ferrix/bin/bgp/README.md

BGP

This crate provides the bgp binary, which runs the Border Gateway Protocol daemon together with its read-only HTTP status interface.

cargo run -- path/to/config.textpb

--check reads the configuration, says whether it is in order, and stops without listening on anything or touching a route: for trying a change before the daemon is restarted with it.

The configuration format is defined in crates/server/proto/config.proto; testdata/server.textpb is a worked example. The daemon is configured entirely through that file: peers, and the routes it originates (originate { prefix: ... }). There is no way to change either at runtime.

When http_endpoint is set the binary also serves:

Path	Purpose
GET /statusz	Status page: configuration, peers and session state, originated and received routes, live event log
GET /api/status	Router ID, ASN, uptime and table sizes
GET /api/config	The configuration file as read at startup
GET /api/peers	Configured peers and established sessions
GET /api/routes/local	Originated routes
GET /api/routes/received	Routes learned from peers; see "Finding a route"
GET /api/routes/rejected	Routes peers announced that were refused, by peer
GET /api/rpki	Validation with the RPKI: the cache, the data and its age
GET /api/kernel	The routes in the kernel: how many, and why the others are not; the machine's interfaces
GET /api/events	Server-sent events for session and route changes
GET /metricz	Prometheus metrics

Everything under it is read-only and unauthenticated. Bind it to loopback or put it behind a reverse proxy.

What it does today

It establishes sessions (both dialling configured peers and accepting connections, with RFC 4271 collision detection), announces its originated routes to every peer with itself as next hop, judges what peers send by each peer's import policy, keeps what passes, chooses a best route for each prefix, and on Linux can put those best routes in the kernel so that the machine forwards by them. Peers whose export says so are told the best of what the others announced, as the roles of RFC 9234 allow.

Who may connect

Connections are taken only from the addresses of configured peers, and from each of those only a few at a time. Anything else is closed unread and counted in bgp_connections_refused_total. accept_unconfigured_peers: true lets any host open a session as any AS; it exists for tests on a closed network.

What is taken from a peer

Every route a peer announces is judged before it goes into the routing table (import in a peer's configuration; see config.proto):

• Refused for any peer unless turned off: prefixes that are not routed on the Internet (private, documentation, multicast, reserved), prefixes longer than /24 or /48, a default route, paths with a private or reserved AS number, paths that do not begin with the peer's own AS, and paths with an AS_SET in them (RFC 9774).
• The next hop has to be the peer itself (accept_third_party_next_hop turns that off; a peer whose role is ROUTE_SERVER is exempt). A next hop is where our traffic will be sent: a peer may draw it to itself, not point it at another member of an exchange or at a machine on some other network of ours.
• The prefixes this router originates, and anything inside them, are not taken from another network, unless the peer's prefixes list names them (a customer with part of our space).
• Always refused: a path with our own AS number in it, and a next hop that is no address to send to (unspecified, loopback, multicast, our own).
• If configured: a list of the prefixes the peer may announce and of the AS numbers its routes may originate from.
• max_prefixes: one route more ends the session with Cease / Maximum Number of Prefixes Reached and keeps the peer down for ten minutes.

A configured peer is one router: a second connection from its address under another BGP Identifier is refused, so the limits above are not multiplied by opening more sessions. And a peer whose sessions keep coming up and failing within a minute is kept down, from the third time for a second and for twice as long each time after, up to five minutes: every such session costs a look through the whole table.

A refused route is dropped, along with the route it would have replaced. bgp_routes_rejected_total counts them by reason, and the status page shows how many routes each peer has and how many were refused.

Passing routes on

A peer whose export has learned_routes: true is told the best route of every prefix the others announced, not only what we originate. The rules are the usual ones and cannot be turned off: not back to the peer a route came from, not to a peer whose AS is on its path already, not from one peer in our own AS to another, not past NO_EXPORT or NO_ADVERTISE. A route goes out with our AS added to its path and ourselves as next hop, without the MED it came with; to a peer in our own AS the path is as it was and LOCAL_PREF goes along; to a ROUTE_SERVER_CLIENT path, MED and (where the client can use it) next hop are left as they came, as a route server sends them (RFC 7947).

