For reading and verification only. You may not copy, use, modify or redistribute this code. Full notice on the main page.
tunnel/server.go
| 1 | package main |
| 2 | |
| 3 | import ( |
| 4 | "bufio" |
| 5 | "context" |
| 6 | "crypto/rand" |
| 7 | "crypto/subtle" |
| 8 | "crypto/tls" |
| 9 | "encoding/hex" |
| 10 | "errors" |
| 11 | "fmt" |
| 12 | "hash/maphash" |
| 13 | "io" |
| 14 | "log/slog" |
| 15 | "net" |
| 16 | "net/netip" |
| 17 | "os" |
| 18 | "strings" |
| 19 | "sync" |
| 20 | "sync/atomic" |
| 21 | "time" |
| 22 | ) |
| 23 | |
| 24 | const ( |
| 25 | // TLS alert descriptions, sent (in the clear, as a fatal alert record) to a client we decide not to serve. |
| 26 | alertInternalError = 80 |
| 27 | alertUnrecognizedName = 112 |
| 28 | ) |
| 29 | |
| 30 | var ( |
| 31 | errOffline = errors.New("device_offline") |
| 32 | errAttachTimeout = errors.New("attach_timeout") |
| 33 | errControlWrite = errors.New("control_write_failed") |
| 34 | errShuttingDown = errors.New("shutting_down") |
| 35 | errByteCap = errors.New("stream byte cap reached") |
| 36 | errHostTaken = errors.New("host_taken") |
| 37 | errGatedBytes = errors.New("gated byte budget spent") |
| 38 | ) |
| 39 | |
| 40 | // pendingStream is an OPEN waiting for its ATTACH: the secret is one use, and only the device that was sent the OPEN knows it. |
| 41 | type pendingStream struct { |
| 42 | cc *ctrl |
| 43 | secret string |
| 44 | ch chan net.Conn // buffered(1): the ATTACHed data connection |
| 45 | } |
| 46 | |
| 47 | // Server is the tunnel node. It never terminates a client's TLS: the only TLS it speaks is the one it holds for its own name, with devices. |
| 48 | type Server struct { |
| 49 | cfg Config |
| 50 | log *slog.Logger |
| 51 | tlsCfg *tls.Config |
| 52 | routes *routeTable |
| 53 | replay *replayCache |
| 54 | |
| 55 | nodeURL string // the htu of a proof: this node's URL |
| 56 | |
| 57 | clk clock |
| 58 | health *cpHealth |
| 59 | |
| 60 | ctx context.Context |
| 61 | cancel context.CancelFunc |
| 62 | bg sync.WaitGroup // the feed loop, the background tasks and the refreshes of connections: ended by cancel |
| 63 | |
| 64 | mu sync.Mutex |
| 65 | pending map[string]*pendingStream |
| 66 | conns map[*ctrl]struct{} // every live control connection, for shutdown |
| 67 | |
| 68 | streams atomic.Int64 // client connections being served, global |
| 69 | anonStreams atomic.Int64 // of them, the ones whose source is neither openai nor marked good: they may not use the share reserved for the others |
| 70 | unauth atomic.Int64 // connections that have not finished their first message |
| 71 | anonUnauth atomic.Int64 // of them, the ones whose source is neither ChatGPT's, nor a device's, nor marked good: they may not use the share that is kept for the others |
| 72 | authQ atomic.Int64 // exchanges that wait for, or are in, a call to authorize |
| 73 | connSeq atomic.Int64 |
| 74 | |
| 75 | authSem chan struct{} // the calls to authorize in flight: at most cfg.AuthorizeMax |
| 76 | |
| 77 | watchMu sync.Mutex |
| 78 | watches map[*feedWatch]struct{} // the exchanges waiting for authorize, each told of the feed events applied meanwhile |
| 79 | |
| 80 | killed atomic.Bool // the global kill switch is on: no device is served until it is cleared |
| 81 | draining atomic.Bool // SIGTERM: no new client and no new device connection, the open streams finish |
| 82 | drainFor atomic.Int64 // the quiet period this drain was started with, in nanoseconds (what a late device is told) |
| 83 | |
| 84 | // hardening (limits.go, hints.go, gated.go, openai_prefixes.go, metrics.go, observer.go) |
| 85 | ips *ipLimits |
| 86 | hints *hintTable |
| 87 | openai *openaiPrefixes |
| 88 | gates *gateTable |
| 89 | met *metrics |
| 90 | sampler *logSampler |
| 91 | seed maphash.Seed // per boot: what Server.short hashes with, so two log lines of one source can be told apart and nothing can be read back |
| 92 | startedAt time.Time |
| 93 | |
| 94 | httpsLn, httpLn net.Listener |
| 95 | metricsSrv *metricsServer |
| 96 | wg sync.WaitGroup |
| 97 | bufPool sync.Pool // the 32 KiB buffers a pump takes for bulk reads |
| 98 | smallBufPool sync.Pool // the 4 KiB buffers a pump waits in a read with |
| 99 | bigBuffersInUse atomic.Int64 // 32 KiB buffers held by a pump now (an idle stream holds none: tests and the memory check read it) |
| 100 | } |
| 101 | |
| 102 | const ( |
| 103 | pumpSmallBuf = 4 << 10 |
| 104 | pumpBigBuf = 32 << 10 |
| 105 | ) |
| 106 | |
| 107 | // NewServer validates cfg and builds the node. It does not listen yet. |
| 108 | func NewServer(cfg Config) (*Server, error) { |
| 109 | if cfg.Logger == nil { |
| 110 | cfg.Logger = slog.New(slog.NewJSONHandler(io.Discard, nil)) |
| 111 | } |
| 112 | if cfg.Now == nil { |
| 113 | cfg.Now = time.Now |
| 114 | } |
| 115 | if cfg.Health == nil { |
| 116 | cfg.Health = &cpHealth{} |
| 117 | } |
| 118 | if err := cfg.validate(); err != nil { |
| 119 | return nil, err |
| 120 | } |
| 121 | ctx, cancel := context.WithCancel(context.Background()) |
| 122 | clk := newClock(cfg.Now, cfg.AfterFunc) |
| 123 | s := &Server{ |
| 124 | cfg: cfg, |
| 125 | log: cfg.Logger, |
| 126 | routes: newRouteTable(), |
| 127 | replay: newReplayCache(replayLimit), |
| 128 | nodeURL: nodeProofURL(cfg.NodeHost, cfg.PublicPort), |
| 129 | clk: clk, |
| 130 | health: cfg.Health, |
| 131 | ctx: ctx, |
| 132 | cancel: cancel, |
| 133 | pending: map[string]*pendingStream{}, |
| 134 | conns: map[*ctrl]struct{}{}, |
| 135 | watches: map[*feedWatch]struct{}{}, |
| 136 | authSem: make(chan struct{}, cfg.AuthorizeMax), |
| 137 | seed: maphash.MakeSeed(), |
| 138 | met: newMetrics(), |
| 139 | sampler: newLogSampler(clk), |
| 140 | hints: newHintTable(clk, cfg.HintTTL, cfg.HintsPerLabel), |
| 141 | gates: newGateTable(clk), |
| 142 | startedAt: clk.Now(), |
| 143 | } |
| 144 | s.ips = newIPLimits(&cfg, clk, s.seed) |
| 145 | s.openai = newOpenAIPrefixes(&cfg, clk, s.met.prefixRefresh.inc) |
| 146 | s.routes.lim = labelLimitsOf(&cfg) |
| 147 | if cfg.HTTPSAddr != "" { |
| 148 | s.tlsCfg = &tls.Config{ |
| 149 | GetCertificate: cfg.CertSource, |
| 150 | MinVersion: tls.VersionTLS13, |
| 151 | NextProtos: []string{policyData.Protocol.ALPN}, |
| 152 | } |
| 153 | } |
| 154 | s.bufPool.New = func() any { b := make([]byte, pumpBigBuf); return &b } |
| 155 | s.smallBufPool.New = func() any { b := make([]byte, pumpSmallBuf); return &b } |
| 156 | return s, nil |
| 157 | } |
| 158 | |
| 159 | // replayLimit bounds the proofs remembered at once (a proof lives for the policy's proof TTL, so this is the number of proofs per minute the node |
| 160 | // admits before it refuses more). |
| 161 | const replayLimit = 1 << 18 |
| 162 | |
| 163 | func (s *Server) now() time.Time { return s.clk.Now() } |
| 164 | |
| 165 | func addrOf(ln net.Listener) string { |
| 166 | if ln == nil { |
| 167 | return "" |
| 168 | } |
| 169 | return ln.Addr().String() |
| 170 | } |
| 171 | |
| 172 | func (s *Server) HTTPSAddr() string { return addrOf(s.httpsLn) } |
| 173 | func (s *Server) HTTPAddr() string { return addrOf(s.httpLn) } |
| 174 | |
| 175 | // Online reports whether the host has a live control connection that opens streams. |
| 176 | func (s *Server) Online(host string) bool { |
| 177 | cc, _ := s.routes.lookup(host) |
| 178 | return cc != nil && cc.canOpen(s.now()) |
| 179 | } |
| 180 | |
| 181 | // Start opens the two listeners (443 and 80: there is no other) and serves until Shutdown. |
| 182 | func (s *Server) Start() error { |
| 183 | // Every socket the node accepts gets TCP keepalive from the policy: a stream whose peer vanished (a computer that slept, a network that changed) ends |
| 184 | // by itself instead of holding its slot until the idle timeout. |
| 185 | lc := s.listenConfig() |
| 186 | listen := func(addr string) (net.Listener, error) { |
| 187 | if addr == "" { |
| 188 | return nil, nil |
| 189 | } |
| 190 | return lc.Listen(context.Background(), "tcp", addr) |
| 191 | } |
| 192 | var err error |
| 193 | if s.httpsLn, err = listen(s.cfg.HTTPSAddr); err != nil { |
| 194 | return fmt.Errorf("https listener: %w", err) |
| 195 | } |
| 196 | if s.httpLn, err = listen(s.cfg.HTTPAddr); err != nil { |
| 197 | _ = s.closeListeners() |
| 198 | return fmt.Errorf("http listener: %w", err) |
| 199 | } |
| 200 | if merr := s.startMetrics(); merr != nil { |
| 201 | // The node serves without its metrics rather than not at all: a port that is taken is not a reason for a fleet's devices to lose their node. |
| 202 | // The operator sees the line and the missing /healthz. |
| 203 | s.log.Warn("metrics_unavailable") |
| 204 | } |
| 205 | // What the node must already have stopped serving is known before the first device connects: the feed's first answer (a revoke that happened while |
| 206 | // the node was down, the kill switch) comes first, bounded so that a late control plane does not keep the node from starting. |
| 207 | s.startBackground() |
| 208 | for _, l := range []struct { |
| 209 | ln net.Listener |
| 210 | fn func(net.Conn) |
| 211 | }{{s.httpsLn, s.handleHTTPS}, {s.httpLn, s.handleHTTP}} { |
| 212 | if l.ln != nil { |
| 213 | s.wg.Add(1) |
| 214 | go s.acceptLoop(l.ln, l.fn) |
| 215 | } |
| 216 | } |
| 217 | s.log.Info("listening", "https", s.HTTPSAddr(), "http", s.HTTPAddr(), "node", s.cfg.NodeID) |
| 218 | return nil |
| 219 | } |
| 220 | |
| 221 | // keepaliveProbes is how many unanswered keepalive probes end a connection. |
| 222 | const keepaliveProbes = 4 |
| 223 | |
| 224 | // listenConfig is how the node's listeners are made: the connections they accept have TCP keepalive on, with the policy's period. |
| 225 | func (s *Server) listenConfig() net.ListenConfig { |
| 226 | return net.ListenConfig{KeepAliveConfig: net.KeepAliveConfig{Enable: true, Idle: s.cfg.TCPKeepalive, Interval: s.cfg.TCPKeepalive, Count: keepaliveProbes}} |
| 227 | } |
| 228 | |
| 229 | func (s *Server) closeListeners() error { |
| 230 | var first error |
| 231 | for _, ln := range []net.Listener{s.httpsLn, s.httpLn} { |
| 232 | if ln != nil { |
| 233 | if err := ln.Close(); err != nil && first == nil { |
| 234 | first = err |
| 235 | } |
| 236 | } |
| 237 | } |
| 238 | return first |
| 239 | } |
| 240 | |
| 241 | // Shutdown stops accepting, tells every device the node is going (CLOSE shutdown) and waits (up to grace) for streams to finish; the streams are |
| 242 | // force-closed when it ends. |
| 243 | func (s *Server) Shutdown(grace time.Duration) { s.shutdown(grace, closeShutdown) } |
| 244 | |
