lnet/lnet/api-ni.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38 #include <lnet/lib-lnet.h>
  39
  40 #ifdef __KERNEL__
  41 #define D_LNI D_CONSOLE
  42 #else
  43 #define D_LNI D_CONFIG
  44 #endif
  45
  46 lnet_t      the_lnet;                           /* THE state of the network */
  47
  48 #ifdef __KERNEL__
  49
  50 static char *ip2nets = "";
  51 CFS_MODULE_PARM(ip2nets, "s", charp, 0444,
  52                 "LNET network <- IP table");
  53
  54 static char *networks = "";
  55 CFS_MODULE_PARM(networks, "s", charp, 0444,
  56                 "local networks");
  57
  58 static char *routes = "";
  59 CFS_MODULE_PARM(routes, "s", charp, 0444,
  60                 "routes to non-local networks");
  61
  62 static char *portals_compatibility = "none";
  63 CFS_MODULE_PARM(portals_compatibility, "s", charp, 0444,
  64                 "wire protocol compatibility: 'strong'|'weak'|'none'");
  65
  66 char *
  67 lnet_get_routes(void)
  68 {
  69         return routes;
  70 }
  71
  72 char *
  73 lnet_get_networks(void)
  74 {
  75         char   *nets;
  76         int     rc;
  77
  78         if (*networks != 0 && *ip2nets != 0) {
  79                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  80                                    "'ip2nets' but not both at once\n");
  81                 return NULL;
  82         }
  83
  84         if (*ip2nets != 0) {
  85                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  86                 return (rc == 0) ? nets : NULL;
  87         }
  88
  89         if (*networks != 0)
  90                 return networks;
  91
  92         return "tcp";
  93 }
  94
  95 int
  96 lnet_get_portals_compatibility(void)
  97 {
  98         if (!strcmp(portals_compatibility, "none")) {
  99                 return 0;
 100         }
 101
 102         if (!strcmp(portals_compatibility, "weak")) {
 103                 LCONSOLE_WARN("Starting in weak portals-compatible mode\n");
 104                 return 1;
 105         }
 106
 107         if (!strcmp(portals_compatibility, "strong")) {
 108                 LCONSOLE_WARN("Starting in strong portals-compatible mode\n");
 109                 return 2;
 110         }
 111
 112         LCONSOLE_ERROR_MSG(0x102, "portals_compatibility=\"%s\" not supported\n",
 113                            portals_compatibility);
 114         return -EINVAL;
 115 }
 116
 117 void
 118 lnet_init_locks(void)
 119 {
 120         spin_lock_init (&the_lnet.ln_lock);
 121         cfs_waitq_init (&the_lnet.ln_waitq);
 122         init_mutex(&the_lnet.ln_lnd_mutex);
 123         init_mutex(&the_lnet.ln_api_mutex);
 124 }
 125
 126 void
 127 lnet_fini_locks(void)
 128 {
 129 }
 130
 131 #else
 132
 133 char *
 134 lnet_get_routes(void)
 135 {
 136         char *str = getenv("LNET_ROUTES");
 137
 138         return (str == NULL) ? "" : str;
 139 }
 140
 141 char *
 142 lnet_get_networks (void)
 143 {
 144         static char       default_networks[256];
 145         char             *networks = getenv ("LNET_NETWORKS");
 146         char             *ip2nets  = getenv ("LNET_IP2NETS");
 147         char             *str;
 148         char             *sep;
 149         int               len;
 150         int               nob;
 151         int               rc;
 152         struct list_head *tmp;
 153
 154 #ifdef NOT_YET
 155         if (networks != NULL && ip2nets != NULL) {
 156                 LCONSOLE_ERROR_MSG(0x103, "Please set EITHER 'LNET_NETWORKS' or"
 157                                    " 'LNET_IP2NETS' but not both at once\n");
 158                 return NULL;
 159         }
 160
 161         if (ip2nets != NULL) {
 162                 rc = lnet_parse_ip2nets(&networks, ip2nets);
 163                 return (rc == 0) ? networks : NULL;
 164         }
 165 #else
 166         ip2nets = NULL;
 167         rc = 0;
 168 #endif
 169         if (networks != NULL)
 170                 return networks;
 171
 172         /* In userland, the default 'networks=' is the list of known net types */
 173
 174         len = sizeof(default_networks);
 175         str = default_networks;
 176         *str = 0;
 177         sep = "";
 178
 179         list_for_each (tmp, &the_lnet.ln_lnds) {
 180                 lnd_t *lnd = list_entry(tmp, lnd_t, lnd_list);
 181
 182                 nob = snprintf(str, len, "%s%s", sep,
 183                                libcfs_lnd2str(lnd->lnd_type));
 184                 len -= nob;
 185                 if (len < 0) {
 186                         /* overflowed the string; leave it where it was */
 187                         *str = 0;
 188                         break;
 189                 }
 190
 191                 str += nob;
 192                 sep = ",";
 193         }
 194
 195         return default_networks;
 196 }
 197
 198 int
 199 lnet_get_portals_compatibility(void)
 200 {
 201         return 0;
 202 }
 203
 204 # ifndef HAVE_LIBPTHREAD
 205
 206 void lnet_init_locks(void)
 207 {
 208         the_lnet.ln_lock = 0;
 209         the_lnet.ln_lnd_mutex = 0;
 210         the_lnet.ln_api_mutex = 0;
 211 }
 212
 213 void lnet_fini_locks(void)
 214 {
 215         LASSERT (the_lnet.ln_api_mutex == 0);
 216         LASSERT (the_lnet.ln_lnd_mutex == 0);
 217         LASSERT (the_lnet.ln_lock == 0);
 218 }
 219
 220 # else
 221
 222 void lnet_init_locks(void)
 223 {
 224         pthread_cond_init(&the_lnet.ln_cond, NULL);
 225         pthread_mutex_init(&the_lnet.ln_lock, NULL);
 226         pthread_mutex_init(&the_lnet.ln_lnd_mutex, NULL);
 227         pthread_mutex_init(&the_lnet.ln_api_mutex, NULL);
 228 }
 229
 230 void lnet_fini_locks(void)
 231 {
 232         pthread_mutex_destroy(&the_lnet.ln_api_mutex);
 233         pthread_mutex_destroy(&the_lnet.ln_lnd_mutex);
 234         pthread_mutex_destroy(&the_lnet.ln_lock);
 235         pthread_cond_destroy(&the_lnet.ln_cond);
 236 }
 237
 238 # endif
 239 #endif
 240
 241 void lnet_assert_wire_constants (void)
 242 {
 243         /* Wire protocol assertions generated by 'wirecheck'
 244          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 245          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 246          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 247
 248         /* Constants... */
 249         CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 250         CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
 251         CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
 252         CLASSERT (LNET_MSG_ACK == 0);
 253         CLASSERT (LNET_MSG_PUT == 1);
 254         CLASSERT (LNET_MSG_GET == 2);
 255         CLASSERT (LNET_MSG_REPLY == 3);
 256         CLASSERT (LNET_MSG_HELLO == 4);
 257
 258         /* Checks for struct ptl_handle_wire_t */
 259         CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
 260         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 261         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 262         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 263         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 264
 265         /* Checks for struct lnet_magicversion_t */
 266         CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
 267         CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
 268         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 269         CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
 270         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 271         CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 272         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 273
