Merge branch '20722-docker-upgrade'
[arvados.git] / lib / crunchrun / container_gateway.go
1 // Copyright (C) The Arvados Authors. All rights reserved.
2 //
3 // SPDX-License-Identifier: AGPL-3.0
4
5 package crunchrun
6
7 import (
8         "context"
9         "crypto/hmac"
10         "crypto/rand"
11         "crypto/rsa"
12         "crypto/sha256"
13         "crypto/tls"
14         "fmt"
15         "io"
16         "net"
17         "net/http"
18         "net/url"
19         "os"
20         "os/exec"
21         "strings"
22         "sync"
23         "syscall"
24         "time"
25
26         "git.arvados.org/arvados.git/lib/controller/rpc"
27         "git.arvados.org/arvados.git/lib/selfsigned"
28         "git.arvados.org/arvados.git/lib/webdavfs"
29         "git.arvados.org/arvados.git/sdk/go/arvados"
30         "git.arvados.org/arvados.git/sdk/go/auth"
31         "git.arvados.org/arvados.git/sdk/go/ctxlog"
32         "git.arvados.org/arvados.git/sdk/go/httpserver"
33         "github.com/creack/pty"
34         "github.com/google/shlex"
35         "github.com/hashicorp/yamux"
36         "golang.org/x/crypto/ssh"
37         "golang.org/x/net/webdav"
38 )
39
40 type GatewayTarget interface {
41         // Command that will execute cmd inside the container
42         InjectCommand(ctx context.Context, detachKeys, username string, usingTTY bool, cmd []string) (*exec.Cmd, error)
43
44         // IP address inside container
45         IPAddress() (string, error)
46 }
47
48 type GatewayTargetStub struct{}
49
50 func (GatewayTargetStub) IPAddress() (string, error) { return "127.0.0.1", nil }
51 func (GatewayTargetStub) InjectCommand(ctx context.Context, detachKeys, username string, usingTTY bool, cmd []string) (*exec.Cmd, error) {
52         return exec.CommandContext(ctx, cmd[0], cmd[1:]...), nil
53 }
54
55 type Gateway struct {
56         ContainerUUID string
57         // Caller should set Address to "", or "host:0" or "host:port"
58         // where host is a known external IP address; port is a
59         // desired port number to listen on; and ":0" chooses an
60         // available dynamic port.
61         //
62         // If Address is "", Start() listens only on the loopback
63         // interface (and changes Address to "127.0.0.1:port").
64         // Otherwise it listens on all interfaces.
65         //
66         // If Address is "host:0", Start() updates Address to
67         // "host:port".
68         Address    string
69         AuthSecret string
70         Target     GatewayTarget
71         Log        interface {
72                 Printf(fmt string, args ...interface{})
73         }
74         // If non-nil, set up a ContainerGatewayTunnel, so that the
75         // controller can connect to us even if our external IP
76         // address is unknown or not routable from controller.
77         ArvadosClient *arvados.Client
78
79         // When a tunnel is connected or reconnected, this func (if
80         // not nil) will be called with the InternalURL of the
81         // controller process at the other end of the tunnel.
82         UpdateTunnelURL func(url string)
83
84         // Source for serving WebDAV requests with
85         // X-Webdav-Source: /log
86         LogCollection arvados.CollectionFileSystem
87
88         sshConfig   ssh.ServerConfig
89         requestAuth string
90         respondAuth string
91 }
92
93 // Start starts an http server that allows authenticated clients to open an
94 // interactive "docker exec" session and (in future) connect to tcp ports
95 // inside the docker container.
