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 (c) 2003, 2010, Oracle and/or its affiliates. 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 char *
  63 lnet_get_routes(void)
  64 {
  65         return routes;
  66 }
  67
  68 char *
  69 lnet_get_networks(void)
  70 {
  71         char   *nets;
  72         int     rc;
  73
  74         if (*networks != 0 && *ip2nets != 0) {
  75                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  76                                    "'ip2nets' but not both at once\n");
  77                 return NULL;
  78         }
  79
  80         if (*ip2nets != 0) {
  81                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  82                 return (rc == 0) ? nets : NULL;
  83         }
  84
  85         if (*networks != 0)
  86                 return networks;
  87
  88         return "tcp";
  89 }
  90
  91 void
  92 lnet_init_locks(void)
  93 {
  94         cfs_spin_lock_init (&the_lnet.ln_lock);
  95         cfs_waitq_init (&the_lnet.ln_waitq);
  96         cfs_init_mutex(&the_lnet.ln_lnd_mutex);
  97         cfs_init_mutex(&the_lnet.ln_api_mutex);
  98 }
  99
 100 void
 101 lnet_fini_locks(void)
 102 {
 103 }
 104
 105 #else
 106
 107 char *
 108 lnet_get_routes(void)
 109 {
 110         char *str = getenv("LNET_ROUTES");
 111
 112         return (str == NULL) ? "" : str;
 113 }
 114
 115 char *
 116 lnet_get_networks (void)
 117 {
 118         static char       default_networks[256];
 119         char             *networks = getenv ("LNET_NETWORKS");
 120         char             *ip2nets  = getenv ("LNET_IP2NETS");
 121         char             *str;
 122         char             *sep;
 123         int               len;
 124         int               nob;
 125         int               rc;
 126         cfs_list_t       *tmp;
 127
 128 #ifdef NOT_YET
 129         if (networks != NULL && ip2nets != NULL) {
 130                 LCONSOLE_ERROR_MSG(0x103, "Please set EITHER 'LNET_NETWORKS' or"
 131                                    " 'LNET_IP2NETS' but not both at once\n");
 132                 return NULL;
 133         }
 134
 135         if (ip2nets != NULL) {
 136                 rc = lnet_parse_ip2nets(&networks, ip2nets);
 137                 return (rc == 0) ? networks : NULL;
 138         }
 139 #else
 140         ip2nets = NULL;
 141         rc = 0;
 142 #endif
 143         if (networks != NULL)
 144                 return networks;
 145
 146         /* In userland, the default 'networks=' is the list of known net types */
 147
 148         len = sizeof(default_networks);
 149         str = default_networks;
 150         *str = 0;
 151         sep = "";
 152
 153         cfs_list_for_each (tmp, &the_lnet.ln_lnds) {
 154                 lnd_t *lnd = cfs_list_entry(tmp, lnd_t, lnd_list);
 155
 156                 nob = snprintf(str, len, "%s%s", sep,
 157                                libcfs_lnd2str(lnd->lnd_type));
 158                 len -= nob;
 159                 if (len < 0) {
 160                         /* overflowed the string; leave it where it was */
 161                         *str = 0;
 162                         break;
 163                 }
 164
 165                 str += nob;
 166                 sep = ",";
 167         }
 168
 169         return default_networks;
 170 }
 171
 172 # ifndef HAVE_LIBPTHREAD
 173
 174 void lnet_init_locks(void)
 175 {
 176         the_lnet.ln_lock = 0;
 177         the_lnet.ln_lnd_mutex = 0;
 178         the_lnet.ln_api_mutex = 0;
 179 }
 180
 181 void lnet_fini_locks(void)
 182 {
 183         LASSERT (the_lnet.ln_api_mutex == 0);
 184         LASSERT (the_lnet.ln_lnd_mutex == 0);
 185         LASSERT (the_lnet.ln_lock == 0);
 186 }
 187
 188 # else
 189
 190 void lnet_init_locks(void)
 191 {
 192         pthread_cond_init(&the_lnet.ln_cond, NULL);
 193         pthread_mutex_init(&the_lnet.ln_lock, NULL);
 194         pthread_mutex_init(&the_lnet.ln_lnd_mutex, NULL);
 195         pthread_mutex_init(&the_lnet.ln_api_mutex, NULL);
 196 }
 197
 198 void lnet_fini_locks(void)
 199 {
 200         pthread_mutex_destroy(&the_lnet.ln_api_mutex);
 201         pthread_mutex_destroy(&the_lnet.ln_lnd_mutex);
 202         pthread_mutex_destroy(&the_lnet.ln_lock);
 203         pthread_cond_destroy(&the_lnet.ln_cond);
 204 }
 205
 206 # endif
 207 #endif
 208
 209 void lnet_assert_wire_constants (void)
 210 {
 211         /* Wire protocol assertions generated by 'wirecheck'
 212          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 213          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 214          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 215
 216         /* Constants... */
 217         CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 218         CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
 219         CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
 220         CLASSERT (LNET_MSG_ACK == 0);
 221         CLASSERT (LNET_MSG_PUT == 1);
 222         CLASSERT (LNET_MSG_GET == 2);
 223         CLASSERT (LNET_MSG_REPLY == 3);
 224         CLASSERT (LNET_MSG_HELLO == 4);
 225
 226         /* Checks for struct ptl_handle_wire_t */
 227         CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
 228         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 229         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 230         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 231         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 232
 233         /* Checks for struct lnet_magicversion_t */
 234         CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
 235         CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
 236         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 237         CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
 238         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 239         CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 240         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 241
 242         /* Checks for struct lnet_hdr_t */
 243         CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
 244         CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 245         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 246         CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
 247         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 248         CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 249         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 250         CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
 251         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 252         CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
 253         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 254         CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
 255         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 256         CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
 257         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 258
 259         /* Ack */