What goes where is decided by roles (RFC 9234). A route learned from a provider, a peer or a route server is marked "only to customer" on the way in and goes on to customers only; a customer's route goes to everyone. So with roles on every peer, learned_routes makes us transit for our customers and nothing else. A peer without a role has no such rule, and is told every route; the daemon warns about that at startup. The top-level export is for every peer that has none of its own:

export { learned_routes: true }
peers { name: "customer-a" role: CUSTOMER ... }
peers { name: "transit-a" role: PROVIDER ... }

Marks. import puts communities on every route taken from a peer, and takes some off first; export does the same for every route announced to one. Where a route was learned, which router it went through:

peers {
  name: "ixp-a-rs"
  import {
    remove_communities: "64601:*"        # nobody else's word for our marks
    add_communities: "64601:1001"        # learned at IXP A
    add_large_communities: "64601:1:1001"
    local_pref: 200
  }
}
export { add_large_communities: "64601:2:1" }   # went through router 1

Patterns for remove_communities are 64601:100, 64601:*, *:*, or with three parts for large communities. import.local_pref is what route selection goes by: of two routes for a prefix the one with the higher is chosen, whatever else they have.

Control communities. With control_communities: true, a route's communities say which peers it is passed on to, in the form Internet exchanges' route servers use (L our AS, P a peer's):

Standard	Large
0:P	L:0:P	not to the peer with AS P
L:P	L:1:P	to the peer with AS P, and then only to peers named so
0:L	L:0:0	to nobody

"Not" wins over "to"; a route with none of them goes to everyone it otherwise would. Standard communities only have room for AS numbers below 65536; the large forms work for any. They are taken off a route as it is passed on, and they hold for routes we originate too. With this on, standard communities L:n mean "to the peer with AS n", so mark routes with large communities (L:2:n and up) instead.

Cost. A decision is a few comparisons on the route's communities and path. Routes that go out alike go in one UPDATE: changes are gathered for 50 ms before they go, and a peer that is given the whole table (when it comes up, or asks for it again) has it gathered first, so that the prefixes that share attributes, which are spread over the table's shards, go together. At most four sessions gather a table at once; the others wait their turn. cargo bench -p bgp-bench --bench transit, with one peer sending 1.23 million routes over a session and another being told of them, on a laptop: in our table after 0.47 s, all of it heard by the other peer after 2.1 s, in UPDATEs of four routes each on average.

Not there yet. Per-client best path selection for a route server (every client is told the one best route); ADD-PATH for learned routes (only the best is sent); route reflection; a way to send pre-built messages to several peers at once.

Route origin validation (RPKI)

With rpki { caches { address: "..." port: 3323 } } in the configuration, the daemon keeps the validated ROA payloads of a validator (Routinator, rpki-client with StayRTR, FORT) over the RPKI-to-Router protocol, and validates every route from peers outside its own AS (RFC 6811):

• invalid (the holder of the addresses authorised another origin, or a shorter prefix): refused, like any route the import policy refuses, unless the peer has accept_rpki_invalid: true;
• valid and not found: taken. Validity does not otherwise change which route is chosen.

The data changes all the time, and the routes follow without the peer having to send them again: a route in the table that becomes invalid is taken out, and a route refused as invalid is kept aside ("held") and taken when it no longer is. If no cache confirms the data for the expire interval (two hours unless the cache or the configuration says otherwise) it is dropped, and nothing is invalid any more.

The status page shows the cache, how fresh the data is, each route's validity, and how many of each peer's routes are held. /api/rpki has the same as JSON; bgp_rpki_vrps{family}, bgp_rpki_aspas and bgp_rpki_updates_total are on /metricz. Refused routes are counted under the reasons RPKI invalid and ASPA invalid.

The connection to the cache is plain TCP, without authentication: run the validator on the same host or on a network you trust.