96 func (gw *Gateway) Start() error {
97         gw.sshConfig = ssh.ServerConfig{
98                 NoClientAuth: true,
99                 PasswordCallback: func(c ssh.ConnMetadata, pass []byte) (*ssh.Permissions, error) {
100                         if c.User() == "_" {
101                                 return nil, nil
102                         }
103                         return nil, fmt.Errorf("cannot specify user %q via ssh client", c.User())
104                 },
105                 PublicKeyCallback: func(c ssh.ConnMetadata, pubKey ssh.PublicKey) (*ssh.Permissions, error) {
106                         if c.User() == "_" {
107                                 return &ssh.Permissions{
108                                         Extensions: map[string]string{
109                                                 "pubkey-fp": ssh.FingerprintSHA256(pubKey),
110                                         },
111                                 }, nil
112                         }
113                         return nil, fmt.Errorf("cannot specify user %q via ssh client", c.User())
114                 },
115         }
116         pvt, err := rsa.GenerateKey(rand.Reader, 2048)
117         if err != nil {
118                 return err
119         }
120         err = pvt.Validate()
121         if err != nil {
122                 return err
123         }
124         signer, err := ssh.NewSignerFromKey(pvt)
125         if err != nil {
126                 return err
127         }
128         gw.sshConfig.AddHostKey(signer)
129
130         // Address (typically provided by arvados-dispatch-cloud) is
131         // HOST:PORT where HOST is our IP address or hostname as seen
132         // from arvados-controller, and PORT is either the desired
133         // port where we should run our gateway server, or "0" if we
134         // should choose an available port.
135         extAddr := gw.Address
136         // Generally we can't know which local interface corresponds
137         // to an externally reachable IP address, so if we expect to
138         // be reachable by external hosts, we listen on all
139         // interfaces.
140         listenHost := ""
141         if extAddr == "" {
142                 // If the dispatcher doesn't tell us our external IP
143                 // address, controller will only be able to connect
144                 // through the tunnel (see runTunnel), so our gateway
145                 // server only needs to listen on the loopback
146                 // interface.
147                 extAddr = "127.0.0.1:0"
148                 listenHost = "127.0.0.1"
149         }
150         extHost, extPort, err := net.SplitHostPort(extAddr)
151         if err != nil {
152                 return err
153         }
154         cert, err := selfsigned.CertGenerator{}.Generate()
155         if err != nil {
156                 return err
157         }
158         h := hmac.New(sha256.New, []byte(gw.AuthSecret))
159         h.Write(cert.Certificate[0])
160         gw.requestAuth = fmt.Sprintf("%x", h.Sum(nil))
161         h.Reset()
162         h.Write([]byte(gw.requestAuth))
163         gw.respondAuth = fmt.Sprintf("%x", h.Sum(nil))
164
165         srv := &httpserver.Server{
166                 Server: http.Server{
167                         Handler: gw,
168                         TLSConfig: &tls.Config{
169                                 Certificates: []tls.Certificate{cert},
170                         },
171                 },
172                 Addr: net.JoinHostPort(listenHost, extPort),
173         }
174         err = srv.Start()
175         if err != nil {
176                 return err
177         }
178         go func() {
179                 err := srv.Wait()
180                 gw.Log.Printf("gateway server stopped: %s", err)
181         }()
182         // Get the port number we are listening on (extPort might be
183         // "0" or a port name, in which case this will be different).
184         _, listenPort, err := net.SplitHostPort(srv.Addr)
185         if err != nil {
186                 return err
187         }
188         // When changing state to Running, the caller will want to set
189         // gateway_address to a "HOST:PORT" that, if controller
190         // connects to it, will reach this gateway server.
191         //
192         // The most likely thing to work is: HOST is our external
193         // hostname/IP as provided by the caller
194         // (arvados-dispatch-cloud) or 127.0.0.1 to indicate
195         // non-tunnel connections aren't available; and PORT is the
196         // port number we are listening on.
197         gw.Address = net.JoinHostPort(extHost, listenPort)
198         gw.Log.Printf("gateway server listening at %s", gw.Address)
199         if gw.ArvadosClient != nil {
200                 go gw.maintainTunnel(gw.Address)
201         }
202         return nil
203 }
204
205 func (gw *Gateway) maintainTunnel(addr string) {
206         for ; ; time.Sleep(5 * time.Second) {
207                 err := gw.runTunnel(addr)
208                 gw.Log.Printf("runTunnel: %s", err)
209         }
210 }
211
212 // runTunnel connects to controller and sets up a tunnel through
213 // which controller can connect to the gateway server at the given
214 // addr.
