lnet/lnet/api-ni.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2013, Intel Corporation.
  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 #ifdef __KERNEL__
  40 #include <linux/log2.h>
  41 #endif
  42
  43 #ifdef __KERNEL__
  44 #define D_LNI D_CONSOLE
  45 #else
  46 #define D_LNI D_CONFIG
  47 #endif
  48
  49 lnet_t      the_lnet;                           /* THE state of the network */
  50 EXPORT_SYMBOL(the_lnet);
  51
  52 #ifdef __KERNEL__
  53
  54 static char *ip2nets = "";
  55 CFS_MODULE_PARM(ip2nets, "s", charp, 0444,
  56                 "LNET network <- IP table");
  57
  58 static char *networks = "";
  59 CFS_MODULE_PARM(networks, "s", charp, 0444,
  60                 "local networks");
  61
  62 static char *routes = "";
  63 CFS_MODULE_PARM(routes, "s", charp, 0444,
  64                 "routes to non-local networks");
  65
  66 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  67 CFS_MODULE_PARM(rnet_htable_size, "i", int, 0444,
  68                 "size of remote network hash table");
  69
  70 char *
  71 lnet_get_routes(void)
  72 {
  73         return routes;
  74 }
  75
  76 char *
  77 lnet_get_networks(void)
  78 {
  79         char   *nets;
  80         int     rc;
  81
  82         if (*networks != 0 && *ip2nets != 0) {
  83                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  84                                    "'ip2nets' but not both at once\n");
  85                 return NULL;
  86         }
  87
  88         if (*ip2nets != 0) {
  89                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  90                 return (rc == 0) ? nets : NULL;
  91         }
  92
  93         if (*networks != 0)
  94                 return networks;
  95
  96         return "tcp";
  97 }
  98
  99 void
 100 lnet_init_locks(void)
 101 {
 102         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 103         cfs_waitq_init(&the_lnet.ln_eq_waitq);
 104         mutex_init(&the_lnet.ln_lnd_mutex);
 105         mutex_init(&the_lnet.ln_api_mutex);
 106 }
 107
 108 void
 109 lnet_fini_locks(void)
 110 {
 111 }
 112
 113 #else
 114
 115 char *
 116 lnet_get_routes(void)
 117 {
 118         char *str = getenv("LNET_ROUTES");
 119
 120         return (str == NULL) ? "" : str;
 121 }
 122
 123 char *
 124 lnet_get_networks (void)
 125 {
 126         static char       default_networks[256];
 127         char             *networks = getenv ("LNET_NETWORKS");
 128         char             *str;
 129         char             *sep;
 130         int               len;
 131         int               nob;
 132         cfs_list_t       *tmp;
 133
 134         if (networks != NULL)
 135                 return networks;
 136
 137         /* In userland, the default 'networks=' is the list of known net types */
 138         len = sizeof(default_networks);
 139         str = default_networks;
 140         *str = 0;
 141         sep = "";
 142
 143         cfs_list_for_each (tmp, &the_lnet.ln_lnds) {
 144                 lnd_t *lnd = cfs_list_entry(tmp, lnd_t, lnd_list);
 145
 146                 nob = snprintf(str, len, "%s%s", sep,
 147                                libcfs_lnd2str(lnd->lnd_type));
 148                 if (nob >= len) {
 149                         /* overflowed the string; leave it where it was */
 150                         *str = 0;
 151                         break;
 152                 }
 153                 len -= nob;
 154                 str += nob;
 155                 sep = ",";
 156         }
 157
 158         return default_networks;
 159 }
 160
 161 # ifndef HAVE_LIBPTHREAD
 162
 163 void lnet_init_locks(void)
 164 {
 165         the_lnet.ln_eq_wait_lock = 0;
 166         the_lnet.ln_lnd_mutex = 0;
 167         the_lnet.ln_api_mutex = 0;
 168 }
 169
 170 void lnet_fini_locks(void)
 171 {
 172         LASSERT(the_lnet.ln_api_mutex == 0);
 173         LASSERT(the_lnet.ln_lnd_mutex == 0);
 174         LASSERT(the_lnet.ln_eq_wait_lock == 0);
 175 }
 176
 177 # else
 178
 179 void lnet_init_locks(void)
 180 {
 181         pthread_cond_init(&the_lnet.ln_eq_cond, NULL);
 182         pthread_mutex_init(&the_lnet.ln_eq_wait_lock, NULL);
 183         pthread_mutex_init(&the_lnet.ln_lnd_mutex, NULL);
 184         pthread_mutex_init(&the_lnet.ln_api_mutex, NULL);
 185 }
 186
 187 void lnet_fini_locks(void)
 188 {
 189         pthread_mutex_destroy(&the_lnet.ln_api_mutex);
 190         pthread_mutex_destroy(&the_lnet.ln_lnd_mutex);
 191         pthread_mutex_destroy(&the_lnet.ln_eq_wait_lock);
 192         pthread_cond_destroy(&the_lnet.ln_eq_cond);
 193 }
 194
 195 # endif
 196 #endif
 197
 198 static int
 199 lnet_create_remote_nets_table(void)
 200 {
 201         int             i;
 202         cfs_list_t      *hash;
 203
 204         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 205         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 206         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 207         if (hash == NULL) {
 208                 CERROR("Failed to create remote nets hash table\n");
 209                 return -ENOMEM;
 210         }
 211
 212         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 213                 CFS_INIT_LIST_HEAD(&hash[i]);
 214         the_lnet.ln_remote_nets_hash = hash;
 215         return 0;
 216 }
 217
 218 static void
 219 lnet_destroy_remote_nets_table(void)
 220 {
 221         int             i;
 222         cfs_list_t      *hash;
 223
 224         if (the_lnet.ln_remote_nets_hash == NULL)
 225                 return;
 226
 227         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 228                 LASSERT(cfs_list_empty(&the_lnet.ln_remote_nets_hash[i]));
 229
 230         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 231                     LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 232         the_lnet.ln_remote_nets_hash = NULL;
 233 }
 234
 235 static void
 236 lnet_destroy_locks(void)
 237 {
 238         if (the_lnet.ln_res_lock != NULL) {
 239                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 240                 the_lnet.ln_res_lock = NULL;
 241         }
 242
 243         if (the_lnet.ln_net_lock != NULL) {
 244                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 245                 the_lnet.ln_net_lock = NULL;
 246         }
 247
 248         lnet_fini_locks();
 249 }
 250
 251 static int
 252 lnet_create_locks(void)
 253 {
 254         lnet_init_locks();
 255
 256         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 257         if (the_lnet.ln_res_lock == NULL)
 258                 goto failed;
 259
 260         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 261         if (the_lnet.ln_net_lock == NULL)
 262                 goto failed;
 263
 264         return 0;
 265
 266  failed:
 267         lnet_destroy_locks();
 268         return -ENOMEM;
 269 }
 270
 271 void lnet_assert_wire_constants (void)
 272 {
 273         /* Wire protocol assertions generated by 'wirecheck'
 274          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 275          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 276          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 277
 278         /* Constants... */
 279         CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 280         CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
 281         CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
 282         CLASSERT (LNET_MSG_ACK == 0);
 283         CLASSERT (LNET_MSG_PUT == 1);
 284         CLASSERT (LNET_MSG_GET == 2);
 285         CLASSERT (LNET_MSG_REPLY == 3);
 286         CLASSERT (LNET_MSG_HELLO == 4);
 287
 288         /* Checks for struct ptl_handle_wire_t */
