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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_LNET
  34
  35 #include <linux/completion.h>
  36 #include <net/sock.h>
  37 #include <lnet/lib-lnet.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                            __u32 peer_ip, int peer_port)
  95 {
  96         switch (rc) {
  97         /* "normal" errors */
  98         case -ECONNREFUSED:
  99                 CNETERR("Connection to %s at host %pI4h on port %d was "
 100                         "refused: check that Lustre is running on that node.\n",
 101                         libcfs_nid2str(peer_nid), &peer_ip, peer_port);
 102                 break;
 103         case -EHOSTUNREACH:
 104         case -ENETUNREACH:
 105                 CNETERR("Connection to %s at host %pI4h "
 106                         "was unreachable: the network or that node may "
 107                         "be down, or Lustre may be misconfigured.\n",
 108                         libcfs_nid2str(peer_nid), &peer_ip);
 109                 break;
 110         case -ETIMEDOUT:
 111                 CNETERR("Connection to %s at host %pI4h on "
 112                         "port %d took too long: that node may be hung "
 113                         "or experiencing high load.\n",
 114                         libcfs_nid2str(peer_nid), &peer_ip, peer_port);
 115                 break;
 116         case -ECONNRESET:
 117                 LCONSOLE_ERROR_MSG(0x11b, "Connection to %s at host %pI4h"
 118                                    " on port %d was reset: "
 119                                    "is it running a compatible version of "
 120                                    "Lustre and is %s one of its NIDs?\n",
 121                                    libcfs_nid2str(peer_nid), &peer_ip,
 122                                    peer_port, libcfs_nid2str(peer_nid));
 123                 break;
 124         case -EPROTO:
 125                 LCONSOLE_ERROR_MSG(0x11c, "Protocol error connecting to %s at "
 126                                    "host %pI4h on port %d: is it running "
 127                                    "a compatible version of Lustre?\n",
 128                                    libcfs_nid2str(peer_nid), &peer_ip,
 129                                    peer_port);
 130                 break;
 131         case -EADDRINUSE:
 132                 LCONSOLE_ERROR_MSG(0x11d, "No privileged ports available to "
 133                                    "connect to %s at host %pI4h on port "
 134                                    "%d\n", libcfs_nid2str(peer_nid),
 135                                    &peer_ip, peer_port);
 136                 break;
 137         default:
 138                 LCONSOLE_ERROR_MSG(0x11e, "Unexpected error %d connecting to %s"
 139                                    " at host %pI4h on port %d\n", rc,
 140                                    libcfs_nid2str(peer_nid),
 141                                    &peer_ip, peer_port);
 142                 break;
 143         }
 144 }
 145 EXPORT_SYMBOL(lnet_connect_console_error);
 146
 147 int
 148 lnet_connect(struct socket **sockp, lnet_nid_t peer_nid,
 149             __u32 local_ip, __u32 peer_ip, int peer_port, struct net *ns)
 150 {
 151         struct lnet_acceptor_connreq cr;
 152         struct socket           *sock;
 153         int                     rc;
 154         int                     port;
 155
 156         BUILD_BUG_ON(sizeof(cr) > 16); /* not too big to be on the stack */
 157
 158         for (port = LNET_ACCEPTOR_MAX_RESERVED_PORT;
 159              port >= LNET_ACCEPTOR_MIN_RESERVED_PORT;
 160              --port) {
 161                 /* Iterate through reserved ports. */
 162
 163                 rc = lnet_sock_connect(&sock,
 164                                        local_ip, port,
 165                                        peer_ip, peer_port, ns);
 166                 if (rc) {
 167                         if (rc == -EADDRINUSE || rc == -EADDRNOTAVAIL)
 168                                 continue;
 169                         goto failed;
 170                 }
 171
 172                 BUILD_BUG_ON(LNET_PROTO_ACCEPTOR_VERSION != 1);
 173
 174                 cr.acr_magic   = LNET_PROTO_ACCEPTOR_MAGIC;
 175                 cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 176                 cr.acr_nid     = peer_nid;
 177
 178                 if (the_lnet.ln_testprotocompat) {
 179                         /* single-shot proto check */
 180                         if (test_and_clear_bit(2, &the_lnet.ln_testprotocompat))
 181                                 cr.acr_version++;
 182                         if (test_and_clear_bit(3, &the_lnet.ln_testprotocompat))
 183                                 cr.acr_magic = LNET_PROTO_MAGIC;
 184                 }
 185
 186                 rc = lnet_sock_write(sock, &cr, sizeof(cr),
 187                                        accept_timeout);
 188                 if (rc != 0)
 189                         goto failed_sock;
 190
 191                 *sockp = sock;
 192                 return 0;
 193         }
 194
 195         rc = -EADDRINUSE;
 196         goto failed;
 197
 198 failed_sock:
 199         sock_release(sock);
 200 failed:
 201         lnet_connect_console_error(rc, peer_nid, peer_ip, peer_port);
 202         return rc;
 203 }
 204 EXPORT_SYMBOL(lnet_connect);
 205
 206 static int
 207 lnet_accept(struct socket *sock, __u32 magic)
 208 {
 209         struct lnet_acceptor_connreq cr;
 210         __u32                   peer_ip;
 211         int                     peer_port;
 212         int                     rc;
 213         int                     flip;
 214         struct lnet_ni *ni;
 215         char                   *str;
 216
 217         LASSERT(sizeof(cr) <= 16);              /* not too big for the stack */
 218
 219         rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
 220         LASSERT(rc == 0);                       /* we succeeded before */
 221
 222         if (!lnet_accept_magic(magic, LNET_PROTO_ACCEPTOR_MAGIC)) {
 223
 224                 if (lnet_accept_magic(magic, LNET_PROTO_MAGIC)) {
 225                         /* future version compatibility!