 274         /* Checks for struct lnet_hdr_t */
 275         CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
 276         CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 277         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 278         CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
 279         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 280         CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 281         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 282         CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
 283         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 284         CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
 285         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 286         CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
 287         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 288         CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
 289         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 290
 291         /* Ack */
 292         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 293         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 294         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 295         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 296         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 297         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 298
 299         /* Put */
 300         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 301         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 302         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 303         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 304         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 305         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 306         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 307         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 308         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 309         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 310
 311         /* Get */
 312         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 313         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 314         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 315         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 316         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 317         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 318         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 319         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 320         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 321         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 322
 323         /* Reply */
 324         CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 325         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 326
 327         /* Hello */
 328         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 329         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 330         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 331         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 332 }
 333
 334 lnd_t *
 335 lnet_find_lnd_by_type (int type)
 336 {
 337         lnd_t              *lnd;
 338         struct list_head   *tmp;
 339
 340         /* holding lnd mutex */
 341         list_for_each (tmp, &the_lnet.ln_lnds) {
 342                 lnd = list_entry(tmp, lnd_t, lnd_list);
 343
 344                 if (lnd->lnd_type == type)
 345                         return lnd;
 346         }
 347
 348         return NULL;
 349 }
 350
 351 void
 352 lnet_register_lnd (lnd_t *lnd)
 353 {
 354         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
 355
 356         LASSERT (the_lnet.ln_init);
 357         LASSERT (libcfs_isknown_lnd(lnd->lnd_type));
 358         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 359
 360         list_add_tail (&lnd->lnd_list, &the_lnet.ln_lnds);
 361         lnd->lnd_refcount = 0;
 362
 363         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 364
 365         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
 366 }
 367
 368 void
 369 lnet_unregister_lnd (lnd_t *lnd)
 370 {
 371         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
 372
 373         LASSERT (the_lnet.ln_init);
 374         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 375         LASSERT (lnd->lnd_refcount == 0);
 376
 377         list_del (&lnd->lnd_list);
 378         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 379
 380         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
 381 }
 382
 383 #ifndef LNET_USE_LIB_FREELIST
 384
 385 int
 386 lnet_descriptor_setup (void)
 387 {
 388         return 0;
 389 }
 390
 391 void
 392 lnet_descriptor_cleanup (void)
 393 {
 394 }
 395
 396 #else
 397
 398 int
 399 lnet_freelist_init (lnet_freelist_t *fl, int n, int size)
 400 {
 401         char *space;
 402
 403         LASSERT (n > 0);
 404
 405         size += offsetof (lnet_freeobj_t, fo_contents);
 406
 407         LIBCFS_ALLOC(space, n * size);
 408         if (space == NULL)
 409                 return (-ENOMEM);
 410
 411         CFS_INIT_LIST_HEAD (&fl->fl_list);
 412         fl->fl_objs = space;
 413         fl->fl_nobjs = n;
 414         fl->fl_objsize = size;
 415
 416         do
 417         {
 418                 memset (space, 0, size);
 419                 list_add ((struct list_head *)space, &fl->fl_list);
 420                 space += size;
 421         } while (--n != 0);
 422
 423         return (0);
 424 }
 425
 426 void
 427 lnet_freelist_fini (lnet_freelist_t *fl)
 428 {
 429         struct list_head *el;
 430         int               count;
 431
 432         if (fl->fl_nobjs == 0)
 433                 return;
 434
 435         count = 0;
 436         for (el = fl->fl_list.next; el != &fl->fl_list; el = el->next)
 437                 count++;
 438
 439         LASSERT (count == fl->fl_nobjs);
 440
 441         LIBCFS_FREE(fl->fl_objs, fl->fl_nobjs * fl->fl_objsize);
 442         memset (fl, 0, sizeof (*fl));
 443 }
 444
 445 int
 446 lnet_descriptor_setup (void)
 447 {
 448         /* NB on failure caller must still call lnet_descriptor_cleanup */
 449         /*               ******                                         */
 450         int        rc;
 451
 452         memset (&the_lnet.ln_free_mes,  0, sizeof (the_lnet.ln_free_mes));
 453         memset (&the_lnet.ln_free_msgs, 0, sizeof (the_lnet.ln_free_msgs));
 454         memset (&the_lnet.ln_free_mds,  0, sizeof (the_lnet.ln_free_mds));
 455         memset (&the_lnet.ln_free_eqs,  0, sizeof (the_lnet.ln_free_eqs));
 456
 457         rc = lnet_freelist_init(&the_lnet.ln_free_mes,
 458                                 MAX_MES, sizeof (lnet_me_t));
 459         if (rc != 0)
 460                 return (rc);
 461
 462         rc = lnet_freelist_init(&the_lnet.ln_free_msgs,
 463                                 MAX_MSGS, sizeof (lnet_msg_t));
 464         if (rc != 0)
 465                 return (rc);
 466
 467         rc = lnet_freelist_init(&the_lnet.ln_free_mds,
 468                                 MAX_MDS, sizeof (lnet_libmd_t));
 469         if (rc != 0)
 470                 return (rc);
 471
 472         rc = lnet_freelist_init(&the_lnet.ln_free_eqs,
 473                                 MAX_EQS, sizeof (lnet_eq_t));
 474         return (rc);
 475 }
 476
 477 void
 478 lnet_descriptor_cleanup (void)