 260         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 261         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 262         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 263         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 264         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 265         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 266
 267         /* Put */
 268         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 269         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 270         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 271         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 272         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 273         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 274         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 275         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 276         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 277         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 278
 279         /* Get */
 280         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 281         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 282         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 283         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 284         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 285         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 286         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 287         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 288         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 289         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 290
 291         /* Reply */
 292         CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 293         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 294
 295         /* Hello */
 296         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 297         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 298         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 299         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 300 }
 301
 302 lnd_t *
 303 lnet_find_lnd_by_type (int type)
 304 {
 305         lnd_t              *lnd;
 306         cfs_list_t         *tmp;
 307
 308         /* holding lnd mutex */
 309         cfs_list_for_each (tmp, &the_lnet.ln_lnds) {
 310                 lnd = cfs_list_entry(tmp, lnd_t, lnd_list);
 311
 312                 if ((int)lnd->lnd_type == type)
 313                         return lnd;
 314         }
 315
 316         return NULL;
 317 }
 318
 319 void
 320 lnet_register_lnd (lnd_t *lnd)
 321 {
 322         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
 323
 324         LASSERT (the_lnet.ln_init);
 325         LASSERT (libcfs_isknown_lnd(lnd->lnd_type));
 326         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 327
 328         cfs_list_add_tail (&lnd->lnd_list, &the_lnet.ln_lnds);
 329         lnd->lnd_refcount = 0;
 330
 331         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 332
 333         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
 334 }
 335
 336 void
 337 lnet_unregister_lnd (lnd_t *lnd)
 338 {
 339         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
 340
 341         LASSERT (the_lnet.ln_init);
 342         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 343         LASSERT (lnd->lnd_refcount == 0);
 344
 345         cfs_list_del (&lnd->lnd_list);
 346         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 347
 348         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
 349 }
 350
 351 #ifndef LNET_USE_LIB_FREELIST
 352
 353 int
 354 lnet_descriptor_setup (void)
 355 {
 356         return 0;
 357 }
 358
 359 void
 360 lnet_descriptor_cleanup (void)
 361 {
 362 }
 363
 364 #else
 365
 366 int
 367 lnet_freelist_init (lnet_freelist_t *fl, int n, int size)
 368 {
 369         char *space;
 370
 371         LASSERT (n > 0);
 372
 373         size += offsetof (lnet_freeobj_t, fo_contents);
 374
 375         LIBCFS_ALLOC(space, n * size);
 376         if (space == NULL)
 377                 return (-ENOMEM);
 378
 379         CFS_INIT_LIST_HEAD (&fl->fl_list);
 380         fl->fl_objs = space;
 381         fl->fl_nobjs = n;
 382         fl->fl_objsize = size;
 383
 384         do
 385         {
 386                 memset (space, 0, size);
 387                 cfs_list_add ((cfs_list_t *)space, &fl->fl_list);
 388                 space += size;
 389         } while (--n != 0);
 390
 391         return (0);
 392 }
 393
 394 void
 395 lnet_freelist_fini (lnet_freelist_t *fl)
 396 {
 397         cfs_list_t       *el;
 398         int               count;
 399
 400         if (fl->fl_nobjs == 0)
 401                 return;
 402
 403         count = 0;
 404         for (el = fl->fl_list.next; el != &fl->fl_list; el = el->next)
 405                 count++;
 406
 407         LASSERT (count == fl->fl_nobjs);
 408
 409         LIBCFS_FREE(fl->fl_objs, fl->fl_nobjs * fl->fl_objsize);
 410         memset (fl, 0, sizeof (*fl));
 411 }
 412
 413 int
 414 lnet_descriptor_setup (void)
 415 {
 416         /* NB on failure caller must still call lnet_descriptor_cleanup */
 417         /*               ******                                         */
 418         int        rc;
 419
 420         memset (&the_lnet.ln_free_mes,  0, sizeof (the_lnet.ln_free_mes));
 421         memset (&the_lnet.ln_free_msgs, 0, sizeof (the_lnet.ln_free_msgs));
 422         memset (&the_lnet.ln_free_mds,  0, sizeof (the_lnet.ln_free_mds));
 423         memset (&the_lnet.ln_free_eqs,  0, sizeof (the_lnet.ln_free_eqs));
 424
 425         rc = lnet_freelist_init(&the_lnet.ln_free_mes,
 426                                 MAX_MES, sizeof (lnet_me_t));
 427         if (rc != 0)
 428                 return (rc);
 429
 430         rc = lnet_freelist_init(&the_lnet.ln_free_msgs,
 431                                 MAX_MSGS, sizeof (lnet_msg_t));
 432         if (rc != 0)
 433                 return (rc);
 434
 435         rc = lnet_freelist_init(&the_lnet.ln_free_mds,
 436                                 MAX_MDS, sizeof (lnet_libmd_t));
 437         if (rc != 0)
 438                 return (rc);
 439
 440         rc = lnet_freelist_init(&the_lnet.ln_free_eqs,
 441                                 MAX_EQS, sizeof (lnet_eq_t));
 442         return (rc);
 443 }
 444
 445 void
 446 lnet_descriptor_cleanup (void)
 447 {
 448         lnet_freelist_fini (&the_lnet.ln_free_mes);
 449         lnet_freelist_fini (&the_lnet.ln_free_msgs);
 450         lnet_freelist_fini (&the_lnet.ln_free_mds);
 451         lnet_freelist_fini (&the_lnet.ln_free_eqs);
 452 }
 453
 454 #endif
 455
 456 __u64
 457 lnet_create_interface_cookie (void)
 458 {
 459         /* NB the interface cookie in wire handles guards against delayed
 460          * replies and ACKs appearing valid after reboot. Initialisation time,
 461          * even if it's only implemented to millisecond resolution is probably
 462          * easily good enough. */
 463         struct timeval tv;
 464         __u64          cookie;
 465 #ifndef __KERNEL__
 466         int            rc = gettimeofday (&tv, NULL);