215 func (gw *Gateway) runTunnel(addr string) error {
216         ctx := auth.NewContext(context.Background(), auth.NewCredentials(gw.ArvadosClient.AuthToken))
217         arpc := rpc.NewConn("", &url.URL{Scheme: "https", Host: gw.ArvadosClient.APIHost}, gw.ArvadosClient.Insecure, rpc.PassthroughTokenProvider)
218         tun, err := arpc.ContainerGatewayTunnel(ctx, arvados.ContainerGatewayTunnelOptions{
219                 UUID:       gw.ContainerUUID,
220                 AuthSecret: gw.AuthSecret,
221         })
222         if err != nil {
223                 return fmt.Errorf("error creating gateway tunnel: %s", err)
224         }
225         mux, err := yamux.Client(tun.Conn, nil)
226         if err != nil {
227                 return fmt.Errorf("error setting up mux client end: %s", err)
228         }
229         if url := tun.Header.Get("X-Arvados-Internal-Url"); url != "" && gw.UpdateTunnelURL != nil {
230                 gw.UpdateTunnelURL(url)
231         }
232         for {
233                 muxconn, err := mux.AcceptStream()
234                 if err != nil {
235                         return err
236                 }
237                 gw.Log.Printf("tunnel connection %d started", muxconn.StreamID())
238                 go func() {
239                         defer muxconn.Close()
240                         gwconn, err := net.Dial("tcp", addr)
241                         if err != nil {
242                                 gw.Log.Printf("tunnel connection %d: error connecting to %s: %s", muxconn.StreamID(), addr, err)
243                                 return
244                         }
245                         defer gwconn.Close()
246                         var wg sync.WaitGroup
247                         wg.Add(2)
248                         go func() {
249                                 defer wg.Done()
250                                 _, err := io.Copy(gwconn, muxconn)
251                                 if err != nil {
252                                         gw.Log.Printf("tunnel connection %d: mux end: %s", muxconn.StreamID(), err)
253                                 }
254                                 gwconn.Close()
255                         }()
256                         go func() {
257                                 defer wg.Done()
258                                 _, err := io.Copy(muxconn, gwconn)
259                                 if err != nil {
260                                         gw.Log.Printf("tunnel connection %d: gateway end: %s", muxconn.StreamID(), err)
261                                 }
262                                 muxconn.Close()
263                         }()
264                         wg.Wait()
265                         gw.Log.Printf("tunnel connection %d finished", muxconn.StreamID())
266                 }()
267         }
268 }
269
270 var webdavMethod = map[string]bool{
271         "GET":      true,
272         "OPTIONS":  true,
273         "PROPFIND": true,
274 }
275
276 func (gw *Gateway) ServeHTTP(w http.ResponseWriter, req *http.Request) {
277         w.Header().Set("Vary", "X-Arvados-Authorization, X-Arvados-Container-Gateway-Uuid, X-Webdav-Prefix, X-Webdav-Source")
278         reqUUID := req.Header.Get("X-Arvados-Container-Gateway-Uuid")
279         if reqUUID == "" {
280                 // older controller versions only send UUID as query param
281                 req.ParseForm()
282                 reqUUID = req.Form.Get("uuid")
283         }
284         if reqUUID != gw.ContainerUUID {
285                 http.Error(w, fmt.Sprintf("misdirected request: meant for %q but received by crunch-run %q", reqUUID, gw.ContainerUUID), http.StatusBadGateway)
286                 return
287         }
288         if req.Header.Get("X-Arvados-Authorization") != gw.requestAuth {
289                 http.Error(w, "bad X-Arvados-Authorization header", http.StatusUnauthorized)
290                 return
291         }
292         w.Header().Set("X-Arvados-Authorization-Response", gw.respondAuth)
293         switch {
294         case req.Method == "POST" && req.Header.Get("Upgrade") == "ssh":
295                 gw.handleSSH(w, req)
296         case req.Header.Get("X-Webdav-Source") == "/log":
297                 if !webdavMethod[req.Method] {
298                         http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
299                         return
300                 }
301                 gw.handleLogsWebDAV(w, req)
302         default:
303                 http.Error(w, "path not found", http.StatusNotFound)
304         }
305 }
306
307 func (gw *Gateway) handleLogsWebDAV(w http.ResponseWriter, r *http.Request) {
308         prefix := r.Header.Get("X-Webdav-Prefix")
309         if !strings.HasPrefix(r.URL.Path, prefix) {
