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, struct lnet_nid *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_nidstr(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_nidstr(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_nidstr(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_nidstr(peer_nid), sa,
 115                                    libcfs_nidstr(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_nidstr(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_nidstr(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_nidstr(peer_nid), sa);
 131                 break;
 132         }
 133 }
 134 EXPORT_SYMBOL(lnet_connect_console_error);
 135
 136 struct socket *
 137 lnet_connect(struct lnet_nid *peer_nid, int interface,
 138              struct sockaddr *peeraddr,
 139              struct net *ns)
 140 {
 141         struct lnet_acceptor_connreq cr1;
 142         struct lnet_acceptor_connreq_v2 cr2;
 143         void *cr;
 144         int crsize;
 145         struct socket *sock;
 146         int rc;
 147         int port;
 148
 149         BUILD_BUG_ON(sizeof(cr) > 16); /* not too big to be on the stack */
 150
 151         LASSERT(peeraddr->sa_family == AF_INET ||
 152                 peeraddr->sa_family == AF_INET6);
 153
 154         for (port = LNET_ACCEPTOR_MAX_RESERVED_PORT;
 155              port >= LNET_ACCEPTOR_MIN_RESERVED_PORT;
 156              --port) {
 157                 /* Iterate through reserved ports. */
 158                 sock = lnet_sock_connect(interface, port, peeraddr, ns);
 159                 if (IS_ERR(sock)) {
 160                         rc = PTR_ERR(sock);
 161                         if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL)
 162                                 continue;
 163                         goto failed;
 164                 }
 165
 166                 BUILD_BUG_ON(LNET_PROTO_ACCEPTOR_VERSION != 1);
 167
 168                 if (nid_is_nid4(peer_nid)) {
 169                         cr1.acr_magic   = LNET_PROTO_ACCEPTOR_MAGIC;
 170                         cr1.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 171                         cr1.acr_nid     = lnet_nid_to_nid4(peer_nid);
 172                         cr = &cr1;
 173                         crsize = sizeof(cr1);
 174
 175                         if (the_lnet.ln_testprotocompat) {
 176                                 /* single-shot proto check */
 177                                 if (test_and_clear_bit(
 178                                             2, &the_lnet.ln_testprotocompat))
 179                                         cr1.acr_version++;
 180                                 if (test_and_clear_bit(
 181                                             3, &the_lnet.ln_testprotocompat))
 182                                         cr1.acr_magic = LNET_PROTO_MAGIC;
 183                         }
 184
 185                 } else {
 186                         cr2.acr_magic   = LNET_PROTO_ACCEPTOR_MAGIC;
 187                         cr2.acr_version = LNET_PROTO_ACCEPTOR_VERSION_16;
 188                         cr2.acr_nid     = *peer_nid;
 189                         cr = &cr2;
 190                         crsize = sizeof(cr2);
 191                 }
 192
 193                 rc = lnet_sock_write(sock, cr, crsize, accept_timeout);
 194                 if (rc != 0)
 195                         goto failed_sock;
 196
 197                 return sock;
 198         }
 199
 200         rc = -EADDRINUSE;
 201         goto failed;
 202
 203 failed_sock:
 204         sock_release(sock);
 205 failed:
 206         lnet_connect_console_error(rc, peer_nid, peeraddr);
 207         return ERR_PTR(rc);
 208 }
 209 EXPORT_SYMBOL(lnet_connect);
 210
 211 static int
 212 lnet_accept(struct socket *sock, __u32 magic)
 213 {
 214         struct lnet_acceptor_connreq cr;
 215         struct lnet_acceptor_connreq_v2 cr2;
 216         struct lnet_nid nid;
 217         struct sockaddr_storage peer;
 218         int peer_version;
 219         int rc;
 220         int flip;
 221         struct lnet_ni *ni;
 222         char *str;
 223
 224         LASSERT(sizeof(cr) <= 16);              /* not too big for the stack */
 225
 226         rc = lnet_sock_getaddr(sock, true, &peer);
 227         if (rc != 0) {
 228                 CERROR("Can't determine new connection's address\n");
 229                 return rc;
 230         }
 231
 232         if (!lnet_accept_magic(magic, LNET_PROTO_ACCEPTOR_MAGIC)) {
 233
 234                 if (lnet_accept_magic(magic, LNET_PROTO_MAGIC)) {
 235                         /* future version compatibility!