| 245 | func (s *Server) shutdown(grace time.Duration, reason string) { |
| 246 | _ = s.closeListeners() |
| 247 | s.stopMetrics() |
| 248 | s.mu.Lock() |
| 249 | live := make([]*ctrl, 0, len(s.conns)) |
| 250 | for cc := range s.conns { |
| 251 | live = append(live, cc) |
| 252 | } |
| 253 | s.mu.Unlock() |
| 254 | for _, cc := range live { |
| 255 | cc.close(reason) |
| 256 | } |
| 257 | done := make(chan struct{}) |
| 258 | go func() { s.wg.Wait(); close(done) }() |
| 259 | select { |
| 260 | case <-done: |
| 261 | case <-time.After(grace): |
| 262 | } |
| 263 | s.mu.Lock() // with goBackground: a task is either counted before the wait below starts, or sees the node ended and does not start |
| 264 | s.cancel() |
| 265 | s.mu.Unlock() |
| 266 | select { |
| 267 | case <-done: |
| 268 | case <-time.After(2 * time.Second): |
| 269 | } |
| 270 | bgDone := make(chan struct{}) |
| 271 | go func() { s.bg.Wait(); close(bgDone) }() |
| 272 | select { |
| 273 | case <-bgDone: |
| 274 | case <-time.After(2 * time.Second): |
| 275 | } |
| 276 | } |
| 277 | |
| 278 | func (s *Server) acceptLoop(ln net.Listener, handle func(net.Conn)) { |
| 279 | defer s.wg.Done() |
| 280 | var delay time.Duration |
| 281 | for { |
| 282 | c, err := ln.Accept() |
| 283 | if err != nil { |
| 284 | if s.ctx.Err() != nil || errors.Is(err, net.ErrClosed) { |
| 285 | return |
| 286 | } |
| 287 | delay = min(max(2*delay, 5*time.Millisecond), time.Second) |
| 288 | s.log.Warn("accept_error", "retry_in_ms", delay.Milliseconds()) |
| 289 | select { |
| 290 | case <-time.After(delay): |
| 291 | case <-s.ctx.Done(): |
| 292 | return |
| 293 | } |
| 294 | continue |
| 295 | } |
| 296 | delay = 0 |
| 297 | s.wg.Add(1) |
| 298 | go handle(c) |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | func tryAcquire(counter *atomic.Int64, limit int64) bool { |
| 303 | for { |
| 304 | cur := counter.Load() |
| 305 | if cur >= limit { |
| 306 | return false |
| 307 | } |
| 308 | if counter.CompareAndSwap(cur, cur+1) { |
| 309 | return true |
| 310 | } |
| 311 | } |
| 312 | } |
| 313 | |
| 314 | // anonUnauthMax is the most unauthenticated connections the sources nobody vouches for may hold together. |
| 315 | func (s *Server) anonUnauthMax() int64 { |
| 316 | return int64(s.cfg.MaxUnauth - s.cfg.MaxUnauth*s.cfg.OpenAIReservedPct/100) |
| 317 | } |
| 318 | |
| 319 | func newStreamID() string { |
| 320 | var b [16]byte |
| 321 | if _, err := rand.Read(b[:]); err != nil { |
| 322 | panic("crypto/rand failed: " + err.Error()) |
| 323 | } |
| 324 | return hex.EncodeToString(b[:]) |
| 325 | } |
| 326 | |
| 327 | // newSecret is a 256-bit one-use attach secret (43 base64url characters). |
| 328 | func newSecret() string { |
| 329 | var b [32]byte |
| 330 | if _, err := rand.Read(b[:]); err != nil { |
| 331 | panic("crypto/rand failed: " + err.Error()) |
| 332 | } |
| 333 | return b64Encode(b[:]) |
| 334 | } |
| 335 | |
| 336 | // newNonce is the 128-bit CHALLENGE nonce (22 base64url characters). |
| 337 | func newNonce() string { |
| 338 | var b [16]byte |
| 339 | if _, err := rand.Read(b[:]); err != nil { |
| 340 | panic("crypto/rand failed: " + err.Error()) |
| 341 | } |
| 342 | return b64Encode(b[:]) |
| 343 | } |
| 344 | |
| 345 | func (s *Server) tapBytes(id, dir string, p []byte) { |
| 346 | if s.cfg.tap != nil { |
| 347 | s.cfg.tap(id, dir, p) |
| 348 | } |
| 349 | } |
| 350 | |
| 351 | // emit records one protocol step for the contracts' transcript checker; it costs nothing when no trace is set. |
| 352 | func (s *Server) emit(at time.Time, ev map[string]any) { |
| 353 | if s.cfg.trace != nil { |
| 354 | ev["t"] = epochSeconds(at) |
| 355 | s.cfg.trace(ev) |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | // sendAlert writes one fatal TLS alert record, then lets the peer drain for a moment so it is not reset before it can read the alert. |
| 360 | func sendAlert(c net.Conn, desc byte) { |
| 361 | _ = c.SetWriteDeadline(time.Now().Add(time.Second)) |
| 362 | _, _ = c.Write([]byte{21, 3, 3, 0, 2, 2, desc}) |
| 363 | if cw, ok := c.(interface{ CloseWrite() error }); ok { |
| 364 | _ = cw.CloseWrite() |
| 365 | } |
| 366 | _ = c.SetReadDeadline(time.Now().Add(time.Second)) |
| 367 | _, _ = io.Copy(io.Discard, io.LimitReader(c, 16<<10)) |
| 368 | } |
| 369 | |
| 370 | func helloReason(err error) string { |
| 371 | switch { |
| 372 | case errors.Is(err, errNotTLS): |
| 373 | return "not_tls" |
| 374 | case errors.Is(err, errNotClientHello), errors.Is(err, errHelloMalformed): |
| 375 | return "bad_hello" |
| 376 | case errors.Is(err, errHelloTooLarge): |
| 377 | return "hello_too_large" |
| 378 | case errors.Is(err, os.ErrDeadlineExceeded): |
| 379 | return "hello_timeout" |
| 380 | case errors.Is(err, io.EOF), errors.Is(err, io.ErrUnexpectedEOF): |
| 381 | return "closed_before_hello" |
| 382 | default: |
| 383 | return "read_error" |
| 384 | } |
| 385 | } |
| 386 | |
| 387 | // peerAddr is the client's address as the contracts write it in an OPEN: no zone, an IPv4-mapped address as the IPv4 address it carries. |
| 388 | func peerAddr(c net.Conn) (netip.Addr, bool) { |
| 389 | ap, err := netip.ParseAddrPort(c.RemoteAddr().String()) |
| 390 | if err != nil { |
| 391 | return netip.Addr{}, false |
| 392 | } |
| 393 | return ap.Addr().WithZone("").Unmap(), true |
| 394 | } |
| 395 | |
| 396 | // ---------------------------------------------------------------- routing table |
| 397 | |
| 398 | // hostEntry is what the node knows about one device host: the control connection that serves it (nil while the device is away) and how many client |
| 399 | // streams are open on it. |
| 400 | type hostEntry struct { |
| 401 | cc *ctrl |
| 402 | streams int64 // client streams open on the host, all slices together |
| 403 | anon int64 // of them, the ones in the anonymous slice (the rest are in the slice of sources the device marked good) |
| 404 | live map[*liveStream]struct{} // the client streams being served on this host: what a revoke, a pause or an expiry must be able to cut |
| 405 | gated atomic.Bool // the control plane's last status for the host's connection is gated: its streams are held to the day's budget (gated.go) |
| 406 | newOnes bucket // the host's budget of new streams per second, for sources that are not marked good |
| 407 | } |
| 408 | |
| 409 | // liveStream is one client stream the node serves for a host, from the moment it is admitted (before its OPEN) to its end. A revoke, a pause, an |
| 410 | // expired connection or the kill switch kills it: whichever connections it holds are closed, and one that is not attached yet is closed the moment it is. |
| 411 | type liveStream struct { |
| 412 | mu sync.Mutex |
| 413 | client net.Conn |
| 414 | dev net.Conn |
| 415 | killed string // why it was killed, "" while it lives |
| 416 | |
| 417 | he *hostEntry // the host's entry: what says, at every chunk, whether the host is gated now (a RENEW may have made it so since the stream began) |
| 418 | usage *dayUsage // the host's gated-day record |
| 419 | anon bool // the stream counts in the host's anonymous slice |
| 420 | } |
| 421 | |
| 422 | // bind gives the stream its two connections. It reports false when the stream was killed before they were ready (the caller closes them). |
| 423 | func (l *liveStream) bind(client, dev net.Conn) bool { |
| 424 | l.mu.Lock() |
| 425 | defer l.mu.Unlock() |
| 426 | if l.killed != "" { |
| 427 | return false |
| 428 | } |
| 429 | l.client, l.dev = client, dev |
| 430 | return true |
| 431 | } |
| 432 | |
| 433 | func (l *liveStream) kill(reason string) { |
| 434 | l.mu.Lock() |
| 435 | if l.killed == "" { |
| 436 | l.killed = reason |
| 437 | } |
| 438 | client, dev := l.client, l.dev |
| 439 | l.mu.Unlock() |
| 440 | if client != nil { |
| 441 | _ = client.Close() |
| 442 | } |
| 443 | if dev != nil { |
| 444 | _ = dev.Close() |
| 445 | } |
| 446 | } |
| 447 | |
| 448 | func (l *liveStream) killedBy() string { |
| 449 | l.mu.Lock() |
| 450 | defer l.mu.Unlock() |
| 451 | return l.killed |
| 452 | } |
| 453 | |
| 454 | // routeTable maps a host to the control connection of the device that owns it. The host comes ONLY from the control plane's answer; the table |
| 455 | // refuses to give one host to two devices and replaces a connection only by a newer one of the SAME subject. It is safe for concurrent use. |
| 456 | type routeTable struct { |
| 457 | mu sync.Mutex |
| 458 | byHost map[string]*hostEntry |
| 459 | bySub map[string]*ctrl |
| 460 | lim labelLimits |
| 461 | } |
| 462 | |
| 463 | // labelLimits are what one device host accepts: how many streams (a cap on the whole, and on what sources that are not marked good may hold of it), |
| 464 | // and how fast new ones may begin when they come from such sources. |
| 465 | type labelLimits struct { |
| 466 | total, anon int |
| 467 | rate, burst float64 |
| 468 | } |
| 469 | |
| 470 | func labelLimitsOf(cfg *Config) labelLimits { |
| 471 | return labelLimits{total: cfg.MaxStreamsPerHost, anon: min(cfg.LabelAnonSlice, cfg.MaxStreamsPerHost), rate: float64(cfg.LabelNewPerS), burst: float64(cfg.LabelNewBurst)} |
| 472 | } |
| 473 | |
| 474 | func newRouteTable() *routeTable { |
| 475 | return &routeTable{byHost: map[string]*hostEntry{}, bySub: map[string]*ctrl{}} |
| 476 | } |
| 477 | |
| 478 | // activate binds cc to host as subject sub. It returns the connection it replaced (the same subject's older one: the caller closes it) or |
| 479 | // errHostTaken when another subject holds the host. bound, when set, runs under the table's lock once cc is bound, so whatever it records is |
| 480 | // ordered like the routing decisions are. |
| 481 | func (rt *routeTable) activate(cc *ctrl, sub, host string, bound func()) (*ctrl, error) { |
| 482 | rt.mu.Lock() |
| 483 | defer rt.mu.Unlock() |
| 484 | he := rt.byHost[host] |
| 485 | if he != nil && he.cc != nil && he.cc != cc && he.cc.subject() != sub { |
| 486 | return nil, errHostTaken |
| 487 | } |
| 488 | var old *ctrl |
| 489 | prevHost := "" |
| 490 | if prev := rt.bySub[sub]; prev != nil && prev != cc { |
| 491 | old, prevHost = prev, prev.hostName() |
| 492 | if pe := rt.byHost[prevHost]; pe != nil && pe.cc == prev { |
| 493 | pe.cc = nil |
| 494 | } |
| 495 | } |
| 496 | he = rt.byHost[host] // looked up again: unbinding the older connection must not have dropped it from under us |
| 497 | if he == nil { |
| 498 | he = &hostEntry{} |
| 499 | rt.byHost[host] = he |
| 500 | } |
| 501 | he.cc = cc |
| 502 | rt.bySub[sub] = cc |
| 503 | if prevHost != "" && prevHost != host { |