 226                          * When LNET unifies protocols over all LNDs, the first
 227                          * thing sent will be a version query.  I send back
 228                          * LNET_PROTO_ACCEPTOR_MAGIC to tell her I'm "old" */
 229
 230                         memset(&cr, 0, sizeof(cr));
 231                         cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
 232                         cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 233                         rc = lnet_sock_write(sock, &cr, sizeof(cr),
 234                                                accept_timeout);
 235
 236                         if (rc != 0)
 237                                 CERROR("Error sending magic+version in response to LNET magic from %pI4h: %d\n",
 238                                        &peer_ip, rc);
 239                         return -EPROTO;
 240                 }
 241
 242                 if (lnet_accept_magic(magic, LNET_PROTO_TCP_MAGIC))
 243                         str = "'old' socknal/tcpnal";
 244                 else
 245                         str = "unrecognised";
 246
 247                 LCONSOLE_ERROR_MSG(0x11f, "Refusing connection from %pI4h"
 248                                    " magic %08x: %s acceptor protocol\n",
 249                                    &peer_ip, magic, str);
 250                 return -EPROTO;
 251         }
 252
 253         flip = (magic != LNET_PROTO_ACCEPTOR_MAGIC);
 254
 255         rc = lnet_sock_read(sock, &cr.acr_version,
 256                               sizeof(cr.acr_version),
 257                               accept_timeout);
 258         if (rc != 0) {
 259                 CERROR("Error %d reading connection request version from "
 260                        "%pI4h\n", rc, &peer_ip);
 261                 return -EIO;
 262         }
 263
 264         if (flip)
 265                 __swab32s(&cr.acr_version);
 266
 267         if (cr.acr_version != LNET_PROTO_ACCEPTOR_VERSION) {
 268                 /* future version compatibility!
 269                  * An acceptor-specific protocol rev will first send a version
 270                  * query.  I send back my current version to tell her I'm
 271                  * "old". */
 272                 int peer_version = cr.acr_version;
 273
 274                 memset(&cr, 0, sizeof(cr));
 275                 cr.acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
 276                 cr.acr_version = LNET_PROTO_ACCEPTOR_VERSION;
 277
 278                 rc = lnet_sock_write(sock, &cr, sizeof(cr),
 279                                        accept_timeout);
 280
 281                 if (rc != 0)
 282                         CERROR("Error sending magic+version in response to version %d from %pI4h: %d\n",
 283                                peer_version, &peer_ip, rc);
 284                 return -EPROTO;
 285         }
 286
 287         rc = lnet_sock_read(sock, &cr.acr_nid,
 288                               sizeof(cr) -
 289                               offsetof(struct lnet_acceptor_connreq, acr_nid),
 290                               accept_timeout);
 291         if (rc != 0) {
 292                 CERROR("Error %d reading connection request from "
 293                        "%pI4h\n", rc, &peer_ip);
 294                 return -EIO;
 295         }
 296
 297         if (flip)
 298                 __swab64s(&cr.acr_nid);
 299
 300         ni = lnet_nid2ni_addref(cr.acr_nid);
