lnet/lnet/acceptor.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #define DEBUG_SUBSYSTEM S_LNET
  33
  34 #include <linux/completion.h>
  35 #include <net/sock.h>
  36 #include <lnet/lib-lnet.h>
  37 #include <linux/sunrpc/addr.h>
  38
  39 static int   accept_port    = 988;
  40 static int   accept_backlog = 127;
  41 static int   accept_timeout = 5;
  42
  43 static struct {
  44         int                     pta_shutdown;
  45         struct socket           *pta_sock;
  46         struct completion       pta_signal;
  47         struct net              *pta_ns;
  48         wait_queue_head_t       pta_waitq;
  49         atomic_t                pta_ready;
  50 #ifdef HAVE_SK_DATA_READY_ONE_ARG
  51         void                    (*pta_odata)(struct sock *);
  52 #else
  53         void                    (*pta_odata)(struct sock *, int);
  54 #endif
  55 } lnet_acceptor_state = {
  56         .pta_shutdown = 1
  57 };
  58
  59 int
  60 lnet_acceptor_port(void)
  61 {
  62         return accept_port;
  63 }
  64
  65 static inline int
  66 lnet_accept_magic(__u32 magic, __u32 constant)
  67 {
  68         return (magic == constant ||
  69                 magic == __swab32(constant));
  70 }
  71
  72 EXPORT_SYMBOL(lnet_acceptor_port);
  73
  74 static char *accept_type = "secure";
  75
  76 module_param_named(accept, accept_type, charp, 0444);
  77 MODULE_PARM_DESC(accept, "Accept connections (secure|all|none)");
  78 module_param(accept_port, int, 0444);
  79 MODULE_PARM_DESC(accept_port, "Acceptor's port (same on all nodes)");
  80 module_param(accept_backlog, int, 0444);
  81 MODULE_PARM_DESC(accept_backlog, "Acceptor's listen backlog");
  82 module_param(accept_timeout, int, 0644);
  83 MODULE_PARM_DESC(accept_timeout, "Acceptor's timeout (seconds)");
  84
  85 int
  86 lnet_acceptor_timeout(void)
  87 {
  88         return accept_timeout;
  89 }
  90 EXPORT_SYMBOL(lnet_acceptor_timeout);
  91
  92 void
  93 lnet_connect_console_error (int rc, lnet_nid_t peer_nid,
  94                             struct sockaddr *sa)
  95 {
  96         switch (rc) {
  97         /* "normal" errors */
  98         case -ECONNREFUSED:
  99                 CNETERR("Connection to %s at host %pISp was refused: check that Lustre is running on that node.\n",
 100                         libcfs_nid2str(peer_nid), sa);
 101                 break;
 102         case -EHOSTUNREACH:
 103         case -ENETUNREACH:
 104                 CNETERR("Connection to %s at host %pIS was unreachable: the network or that node may be down, or Lustre may be misconfigured.\n",
 105                         libcfs_nid2str(peer_nid), sa);
 106                 break;
 107         case -ETIMEDOUT:
 108                 CNETERR("Connection to %s at host %pISp took too long: that node may be hung or experiencing high load.\n",
 109                         libcfs_nid2str(peer_nid), sa);
 110                 break;
 111         case -ECONNRESET:
 112                 LCONSOLE_ERROR_MSG(0x11b,
 113                                    "Connection to %s at host %pISp was reset: is it running a compatible version of Lustre and is %s one of its NIDs?\n",
 114                                    libcfs_nid2str(peer_nid), sa,
 115                                    libcfs_nid2str(peer_nid));
 116                 break;
 117         case -EPROTO:
 118                 LCONSOLE_ERROR_MSG(0x11c,
 119                                    "Protocol error connecting to %s at host %pISp: is it running a compatible version of Lustre?\n",
 120                                    libcfs_nid2str(peer_nid), sa);
 121                 break;
 122         case -EADDRINUSE:
 123                 LCONSOLE_ERROR_MSG(0x11d,