| 504 | if pe := rt.byHost[prevHost]; pe != nil { |
| 505 | rt.dropIfIdle(prevHost, pe) |
| 506 | } |
| 507 | } |
| 508 | cc.bind(sub, host) |
| 509 | if bound != nil { |
| 510 | bound() |
| 511 | } |
| 512 | return old, nil |
| 513 | } |
| 514 | |
| 515 | // remove unbinds cc (it never touches a newer connection that replaced it). |
| 516 | func (rt *routeTable) remove(cc *ctrl) { |
| 517 | rt.mu.Lock() |
| 518 | defer rt.mu.Unlock() |
| 519 | sub, host := cc.subject(), cc.hostName() |
| 520 | if rt.bySub[sub] == cc { |
| 521 | delete(rt.bySub, sub) |
| 522 | } |
| 523 | if he := rt.byHost[host]; he != nil && he.cc == cc { |
| 524 | he.cc = nil |
| 525 | rt.dropIfIdle(host, he) |
| 526 | } |
| 527 | } |
| 528 | |
| 529 | func (rt *routeTable) dropIfIdle(host string, he *hostEntry) { |
| 530 | if he.cc == nil && he.streams == 0 { |
| 531 | delete(rt.byHost, host) |
| 532 | } |
| 533 | } |
| 534 | |
| 535 | func (rt *routeTable) lookup(host string) (*ctrl, *hostEntry) { |
| 536 | rt.mu.Lock() |
| 537 | defer rt.mu.Unlock() |
| 538 | he := rt.byHost[host] |
| 539 | if he == nil { |
| 540 | return nil, nil |
| 541 | } |
| 542 | return he.cc, he |
| 543 | } |
| 544 | |
| 545 | // The reasons routes.acquire gives for a stream it does not admit. |
| 546 | const ( |
| 547 | admitUnknown = "unknown_host" // no device is online under the name |
| 548 | admitHostLimit = "host_limit" // the host holds all the streams its slice allows |
| 549 | admitLabelRate = "label_rate" // the host is being asked for new streams faster than it accepts from sources that are not marked good |
| 550 | ) |
| 551 | |
| 552 | // acquire counts one more client stream on the host and registers it so that it can be killed. A source the device marked good (hint) may use any of |
| 553 | // the host's cap; any other source only the anonymous slice, which is what leaves room for the good ones when an attacker holds streams, and its new |
| 554 | // streams are also paced per host. It returns the connection to OPEN on, or why the stream is refused. |
| 555 | func (rt *routeTable) acquire(host string, hint bool, now time.Time) (*ctrl, *liveStream, string) { |
| 556 | rt.mu.Lock() |
| 557 | defer rt.mu.Unlock() |
| 558 | he := rt.byHost[host] |
| 559 | if he == nil || he.cc == nil { |
| 560 | return nil, nil, admitUnknown |
| 561 | } |
| 562 | if he.streams >= int64(rt.lim.total) || (!hint && he.anon >= int64(rt.lim.anon)) { |
| 563 | return nil, nil, admitHostLimit |
| 564 | } |
| 565 | if !hint && !he.newOnes.take(now, rt.lim.rate, rt.lim.burst) { |
| 566 | return nil, nil, admitLabelRate |
| 567 | } |
| 568 | he.streams++ |
| 569 | if !hint { |
| 570 | he.anon++ |
| 571 | } |
| 572 | ls := &liveStream{he: he, anon: !hint} |
| 573 | if he.live == nil { |
| 574 | he.live = map[*liveStream]struct{}{} |
| 575 | } |
| 576 | he.live[ls] = struct{}{} |
| 577 | return he.cc, ls, "" |
| 578 | } |
| 579 | |
| 580 | func (rt *routeTable) release(host string, ls *liveStream) { |
| 581 | rt.mu.Lock() |
| 582 | defer rt.mu.Unlock() |
| 583 | he := ls.he |
| 584 | he.streams-- |
| 585 | if ls.anon { |
| 586 | he.anon-- |
| 587 | } |
| 588 | delete(he.live, ls) |
| 589 | if rt.byHost[host] == he { |
| 590 | rt.dropIfIdle(host, he) |
| 591 | } |
| 592 | } |
| 593 | |
| 594 | // setGated records, for the host cc serves, whether the control plane last called it gated. A connection that has been replaced says nothing. |
| 595 | func (rt *routeTable) setGated(cc *ctrl, gated bool) { |
| 596 | rt.mu.Lock() |
| 597 | defer rt.mu.Unlock() |
| 598 | if he := rt.byHost[cc.hostName()]; he != nil && he.cc == cc { |
| 599 | he.gated.Store(gated) |
| 600 | } |
| 601 | } |
| 602 | |
| 603 | // onlineCount is the number of device hosts with a control connection bound to them. |
| 604 | func (rt *routeTable) onlineCount() int { |
| 605 | rt.mu.Lock() |
| 606 | defer rt.mu.Unlock() |
| 607 | n := 0 |
| 608 | for _, he := range rt.byHost { |
| 609 | if he.cc != nil { |
| 610 | n++ |
| 611 | } |
| 612 | } |
| 613 | return n |
| 614 | } |
| 615 | |
| 616 | // killStreams kills the client streams being served on host. Unless force, it does nothing while a connection of the device is bound to the host (an |
| 617 | // expiry that raced a newer connection must not cut what that one serves). |
| 618 | func (rt *routeTable) killStreams(host, reason string, force bool) int { |
| 619 | rt.mu.Lock() |
| 620 | var doomed []*liveStream |
| 621 | if he := rt.byHost[host]; he != nil && (force || he.cc == nil) { |
| 622 | for ls := range he.live { |
| 623 | doomed = append(doomed, ls) |
| 624 | } |
| 625 | } |
| 626 | rt.mu.Unlock() |
| 627 | for _, ls := range doomed { |
| 628 | ls.kill(reason) |
| 629 | } |
| 630 | return len(doomed) |
| 631 | } |
| 632 | |
| 633 | // killAllStreams kills every client stream on every host (the kill switch, a feed reset). |
| 634 | func (rt *routeTable) killAllStreams(reason string) { |
| 635 | rt.mu.Lock() |
| 636 | var doomed []*liveStream |
| 637 | for _, he := range rt.byHost { |
| 638 | for ls := range he.live { |
| 639 | doomed = append(doomed, ls) |
| 640 | } |
| 641 | } |
| 642 | rt.mu.Unlock() |
| 643 | for _, ls := range doomed { |
| 644 | ls.kill(reason) |
| 645 | } |
| 646 | } |
| 647 | |
| 648 | // hostHasLabel reports whether host is <label>.<something>. |
| 649 | func hostHasLabel(host, label string) bool { |
| 650 | return len(host) > len(label) && host[len(label)] == '.' && host[:len(label)] == label |
| 651 | } |
| 652 | |
| 653 | // hostsOfLabel are the hosts the table knows (a connection bound to them, or streams still open on them) whose label is label. A host is |
| 654 | // <label>.<zone>, and the label is the first 20 characters: the control plane's answer is the only thing that ever put a host in the table. |
| 655 | func (rt *routeTable) hostsOfLabel(label string) []string { |
| 656 | rt.mu.Lock() |
| 657 | defer rt.mu.Unlock() |
| 658 | var hosts []string |
| 659 | for host := range rt.byHost { |
| 660 | if hostHasLabel(host, label) { |
| 661 | hosts = append(hosts, host) |
| 662 | } |
| 663 | } |
| 664 | return hosts |
| 665 | } |
| 666 | |
| 667 | // controls are the connections bound to a host of label, and every connection when label is "". |
| 668 | func (rt *routeTable) controls(label string) []*ctrl { |
| 669 | rt.mu.Lock() |
| 670 | defer rt.mu.Unlock() |
| 671 | var out []*ctrl |
| 672 | for host, he := range rt.byHost { |
| 673 | if he.cc != nil && (label == "" || hostHasLabel(host, label)) { |
| 674 | out = append(out, he.cc) |
| 675 | } |
| 676 | } |
| 677 | return out |
| 678 | } |
| 679 | |
| 680 | // ---------------------------------------------------------------- the control connection of one device |
| 681 | |
| 682 | // ctrl is one control connection of a device: from HELLO to its close. |
| 683 | type ctrl struct { |
| 684 | s *Server |
| 685 | id string |
| 686 | conn net.Conn |
| 687 | wmu sync.Mutex |
| 688 | |
| 689 | mu sync.Mutex |
| 690 | sub, host string |
| 691 | status string |
| 692 | plan string // paid or trial, as the control plane last said (the grace is for paid) |
| 693 | jti string // the session the device last authenticated with: what a refresh of this connection tells the control plane |
| 694 | authAt int64 // the control plane's own time (epoch s) at its last answer for this connection: a feed event older than it is already in it |
| 695 | validUntil time.Time // what the control plane granted, never changed by the node |
| 696 | graceUntil time.Time // a paid connection's end while the control plane does not answer (zero when it is not in its grace) |
| 697 | renewDown bool // the last RENEW got no answer from the control plane (a timeout, a 5xx, a 429) |
| 698 | refused bool // the last RENEW got an answer from the control plane, and it was no: the control plane being down later does not extend it |
| 699 | ready bool // an OK was sent: the connection is in the routing table |
| 700 | lastRenew time.Time |
| 701 | stop func() bool // the timer that ends the connection at validUntil (or at the end of its grace) |
| 702 | |
| 703 | refreshing, refreshAgain bool // a refresh of this connection (feed pause/resume) is running, and another event came while it ran |
| 704 | |
| 705 | done chan struct{} |
| 706 | once sync.Once |
| 707 | reason atomic.Value // string, why it was closed |
| 708 | } |
| 709 | |
| 710 | func (s *Server) newCtrl(c net.Conn) *ctrl { |
| 711 | cc := &ctrl{s: s, id: fmt.Sprintf("c%d", s.connSeq.Add(1)), conn: c, done: make(chan struct{})} |
| 712 | s.mu.Lock() |
| 713 | s.conns[cc] = struct{}{} |
| 714 | s.mu.Unlock() |
| 715 | return cc |
| 716 | } |
| 717 | |
| 718 | func (c *ctrl) subject() string { c.mu.Lock(); defer c.mu.Unlock(); return c.sub } |
| 719 | func (c *ctrl) hostName() string { c.mu.Lock(); defer c.mu.Unlock(); return c.host } |
| 720 | |
| 721 | func (c *ctrl) bind(sub, host string) { |
| 722 | c.mu.Lock() |
| 723 | c.sub, c.host = sub, host |
| 724 | c.mu.Unlock() |
| 725 | } |
| 726 | |
| 727 | // setReady records what the control plane decided: the connection serves streams from now on, until validUntil. A new decision ends any grace. |
| 728 | func (c *ctrl) setReady(resp AuthorizeResponse, jti string, validUntil time.Time) { |
| 729 | c.mu.Lock() |
| 730 | c.ready, c.status, c.plan, c.jti, c.authAt, c.validUntil = true, resp.Status, resp.Plan, jti, resp.ServerTime, validUntil |
| 731 | c.graceUntil, c.renewDown, c.refused = time.Time{}, false, false |
| 732 | c.mu.Unlock() |
| 733 | } |
| 734 | |
| 735 | // setStatus records a status the control plane gave in an answer that does not extend the connection (a refresh after a pause or a resume). |
| 736 | func (c *ctrl) setStatus(resp AuthorizeResponse) { |
| 737 | c.mu.Lock() |
| 738 | c.status, c.plan, c.authAt = resp.Status, resp.Plan, resp.ServerTime |
| 739 | c.mu.Unlock() |
| 740 | } |
| 741 | |
| 742 | // pause stops this connection from opening streams until the control plane says otherwise. |
| 743 | func (c *ctrl) pause() { |
| 744 | c.mu.Lock() |
| 745 | c.status = statusPaused |
| 746 | c.mu.Unlock() |
| 747 | } |
| 748 | |
| 749 | // markRenew records how the control plane came out of a RENEW that was not granted: it did not answer (down), or it answered no (refused). |
| 750 | func (c *ctrl) markRenew(down, refused bool) { |
| 751 | c.mu.Lock() |
| 752 | c.renewDown, c.refused = down, refused |
| 753 | c.mu.Unlock() |
| 754 | } |
| 755 | |
| 756 | // canOpen reports whether a client stream may be opened on this connection at now: an OK was sent, the device is not paused and valid_until has not |