 236                          * When LNET unifies protocols over all LNDs, the first
 237                          * thing sent will be a version query.  I send back
 238                          * LNET_PROTO_ACCEPTOR_MAGIC to tell her I'm "old" */
 239
 240                         memset(&cr, 0, sizeof(cr));
 241                         cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
 242                         cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 243                         rc = lnet_sock_write(sock, &cr, sizeof(cr),
 244                                                accept_timeout);
 245
 246                         if (rc != 0)
 247                                 CERROR("Error sending magic+version in response to LNET magic from %pIS: %d\n",
 248                                        &peer, rc);
 249                         return -EPROTO;
 250                 }
 251
 252                 if (lnet_accept_magic(magic, LNET_PROTO_TCP_MAGIC))
 253                         str = "'old' socknal/tcpnal";
 254                 else
 255                         str = "unrecognised";
 256
 257                 LCONSOLE_ERROR_MSG(0x11f, "Refusing connection from %pIS"
 258                                    " magic %08x: %s acceptor protocol\n",
 259                                    &peer, magic, str);
 260                 return -EPROTO;
 261         }
 262
 263         flip = (magic != LNET_PROTO_ACCEPTOR_MAGIC);
 264
 265         rc = lnet_sock_read(sock, &cr.acr_version,
 266                               sizeof(cr.acr_version),
 267                               accept_timeout);
 268         if (rc != 0) {
 269                 CERROR("Error %d reading connection request version from %pIS\n",
 270                        rc, &peer);
 271                 return -EIO;
 272         }
 273
 274         if (flip)
 275                 __swab32s(&cr.acr_version);
 276
 277         switch (cr.acr_version) {
 278         default:
 279                 /* future version compatibility!
 280                  * An acceptor-specific protocol rev will first send a version
 281                  * query.  I send back my current version to tell her I'm
 282                  * "old". */
 283                 peer_version = cr.acr_version;
 284
 285                 memset(&cr, 0, sizeof(cr));
 286                 cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
 287                 cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 288
 289                 rc = lnet_sock_write(sock, &cr, sizeof(cr),
 290                                        accept_timeout);
 291
 292                 if (rc != 0)
 293                         CERROR("Error sending magic+version in response to version %d from %pIS: %d\n",
 294                                peer_version, &peer, rc);
 295                 return -EPROTO;
 296
 297         case LNET_PROTO_ACCEPTOR_VERSION:
 298
 299                 rc = lnet_sock_read(sock, &cr.acr_nid,
 300                                     sizeof(cr) -
 301                                     offsetof(struct lnet_acceptor_connreq,
 302                                              acr_nid),
 303                                     accept_timeout);
 304                 if (rc)
 305                         break;
 306                 if (flip)
 307                         __swab64s(&cr.acr_nid);
 308
 309                 lnet_nid4_to_nid(cr.acr_nid, &nid);
 310                 break;
 311
 312         case LNET_PROTO_ACCEPTOR_VERSION_16:
 313                 rc = lnet_sock_read(sock, &cr2.acr_nid,
 314                                     sizeof(cr2) -
 315                                     offsetof(struct lnet_acceptor_connreq_v2,
 316                                              acr_nid),
 317                                     accept_timeout);
 318                 if (rc)
 319                         break;
 320                 nid = cr2.acr_nid;
 321                 break;
 322         }
 323         if (rc != 0) {
 324                 CERROR("Error %d reading connection request from %pIS\n",
 325                        rc, &peer);
 326                 return -EIO;
 327         }
 328
 329         ni = lnet_nid_to_ni_addref(&nid);
 330         if (ni == NULL ||               /* no matching net */
 331             !nid_same(&ni->ni_nid, &nid)) {
 332                 /* right NET, wrong NID! */
 333                 if (ni != NULL)
 334                         lnet_ni_decref(ni);
 335                 LCONSOLE_ERROR_MSG(0x120,
 336                                    "Refusing connection from %pIS for %s: No matching NI\n",