 124                                    "No privileged ports available to connect to %s at host %pISp\n",
 125                                    libcfs_nid2str(peer_nid), sa);
 126                 break;
 127         default:
 128                 LCONSOLE_ERROR_MSG(0x11e,
 129                                    "Unexpected error %d connecting to %s at host %pISp\n",
 130                                    rc, libcfs_nid2str(peer_nid), sa);
 131                 break;
 132         }
 133 }
 134 EXPORT_SYMBOL(lnet_connect_console_error);
 135
 136 struct socket *
 137 lnet_connect(lnet_nid_t peer_nid, int interface, struct sockaddr *peeraddr,
 138              struct net *ns)
 139 {
 140         struct lnet_acceptor_connreq cr;
 141         struct socket           *sock;
 142         int                     rc;
 143         int                     port;
 144
 145         BUILD_BUG_ON(sizeof(cr) > 16); /* not too big to be on the stack */
 146
 147         LASSERT(peeraddr->sa_family == AF_INET ||
 148                 peeraddr->sa_family == AF_INET6);
 149
 150         for (port = LNET_ACCEPTOR_MAX_RESERVED_PORT;
 151              port >= LNET_ACCEPTOR_MIN_RESERVED_PORT;
 152              --port) {
 153                 /* Iterate through reserved ports. */
 154                 sock = lnet_sock_connect(interface, port, peeraddr, ns);
 155                 if (IS_ERR(sock)) {
 156                         rc = PTR_ERR(sock);
 157                         if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL)
 158                                 continue;
 159                         goto failed;
 160                 }
 161
 162                 BUILD_BUG_ON(LNET_PROTO_ACCEPTOR_VERSION != 1);
 163
 164                 cr.acr_magic   = LNET_PROTO_ACCEPTOR_MAGIC;
 165                 cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 166                 cr.acr_nid     = peer_nid;
 167
 168                 if (the_lnet.ln_testprotocompat) {
 169                         /* single-shot proto check */
 170                         if (test_and_clear_bit(2, &the_lnet.ln_testprotocompat))
 171                                 cr.acr_version++;
 172                         if (test_and_clear_bit(3, &the_lnet.ln_testprotocompat))
 173                                 cr.acr_magic = LNET_PROTO_MAGIC;
 174                 }
 175
 176                 rc = lnet_sock_write(sock, &cr, sizeof(cr),
 177                                        accept_timeout);
 178                 if (rc != 0)
 179                         goto failed_sock;
 180
 181                 return sock;
 182         }
 183
 184         rc = -EADDRINUSE;
 185         goto failed;
 186
 187 failed_sock:
 188         sock_release(sock);
 189 failed:
 190         lnet_connect_console_error(rc, peer_nid, peeraddr);
 191         return ERR_PTR(rc);
 192 }
 193 EXPORT_SYMBOL(lnet_connect);
 194
 195 static int
 196 lnet_accept(struct socket *sock, __u32 magic)
 197 {
 198         struct lnet_acceptor_connreq cr;
 199         struct sockaddr_storage peer;
 200         int                     rc;
 201         int                     flip;
 202         struct lnet_ni *ni;
 203         char                   *str;
 204
 205         LASSERT(sizeof(cr) <= 16);              /* not too big for the stack */
 206
 207         rc = lnet_sock_getaddr(sock, true, &peer);
 208         if (rc != 0) {
 209                 CERROR("Can't determine new connection's address\n");
 210                 return rc;
 211         }
 212
 213         if (!lnet_accept_magic(magic, LNET_PROTO_ACCEPTOR_MAGIC)) {
 214
 215                 if (lnet_accept_magic(magic, LNET_PROTO_MAGIC)) {
 216                         /* future version compatibility!