 289         CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
 290         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 291         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 292         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 293         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 294
 295         /* Checks for struct lnet_magicversion_t */
 296         CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
 297         CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
 298         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 299         CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
 300         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 301         CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 302         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 303
 304         /* Checks for struct lnet_hdr_t */
 305         CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
 306         CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 307         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 308         CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
 309         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 310         CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 311         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 312         CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
 313         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 314         CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
 315         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 316         CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
 317         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 318         CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
 319         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 320
 321         /* Ack */
 322         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 323         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 324         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 325         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 326         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 327         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 328
 329         /* Put */
 330         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 331         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 332         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 333         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 334         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 335         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 336         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 337         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 338         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 339         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 340
 341         /* Get */
 342         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 343         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 344         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 345         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 346         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 347         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 348         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 349         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 350         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 351         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 352
 353         /* Reply */
 354         CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 355         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 356
 357         /* Hello */
 358         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 359         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 360         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 361         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 362 }
 363
 364 lnd_t *
 365 lnet_find_lnd_by_type (int type)
 366 {
 367         lnd_t              *lnd;
 368         cfs_list_t         *tmp;
 369
 370         /* holding lnd mutex */
 371         cfs_list_for_each (tmp, &the_lnet.ln_lnds) {
 372                 lnd = cfs_list_entry(tmp, lnd_t, lnd_list);
 373
 374                 if ((int)lnd->lnd_type == type)
 375                         return lnd;
 376         }
 377
 378         return NULL;
 379 }
 380
 381 void
 382 lnet_register_lnd (lnd_t *lnd)
 383 {
 384         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
 385
 386         LASSERT (the_lnet.ln_init);
 387         LASSERT (libcfs_isknown_lnd(lnd->lnd_type));
 388         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 389
 390         cfs_list_add_tail (&lnd->lnd_list, &the_lnet.ln_lnds);
 391         lnd->lnd_refcount = 0;
 392
 393         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 394
 395         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
 396 }
 397 EXPORT_SYMBOL(lnet_register_lnd);
 398
 399 void
 400 lnet_unregister_lnd (lnd_t *lnd)
 401 {
 402         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
 403
 404         LASSERT (the_lnet.ln_init);
 405         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 406         LASSERT (lnd->lnd_refcount == 0);
 407
 408         cfs_list_del (&lnd->lnd_list);
 409         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 410
 411         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
 412 }
 413 EXPORT_SYMBOL(lnet_unregister_lnd);
 414
 415 void
 416 lnet_counters_get(lnet_counters_t *counters)
 417 {
 418         lnet_counters_t *ctr;
 419         int             i;
 420
 421         memset(counters, 0, sizeof(*counters));
 422
 423         lnet_net_lock(LNET_LOCK_EX);
 424
 425         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 426                 counters->msgs_max     += ctr->msgs_max;
 427                 counters->msgs_alloc   += ctr->msgs_alloc;
 428                 counters->errors       += ctr->errors;
 429                 counters->send_count   += ctr->send_count;
 430                 counters->recv_count   += ctr->recv_count;
 431                 counters->route_count  += ctr->route_count;
 432                 counters->drop_length  += ctr->drop_length;
 433                 counters->send_length  += ctr->send_length;
 434                 counters->recv_length  += ctr->recv_length;
 435                 counters->route_length += ctr->route_length;
 436                 counters->drop_length  += ctr->drop_length;
 437
 438         }
 439         lnet_net_unlock(LNET_LOCK_EX);
 440 }
 441 EXPORT_SYMBOL(lnet_counters_get);
 442
 443 void
 444 lnet_counters_reset(void)
 445 {
 446         lnet_counters_t *counters;
 447         int             i;
 448
 449         lnet_net_lock(LNET_LOCK_EX);
 450
 451         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 452                 memset(counters, 0, sizeof(lnet_counters_t));
 453
 454         lnet_net_unlock(LNET_LOCK_EX);
 455 }
 456 EXPORT_SYMBOL(lnet_counters_reset);
 457
 458 #ifdef LNET_USE_LIB_FREELIST
 459
 460 int
 461 lnet_freelist_init (lnet_freelist_t *fl, int n, int size)
 462 {
 463         char *space;
 464
 465         LASSERT (n > 0);
 466
 467         size += offsetof (lnet_freeobj_t, fo_contents);
 468
 469         LIBCFS_ALLOC(space, n * size);
 470         if (space == NULL)
 471                 return (-ENOMEM);
 472
 473         CFS_INIT_LIST_HEAD (&fl->fl_list);