Roles and route leaks (RFC 9234)

role on a peer says what the peer is to us: PROVIDER, CUSTOMER, PEER, ROUTE_SERVER or ROUTE_SERVER_CLIENT. It is not for peers in our own AS. With it:

• The two ends have to agree. Our OPEN says what we are to the peer (its provider, if it is our customer). A peer whose OPEN says it is something else than configured is refused with OPEN Message Error / Role Mismatch: one side thinking "customer" and the other "peer" is how leaks happen. A peer that says nothing (it does not implement the RFC) is accepted, unless require_role: true. The status page shows the role, with a tick when the peer confirmed it.
• Routes are marked "only to customer" (the OTC attribute) when they come in from a provider, a peer or a route server, and when ours go out to a customer, a peer or a route server client. The mark stays on the route.
• A marked route coming the wrong way is a leak and is refused: from a customer or a route server client always, and from a peer unless the peer put the mark there itself. The reason shown is route leak (Only to Customer). No setting turns this off, as the RFC requires.

This catches leaks even where nobody on the path has an ASPA, as long as the networks on either side of the leaker mark their routes.

AS path verification (ASPA)

Origin validation says who may originate a prefix. It says nothing about the rest of the path, so it does not catch a route leak (a network passing its provider's or peer's routes on to another provider or peer), nor a hijack that keeps the right origin at the end of a made-up path. ASPAs do: in one, a network lists its providers, and a path that goes from that network to anyone else, where only a provider could come next, is invalid (draft-ietf-sidrops-aspa-verification).

Two things are needed for paths to be verified:

• a validator that hands out ASPAs, which takes version 2 of the RPKI-to-Router protocol (draft-ietf-sidrops-8210bis). The daemon asks for version 2 and falls back to 1 or 0 if that is what the validator speaks; the status page says which is in use, and warns when it is one without ASPAs;
• role on the peer, saying what the peer is to us, because what a path may look like depends on who hands it over:

role	The peer	Its paths
CUSTOMER	buys transit from us	must lead up through providers all the way
PEER	exchanges its own and its customers' routes with us	the same
ROUTE_SERVER	is an exchange's route server	the same; the path need not begin with the route server's AS
ROUTE_SERVER_CLIENT	is a client of ours, where we are the route server	the same
PROVIDER	sells us transit	may lead up, across one peering, and down

A peer without a role has its origins validated and its paths not.

A path is invalid (refused, unless the peer has accept_aspa_invalid: true), valid, or unknown, which is what most paths are while few networks have an ASPA; unknown ones are taken. A path with an AS_SET in it is invalid, and so is one that does not begin with the peer's AS (other than from a route server). As with origins, routes follow the data: held while invalid, taken when they no longer are.

rpki { caches { address: "127.0.0.1" port: 3323 } }
peers {
    name: "transit"
    endpoint { address: "192.0.2.1" }
    remote_asn: 64496
    role: PROVIDER
}

IPv4 over a link that has only IPv6 addresses (RFC 8950)

A session over IPv6 carries IPv4 routes too, with IPv6 next hops, if the peer agrees: the daemon offers the Extended Next Hop capability on every session over IPv6 whose address families include IPv4 unicast, and nothing has to be configured for it on this side. (On FRR it is neighbor ... capability extended-nexthop, on BIRD extended next hop on in the IPv4 channel.) The peers table on the status page says "IPv4 with IPv6 next hops" for a session that agreed on it.

Our IPv4 routes then go out with the session's own IPv6 address as next hop, and the peer's arrive with its IPv6 address, or with a global and a link-local one. In the kernel such a route is an IPv4 route by way of an IPv6 neighbor (10.2.0.0/24 via inet6 2001:db8::2 dev eth0, Linux 5.2 and later).

