apache/modules/proxy

mod_proxy_beacon

Self-registering reverse-proxy balancer membership: backend servers announce
themselves to a front-end proxy over UDP datagrams, and the proxy adds, enables,
and evicts them as live balancer members. See
docs/manual/mod/mod_proxy_beacon.xml for the user-facing directive reference;
this file describes the architecture.

The transport is plain APR UDP sockets -- small, periodic, fire-and-forget
announcements with no third-party dependency.


Dependencies:
  Requires mod_watchdog and mod_proxy_balancer. No external library: the UDP
  transport uses APR sockets and authentication uses SipHash from APR-util (no
  OpenSSL dependency).


Transport (unicast UDP, not multicast):
  Each backend sends a point-to-point UDP datagram (apr_socket_sendto) to the
  proxy's configured address. This is deliberately UNICAST -- unlike
  mod_heartbeat, which multicasts to a group. Multicast is filtered on most
  networks and does not traverse the public Internet, so a cross-host control
  plane must be unicast. The socket-handling pattern is adapted from
  mod_heartmonitor's unicast receive path.

  Datagrams are fire-and-forget. The design tolerates UDP loss and reordering
  without a reconnect or framing layer: a lost announcement just delays an update
  by one interval (announcements are periodic and idempotent), and an out-of-order
  datagram is rejected by the per-url monotonic timestamp (see Authentication).


Roles:
  Backend  = sender   (ProxyBeaconAddress): sends to the proxy, advertising its
             own routable URL (ProxyBeaconAdvertise) periodically.
  Proxy    = receiver (ProxyBeaconListen): binds a stable address, receives
             announcements, and manages balancer://<name> (ProxyBeaconBalancer).

  The proxy is the fixed unicast rendezvous; each backend is configured with its
  address. One proxy receiver serves many backend senders. A leading scheme in
  an address (e.g. tcp://) is accepted and ignored.

  ProxyBeaconListen's host and port are both optional: an omitted host/port (or
  no argument at all) is inherited from the server's own address/port. Since UDP
  and TCP are independent port spaces, the beacon socket can share the server's
  service port without colliding with its TCP listener -- so a backend can beacon
  to the proxy at its real service endpoint. (The listener binds in an
  unprivileged watchdog child, so privileged ports like 80/443 cannot be shared
  this way.) The sender's ProxyBeaconAddress targets the proxy and must be given
  in full.


Where it runs:
  All background work runs in a single mod_watchdog SINGLETON callback -- exactly
  one child process owns the socket (so only one process binds the receive port
  and performs membership changes). The watchdog fires every ~100ms:

    STARTING  open the UDP socket; receiver binds, sender resolves its dest.
    RUNNING   sender:   send a throttled announcement (ProxyBeaconInterval).
              receiver: drain the socket, then run a throttled (~1s) eviction
                        sweep.
    STOPPING  close the socket.

  Because this relies on a singleton watchdog child, it is inactive under the
  prefork MPM.


Adding/removing members (reuses the balancer-manager path):
  Runtime worker *removal* does not exist in httpd, so membership is managed
  entirely as status-flag changes via mod_proxy_balancer's exported
  balancer_manage() optional function -- the same code the balancer-manager web
  UI calls. The receiver synthesizes a minimal request_rec (no real client
  request exists in the watchdog thread; see beacon_make_fake_request) and calls:

    add     b=<name> b_nwrkr=<url> b_wyes=1 b_nwrkr_flags=B
                                                (worker created, disabled, marked)
    enable  b=<name> w=<url> w_status_D=0       (cleared -> serves traffic)
    evict   b=<name> w=<url> w_status_D=1       (disabled -> out of rotation)

  Adds bump the balancer's shm "updated" timestamp, so every other child picks
  up the new worker through the normal ap_proxy_sync_balancer() path -- the
  change made in the singleton propagates fleet-wide. The target balancer must
  have spare slots (ProxySet growth / BalancerGrowth).