 474         fl->fl_objs = space;
 475         fl->fl_nobjs = n;
 476         fl->fl_objsize = size;
 477
 478         do
 479         {
 480                 memset (space, 0, size);
 481                 cfs_list_add ((cfs_list_t *)space, &fl->fl_list);
 482                 space += size;
 483         } while (--n != 0);
 484
 485         return (0);
 486 }
 487
 488 void
 489 lnet_freelist_fini (lnet_freelist_t *fl)
 490 {
 491         cfs_list_t       *el;
 492         int               count;
 493
 494         if (fl->fl_nobjs == 0)
 495                 return;
 496
 497         count = 0;
 498         for (el = fl->fl_list.next; el != &fl->fl_list; el = el->next)
 499                 count++;
 500
 501         LASSERT (count == fl->fl_nobjs);
 502
 503         LIBCFS_FREE(fl->fl_objs, fl->fl_nobjs * fl->fl_objsize);
 504         memset (fl, 0, sizeof (*fl));
 505 }
 506
 507 #endif /* LNET_USE_LIB_FREELIST */
 508
 509 __u64
 510 lnet_create_interface_cookie (void)
 511 {
 512         /* NB the interface cookie in wire handles guards against delayed
 513          * replies and ACKs appearing valid after reboot. Initialisation time,
 514          * even if it's only implemented to millisecond resolution is probably
 515          * easily good enough. */
 516         struct timeval tv;
 517         __u64          cookie;
 518 #ifndef __KERNEL__
 519         int            rc = gettimeofday (&tv, NULL);
 520         LASSERT (rc == 0);
 521 #else
 522         cfs_gettimeofday(&tv);
 523 #endif
 524         cookie = tv.tv_sec;
 525         cookie *= 1000000;
 526         cookie += tv.tv_usec;
 527         return cookie;
 528 }
 529
 530 static char *
 531 lnet_res_type2str(int type)
 532 {
 533         switch (type) {
 534         default:
 535                 LBUG();
 536         case LNET_COOKIE_TYPE_MD:
 537                 return "MD";
 538         case LNET_COOKIE_TYPE_ME:
 539                 return "ME";
 540         case LNET_COOKIE_TYPE_EQ:
 541                 return "EQ";
 542         }
 543 }
 544
 545 void
 546 lnet_res_container_cleanup(struct lnet_res_container *rec)
 547 {
 548         int     count = 0;
 549
 550         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 551                 return;
 552
 553         while (!cfs_list_empty(&rec->rec_active)) {
 554                 cfs_list_t *e = rec->rec_active.next;
 555
 556                 cfs_list_del_init(e);
 557                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 558                         lnet_eq_free(cfs_list_entry(e, lnet_eq_t, eq_list));
 559
 560                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 561                         lnet_md_free(cfs_list_entry(e, lnet_libmd_t, md_list));
 562
 563                 } else { /* NB: Active MEs should be attached on portals */
 564                         LBUG();
 565                 }
 566                 count++;
 567         }
 568
 569         if (count > 0) {
 570                 /* Found alive MD/ME/EQ, user really should unlink/free
 571                  * all of them before finalize LNet, but if someone didn't,
 572                  * we have to recycle garbage for him */
 573                 CERROR("%d active elements on exit of %s container\n",
 574                        count, lnet_res_type2str(rec->rec_type));
 575         }
 576
 577 #ifdef LNET_USE_LIB_FREELIST
 578         lnet_freelist_fini(&rec->rec_freelist);
 579 #endif
 580         if (rec->rec_lh_hash != NULL) {
 581                 LIBCFS_FREE(rec->rec_lh_hash,
 582                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 583                 rec->rec_lh_hash = NULL;
 584         }
 585
 586         rec->rec_type = 0; /* mark it as finalized */
 587 }
 588
 589 int
 590 lnet_res_container_setup(struct lnet_res_container *rec,
 591                          int cpt, int type, int objnum, int objsz)
 592 {
 593         int     rc = 0;
 594         int     i;
 595
 596         LASSERT(rec->rec_type == 0);
 597
 598         rec->rec_type = type;
 599         CFS_INIT_LIST_HEAD(&rec->rec_active);
 600
 601 #ifdef LNET_USE_LIB_FREELIST
 602         memset(&rec->rec_freelist, 0, sizeof(rec->rec_freelist));
 603         rc = lnet_freelist_init(&rec->rec_freelist, objnum, objsz);
 604         if (rc != 0)
 605                 goto out;
 606 #endif
 607         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 608
 609         /* Arbitrary choice of hash table size */
 610         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 611                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 612         if (rec->rec_lh_hash == NULL) {
 613                 rc = -ENOMEM;
 614                 goto out;
 615         }
 616
 617         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 618                 CFS_INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 619
 620         return 0;
 621
 622 out:
 623         CERROR("Failed to setup %s resource container\n",
 624                lnet_res_type2str(type));
 625         lnet_res_container_cleanup(rec);
 626         return rc;
 627 }
 628
 629 static void
 630 lnet_res_containers_destroy(struct lnet_res_container **recs)
 631 {
 632         struct lnet_res_container       *rec;
 633         int                             i;
 634
 635         cfs_percpt_for_each(rec, i, recs)
 636                 lnet_res_container_cleanup(rec);
 637
 638         cfs_percpt_free(recs);
 639 }
 640
 641 static struct lnet_res_container **
 642 lnet_res_containers_create(int type, int objnum, int objsz)
 643 {
 644         struct lnet_res_container       **recs;
 645         struct lnet_res_container       *rec;
 646         int                             rc;
 647         int                             i;
 648
 649         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 650         if (recs == NULL) {
 651                 CERROR("Failed to allocate %s resource containers\n",
 652                        lnet_res_type2str(type));
 653                 return NULL;
 654         }
 655
 656         cfs_percpt_for_each(rec, i, recs) {
 657                 rc = lnet_res_container_setup(rec, i, type, objnum, objsz);
 658                 if (rc != 0) {
 659                         lnet_res_containers_destroy(recs);
 660                         return NULL;
 661                 }
 662         }
 663
 664         return recs;
 665 }
 666
 667 lnet_libhandle_t *
 668 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 669 {
 670         /* ALWAYS called with lnet_res_lock held */
 671         cfs_list_t              *head;
 672         lnet_libhandle_t        *lh;
 673         unsigned int            hash;
 674
 675         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
 676                 return NULL;
 677
 678         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 679         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 680
 681         cfs_list_for_each_entry(lh, head, lh_hash_chain) {
 682                 if (lh->lh_cookie == cookie)
 683                         return lh;
 684         }
 685
 686         return NULL;
 687 }
 688
 689 void
 690 lnet_res_lh_initialize(struct lnet_res_container *rec, lnet_libhandle_t *lh)
 691 {
 692         /* ALWAYS called with lnet_res_lock held */
 693         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 694         unsigned int    hash;
 695
 696         lh->lh_cookie = rec->rec_lh_cookie;
 697         rec->rec_lh_cookie += 1 << ibits;
 698
 699         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 700
 701         cfs_list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 702 }
 703
 704 #ifndef __KERNEL__
 705 /**
 706  * Reserved API - do not use.