 337                                    &peer, libcfs_nidstr(&nid));
 338                 return -EPERM;
 339         }
 340
 341         if (ni->ni_net->net_lnd->lnd_accept == NULL) {
 342                 /* This catches a request for the loopback LND */
 343                 lnet_ni_decref(ni);
 344                 LCONSOLE_ERROR_MSG(0x121,
 345                                    "Refusing connection from %pIS for %s: NI doesn not accept IP connections\n",
 346                                   &peer, libcfs_nidstr(&nid));
 347                 return -EPERM;
 348         }
 349
 350         CDEBUG(D_NET, "Accept %s from %pI4h\n", libcfs_nidstr(&nid), &peer);
 351
 352         rc = ni->ni_net->net_lnd->lnd_accept(ni, sock);
 353
 354         lnet_ni_decref(ni);
 355         return rc;
 356 }
 357
 358 #ifdef HAVE_SK_DATA_READY_ONE_ARG
 359 static void lnet_acceptor_ready(struct sock *sk)
 360 #else
 361 static void lnet_acceptor_ready(struct sock *sk, int len)
 362 #endif
 363 {
 364         /* Ensure pta_odata has actually been set before calling it */
 365         rmb();
 366 #ifdef HAVE_SK_DATA_READY_ONE_ARG
 367         lnet_acceptor_state.pta_odata(sk);
 368 #else
 369         lnet_acceptor_state.pta_odata(sk, 0);
 370 #endif
 371
 372         atomic_set(&lnet_acceptor_state.pta_ready, 1);
 373         wake_up(&lnet_acceptor_state.pta_waitq);
 374 }
 375
 376 static int
 377 lnet_acceptor(void *arg)
 378 {
 379         struct socket  *newsock;
 380         int            rc;
 381         __u32          magic;
 382         struct sockaddr_storage peer;
 383         int            secure = (int)((uintptr_t)arg);
 384
 385         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 386
 387         lnet_acceptor_state.pta_sock =
 388                 lnet_sock_listen(accept_port, accept_backlog,
 389                                  lnet_acceptor_state.pta_ns);
 390         if (IS_ERR(lnet_acceptor_state.pta_sock)) {
 391                 rc = PTR_ERR(lnet_acceptor_state.pta_sock);
 392                 if (rc == -EADDRINUSE)
 393                         LCONSOLE_ERROR_MSG(0x122, "Can't start acceptor on port"
 394                                            " %d: port already in use\n",
 395                                            accept_port);
 396                 else
 397                         LCONSOLE_ERROR_MSG(0x123, "Can't start acceptor on port "
 398                                            "%d: unexpected error %d\n",
 399                                            accept_port, rc);
 400
 401                 lnet_acceptor_state.pta_sock = NULL;
 402         } else {
 403                 rc = 0;
 404                 LCONSOLE(0, "Accept %s, port %d\n", accept_type, accept_port);
 405                 init_waitqueue_head(&lnet_acceptor_state.pta_waitq);
 406                 lnet_acceptor_state.pta_odata =
 407                         lnet_acceptor_state.pta_sock->sk->sk_data_ready;
 408                 /* ensure pta_odata gets set before there is any chance of
 409                  * lnet_accept_ready() trying to read it.
 410                  */
 411                 wmb();
 412                 lnet_acceptor_state.pta_sock->sk->sk_data_ready =
 413                         lnet_acceptor_ready;
 414                 atomic_set(&lnet_acceptor_state.pta_ready, 1);
 415         }
 416
 417         /* set init status and unblock parent */
 418         lnet_acceptor_state.pta_shutdown = rc;
 419         complete(&lnet_acceptor_state.pta_signal);
 420
 421         if (rc != 0)
 422                 return rc;
 423
 424         while (!lnet_acceptor_state.pta_shutdown) {
 425
 426                 wait_event_idle(lnet_acceptor_state.pta_waitq,
 427                                 lnet_acceptor_state.pta_shutdown ||
 428                                 atomic_read(&lnet_acceptor_state.pta_ready));
 429                 if (!atomic_read(&lnet_acceptor_state.pta_ready))
 430                         continue;
 431                 atomic_set(&lnet_acceptor_state.pta_ready, 0);
 432                 rc = kernel_accept(lnet_acceptor_state.pta_sock, &newsock,
 433                                    SOCK_NONBLOCK);
 434                 if (rc != 0) {
 435                         if (rc != -EAGAIN) {
 436                                 CWARN("Accept error %d: pausing...\n", rc);
 437                                 schedule_timeout_uninterruptible(
 438                                         cfs_time_seconds(1));
 439                         }
 440                         continue;
 441                 }
 442