 301         if (ni == NULL ||               /* no matching net */
 302             ni->ni_nid != cr.acr_nid) { /* right NET, wrong NID! */
 303                 if (ni != NULL)
 304                         lnet_ni_decref(ni);
 305                 LCONSOLE_ERROR_MSG(0x120, "Refusing connection from %pI4h "
 306                                    "for %s: No matching NI\n",
 307                                    &peer_ip, libcfs_nid2str(cr.acr_nid));
 308                 return -EPERM;
 309         }
 310
 311         if (ni->ni_net->net_lnd->lnd_accept == NULL) {
 312                 /* This catches a request for the loopback LND */
 313                 lnet_ni_decref(ni);
 314                 LCONSOLE_ERROR_MSG(0x121, "Refusing connection from %pI4h "
 315                                   "for %s: NI doesn not accept IP connections\n",
 316                                   &peer_ip, libcfs_nid2str(cr.acr_nid));
 317                 return -EPERM;
 318         }
 319
 320         CDEBUG(D_NET, "Accept %s from %pI4h\n",
 321                libcfs_nid2str(cr.acr_nid), &peer_ip);
 322
 323         rc = ni->ni_net->net_lnd->lnd_accept(ni, sock);
 324
 325         lnet_ni_decref(ni);
 326         return rc;
 327 }
 328
 329 #ifdef HAVE_SK_DATA_READY_ONE_ARG
 330 static void lnet_acceptor_ready(struct sock *sk)
 331 #else
 332 static void lnet_acceptor_ready(struct sock *sk, int len)
 333 #endif
 334 {
 335         /* Ensure pta_odata has actually been set before calling it */
 336         rmb();
 337 #ifdef HAVE_SK_DATA_READY_ONE_ARG
 338         lnet_acceptor_state.pta_odata(sk);
 339 #else
 340         lnet_acceptor_state.pta_odata(sk, 0);
 341 #endif
 342
 343         atomic_set(&lnet_acceptor_state.pta_ready, 1);
 344         wake_up(&lnet_acceptor_state.pta_waitq);
 345 }
 346
 347 static int
 348 lnet_acceptor(void *arg)
 349 {
 350         struct socket  *newsock;
 351         int            rc;
 352         __u32          magic;
 353         __u32          peer_ip;
 354         int            peer_port;
 355         int            secure = (int)((uintptr_t)arg);
 356
 357         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 358
 359         rc = lnet_sock_listen(&lnet_acceptor_state.pta_sock,
 360                               0, accept_port, accept_backlog,
 361                               lnet_acceptor_state.pta_ns);
 362         if (rc != 0) {
 363                 if (rc == -EADDRINUSE)
 364                         LCONSOLE_ERROR_MSG(0x122, "Can't start acceptor on port"
 365                                            " %d: port already in use\n",
 366                                            accept_port);
 367                 else
 368                         LCONSOLE_ERROR_MSG(0x123, "Can't start acceptor on port "
 369                                            "%d: unexpected error %d\n",
 370                                            accept_port, rc);
 371
 372                 lnet_acceptor_state.pta_sock = NULL;
 373         } else {
 374                 LCONSOLE(0, "Accept %s, port %d\n", accept_type, accept_port);
 375                 init_waitqueue_head(&lnet_acceptor_state.pta_waitq);
 376                 lnet_acceptor_state.pta_odata =
 377                         lnet_acceptor_state.pta_sock->sk->sk_data_ready;
 378                 /* ensure pta_odata gets set before there is any chance of
 379                  * lnet_accept_ready() trying to read it.