 467         LASSERT (rc == 0);
 468 #else
 469         cfs_gettimeofday(&tv);
 470 #endif
 471         cookie = tv.tv_sec;
 472         cookie *= 1000000;
 473         cookie += tv.tv_usec;
 474         return cookie;
 475 }
 476
 477 int
 478 lnet_setup_handle_hash (void)
 479 {
 480         int       i;
 481
 482         /* Arbitrary choice of hash table size */
 483 #ifdef __KERNEL__
 484         the_lnet.ln_lh_hash_size =
 485                 CFS_PAGE_SIZE / sizeof (cfs_list_t);
 486 #else
 487         the_lnet.ln_lh_hash_size = (MAX_MES + MAX_MDS + MAX_EQS)/4;
 488 #endif
 489         LIBCFS_ALLOC(the_lnet.ln_lh_hash_table,
 490                      the_lnet.ln_lh_hash_size * sizeof (cfs_list_t));
 491         if (the_lnet.ln_lh_hash_table == NULL)
 492                 return (-ENOMEM);
 493
 494         for (i = 0; i < the_lnet.ln_lh_hash_size; i++)
 495                 CFS_INIT_LIST_HEAD (&the_lnet.ln_lh_hash_table[i]);
 496
 497         the_lnet.ln_next_object_cookie = LNET_COOKIE_TYPES;
 498
 499         return (0);
 500 }
 501
 502 void
 503 lnet_cleanup_handle_hash (void)
 504 {
 505         if (the_lnet.ln_lh_hash_table == NULL)
 506                 return;
 507
 508         LIBCFS_FREE(the_lnet.ln_lh_hash_table,
 509                     the_lnet.ln_lh_hash_size * sizeof (cfs_list_t));
 510 }
 511
 512 lnet_libhandle_t *
 513 lnet_lookup_cookie (__u64 cookie, int type)
 514 {
 515         /* ALWAYS called with LNET_LOCK held */
 516         cfs_list_t          *list;
 517         cfs_list_t          *el;
 518         unsigned int         hash;
 519
 520         if ((cookie & (LNET_COOKIE_TYPES - 1)) != type)
 521                 return (NULL);
 522
 523         hash = ((unsigned int)cookie) % the_lnet.ln_lh_hash_size;
 524         list = &the_lnet.ln_lh_hash_table[hash];
 525
 526         cfs_list_for_each (el, list) {
 527                 lnet_libhandle_t *lh = cfs_list_entry (el, lnet_libhandle_t,
 528                                                       lh_hash_chain);
 529
 530                 if (lh->lh_cookie == cookie)
 531                         return (lh);
 532         }
 533
 534         return (NULL);
 535 }
 536
 537 void
 538 lnet_initialise_handle (lnet_libhandle_t *lh, int type)
 539 {
 540         /* ALWAYS called with LNET_LOCK held */
 541         unsigned int    hash;
 542
 543         LASSERT (type >= 0 && type < LNET_COOKIE_TYPES);
 544         lh->lh_cookie = the_lnet.ln_next_object_cookie | type;
 545         the_lnet.ln_next_object_cookie += LNET_COOKIE_TYPES;
 546
 547         hash = ((unsigned int)lh->lh_cookie) % the_lnet.ln_lh_hash_size;
 548         cfs_list_add (&lh->lh_hash_chain, &the_lnet.ln_lh_hash_table[hash]);
 549 }
 550
 551 void
 552 lnet_invalidate_handle (lnet_libhandle_t *lh)
 553 {
 554         /* ALWAYS called with LNET_LOCK held */
 555         cfs_list_del (&lh->lh_hash_chain);
 556 }
 557
 558 cfs_list_t *
 559 lnet_portal_mhash_alloc(void)
 560 {
 561         cfs_list_t       *mhash;
 562         int               i;
 563
 564         LIBCFS_ALLOC(mhash, sizeof(cfs_list_t) * LNET_PORTAL_HASH_SIZE);
 565         if (mhash == NULL)
 566                 return NULL;
 567
 568         for (i = 0; i < LNET_PORTAL_HASH_SIZE; i++)
 569                 CFS_INIT_LIST_HEAD(&mhash[i]);
 570
 571         return mhash;
 572 }
 573
 574 void
 575 lnet_portal_mhash_free(cfs_list_t *mhash)
 576 {
 577         int     i;
 578
 579         for (i = 0; i < LNET_PORTAL_HASH_SIZE; i++) {
 580                 while (!cfs_list_empty(&mhash[i])) {
 581                         lnet_me_t *me = cfs_list_entry(mhash[i].next,
 582                                                        lnet_me_t, me_list);
 583                         CERROR ("Active ME %p on exit portal mhash\n", me);
 584                         cfs_list_del(&me->me_list);
 585                         lnet_me_free(me);
 586                 }
 587         }
 588         LIBCFS_FREE(mhash, sizeof(cfs_list_t) * LNET_PORTAL_HASH_SIZE);
 589 }
 590
 591 int
 592 lnet_init_finalizers(void)
 593 {
 594 #ifdef __KERNEL__
 595         int    i;
 596
 597         the_lnet.ln_nfinalizers = (int) cfs_num_online_cpus();
 598
 599         LIBCFS_ALLOC(the_lnet.ln_finalizers,
 600                      the_lnet.ln_nfinalizers *
 601                      sizeof(*the_lnet.ln_finalizers));
 602         if (the_lnet.ln_finalizers == NULL) {
 603                 CERROR("Can't allocate ln_finalizers\n");
 604                 return -ENOMEM;
 605         }
 606
 607         for (i = 0; i < the_lnet.ln_nfinalizers; i++)
 608                 the_lnet.ln_finalizers[i] = NULL;
 609 #else
 610         the_lnet.ln_finalizing = 0;
 611 #endif
 612
 613         CFS_INIT_LIST_HEAD(&the_lnet.ln_finalizeq);
 614         return 0;
 615 }
 616
 617 void
 618 lnet_fini_finalizers(void)
 619 {
 620 #ifdef __KERNEL__
 621         int    i;
 622
 623         for (i = 0; i < the_lnet.ln_nfinalizers; i++)
 624                 LASSERT (the_lnet.ln_finalizers[i] == NULL);
 625
 626         LIBCFS_FREE(the_lnet.ln_finalizers,
 627                     the_lnet.ln_nfinalizers *
 628                     sizeof(*the_lnet.ln_finalizers));
 629 #else
 630         LASSERT (!the_lnet.ln_finalizing);
 631 #endif
 632         LASSERT (cfs_list_empty(&the_lnet.ln_finalizeq));
 633 }
 634
 635 #ifndef __KERNEL__
 636 /* Temporary workaround to allow uOSS and test programs force server
 637  * mode in userspace. See comments near ln_server_mode_flag in
 638  * lnet/lib-types.h */
 639
 640 void
 641 lnet_server_mode() {
 642         the_lnet.ln_server_mode_flag = 1;
 643 }
 644 #endif
 645
 646 int
 647 lnet_prepare(lnet_pid_t requested_pid)
 648 {
 649         /* Prepare to bring up the network */
 650         int               rc = 0;
 651         int               i;
 652
 653         LASSERT (the_lnet.ln_refcount == 0);
 654
 655         the_lnet.ln_routing = 0;
 656
 657 #ifdef __KERNEL__
 658         LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 659         the_lnet.ln_pid = requested_pid;
 660 #else
 661         if (the_lnet.ln_server_mode_flag) {/* server case (uOSS) */
 662                 LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 663
 664                 if (cfs_curproc_uid())/* Only root can run user-space server */
 665                         return -EPERM;
 666                 the_lnet.ln_pid = requested_pid;
 667
 668         } else {/* client case (liblustre) */
 669
 670                 /* My PID must be unique on this node and flag I'm userspace */
 671                 the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
 672         }
 673 #endif
 674
 675         rc = lnet_descriptor_setup();
 676         if (rc != 0)
 677                 goto failed0;
 678
 679         memset(&the_lnet.ln_counters, 0,
 680                sizeof(the_lnet.ln_counters));
 681
 682         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_msgs);
 683         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_mds);
 684         CFS_INIT_LIST_HEAD (&the_lnet.ln_active_eqs);
 685         CFS_INIT_LIST_HEAD (&the_lnet.ln_test_peers);
 686         CFS_INIT_LIST_HEAD (&the_lnet.ln_nis);