310                 http.Error(w, "X-Webdav-Prefix header is not a prefix of the requested path", http.StatusBadRequest)
311                 return
312         }
313         if gw.LogCollection == nil {
314                 http.Error(w, "Not found", http.StatusNotFound)
315                 return
316         }
317         wh := webdav.Handler{
318                 Prefix: prefix,
319                 FileSystem: &webdavfs.FS{
320                         FileSystem:    gw.LogCollection,
321                         Prefix:        "",
322                         Writing:       false,
323                         AlwaysReadEOF: r.Method == "PROPFIND",
324                 },
325                 LockSystem: webdavfs.NoLockSystem,
326                 Logger:     gw.webdavLogger,
327         }
328         wh.ServeHTTP(w, r)
329 }
330
331 func (gw *Gateway) webdavLogger(r *http.Request, err error) {
332         if err != nil && !os.IsNotExist(err) {
333                 ctxlog.FromContext(r.Context()).WithError(err).Info("error reported by webdav handler")
334         } else {
335                 ctxlog.FromContext(r.Context()).WithError(err).Debug("webdav request log")
336         }
337 }
338
339 // handleSSH connects to an SSH server that allows the caller to run
340 // interactive commands as root (or any other desired user) inside the
341 // container. The tunnel itself can only be created by an
342 // authenticated caller, so the SSH server itself is wide open (any
343 // password or key will be accepted).
344 //
345 // Requests must have path "/ssh" and the following headers:
346 //
347 // Connection: upgrade
348 // Upgrade: ssh
349 // X-Arvados-Target-Uuid: uuid of container
350 // X-Arvados-Authorization: must match
351 // hmac(AuthSecret,certfingerprint) (this prevents other containers
352 // and shell nodes from connecting directly)
353 //
354 // Optional headers:
355 //
356 // X-Arvados-Detach-Keys: argument to "docker exec --detach-keys",
357 // e.g., "ctrl-p,ctrl-q"
358 // X-Arvados-Login-Username: argument to "docker exec --user": account
359 // used to run command(s) inside the container.
360 func (gw *Gateway) handleSSH(w http.ResponseWriter, req *http.Request) {
361         req.ParseForm()
362         detachKeys := req.Form.Get("detach_keys")
363         username := req.Form.Get("login_username")
364         if username == "" {
365                 username = "root"
366         }
367         hj, ok := w.(http.Hijacker)
368         if !ok {
369                 http.Error(w, "ResponseWriter does not support connection upgrade", http.StatusInternalServerError)
370                 return
371         }
372         netconn, _, err := hj.Hijack()
373         if !ok {
374                 http.Error(w, err.Error(), http.StatusInternalServerError)
375                 return
376         }
377         defer netconn.Close()
378         w.Header().Set("Connection", "upgrade")
379         w.Header().Set("Upgrade", "ssh")
380         netconn.Write([]byte("HTTP/1.1 101 Switching Protocols\r\n"))
381         w.Header().Write(netconn)
382         netconn.Write([]byte("\r\n"))
383
384         ctx := req.Context()
385
386         conn, newchans, reqs, err := ssh.NewServerConn(netconn, &gw.sshConfig)
387         if err == io.EOF {
388                 return
389         } else if err != nil {
390                 gw.Log.Printf("ssh.NewServerConn: %s", err)
391                 return
392         }
393         defer conn.Close()
394         go ssh.DiscardRequests(reqs)
395         for newch := range newchans {
396                 switch newch.ChannelType() {
397                 case "direct-tcpip":
398                         go gw.handleDirectTCPIP(ctx, newch)
399                 case "session":
400                         go gw.handleSession(ctx, newch, detachKeys, username)
401                 default:
402                         go newch.Reject(ssh.UnknownChannelType, fmt.Sprintf("unsupported channel type %q", newch.ChannelType()))
403                 }
404         }
405 }
406
407 func (gw *Gateway) handleDirectTCPIP(ctx context.Context, newch ssh.NewChannel) {
408         ch, reqs, err := newch.Accept()
409         if err != nil {
410                 gw.Log.Printf("accept direct-tcpip channel: %s", err)
411                 return
412         }
413         defer ch.Close()
414         go ssh.DiscardRequests(reqs)
415
416         // RFC 4254 7.2 (copy of channelOpenDirectMsg in
417         // golang.org/x/crypto/ssh)
418         var msg struct {
419                 Raddr string
420                 Rport uint32
421                 Laddr string
422                 Lport uint32
423         }
424         err = ssh.Unmarshal(newch.ExtraData(), &msg)
425         if err != nil {
426                 fmt.Fprintf(ch.Stderr(), "unmarshal direct-tcpip extradata: %s\n", err)
427                 return
