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         LASSERT(rc == 0);                       /* we succeeded before */
 209
 210         if (!lnet_accept_magic(magic, LNET_PROTO_ACCEPTOR_MAGIC)) {
 211
 212                 if (lnet_accept_magic(magic, LNET_PROTO_MAGIC)) {
 213                         /* future version compatibility!
 214                          * When LNET unifies protocols over all LNDs, the first
 215                          * thing sent will be a version query.  I send back
 216                          * LNET_PROTO_ACCEPTOR_MAGIC to tell her I'm "old" */
 217
 218                         memset(&cr, 0, sizeof(cr));
 219                         cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
 220                         cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 221                         rc = lnet_sock_write(sock, &cr, sizeof(cr),
 222                                                accept_timeout);
 223
 224                         if (rc != 0)
 225                                 CERROR("Error sending magic+version in response to LNET magic from %pIS: %d\n",
 226                                        &peer, rc);
 227                         return -EPROTO;
 228                 }
 229
 230                 if (lnet_accept_magic(magic, LNET_PROTO_TCP_MAGIC))
 231                         str = "'old' socknal/tcpnal";
 232                 else
 233                         str = "unrecognised";
 234
 235                 LCONSOLE_ERROR_MSG(0x11f, "Refusing connection from %pIS"
 236                                    " magic %08x: %s acceptor protocol\n",
 237                                    &peer, magic, str);
 238                 return -EPROTO;
 239         }
 240
 241         flip = (magic != LNET_PROTO_ACCEPTOR_MAGIC);
 242
 243         rc = lnet_sock_read(sock, &cr.acr_version,
 244                               sizeof(cr.acr_version),
 245                               accept_timeout);
 246         if (rc != 0) {
 247                 CERROR("Error %d reading connection request version from %pIS\n",
 248                        rc, &peer);
 249                 return -EIO;
 250         }
 251
 252         if (flip)
 253                 __swab32s(&cr.acr_version);
 254
 255         if (cr.acr_version != LNET_PROTO_ACCEPTOR_VERSION) {
 256                 /* future version compatibility!
 257                  * An acceptor-specific protocol rev will first send a version
 258                  * query.  I send back my current version to tell her I'm
 259                  * "old". */
 260                 int peer_version = cr.acr_version;
 261
 262                 memset(&cr, 0, sizeof(cr));
 263                 cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
 264                 cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 265
 266                 rc = lnet_sock_write(sock, &cr, sizeof(cr),
 267                                        accept_timeout);
 268
 269                 if (rc != 0)
 270                         CERROR("Error sending magic+version in response to version %d from %pIS: %d\n",
 271                                peer_version, &peer, rc);
 272                 return -EPROTO;
 273         }
 274
 275         rc = lnet_sock_read(sock, &cr.acr_nid,
 276                               sizeof(cr) -
 277                               offsetof(struct lnet_acceptor_connreq, acr_nid),
 278                               accept_timeout);
 279         if (rc != 0) {
 280                 CERROR("Error %d reading connection request from %pIS\n",
 281                        rc, &peer);
 282                 return -EIO;
 283         }
 284
 285         if (flip)
 286                 __swab64s(&cr.acr_nid);
 287
 288         ni = lnet_nid2ni_addref(cr.acr_nid);
 289         if (ni == NULL ||               /* no matching net */
 290             ni->ni_nid != cr.acr_nid) { /* right NET, wrong NID! */
 291                 if (ni != NULL)
 292                         lnet_ni_decref(ni);
 293                 LCONSOLE_ERROR_MSG(0x120,
 294                                    "Refusing connection from %pIS for %s: No matching NI\n",
 295                                    &peer, libcfs_nid2str(cr.acr_nid));
 296                 return -EPERM;
 297         }
 298
 299         if (ni->ni_net->net_lnd->lnd_accept == NULL) {
 300                 /* This catches a request for the loopback LND */
 301                 lnet_ni_decref(ni);