 380                  */
 381                 wmb();
 382                 lnet_acceptor_state.pta_sock->sk->sk_data_ready =
 383                         lnet_acceptor_ready;
 384                 atomic_set(&lnet_acceptor_state.pta_ready, 1);
 385         }
 386
 387         /* set init status and unblock parent */
 388         lnet_acceptor_state.pta_shutdown = rc;
 389         complete(&lnet_acceptor_state.pta_signal);
 390
 391         if (rc != 0)
 392                 return rc;
 393
 394         while (!lnet_acceptor_state.pta_shutdown) {
 395
 396                 wait_event_idle(lnet_acceptor_state.pta_waitq,
 397                                 lnet_acceptor_state.pta_shutdown ||
 398                                 atomic_read(&lnet_acceptor_state.pta_ready));
 399                 if (!atomic_read(&lnet_acceptor_state.pta_ready))
 400                         continue;
 401                 atomic_set(&lnet_acceptor_state.pta_ready, 0);
 402                 rc = kernel_accept(lnet_acceptor_state.pta_sock, &newsock,
 403                                    SOCK_NONBLOCK);
 404                 if (rc != 0) {
 405                         if (rc != -EAGAIN) {
 406                                 CWARN("Accept error %d: pausing...\n", rc);
 407                                 schedule_timeout_uninterruptible(
 408                                         cfs_time_seconds(1));
 409                         }
 410                         continue;
 411                 }
 412
 413                 /* make sure we call lnet_sock_accept() again, until it fails */
 414                 atomic_set(&lnet_acceptor_state.pta_ready, 1);
 415
 416                 rc = lnet_sock_getaddr(newsock, true, &peer_ip, &peer_port);
 417                 if (rc != 0) {
 418                         CERROR("Can't determine new connection's address\n");
 419                         goto failed;
 420                 }
 421
 422                 if (secure && peer_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
 423                         CERROR("Refusing connection from %pI4h: "
 424                                "insecure port %d\n", &peer_ip, peer_port);
 425                         goto failed;
 426                 }
 427
 428                 rc = lnet_sock_read(newsock, &magic, sizeof(magic),
 429                                       accept_timeout);
 430                 if (rc != 0) {
 431                         CERROR("Error %d reading connection request from "
 432                                "%pI4h\n", rc, &peer_ip);
 433                         goto failed;
 434                 }
 435
 436                 rc = lnet_accept(newsock, magic);
 437                 if (rc != 0)
 438                         goto failed;
 439
 440                 continue;
 441
 442 failed:
 443                 sock_release(newsock);
 444         }
 445
 446         lnet_acceptor_state.pta_sock->sk->sk_data_ready =
 447                 lnet_acceptor_state.pta_odata;
 448         sock_release(lnet_acceptor_state.pta_sock);
 449         lnet_acceptor_state.pta_sock = NULL;
 450
 451         CDEBUG(D_NET, "Acceptor stopping\n");
 452
 453         /* unblock lnet_acceptor_stop() */
 454         complete(&lnet_acceptor_state.pta_signal);
 455         return 0;
 456 }
 457
 458 static inline int
 459 accept2secure(const char *acc, long *sec)
 460 {
 461         if (!strcmp(acc, "secure")) {
 462                 *sec = 1;
 463                 return 1;
 464         } else if (!strcmp(acc, "all")) {
 465                 *sec = 0;
 466                 return 1;
 467         } else if (!strcmp(acc, "none")) {
 468                 return 0;
 469         } else {
 470                 LCONSOLE_ERROR_MSG(0x124, "Can't parse 'accept=\"%s\"'\n",
 471                                    acc);
 472                 return -EINVAL;
 473         }
 474 }
 475
 476 int
 477 lnet_acceptor_start(void)
 478 {
 479         struct task_struct *task;
 480         int  rc;
 481         long rc2;
 482         long secure;
 483
 484         /* if acceptor is already running return immediately */
 485         if (!lnet_acceptor_state.pta_shutdown)
 486                 return 0;
 487
 488         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 489
 490         init_completion(&lnet_acceptor_state.pta_signal);
 491         rc = accept2secure(accept_type, &secure);
 492         if (rc <= 0)
 493                 return rc;
 494
 495         if (lnet_count_acceptor_nets() == 0)  /* not required */
 496                 return 0;
 497         if (current->nsproxy && current->nsproxy->net_ns)
 498                 lnet_acceptor_state.pta_ns = current->nsproxy->net_ns;
 499         else
 500                 lnet_acceptor_state.pta_ns = &init_net;
 501         task = kthread_run(lnet_acceptor, (void *)(uintptr_t)secure,
 502                            "acceptor_%03ld", secure);
 503         if (IS_ERR(task)) {
 504                 rc2 = PTR_ERR(task);
 505                 CERROR("Can't start acceptor thread: %ld\n", rc2);
 506                 return -ESRCH;
 507         }
 508
 509         /* wait for acceptor to startup */
 510         wait_for_completion(&lnet_acceptor_state.pta_signal);
 511
 512         if (!lnet_acceptor_state.pta_shutdown) {
 513                 /* started OK */
 514                 LASSERT(lnet_acceptor_state.pta_sock != NULL);
 515                 return 0;
 516         }
 517
 518         LASSERT(lnet_acceptor_state.pta_sock == NULL);
 519
 520         return -ENETDOWN;
 521 }
 522
 523 void
 524 lnet_acceptor_stop(void)
 525 {
 526         if (lnet_acceptor_state.pta_shutdown) /* not running */
 527                 return;
 528
 529         lnet_acceptor_state.pta_shutdown = 1;
 530         wake_up(&lnet_acceptor_state.pta_waitq);
 531
 532         /* block until acceptor signals exit */
 533         wait_for_completion(&lnet_acceptor_state.pta_signal);
 534 }