  Per-backend state (last-seen time, added/evicted flags) is kept in a hash in
  the singleton's own pool -- no shared memory is needed, since the singleton is
  the only process that adds or evicts. A re-announcement after eviction
  re-enables the worker; the flags make these transitions idempotent.


Message format:
  Plain ASCII, space-separated key=value tokens, one datagram per announcement:

    BEACON url=http://host:port host=<h> pid=<n> seq=<n> ts=<usec> mac=<hex>

  url= is the routable backend origin the proxy adds as a BalancerMember. ts=
  (microseconds since the epoch) and mac= authenticate the message and are always
  present, since a shared secret is required (see below). host=/pid=/seq= are
  informational.

Member identity:
  url= names an ORIGIN -- scheme://host[:port] -- and nothing else.  The receiver
  canonicalizes it (lower-case scheme and host, the scheme's default port dropped,
  one trailing "/" ignored) and uses that single string as the member's identity:
  the key of its lease/replay record in ctx->seen, the worker name handed to
  balancer_manage(), and the name used for every later enable/evict.  Keeping
  those equal matters because the balancer finds a worker by LONGEST-PREFIX match
  (ap_proxy_get_worker) while ap_proxy_define_worker() stores a normalized name:
  tracking the raw announced string made "http://h:80" never find its worker
  "http://h" (member added but left disabled, add still logged as a success), and
  let "http://h/one" and "http://h/one/two" hold two leases over one worker.

  Anything that is not an origin -- a path other than "/", query, fragment,
  userinfo, a bad port, or the "unix:/path|url" worker syntax (no host part) -- is
  dropped with a rate-limited "unacceptable url" log.  ProxyBeaconAdvertise applies
  the same rule at config time and announces the canonical form.  After an add,
  and before any enable/evict, the worker is looked up again and the operation is
  reported as failed if it is not there, since balancer_manage() returns success
  when it finds no such worker.  Path-based routing belongs in ProxyPass on the
  proxy, not in the advertised URL.

  A UDP datagram is delivered whole, so the receiver reads each message into a
  fixed stack buffer and NUL-terminates it (buf[len] = '\0') before parsing.
  The sender transmits strlen(msg) bytes (no trailing NUL on the wire).


Ownership and watchdog hand-off:
  Every worker beacon creates is marked PROXY_WORKER_BEACON (status 'B', "Bcn" in
  the balancer-manager) in the worker's SHARED state.  Beacon enables/evicts only
  marked workers; a same-named worker without the mark (configured, or added by
  an administrator) is left alone and the announcement is dropped ("member not
  managed by beacon").

  The leases (ctx->seen) are private to the process that owns the watchdog
  singleton, but the members they describe live in shared memory and outlive that
  process.  When ownership moves (owner crashed, was recycled, or the server was
  restarted), the new owner starts with an empty table and would never evict a
  member that has since gone silent.  So, once at the start of ownership
  (beacon_reconcile(), first RUNNING tick of a listener), it syncs the balancer
  (ap_proxy_sync_balancer()) and adopts every marked member: last_seen = now, so
  a silent member is evicted one ProxyBeaconTimeout later; a member that is
  already DISABLED is recorded as evicted, so its next announcement re-enables it.

  Not recovered: the per-member replay high-water mark (last_ts restarts at 0).
  The global freshness window (ProxyBeaconMaxSkew) still applies, so a captured
  announcement can be replayed once per member after a hand-off, refreshing that
  one lease for one timeout.  Keeping that state across hand-off would need a
  shared lease table; the marker + reconcile design was chosen as the smaller,
  ABI-safe change (the status field is existing shared state, bit 0x1000 is new).

  A new owner can only start once the watchdog-callback mutex is released by the
  dead one.  If the default mechanism is not released when its holder is killed on
  a platform, use e.g. "Mutex sysvsem watchdog-callback".

