1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
5 // Package service provides a cmd.Handler that brings up a system service.
22 "git.arvados.org/arvados.git/lib/cmd"
23 "git.arvados.org/arvados.git/lib/config"
24 "git.arvados.org/arvados.git/sdk/go/arvados"
25 "git.arvados.org/arvados.git/sdk/go/ctxlog"
26 "git.arvados.org/arvados.git/sdk/go/health"
27 "git.arvados.org/arvados.git/sdk/go/httpserver"
28 "github.com/coreos/go-systemd/daemon"
29 "github.com/julienschmidt/httprouter"
30 "github.com/prometheus/client_golang/prometheus"
31 "github.com/sirupsen/logrus"
34 type Handler interface {
37 // Done returns a channel that closes when the handler shuts
38 // itself down, or nil if this never happens.
39 Done() <-chan struct{}
42 type NewHandlerFunc func(_ context.Context, _ *arvados.Cluster, token string, registry *prometheus.Registry) Handler
45 newHandler NewHandlerFunc
46 svcName arvados.ServiceName
47 ctx context.Context // enables tests to shutdown service; no public API yet
50 var requestQueueDumpCheckInterval = time.Minute
52 // Command returns a cmd.Handler that loads site config, calls
53 // newHandler with the current cluster and node configs, and brings up
54 // an http server with the returned handler.
56 // The handler is wrapped with server middleware (adding X-Request-ID
57 // headers, logging requests/responses, etc).
58 func Command(svcName arvados.ServiceName, newHandler NewHandlerFunc) cmd.Handler {
60 newHandler: newHandler,
62 ctx: context.Background(),
66 func (c *command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
67 log := ctxlog.New(stderr, "json", "info")
72 log.WithError(err).Error("exiting")
76 flags := flag.NewFlagSet("", flag.ContinueOnError)
77 flags.SetOutput(stderr)
79 loader := config.NewLoader(stdin, log)
80 loader.SetupFlags(flags)
82 // prog is [keepstore, keep-web, git-httpd, ...] but the
83 // legacy config flags are [-legacy-keepstore-config,
84 // -legacy-keepweb-config, -legacy-git-httpd-config, ...]
85 legacyFlag := "-legacy-" + strings.Replace(prog, "keep-", "keep", 1) + "-config"
86 args = loader.MungeLegacyConfigArgs(log, args, legacyFlag)
88 versionFlag := flags.Bool("version", false, "Write version information to stdout and exit 0")
89 pprofAddr := flags.String("pprof", "", "Serve Go profile data at `[addr]:port`")
90 if ok, code := cmd.ParseFlags(flags, prog, args, "", stderr); !ok {
92 } else if *versionFlag {
93 return cmd.Version.RunCommand(prog, args, stdin, stdout, stderr)
98 log.Println(http.ListenAndServe(*pprofAddr, nil))
102 if strings.HasSuffix(prog, "controller") {
103 // Some config-loader checks try to make API calls via
104 // controller. Those can't be expected to work if this
105 // process _is_ the controller: we haven't started an
107 loader.SkipAPICalls = true
110 cfg, err := loader.Load()
114 cluster, err := cfg.GetCluster("")
119 // Now that we've read the config, replace the bootstrap
120 // logger with a new one according to the logging config.