 687         CFS_INIT_LIST_HEAD (&the_lnet.ln_zombie_nis);
 688         CFS_INIT_LIST_HEAD (&the_lnet.ln_remote_nets);
 689         CFS_INIT_LIST_HEAD (&the_lnet.ln_routers);
 690
 691         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 692
 693         lnet_init_rtrpools();
 694
 695         rc = lnet_setup_handle_hash ();
 696         if (rc != 0)
 697                 goto failed0;
 698
 699         rc = lnet_create_peer_table();
 700         if (rc != 0)
 701                 goto failed1;
 702
 703         rc = lnet_init_finalizers();
 704         if (rc != 0)
 705                 goto failed2;
 706
 707         the_lnet.ln_nportals = MAX_PORTALS;
 708         LIBCFS_ALLOC(the_lnet.ln_portals,
 709                      the_lnet.ln_nportals *
 710                      sizeof(*the_lnet.ln_portals));
 711         if (the_lnet.ln_portals == NULL) {
 712                 rc = -ENOMEM;
 713                 goto failed3;
 714         }
 715
 716         for (i = 0; i < the_lnet.ln_nportals; i++) {
 717                 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_mlist));
 718                 CFS_INIT_LIST_HEAD(&(the_lnet.ln_portals[i].ptl_msgq));
 719                 the_lnet.ln_portals[i].ptl_options = 0;
 720         }
 721
 722         return 0;
 723
 724  failed3:
 725         lnet_fini_finalizers();
 726  failed2:
 727         lnet_destroy_peer_table();
 728  failed1:
 729         lnet_cleanup_handle_hash();
 730  failed0:
 731         lnet_descriptor_cleanup();
 732         return rc;
 733 }
 734
 735 int
 736 lnet_unprepare (void)
 737 {
 738         int       idx;
 739
 740         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 741          * have shut down already, so it is safe to unlink and free all
 742          * descriptors, even those that appear committed to a network op (eg MD
 743          * with non-zero pending count) */
 744
 745         lnet_fail_nid(LNET_NID_ANY, 0);
 746
 747         LASSERT (cfs_list_empty(&the_lnet.ln_test_peers));
 748         LASSERT (the_lnet.ln_refcount == 0);
 749         LASSERT (cfs_list_empty(&the_lnet.ln_nis));
 750         LASSERT (cfs_list_empty(&the_lnet.ln_zombie_nis));
 751         LASSERT (the_lnet.ln_nzombie_nis == 0);
 752
 753         for (idx = 0; idx < the_lnet.ln_nportals; idx++) {
 754                 lnet_portal_t *ptl = &the_lnet.ln_portals[idx];
 755
 756                 LASSERT (cfs_list_empty(&ptl->ptl_msgq));
 757
 758                 while (!cfs_list_empty(&ptl->ptl_mlist)) {
 759                         lnet_me_t *me = cfs_list_entry(ptl->ptl_mlist.next,
 760                                                        lnet_me_t, me_list);
 761                         CERROR ("Active ME %p on exit\n", me);
 762                         cfs_list_del (&me->me_list);
 763                         lnet_me_free (me);
 764                 }
 765
 766                 if (ptl->ptl_mhash != NULL) {
 767                         LASSERT (lnet_portal_is_unique(ptl));
 768                         lnet_portal_mhash_free(ptl->ptl_mhash);
 769                 }
 770         }
 771
 772         while (!cfs_list_empty (&the_lnet.ln_active_mds)) {
 773                 lnet_libmd_t *md = cfs_list_entry (the_lnet.ln_active_mds.next,
 774                                                    lnet_libmd_t, md_list);
 775
 776                 CERROR ("Active MD %p on exit\n", md);
 777                 cfs_list_del_init (&md->md_list);
 778                 lnet_md_free (md);
 779         }
 780
 781         while (!cfs_list_empty (&the_lnet.ln_active_eqs)) {
 782                 lnet_eq_t *eq = cfs_list_entry (the_lnet.ln_active_eqs.next,
 783                                                 lnet_eq_t, eq_list);
 784
 785                 CERROR ("Active EQ %p on exit\n", eq);
 786                 cfs_list_del (&eq->eq_list);
 787                 lnet_eq_free (eq);
 788         }
 789
 790         while (!cfs_list_empty (&the_lnet.ln_active_msgs)) {
 791                 lnet_msg_t *msg = cfs_list_entry (the_lnet.ln_active_msgs.next,
 792                                                   lnet_msg_t, msg_activelist);
 793
 794                 CERROR ("Active msg %p on exit\n", msg);
 795                 LASSERT (msg->msg_onactivelist);
 796                 msg->msg_onactivelist = 0;
 797                 cfs_list_del (&msg->msg_activelist);
 798                 lnet_msg_free (msg);
 799         }
 800
 801         LIBCFS_FREE(the_lnet.ln_portals,
 802                     the_lnet.ln_nportals * sizeof(*the_lnet.ln_portals));
 803
 804         lnet_free_rtrpools();
 805         lnet_fini_finalizers();
 806         lnet_destroy_peer_table();
 807         lnet_cleanup_handle_hash();
 808         lnet_descriptor_cleanup();
 809
 810         return (0);
 811 }
 812
 813 lnet_ni_t  *
 814 lnet_net2ni_locked (__u32 net)
 815 {
 816         cfs_list_t       *tmp;
 817         lnet_ni_t        *ni;
 818
 819         cfs_list_for_each (tmp, &the_lnet.ln_nis) {
 820                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 821
 822                 if (LNET_NIDNET(ni->ni_nid) == net) {
 823                         lnet_ni_addref_locked(ni);
 824                         return ni;
 825                 }
 826         }
 827
 828         return NULL;
 829 }
 830
 831 int
 832 lnet_islocalnet (__u32 net)
 833 {
 834         lnet_ni_t        *ni;
 835
 836         LNET_LOCK();
 837         ni = lnet_net2ni_locked(net);
 838         if (ni != NULL)
 839                 lnet_ni_decref_locked(ni);
 840         LNET_UNLOCK();
 841
 842         return ni != NULL;
 843 }
 844
 845 lnet_ni_t  *
 846 lnet_nid2ni_locked (lnet_nid_t nid)
 847 {
 848         cfs_list_t       *tmp;
 849         lnet_ni_t        *ni;
 850
 851         cfs_list_for_each (tmp, &the_lnet.ln_nis) {
 852                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 853
 854                 if (ni->ni_nid == nid) {
 855                         lnet_ni_addref_locked(ni);
 856                         return ni;
 857                 }
 858         }
 859
 860         return NULL;
 861 }
 862
 863 int
 864 lnet_islocalnid (lnet_nid_t nid)
 865 {
 866         lnet_ni_t     *ni;
 867
 868         LNET_LOCK();
 869         ni = lnet_nid2ni_locked(nid);
 870         if (ni != NULL)
 871                 lnet_ni_decref_locked(ni);
 872         LNET_UNLOCK();
 873
 874         return ni != NULL;
 875 }
 876
 877 int
 878 lnet_count_acceptor_nis (void)
 879 {
 880         /* Return the # of NIs that need the acceptor. */
 881         int            count = 0;
 882 #if defined(__KERNEL__) || defined(HAVE_LIBPTHREAD)
 883         cfs_list_t    *tmp;
 884         lnet_ni_t     *ni;
 885
 886         LNET_LOCK();
 887         cfs_list_for_each (tmp, &the_lnet.ln_nis) {
 888                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 889
 890                 if (ni->ni_lnd->lnd_accept != NULL)
 891                         count++;
 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 (cfs_list_empty(&the_lnet.ln_zombie_nis));
 913         LASSERT (the_lnet.ln_nzombie_nis == 0);
 914         LASSERT (cfs_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 (!cfs_list_empty(&the_lnet.ln_nis)) {