 217                          * When LNET unifies protocols over all LNDs, the first
 218                          * thing sent will be a version query.  I send back
 219                          * LNET_PROTO_ACCEPTOR_MAGIC to tell her I'm "old" */
 220
 221                         memset(&cr, 0, sizeof(cr));
 222                         cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
 223                         cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 224                         rc = lnet_sock_write(sock, &cr, sizeof(cr),
 225                                                accept_timeout);
 226
 227                         if (rc != 0)
 228                                 CERROR("Error sending magic+version in response to LNET magic from %pIS: %d\n",
 229                                        &peer, rc);
 230                         return -EPROTO;
 231                 }
 232
 233                 if (lnet_accept_magic(magic, LNET_PROTO_TCP_MAGIC))
 234                         str = "'old' socknal/tcpnal";
 235                 else
 236                         str = "unrecognised";
 237
 238                 LCONSOLE_ERROR_MSG(0x11f, "Refusing connection from %pIS"
 239                                    " magic %08x: %s acceptor protocol\n",
 240                                    &peer, magic, str);
 241                 return -EPROTO;
 242         }
 243
 244         flip = (magic != LNET_PROTO_ACCEPTOR_MAGIC);
 245
 246         rc = lnet_sock_read(sock, &cr.acr_version,
 247                               sizeof(cr.acr_version),
 248                               accept_timeout);
 249         if (rc != 0) {
 250                 CERROR("Error %d reading connection request version from %pIS\n",
 251                        rc, &peer);
 252                 return -EIO;
 253         }
 254
 255         if (flip)
 256                 __swab32s(&cr.acr_version);
 257
 258         if (cr.acr_version != LNET_PROTO_ACCEPTOR_VERSION) {
 259                 /* future version compatibility!
 260                  * An acceptor-specific protocol rev will first send a version
 261                  * query.  I send back my current version to tell her I'm
 262                  * "old". */
 263                 int peer_version = cr.acr_version;
 264
 265                 memset(&cr, 0, sizeof(cr));
 266                 cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
 267                 cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 268
 269                 rc = lnet_sock_write(sock, &cr, sizeof(cr),
 270                                        accept_timeout);
 271
 272                 if (rc != 0)
 273                         CERROR("Error sending magic+version in response to version %d from %pIS: %d\n",
 274                                peer_version, &peer, rc);
 275                 return -EPROTO;
 276         }
 277
 278         rc = lnet_sock_read(sock, &cr.acr_nid,
 279                               sizeof(cr) -
 280                               offsetof(struct lnet_acceptor_connreq, acr_nid),
 281                               accept_timeout);
 282         if (rc != 0) {
 283                 CERROR("Error %d reading connection request from %pIS\n",
 284                        rc, &peer);
 285                 return -EIO;
 286         }
 287
 288         if (flip)
 289                 __swab64s(&cr.acr_nid);
 290
 291         ni = lnet_nid2ni_addref(cr.acr_nid);
 292         if (ni == NULL ||               /* no matching net */
 293             ni->ni_nid != cr.acr_nid) { /* right NET, wrong NID! */
 294                 if (ni != NULL)
 295                         lnet_ni_decref(ni);
 296                 LCONSOLE_ERROR_MSG(0x120,
 297                                    "Refusing connection from %pIS for %s: No matching NI\n",
 298                                    &peer, libcfs_nid2str(cr.acr_nid));
 299                 return -EPERM;
 300         }
 301
 302         if (ni->ni_net->net_lnd->lnd_accept == NULL) {
 303                 /* This catches a request for the loopback LND */
 304                 lnet_ni_decref(ni);
 305                 LCONSOLE_ERROR_MSG(0x121,
 306                                    "Refusing connection from %pIS for %s: NI doesn not accept IP connections\n",
 307                                   &peer, libcfs_nid2str(cr.acr_nid));