 479 {
 480         lnet_freelist_fini (&the_lnet.ln_free_mes);
 481         lnet_freelist_fini (&the_lnet.ln_free_msgs);
 482         lnet_freelist_fini (&the_lnet.ln_free_mds);
 483         lnet_freelist_fini (&the_lnet.ln_free_eqs);
 484 }
 485
 486 #endif
 487
 488 __u64
 489 lnet_create_interface_cookie (void)
 490 {
 491         /* NB the interface cookie in wire handles guards against delayed
 492          * replies and ACKs appearing valid after reboot. Initialisation time,
 493          * even if it's only implemented to millisecond resolution is probably
 494          * easily good enough. */
 495         struct timeval tv;
 496         __u64          cookie;
 497 #ifndef __KERNEL__
 498         int            rc = gettimeofday (&tv, NULL);
 499         LASSERT (rc == 0);
 500 #else
 501         do_gettimeofday(&tv);
 502 #endif
 503         cookie = tv.tv_sec;
 504         cookie *= 1000000;
 505         cookie += tv.tv_usec;
 506         return cookie;
 507 }
 508
 509 int
 510 lnet_setup_handle_hash (void)
 511 {
 512         int       i;
 513
 514         /* Arbitrary choice of hash table size */
 515 #ifdef __KERNEL__
 516         the_lnet.ln_lh_hash_size = CFS_PAGE_SIZE / sizeof (struct list_head);
 517 #else
 518         the_lnet.ln_lh_hash_size = (MAX_MES + MAX_MDS + MAX_EQS)/4;
 519 #endif
 520         LIBCFS_ALLOC(the_lnet.ln_lh_hash_table,
 521                      the_lnet.ln_lh_hash_size * sizeof (struct list_head));
 522         if (the_lnet.ln_lh_hash_table == NULL)
 523                 return (-ENOMEM);
 524
 525         for (i = 0; i < the_lnet.ln_lh_hash_size; i++)
 526                 CFS_INIT_LIST_HEAD (&the_lnet.ln_lh_hash_table[i]);
 527
 528         the_lnet.ln_next_object_cookie = LNET_COOKIE_TYPES;
 529
 530         return (0);
 531 }
 532
 533 void
 534 lnet_cleanup_handle_hash (void)
 535 {
 536         if (the_lnet.ln_lh_hash_table == NULL)
 537                 return;
 538
 539         LIBCFS_FREE(the_lnet.ln_lh_hash_table,
 540                     the_lnet.ln_lh_hash_size * sizeof (struct list_head));
 541 }
 542
 543 lnet_libhandle_t *
 544 lnet_lookup_cookie (__u64 cookie, int type)
 545 {
 546         /* ALWAYS called with LNET_LOCK held */
 547         struct list_head    *list;
 548         struct list_head    *el;
 549         unsigned int         hash;
 550
 551         if ((cookie & (LNET_COOKIE_TYPES - 1)) != type)
 552                 return (NULL);
 553
 554         hash = ((unsigned int)cookie) % the_lnet.ln_lh_hash_size;
 555         list = &the_lnet.ln_lh_hash_table[hash];
 556
 557         list_for_each (el, list) {
 558                 lnet_libhandle_t *lh = list_entry (el, lnet_libhandle_t,
 559                                                   lh_hash_chain);
 560
 561                 if (lh->lh_cookie == cookie)
 562                         return (lh);
 563         }
 564
 565         return (NULL);
 566 }
 567
 568 void
 569 lnet_initialise_handle (lnet_libhandle_t *lh, int type)
 570 {
 571         /* ALWAYS called with LNET_LOCK held */
 572         unsigned int    hash;
 573
 574         LASSERT (type >= 0 && type < LNET_COOKIE_TYPES);
 575         lh->lh_cookie = the_lnet.ln_next_object_cookie | type;
 576         the_lnet.ln_next_object_cookie += LNET_COOKIE_TYPES;
 577
 578         hash = ((unsigned int)lh->lh_cookie) % the_lnet.ln_lh_hash_size;
 579         list_add (&lh->lh_hash_chain, &the_lnet.ln_lh_hash_table[hash]);
 580 }
 581
 582 void
 583 lnet_invalidate_handle (lnet_libhandle_t *lh)
 584 {
 585         /* ALWAYS called with LNET_LOCK held */
 586         list_del (&lh->lh_hash_chain);
 587 }
 588
 589 int
 590 lnet_init_finalizers(void)
 591 {
 592 #ifdef __KERNEL__
 593         int    i;
 594
 595         the_lnet.ln_nfinalizers = num_online_cpus();
 596
 597         LIBCFS_ALLOC(the_lnet.ln_finalizers,
 598                      the_lnet.ln_nfinalizers *
 599                      sizeof(*the_lnet.ln_finalizers));
 600         if (the_lnet.ln_finalizers == NULL) {
 601                 CERROR("Can't allocate ln_finalizers\n");
 602                 return -ENOMEM;
 603         }
 604
 605         for (i = 0; i < the_lnet.ln_nfinalizers; i++)
 606                 the_lnet.ln_finalizers[i] = NULL;
 607 #else
 608         the_lnet.ln_finalizing = 0;
 609 #endif
 610
 611         CFS_INIT_LIST_HEAD(&the_lnet.ln_finalizeq);
 612         return 0;
 613 }
 614
 615 void
 616 lnet_fini_finalizers(void)
 617 {
 618 #ifdef __KERNEL__
 619         int    i;
 620
 621         for (i = 0; i < the_lnet.ln_nfinalizers; i++)
 622                 LASSERT (the_lnet.ln_finalizers[i] == NULL);
 623
 624         LIBCFS_FREE(the_lnet.ln_finalizers,
 625                     the_lnet.ln_nfinalizers *
 626                     sizeof(*the_lnet.ln_finalizers));
 627 #else
 628         LASSERT (!the_lnet.ln_finalizing);
 629 #endif
 630         LASSERT (list_empty(&the_lnet.ln_finalizeq));
 631 }
 632
 633 #ifndef __KERNEL__
 634 /* Temporary workaround to allow uOSS and test programs force server
 635  * mode in userspace. See comments near ln_server_mode_flag in
 636  * lnet/lib-types.h */
 637
 638 void
 639 lnet_server_mode() {
 640         the_lnet.ln_server_mode_flag = 1;
 641 }
 642 #endif
 643
 644 int
 645 lnet_prepare(lnet_pid_t requested_pid)
 646 {
 647         /* Prepare to bring up the network */
 648         int               rc = 0;
 649         int               i;
 650
 651         LASSERT (the_lnet.ln_refcount == 0);
 652
 653         the_lnet.ln_routing = 0;
 654
 655 #ifdef __KERNEL__
 656         LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 657         the_lnet.ln_pid = requested_pid;
 658 #else
 659         if (the_lnet.ln_server_mode_flag) {/* server case (uOSS) */
 660                 LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 661
 662                 if (cfs_curproc_uid())/* Only root can run user-space server */
 663                         return -EPERM;
 664                 the_lnet.ln_pid = requested_pid;
 665
 666         } else {/* client case (liblustre) */
 667
 668                 /* My PID must be unique on this node and flag I'm userspace */
 669                 the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
 670         }
 671 #endif
 672
 673         rc = lnet_descriptor_setup();
 674         if (rc != 0)
 675                 goto failed0;
 676
 677         memset(&the_lnet.ln_counters, 0,
 678                sizeof(the_lnet.ln_counters));
 679
 680         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_msgs);
 681         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_mds);
 682         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_eqs);
 683         CFS_INIT_LIST_HEAD (&the_lnet.ln_test_peers);
 684         CFS_INIT_LIST_HEAD (&the_lnet.ln_nis);
 685         CFS_INIT_LIST_HEAD (&the_lnet.ln_zombie_nis);
 686         CFS_INIT_LIST_HEAD (&the_lnet.ln_remote_nets);
 687         CFS_INIT_LIST_HEAD (&the_lnet.ln_routers);
 688
 689         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 690
 691         lnet_init_rtrpools();
 692
 693         rc = lnet_setup_handle_hash ();
 694         if (rc != 0)
 695                 goto failed0;
 696
 697         rc = lnet_create_peer_table();
 698         if (rc != 0)
 699                 goto failed1;
 700
 701         rc = lnet_init_finalizers();
 702         if (rc != 0)
 703                 goto failed2;
 704
 705         the_lnet.ln_nportals = MAX_PORTALS;
 706         LIBCFS_ALLOC(the_lnet.ln_portals,
 707                      the_lnet.ln_nportals *
 708                      sizeof(*the_lnet.ln_portals));