 443                 /* make sure we call lnet_sock_accept() again, until it fails */
 444                 atomic_set(&lnet_acceptor_state.pta_ready, 1);
 445
 446                 rc = lnet_sock_getaddr(newsock, true, &peer);
 447                 if (rc != 0) {
 448                         CERROR("Can't determine new connection's address\n");
 449                         goto failed;
 450                 }
 451
 452                 if (secure &&
 453                     rpc_get_port((struct sockaddr *)&peer) >
 454                     LNET_ACCEPTOR_MAX_RESERVED_PORT) {
 455                         CERROR("Refusing connection from %pISp: insecure port.\n",
 456                                &peer);
 457                         goto failed;
 458                 }
 459
 460                 rc = lnet_sock_read(newsock, &magic, sizeof(magic),
 461                                       accept_timeout);
 462                 if (rc != 0) {
 463                         CERROR("Error %d reading connection request from %pIS\n",
 464                                rc, &peer);
 465                         goto failed;
 466                 }
 467
 468                 rc = lnet_accept(newsock, magic);
 469                 if (rc != 0)
 470                         goto failed;
 471
 472                 continue;
 473
 474 failed:
 475                 sock_release(newsock);
 476         }
 477
 478         lnet_acceptor_state.pta_sock->sk->sk_data_ready =
 479                 lnet_acceptor_state.pta_odata;
 480         sock_release(lnet_acceptor_state.pta_sock);
 481         lnet_acceptor_state.pta_sock = NULL;
 482
 483         CDEBUG(D_NET, "Acceptor stopping\n");
 484
 485         /* unblock lnet_acceptor_stop() */
 486         complete(&lnet_acceptor_state.pta_signal);
 487         return 0;
 488 }
 489
 490 static inline int
 491 accept2secure(const char *acc, long *sec)
 492 {
 493         if (!strcmp(acc, "secure")) {
 494                 *sec = 1;
 495                 return 1;
 496         } else if (!strcmp(acc, "all")) {
 497                 *sec = 0;
 498                 return 1;
 499         } else if (!strcmp(acc, "none")) {
 500                 return 0;
 501         } else {
 502                 LCONSOLE_ERROR_MSG(0x124, "Can't parse 'accept=\"%s\"'\n",
 503                                    acc);
 504                 return -EINVAL;
 505         }
 506 }
 507
 508 int
 509 lnet_acceptor_start(void)
 510 {
 511         struct task_struct *task;
 512         int  rc;
 513         long rc2;
 514         long secure;
 515
 516         /* if acceptor is already running return immediately */
 517         if (!lnet_acceptor_state.pta_shutdown)
 518                 return 0;
 519
 520         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 521
 522         init_completion(&lnet_acceptor_state.pta_signal);
 523         rc = accept2secure(accept_type, &secure);
 524         if (rc <= 0)
 525                 return rc;
 526
 527         if (lnet_count_acceptor_nets() == 0)  /* not required */
 528                 return 0;
 529         if (current->nsproxy && current->nsproxy->net_ns)
 530                 lnet_acceptor_state.pta_ns = current->nsproxy->net_ns;
 531         else
 532                 lnet_acceptor_state.pta_ns = &init_net;
 533         task = kthread_run(lnet_acceptor, (void *)(uintptr_t)secure,
 534                            "acceptor_%03ld", secure);
 535         if (IS_ERR(task)) {
 536                 rc2 = PTR_ERR(task);
 537                 CERROR("Can't start acceptor thread: %ld\n", rc2);
 538                 return -ESRCH;
 539         }
 540
 541         /* wait for acceptor to startup */
 542         wait_for_completion(&lnet_acceptor_state.pta_signal);
 543
 544         if (!lnet_acceptor_state.pta_shutdown) {
 545                 /* started OK */
 546                 LASSERT(lnet_acceptor_state.pta_sock != NULL);
 547                 return 0;
 548         }
 549
 550         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 551
 552         return -ENETDOWN;
 553 }
 554
 555 void
 556 lnet_acceptor_stop(void)
 557 {
 558         if (lnet_acceptor_state.pta_shutdown) /* not running */
 559                 return;
 560
 561         /* If still required, return immediately */
 562         if (the_lnet.ln_refcount && lnet_count_acceptor_nets() > 0)
 563                 return;
 564
 565         lnet_acceptor_state.pta_shutdown = 1;
 566         wake_up(&lnet_acceptor_state.pta_waitq);
 567
 568         /* block until acceptor signals exit */
 569         wait_for_completion(&lnet_acceptor_state.pta_signal);
 570 }