| 757 | // come. A paused connection stays up (a resume arrives as a statement) but opens nothing. |
| 758 | func (c *ctrl) canOpen(now time.Time) bool { |
| 759 | c.mu.Lock() |
| 760 | defer c.mu.Unlock() |
| 761 | return c.ready && c.status != statusPaused && now.Before(c.validUntil) |
| 762 | } |
| 763 | |
| 764 | // tooSoon records a RENEW at now and reports whether the previous one was less than gap ago. |
| 765 | func (c *ctrl) tooSoon(now time.Time, gap time.Duration) bool { |
| 766 | c.mu.Lock() |
| 767 | defer c.mu.Unlock() |
| 768 | if !c.lastRenew.IsZero() && now.Sub(c.lastRenew) < gap { |
| 769 | return true |
| 770 | } |
| 771 | c.lastRenew = now |
| 772 | return false |
| 773 | } |
| 774 | |
| 775 | func (c *ctrl) isReady() bool { c.mu.Lock(); defer c.mu.Unlock(); return c.ready } |
| 776 | |
| 777 | func (c *ctrl) writeLocked(line string, at time.Time) error { |
| 778 | err := c.sendLocked(line) |
| 779 | c.s.emit(at, map[string]any{"conn": c.id, "ev": "send", "line": line}) |
| 780 | return err |
| 781 | } |
| 782 | |
| 783 | // sendLocked puts one line on the wire without recording it, for a caller that records it itself. |
| 784 | func (c *ctrl) sendLocked(line string) error { |
| 785 | _ = c.conn.SetWriteDeadline(time.Now().Add(c.s.cfg.WriteTimeout)) |
| 786 | _, err := io.WriteString(c.conn, line+"\n") |
| 787 | return err |
| 788 | } |
| 789 | |
| 790 | func (c *ctrl) write(line string) error { return c.writeAt(line, c.s.now()) } |
| 791 | |
| 792 | // writeAt writes one line; at is the moment recorded in the trace (the moment the decision to write it was taken). |
| 793 | func (c *ctrl) writeAt(line string, at time.Time) error { |
| 794 | c.wmu.Lock() |
| 795 | defer c.wmu.Unlock() |
| 796 | return c.writeLocked(line, at) |
| 797 | } |
| 798 | |
| 799 | // close ends the connection for reason (a CLOSE reason, or another word for the log). A reason the device is told (CLOSE <reason>) is sent first, |
| 800 | // unless the writer is stuck. |
| 801 | func (c *ctrl) close(reason string) { |
| 802 | c.once.Do(func() { |
| 803 | c.reason.Store(reason) |
| 804 | close(c.done) |
| 805 | c.mu.Lock() |
| 806 | if c.stop != nil { |
| 807 | c.stop() |
| 808 | } |
| 809 | c.mu.Unlock() |
| 810 | // The CLOSE line goes out under the write lock, so it follows an OK that is still being written; a writer that is stuck is not waited for. |
| 811 | told, ready := false, false |
| 812 | if c.lockWrite(closeWait) { |
| 813 | if ready = c.isReady(); ready && inList(policyData.Protocol.CloseReasons, reason) { |
| 814 | _ = c.writeLocked(closeLine(reason), c.s.now()) |
| 815 | told = true |
| 816 | } |
| 817 | c.wmu.Unlock() |
| 818 | } else { |
| 819 | ready = c.isReady() |
| 820 | } |
| 821 | if ready && !told { |
| 822 | // The transport ended without a CLOSE line (the device went away, the write was stuck). The recording still has to know this |
| 823 | // connection is over, so a trace gets a marker that never went on the wire. |
| 824 | c.s.emit(c.s.now(), map[string]any{"conn": c.id, "ev": "send", "line": "CLOSE", "wire": false}) |
| 825 | } |
| 826 | _ = c.conn.Close() |
| 827 | host := c.hostName() |
| 828 | c.s.routes.remove(c) |
| 829 | c.s.mu.Lock() |
| 830 | delete(c.s.conns, c) |
| 831 | c.s.mu.Unlock() |
| 832 | // A connection that ends because its grant ended (not because the device went away) takes the streams of its host with it, unless a newer |
| 833 | // connection of the device already serves them. |
| 834 | if ready && endsStreams(reason) { |
| 835 | c.s.routes.killStreams(host, reason, false) |
| 836 | } |
| 837 | }) |
| 838 | } |
| 839 | |
| 840 | // endsStreams says whether a connection closed for reason takes the host's open streams with it: the grant ended (expired, revoked, released) or the |
| 841 | // kill switch is on. A device that went away, a replaced connection and a node going down leave them to their own limits. |
| 842 | func endsStreams(reason string) bool { |
| 843 | switch reason { |
| 844 | case closeExpired, closeRevoked, closeReleased, closeKilled: |
| 845 | return true |
| 846 | } |
| 847 | return false |
| 848 | } |
| 849 | |
| 850 | // closeWait is how long a close waits for a writer that holds the connection's write lock before it gives up on the CLOSE line. |
| 851 | const closeWait = 250 * time.Millisecond |
| 852 | |
| 853 | func (c *ctrl) lockWrite(wait time.Duration) bool { |
| 854 | deadline := time.Now().Add(wait) |
| 855 | for !c.wmu.TryLock() { |
| 856 | if time.Now().After(deadline) { |
| 857 | return false |
| 858 | } |
| 859 | time.Sleep(time.Millisecond) |
| 860 | } |
| 861 | return true |
| 862 | } |
| 863 | |
| 864 | func (c *ctrl) closeReason() string { |
| 865 | if r, ok := c.reason.Load().(string); ok { |
| 866 | return r |
| 867 | } |
| 868 | return "" |
| 869 | } |
| 870 | |
| 871 | // arm schedules the end of the connection at validUntil (CLOSE expired); a RENEW arms it again for the new moment. |
| 872 | func (c *ctrl) arm(validUntil time.Time) { |
| 873 | c.mu.Lock() |
| 874 | defer c.mu.Unlock() |
| 875 | if c.stop != nil { |
| 876 | c.stop() |
| 877 | } |
| 878 | c.stop = c.s.clk.AfterFunc(validUntil.Sub(c.s.now()), func() { c.expire(validUntil) }) |
| 879 | } |
| 880 | |
| 881 | // expire is the end of the grant the timer was armed for. A grant a RENEW replaced since is not this one, and nothing happens. Otherwise the |
| 882 | // connection closes, except for a paid one while the control plane does not answer: it keeps its connection and its streams (but opens no new |
| 883 | // stream: canOpen is false from valid_until) for the grace, and closes when that is over. A trial or a gated connection is never extended, and |
| 884 | // neither is a paid one the control plane did answer (its RENEW was refused, or never came). |
| 885 | func (c *ctrl) expire(validUntil time.Time) { |
| 886 | c.mu.Lock() |
| 887 | current := c.validUntil.Equal(validUntil) |
| 888 | inGrace := !c.graceUntil.IsZero() |
| 889 | paid := c.plan == "paid" |
| 890 | down := c.renewDown |
| 891 | refused := c.refused |
| 892 | c.mu.Unlock() |
| 893 | if !current { |
| 894 | return |
| 895 | } |
| 896 | now := c.s.now() |
| 897 | if !inGrace && paid && !refused && c.s.cfg.Grace > 0 && (down || c.s.health.down(now, c.s.cfg.DownQuiet)) { |
| 898 | graceUntil := validUntil.Add(c.s.cfg.Grace) |
| 899 | c.mu.Lock() |
| 900 | c.graceUntil = graceUntil |
| 901 | if c.stop != nil { |
| 902 | c.stop() |
| 903 | } |
| 904 | c.stop = c.s.clk.AfterFunc(graceUntil.Sub(now), func() { c.expire(validUntil) }) |
| 905 | c.mu.Unlock() |
| 906 | c.s.log.Warn("control_grace", "conn", c.id, "grace_s", int64(c.s.cfg.Grace/time.Second)) |
| 907 | return |
| 908 | } |
| 909 | c.close(closeExpired) |
| 910 | } |
| 911 | |
| 912 | // bufConn serves reads through the bufio.Reader that parsed the first protocol line, so no byte the peer sent right after it is lost. |
| 913 | type bufConn struct { |
| 914 | net.Conn |
| 915 | r *bufio.Reader |
| 916 | } |
| 917 | |
| 918 | func (b *bufConn) Read(p []byte) (int, error) { return b.r.Read(p) } |
| 919 | |
| 920 | func (b *bufConn) CloseWrite() error { |
| 921 | if cw, ok := b.Conn.(interface{ CloseWrite() error }); ok { |
| 922 | return cw.CloseWrite() |
| 923 | } |
| 924 | return b.Conn.Close() |
| 925 | } |
| 926 | |
| 927 | // prefixConn replays the bytes already read from a connection before reading on: the ClientHello that was peeked to learn the SNI. |
| 928 | type prefixConn struct { |
| 929 | net.Conn |
| 930 | prefix []byte |
| 931 | } |
| 932 | |
| 933 | func (p *prefixConn) Read(b []byte) (int, error) { |
| 934 | if len(p.prefix) > 0 { |
| 935 | n := copy(b, p.prefix) |
| 936 | p.prefix = p.prefix[n:] |
| 937 | return n, nil |
| 938 | } |
| 939 | return p.Conn.Read(b) |
| 940 | } |
| 941 | |
| 942 | // ---------------------------------------------------------------- :443 |
| 943 | |
| 944 | // sourceOf is the source address of an accepted connection and whether it is in the OpenAI list. |
| 945 | func (s *Server) sourceOf(c net.Conn) (peer netip.Addr, ok, openai bool) { |
| 946 | if s.cfg.peerOf != nil { |
| 947 | peer, ok = s.cfg.peerOf(c) |
| 948 | } else { |
| 949 | peer, ok = peerAddr(c) |
| 950 | } |
| 951 | return peer, ok, ok && s.openai.contains(peer) |
| 952 | } |
| 953 | |
| 954 | // short is the per-boot salted short hash of a word (an address, a name): how a log line tells two of them apart without holding either. The salt |
| 955 | // lives in memory only, so the hash cannot be read back, and it changes at every start, so it cannot be followed from one boot to the next. |
| 956 | func (s *Server) short(word string) string { |
| 957 | return fmt.Sprintf("%08x", uint32(maphash.String(s.seed, word))) |
| 958 | } |
| 959 | |
| 960 | // refused counts a connection or stream the node did not serve and writes its log line, at most one a second per reason: a flood must not become a flood |
| 961 | // of log lines (the count is in the metrics, and the line says how many were left out). The line never names a source or a name. |
| 962 | func (s *Server) refused(level slog.Level, msg, reason string, attrs ...any) { |
| 963 | s.met.rejected.inc(reason) |
| 964 | if ok, cut := s.sampler.allow(msg + "/" + reason); ok { |
| 965 | attrs = append(attrs, "reason", reason) |
| 966 | if cut > 0 { |
| 967 | attrs = append(attrs, "suppressed", cut) |
| 968 | } |
| 969 | s.log.Log(context.Background(), level, msg, attrs...) |
| 970 | } |
| 971 | } |
| 972 | |
| 973 | // handleHTTPS is the one public listener. It reads only the cleartext ClientHello for the SNI: the node's own name is the control handler (the |
| 974 | // device speaks protocol v2 to it over TLS with ALPN all-tunnel/2), any other name is a client of a device and is spliced as raw bytes. |
| 975 | // |
| 976 | // Before a byte is read, the source is admitted: a source that is banned, that opens connections faster than the policy lets it, or that already holds |
| 977 | // the most it may is dropped without an answer (limits.go). |
| 978 | func (s *Server) handleHTTPS(c net.Conn) { |
| 979 | defer s.wg.Done() |
| 980 | handedOff := false // an ATTACHed data connection belongs to its stream from then on |
| 981 | defer func() { |
| 982 | if !handedOff { |
| 983 | _ = c.Close() |
| 984 | } |
| 985 | }() |
| 986 | id := newStreamID() |
| 987 | start := s.now() |
| 988 | |
| 989 | peer, havePeer, fromOpenAI := s.sourceOf(c) |
| 990 | slotDone := func() {} |
| 991 | if havePeer { |