 921                 ni = cfs_list_entry(the_lnet.ln_nis.next,
 922                                     lnet_ni_t, ni_list);
 923                 cfs_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 (cfs_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 = cfs_list_entry(the_lnet.ln_zombie_nis.next,
 969                                     lnet_ni_t, ni_list);
 970                 cfs_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 (!cfs_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         cfs_list_t         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 (!cfs_list_empty(&nilist)) {
1024                 ni = cfs_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 = cfs_request_module("%s",
1036                                                 libcfs_lnd2modname(lnd_type));
1037                         LNET_MUTEX_DOWN(&the_lnet.ln_lnd_mutex);
1038
1039                         lnd = lnet_find_lnd_by_type(lnd_type);
1040                         if (lnd == NULL) {
1041                                 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1042                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1043                                        libcfs_lnd2str(lnd_type),
1044                                        libcfs_lnd2modname(lnd_type), rc);
1045 #ifndef HAVE_MODULE_LOADING_SUPPORT
1046                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1047                                          "compiled with kernel module "
1048                                          "loading support.");
1049 #endif
1050                                 goto failed;
1051                         }
1052                 }
1053 #else
1054                 if (lnd == NULL) {
1055                         LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1056                         CERROR("LND %s not supported\n",
1057                                libcfs_lnd2str(lnd_type));
1058                         goto failed;
1059                 }
1060 #endif
1061
1062                 ni->ni_refcount = 1;
1063
1064                 LNET_LOCK();
1065                 lnd->lnd_refcount++;
1066                 LNET_UNLOCK();
1067
1068                 ni->ni_lnd = lnd;
1069
1070                 rc = (lnd->lnd_startup)(ni);
1071
1072                 LNET_MUTEX_UP(&the_lnet.ln_lnd_mutex);
1073
1074                 if (rc != 0) {
1075                         LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
1076                                            "\n",
1077                                            rc, libcfs_lnd2str(lnd->lnd_type));
1078                         LNET_LOCK();
1079                         lnd->lnd_refcount--;
1080                         LNET_UNLOCK();
1081                         goto failed;
1082                 }
1083
1084                 LASSERT (ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1085
1086                 cfs_list_del(&ni->ni_list);
1087
1088                 LNET_LOCK();
1089                 cfs_list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1090                 LNET_UNLOCK();
1091
1092                 if (lnd->lnd_type == LOLND) {
1093                         lnet_ni_addref(ni);
1094                         LASSERT (the_lnet.ln_loni == NULL);
1095                         the_lnet.ln_loni = ni;
1096                         continue;
1097                 }
1098
1099 #ifndef __KERNEL__
1100                 if (lnd->lnd_wait != NULL) {
1101                         if (the_lnet.ln_eqwaitni == NULL) {
1102                                 lnet_ni_addref(ni);
1103                                 the_lnet.ln_eqwaitni = ni;
1104                         }
1105                 } else {
1106 # ifndef HAVE_LIBPTHREAD
1107                         LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1108                                            "single-threaded runtime\n",
1109                                            libcfs_lnd2str(lnd_type));
1110                         goto failed;
1111 # endif
1112                 }
1113 #endif
1114                 if (ni->ni_peertxcredits == 0 ||
1115                     ni->ni_maxtxcredits == 0) {
1116                         LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1117                                            libcfs_lnd2str(lnd->lnd_type),
1118                                            ni->ni_peertxcredits == 0 ?
1119                                            "" : "per-peer ");
1120                         goto failed;
1121                 }
1122
1123                 ni->ni_txcredits = ni->ni_mintxcredits = ni->ni_maxtxcredits;
1124
1125                 CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1126                        libcfs_nid2str(ni->ni_nid),
1127                        ni->ni_peertxcredits, ni->ni_txcredits,
1128                        ni->ni_peerrtrcredits, ni->ni_peertimeout);
1129
1130                 nicount++;
1131         }
1132
1133         if (the_lnet.ln_eqwaitni != NULL && nicount > 1) {
1134                 lnd_type = the_lnet.ln_eqwaitni->ni_lnd->lnd_type;
1135                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1136                                    "\n",
1137                                    libcfs_lnd2str(lnd_type));
1138                 goto failed;
1139         }
1140
1141         return 0;
1142
1143  failed:
1144         lnet_shutdown_lndnis();
1145
1146         while (!cfs_list_empty(&nilist)) {
1147                 ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
1148                 cfs_list_del(&ni->ni_list);
1149                 LIBCFS_FREE(ni, sizeof(*ni));
1150         }
1151
1152         return -ENETDOWN;
1153 }
1154
1155 int
1156 LNetInit(void)
1157 {
1158         lnet_assert_wire_constants ();
1159         LASSERT (!the_lnet.ln_init);
1160
1161         memset(&the_lnet, 0, sizeof(the_lnet));
1162
1163         lnet_init_locks();
1164         the_lnet.ln_refcount = 0;
1165         the_lnet.ln_init = 1;
1166         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1167         CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
1168         CFS_INIT_LIST_HEAD(&the_lnet.ln_zombie_rcd);
1169
1170 #ifdef __KERNEL__
1171         /* All LNDs apart from the LOLND are in separate modules.  They
1172          * register themselves when their module loads, and unregister
1173          * themselves when their module is unloaded. */
1174 #else
1175         /* Register LNDs
1176          * NB the order here determines default 'networks=' order */
1177 # ifdef CRAY_XT3
1178         LNET_REGISTER_ULND(the_ptllnd);
1179 # endif
1180 # ifdef HAVE_LIBPTHREAD
1181         LNET_REGISTER_ULND(the_tcplnd);
1182 # endif
1183 #endif
1184         lnet_register_lnd(&the_lolnd);
1185         return 0;
1186 }
1187
1188 void
1189 LNetFini(void)
1190 {
1191         LASSERT (the_lnet.ln_init);
1192         LASSERT (the_lnet.ln_refcount == 0);
1193
1194         while (!cfs_list_empty(&the_lnet.ln_lnds))
1195                 lnet_unregister_lnd(cfs_list_entry(the_lnet.ln_lnds.next,
1196                                                    lnd_t, lnd_list));
1197         lnet_fini_locks();
1198