 707  * Temporary workaround to allow uOSS and test programs force server
 708  * mode in userspace. See comments near ln_server_mode_flag in
 709  * lnet/lib-types.h */
 710
 711 void
 712 lnet_server_mode() {
 713         the_lnet.ln_server_mode_flag = 1;
 714 }
 715 #endif
 716
 717 int lnet_unprepare(void);
 718
 719 int
 720 lnet_prepare(lnet_pid_t requested_pid)
 721 {
 722         /* Prepare to bring up the network */
 723         struct lnet_res_container **recs;
 724         int                       rc = 0;
 725
 726         LASSERT (the_lnet.ln_refcount == 0);
 727
 728         the_lnet.ln_routing = 0;
 729
 730 #ifdef __KERNEL__
 731         LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 732         the_lnet.ln_pid = requested_pid;
 733 #else
 734         if (the_lnet.ln_server_mode_flag) {/* server case (uOSS) */
 735                 LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 736
 737                 if (cfs_curproc_uid())/* Only root can run user-space server */
 738                         return -EPERM;
 739                 the_lnet.ln_pid = requested_pid;
 740
 741         } else {/* client case (liblustre) */
 742
 743                 /* My PID must be unique on this node and flag I'm userspace */
 744                 the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
 745         }
 746 #endif
 747
 748         CFS_INIT_LIST_HEAD(&the_lnet.ln_test_peers);
 749         CFS_INIT_LIST_HEAD(&the_lnet.ln_nis);
 750         CFS_INIT_LIST_HEAD(&the_lnet.ln_nis_cpt);
 751         CFS_INIT_LIST_HEAD(&the_lnet.ln_nis_zombie);
 752         CFS_INIT_LIST_HEAD(&the_lnet.ln_routers);
 753
 754         rc = lnet_create_remote_nets_table();
 755         if (rc != 0)
 756                 goto failed;
 757
 758         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 759
 760         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
 761                                                 sizeof(lnet_counters_t));
 762         if (the_lnet.ln_counters == NULL) {
 763                 CERROR("Failed to allocate counters for LNet\n");
 764                 rc = -ENOMEM;
 765                 goto failed;
 766         }
 767
 768         rc = lnet_peer_tables_create();
 769         if (rc != 0)
 770                 goto failed;
 771
 772         rc = lnet_msg_containers_create();
 773         if (rc != 0)
 774                 goto failed;
 775
 776         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 777                                       LNET_COOKIE_TYPE_EQ, LNET_FL_MAX_EQS,
 778                                       sizeof(lnet_eq_t));
 779         if (rc != 0)
 780                 goto failed;
 781
 782         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME, LNET_FL_MAX_MES,
 783                                           sizeof(lnet_me_t));
 784         if (recs == NULL)
 785                 goto failed;
 786
 787         the_lnet.ln_me_containers = recs;
 788
 789         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD, LNET_FL_MAX_MDS,
 790                                           sizeof(lnet_libmd_t));
 791         if (recs == NULL)
 792                 goto failed;
 793
 794         the_lnet.ln_md_containers = recs;
 795
 796         rc = lnet_portals_create();
 797         if (rc != 0) {
 798                 CERROR("Failed to create portals for LNet: %d\n", rc);
 799                 goto failed;
 800         }
 801
 802         return 0;
 803
 804  failed:
 805         lnet_unprepare();
 806         return rc;
 807 }
 808
 809 int
 810 lnet_unprepare (void)
 811 {
 812         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 813          * have shut down already, so it is safe to unlink and free all
 814          * descriptors, even those that appear committed to a network op (eg MD
 815          * with non-zero pending count) */
 816
 817         lnet_fail_nid(LNET_NID_ANY, 0);
 818
 819         LASSERT(the_lnet.ln_refcount == 0);
 820         LASSERT(cfs_list_empty(&the_lnet.ln_test_peers));
 821         LASSERT(cfs_list_empty(&the_lnet.ln_nis));
 822         LASSERT(cfs_list_empty(&the_lnet.ln_nis_cpt));
 823         LASSERT(cfs_list_empty(&the_lnet.ln_nis_zombie));
 824
 825         lnet_portals_destroy();
 826
 827         if (the_lnet.ln_md_containers != NULL) {
 828                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 829                 the_lnet.ln_md_containers = NULL;
 830         }
 831
 832         if (the_lnet.ln_me_containers != NULL) {
 833                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 834                 the_lnet.ln_me_containers = NULL;
 835         }
 836
 837         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 838
 839         lnet_msg_containers_destroy();
 840         lnet_peer_tables_destroy();
 841         lnet_rtrpools_free();
 842
 843         if (the_lnet.ln_counters != NULL) {
 844                 cfs_percpt_free(the_lnet.ln_counters);
 845                 the_lnet.ln_counters = NULL;
 846         }
 847         lnet_destroy_remote_nets_table();
 848
 849         return 0;
 850 }
 851
 852 lnet_ni_t  *
 853 lnet_net2ni_locked(__u32 net, int cpt)
 854 {
 855         cfs_list_t      *tmp;
 856         lnet_ni_t       *ni;
 857
 858         LASSERT(cpt != LNET_LOCK_EX);
 859
 860         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 861                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 862
 863                 if (LNET_NIDNET(ni->ni_nid) == net) {
 864                         lnet_ni_addref_locked(ni, cpt);
 865                         return ni;
 866                 }
 867         }
 868
 869         return NULL;
 870 }
 871
 872 lnet_ni_t *
 873 lnet_net2ni(__u32 net)
 874 {