| 992 | rel, why := s.ips.admit(peer, fromOpenAI) |
| 993 | if why != "" { |
| 994 | s.refused(slog.LevelInfo, "conn_reject", why, "stream", id) |
| 995 | return |
| 996 | } |
| 997 | slotDone = rel |
| 998 | } |
| 999 | defer slotDone() |
| 1000 | |
| 1001 | // The unauthenticated budget has a reserve, as the stream budget has: a source that is neither an address of the OpenAI list, nor one that holds (or just held) a device's |
| 1002 | // control connection, nor one a device marked good, may not take the share kept for those (policy node_limits.openai_reserved_pct). Twenty anonymous addresses holding the |
| 1003 | // most a source may hold fill the anonymous share, never the whole budget: ChatGPT's clients and the devices that reconnect (they are unauthenticated until their PROOF) go on. |
| 1004 | anon := !(havePeer && (fromOpenAI || s.ips.knownDevice(peer) || s.hints.goodAnywhere(peer))) |
| 1005 | if anon && !tryAcquire(&s.anonUnauth, s.anonUnauthMax()) { |
| 1006 | s.refused(slog.LevelWarn, "conn_reject", "unauth_anon_limit") |
| 1007 | return |
| 1008 | } |
| 1009 | if !tryAcquire(&s.unauth, int64(s.cfg.MaxUnauth)) { |
| 1010 | if anon { |
| 1011 | s.anonUnauth.Add(-1) |
| 1012 | } |
| 1013 | s.refused(slog.LevelWarn, "conn_reject", "unauth_limit") |
| 1014 | return |
| 1015 | } |
| 1016 | released := false |
| 1017 | release := func() { |
| 1018 | if !released { |
| 1019 | released = true |
| 1020 | s.unauth.Add(-1) |
| 1021 | if anon { |
| 1022 | s.anonUnauth.Add(-1) |
| 1023 | } |
| 1024 | } |
| 1025 | } |
| 1026 | defer release() |
| 1027 | var unmark func() |
| 1028 | // strike is a source that sent something no client of a device sends: a name that is not a device online, or a ClientHello that is no TLS. Too |
| 1029 | // many different ones and the source is dropped at accept for a while (an address of the OpenAI list is never struck: ChatGPT retrying an |
| 1030 | // offline device is not a scan). |
| 1031 | strike := func(name string) { |
| 1032 | // An address a device marked good (it just completed authenticated calls) is not a scanner, whatever names it asks for: at fleet scale a shared |
| 1033 | // egress address asks for many devices that are offline at that moment. |
| 1034 | if havePeer && !s.hints.goodAnywhere(peer) && s.ips.strike(peer, name, fromOpenAI) { |
| 1035 | s.met.bansTotal.Add(1) |
| 1036 | s.log.Info("source_banned", "source", s.short(peer.String()), "ban_s", int64(s.cfg.UnknownSNIBan/time.Second)) |
| 1037 | } |
| 1038 | } |
| 1039 | |
| 1040 | peeked, sni, err := readClientHello(c, s.cfg.HelloTimeout) |
| 1041 | if err != nil && !errors.Is(err, errBadSNI) { |
| 1042 | // A ClientHello that is truncated, malformed, too large or never finished never produces an OPEN. |
| 1043 | s.emit(start, map[string]any{"ev": "client_hello", "sni": "", "complete": false}) |
| 1044 | reason := helloReason(err) |
| 1045 | s.refused(slog.LevelInfo, "conn_reject", reason, "stream", id, "peeked_bytes", len(peeked)) |
| 1046 | if reason == "not_tls" || reason == "bad_hello" || reason == "hello_too_large" { |
| 1047 | strike("") |
| 1048 | } |
| 1049 | return |
| 1050 | } |
| 1051 | if err != nil || sni == "" { |
| 1052 | reason := "no_sni" |
| 1053 | if err != nil { |
| 1054 | reason = "bad_sni" |
| 1055 | } |
| 1056 | s.emit(s.now(), map[string]any{"ev": "tls_alert", "sni": ""}) |
| 1057 | s.refused(slog.LevelInfo, "conn_reject", reason, "stream", id, "peeked_bytes", len(peeked)) |
| 1058 | strike("") |
| 1059 | sendAlert(c, alertUnrecognizedName) |
| 1060 | return |
| 1061 | } |
| 1062 | if sni == s.cfg.NodeHost { |
| 1063 | // A connection of the node's own name that proves who it is (a device's control connection after its proof, a data connection after its ATTACH) |
| 1064 | // stops being an unknown source's: it gives back its slot and the token it took, and a control connection marks its address as a device's for as |
| 1065 | // long as it lives, so that the connections of the streams it serves are not held to what an anonymous source gets (limits.go). |
| 1066 | defer func() { |
| 1067 | if unmark != nil { |
| 1068 | unmark() |
| 1069 | } |
| 1070 | }() |
| 1071 | handedOff = s.serveOwnName(c, peeked, func(control bool) { |
| 1072 | release() |
| 1073 | if havePeer { |
| 1074 | s.ips.refund(peer, fromOpenAI) |
| 1075 | if control && unmark == nil { |
| 1076 | unmark = s.ips.markDevice(peer) |
| 1077 | } |
| 1078 | } |
| 1079 | slotDone() |
| 1080 | }) |
| 1081 | return |
| 1082 | } |
| 1083 | release() |
| 1084 | s.met.observeClientHello(peeked) |
| 1085 | s.serveClient(c, id, sni, peeked, start, peer, havePeer, strike) |
| 1086 | } |
| 1087 | |
| 1088 | // classOf names the class of a client stream's source for the OPEN: the configured classifier when there is one, otherwise openai for an address of |
| 1089 | // the OpenAI list, hint for one the device marked, anon for the rest. marked says whether the device marked the address (what puts the stream in the |
| 1090 | // host's hint slice, whatever class the OPEN says). |
| 1091 | func (s *Server) classOf(peer netip.Addr, havePeer bool, host string) (class string, marked bool) { |
| 1092 | class = classAnon |
| 1093 | if havePeer { |
| 1094 | marked = s.hints.good(host, peer) |
| 1095 | } |
| 1096 | switch { |
| 1097 | case s.cfg.Classify != nil && havePeer: |
| 1098 | if got := s.cfg.Classify(peer, host); inList(policyData.Protocol.SourceClasses, got) { |
| 1099 | class = got |
| 1100 | } |
| 1101 | case havePeer && s.openai.contains(peer): |
| 1102 | class = classOpenAI |
| 1103 | case marked: |
| 1104 | class = classHint |
| 1105 | } |
| 1106 | return class, marked || class == classHint |
| 1107 | } |
| 1108 | |
| 1109 | // serveClient splices one client of a device. Everything the node learned is the SNI and the ClientHello's bytes, and the OPEN is sent only now |
| 1110 | // that the ClientHello is complete and well formed. |
| 1111 | func (s *Server) serveClient(c net.Conn, id, sni string, peeked []byte, start time.Time, peer netip.Addr, havePeer bool, strike func(name string)) { |
| 1112 | if s.cfg.divert != nil && s.cfg.divert(c, peeked, s.log) { |
| 1113 | return |
| 1114 | } |
| 1115 | class, marked := s.classOf(peer, havePeer, sni) |
| 1116 | peerText := "0.0.0.0" |
| 1117 | if havePeer { |
| 1118 | peerText = peer.String() |
| 1119 | } |
| 1120 | hello := map[string]any{"ev": "client_hello", "sni": sni, "complete": true, "peer_ip": peerText, "class": class} |
| 1121 | s.emit(s.now(), hello) |
| 1122 | |
| 1123 | if s.draining.Load() { |
| 1124 | // A draining node takes no new client: the TLS alert, never an OPEN. (The client's ClientHello is not forwarded anywhere.) |
| 1125 | s.refused(slog.LevelInfo, "stream_reject", "draining", "stream", id, "peeked_bytes", len(peeked)) |
| 1126 | s.alert(c, sni, alertInternalError) |
| 1127 | return |
| 1128 | } |
| 1129 | // The global budget: every stream counts in it, and a stream whose source is neither ChatGPT's (the OpenAI list) nor marked good by a device may not |
| 1130 | // take the share that is reserved for those (policy node_limits.openai_reserved_pct). |
| 1131 | if !tryAcquire(&s.streams, int64(s.cfg.MaxStreams)) { |
| 1132 | s.refused(slog.LevelWarn, "stream_reject", "global_limit", "stream", id) |
| 1133 | s.alert(c, sni, alertInternalError) |
| 1134 | return |
| 1135 | } |
| 1136 | defer s.streams.Add(-1) |
| 1137 | if class == classAnon && !marked { |
| 1138 | if !tryAcquire(&s.anonStreams, int64(s.cfg.MaxStreams-s.cfg.MaxStreams*s.cfg.OpenAIReservedPct/100)) { |
| 1139 | s.refused(slog.LevelWarn, "stream_reject", "anon_global_limit", "stream", id) |
| 1140 | s.alert(c, sni, alertInternalError) |
| 1141 | return |
| 1142 | } |
| 1143 | defer s.anonStreams.Add(-1) |
| 1144 | } |
| 1145 | |
| 1146 | cc, live, why := s.routes.acquire(sni, marked, s.now()) |
| 1147 | if why != "" { |
| 1148 | desc := byte(alertInternalError) |
| 1149 | if why == admitUnknown { |
| 1150 | desc = alertUnrecognizedName |
| 1151 | strike(sni) |
| 1152 | } |
| 1153 | s.refused(slog.LevelInfo, "stream_reject", why, "stream", id, "peeked_bytes", len(peeked)) |
| 1154 | s.alert(c, sni, desc) |
| 1155 | return |
| 1156 | } |
| 1157 | defer s.routes.release(sni, live) |
| 1158 | // The host's record of today's gated budget: every stream holds it, so that a host that turns gated while the stream is open is held to it too. |
| 1159 | live.usage = s.gates.acquire(sni) |
| 1160 | defer s.gates.release(live.usage) |
| 1161 | if live.he.gated.Load() { |
| 1162 | if ok, reason := live.usage.admit(s.now(), s.cfg.GatedBytesPerDay, s.cfg.GatedStreamsPerDay); !ok { |
| 1163 | s.refused(slog.LevelInfo, "stream_reject", reason, "stream", id, "peeked_bytes", len(peeked)) |
| 1164 | s.alert(c, sni, alertInternalError) |
| 1165 | return |
| 1166 | } |
| 1167 | } |
| 1168 | |
| 1169 | dc, err := s.openStream(cc, id, peerText, class) |
| 1170 | if err != nil { |
| 1171 | s.refused(slog.LevelInfo, "stream_reject", err.Error(), "stream", id, "peeked_bytes", len(peeked), "wait_ms", s.now().Sub(start).Milliseconds()) |
| 1172 | s.alert(c, sni, alertInternalError) |
| 1173 | return |
| 1174 | } |
| 1175 | defer dc.Close() |
| 1176 | if !live.bind(c, dc) { |
| 1177 | // Cut (a revoke, a pause) between the OPEN and the ATTACH: nothing of the client's goes to the device. |
| 1178 | s.refused(slog.LevelInfo, "stream_reject", live.killedBy(), "stream", id, "peeked_bytes", len(peeked)) |
| 1179 | s.alert(c, sni, alertInternalError) |
| 1180 | return |
| 1181 | } |
| 1182 | s.met.streamsOpened.Add(1) |
| 1183 | s.log.Info("stream_open", "stream", id, "hello_bytes", len(peeked), "setup_ms", s.now().Sub(start).Milliseconds(), "class", class) |
| 1184 | res := s.splice(id, c, dc, peeked, live) |
| 1185 | if why := live.killedBy(); why != "" { |
| 1186 | res.reason = why |
| 1187 | } |
| 1188 | s.met.streamEnd.inc(res.reason) |
| 1189 | s.log.Info("stream_close", "stream", id, "up_bytes", res.up, "down_bytes", res.down, |
| 1190 | "duration_ms", s.now().Sub(start).Milliseconds(), "reason", res.reason) |
| 1191 | } |
| 1192 | |
| 1193 | func (s *Server) alert(c net.Conn, sni string, desc byte) { |
| 1194 | s.emit(s.now(), map[string]any{"ev": "tls_alert", "sni": sni}) |
| 1195 | sendAlert(c, desc) |
| 1196 | } |
| 1197 | |
| 1198 | // openStream registers a pending stream, tells the device (OPEN over its control connection) and waits for its ATTACH. The returned connection has |
| 1199 | // had its ATTACH line consumed; the caller's first write is the first byte of the raw stream (there is deliberately no reply line in between, so a |