1199         the_lnet.ln_init = 0;
1200 }
1201
1202 int
1203 LNetNIInit(lnet_pid_t requested_pid)
1204 {
1205         int         im_a_router = 0;
1206         int         rc;
1207
1208         LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1209
1210         LASSERT (the_lnet.ln_init);
1211         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1212
1213         if (the_lnet.ln_refcount > 0) {
1214                 rc = the_lnet.ln_refcount++;
1215                 goto out;
1216         }
1217
1218         lnet_get_tunables();
1219
1220         if (requested_pid == LNET_PID_ANY) {
1221                 /* Don't instantiate LNET just for me */
1222                 rc = -ENETDOWN;
1223                 goto failed0;
1224         }
1225
1226         rc = lnet_prepare(requested_pid);
1227         if (rc != 0)
1228                 goto failed0;
1229
1230         rc = lnet_startup_lndnis();
1231         if (rc != 0)
1232                 goto failed1;
1233
1234         rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1235         if (rc != 0)
1236                 goto failed2;
1237
1238         rc = lnet_check_routes();
1239         if (rc != 0)
1240                 goto failed2;
1241
1242         rc = lnet_alloc_rtrpools(im_a_router);
1243         if (rc != 0)
1244                 goto failed2;
1245
1246         rc = lnet_acceptor_start();
1247         if (rc != 0)
1248                 goto failed2;
1249
1250         the_lnet.ln_refcount = 1;
1251         /* Now I may use my own API functions... */
1252
1253         /* NB router checker needs the_lnet.ln_ping_info in
1254          * lnet_router_checker -> lnet_update_ni_status */
1255         rc = lnet_ping_target_init();
1256         if (rc != 0)
1257                 goto failed3;
1258
1259         rc = lnet_router_checker_start();
1260         if (rc != 0)
1261                 goto failed4;
1262
1263         lnet_proc_init();
1264         goto out;
1265
1266  failed4:
1267         lnet_ping_target_fini();
1268  failed3:
1269         the_lnet.ln_refcount = 0;
1270         lnet_acceptor_stop();
1271  failed2:
1272         lnet_destroy_routes();
1273         lnet_shutdown_lndnis();
1274  failed1:
1275         lnet_unprepare();
1276  failed0:
1277         LASSERT (rc < 0);
1278  out:
1279         LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1280         return rc;
1281 }
1282
1283 int
1284 LNetNIFini()
1285 {
1286         LNET_MUTEX_DOWN(&the_lnet.ln_api_mutex);
1287
1288         LASSERT (the_lnet.ln_init);
1289         LASSERT (the_lnet.ln_refcount > 0);
1290
1291         if (the_lnet.ln_refcount != 1) {
1292                 the_lnet.ln_refcount--;
1293         } else {
1294                 LASSERT (!the_lnet.ln_niinit_self);
1295
1296                 lnet_proc_fini();
1297                 lnet_router_checker_stop();
1298                 lnet_ping_target_fini();
1299
1300                 /* Teardown fns that use my own API functions BEFORE here */
1301                 the_lnet.ln_refcount = 0;
1302
1303                 lnet_acceptor_stop();
1304                 lnet_destroy_routes();
1305                 lnet_shutdown_lndnis();
1306                 lnet_unprepare();
1307         }
1308
1309         LNET_MUTEX_UP(&the_lnet.ln_api_mutex);
1310         return 0;
1311 }
1312
1313 int
1314 LNetCtl(unsigned int cmd, void *arg)
1315 {
1316         struct libcfs_ioctl_data *data = arg;
1317         lnet_process_id_t         id = {0};
1318         lnet_ni_t                *ni;
1319         int                       rc;
1320
1321         LASSERT (the_lnet.ln_init);
1322         LASSERT (the_lnet.ln_refcount > 0);
1323
1324         switch (cmd) {
1325         case IOC_LIBCFS_GET_NI:
1326                 rc = LNetGetId(data->ioc_count, &id);
1327                 data->ioc_nid = id.nid;
1328                 return rc;
1329
1330         case IOC_LIBCFS_FAIL_NID:
1331                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1332
1333         case IOC_LIBCFS_ADD_ROUTE:
1334                 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1335                                     data->ioc_nid);
1336                 return (rc != 0) ? rc : lnet_check_routes();
1337
1338         case IOC_LIBCFS_DEL_ROUTE:
1339                 return lnet_del_route(data->ioc_net, data->ioc_nid);
1340
1341         case IOC_LIBCFS_GET_ROUTE:
1342                 return lnet_get_route(data->ioc_count,
1343                                       &data->ioc_net, &data->ioc_count,
1344                                       &data->ioc_nid, &data->ioc_flags);
1345         case IOC_LIBCFS_NOTIFY_ROUTER:
1346                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1347                                    cfs_time_current() -
1348                                    cfs_time_seconds(cfs_time_current_sec() -
1349                                                     (time_t)data->ioc_u64[0]));
1350
1351         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1352                 /* This can be removed once lustre stops calling it */
1353                 return 0;
1354
1355         case IOC_LIBCFS_LNET_DIST:
1356                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1357                 if (rc < 0 && rc != -EHOSTUNREACH)
1358                         return rc;
1359
1360                 data->ioc_u32[0] = rc;
1361                 return 0;
1362
1363         case IOC_LIBCFS_TESTPROTOCOMPAT:
1364                 LNET_LOCK();
1365                 the_lnet.ln_testprotocompat = data->ioc_flags;
1366                 LNET_UNLOCK();
1367                 return 0;
1368
1369         case IOC_LIBCFS_PING:
1370                 id.nid = data->ioc_nid;
1371                 id.pid = data->ioc_u32[0];
1372                 rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
1373                                (lnet_process_id_t *)data->ioc_pbuf1,
1374                                data->ioc_plen1/sizeof(lnet_process_id_t));
1375                 if (rc < 0)
1376                         return rc;
1377                 data->ioc_count = rc;
1378                 return 0;
1379
1380         case IOC_LIBCFS_DEBUG_PEER: {
1381                 /* CAVEAT EMPTOR: this one designed for calling directly; not
1382                  * via an ioctl */
1383                 id = *((lnet_process_id_t *) arg);
1384
1385                 lnet_debug_peer(id.nid);
1386
1387                 ni = lnet_net2ni(LNET_NIDNET(id.nid));
1388                 if (ni == NULL) {
1389                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(id));
1390                 } else {
1391                         if (ni->ni_lnd->lnd_ctl == NULL) {
1392                                 CDEBUG(D_WARNING, "No ctl for %s\n",
1393                                        libcfs_id2str(id));
1394                         } else {
1395                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1396                         }
1397
1398                         lnet_ni_decref(ni);
1399                 }
1400                 return 0;
1401         }
1402
1403         default:
1404                 ni = lnet_net2ni(data->ioc_net);
1405                 if (ni == NULL)
1406                         return -EINVAL;