 308                 return -EPERM;
 309         }
 310
 311         CDEBUG(D_NET, "Accept %s from %pI4h\n",
 312                libcfs_nid2str(cr.acr_nid), &peer);
 313
 314         rc = ni->ni_net->net_lnd->lnd_accept(ni, sock);
 315
 316         lnet_ni_decref(ni);
 317         return rc;
 318 }
 319
 320 #ifdef HAVE_SK_DATA_READY_ONE_ARG
 321 static void lnet_acceptor_ready(struct sock *sk)
 322 #else
 323 static void lnet_acceptor_ready(struct sock *sk, int len)
 324 #endif
 325 {
 326         /* Ensure pta_odata has actually been set before calling it */
 327         rmb();
 328 #ifdef HAVE_SK_DATA_READY_ONE_ARG
 329         lnet_acceptor_state.pta_odata(sk);
 330 #else
 331         lnet_acceptor_state.pta_odata(sk, 0);
 332 #endif
 333
 334         atomic_set(&lnet_acceptor_state.pta_ready, 1);
 335         wake_up(&lnet_acceptor_state.pta_waitq);
 336 }
 337
 338 static int
 339 lnet_acceptor(void *arg)
 340 {
 341         struct socket  *newsock;
 342         int            rc;
 343         __u32          magic;
 344         struct sockaddr_storage peer;
 345         int            secure = (int)((uintptr_t)arg);
 346
 347         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 348
 349         lnet_acceptor_state.pta_sock =
 350                 lnet_sock_listen(accept_port, accept_backlog,
 351                                  lnet_acceptor_state.pta_ns);
 352         if (IS_ERR(lnet_acceptor_state.pta_sock)) {
 353                 rc = PTR_ERR(lnet_acceptor_state.pta_sock);
 354                 if (rc == -EADDRINUSE)
 355                         LCONSOLE_ERROR_MSG(0x122, "Can't start acceptor on port"
 356                                            " %d: port already in use\n",
 357                                            accept_port);
 358                 else
 359                         LCONSOLE_ERROR_MSG(0x123, "Can't start acceptor on port "
 360                                            "%d: unexpected error %d\n",
 361                                            accept_port, rc);
 362
 363                 lnet_acceptor_state.pta_sock = NULL;
 364         } else {
 365                 rc = 0;
 366                 LCONSOLE(0, "Accept %s, port %d\n", accept_type, accept_port);
 367                 init_waitqueue_head(&lnet_acceptor_state.pta_waitq);
 368                 lnet_acceptor_state.pta_odata =
 369                         lnet_acceptor_state.pta_sock->sk->sk_data_ready;
 370                 /* ensure pta_odata gets set before there is any chance of
 371                  * lnet_accept_ready() trying to read it.
 372                  */
 373                 wmb();
 374                 lnet_acceptor_state.pta_sock->sk->sk_data_ready =
 375                         lnet_acceptor_ready;
 376                 atomic_set(&lnet_acceptor_state.pta_ready, 1);
 377         }
 378
 379         /* set init status and unblock parent */
 380         lnet_acceptor_state.pta_shutdown = rc;
 381         complete(&lnet_acceptor_state.pta_signal);
 382
 383         if (rc != 0)
 384                 return rc;
 385
 386         while (!lnet_acceptor_state.pta_shutdown) {
 387
 388                 wait_event_idle(lnet_acceptor_state.pta_waitq,
 389                                 lnet_acceptor_state.pta_shutdown ||
 390                                 atomic_read(&lnet_acceptor_state.pta_ready));
 391                 if (!atomic_read(&lnet_acceptor_state.pta_ready))
 392                         continue;
 393                 atomic_set(&lnet_acceptor_state.pta_ready, 0);
 394                 rc = kernel_accept(lnet_acceptor_state.pta_sock, &newsock,
 395                                    SOCK_NONBLOCK);
 396                 if (rc != 0) {
 397                         if (rc != -EAGAIN) {
 398                                 CWARN("Accept error %d: pausing...\n", rc);
 399                                 schedule_timeout_uninterruptible(
 400                                         cfs_time_seconds(1));
 401                         }
 402                         continue;
 403                 }
 404
 405                 /* make sure we call lnet_sock_accept() again, until it fails */
 406                 atomic_set(&lnet_acceptor_state.pta_ready, 1);
 407
 408                 rc = lnet_sock_getaddr(newsock, true, &peer);
 409                 if (rc != 0) {