 875         lnet_ni_t *ni;
 876
 877         lnet_net_lock(0);
 878         ni = lnet_net2ni_locked(net, 0);
 879         lnet_net_unlock(0);
 880
 881         return ni;
 882 }
 883 EXPORT_SYMBOL(lnet_net2ni);
 884
 885 static unsigned int
 886 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 887 {
 888         __u64           key = nid;
 889         unsigned int    val;
 890
 891         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 892
 893         if (number == 1)
 894                 return 0;
 895
 896         val = cfs_hash_long(key, LNET_CPT_BITS);
 897         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 898         if (val < number)
 899                 return val;
 900
 901         return (unsigned int)(key + val + (val >> 1)) % number;
 902 }
 903
 904 int
 905 lnet_cpt_of_nid_locked(lnet_nid_t nid)
 906 {
 907         struct lnet_ni *ni;
 908
 909         /* must called with hold of lnet_net_lock */
 910         if (LNET_CPT_NUMBER == 1)
 911                 return 0; /* the only one */
 912
 913         /* take lnet_net_lock(any) would be OK */
 914         if (!cfs_list_empty(&the_lnet.ln_nis_cpt)) {
 915                 cfs_list_for_each_entry(ni, &the_lnet.ln_nis_cpt, ni_cptlist) {
 916                         if (LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid))
 917                                 continue;
 918
 919                         LASSERT(ni->ni_cpts != NULL);
 920                         return ni->ni_cpts[lnet_nid_cpt_hash
 921                                            (nid, ni->ni_ncpts)];
 922                 }
 923         }
 924
 925         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 926 }
 927
 928 int
 929 lnet_cpt_of_nid(lnet_nid_t nid)
 930 {
 931         int     cpt;
 932         int     cpt2;
 933
 934         if (LNET_CPT_NUMBER == 1)
 935                 return 0; /* the only one */
 936
 937         if (cfs_list_empty(&the_lnet.ln_nis_cpt))
 938                 return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 939
 940         cpt = lnet_net_lock_current();
 941         cpt2 = lnet_cpt_of_nid_locked(nid);
 942         lnet_net_unlock(cpt);
 943
 944         return cpt2;
 945 }
 946 EXPORT_SYMBOL(lnet_cpt_of_nid);
 947
 948 int
 949 lnet_islocalnet(__u32 net)
 950 {
 951         struct lnet_ni  *ni;
 952         int             cpt;
 953
 954         cpt = lnet_net_lock_current();
 955
 956         ni = lnet_net2ni_locked(net, cpt);
 957         if (ni != NULL)
 958                 lnet_ni_decref_locked(ni, cpt);
 959
 960         lnet_net_unlock(cpt);
 961
 962         return ni != NULL;
 963 }
 964
 965 lnet_ni_t  *
 966 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 967 {
 968         struct lnet_ni  *ni;
 969         cfs_list_t      *tmp;
 970
 971         LASSERT(cpt != LNET_LOCK_EX);
 972
 973         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 974                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 975
 976                 if (ni->ni_nid == nid) {
 977                         lnet_ni_addref_locked(ni, cpt);
 978                         return ni;
 979                 }
 980         }
 981
 982         return NULL;
 983 }
 984
 985 int
 986 lnet_islocalnid(lnet_nid_t nid)
 987 {
 988         struct lnet_ni  *ni;
 989         int             cpt;
 990
 991         cpt = lnet_net_lock_current();
 992         ni = lnet_nid2ni_locked(nid, cpt);
 993         if (ni != NULL)
 994                 lnet_ni_decref_locked(ni, cpt);
 995         lnet_net_unlock(cpt);
 996
 997         return ni != NULL;
 998 }
 999
1000 int
1001 lnet_count_acceptor_nis (void)
1002 {
1003         /* Return the # of NIs that need the acceptor. */
1004         int             count = 0;
1005 #if defined(__KERNEL__) || defined(HAVE_LIBPTHREAD)
1006         cfs_list_t      *tmp;
1007         struct lnet_ni  *ni;
1008         int             cpt;
1009
1010         cpt = lnet_net_lock_current();
1011         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
1012                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
1013
1014                 if (ni->ni_lnd->lnd_accept != NULL)
1015                         count++;
1016         }
1017
1018         lnet_net_unlock(cpt);
1019
1020 #endif /* defined(__KERNEL__) || defined(HAVE_LIBPTHREAD) */
1021         return count;
1022 }
1023
1024 static int
1025 lnet_ni_tq_credits(lnet_ni_t *ni)
1026 {
1027         int     credits;
1028
1029         LASSERT(ni->ni_ncpts >= 1);
1030
1031         if (ni->ni_ncpts == 1)
1032                 return ni->ni_maxtxcredits;
1033
1034         credits = ni->ni_maxtxcredits / ni->ni_ncpts;
1035         credits = max(credits, 8 * ni->ni_peertxcredits);
1036         credits = min(credits, ni->ni_maxtxcredits);
1037
1038         return credits;
1039 }
1040
1041 void
1042 lnet_shutdown_lndnis (void)
1043 {
1044         int                i;
1045         int                islo;
1046         lnet_ni_t         *ni;
1047
1048         /* NB called holding the global mutex */
1049
1050         /* All quiet on the API front */
1051         LASSERT(!the_lnet.ln_shutdown);
1052         LASSERT(the_lnet.ln_refcount == 0);
1053         LASSERT(cfs_list_empty(&the_lnet.ln_nis_zombie));
1054
1055         lnet_net_lock(LNET_LOCK_EX);
1056         the_lnet.ln_shutdown = 1;       /* flag shutdown */
1057
1058         /* Unlink NIs from the global table */
1059         while (!cfs_list_empty(&the_lnet.ln_nis)) {
1060                 ni = cfs_list_entry(the_lnet.ln_nis.next,
1061                                     lnet_ni_t, ni_list);
1062                 /* move it to zombie list and nobody can find it anymore */
1063                 cfs_list_move(&ni->ni_list, &the_lnet.ln_nis_zombie);