Peers at link-local addresses

A link with no addresses but the link-local ones every interface has (fe80::/10) is enough for a session. Such an address means something only on its link, so the peer is the address and the interface together:

peers {
  name: "upstream"
  endpoint { address: "fe80::1%eno3" }
  remote_asn: 64602
}

• The interface is looked up by its name each time it is needed: one that is not there yet, or is made anew with another number (a tunnel that is restarted), is found when it is. Until then the peer's state says there is no such interface.
• The same address on two interfaces is two peers. A link-local address without an interface is refused, as is an interface on any other address.
• To take the peer's connections the daemon must listen on :: (as it does by default), not on one global address.
• Both address families go over the session (IPv4 as above, if the peer agrees). Our routes go with our link-local address as next hop, in both places MP_REACH_NLRI has for one, as FRR sends its own. The peer's next hop is taken in any of the forms in use: the link-local address alone, twice, or after ::. In the kernel the routes go by way of that address on the session's interface (10.2.0.0/24 via inet6 fe80::1 dev eno3).
• local_address, if set to a link-local address of ours, is taken to be on the peer's interface.

With FRR, configure the peer by interface (neighbor eno3 interface remote-as ..., "BGP unnumbered"). FRR then learns who is on the link from IPv6 router advertisements, which this daemon does not send: run radvd (or the like) on the interface beside it. FRR's older neighbor fe80::1 interface eno3 brings the session up but FRR 10.5 then takes every route from us for unreachable. BIRD needs only neighbor fe80::1%eno3 as ....

Forwarding: routes in the kernel (Linux)

With kernel {} in the configuration the daemon puts the best route of every prefix in the kernel's routing table, over netlink, and keeps it there as the best routes change. Without it the daemon only speaks BGP and the machine's forwarding is not touched.

kernel {
    table: 254     # the main table; the default
    metric: 20     # the default
}

• What is installed. A best route whose next hop is a neighbor: an address in the network of one of the machine's interfaces that is up. For an IPv6 route with a link-local next hop, the interface is the one the peer is a neighbor on. A route whose next hop is further away (a route from a peer in our own AS that did not set itself as next hop, say) cannot be installed: looking up the way to such a next hop through other routes is not done yet.

• Routes that cannot be installed are not chosen. With kernel {}, a route whose next hop cannot be reached is not taken into the table, so that it cannot be the best route of its prefix while another peer has one that works (RFC 4271 §9.1.2.1). It is held aside, counted as refused for "next hop unreachable" and shown as held on the status page, and taken as soon as the machine's interfaces change so that it can be reached; a route in the table whose next hop goes away is held in the same way. Without kernel {} the daemon does not know what can be reached, and takes every route.

• Nexthop objects. Where the kernel has them (Linux 5.3 and later) a gateway is put there once, as a nexthop object, and routes name it; ip route shows such routes with nhid, and ip nexthop the objects. The kernel then checks the gateway once and not for every route, which makes a full table about a third faster to install, and lets all of them be removed at once when the daemon stops. The objects the daemon makes are its own: it does not use anyone else's, puts its own back if one is removed, and removes them when it stops. without_nexthop_objects: true has every route carry its own gateway instead, as on older kernels.

• What is the machine's own. A route for a network the machine is itself on, or for part of one, is never installed, whatever the policy let into the table: the kernel reaches those directly. A peer on a shared LAN that announces part of the LAN would otherwise take the traffic to its neighbors, and our sessions with them. The status page counts such routes as "inside a network of this machine".

• What is ours. Routes are marked proto bgp. Every route with that mark in the configured table with the configured metric is taken to be this daemon's, and no other route is ever changed or removed. So do not run two daemons that install BGP routes into the same table with the same metric.

• Starting and stopping. At startup, routes of ours left by an earlier run are kept, so that the machine goes on forwarding while the peers send their tables again; those that no peer has announced again after stale_seconds (120 unless set) are removed. On SIGINT or SIGTERM the daemon closes its sessions and then removes its routes. After a crash they stay until the next start.

• Interfaces. The daemon learns the machine's interfaces and addresses from the kernel and is told when they change; a second after the last change, and five seconds after the first at the latest, it works all routes out again, since any next hop may resolve differently.

• When the kernel does not answer. The kernel says which requests it refused. If it had more to say than there was room for, or says nothing for ten seconds, the daemon asks it what it has of our routes and puts right what is not as it should be.