121 log = ctxlog.New(stderr, cluster.SystemLogs.Format, cluster.SystemLogs.LogLevel)
122 logger := log.WithFields(logrus.Fields{
124 "ClusterID": cluster.ClusterID,
126 ctx := ctxlog.Context(c.ctx, logger)
128 listenURL, internalURL, err := getListenAddr(cluster.Services, c.svcName, log)
132 ctx = context.WithValue(ctx, contextKeyURL{}, internalURL)
134 reg := prometheus.NewRegistry()
135 loader.RegisterMetrics(reg)
137 // arvados_version_running{version="1.2.3~4"} 1.0
138 mVersion := prometheus.NewGaugeVec(prometheus.GaugeOpts{
139 Namespace: "arvados",
140 Name: "version_running",
141 Help: "Indicated version is running.",
142 }, []string{"version"})
143 mVersion.WithLabelValues(cmd.Version.String()).Set(1)
144 reg.MustRegister(mVersion)
146 handler := c.newHandler(ctx, cluster, cluster.SystemRootToken, reg)
147 if err = handler.CheckHealth(); err != nil {
151 instrumented := httpserver.Instrument(reg, log,
152 httpserver.HandlerWithDeadline(cluster.API.RequestTimeout.Duration(),
153 httpserver.AddRequestIDs(
154 httpserver.Inspect(reg, cluster.ManagementToken,
155 httpserver.LogRequests(
156 interceptHealthReqs(cluster.ManagementToken, handler.CheckHealth,
157 httpserver.NewRequestLimiter(cluster.API.MaxConcurrentRequests, handler, reg)))))))
158 srv := &httpserver.Server{
160 Handler: ifCollectionInHost(instrumented, instrumented.ServeAPI(cluster.ManagementToken, instrumented)),
161 BaseContext: func(net.Listener) context.Context { return ctx },
163 Addr: listenURL.Host,
165 if listenURL.Scheme == "https" || listenURL.Scheme == "wss" {
166 tlsconfig, err := makeTLSConfig(cluster, logger)
168 logger.WithError(err).Errorf("cannot start %s service on %s", c.svcName, listenURL.String())
171 srv.TLSConfig = tlsconfig
177 logger.WithFields(logrus.Fields{
180 "Service": c.svcName,
181 "Version": cmd.Version.String(),
183 if _, err := daemon.SdNotify(false, "READY=1"); err != nil {
184 logger.WithError(err).Errorf("error notifying init daemon")
187 // Shut down server if caller cancels context
192 // Shut down server if handler dies
196 go c.requestQueueDumpCheck(cluster, prog, reg, &srv.Server, logger)
204 // If SystemLogs.RequestQueueDumpDirectory is set, monitor the
205 // server's incoming HTTP request queue size. When it exceeds 90% of
206 // API.MaxConcurrentRequests, write the /_inspect/requests data to a
207 // JSON file in the specified directory.
208 func (c *command) requestQueueDumpCheck(cluster *arvados.Cluster, prog string, reg *prometheus.Registry, srv *http.Server, logger logrus.FieldLogger) {
209 outdir := cluster.SystemLogs.RequestQueueDumpDirectory
210 if outdir == "" || cluster.ManagementToken == "" {
213 logger = logger.WithField("worker", "RequestQueueDump")
214 outfile := outdir + "/" + prog + "-requests.json"
215 for range time.NewTicker(requestQueueDumpCheckInterval).C {
216 mfs, err := reg.Gather()
218 logger.WithError(err).Warn("error getting metrics")
222 for _, mf := range mfs {
223 if mf.Name != nil && *mf.Name == "arvados_concurrent_requests" && len(mf.Metric) == 1 {
224 n := int(mf.Metric[0].GetGauge().GetValue())
225 if n > 0 && n >= cluster.API.MaxConcurrentRequests*9/10 {
232 req, err := http.NewRequest("GET", "/_inspect/requests", nil)
234 logger.WithError(err).Warn("error in http.NewRequest")
237 req.Header.Set("Authorization", "Bearer "+cluster.ManagementToken)
238 resp := httptest.NewRecorder()
239 srv.Handler.ServeHTTP(resp, req)
240 if code := resp.Result().StatusCode; code != http.StatusOK {
241 logger.WithField("StatusCode", code).Warn("error getting /_inspect/requests")
244 err = os.WriteFile(outfile, resp.Body.Bytes(), 0777)
246 logger.WithError(err).Warn("error writing file")
253 // If an incoming request's target vhost has an embedded collection
254 // UUID or PDH, handle it with hTrue, otherwise handle it with
257 // Facilitates routing "http://collections.example/metrics" to metrics
258 // and "http://{uuid}.collections.example/metrics" to a file in a
260 func ifCollectionInHost(hTrue, hFalse http.Handler) http.Handler {
261 return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
262 if arvados.CollectionIDFromDNSName(r.Host) != "" {
263 hTrue.ServeHTTP(w, r)
265 hFalse.ServeHTTP(w, r)
270 func interceptHealthReqs(mgtToken string, checkHealth func() error, next http.Handler) http.Handler {
271 mux := httprouter.New()
272 mux.Handler("GET", "/_health/ping", &health.Handler{
275 Routes: health.Routes{"ping": checkHealth},
278 return ifCollectionInHost(next, mux)
281 // Determine listenURL (addr:port where server should bind) and
282 // internalURL (target url that client should connect to) for a
285 // If the config does not specify ListenURL, we check all of the
286 // configured InternalURLs. If there is exactly one that matches our
287 // hostname, or exactly one that matches a local interface address,
288 // then we use that as listenURL.