 709         if (the_lnet.ln_portals == NULL) {
 710                 rc = -ENOMEM;
 711                 goto failed3;
 712         }
 713
 714         for (i = 0; i < the_lnet.ln_nportals; i++) {
 715                 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_ml));
 716                 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_msgq));
 717                 the_lnet.ln_portals[i].ptl_options = 0;
 718         }
 719
 720         return 0;
 721
 722  failed3:
 723         lnet_fini_finalizers();
 724  failed2:
 725         lnet_destroy_peer_table();
 726  failed1:
 727         lnet_cleanup_handle_hash();
 728  failed0:
 729         lnet_descriptor_cleanup();
 730         return rc;
 731 }
 732
 733 int
 734 lnet_unprepare (void)
 735 {
 736         int       idx;
 737
 738         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 739          * have shut down already, so it is safe to unlink and free all
 740          * descriptors, even those that appear committed to a network op (eg MD
 741          * with non-zero pending count) */
 742
 743         lnet_fail_nid(LNET_NID_ANY, 0);
 744
 745         LASSERT (list_empty(&the_lnet.ln_test_peers));
 746         LASSERT (the_lnet.ln_refcount == 0);
 747         LASSERT (list_empty(&the_lnet.ln_nis));
 748         LASSERT (list_empty(&the_lnet.ln_zombie_nis));
 749         LASSERT (the_lnet.ln_nzombie_nis == 0);
 750
 751         for (idx = 0; idx < the_lnet.ln_nportals; idx++) {
 752                 LASSERT (list_empty(&the_lnet.ln_portals[idx].ptl_msgq));
 753
 754                 while (!list_empty (&the_lnet.ln_portals[idx].ptl_ml)) {
 755                         lnet_me_t *me = list_entry (the_lnet.ln_portals[idx].ptl_ml.next,
 756                                                     lnet_me_t, me_list);
 757
 758                         CERROR ("Active me %p on exit\n", me);
 759                         list_del (&me->me_list);
 760                         lnet_me_free (me);
 761                 }
 762         }
 763
 764         while (!list_empty (&the_lnet.ln_active_mds)) {
 765                 lnet_libmd_t *md = list_entry (the_lnet.ln_active_mds.next,
 766                                                lnet_libmd_t, md_list);
 767
 768                 CERROR ("Active md %p on exit\n", md);
 769                 list_del_init (&md->md_list);
 770                 lnet_md_free (md);
 771         }
 772
 773         while (!list_empty (&the_lnet.ln_active_eqs)) {
 774                 lnet_eq_t *eq = list_entry (the_lnet.ln_active_eqs.next,
 775                                             lnet_eq_t, eq_list);
 776
 777                 CERROR ("Active eq %p on exit\n", eq);
 778                 list_del (&eq->eq_list);
 779                 lnet_eq_free (eq);
 780         }
 781
 782         while (!list_empty (&the_lnet.ln_active_msgs)) {
 783                 lnet_msg_t *msg = list_entry (the_lnet.ln_active_msgs.next,
 784                                               lnet_msg_t, msg_activelist);
 785
 786                 CERROR ("Active msg %p on exit\n", msg);
 787                 LASSERT (msg->msg_onactivelist);
 788                 msg->msg_onactivelist = 0;
 789                 list_del (&msg->msg_activelist);
 790                 lnet_msg_free (msg);
 791         }
 792
 793         LIBCFS_FREE(the_lnet.ln_portals,
 794                     the_lnet.ln_nportals * sizeof(*the_lnet.ln_portals));
 795
 796         lnet_free_rtrpools();
 797         lnet_fini_finalizers();
 798         lnet_destroy_peer_table();
 799         lnet_cleanup_handle_hash();
 800         lnet_descriptor_cleanup();
 801
 802         return (0);
 803 }
 804
 805 lnet_ni_t  *
 806 lnet_net2ni_locked (__u32 net)
 807 {
 808         struct list_head *tmp;
 809         lnet_ni_t        *ni;
 810
 811         list_for_each (tmp, &the_lnet.ln_nis) {
 812                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 813
 814                 if (lnet_ptlcompat_matchnet(LNET_NIDNET(ni->ni_nid), net)) {
 815                         lnet_ni_addref_locked(ni);
 816                         return ni;
 817                 }
 818         }
 819
 820         return NULL;
 821 }
 822
 823 int
 824 lnet_islocalnet (__u32 net)
 825 {
 826         lnet_ni_t        *ni;
 827
 828         LNET_LOCK();
 829         ni = lnet_net2ni_locked(net);
 830         if (ni != NULL)
 831                 lnet_ni_decref_locked(ni);
 832         LNET_UNLOCK();
 833
 834         return ni != NULL;
 835 }
 836
 837 lnet_ni_t  *
 838 lnet_nid2ni_locked (lnet_nid_t nid)
 839 {
 840         struct list_head *tmp;
 841         lnet_ni_t        *ni;
 842
 843         list_for_each (tmp, &the_lnet.ln_nis) {
 844                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 845
 846                 if (lnet_ptlcompat_matchnid(ni->ni_nid, nid)) {
 847                         lnet_ni_addref_locked(ni);
 848                         return ni;
 849                 }
 850         }
 851
 852         return NULL;
 853 }
 854
 855 int
 856 lnet_islocalnid (lnet_nid_t nid)
 857 {
 858         lnet_ni_t     *ni;
 859
 860         LNET_LOCK();
 861         ni = lnet_nid2ni_locked(nid);
 862         if (ni != NULL)
 863                 lnet_ni_decref_locked(ni);
 864         LNET_UNLOCK();
 865
 866         return ni != NULL;
 867 }
 868
 869 int
 870 lnet_count_acceptor_nis (lnet_ni_t **first_ni)
 871 {
 872         /* Return the # of NIs that need the acceptor.  Return the first one in
 873          * *first_ni so the acceptor can pass it connections "blind" to retain
 874          * binary compatibility. */
 875         int                count = 0;
 876 #if defined(__KERNEL__) || defined(HAVE_LIBPTHREAD)
 877         struct list_head  *tmp;
 878         lnet_ni_t         *ni;
 879
 880         LNET_LOCK();
 881         list_for_each (tmp, &the_lnet.ln_nis) {
 882                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 883
 884                 if (ni->ni_lnd->lnd_accept != NULL) {
 885                         /* This LND uses the acceptor */
 886                         if (count == 0 && first_ni != NULL) {
 887                                 lnet_ni_addref_locked(ni);
 888                                 *first_ni = ni;
 889                         }
 890                         count++;
 891                 }
 892         }
 893
 894         LNET_UNLOCK();
 895
 896 #endif /* defined(__KERNEL__) || defined(HAVE_LIBPTHREAD) */
 897         return count;
 898 }
 899
 900 void
 901 lnet_shutdown_lndnis (void)
 902 {
 903         int                i;
 904         int                islo;
 905         lnet_ni_t         *ni;
 906
 907         /* NB called holding the global mutex */
 908
 909         /* All quiet on the API front */
 910         LASSERT (!the_lnet.ln_shutdown);
 911         LASSERT (the_lnet.ln_refcount == 0);
 912         LASSERT (list_empty(&the_lnet.ln_zombie_nis));
 913         LASSERT (the_lnet.ln_nzombie_nis == 0);
 914         LASSERT (list_empty(&the_lnet.ln_remote_nets));
 915
 916         LNET_LOCK();
 917         the_lnet.ln_shutdown = 1;               /* flag shutdown */
 918
 919         /* Unlink NIs from the global table */
 920         while (!list_empty(&the_lnet.ln_nis)) {
 921                 ni = list_entry(the_lnet.ln_nis.next,
 922                                 lnet_ni_t, ni_list);
 923                 list_del (&ni->ni_list);
 924
 925                 the_lnet.ln_nzombie_nis++;
 926                 lnet_ni_decref_locked(ni); /* drop apini's ref */
 927         }
 928
 929         /* Drop the cached eqwait NI. */