1407
1408                 if (ni->ni_lnd->lnd_ctl == NULL)
1409                         rc = -EINVAL;
1410                 else
1411                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1412
1413                 lnet_ni_decref(ni);
1414                 return rc;
1415         }
1416         /* not reached */
1417 }
1418
1419 int
1420 LNetGetId(unsigned int index, lnet_process_id_t *id)
1421 {
1422         lnet_ni_t        *ni;
1423         cfs_list_t       *tmp;
1424         int               rc = -ENOENT;
1425
1426         LASSERT (the_lnet.ln_init);
1427         LASSERT (the_lnet.ln_refcount > 0);
1428
1429         LNET_LOCK();
1430
1431         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
1432                 if (index-- != 0)
1433                         continue;
1434
1435                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
1436
1437                 id->nid = ni->ni_nid;
1438                 id->pid = the_lnet.ln_pid;
1439                 rc = 0;
1440                 break;
1441         }
1442
1443         LNET_UNLOCK();
1444
1445         return rc;
1446 }
1447
1448 void
1449 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1450 {
1451         snprintf(str, len, LPX64, h.cookie);
1452 }
1453
1454 static int
1455 lnet_create_ping_info(void)
1456 {
1457         int               i;
1458         int               n;
1459         int               rc;
1460         unsigned int      infosz;
1461         lnet_ni_t        *ni;
1462         lnet_process_id_t id;
1463         lnet_ping_info_t *pinfo;
1464
1465         for (n = 0; ; n++) {
1466                 rc = LNetGetId(n, &id);
1467                 if (rc == -ENOENT)
1468                         break;
1469
1470                 LASSERT (rc == 0);
1471         }
1472
1473         infosz = offsetof(lnet_ping_info_t, pi_ni[n]);
1474         LIBCFS_ALLOC(pinfo, infosz);
1475         if (pinfo == NULL) {
1476                 CERROR("Can't allocate ping info[%d]\n", n);
1477                 return -ENOMEM;
1478         }
1479
1480         pinfo->pi_nnis    = n;
1481         pinfo->pi_pid     = the_lnet.ln_pid;
1482         pinfo->pi_magic   = LNET_PROTO_PING_MAGIC;
1483         pinfo->pi_version = LNET_PROTO_PING_VERSION;
1484
1485         for (i = 0; i < n; i++) {
1486                 lnet_ni_status_t *ns = &pinfo->pi_ni[i];
1487
1488                 rc = LNetGetId(i, &id);
1489                 LASSERT (rc == 0);
1490
1491                 ns->ns_nid    = id.nid;
1492                 ns->ns_status = LNET_NI_STATUS_UP;
1493
1494                 LNET_LOCK();
1495
1496                 ni = lnet_nid2ni_locked(id.nid);
1497                 LASSERT (ni != NULL);
1498                 LASSERT (ni->ni_status == NULL);
1499                 ni->ni_status = ns;
1500                 lnet_ni_decref_locked(ni);
1501
1502                 LNET_UNLOCK();
1503         }
1504
1505         the_lnet.ln_ping_info = pinfo;
1506         return 0;
1507 }
1508
1509 static void
1510 lnet_destroy_ping_info(void)
1511 {
1512         lnet_ni_t *ni;
1513
1514         LNET_LOCK();
1515
1516         cfs_list_for_each_entry (ni, &the_lnet.ln_nis, ni_list) {
1517                 ni->ni_status = NULL;
1518         }
1519
1520         LNET_UNLOCK();
1521
1522         LIBCFS_FREE(the_lnet.ln_ping_info,
1523                     offsetof(lnet_ping_info_t,
1524                              pi_ni[the_lnet.ln_ping_info->pi_nnis]));
1525         the_lnet.ln_ping_info = NULL;
1526         return;
1527 }
1528
1529 int
1530 lnet_ping_target_init(void)
1531 {
1532         lnet_md_t         md = {0};
1533         lnet_handle_me_t  meh;
1534         lnet_process_id_t id;
1535         int               rc;
1536         int               rc2;
1537         int               infosz;
1538
1539         rc = lnet_create_ping_info();
1540         if (rc != 0)
1541                 return rc;
1542
1543         /* We can have a tiny EQ since we only need to see the unlink event on
1544          * teardown, which by definition is the last one! */
1545         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1546         if (rc != 0) {
1547                 CERROR("Can't allocate ping EQ: %d\n", rc);
1548                 goto failed_0;
1549         }
1550
1551         memset(&id, 0, sizeof(lnet_process_id_t));
1552         id.nid = LNET_NID_ANY;
1553         id.pid = LNET_PID_ANY;
1554
1555         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1556                           LNET_PROTO_PING_MATCHBITS, 0,
1557                           LNET_UNLINK, LNET_INS_AFTER,
1558                           &meh);
1559         if (rc != 0) {
1560                 CERROR("Can't create ping ME: %d\n", rc);
1561                 goto failed_1;
1562         }
1563
1564         /* initialize md content */
1565         infosz = offsetof(lnet_ping_info_t,
1566                           pi_ni[the_lnet.ln_ping_info->pi_nnis]);
1567         md.start     = the_lnet.ln_ping_info;
1568         md.length    = infosz;
1569         md.threshold = LNET_MD_THRESH_INF;
1570         md.max_size  = 0;
1571         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1572                        LNET_MD_MANAGE_REMOTE;
1573         md.user_ptr  = NULL;
1574         md.eq_handle = the_lnet.ln_ping_target_eq;
1575
1576         rc = LNetMDAttach(meh, md,
1577                           LNET_RETAIN,
1578                           &the_lnet.ln_ping_target_md);
1579         if (rc != 0) {
1580                 CERROR("Can't attach ping MD: %d\n", rc);
1581                 goto failed_2;
1582         }
1583
1584         return 0;
1585
1586  failed_2:
1587         rc2 = LNetMEUnlink(meh);
1588         LASSERT (rc2 == 0);
1589  failed_1:
1590         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1591         LASSERT (rc2 == 0);
1592  failed_0:
1593         lnet_destroy_ping_info();
1594         return rc;
1595 }
1596
1597 void
1598 lnet_ping_target_fini(void)
1599 {
1600         lnet_event_t    event;
1601         int             rc;
1602         int             which;
1603         int             timeout_ms = 1000;
1604         cfs_sigset_t    blocked = cfs_block_allsigs();
1605
1606         LNetMDUnlink(the_lnet.ln_ping_target_md);
1607         /* NB md could be busy; this just starts the unlink */
1608
1609         for (;;) {
1610                 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1611                                 timeout_ms, &event, &which);
1612
1613                 /* I expect overflow... */
1614                 LASSERT (rc >= 0 || rc == -EOVERFLOW);
1615
1616                 if (rc == 0) {
1617                         /* timed out: provide a diagnostic */
1618                         CWARN("Still waiting for ping MD to unlink\n");
1619                         timeout_ms *= 2;
1620                         continue;
1621                 }
1622
1623                 /* Got a valid event */
1624                 if (event.unlinked)
1625                         break;
1626         }
1627
1628         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1629         LASSERT (rc == 0);
1630         lnet_destroy_ping_info();
1631         cfs_restore_sigs(blocked);