| 1200 | // device can hand the connection straight to its TLS layer). |
| 1201 | func (s *Server) openStream(cc *ctrl, id, peer, class string) (net.Conn, error) { |
| 1202 | at := s.now() |
| 1203 | if !cc.canOpen(at) { |
| 1204 | return nil, errOffline |
| 1205 | } |
| 1206 | secret := newSecret() |
| 1207 | p := &pendingStream{cc: cc, secret: secret, ch: make(chan net.Conn, 1)} |
| 1208 | s.mu.Lock() |
| 1209 | s.pending[id] = p |
| 1210 | s.mu.Unlock() |
| 1211 | |
| 1212 | if err := cc.writeAt(formatLine(message{"verb": "OPEN", "stream_id": id, "secret": secret, "peer_ip": peer, "class": class}), at); err != nil { |
| 1213 | s.dropPending(id) |
| 1214 | cc.close("write_error") |
| 1215 | return nil, errControlWrite |
| 1216 | } |
| 1217 | timer := time.NewTimer(s.cfg.AttachTimeout) |
| 1218 | defer timer.Stop() |
| 1219 | done := cc.done |
| 1220 | wait: |
| 1221 | for { |
| 1222 | select { |
| 1223 | case dc := <-p.ch: |
| 1224 | return dc, nil |
| 1225 | case <-timer.C: |
| 1226 | break wait |
| 1227 | case <-done: |
| 1228 | if cc.closeReason() != closeReplaced { |
| 1229 | break wait |
| 1230 | } |
| 1231 | // A newer connection of the same device took over: this OPEN was sent with the old one, and its attach secret is still good. |
| 1232 | done = nil |
| 1233 | case <-s.ctx.Done(): |
| 1234 | break wait |
| 1235 | } |
| 1236 | } |
| 1237 | // Failed: unregister, but an ATTACH may have just landed. |
| 1238 | if !s.dropPending(id) { |
| 1239 | return <-p.ch, nil |
| 1240 | } |
| 1241 | switch { |
| 1242 | case s.ctx.Err() != nil: |
| 1243 | return nil, errShuttingDown |
| 1244 | case cc.closeReason() != "" && cc.closeReason() != closeReplaced: |
| 1245 | return nil, errOffline |
| 1246 | } |
| 1247 | return nil, errAttachTimeout |
| 1248 | } |
| 1249 | |
| 1250 | // dropPending removes id; it reports whether the entry was still there. |
| 1251 | func (s *Server) dropPending(id string) bool { |
| 1252 | s.mu.Lock() |
| 1253 | defer s.mu.Unlock() |
| 1254 | _, ok := s.pending[id] |
| 1255 | delete(s.pending, id) |
| 1256 | return ok |
| 1257 | } |
| 1258 | |
| 1259 | // ---------------------------------------------------------------- splice |
| 1260 | |
| 1261 | type spliceResult struct { |
| 1262 | up, down int64 // client->device, device->client |
| 1263 | reason string |
| 1264 | } |
| 1265 | |
| 1266 | func classifyErr(err error) string { |
| 1267 | switch { |
| 1268 | case err == nil: |
| 1269 | return "closed" |
| 1270 | case errors.Is(err, errByteCap): |
| 1271 | return "byte_cap" |
| 1272 | case errors.Is(err, errGatedBytes): |
| 1273 | return "gated_bytes" |
| 1274 | case errors.Is(err, os.ErrDeadlineExceeded): |
| 1275 | return "idle" |
| 1276 | default: |
| 1277 | return "conn_error" |
| 1278 | } |
| 1279 | } |
| 1280 | |
| 1281 | func closeWrite(c net.Conn) { |
| 1282 | if cw, ok := c.(interface{ CloseWrite() error }); ok { |
| 1283 | _ = cw.CloseWrite() |
| 1284 | return |
| 1285 | } |
| 1286 | _ = c.Close() |
| 1287 | } |
| 1288 | |
| 1289 | // pump copies src to dst with per-read idle deadlines and the shared byte budget. A clean EOF half-closes dst and returns nil. While the stream's host |
| 1290 | // is gated, every chunk is also charged to the host's budget of the day, and the stream ends where the budget does. watch, when set, reads the |
| 1291 | // device's first bytes for the handshake observer. |
| 1292 | func (s *Server) pump(id, dir string, dst, src net.Conn, ctr, total *atomic.Int64, live *liveStream, watch *serverHelloWatch) error { |
| 1293 | // A pump spends the whole life of an idle stream blocked in a read, so the buffer it waits with is a small one (4 KiB); it takes a 32 KiB one only for a read that filled |
| 1294 | // the small one (a bulk transfer) and gives it back when a read comes back short. An idle stream then holds about 8 KiB of buffer instead of 64 KiB (the acceptance is under 64 KiB of memory per stream). |
| 1295 | sp := s.smallBufPool.Get().(*[]byte) |
| 1296 | small := *sp |
| 1297 | defer s.smallBufPool.Put(sp) |
| 1298 | var bp *[]byte // the big buffer, while this pump holds one |
| 1299 | giveBack := func() { |
| 1300 | if bp != nil { |
| 1301 | s.bigBuffersInUse.Add(-1) |
| 1302 | s.bufPool.Put(bp) |
| 1303 | bp = nil |
| 1304 | } |
| 1305 | } |
| 1306 | defer giveBack() |
| 1307 | buf := small |
| 1308 | for { |
| 1309 | _ = src.SetReadDeadline(time.Now().Add(s.cfg.StreamIdle)) |
| 1310 | n, rerr := src.Read(buf) |
| 1311 | if n > 0 { |
| 1312 | chunk, spent := buf[:n], false |
| 1313 | if live != nil && live.he.gated.Load() { |
| 1314 | if got := live.usage.charge(s.now(), int64(n), s.cfg.GatedBytesPerDay); got < int64(n) { |
| 1315 | chunk, spent = buf[:got], true |
| 1316 | } |
| 1317 | } |
| 1318 | if total.Add(int64(len(chunk))) > s.cfg.MaxStreamBytes { |
| 1319 | return errByteCap |
| 1320 | } |
| 1321 | if len(chunk) > 0 { |
| 1322 | s.tapBytes(id, dir, chunk) |
| 1323 | if watch != nil { |
| 1324 | watch.feed(chunk) |
| 1325 | } |
| 1326 | _ = dst.SetWriteDeadline(time.Now().Add(s.cfg.StreamIdle)) |
| 1327 | if _, werr := dst.Write(chunk); werr != nil { |
| 1328 | return werr |
| 1329 | } |
| 1330 | ctr.Add(int64(len(chunk))) |
| 1331 | s.met.bytes.add(dir, int64(len(chunk))) |
| 1332 | } |
| 1333 | if spent { |
| 1334 | s.met.gatedCut.Add(1) |
| 1335 | return errGatedBytes |
| 1336 | } |
| 1337 | } |
| 1338 | if rerr != nil { |
| 1339 | if rerr == io.EOF { |
| 1340 | closeWrite(dst) |
| 1341 | return nil |
| 1342 | } |
| 1343 | return rerr |
| 1344 | } |
| 1345 | switch { // the buffer for the next read: bulk traffic gets the big one, a trickle (and the wait of an idle stream) the small one |
| 1346 | case bp == nil && n == len(small): |
| 1347 | bp = s.bufPool.Get().(*[]byte) |
| 1348 | s.bigBuffersInUse.Add(1) |
| 1349 | buf = *bp |
| 1350 | case bp != nil && n < len(small): |
| 1351 | giveBack() |
| 1352 | buf = small |
| 1353 | } |
| 1354 | } |
| 1355 | } |
| 1356 | |
| 1357 | // splice copies bytes both ways between client and dev until both sides are done, one fails, or a limit trips. preload (the peeked ClientHello) is |
| 1358 | // written to dev before anything else from the client. Neither side's bytes are interpreted. |
| 1359 | func (s *Server) splice(id string, client, dev net.Conn, preload []byte, live *liveStream) spliceResult { |
| 1360 | var up, down, total atomic.Int64 |
| 1361 | var rmu sync.Mutex |
| 1362 | reason := "" |
| 1363 | setReason := func(r string) { |
| 1364 | rmu.Lock() |
| 1365 | if reason == "" { |
| 1366 | reason = r |
| 1367 | } |
| 1368 | rmu.Unlock() |
| 1369 | } |
| 1370 | closeBoth := func() { _ = client.Close(); _ = dev.Close() } |
| 1371 | |
| 1372 | age := time.AfterFunc(s.cfg.StreamMaxAge, func() { setReason("max_age"); closeBoth() }) |
| 1373 | defer age.Stop() |
| 1374 | stopWatch := make(chan struct{}) |
| 1375 | defer close(stopWatch) |
| 1376 | go func() { |
| 1377 | select { |
| 1378 | case <-s.ctx.Done(): |
| 1379 | setReason("shutdown") |
| 1380 | closeBoth() |
| 1381 | case <-stopWatch: |
| 1382 | } |
| 1383 | }() |
| 1384 | |
| 1385 | results := make(chan error, 2) |
| 1386 | go func() { // client -> device |
| 1387 | if len(preload) > 0 { |
| 1388 | if live.he.gated.Load() && live.usage.charge(s.now(), int64(len(preload)), s.cfg.GatedBytesPerDay) < int64(len(preload)) { |
| 1389 | // a ClientHello cut short is of no use to anyone: nothing of it goes to the device |
| 1390 | s.met.gatedCut.Add(1) |
| 1391 | results <- errGatedBytes |
| 1392 | return |
| 1393 | } |
| 1394 | _ = dev.SetWriteDeadline(time.Now().Add(s.cfg.WriteTimeout)) |
| 1395 | if _, err := dev.Write(preload); err != nil { |
| 1396 | results <- err |
| 1397 | return |
| 1398 | } |
| 1399 | s.tapBytes(id, "up", preload) |
| 1400 | up.Add(int64(len(preload))) |
| 1401 | total.Add(int64(len(preload))) |
| 1402 | s.met.bytes.add("up", int64(len(preload))) |
| 1403 | } |
| 1404 | results <- s.pump(id, "up", dev, client, &up, &total, live, nil) |
| 1405 | }() |
| 1406 | go func() { // device -> client |
| 1407 | results <- s.pump(id, "down", client, dev, &down, &total, live, newServerHelloWatch(s.met)) |
| 1408 | }() |
| 1409 | |
| 1410 | first := <-results |
| 1411 | if first == nil { |
| 1412 | // One direction ended cleanly (and was half-closed): the other gets a bounded time to finish. |
| 1413 | t := time.NewTimer(s.cfg.HalfCloseTimeout) |
| 1414 | select { |
| 1415 | case second := <-results: |
| 1416 | setReason(classifyErr(second)) |
| 1417 | case <-t.C: |
| 1418 | setReason("half_close_timeout") |
| 1419 | closeBoth() |
| 1420 | <-results |
| 1421 | } |
| 1422 | t.Stop() |
| 1423 | } else { |
| 1424 | setReason(classifyErr(first)) |
| 1425 | closeBoth() |
| 1426 | <-results |
| 1427 | } |
| 1428 | closeBoth() |
| 1429 | rmu.Lock() |
| 1430 | defer rmu.Unlock() |
| 1431 | return spliceResult{up: up.Load(), down: down.Load(), reason: reason} |
| 1432 | } |
| 1433 | |
| 1434 | // ---------------------------------------------------------------- the node's own name: control and data connections of devices |
| 1435 | |
| 1436 | func readLine(br *bufio.Reader) (string, error) { |
| 1437 | b, err := br.ReadSlice('\n') |
| 1438 | if err != nil { |
| 1439 | return "", err |
| 1440 | } |
| 1441 | return string(b[:len(b)-1]), nil |
| 1442 | } |
| 1443 | |
| 1444 | func writeLineThenClose(c net.Conn, line string) { |
| 1445 | _ = c.SetWriteDeadline(time.Now().Add(2 * time.Second)) |
| 1446 | _, _ = io.WriteString(c, line+"\n") |
| 1447 | _ = c.Close() |
| 1448 | } |
| 1449 | |
| 1450 | // serveOwnName terminates the DEVICE-facing TLS on the node's own name (its public-CA certificate, TLS 1.3, ALPN all-tunnel/2), reads the first line |
| 1451 | // and dispatches HELLO or ATTACH. This is the tunnel's own connection to its own client, never a client's TLS. It reports whether the connection was |
| 1452 | // handed to a stream (an ATTACH that matched), in which case the caller must not close it. |
| 1453 | func (s *Server) serveOwnName(raw net.Conn, peeked []byte, release func(control bool)) bool { |
| 1454 | tc := tls.Server(&prefixConn{Conn: raw, prefix: peeked}, s.tlsCfg) |
| 1455 | _ = tc.SetDeadline(time.Now().Add(s.cfg.FirstLineTimeout)) |
| 1456 | if err := tc.Handshake(); err != nil { |
| 1457 | s.log.Debug("control_tls_failed") |
| 1458 | return false |
| 1459 | } |
| 1460 | if tc.ConnectionState().NegotiatedProtocol != policyData.Protocol.ALPN { |
| 1461 | s.log.Info("control_reject", "reason", "bad_alpn") |
| 1462 | return false |