 930         if (the_lnet.ln_eqwaitni != NULL) {
 931                 lnet_ni_decref_locked(the_lnet.ln_eqwaitni);
 932                 the_lnet.ln_eqwaitni = NULL;
 933         }
 934
 935         /* Drop the cached loopback NI. */
 936         if (the_lnet.ln_loni != NULL) {
 937                 lnet_ni_decref_locked(the_lnet.ln_loni);
 938                 the_lnet.ln_loni = NULL;
 939         }
 940
 941         LNET_UNLOCK();
 942
 943         /* Clear lazy portals and drop delayed messages which hold refs
 944          * on their lnet_msg_t::msg_rxpeer */
 945         for (i = 0; i < the_lnet.ln_nportals; i++)
 946                 LNetClearLazyPortal(i);
 947
 948         /* Clear the peer table and wait for all peers to go (they hold refs on
 949          * their NIs) */
 950         lnet_clear_peer_table();
 951
 952         LNET_LOCK();
 953         /* Now wait for the NI's I just nuked to show up on apini_zombie_nis
 954          * and shut them down in guaranteed thread context */
 955         i = 2;
 956         while (the_lnet.ln_nzombie_nis != 0) {
 957
 958                 while (list_empty(&the_lnet.ln_zombie_nis)) {
 959                         LNET_UNLOCK();
 960                         ++i;
 961                         if ((i & (-i)) == i)
 962                                 CDEBUG(D_WARNING,"Waiting for %d zombie NIs\n",
 963                                        the_lnet.ln_nzombie_nis);
 964                         cfs_pause(cfs_time_seconds(1));
 965                         LNET_LOCK();
 966                 }
 967
 968                 ni = list_entry(the_lnet.ln_zombie_nis.next,
 969                                 lnet_ni_t, ni_list);
 970                 list_del(&ni->ni_list);
 971                 ni->ni_lnd->lnd_refcount--;
 972
 973                 LNET_UNLOCK();
 974
 975                 islo = ni->ni_lnd->lnd_type == LOLND;
 976
 977                 LASSERT (!in_interrupt ());
 978                 (ni->ni_lnd->lnd_shutdown)(ni);
 979
 980                 /* can't deref lnd anymore now; it might have unregistered
 981                  * itself...  */
 982
 983                 if (!islo)
 984                         CDEBUG(D_LNI, "Removed LNI %s\n",
 985                                libcfs_nid2str(ni->ni_nid));
 986
 987                 LIBCFS_FREE(ni, sizeof(*ni));
 988
 989                 LNET_LOCK();
 990                 the_lnet.ln_nzombie_nis--;
 991         }
 992
 993         the_lnet.ln_shutdown = 0;
 994         LNET_UNLOCK();
 995
 996         if (the_lnet.ln_network_tokens != NULL) {
 997                 LIBCFS_FREE(the_lnet.ln_network_tokens,
 998                             the_lnet.ln_network_tokens_nob);
 999                 the_lnet.ln_network_tokens = NULL;
1000         }
1001 }
1002
1003 int
1004 lnet_startup_lndnis (void)
1005 {
1006         lnd_t             *lnd;
1007         lnet_ni_t         *ni;
1008         struct list_head   nilist;
1009         int                rc = 0;
1010         int                lnd_type;
1011         int                nicount = 0;
1012         char              *nets = lnet_get_networks();
1013
1014         CFS_INIT_LIST_HEAD(&nilist);
1015
1016         if (nets == NULL)
1017                 goto failed;
1018
1019         rc = lnet_parse_networks(&nilist, nets);
1020         if (rc != 0)
1021                 goto failed;
1022
1023         while (!list_empty(&nilist)) {
1024                 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
1025                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
1026
1027                 LASSERT (libcfs_isknown_lnd(lnd_type));
1028
1029                 LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
1030                 lnd = lnet_find_lnd_by_type(lnd_type);
1031
1032 #ifdef __KERNEL__
1033                 if (lnd == NULL) {
1034                         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1035                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
1036                         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
1037
1038                         lnd = lnet_find_lnd_by_type(lnd_type);
1039                         if (lnd == NULL) {
1040                                 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1041                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1042                                        libcfs_lnd2str(lnd_type),
1043                                        libcfs_lnd2modname(lnd_type), rc);
1044 #ifndef CONFIG_KMOD
1045                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1046                                          "compiled with CONFIG_KMOD set for "
1047                                          "automatic module loading.");
1048 #endif
1049                                 goto failed;
1050                         }
1051                 }
1052 #else
1053                 if (lnd == NULL) {
1054                         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1055                         CERROR("LND %s not supported\n",
1056                                libcfs_lnd2str(lnd_type));
1057                         goto failed;
1058                 }
1059 #endif
1060
1061                 ni->ni_refcount = 1;
1062
1063                 LNET_LOCK();
1064                 lnd->lnd_refcount++;
1065                 LNET_UNLOCK();
1066
1067                 ni->ni_lnd = lnd;
1068
1069                 rc = (lnd->lnd_startup)(ni);
1070
1071                 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1072
1073                 if (rc != 0) {
1074                         LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
1075                                            "\n",
1076                                            rc, libcfs_lnd2str(lnd->lnd_type));
1077                         LNET_LOCK();
1078                         lnd->lnd_refcount--;
1079                         LNET_UNLOCK();
1080                         goto failed;
1081                 }
1082
1083                 LASSERT (ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1084
1085                 list_del(&ni->ni_list);
1086
1087                 LNET_LOCK();
1088                 list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1089                 LNET_UNLOCK();
1090
1091                 if (lnd->lnd_type == LOLND) {
1092                         lnet_ni_addref(ni);
1093                         LASSERT (the_lnet.ln_loni == NULL);
1094                         the_lnet.ln_loni = ni;
1095                         continue;
1096                 }
1097
1098 #ifndef __KERNEL__
1099                 if (lnd->lnd_wait != NULL) {
1100                         if (the_lnet.ln_eqwaitni == NULL) {
1101                                 lnet_ni_addref(ni);
1102                                 the_lnet.ln_eqwaitni = ni;
1103                         }
1104                 } else {
1105 # ifndef HAVE_LIBPTHREAD
1106                         LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1107                                            "single-threaded runtime\n",
1108                                            libcfs_lnd2str(lnd_type));
1109                         goto failed;
1110 # endif
1111                 }
1112 #endif
1113                 if (ni->ni_peertxcredits == 0 ||
1114                     ni->ni_maxtxcredits == 0) {
1115                         LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1116                                            libcfs_lnd2str(lnd->lnd_type),
1117                                            ni->ni_peertxcredits == 0 ?