Operator control:
  Beacon owns the DISABLED flag of its members (lease expiry sets it, a new
  announcement clears it).  It never touches STOPPED, DRAIN or the hcheck
  failure flag, so an administrator who needs a member held out of rotation uses
  Stopped/Drain; a manual "Disable" is overridden by the next announcement after
  an eviction.  (After a hand-off a DISABLED beacon member is assumed to have been
  evicted, for the same reason.)

Receiver validation:
  Datagrams > BEACON_MAX_MSG_LEN (2048) are dropped, not truncated (recvfrom is
  asked for one byte more to tell); a datagram containing a NUL is dropped.  The
  MAC is verified on every path, including log-only (no ProxyBeaconBalancer).
  After the MAC: strict grammar (beacon_frame_check(): "BEACON " marker,
  single-space key=value tokens, known fields not repeated), ts= strictly 1-18
  decimal digits (beacon_parse_ts()), freshness compared in microseconds so a
  sub-second ProxyBeaconMaxSkew is honoured.  The url must be printable ASCII
  and an origin (see Member identity).

Capacity:
  At most BEACON_MAX_MEMBERS (256) urls are tracked.  beacon_sweep() also frees
  records that are evicted or whose add failed once they have been silent for
  2 * max_skew + 1s: the record doubles as the replay high-water mark, and a
  message accepted at time A carries ts within max_skew of A, so it is stale after
  A + 2 * max_skew; nothing is lost by forgetting it then.  A reclaimed member that
  announces again is re-adopted from the balancer (try_add finds the marked worker).
  Records come from a free list: memory is bounded by the cap however often they
  are recycled.  A flood of distinct urls from a secret holder can still fill the
  table until it falls silent.

Socket open:
  A failed open/bind/resolve is retried with back-off (30s doubling to 10min),
  not on every tick and not never.  The address lookup is synchronous on the one
  watchdog thread shared by all beacon contexts; prefer IP literals.

Authentication (ProxyBeaconSecret, REQUIRED):
  ProxyBeaconSecret must be set on every participating server (the receiver and
  every sender); startup fails otherwise. There is no unauthenticated mode:
  without a MAC, anyone who can reach the receiver port could announce a URL and
  hijack client traffic, and a UDP source address is trivially spoofable, so
  authentication matters at least as much here as it would over a connection.

  With ProxyBeaconSecret set identically on the proxy and all backends, each
  announcement is signed with a SipHash-2-4 MAC over the message prefix (the key
  is derived from the passphrase via apr_md5, as in mod_session_crypto). The
  receiver recomputes the MAC (constant-time compare) and rejects anything that
  does not match -- before the URL is parsed or acted on. Replay protection has
  two parts, both keyed on the signed ts= (microseconds): (1) a freshness window
  (ProxyBeaconMaxSkew) rejects messages whose timestamp is too far from now,
  which assumes proxy and backend clocks are roughly synchronized (NTP); (2) a
  per-url high-water mark rejects any ts= that does not strictly exceed the last
  one accepted for that url, so a byte-identical replay (same ts, same MAC) --
  e.g. replaying a dead backend's last announcement to keep it from being evicted
  -- is dropped even within the freshness window. Microsecond granularity ensures
  genuine sub-second announcements always advance. seq= is NOT used for replay
  (it resets when a backend restarts); the wall-clock ts= does not.

  "hex:<32 hex digits>" is the 128-bit SipHash key itself (openssl rand -hex 16):
  no passphrase, nothing to guess.  Any other value is a passphrase of at least 16
  characters; the key is then MD5(passphrase) and the tag is 64 bits, so a
  captured announcement permits offline guessing at full speed.  A valid MAC shows possession of the
  secret, nothing more: every holder can register any reachable origin, speak for
  other members and consume slots/records.  The secret is a membership-
  administrator credential, not a boundary between partially trusted workloads.

  Announcements are authenticated, not encrypted -- the payload is operational
  metadata, not secret. For transport confidentiality, DTLS would be a separate,
  orthogonal future layer.