 302                 LCONSOLE_ERROR_MSG(0x121,
 303                                    "Refusing connection from %pIS for %s: NI doesn not accept IP connections\n",
 304                                   &peer, libcfs_nid2str(cr.acr_nid));
 305                 return -EPERM;
 306         }
 307
 308         CDEBUG(D_NET, "Accept %s from %pI4h\n",
 309                libcfs_nid2str(cr.acr_nid), &peer);
 310
 311         rc = ni->ni_net->net_lnd->lnd_accept(ni, sock);
 312
 313         lnet_ni_decref(ni);
 314         return rc;
 315 }
 316
 317 #ifdef HAVE_SK_DATA_READY_ONE_ARG
 318 static void lnet_acceptor_ready(struct sock *sk)
 319 #else
 320 static void lnet_acceptor_ready(struct sock *sk, int len)
 321 #endif
 322 {
 323         /* Ensure pta_odata has actually been set before calling it */
 324         rmb();
 325 #ifdef HAVE_SK_DATA_READY_ONE_ARG
 326         lnet_acceptor_state.pta_odata(sk);
 327 #else
 328         lnet_acceptor_state.pta_odata(sk, 0);
 329 #endif
 330
 331         atomic_set(&lnet_acceptor_state.pta_ready, 1);
 332         wake_up(&lnet_acceptor_state.pta_waitq);
 333 }
 334
 335 static int
 336 lnet_acceptor(void *arg)
 337 {
 338         struct socket  *newsock;
 339         int            rc;
 340         __u32          magic;
 341         struct sockaddr_storage peer;
 342         int            secure = (int)((uintptr_t)arg);
 343
 344         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 345
 346         lnet_acceptor_state.pta_sock =
 347                 lnet_sock_listen(accept_port, accept_backlog,
 348                                  lnet_acceptor_state.pta_ns);
 349         if (IS_ERR(lnet_acceptor_state.pta_sock)) {
 350                 rc = PTR_ERR(lnet_acceptor_state.pta_sock);
 351                 if (rc == -EADDRINUSE)
 352                         LCONSOLE_ERROR_MSG(0x122, "Can't start acceptor on port"
 353                                            " %d: port already in use\n",
 354                                            accept_port);
 355                 else
 356                         LCONSOLE_ERROR_MSG(0x123, "Can't start acceptor on port "
 357                                            "%d: unexpected error %d\n",
 358                                            accept_port, rc);
 359
 360                 lnet_acceptor_state.pta_sock = NULL;
 361         } else {
 362                 rc = 0;
 363                 LCONSOLE(0, "Accept %s, port %d\n", accept_type, accept_port);
 364                 init_waitqueue_head(&lnet_acceptor_state.pta_waitq);
 365                 lnet_acceptor_state.pta_odata =
 366                         lnet_acceptor_state.pta_sock->sk->sk_data_ready;
 367                 /* ensure pta_odata gets set before there is any chance of
 368                  * lnet_accept_ready() trying to read it.
 369                  */
 370                 wmb();
 371                 lnet_acceptor_state.pta_sock->sk->sk_data_ready =
 372                         lnet_acceptor_ready;
 373                 atomic_set(&lnet_acceptor_state.pta_ready, 1);
 374         }
 375
 376         /* set init status and unblock parent */
 377         lnet_acceptor_state.pta_shutdown = rc;
 378         complete(&lnet_acceptor_state.pta_signal);
 379
 380         if (rc != 0)
 381                 return rc;
 382
 383         while (!lnet_acceptor_state.pta_shutdown) {
 384
 385                 wait_event_idle(lnet_acceptor_state.pta_waitq,
 386                                 lnet_acceptor_state.pta_shutdown ||
 387                                 atomic_read(&lnet_acceptor_state.pta_ready));
 388                 if (!atomic_read(&lnet_acceptor_state.pta_ready))
 389                         continue;
 390                 atomic_set(&lnet_acceptor_state.pta_ready, 0);
 391                 rc = kernel_accept(lnet_acceptor_state.pta_sock, &newsock,
 392                                    SOCK_NONBLOCK);
 393                 if (rc != 0) {
 394                         if (rc != -EAGAIN) {
 395                                 CWARN("Accept error %d: pausing...\n", rc);
 396                                 schedule_timeout_uninterruptible(
 397                                         cfs_time_seconds(1));
 398                         }
 399                         continue;
 400                 }
 401
 402                 /* make sure we call lnet_sock_accept() again, until it fails */
 403                 atomic_set(&lnet_acceptor_state.pta_ready, 1);