1118                                            "" : "per-peer ");
1119                         goto failed;
1120                 }
1121
1122                 ni->ni_txcredits = ni->ni_mintxcredits = ni->ni_maxtxcredits;
1123
1124                 CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1125                        libcfs_nid2str(ni->ni_nid),
1126                        ni->ni_peertxcredits, ni->ni_txcredits,
1127                        ni->ni_peerrtrcredits, ni->ni_peertimeout);
1128
1129                 /* Handle nidstrings for network 0 just like this one */
1130                 if (the_lnet.ln_ptlcompat > 0) {
1131                         if (nicount > 0) {
1132                                 LCONSOLE_ERROR_MSG(0x108, "Can't run > 1 "
1133                                        "network when portals_compatibility is "
1134                                        "set\n");
1135                                 goto failed;
1136                         }
1137                         libcfs_setnet0alias(lnd->lnd_type);
1138                 }
1139
1140                 nicount++;
1141         }
1142
1143         if (the_lnet.ln_eqwaitni != NULL && nicount > 1) {
1144                 lnd_type = the_lnet.ln_eqwaitni->ni_lnd->lnd_type;
1145                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1146                                    "\n",
1147                                    libcfs_lnd2str(lnd_type));
1148                 goto failed;
1149         }
1150
1151         return 0;
1152
1153  failed:
1154         lnet_shutdown_lndnis();
1155
1156         while (!list_empty(&nilist)) {
1157                 ni = list_entry(nilist.next, lnet_ni_t, ni_list);
1158                 list_del(&ni->ni_list);
1159                 LIBCFS_FREE(ni, sizeof(*ni));
1160         }
1161
1162         return -ENETDOWN;
1163 }
1164
1165 int
1166 LNetInit(void)
1167 {
1168         int    rc;
1169
1170         lnet_assert_wire_constants ();
1171         LASSERT (!the_lnet.ln_init);
1172
1173         memset(&the_lnet, 0, sizeof(the_lnet));
1174
1175         rc = lnet_get_portals_compatibility();
1176         if (rc < 0)
1177                 return rc;
1178
1179         lnet_init_locks();
1180         CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
1181         the_lnet.ln_ptlcompat = rc;
1182         the_lnet.ln_refcount = 0;
1183         the_lnet.ln_init = 1;
1184
1185 #ifdef __KERNEL__
1186         /* All LNDs apart from the LOLND are in separate modules.  They
1187          * register themselves when their module loads, and unregister
1188          * themselves when their module is unloaded. */
1189 #else
1190         /* Register LNDs
1191          * NB the order here determines default 'networks=' order */
1192 # ifdef CRAY_XT3
1193         LNET_REGISTER_ULND(the_ptllnd);
1194 # endif
1195 # ifdef HAVE_LIBPTHREAD
1196         LNET_REGISTER_ULND(the_tcplnd);
1197 # endif
1198 #endif
1199         lnet_register_lnd(&the_lolnd);
1200         return 0;
1201 }
1202
1203 void
1204 LNetFini(void)
1205 {
1206         LASSERT (the_lnet.ln_init);
1207         LASSERT (the_lnet.ln_refcount == 0);
1208
1209         while (!list_empty(&the_lnet.ln_lnds))
1210                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1211                                                lnd_t, lnd_list));
1212         lnet_fini_locks();
1213
1214         the_lnet.ln_init = 0;
1215 }
1216
1217 int
1218 LNetNIInit(lnet_pid_t requested_pid)
1219 {
1220         int         im_a_router = 0;
1221         int         rc;
1222
1223         LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1224
1225         LASSERT (the_lnet.ln_init);
1226         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1227
1228         if (the_lnet.ln_refcount > 0) {
1229                 rc = the_lnet.ln_refcount++;
1230                 goto out;
1231         }
1232
1233         lnet_get_tunables();
1234
1235         if (requested_pid == LNET_PID_ANY) {
1236                 /* Don't instantiate LNET just for me */
1237                 rc = -ENETDOWN;
1238                 goto failed0;
1239         }
1240
1241         rc = lnet_prepare(requested_pid);
1242         if (rc != 0)
1243                 goto failed0;
1244
1245         rc = lnet_startup_lndnis();
1246         if (rc != 0)
1247                 goto failed1;
1248
1249         rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1250         if (rc != 0)
1251                 goto failed2;
1252
1253         rc = lnet_check_routes();
1254         if (rc != 0)
1255                 goto failed2;
1256
1257         rc = lnet_alloc_rtrpools(im_a_router);
1258         if (rc != 0)
1259                 goto failed2;
1260
1261         rc = lnet_acceptor_start();
1262         if (rc != 0)
1263                 goto failed2;
1264
1265         the_lnet.ln_refcount = 1;
1266         /* Now I may use my own API functions... */
1267
1268         rc = lnet_router_checker_start();
1269         if (rc != 0)
1270                 goto failed3;
1271
1272         rc = lnet_ping_target_init();
1273         if (rc != 0)
1274                 goto failed4;
1275
1276         lnet_proc_init();
1277         goto out;
1278
1279  failed4:
1280         lnet_router_checker_stop();
1281  failed3:
1282         the_lnet.ln_refcount = 0;
1283         lnet_acceptor_stop();
1284  failed2:
1285         lnet_destroy_routes();
1286         lnet_shutdown_lndnis();
1287  failed1:
1288         lnet_unprepare();
1289  failed0:
1290         LASSERT (rc < 0);
1291  out:
1292         LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1293         return rc;
1294 }
1295
1296 int
1297 LNetNIFini()
1298 {
1299         LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1300
1301         LASSERT (the_lnet.ln_init);
1302         LASSERT (the_lnet.ln_refcount > 0);
1303
1304         if (the_lnet.ln_refcount != 1) {
1305                 the_lnet.ln_refcount--;
1306         } else {
1307                 LASSERT (!the_lnet.ln_niinit_self);
1308
1309                 lnet_proc_fini();
1310                 lnet_ping_target_fini();
1311                 lnet_router_checker_stop();
1312
1313                 /* Teardown fns that use my own API functions BEFORE here */
1314                 the_lnet.ln_refcount = 0;
1315
1316                 lnet_acceptor_stop();
1317                 lnet_destroy_routes();
1318                 lnet_shutdown_lndnis();
1319                 lnet_unprepare();
1320         }
1321
1322         LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1323         return 0;
1324 }
1325
1326 int
1327 LNetCtl(unsigned int cmd, void *arg)
1328 {
1329         struct libcfs_ioctl_data *data = arg;
1330         lnet_process_id_t         id;
1331         lnet_ni_t                *ni;
1332         int                       rc;
1333
1334         LASSERT (the_lnet.ln_init);
1335         LASSERT (the_lnet.ln_refcount > 0);
1336
1337         switch (cmd) {
1338         case IOC_LIBCFS_GET_NI:
1339                 rc = LNetGetId(data->ioc_count, &id);
1340                 data->ioc_nid = id.nid;
1341                 return rc;
1342
1343         case IOC_LIBCFS_FAIL_NID:
1344                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1345
1346         case IOC_LIBCFS_ADD_ROUTE:
1347                 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1348                                     data->ioc_nid);
1349                 return (rc != 0) ? rc : lnet_check_routes();
1350
1351         case IOC_LIBCFS_DEL_ROUTE:
1352                 return lnet_del_route(data->ioc_net, data->ioc_nid);
1353
1354         case IOC_LIBCFS_GET_ROUTE:
1355                 return lnet_get_route(data->ioc_count,
1356                                       &data->ioc_net, &data->ioc_count,
1357                                       &data->ioc_nid, &data->ioc_flags);
1358         case IOC_LIBCFS_NOTIFY_ROUTER:
