route server creates
signatures per client including its own AS in the BGPsec_PATH,
forward signing to each client AS, see . The route server uses a pCount of 0 to not increase the effective AS hop count, thereby
retaining the intent of 'transparency'.
If it is known that a BGPsec neighbor is a transparent route
server, or otherwise may validly use a pCount of 0 (e.g., see ), the router SHOULD be configured to accept and
process a received pCount of 0. Routers MUST disallow a pCount of 0 by default.
To prevent exposure of the internals of the BGP confederations , a BGPsec speaker exporting to a non-member
removes all intra-confederation Secure_Path Segments. Therefore,
signing within the confederation will not cause external confusion
even if non-unique private ASes are used.
For protection from attacks replaying BGP data on the order of a
day or longer old, rekeying routers with new keys (previously)
provisioned in the RPKI is sufficient. For one approach, see
.
A router that once negotiated (and/or sent) BGPsec should not be
expected to always do so.
Like the DNS, the Global RPKI presents only a loosely consistent
view, depending on timing, updating, fetching, etc. Thus, one
cache or router may have different data about a particular prefix
or router than another cache or router. There is no 'fix' for
this, it is the nature of distributed data with distributed
caches.
Operators who manage certificates SHOULD have RPKI Ghostbuster
Records (see ), signed
indirectly by end entity certificates, for those certificates on
which others' routing depends for certificate and/or ROA
validation.
Operators should be aware of impending algorithm transitions,
which will be rare and slow-paced, see .
They should work with their vendors to ensure support for new
algorithms.
As a router must evaluate certificates and ROAs that are time
dependent, routers' clocks MUST be correct to a tolerance of
approximately an hour. The common approach is for operators to
deploy servers that provide time service, such as , to client routers.
If a router has reason to believe its clock is seriously
incorrect, e.g., it has a time earlier than 2011, it SHOULD NOT
attempt to validate incoming updates. It SHOULD defer validation
until it believes it is within reasonable time tolerance.
This document describes operational considerations for the
deployment of BGPsec. The security considerations for BGPsec are
described in .
This document does not require any IANA actions.
BGPsec Protocol Specification
Router Keying for BGPsec
BGPsec Router Certificate Rollover
BGPsec Considerations for Autonomous System (AS) Migration
The author wishes to thank Thomas King, Arnold Nipper, Alvaro
Retana, and the BGPsec design group.