| 1463 | } |
| 1464 | br := bufio.NewReaderSize(tc, policyData.Protocol.LineMaxBytes) |
| 1465 | line, err := readLine(br) |
| 1466 | if err != nil { |
| 1467 | if errors.Is(err, bufio.ErrBufferFull) { |
| 1468 | writeLineThenClose(tc, errLine(codeLineTooLong)) |
| 1469 | } else { |
| 1470 | s.log.Info("control_reject", "reason", "bad_first_line") |
| 1471 | } |
| 1472 | return false |
| 1473 | } |
| 1474 | first, perr := parseLine(line, dirD2N) |
| 1475 | if perr != nil { |
| 1476 | // An ATTACH or a HELLO with a bad field answers with its own word; anything else is not a command. |
| 1477 | code := errCode(perr) |
| 1478 | if v, _, _ := strings.Cut(line, " "); v != "HELLO" && v != "ATTACH" { |
| 1479 | code = codeBadCommand // contract section 2: a first line that is not HELLO or ATTACH is bad_command, whatever else is wrong with it |
| 1480 | } |
| 1481 | s.log.Info("control_reject", "reason", code) |
| 1482 | writeLineThenClose(tc, errLine(code)) |
| 1483 | return false |
| 1484 | } |
| 1485 | switch first.str("verb") { |
| 1486 | case "HELLO": |
| 1487 | cc := s.newCtrl(tc) |
| 1488 | s.emit(s.now(), map[string]any{"ev": "recv", "conn": cc.id, "line": line}) |
| 1489 | if s.draining.Load() { |
| 1490 | // A draining node takes no new device either: it is told to try the other node. (An ATTACH for a stream that already has its OPEN is served.) |
| 1491 | s.log.Info("control_reject", "conn", cc.id, "reason", "draining") |
| 1492 | _ = cc.write(errLine(codeControlUnavailable)) |
| 1493 | cc.close("refused") |
| 1494 | return false |
| 1495 | } |
| 1496 | s.serveControl(cc, br, first, release) |
| 1497 | return false |
| 1498 | case "ATTACH": |
| 1499 | return s.serveAttach(tc, br, first, line, fmt.Sprintf("d%d", s.connSeq.Add(1)), release) |
| 1500 | } |
| 1501 | s.log.Info("control_reject", "reason", codeBadCommand) |
| 1502 | writeLineThenClose(tc, errLine(codeBadCommand)) |
| 1503 | return false |
| 1504 | } |
| 1505 | |
| 1506 | // exchange is one authentication in flight on a control connection: the HELLO that opened it, or a RENEW. |
| 1507 | type exchange struct { |
| 1508 | kind string // "hello" or "renew" |
| 1509 | token string |
| 1510 | session *SessionClaims |
| 1511 | nonce string |
| 1512 | deadline time.Time |
| 1513 | } |
| 1514 | |
| 1515 | // refuse answers a failed HELLO or RENEW (or PROOF) with ERR. A refused HELLO closes the connection; a refused RENEW keeps it, with its old |
| 1516 | // valid_until. |
| 1517 | func (s *Server) refuse(cc *ctrl, kind, code string) bool { |
| 1518 | _ = cc.write(errLine(code)) |
| 1519 | if kind == "hello" { |
| 1520 | cc.close("refused") |
| 1521 | return false |
| 1522 | } |
| 1523 | return true |
| 1524 | } |
| 1525 | |
| 1526 | // beginExchange verifies the session of a HELLO or RENEW and, when it holds, sends the CHALLENGE. A RENEW must be for the same subject as the |
| 1527 | // connection. It reports whether the connection goes on. |
| 1528 | func (s *Server) beginExchange(cc *ctrl, ex *exchange, kind, token string) bool { |
| 1529 | *ex = exchange{} |
| 1530 | now := s.now() |
| 1531 | session, err := verifySession(token, now, s.cfg.Keys, s.cfg.ClockSkew) |
| 1532 | if err != nil { |
| 1533 | code := errFor(tokenWord(err), codeAuth) |
| 1534 | if kh, ok := s.cfg.Keys.(keyHealth); ok && tokenWord(err) == tokUnknownKID && !kh.Healthy() { |
| 1535 | // The node cannot vouch for the control plane's keys (stale beyond the policy's window): not the device's fault, so not auth. |
| 1536 | code = codeControlUnavailable |
| 1537 | if kind == "renew" { |
| 1538 | // ... and it is the control plane being unreachable, which is what a paid device's grace is for: without this the connection closed at valid_until although the |
| 1539 | // device did everything right (its session was refused only because the node could not read the control plane's keys). |
| 1540 | cc.markRenew(true, false) |
| 1541 | } |
| 1542 | } |
| 1543 | s.log.Info("session_refused", "conn", cc.id, "kind", kind, "reason", tokenWord(err), "answer", code) |
| 1544 | return s.refuse(cc, kind, code) |
| 1545 | } |
| 1546 | if kind == "renew" && session.Subject != cc.subject() { |
| 1547 | s.log.Info("session_refused", "conn", cc.id, "kind", kind, "reason", "other_subject") |
| 1548 | return s.refuse(cc, kind, codeAuth) |
| 1549 | } |
| 1550 | *ex = exchange{kind: kind, token: token, session: session, nonce: newNonce(), deadline: now.Add(s.cfg.FirstLineTimeout)} |
| 1551 | if err := cc.write(formatLine(message{"verb": "CHALLENGE", "nonce": ex.nonce})); err != nil { |
| 1552 | cc.close("write_error") |
| 1553 | return false |
| 1554 | } |
| 1555 | return true |
| 1556 | } |
| 1557 | |
| 1558 | // serveControl runs a control connection from its HELLO to its end. release gives back the unauthenticated slot once the device has authenticated. |
| 1559 | func (s *Server) serveControl(cc *ctrl, br *bufio.Reader, hello message, release func(control bool)) { |
| 1560 | tc := cc.conn |
| 1561 | defer cc.close("closed") |
| 1562 | defer s.routes.remove(cc) // idempotent: a close that raced an activation must not leave a route behind |
| 1563 | start := s.now() |
| 1564 | ex := &exchange{} |
| 1565 | if !s.beginExchange(cc, ex, "hello", hello.str("session")) { |
| 1566 | return |
| 1567 | } |
| 1568 | reason := "" |
| 1569 | for { |
| 1570 | deadline := s.now().Add(s.cfg.DeadPeer) |
| 1571 | if !cc.isReady() || ex.nonce != "" { |
| 1572 | deadline = s.now().Add(s.cfg.FirstLineTimeout) |
| 1573 | if !ex.deadline.IsZero() && ex.deadline.Before(deadline) { |
| 1574 | deadline = ex.deadline |
| 1575 | } |
| 1576 | } |
| 1577 | _ = tc.SetReadDeadline(deadline) |
| 1578 | line, err := readLine(br) |
| 1579 | if err != nil { |
| 1580 | switch { |
| 1581 | case errors.Is(err, bufio.ErrBufferFull): |
| 1582 | _ = cc.write(errLine(codeLineTooLong)) |
| 1583 | reason = closeProtocol |
| 1584 | case errors.Is(err, os.ErrDeadlineExceeded) && ex.nonce != "" && cc.isReady(): |
| 1585 | // The device never answered a RENEW's CHALLENGE: the exchange is dropped, the connection and its old valid_until stay. |
| 1586 | *ex = exchange{} |
| 1587 | continue |
| 1588 | case errors.Is(err, os.ErrDeadlineExceeded): |
| 1589 | reason = closeIdle |
| 1590 | default: |
| 1591 | reason = "closed" |
| 1592 | } |
| 1593 | cc.close(reason) |
| 1594 | s.log.Info("control_down", "conn", cc.id, "duration_ms", s.now().Sub(start).Milliseconds(), "reason", cc.closeReason()) |
| 1595 | return |
| 1596 | } |
| 1597 | s.emit(s.now(), map[string]any{"ev": "recv", "conn": cc.id, "line": line}) |
| 1598 | m, perr := parseLine(line, dirD2N) |
| 1599 | if perr != nil { |
| 1600 | _ = cc.write(errLine(errCode(perr))) |
| 1601 | cc.close(closeProtocol) |
| 1602 | return |
| 1603 | } |
| 1604 | switch m.str("verb") { |
| 1605 | case "PROOF": |
| 1606 | if ex.nonce == "" { |
| 1607 | _ = cc.write(errLine(codeBadCommand)) |
| 1608 | cc.close(closeProtocol) |
| 1609 | return |
| 1610 | } |
| 1611 | if !s.finishExchange(cc, ex, m.str("proof"), release) { |
| 1612 | return |
| 1613 | } |
| 1614 | case "RENEW": |
| 1615 | if !cc.isReady() || ex.nonce != "" { |
| 1616 | _ = cc.write(errLine(codeBadCommand)) |
| 1617 | cc.close(closeProtocol) |
| 1618 | return |
| 1619 | } |
| 1620 | if cc.tooSoon(s.now(), s.cfg.MinRenewGap) { |
| 1621 | // a device cannot turn one connection into a stream of authorize calls: the connection and its old valid_until stay |
| 1622 | if err := cc.write(errLine(codeBusy)); err != nil { |
| 1623 | cc.close("write_error") |
| 1624 | return |
| 1625 | } |
| 1626 | continue |
| 1627 | } |
| 1628 | if !s.beginExchange(cc, ex, "renew", m.str("session")) { |
| 1629 | return |
| 1630 | } |
| 1631 | case "PING": |
| 1632 | if err := cc.write(formatLine(message{"verb": "PONG"})); err != nil { |
| 1633 | cc.close("write_error") |
| 1634 | return |
| 1635 | } |
| 1636 | case "HINT": |
| 1637 | if cc.isReady() { |
| 1638 | var addrs []netip.Addr |
| 1639 | for _, a := range m["addresses"].([]any) { |
| 1640 | if addr, err := netip.ParseAddr(a.(string)); err == nil { |
| 1641 | addrs = append(addrs, addr) |
| 1642 | } |
| 1643 | } |
| 1644 | // filed under the host the control plane gave THIS connection: a device can mark addresses for no other host than its own |
| 1645 | s.hints.mark(cc.hostName(), addrs) |
| 1646 | s.met.hintsMarked.Add(1) |
| 1647 | if s.cfg.OnHint != nil { |
| 1648 | s.cfg.OnHint(cc.hostName(), addrs) |
| 1649 | } |
| 1650 | } |
| 1651 | case "CLOSE": |
| 1652 | cc.close("device_close") |
| 1653 | s.log.Info("control_down", "conn", cc.id, "duration_ms", s.now().Sub(start).Milliseconds(), "reason", "device_close") |
| 1654 | return |
| 1655 | default: // HELLO again, or anything else the device must not send here |
| 1656 | _ = cc.write(errLine(codeBadCommand)) |
| 1657 | cc.close(closeProtocol) |
| 1658 | return |
| 1659 | } |
| 1660 | } |
| 1661 | } |
| 1662 | |
| 1663 | // finishExchange checks the PROOF of the exchange in flight and, when it holds, asks the control plane what the device may do and answers OK. It |
| 1664 | // reports whether the connection goes on. |
| 1665 | func (s *Server) finishExchange(cc *ctrl, ex *exchange, proof string, release func(control bool)) bool { |
| 1666 | cur := *ex |
| 1667 | *ex = exchange{} // the nonce is spent whatever the verdict |
| 1668 | now := s.now() |
| 1669 | pc, err := verifyProof(proof, now, cur.token, cur.session, cur.nonce, s.nodeURL, s.cfg.ClockSkew) |
| 1670 | if err == nil && !s.replay.spend(pc.JTI, pc.IssuedAt, now, s.cfg.ClockSkew) { |
| 1671 | err = tokFail(tokReplayedJTI) |
| 1672 | } |
| 1673 | if err != nil { |
| 1674 | s.log.Info("proof_refused", "conn", cc.id, "kind", cur.kind, "reason", tokenWord(err)) |
| 1675 | return s.refuse(cc, cur.kind, errFor(tokenWord(err), codeBadProof)) |
| 1676 | } |
| 1677 | |
| 1678 | // The device has proven itself, so it no longer holds a pre-authentication slot: a slow control plane must not starve the ClientHellos of |
| 1679 | // clients and the ATTACHes of devices that are already connected. What waits for authorize is bounded on its own. |
| 1680 | if cur.kind == "hello" { |
| 1681 | release(true) |
| 1682 | } |
| 1683 | if s.killed.Load() { |
| 1684 | // The kill switch is on: nobody is served, and nothing is asked of the control plane until it is cleared. |
| 1685 | s.log.Info("authorize_skipped", "conn", cc.id, "kind", cur.kind, "reason", "kill_switch") |
| 1686 | return s.refuse(cc, cur.kind, codeControlUnavailable) |
| 1687 | } |
| 1688 | |