1359                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1360                                    (time_t)data->ioc_u64[0]);
1361
1362         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1363                 return the_lnet.ln_ptlcompat;
1364
1365         case IOC_LIBCFS_LNET_DIST:
1366                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1367                 if (rc < 0 && rc != -EHOSTUNREACH)
1368                         return rc;
1369
1370                 data->ioc_u32[0] = rc;
1371                 return 0;
1372
1373         case IOC_LIBCFS_TESTPROTOCOMPAT:
1374                 LNET_LOCK();
1375                 the_lnet.ln_testprotocompat = data->ioc_flags;
1376                 LNET_UNLOCK();
1377                 return 0;
1378
1379         case IOC_LIBCFS_PING:
1380                 rc = lnet_ping((lnet_process_id_t) {.nid = data->ioc_nid,
1381                                                     .pid = data->ioc_u32[0]},
1382                                data->ioc_u32[1], /* timeout */
1383                                (lnet_process_id_t *)data->ioc_pbuf1,
1384                                data->ioc_plen1/sizeof(lnet_process_id_t));
1385                 if (rc < 0)
1386                         return rc;
1387                 data->ioc_count = rc;
1388                 return 0;
1389
1390         case IOC_LIBCFS_DEBUG_PEER: {
1391                 /* CAVEAT EMPTOR: this one designed for calling directly; not
1392                  * via an ioctl */
1393                 lnet_process_id_t *id = arg;
1394
1395                 lnet_debug_peer(id->nid);
1396
1397                 ni = lnet_net2ni(LNET_NIDNET(id->nid));
1398                 if (ni == NULL) {
1399                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(*id));
1400                 } else {
1401                         if (ni->ni_lnd->lnd_ctl == NULL) {
1402                                 CDEBUG(D_WARNING, "No ctl for %s\n",
1403                                        libcfs_id2str(*id));
1404                         } else {
1405                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1406                         }
1407
1408                         lnet_ni_decref(ni);
1409                 }
1410                 return 0;
1411         }
1412
1413         default:
1414                 ni = lnet_net2ni(data->ioc_net);
1415                 if (ni == NULL)
1416                         return -EINVAL;
1417
1418                 if (ni->ni_lnd->lnd_ctl == NULL)
1419                         rc = -EINVAL;
1420                 else
1421                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1422
1423                 lnet_ni_decref(ni);
1424                 return rc;
1425         }
1426         /* not reached */
1427 }
1428
1429 int
1430 LNetGetId(unsigned int index, lnet_process_id_t *id)
1431 {
1432         lnet_ni_t        *ni;
1433         struct list_head *tmp;
1434         int               rc = -ENOENT;
1435
1436         LASSERT (the_lnet.ln_init);
1437         LASSERT (the_lnet.ln_refcount > 0);
1438
1439         LNET_LOCK();
1440
1441         list_for_each(tmp, &the_lnet.ln_nis) {
1442                 if (index-- != 0)
1443                         continue;
1444
1445                 ni = list_entry(tmp, lnet_ni_t, ni_list);
1446
1447                 id->nid = ni->ni_nid;
1448                 id->pid = the_lnet.ln_pid;
1449                 rc = 0;
1450                 break;
1451         }
1452
1453         LNET_UNLOCK();
1454
1455         return rc;
1456 }
1457
1458 void
1459 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1460 {
1461         snprintf(str, len, LPX64, h.cookie);
1462 }
1463
1464
1465 int
1466 lnet_ping_target_init(void)
1467 {
1468         lnet_handle_me_t  meh;
1469         lnet_process_id_t id;
1470         int               rc;
1471         int               rc2;
1472         int               n;
1473         int               infosz;
1474         int               i;
1475
1476         for (n = 0; ; n++) {
1477                 rc = LNetGetId(n, &id);
1478                 if (rc == -ENOENT)
1479                         break;
1480
1481                 LASSERT (rc == 0);
1482         }
1483
1484         infosz = offsetof(lnet_ping_info_t, pi_nid[n]);
1485         LIBCFS_ALLOC(the_lnet.ln_ping_info, infosz);
1486         if (the_lnet.ln_ping_info == NULL) {
1487                 CERROR("Can't allocate ping info[%d]\n", n);
1488                 return -ENOMEM;
1489         }
1490
1491         the_lnet.ln_ping_info->pi_magic   = LNET_PROTO_PING_MAGIC;
1492         the_lnet.ln_ping_info->pi_version = LNET_PROTO_PING_VERSION;
1493         the_lnet.ln_ping_info->pi_pid     = the_lnet.ln_pid;
1494         the_lnet.ln_ping_info->pi_nnids   = n;
1495
1496         for (i = 0; i < n; i++) {
1497                 rc = LNetGetId(i, &id);
1498                 LASSERT (rc == 0);
1499                 the_lnet.ln_ping_info->pi_nid[i] = id.nid;
1500         }
1501
1502         /* We can have a tiny EQ since we only need to see the unlink event on
1503          * teardown, which by definition is the last one! */
1504         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1505         if (rc != 0) {
1506                 CERROR("Can't allocate ping EQ: %d\n", rc);
1507                 goto failed_0;
1508         }
1509
1510         rc = LNetMEAttach(LNET_RESERVED_PORTAL,
1511                           (lnet_process_id_t){.nid = LNET_NID_ANY,
1512                                               .pid = LNET_PID_ANY},
1513                           LNET_PROTO_PING_MATCHBITS, 0LL,
1514                           LNET_UNLINK, LNET_INS_AFTER,
1515                           &meh);
1516         if (rc != 0) {
1517                 CERROR("Can't create ping ME: %d\n", rc);
1518                 goto failed_1;
1519         }
1520
1521         rc = LNetMDAttach(meh,
1522                           (lnet_md_t){.start = the_lnet.ln_ping_info,
1523                                       .length = infosz,
1524                                       .threshold = LNET_MD_THRESH_INF,
1525                                       .options = (LNET_MD_OP_GET |
1526                                                   LNET_MD_TRUNCATE |
1527                                                   LNET_MD_MANAGE_REMOTE),
1528                                       .eq_handle = the_lnet.ln_ping_target_eq},
1529                           LNET_RETAIN,
1530                           &the_lnet.ln_ping_target_md);
1531         if (rc != 0) {
1532                 CERROR("Can't attach ping MD: %d\n", rc);
1533                 goto failed_2;
1534         }
1535
1536         return 0;
1537
1538  failed_2:
1539         rc2 = LNetMEUnlink(meh);
1540         LASSERT (rc2 == 0);
1541  failed_1:
1542         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1543         LASSERT (rc2 == 0);
1544  failed_0:
1545         LIBCFS_FREE(the_lnet.ln_ping_info, infosz);
1546
1547         return rc;
1548 }
1549
1550 void
1551 lnet_ping_target_fini(void)
1552 {
1553         lnet_event_t    event;
1554         int             rc;
1555         int             which;
1556         int             timeout_ms = 1000;
1557         cfs_sigset_t    blocked = cfs_block_allsigs();
1558
1559         LNetMDUnlink(the_lnet.ln_ping_target_md);
1560         /* NB md could be busy; this just starts the unlink */
1561
1562         for (;;) {
1563                 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1564                                 timeout_ms, &event, &which);
1565
1566                 /* I expect overflow... */
1567                 LASSERT (rc >= 0 || rc == -EOVERFLOW);
1568
1569                 if (rc == 0) {
1570                         /* timed out: provide a diagnostic */
1571                         CWARN("Still waiting for ping MD to unlink\n");
1572                         timeout_ms *= 2;
1573                         continue;
1574                 }
1575
1576                 /* Got a valid event */