428         }
429         switch msg.Raddr {
430         case "localhost", "0.0.0.0", "127.0.0.1", "::1", "::":
431         default:
432                 fmt.Fprintf(ch.Stderr(), "cannot forward to ports on %q, only localhost\n", msg.Raddr)
433                 return
434         }
435
436         dstaddr, err := gw.Target.IPAddress()
437         if err != nil {
438                 fmt.Fprintf(ch.Stderr(), "container has no IP address: %s\n", err)
439                 return
440         } else if dstaddr == "" {
441                 fmt.Fprintf(ch.Stderr(), "container has no IP address\n")
442                 return
443         }
444
445         dst := net.JoinHostPort(dstaddr, fmt.Sprintf("%d", msg.Rport))
446         tcpconn, err := net.Dial("tcp", dst)
447         if err != nil {
448                 fmt.Fprintf(ch.Stderr(), "%s: %s\n", dst, err)
449                 return
450         }
451         go func() {
452                 n, _ := io.Copy(ch, tcpconn)
453                 ctxlog.FromContext(ctx).Debugf("tcpip: sent %d bytes\n", n)
454                 ch.CloseWrite()
455         }()
456         n, _ := io.Copy(tcpconn, ch)
457         ctxlog.FromContext(ctx).Debugf("tcpip: received %d bytes\n", n)
458 }
459
460 func (gw *Gateway) handleSession(ctx context.Context, newch ssh.NewChannel, detachKeys, username string) {
461         ch, reqs, err := newch.Accept()
462         if err != nil {
463                 gw.Log.Printf("error accepting session channel: %s", err)
464                 return
465         }
466         defer ch.Close()
467
468         var pty0, tty0 *os.File
469         // Where to send errors/messages for the client to see
470         logw := io.Writer(ch.Stderr())
471         // How to end lines when sending errors/messages to the client
472         // (changes to \r\n when using a pty)
473         eol := "\n"
474         // Env vars to add to child process
475         termEnv := []string(nil)
476
477         started := 0
478         wantClose := make(chan struct{})
479         for {
480                 var req *ssh.Request
481                 select {
482                 case r, ok := <-reqs:
483                         if !ok {
484                                 return
485                         }
486                         req = r
487                 case <-wantClose:
488                         return
489                 }
490                 ok := false
491                 switch req.Type {
492                 case "shell", "exec":
493                         if started++; started != 1 {
494                                 // RFC 4254 6.5: "Only one of these
495                                 // requests can succeed per channel."
496                                 break
497                         }
498                         ok = true
499                         var payload struct {
500                                 Command string
501                         }
502                         ssh.Unmarshal(req.Payload, &payload)
503                         execargs, err := shlex.Split(payload.Command)
504                         if err != nil {
505                                 fmt.Fprintf(logw, "error parsing supplied command: %s"+eol, err)
506                                 return
507                         }
508                         if len(execargs) == 0 {
509                                 execargs = []string{"/bin/bash", "-login"}
510                         }
511                         go func() {
512                                 var resp struct {
513                                         Status uint32
514                                 }
515                                 defer func() {
516                                         ch.SendRequest("exit-status", false, ssh.Marshal(&resp))
517                                         close(wantClose)
518                                 }()
519
520                                 cmd, err := gw.Target.InjectCommand(ctx, detachKeys, username, tty0 != nil, execargs)
521                                 if err != nil {
522                                         fmt.Fprintln(ch.Stderr(), err)
523                                         ch.CloseWrite()
524                                         resp.Status = 1
525                                         return
526                                 }
527                                 if tty0 != nil {
528                                         cmd.Stdin = tty0
529                                         cmd.Stdout = tty0
530                                         cmd.Stderr = tty0
531                                         go io.Copy(ch, pty0)
532                                         go io.Copy(pty0, ch)
533                                         // Send our own debug messages to tty as well.