1064                 lnet_ni_decref_locked(ni, 0);   /* drop ln_nis' ref */
1065
1066                 if (!cfs_list_empty(&ni->ni_cptlist)) {
1067                         cfs_list_del_init(&ni->ni_cptlist);
1068                         lnet_ni_decref_locked(ni, 0);
1069                 }
1070         }
1071
1072         /* Drop the cached eqwait NI. */
1073         if (the_lnet.ln_eq_waitni != NULL) {
1074                 lnet_ni_decref_locked(the_lnet.ln_eq_waitni, 0);
1075                 the_lnet.ln_eq_waitni = NULL;
1076         }
1077
1078         /* Drop the cached loopback NI. */
1079         if (the_lnet.ln_loni != NULL) {
1080                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1081                 the_lnet.ln_loni = NULL;
1082         }
1083
1084         lnet_net_unlock(LNET_LOCK_EX);
1085
1086         /* Clear lazy portals and drop delayed messages which hold refs
1087          * on their lnet_msg_t::msg_rxpeer */
1088         for (i = 0; i < the_lnet.ln_nportals; i++)
1089                 LNetClearLazyPortal(i);
1090
1091         /* Clear the peer table and wait for all peers to go (they hold refs on
1092          * their NIs) */
1093         lnet_peer_tables_cleanup();
1094
1095         lnet_net_lock(LNET_LOCK_EX);
1096         /* Now wait for the NI's I just nuked to show up on ln_zombie_nis
1097          * and shut them down in guaranteed thread context */
1098         i = 2;
1099         while (!cfs_list_empty(&the_lnet.ln_nis_zombie)) {
1100                 int     *ref;
1101                 int     j;
1102
1103                 ni = cfs_list_entry(the_lnet.ln_nis_zombie.next,
1104                                     lnet_ni_t, ni_list);
1105                 cfs_list_del_init(&ni->ni_list);
1106                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1107                         if (*ref == 0)
1108                                 continue;
1109                         /* still busy, add it back to zombie list */
1110                         cfs_list_add(&ni->ni_list, &the_lnet.ln_nis_zombie);
1111                         break;
1112                 }
1113
1114                 while (!cfs_list_empty(&ni->ni_list)) {
1115                         lnet_net_unlock(LNET_LOCK_EX);
1116                         ++i;
1117                         if ((i & (-i)) == i) {
1118                                 CDEBUG(D_WARNING,
1119                                        "Waiting for zombie LNI %s\n",
1120                                        libcfs_nid2str(ni->ni_nid));
1121                         }
1122                         cfs_pause(cfs_time_seconds(1));
1123                         lnet_net_lock(LNET_LOCK_EX);
1124                         continue;
1125                 }
1126
1127                 ni->ni_lnd->lnd_refcount--;
1128                 lnet_net_unlock(LNET_LOCK_EX);
1129
1130                 islo = ni->ni_lnd->lnd_type == LOLND;
1131
1132                 LASSERT (!cfs_in_interrupt ());
1133                 (ni->ni_lnd->lnd_shutdown)(ni);
1134
1135                 /* can't deref lnd anymore now; it might have unregistered
1136                  * itself...  */
1137
1138                 if (!islo)
1139                         CDEBUG(D_LNI, "Removed LNI %s\n",
1140                                libcfs_nid2str(ni->ni_nid));
1141
1142                 lnet_ni_free(ni);
1143                 lnet_net_lock(LNET_LOCK_EX);
1144         }
1145
1146         the_lnet.ln_shutdown = 0;
1147         lnet_net_unlock(LNET_LOCK_EX);
1148
1149         if (the_lnet.ln_network_tokens != NULL) {
1150                 LIBCFS_FREE(the_lnet.ln_network_tokens,
1151                             the_lnet.ln_network_tokens_nob);
1152                 the_lnet.ln_network_tokens = NULL;
1153         }
1154 }
1155
1156 int
1157 lnet_startup_lndnis (void)
1158 {
1159         lnd_t                   *lnd;
1160         struct lnet_ni          *ni;
1161         struct lnet_tx_queue    *tq;
1162         cfs_list_t              nilist;
1163         int                     i;
1164         int                rc = 0;
1165         int                lnd_type;
1166         int                nicount = 0;
1167         char              *nets = lnet_get_networks();
1168
1169         CFS_INIT_LIST_HEAD(&nilist);
1170
1171         if (nets == NULL)
1172                 goto failed;
1173
1174         rc = lnet_parse_networks(&nilist, nets);
1175         if (rc != 0)
1176                 goto failed;
1177
1178         while (!cfs_list_empty(&nilist)) {
1179                 ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
1180                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
1181
1182                 LASSERT (libcfs_isknown_lnd(lnd_type));
1183
1184                 if (lnd_type == CIBLND    ||
1185                     lnd_type == OPENIBLND ||
1186                     lnd_type == IIBLND    ||
1187                     lnd_type == VIBLND) {
1188                         CERROR("LND %s obsoleted\n",
1189                                libcfs_lnd2str(lnd_type));
1190                         goto failed;
1191                 }
1192
1193                 LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
1194                 lnd = lnet_find_lnd_by_type(lnd_type);
1195
1196 #ifdef __KERNEL__
1197                 if (lnd == NULL) {
1198                         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1199                         rc = cfs_request_module("%s",
1200                                                 libcfs_lnd2modname(lnd_type));
1201                         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
1202
1203                         lnd = lnet_find_lnd_by_type(lnd_type);
1204                         if (lnd == NULL) {