| 1689 | // The control plane decides; nothing the device said names a host. Until its answer is acted on the feed is listened to: an event the feed |
| 1690 | // applies while the call is in flight is not in the answer, and cannot find a connection that is not yet in the routing table. |
| 1691 | watch := s.watchFeed() |
| 1692 | defer watch.stop() |
| 1693 | req := AuthorizeRequest{DeviceID: cur.session.Subject, SessionJTI: cur.session.JTI, Reason: cur.kind, Node: s.cfg.NodeID} |
| 1694 | s.emit(s.now(), map[string]any{"conn": cc.id, "ev": "cp", "call": "authorize", "body": req.toMap()}) |
| 1695 | resp, err := s.authorize(req) |
| 1696 | if err == nil { |
| 1697 | if cerr := checkAuthorize(resp, req); cerr != nil { |
| 1698 | s.log.Warn("authorize_answer_refused", "conn", cc.id, "reason", cerr.Error()) |
| 1699 | s.health.bad(s.now()) |
| 1700 | err = &ControlError{Status: 502, Code: "bad_response"} |
| 1701 | } |
| 1702 | } |
| 1703 | if err != nil { |
| 1704 | status, word := 503, "control_unavailable" |
| 1705 | if ce, ok := err.(*ControlError); ok { |
| 1706 | status, word = ce.Status, ce.Code |
| 1707 | } |
| 1708 | s.emit(s.now(), map[string]any{"conn": cc.id, "ev": "cp_reply", "status": status, "body": map[string]any{"ok": false, "error": word}}) |
| 1709 | code := errForControl(err) |
| 1710 | if cur.kind == "renew" && !isLocalFailure(err) { |
| 1711 | // A RENEW this node could not even send (its own queue was full or its turn did not come) says nothing about the control plane. |
| 1712 | cc.markRenew(code == codeControlUnavailable, code != codeControlUnavailable) |
| 1713 | } |
| 1714 | s.log.Info("authorize_refused", "conn", cc.id, "kind", cur.kind, "reason", code) |
| 1715 | return s.refuse(cc, cur.kind, code) |
| 1716 | } |
| 1717 | s.emit(s.now(), map[string]any{"conn": cc.id, "ev": "cp_reply", "status": 200, "body": resp.toMap()}) |
| 1718 | if resp.E2EKill { |
| 1719 | // The device is told first (the switch closes every connection, this one included), then nobody is served until it is cleared. |
| 1720 | s.log.Warn("kill_switch_seen", "conn", cc.id, "via", "authorize") |
| 1721 | s.refuse(cc, cur.kind, codeControlUnavailable) |
| 1722 | s.setKilled(true) |
| 1723 | return false |
| 1724 | } |
| 1725 | |
| 1726 | validUntil := time.Unix(resp.ValidUntil, 0) |
| 1727 | if !validUntil.After(s.now()) { |
| 1728 | // A grant that has already ended is never relayed. |
| 1729 | return s.refuse(cc, cur.kind, codeAuth) |
| 1730 | } |
| 1731 | if v := watch.verdict(resp.Host, resp.ServerTime); v.cut != "" { |
| 1732 | // A revoke, a release or the end of the license was applied while the call was in flight: the answer is older than it. Nothing is served. |
| 1733 | s.log.Info("authorize_overtaken", "conn", cc.id, "kind", cur.kind, "reason", v.cut) |
| 1734 | if cur.kind != "hello" { |
| 1735 | cc.close(v.cut) |
| 1736 | return false |
| 1737 | } |
| 1738 | return s.refuse(cc, cur.kind, v.err) |
| 1739 | } |
| 1740 | ok := okBody{Node: s.cfg.NodeID, Host: resp.Host, Status: resp.Status, ValidUntil: resp.ValidUntil, ServerTime: resp.ServerTime, Policy: boundWindows(resp.Policy)} |
| 1741 | |
| 1742 | // Publish and answer under the write lock, so no OPEN can be sent before the OK line. |
| 1743 | cc.wmu.Lock() |
| 1744 | var old *ctrl |
| 1745 | recorded := false |
| 1746 | if cur.kind == "hello" { |
| 1747 | // The OK is recorded under the routing table's lock, so two HELLOs of one subject that race are recorded in the order they were routed. |
| 1748 | old, err = s.routes.activate(cc, cur.session.Subject, resp.Host, func() { |
| 1749 | recorded = true |
| 1750 | s.emit(s.now(), map[string]any{"conn": cc.id, "ev": "send", "line": ok.line()}) |
| 1751 | }) |
| 1752 | if err != nil { |
| 1753 | cc.wmu.Unlock() |
| 1754 | s.log.Warn("host_refused", "conn", cc.id, "reason", err.Error()) |
| 1755 | return s.refuse(cc, cur.kind, codeAuth) |
| 1756 | } |
| 1757 | } else if cc.hostName() != resp.Host { |
| 1758 | // The control plane now says this device owns another host: the one this connection serves is no longer its own. Fail closed. |
| 1759 | cc.wmu.Unlock() |
| 1760 | s.log.Warn("renew_host_changed", "conn", cc.id) |
| 1761 | cc.close(closeRevoked) |
| 1762 | return false |
| 1763 | } |
| 1764 | cc.setReady(resp, cur.session.JTI, validUntil) |
| 1765 | s.routes.setGated(cc, resp.Status == statusGated) // every status the control plane gives, a RENEW's included, decides what the host's streams are held to |
| 1766 | // The connection is in the routing table now, so the feed reaches it directly; what it applied before that is taken from the watch. A pause is |
| 1767 | // put back at once (setReady wrote the answer's status over it), and the control plane is asked again. |
| 1768 | late := watch.verdict(resp.Host, resp.ServerTime) |
| 1769 | if late.paused { |
| 1770 | cc.pause() |
| 1771 | } |
| 1772 | var werr error |
| 1773 | if recorded { |
| 1774 | werr = cc.sendLocked(ok.line()) |
| 1775 | } else { |
| 1776 | werr = cc.writeLocked(ok.line(), s.now()) |
| 1777 | } |
| 1778 | if werr == nil { |
| 1779 | werr = s.followUp(cc, resp) |
| 1780 | } |
| 1781 | cc.wmu.Unlock() |
| 1782 | if old != nil { |
| 1783 | old.close(closeReplaced) |
| 1784 | } |
| 1785 | if werr != nil { |
| 1786 | cc.close("write_error") |
| 1787 | return false |
| 1788 | } |
| 1789 | if late.cut != "" { |
| 1790 | // Applied between the check above and the connection joining the routing table: the OK is out, and the CLOSE follows it. |
| 1791 | cc.close(late.cut) |
| 1792 | s.routes.killStreams(resp.Host, late.cut, true) |
| 1793 | return false |
| 1794 | } |
| 1795 | if late.refresh { |
| 1796 | if late.paused { |
| 1797 | s.routes.killStreams(resp.Host, closePaused, true) |
| 1798 | } |
| 1799 | s.refreshAsync(cc) |
| 1800 | } |
| 1801 | cc.arm(validUntil) |
| 1802 | if cur.kind == "hello" { |
| 1803 | s.log.Info("control_up", "conn", cc.id, "status", resp.Status, "plan", resp.Plan, "replaced_previous", old != nil) |
| 1804 | } else { |
| 1805 | s.log.Info("control_renewed", "conn", cc.id, "status", resp.Status, "plan", resp.Plan) |
| 1806 | } |
| 1807 | return true |
| 1808 | } |
| 1809 | |
| 1810 | // followUp writes what goes right after an OK, under the connection's write lock so that no OPEN comes between: the control plane's signed access |
| 1811 | // statement, forwarded as it is (the device verifies it, the node cannot change a letter of it without breaking it), and DRAIN when the node is |
| 1812 | // already draining (a connection that was authenticating when SIGTERM came). |
| 1813 | func (s *Server) followUp(cc *ctrl, resp AuthorizeResponse) error { |
| 1814 | if line := statementLine(resp.Access); resp.Access != "" && line != "" { |
| 1815 | if err := cc.writeLocked(line, s.now()); err != nil { |
| 1816 | return err |
| 1817 | } |
| 1818 | } |
| 1819 | if s.draining.Load() { |
| 1820 | return cc.writeLocked(drainLine(time.Duration(s.drainFor.Load())), s.now()) |
| 1821 | } |
| 1822 | return nil |
| 1823 | } |
| 1824 | |
| 1825 | // The two ways authorize fails without having asked the control plane anything: the node's own queue was full, or its turn did not come in time. |
| 1826 | var ( |
| 1827 | errAuthorizeFull = &ControlError{Status: 503, Code: codeControlUnavailable} |
| 1828 | errAuthorizeWait = &ControlError{Status: 503, Code: codeControlUnavailable} |
| 1829 | ) |
| 1830 | |
| 1831 | func isLocalFailure(err error) bool { return err == errAuthorizeFull || err == errAuthorizeWait } |
| 1832 | |
| 1833 | // authorize asks the control plane for the verdict on one exchange. At most cfg.AuthorizeMax calls run at once; the exchanges waiting for a turn |
| 1834 | // are bounded by the pre-authentication limit and share the call's time limit, and beyond either the device is told control_unavailable. What the |
| 1835 | // call itself came to is recorded, so the node knows whether its control plane answers. |
| 1836 | func (s *Server) authorize(req AuthorizeRequest) (AuthorizeResponse, error) { |
| 1837 | if !tryAcquire(&s.authQ, int64(s.cfg.MaxUnauth)) { |
| 1838 | s.met.authorize.inc("unavailable") |
| 1839 | return AuthorizeResponse{}, errAuthorizeFull |
| 1840 | } |
| 1841 | defer s.authQ.Add(-1) |
| 1842 | ctx, cancel := context.WithTimeout(s.ctx, s.cfg.AuthorizeTimeout) |
| 1843 | defer cancel() |
| 1844 | select { |
| 1845 | case s.authSem <- struct{}{}: |
| 1846 | defer func() { <-s.authSem }() |
| 1847 | case <-ctx.Done(): |
| 1848 | s.met.authorize.inc("unavailable") |
| 1849 | return AuthorizeResponse{}, errAuthorizeWait |
| 1850 | } |
| 1851 | resp, err := s.cfg.Control.Authorize(ctx, req) |
| 1852 | s.health.record(s.now(), err) |
| 1853 | switch { |
| 1854 | case err == nil: |
| 1855 | s.met.authorize.inc("ok") |
| 1856 | case isRefusal(err): |
| 1857 | s.met.authorize.inc("refused") |
| 1858 | default: |
| 1859 | s.met.authorize.inc("unavailable") |
| 1860 | } |
| 1861 | return resp, err |
| 1862 | } |
| 1863 | |
| 1864 | // serveAttach hands a data connection to the stream its ATTACH names. The secret is one use: a wrong one is refused and does not consume the stream, |
| 1865 | // a used or unknown id is unknown_stream. On success there is NO reply: the next byte the node sends is the first byte of the client's stream. |
| 1866 | func (s *Server) serveAttach(tc *tls.Conn, br *bufio.Reader, m message, line, connName string, release func(control bool)) bool { |
| 1867 | id, secret := m.str("stream_id"), m.str("secret") |
| 1868 | verdict := func(result string) { |
| 1869 | s.emit(s.now(), map[string]any{"conn": connName, "ev": "attach", "line": line, "result": result}) |
| 1870 | } |
| 1871 | s.mu.Lock() |
| 1872 | p := s.pending[id] |
| 1873 | switch { |
| 1874 | case p == nil: |
| 1875 | s.mu.Unlock() |
| 1876 | verdict("err") |
| 1877 | s.log.Info("attach_reject", "reason", codeUnknownStream) |
| 1878 | writeLineThenClose(tc, errLine(codeUnknownStream)) |
| 1879 | return false |
| 1880 | case subtle.ConstantTimeCompare([]byte(p.secret), []byte(secret)) != 1: |
| 1881 | s.mu.Unlock() |
| 1882 | verdict("err") |
| 1883 | s.log.Info("attach_reject", "reason", codeAuth, "stream", id) |
| 1884 | writeLineThenClose(tc, errLine(codeAuth)) |
| 1885 | return false |
| 1886 | } |
| 1887 | delete(s.pending, id) // single use |
| 1888 | _ = tc.SetDeadline(time.Time{}) |
| 1889 | verdict("ok") |
| 1890 | p.ch <- &bufConn{Conn: tc, r: br} // buffered(1): never blocks; the stream goroutine owns tc from here |
| 1891 | s.mu.Unlock() |
| 1892 | release(false) |
| 1893 | return true |
| 1894 | } |