• Privileges. Changing routes needs the CAP_NET_ADMIN capability, and listening on port 179 needs CAP_NET_BIND_SERVICE; neither needs root. Without the first the daemon stops as soon as the kernel refuses it.

  setcap cap_net_admin,cap_net_bind_service+ep /usr/local/bin/bgp
  

ip route show proto bgp shows what is there. /api/kernel and the status page show how many routes are installed, how many are not and why, the last refusal by the kernel, and the interfaces; /metricz has bgp_kernel_routes, bgp_kernel_routes_unresolved, bgp_kernel_routes_refused, bgp_kernel_route_changes_total{change} and bgp_kernel_errors_total.

To forward packets for others the machine also needs net.ipv4.ip_forward=1 and net.ipv6.conf.all.forwarding=1; the daemon does not set them.

How long it takes

A million IPv4 routes, about the size of the Internet's table, go into the kernel in a little over a second and come out in half a second (release build, in a virtual machine on an Apple M-series laptop; the test many_routes_are_installed_and_removed with FERRIX_ROUTES=1000000). About three quarters of that is the kernel's own work on each route. What keeps the rest small: requests go to the kernel hundreds to a write, with one acknowledgement for each write and not for each route; a task of its own does the writing, so the next batch is worked out while the kernel is busy with the last; and what a next hop resolves to is worked out once for each next hop, not once for each route.

Testing against a real kernel

The tests that change a kernel's routes are not run by cargo test. scripts/kernel-tests.sh runs them on Linux, each in a network namespace made for it (scripts/in-namespace.sh runs any one command that way): the route installer against ip route, the daemon with a peer, and two daemons that route for each other with pings across. From macOS, scripts/linux-vm.sh runs it (or any command, such as cargo test) in a Lima virtual machine that it creates on first use.

Interoperability

scripts/interop/lab.sh runs the daemon against FRR, BIRD and Routinator, each in a network namespace of its own on one exchange LAN, and checks what every side ends up with. It needs Linux (scripts/linux-vm.sh scripts/interop/lab.sh from macOS; scripts/interop/install.sh installs the others on Ubuntu). Last run against FRR 10.5.1, BIRD 2.18 and Routinator 0.15.2. What it covers:

Phase	What is checked
basics	Sessions over IPv4 and IPv6; routes both ways with path, next hop, communities and large communities intact; four-octet AS numbers; ADD-PATH from BIRD; pings across by the installed routes
roles	RFC 9234 roles agree with FRR (peer) and BIRD (provider); Only to Customer is set where it should be, in both directions
rpki	RTR version 2 with Routinator; an invalid route from FRR is refused and held, and taken after Routinator restarts with data that allows it
refresh	Routes added and withdrawn by FRR; FRR's ROUTE-REFRESH answered; BIRD re-sending
flaps	Five session resets; FRR and BIRD killed and restarted; a cut wire noticed by the hold timer
restart	The daemon stopped in good order (kernel emptied, peers told) and killed (routes kept, stale one cleared by the next start)
scale	A table from BIRD (SCALE_ROUTES, a million for the Internet's size) into the table and the kernel; the daemon killed and restarted under it; the table withdrawn
limits	A peer whose role does not fit gets no session; a peer over its prefix limit gets Cease / Maximum Number of Prefixes and is held down
hijacks	FRR announces our own prefix, a route whose next hop is BIRD, and the exchange LAN's own prefix: none is taken
nexthops	A better route from FRR whose next hop cannot be reached is held and BIRD's is chosen and installed; they swap when the next hop becomes a neighbor, and swap back when it stops being one
transit	FRR as our customer and BIRD as our provider: FRR's routes reach BIRD with our AS on the path, BIRD's reach FRR marked only to customer, one of FRR's is kept from BIRD by a community, and packets go between their networks through us
v6only	The exchange LAN without IPv4 addresses: one IPv6 session with each router for both families, IPv4 routes with IPv6 next hops (RFC 8950) in BGP and in the kernel, IPv4 pings across
linklocal	The exchange LAN with link-local addresses only, and the same ones on a second interface: sessions with FRR (by interface, with radvd beside the daemon) and BIRD at fe80::…%ix0, both families in BGP and in the kernel, pings across, the interface taken down and up

scripts/interop/real-rpki.sh checks the RTR client against a Routinator that has the real RPKI, ASPAs included: the daemon has to end up with as many VRPs and ASPAs as the validator says it has.