1205                                 LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1206                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1207                                        libcfs_lnd2str(lnd_type),
1208                                        libcfs_lnd2modname(lnd_type), rc);
1209 #ifndef HAVE_MODULE_LOADING_SUPPORT
1210                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1211                                          "compiled with kernel module "
1212                                          "loading support.");
1213 #endif
1214                                 goto failed;
1215                         }
1216                 }
1217 #else
1218                 if (lnd == NULL) {
1219                         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1220                         CERROR("LND %s not supported\n",
1221                                libcfs_lnd2str(lnd_type));
1222                         goto failed;
1223                 }
1224 #endif
1225
1226                 lnet_net_lock(LNET_LOCK_EX);
1227                 lnd->lnd_refcount++;
1228                 lnet_net_unlock(LNET_LOCK_EX);
1229
1230                 ni->ni_lnd = lnd;
1231
1232                 rc = (lnd->lnd_startup)(ni);
1233
1234                 LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1235
1236                 if (rc != 0) {
1237                         LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
1238                                            "\n",
1239                                            rc, libcfs_lnd2str(lnd->lnd_type));
1240                         lnet_net_lock(LNET_LOCK_EX);
1241                         lnd->lnd_refcount--;
1242                         lnet_net_unlock(LNET_LOCK_EX);
1243                         goto failed;
1244                 }
1245
1246                 LASSERT (ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1247
1248                 cfs_list_del(&ni->ni_list);
1249
1250                 lnet_net_lock(LNET_LOCK_EX);
1251                 /* refcount for ln_nis */
1252                 lnet_ni_addref_locked(ni, 0);
1253                 cfs_list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1254                 if (ni->ni_cpts != NULL) {
1255                         cfs_list_add_tail(&ni->ni_cptlist,
1256                                           &the_lnet.ln_nis_cpt);
1257                         lnet_ni_addref_locked(ni, 0);
1258                 }
1259
1260                 lnet_net_unlock(LNET_LOCK_EX);
1261
1262                 if (lnd->lnd_type == LOLND) {
1263                         lnet_ni_addref(ni);
1264                         LASSERT (the_lnet.ln_loni == NULL);
1265                         the_lnet.ln_loni = ni;
1266                         continue;
1267                 }
1268
1269 #ifndef __KERNEL__
1270                 if (lnd->lnd_wait != NULL) {
1271                         if (the_lnet.ln_eq_waitni == NULL) {
1272                                 lnet_ni_addref(ni);
1273                                 the_lnet.ln_eq_waitni = ni;
1274                         }
1275                 } else {
1276 # ifndef HAVE_LIBPTHREAD
1277                         LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1278                                            "single-threaded runtime\n",
1279                                            libcfs_lnd2str(lnd_type));
1280                         goto failed;
1281 # endif
1282                 }
1283 #endif
1284                 if (ni->ni_peertxcredits == 0 ||
1285                     ni->ni_maxtxcredits == 0) {
1286                         LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1287                                            libcfs_lnd2str(lnd->lnd_type),
1288                                            ni->ni_peertxcredits == 0 ?
1289                                            "" : "per-peer ");
1290                         goto failed;
1291                 }
1292
1293                 cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1294                         tq->tq_credits_min =
1295                         tq->tq_credits_max =
1296                         tq->tq_credits = lnet_ni_tq_credits(ni);
1297                 }
1298
1299                 CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1300                        libcfs_nid2str(ni->ni_nid), ni->ni_peertxcredits,
1301                        lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1302                        ni->ni_peerrtrcredits, ni->ni_peertimeout);
1303
1304                 nicount++;
1305         }
1306
1307         if (the_lnet.ln_eq_waitni != NULL && nicount > 1) {
1308                 lnd_type = the_lnet.ln_eq_waitni->ni_lnd->lnd_type;
1309                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1310                                    "\n",
1311                                    libcfs_lnd2str(lnd_type));
1312                 goto failed;
1313         }
1314
1315         return 0;
1316
1317  failed:
1318         lnet_shutdown_lndnis();
1319
1320         while (!cfs_list_empty(&nilist)) {
1321                 ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
1322                 cfs_list_del(&ni->ni_list);
1323                 lnet_ni_free(ni);
1324         }
1325
1326         return -ENETDOWN;
1327 }
1328
1329 /**
1330  * Initialize LNet library.
1331  *
1332  * Only userspace program needs to call this function - it's automatically
1333  * called in the kernel at module loading time. Caller has to call LNetFini()
1334  * after a call to LNetInit(), if and only if the latter returned 0. It must
1335  * be called exactly once.
1336  *
1337  * \return 0 on success, and -ve on failures.