290 // Note that listenURL and internalURL may use different protocols
291 // (e.g., listenURL is http, but the service sits behind a proxy, so
292 // clients connect using https).
293 func getListenAddr(svcs arvados.Services, prog arvados.ServiceName, log logrus.FieldLogger) (arvados.URL, arvados.URL, error) {
294 svc, ok := svcs.Map()[prog]
296 return arvados.URL{}, arvados.URL{}, fmt.Errorf("unknown service name %q", prog)
299 if want := os.Getenv("ARVADOS_SERVICE_INTERNAL_URL"); want != "" {
300 url, err := url.Parse(want)
302 return arvados.URL{}, arvados.URL{}, fmt.Errorf("$ARVADOS_SERVICE_INTERNAL_URL (%q): %s", want, err)
307 for internalURL, conf := range svc.InternalURLs {
308 if internalURL.String() == url.String() {
309 listenURL := conf.ListenURL
310 if listenURL.Host == "" {
311 listenURL = internalURL
313 return listenURL, internalURL, nil
316 log.Warnf("possible configuration error: listening on %s (from $ARVADOS_SERVICE_INTERNAL_URL) even though configuration does not have a matching InternalURLs entry", url)
317 internalURL := arvados.URL(*url)
318 return internalURL, internalURL, nil
322 for internalURL, conf := range svc.InternalURLs {
323 listenURL := conf.ListenURL
324 if listenURL.Host == "" {
325 // If ListenURL is not specified, assume
326 // InternalURL is also usable as the listening
327 // proto/addr/port (i.e., simple case with no
328 // intermediate proxy/routing)
329 listenURL = internalURL
331 listenAddr := listenURL.Host
332 if _, _, err := net.SplitHostPort(listenAddr); err != nil {
333 // url "https://foo.example/" (with no
334 // explicit port name/number) means listen on
335 // the well-known port for the specified
336 // protocol, "foo.example:https".
337 port := listenURL.Scheme
338 if port == "ws" || port == "wss" {
339 port = "http" + port[2:]
341 listenAddr = net.JoinHostPort(listenAddr, port)
343 listener, err := net.Listen("tcp", listenAddr)
346 return listenURL, internalURL, nil
347 } else if strings.Contains(err.Error(), "cannot assign requested address") {
348 // If 'Host' specifies a different server than
349 // the current one, it'll resolve the hostname
350 // to IP address, and then fail because it
351 // can't bind an IP address it doesn't own.
354 errors = append(errors, fmt.Sprintf("%s: %s", listenURL, err))
358 return arvados.URL{}, arvados.URL{}, fmt.Errorf("could not enable the %q service on this host: %s", prog, strings.Join(errors, "; "))
360 return arvados.URL{}, arvados.URL{}, fmt.Errorf("configuration does not enable the %q service on this host", prog)
363 type contextKeyURL struct{}
365 func URLFromContext(ctx context.Context) (arvados.URL, bool) {
366 u, ok := ctx.Value(contextKeyURL{}).(arvados.URL)