Not covered: routers other than FRR and BIRD; IPv6 routes over an IPv4 session (the daemon does not send them); an interface that is removed and made anew under a session; long runs.

Finding a route

The status page has a look-up box over the received routes, and /api/routes/received takes the same questions:

Query	Answer
?prefix=193.0.0.0/21	Every route for that prefix, the best first
?address=193.0.0.1	Every prefix that covers the address, most specific first
?within=193.0.0.0/16	That prefix and everything inside it
&peer=NAME	Only routes from that peer (its name or its address)
&best=true	Only the route chosen for each prefix
&limit=N	At most N routes (1000 unless given, 100000 at most)

within, peer and best on their own look through the whole table, which on a full table takes a moment.

Seeing and alerting on refused routes

A refused route is not in the table. What is kept instead, for each peer, is how many were refused for each reason and the latest 200 of them with their AS paths. The status page lists them under "Refused routes"; /api/routes/rejected[?peer=NAME] has them as JSON. The record is kept when the session ends, so a peer that was cut off can still be looked into. The log has a line for a peer's first refused route and then at most one a minute.

For alerting, /metricz has:

Metric	Meaning
bgp_routes_rejected_total{peer,reason}	Routes refused, by peer and why
bgp_peer_routes{peer}	Routes an established peer has in the table
bgp_peer_max_prefixes{peer}	The peer's limit, where it has one
bgp_prefix_limit_exceeded_total{peer}	Sessions ended for going over the limit
bgp_connections_refused_total{reason}	Connections closed without a session
bgp_routes_exported_total{change}	Learned routes announced to, and withdrawn from, other peers
bgp_peer_up{peer}	1 while the session with a configured peer is Established, else 0
bgp_build_info{version}	Always 1; says which version is running

Rules to start from:

groups:
  - name: bgp
    rules:
      # A peer is announcing what its policy does not allow: a leak or a
      # hijack on its side, or a list on ours that is out of date.
      - alert: BgpRoutesRefused
        expr: sum by (peer, reason) (increase(bgp_routes_rejected_total[10m])) > 0
        labels: { severity: warning }
      # A route that has been through our own AS came back: a loop, or
      # someone announcing our AS number.
      - alert: BgpOwnAsInPath
        expr: increase(bgp_routes_rejected_total{reason="own AS in path"}[10m]) > 0
        labels: { severity: page }
      # A peer is close to its limit; the limit may need raising before the
      # session is cut.
      - alert: BgpPeerNearPrefixLimit
        expr: bgp_peer_routes / on (peer) bgp_peer_max_prefixes > 0.8
        for: 10m
        labels: { severity: warning }
      - alert: BgpPrefixLimitExceeded
        expr: increase(bgp_prefix_limit_exceeded_total[1h]) > 0
        labels: { severity: page }
      # A configured peer has no session.
      - alert: BgpPeerDown
        expr: bgp_peer_up == 0
        for: 5m
        labels: { severity: page }
      # Someone is trying to have our own prefixes, or our traffic, go
      # where they should not.
      - alert: BgpHijackAttempt
        expr: increase(bgp_routes_rejected_total{reason=~"prefix is our own|next hop is not the peer"}[10m]) > 0
        labels: { severity: page }
      # The kernel is refusing routes, or its answers are being lost.
      - alert: BgpKernelErrors
        expr: increase(bgp_kernel_errors_total[10m]) > 0
        labels: { severity: warning }

Trying it without a router

fake_peer is a minimal BGP speaker that connects to the daemon, announces a prefix, prints every UPDATE it receives and, with --echo, re-announces each received prefix back with its own ASN prepended (dropping NO_EXPORT routes). The prefix it announces shows up in the received table. The ones it echoes have been through our AS already, so they are refused as loops and counted in the peer's refused routes:

cargo run -- testdata/server.textpb
cargo run -p bgp-integration-tests --example fake_peer -- \
    --server 127.0.0.1:179 --asn 65001 --announce 203.0.113.0/24 --echo