 410                         CERROR("Can't determine new connection's address\n");
 411                         goto failed;
 412                 }
 413
 414                 if (secure &&
 415                     rpc_get_port((struct sockaddr *)&peer) >
 416                     LNET_ACCEPTOR_MAX_RESERVED_PORT) {
 417                         CERROR("Refusing connection from %pISp: insecure port.\n",
 418                                &peer);
 419                         goto failed;
 420                 }
 421
 422                 rc = lnet_sock_read(newsock, &magic, sizeof(magic),
 423                                       accept_timeout);
 424                 if (rc != 0) {
 425                         CERROR("Error %d reading connection request from %pIS\n",
 426                                rc, &peer);
 427                         goto failed;
 428                 }
 429
 430                 rc = lnet_accept(newsock, magic);
 431                 if (rc != 0)
 432                         goto failed;
 433
 434                 continue;
 435
 436 failed:
 437                 sock_release(newsock);
 438         }
 439
 440         lnet_acceptor_state.pta_sock->sk->sk_data_ready =
 441                 lnet_acceptor_state.pta_odata;
 442         sock_release(lnet_acceptor_state.pta_sock);
 443         lnet_acceptor_state.pta_sock = NULL;
 444
 445         CDEBUG(D_NET, "Acceptor stopping\n");
 446
 447         /* unblock lnet_acceptor_stop() */
 448         complete(&lnet_acceptor_state.pta_signal);
 449         return 0;
 450 }
 451
 452 static inline int
 453 accept2secure(const char *acc, long *sec)
 454 {
 455         if (!strcmp(acc, "secure")) {
 456                 *sec = 1;
 457                 return 1;
 458         } else if (!strcmp(acc, "all")) {
 459                 *sec = 0;
 460                 return 1;
 461         } else if (!strcmp(acc, "none")) {
 462                 return 0;
 463         } else {
 464                 LCONSOLE_ERROR_MSG(0x124, "Can't parse 'accept=\"%s\"'\n",
 465                                    acc);
 466                 return -EINVAL;
 467         }
 468 }
 469
 470 int
 471 lnet_acceptor_start(void)
 472 {
 473         struct task_struct *task;
 474         int  rc;
 475         long rc2;
 476         long secure;
 477
 478         /* if acceptor is already running return immediately */
 479         if (!lnet_acceptor_state.pta_shutdown)
 480                 return 0;
 481
 482         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 483
 484         init_completion(&lnet_acceptor_state.pta_signal);
 485         rc = accept2secure(accept_type, &secure);
 486         if (rc <= 0)
 487                 return rc;
 488
 489         if (lnet_count_acceptor_nets() == 0)  /* not required */
 490                 return 0;
 491         if (current->nsproxy && current->nsproxy->net_ns)
 492                 lnet_acceptor_state.pta_ns = current->nsproxy->net_ns;
 493         else
 494                 lnet_acceptor_state.pta_ns = &init_net;
 495         task = kthread_run(lnet_acceptor, (void *)(uintptr_t)secure,
 496                            "acceptor_%03ld", secure);
 497         if (IS_ERR(task)) {
 498                 rc2 = PTR_ERR(task);
 499                 CERROR("Can't start acceptor thread: %ld\n", rc2);
 500                 return -ESRCH;
 501         }
 502
 503         /* wait for acceptor to startup */
 504         wait_for_completion(&lnet_acceptor_state.pta_signal);
 505
 506         if (!lnet_acceptor_state.pta_shutdown) {
 507                 /* started OK */
 508                 LASSERT(lnet_acceptor_state.pta_sock != NULL);
 509                 return 0;
 510         }
 511
 512         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 513
 514         return -ENETDOWN;
 515 }
 516
 517 void
 518 lnet_acceptor_stop(void)
 519 {
 520         if (lnet_acceptor_state.pta_shutdown) /* not running */
 521                 return;
 522
 523         /* If still required, return immediately */
 524         if (the_lnet.ln_refcount && lnet_count_acceptor_nets() > 0)
 525                 return;
 526
 527         lnet_acceptor_state.pta_shutdown = 1;
 528         wake_up(&lnet_acceptor_state.pta_waitq);
 529
 530         /* block until acceptor signals exit */
 531         wait_for_completion(&lnet_acceptor_state.pta_signal);
 532 }