1632 }
1633
1634 int
1635 lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1636 {
1637         lnet_handle_eq_t     eqh;
1638         lnet_handle_md_t     mdh;
1639         lnet_event_t         event;
1640         lnet_md_t            md = {0};
1641         int                  which;
1642         int                  unlinked = 0;
1643         int                  replied = 0;
1644         const int            a_long_time = 60000; /* mS */
1645         int                  infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
1646         lnet_ping_info_t    *info;
1647         lnet_process_id_t    tmpid;
1648         int                  i;
1649         int                  nob;
1650         int                  rc;
1651         int                  rc2;
1652         cfs_sigset_t         blocked;
1653
1654         if (n_ids <= 0 ||
1655             id.nid == LNET_NID_ANY ||
1656             timeout_ms > 500000 ||              /* arbitrary limit! */
1657             n_ids > 20)                         /* arbitrary limit! */
1658                 return -EINVAL;
1659
1660         if (id.pid == LNET_PID_ANY)
1661                 id.pid = LUSTRE_SRV_LNET_PID;
1662
1663         LIBCFS_ALLOC(info, infosz);
1664         if (info == NULL)
1665                 return -ENOMEM;
1666
1667         /* NB 2 events max (including any unlink event) */
1668         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1669         if (rc != 0) {
1670                 CERROR("Can't allocate EQ: %d\n", rc);
1671                 goto out_0;
1672         }
1673
1674         /* initialize md content */
1675         md.start     = info;
1676         md.length    = infosz;
1677         md.threshold = 2; /*GET/REPLY*/
1678         md.max_size  = 0;
1679         md.options   = LNET_MD_TRUNCATE;
1680         md.user_ptr  = NULL;
1681         md.eq_handle = eqh;
1682
1683         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
1684         if (rc != 0) {
1685                 CERROR("Can't bind MD: %d\n", rc);
1686                 goto out_1;
1687         }
1688
1689         rc = LNetGet(LNET_NID_ANY, mdh, id,
1690                      LNET_RESERVED_PORTAL,
1691                      LNET_PROTO_PING_MATCHBITS, 0);
1692
1693         if (rc != 0) {
1694                 /* Don't CERROR; this could be deliberate! */
1695
1696                 rc2 = LNetMDUnlink(mdh);
1697                 LASSERT (rc2 == 0);
1698
1699                 /* NB must wait for the UNLINK event below... */
1700                 unlinked = 1;
1701                 timeout_ms = a_long_time;
1702         }
1703
1704         do {
1705                 /* MUST block for unlink to complete */
1706                 if (unlinked)
1707                         blocked = cfs_block_allsigs();
1708
1709                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
1710
1711                 if (unlinked)
1712                         cfs_restore_sigs(blocked);
1713
1714                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
1715                        (rc2 <= 0) ? -1 : event.type,
1716                        (rc2 <= 0) ? -1 : event.status,
1717                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
1718
1719                 LASSERT (rc2 != -EOVERFLOW);     /* can't miss anything */
1720
1721                 if (rc2 <= 0 || event.status != 0) {
1722                         /* timeout or error */
1723                         if (!replied && rc == 0)
1724                                 rc = (rc2 < 0) ? rc2 :
1725                                      (rc2 == 0) ? -ETIMEDOUT :
1726                                      event.status;
1727
1728                         if (!unlinked) {
1729                                 /* Ensure completion in finite time... */
1730                                 LNetMDUnlink(mdh);
1731                                 /* No assertion (racing with network) */
1732                                 unlinked = 1;
1733                                 timeout_ms = a_long_time;
1734                         } else if (rc2 == 0) {
1735                                 /* timed out waiting for unlink */
1736                                 CWARN("ping %s: late network completion\n",
1737                                       libcfs_id2str(id));
1738                         }
1739                 } else if (event.type == LNET_EVENT_REPLY) {
1740                         replied = 1;
1741                         rc = event.mlength;
1742                 }
1743
1744         } while (rc2 <= 0 || !event.unlinked);
1745
1746         if (!replied) {
1747                 if (rc >= 0)
1748                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
1749                               libcfs_id2str(id));
1750                 rc = -EIO;
1751                 goto out_1;
1752         }
1753
1754         nob = rc;
1755         LASSERT (nob >= 0 && nob <= infosz);
1756
1757         rc = -EPROTO;                           /* if I can't parse... */
1758
1759         if (nob < 8) {
1760                 /* can't check magic/version */
1761                 CERROR("%s: ping info too short %d\n",
1762                        libcfs_id2str(id), nob);
1763                 goto out_1;
1764         }
1765
1766         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
1767                 lnet_swap_pinginfo(info);
1768         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
1769                 CERROR("%s: Unexpected magic %08x\n",
1770                        libcfs_id2str(id), info->pi_magic);
1771                 goto out_1;
1772         }
1773
1774         if (info->pi_version != LNET_PROTO_PING_VERSION) {
1775                 CERROR("%s: Unexpected version 0x%x\n",
1776                        libcfs_id2str(id), info->pi_version);
1777                 goto out_1;
1778         }
1779
1780         if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
1781                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
1782                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
1783                 goto out_1;
1784         }
1785
1786         if (info->pi_nnis < n_ids)
1787                 n_ids = info->pi_nnis;
1788
1789         if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
1790                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
1791                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
1792                 goto out_1;
1793         }
1794
1795         rc = -EFAULT;                           /* If I SEGV... */
1796
1797         for (i = 0; i < n_ids; i++) {
1798                 tmpid.pid = info->pi_pid;
1799                 tmpid.nid = info->pi_ni[i].ns_nid;
1800 #ifdef __KERNEL__
1801                 if (cfs_copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
1802                         goto out_1;
1803 #else
1804                 ids[i] = tmpid;
1805 #endif
1806         }
1807         rc = info->pi_nnis;
1808
1809  out_1:
1810         rc2 = LNetEQFree(eqh);
1811         if (rc2 != 0)
1812                 CERROR("rc2 %d\n", rc2);
1813         LASSERT (rc2 == 0);
1814
1815  out_0:
1816         LIBCFS_FREE(info, infosz);
1817         return rc;
1818 }