 404
 405                 rc = lnet_sock_getaddr(newsock, true, &peer);
 406                 if (rc != 0) {
 407                         CERROR("Can't determine new connection's address\n");
 408                         goto failed;
 409                 }
 410
 411                 if (secure &&
 412                     rpc_get_port((struct sockaddr *)&peer) >
 413                     LNET_ACCEPTOR_MAX_RESERVED_PORT) {
 414                         CERROR("Refusing connection from %pISp: insecure port.\n",
 415                                &peer);
 416                         goto failed;
 417                 }
 418
 419                 rc = lnet_sock_read(newsock, &magic, sizeof(magic),
 420                                       accept_timeout);
 421                 if (rc != 0) {
 422                         CERROR("Error %d reading connection request from %pIS\n",
 423                                rc, &peer);
 424                         goto failed;
 425                 }
 426
 427                 rc = lnet_accept(newsock, magic);
 428                 if (rc != 0)
 429                         goto failed;
 430
 431                 continue;
 432
 433 failed:
 434                 sock_release(newsock);
 435         }
 436
 437         lnet_acceptor_state.pta_sock->sk->sk_data_ready =
 438                 lnet_acceptor_state.pta_odata;
 439         sock_release(lnet_acceptor_state.pta_sock);
 440         lnet_acceptor_state.pta_sock = NULL;
 441
 442         CDEBUG(D_NET, "Acceptor stopping\n");
 443
 444         /* unblock lnet_acceptor_stop() */
 445         complete(&lnet_acceptor_state.pta_signal);
 446         return 0;
 447 }
 448
 449 static inline int
 450 accept2secure(const char *acc, long *sec)
 451 {
 452         if (!strcmp(acc, "secure")) {
 453                 *sec = 1;
 454                 return 1;
 455         } else if (!strcmp(acc, "all")) {
 456                 *sec = 0;
 457                 return 1;
 458         } else if (!strcmp(acc, "none")) {
 459                 return 0;
 460         } else {
 461                 LCONSOLE_ERROR_MSG(0x124, "Can't parse 'accept=\"%s\"'\n",
 462                                    acc);
 463                 return -EINVAL;
 464         }
 465 }
 466
 467 int
 468 lnet_acceptor_start(void)
 469 {
 470         struct task_struct *task;
 471         int  rc;
 472         long rc2;
 473         long secure;
 474
 475         /* if acceptor is already running return immediately */
 476         if (!lnet_acceptor_state.pta_shutdown)
 477                 return 0;
 478
 479         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 480
 481         init_completion(&lnet_acceptor_state.pta_signal);
 482         rc = accept2secure(accept_type, &secure);
 483         if (rc <= 0)
 484                 return rc;
 485
 486         if (lnet_count_acceptor_nets() == 0)  /* not required */
 487                 return 0;
 488         if (current->nsproxy && current->nsproxy->net_ns)
 489                 lnet_acceptor_state.pta_ns = current->nsproxy->net_ns;
 490         else
 491                 lnet_acceptor_state.pta_ns = &init_net;
 492         task = kthread_run(lnet_acceptor, (void *)(uintptr_t)secure,
 493                            "acceptor_%03ld", secure);
 494         if (IS_ERR(task)) {
 495                 rc2 = PTR_ERR(task);
 496                 CERROR("Can't start acceptor thread: %ld\n", rc2);
 497                 return -ESRCH;
 498         }
 499
 500         /* wait for acceptor to startup */
 501         wait_for_completion(&lnet_acceptor_state.pta_signal);
 502
 503         if (!lnet_acceptor_state.pta_shutdown) {
 504                 /* started OK */
 505                 LASSERT(lnet_acceptor_state.pta_sock != NULL);
 506                 return 0;
 507         }
 508
 509         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 510
 511         return -ENETDOWN;
 512 }
 513
 514 void
 515 lnet_acceptor_stop(void)
 516 {
 517         if (lnet_acceptor_state.pta_shutdown) /* not running */
 518                 return;
 519
 520         /* If still required, return immediately */
 521         if (the_lnet.ln_refcount && lnet_count_acceptor_nets() > 0)
 522                 return;
 523
 524         lnet_acceptor_state.pta_shutdown = 1;
 525         wake_up(&lnet_acceptor_state.pta_waitq);
 526
 527         /* block until acceptor signals exit */
 528         wait_for_completion(&lnet_acceptor_state.pta_signal);
 529 }