534                                         logw = tty0
535                                 } else {
536                                         // StdinPipe may seem
537                                         // superfluous here, but it's
538                                         // not: it causes cmd.Run() to
539                                         // return when the subprocess
540                                         // exits. Without it, Run()
541                                         // waits for stdin to close,
542                                         // which causes "ssh ... echo
543                                         // ok" (with the client's
544                                         // stdin connected to a
545                                         // terminal or something) to
546                                         // hang.
547                                         stdin, err := cmd.StdinPipe()
548                                         if err != nil {
549                                                 fmt.Fprintln(ch.Stderr(), err)
550                                                 ch.CloseWrite()
551                                                 resp.Status = 1
552                                                 return
553                                         }
554                                         go func() {
555                                                 io.Copy(stdin, ch)
556                                                 stdin.Close()
557                                         }()
558                                         cmd.Stdout = ch
559                                         cmd.Stderr = ch.Stderr()
560                                 }
561                                 cmd.SysProcAttr = &syscall.SysProcAttr{
562                                         Setctty: tty0 != nil,
563                                         Setsid:  true,
564                                 }
565                                 cmd.Env = append(os.Environ(), termEnv...)
566                                 err = cmd.Run()
567                                 if exiterr, ok := err.(*exec.ExitError); ok {
568                                         if status, ok := exiterr.Sys().(syscall.WaitStatus); ok {
569                                                 resp.Status = uint32(status.ExitStatus())
570                                         }
571                                 } else if err != nil {
572                                         // Propagate errors like `exec: "docker": executable file not found in $PATH`
573                                         fmt.Fprintln(ch.Stderr(), err)
574                                 }
575                                 errClose := ch.CloseWrite()
576                                 if resp.Status == 0 && (err != nil || errClose != nil) {
577                                         resp.Status = 1
578                                 }
579                         }()
580                 case "pty-req":
581                         eol = "\r\n"
582                         p, t, err := pty.Open()
583                         if err != nil {
584                                 fmt.Fprintf(ch.Stderr(), "pty failed: %s"+eol, err)
585                                 break
586                         }
587                         defer p.Close()
588                         defer t.Close()
589                         pty0, tty0 = p, t
590                         ok = true
591                         var payload struct {
592                                 Term string
593                                 Cols uint32
594                                 Rows uint32
595                                 X    uint32
596                                 Y    uint32
597                         }
598                         ssh.Unmarshal(req.Payload, &payload)
599                         termEnv = []string{"TERM=" + payload.Term, "USE_TTY=1"}
600                         err = pty.Setsize(pty0, &pty.Winsize{Rows: uint16(payload.Rows), Cols: uint16(payload.Cols), X: uint16(payload.X), Y: uint16(payload.Y)})
601                         if err != nil {
602                                 fmt.Fprintf(logw, "pty-req: setsize failed: %s"+eol, err)
603                         }
604                 case "window-change":
605                         var payload struct {
606                                 Cols uint32
607                                 Rows uint32
608                                 X    uint32
609                                 Y    uint32
610                         }
611                         ssh.Unmarshal(req.Payload, &payload)
612                         err := pty.Setsize(pty0, &pty.Winsize{Rows: uint16(payload.Rows), Cols: uint16(payload.Cols), X: uint16(payload.X), Y: uint16(payload.Y)})
613                         if err != nil {
614                                 fmt.Fprintf(logw, "window-change: setsize failed: %s"+eol, err)
615                                 break
616                         }
617                         ok = true
618                 case "env":
619                         // TODO: implement "env"
620                         // requests by setting env
621                         // vars in the docker-exec
622                         // command (not docker-exec's
623                         // own environment, which
624                         // would be a gaping security
625                         // hole).
626                 default:
627                         // fmt.Fprintf(logw, "declined request %q on ssh channel"+eol, req.Type)
628                 }
629                 if req.WantReply {
630                         req.Reply(ok, nil)
631                 }
632         }
633 }