1577                 if (event.unlinked)
1578                         break;
1579         }
1580
1581         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1582         LASSERT (rc == 0);
1583
1584         LIBCFS_FREE(the_lnet.ln_ping_info,
1585                     offsetof(lnet_ping_info_t,
1586                              pi_nid[the_lnet.ln_ping_info->pi_nnids]));
1587
1588         cfs_restore_sigs(blocked);
1589 }
1590
1591 int
1592 lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1593 {
1594         lnet_handle_eq_t     eqh;
1595         lnet_handle_md_t     mdh;
1596         lnet_event_t         event;
1597         int                  which;
1598         int                  unlinked = 0;
1599         int                  replied = 0;
1600         const int            a_long_time = 60000; /* mS */
1601         int                  infosz = offsetof(lnet_ping_info_t, pi_nid[n_ids]);
1602         lnet_ping_info_t    *info;
1603         lnet_process_id_t    tmpid;
1604         int                  i;
1605         int                  nob;
1606         int                  rc;
1607         int                  rc2;
1608         cfs_sigset_t         blocked;
1609
1610         if (n_ids <= 0 ||
1611             id.nid == LNET_NID_ANY ||
1612             timeout_ms > 500000 ||              /* arbitrary limit! */
1613             n_ids > 20)                         /* arbitrary limit! */
1614                 return -EINVAL;
1615
1616         if (id.pid == LNET_PID_ANY)
1617                 id.pid = LUSTRE_SRV_LNET_PID;
1618
1619         LIBCFS_ALLOC(info, infosz);
1620         if (info == NULL)
1621                 return -ENOMEM;
1622
1623         /* NB 2 events max (including any unlink event) */
1624         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1625         if (rc != 0) {
1626                 CERROR("Can't allocate EQ: %d\n", rc);
1627                 goto out_0;
1628         }
1629
1630         rc = LNetMDBind((lnet_md_t){.start = info,
1631                                     .length = infosz,
1632                                     .threshold = 2, /* GET/REPLY */
1633                                     .options = LNET_MD_TRUNCATE,
1634                                     .eq_handle = eqh},
1635                         LNET_UNLINK,
1636                         &mdh);
1637         if (rc != 0) {
1638                 CERROR("Can't bind MD: %d\n", rc);
1639                 goto out_1;
1640         }
1641
1642         rc = LNetGet(LNET_NID_ANY, mdh, id,
1643                      LNET_RESERVED_PORTAL,
1644                      LNET_PROTO_PING_MATCHBITS, 0);
1645
1646         if (rc != 0) {
1647                 /* Don't CERROR; this could be deliberate! */
1648
1649                 rc2 = LNetMDUnlink(mdh);
1650                 LASSERT (rc2 == 0);
1651
1652                 /* NB must wait for the UNLINK event below... */
1653                 unlinked = 1;
1654                 timeout_ms = a_long_time;
1655         }
1656
1657         do {
1658                 /* MUST block for unlink to complete */
1659                 if (unlinked)
1660                         blocked = cfs_block_allsigs();
1661
1662                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
1663
1664                 if (unlinked)
1665                         cfs_restore_sigs(blocked);
1666
1667                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
1668                        (rc2 <= 0) ? -1 : event.type,
1669                        (rc2 <= 0) ? -1 : event.status,
1670                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
1671
1672                 LASSERT (rc2 != -EOVERFLOW);     /* can't miss anything */
1673
1674                 if (rc2 <= 0 || event.status != 0) {
1675                         /* timeout or error */
1676                         if (!replied && rc == 0)
1677                                 rc = (rc2 < 0) ? rc2 :
1678                                      (rc2 == 0) ? -ETIMEDOUT :
1679                                      event.status;
1680
1681                         if (!unlinked) {
1682                                 /* Ensure completion in finite time... */
1683                                 LNetMDUnlink(mdh);
1684                                 /* No assertion (racing with network) */
1685                                 unlinked = 1;
1686                                 timeout_ms = a_long_time;
1687                         } else if (rc2 == 0) {
1688                                 /* timed out waiting for unlink */
1689                                 CWARN("ping %s: late network completion\n",
1690                                       libcfs_id2str(id));
1691                         }
1692
1693                 } else if (event.type == LNET_EVENT_REPLY) {
1694                         replied = 1;
1695                         rc = event.mlength;
1696                 }
1697
1698         } while (rc2 <= 0 || !event.unlinked);
1699
1700         if (!replied) {
1701                 if (rc >= 0)
1702                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
1703                               libcfs_id2str(id));
1704                 rc = -EIO;
1705                 goto out_1;
1706         }
1707
1708         nob = rc;
1709         LASSERT (nob >= 0 && nob <= infosz);
1710
1711         rc = -EPROTO;                           /* if I can't parse... */
1712
1713         if (nob < 8) {
1714                 /* can't check magic/version */
1715                 CERROR("%s: ping info too short %d\n",
1716                        libcfs_id2str(id), nob);
1717                 goto out_1;
1718         }
1719
1720         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
1721                 /* NB I might be swabbing garbage until I check below, but it
1722                  * doesn't matter */
1723                 __swab32s(&info->pi_version);
1724                 __swab32s(&info->pi_pid);
1725                 __swab32s(&info->pi_nnids);
1726                 for (i = 0; i < info->pi_nnids && i < n_ids; i++)
1727                         __swab64s(&info->pi_nid[i]);
1728
1729         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
1730                 CERROR("%s: Unexpected magic %08x\n",
1731                        libcfs_id2str(id), info->pi_magic);
1732                 goto out_1;
1733         }
1734
1735         if (info->pi_version != LNET_PROTO_PING_VERSION) {
1736                 CERROR("%s: Unexpected version 0x%x\n",
1737                        libcfs_id2str(id), info->pi_version);
1738                 goto out_1;
1739         }
1740
1741         if (nob < offsetof(lnet_ping_info_t, pi_nid[0])) {
1742                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
1743                        nob, (int)offsetof(lnet_ping_info_t, pi_nid[0]));
1744                 goto out_1;
1745         }
1746
1747         if (info->pi_nnids < n_ids)
1748                 n_ids = info->pi_nnids;
1749
1750         if (nob < offsetof(lnet_ping_info_t, pi_nid[n_ids])) {
1751                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
1752                        nob, (int)offsetof(lnet_ping_info_t, pi_nid[n_ids]));
1753                 goto out_1;
1754         }
1755
1756         rc = -EFAULT;                           /* If I SEGV... */
1757
1758         for (i = 0; i < n_ids; i++) {
1759                 tmpid.pid = info->pi_pid;
1760                 tmpid.nid = info->pi_nid[i];
1761 #ifdef __KERNEL__
1762                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
1763                         goto out_1;
1764 #else
1765                 ids[i] = tmpid;
1766 #endif
1767         }
1768         rc = info->pi_nnids;
1769
1770  out_1:
1771         rc2 = LNetEQFree(eqh);
1772         if (rc2 != 0)
1773                 CERROR("rc2 %d\n", rc2);
1774         LASSERT (rc2 == 0);
1775
1776  out_0:
1777         LIBCFS_FREE(info, infosz);
1778         return rc;
1779 }