1338  */
1339 int
1340 LNetInit(void)
1341 {
1342         int     rc;
1343
1344         lnet_assert_wire_constants();
1345         LASSERT(!the_lnet.ln_init);
1346
1347         memset(&the_lnet, 0, sizeof(the_lnet));
1348
1349         /* refer to global cfs_cpt_table for now */
1350         the_lnet.ln_cpt_table   = cfs_cpt_table;
1351         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1352
1353         LASSERT(the_lnet.ln_cpt_number > 0);
1354         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1355                 /* we are under risk of consuming all lh_cookie */
1356                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
1357                        "please change setting of CPT-table and retry\n",
1358                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1359                 return -1;
1360         }
1361
1362         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1363                 the_lnet.ln_cpt_bits++;
1364
1365         rc = lnet_create_locks();
1366         if (rc != 0) {
1367                 CERROR("Can't create LNet global locks: %d\n", rc);
1368                 return -1;
1369         }
1370
1371         the_lnet.ln_refcount = 0;
1372         the_lnet.ln_init = 1;
1373         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1374         CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
1375         CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1376         CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1377
1378 #ifdef __KERNEL__
1379         /* The hash table size is the number of bits it takes to express the set
1380          * ln_num_routes, minus 1 (better to under estimate than over so we
1381          * don't waste memory). */
1382         if (rnet_htable_size <= 0)
1383                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
1384         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
1385                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
1386         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
1387                                            order_base_2(rnet_htable_size) - 1);
1388
1389         /* All LNDs apart from the LOLND are in separate modules.  They
1390          * register themselves when their module loads, and unregister
1391          * themselves when their module is unloaded. */
1392 #else
1393         the_lnet.ln_remote_nets_hbits = 8;
1394
1395         /* Register LNDs
1396          * NB the order here determines default 'networks=' order */
1397 # ifdef HAVE_LIBPTHREAD
1398         LNET_REGISTER_ULND(the_tcplnd);
1399 # endif
1400 #endif
1401         lnet_register_lnd(&the_lolnd);
1402         return 0;
1403 }
1404 EXPORT_SYMBOL(LNetInit);
1405
1406 /**
1407  * Finalize LNet library.
1408  *
1409  * Only userspace program needs to call this function. It can be called
1410  * at most once.
1411  *
1412  * \pre LNetInit() called with success.
1413  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1414  */
1415 void
1416 LNetFini(void)
1417 {
1418         LASSERT(the_lnet.ln_init);
1419         LASSERT(the_lnet.ln_refcount == 0);
1420
1421         while (!cfs_list_empty(&the_lnet.ln_lnds))
1422                 lnet_unregister_lnd(cfs_list_entry(the_lnet.ln_lnds.next,
1423                                                    lnd_t, lnd_list));
1424         lnet_destroy_locks();
1425
1426         the_lnet.ln_init = 0;
1427 }
1428 EXPORT_SYMBOL(LNetFini);
1429
1430 /**
1431  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1432  *
1433  * Userspace program should call this after a successful call to LNetInit().
1434  * Users must call this function at least once before any other functions.
1435  * For each successful call there must be a corresponding call to
1436  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1437  * ignored.
1438  *
1439  * The PID used by LNet may be different from the one requested.
1440  * See LNetGetId().
1441  *
1442  * \param requested_pid PID requested by the caller.
1443  *
1444  * \return >= 0 on success, and < 0 error code on failures.
1445  */
1446 int
1447 LNetNIInit(lnet_pid_t requested_pid)
1448 {
1449         int         im_a_router = 0;
1450         int         rc;
1451
1452         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
1453
1454         LASSERT (the_lnet.ln_init);
1455         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1456
1457         if (the_lnet.ln_refcount > 0) {
1458                 rc = the_lnet.ln_refcount++;
1459                 goto out;
1460         }
1461
1462         lnet_get_tunables();
1463
1464         if (requested_pid == LNET_PID_ANY) {
1465                 /* Don't instantiate LNET just for me */
1466                 rc = -ENETDOWN;
1467                 goto failed0;
1468         }
1469
1470         rc = lnet_prepare(requested_pid);
1471         if (rc != 0)
1472                 goto failed0;
1473
1474         rc = lnet_startup_lndnis();
1475         if (rc != 0)
1476                 goto failed1;
1477
1478         rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1479         if (rc != 0)
1480                 goto failed2;
1481
1482         rc = lnet_check_routes();
1483         if (rc != 0)
1484                 goto failed2;
1485
1486         rc = lnet_rtrpools_alloc(im_a_router);
1487         if (rc != 0)
1488                 goto failed2;
1489
1490         rc = lnet_acceptor_start();
1491         if (rc != 0)
1492                 goto failed2;
1493
1494         the_lnet.ln_refcount = 1;
1495         /* Now I may use my own API functions... */
1496
1497         /* NB router checker needs the_lnet.ln_ping_info in
1498          * lnet_router_checker -> lnet_update_ni_status_locked */
1499         rc = lnet_ping_target_init();
1500         if (rc != 0)
1501                 goto failed3;
1502
1503         rc = lnet_router_checker_start();
1504         if (rc != 0)
1505                 goto failed4;
1506
1507         lnet_proc_init();
1508         goto out;
1509
1510  failed4:
1511         lnet_ping_target_fini();
1512  failed3:
1513         the_lnet.ln_refcount = 0;
1514         lnet_acceptor_stop();
1515  failed2:
1516         lnet_destroy_routes();
1517         lnet_shutdown_lndnis();
1518  failed1:
1519         lnet_unprepare();
1520  failed0:
1521         LASSERT (rc < 0);
1522  out:
1523         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1524         return rc;
1525 }
1526 EXPORT_SYMBOL(LNetNIInit);
1527
1528 /**
1529  * Stop LNet interfaces, routing, and forwarding.
1530  *
1531  * Users must call this function once for each successful call to LNetNIInit().
1532  * Once the LNetNIFini() operation has been started, the results of pending
1533  * API operations are undefined.
1534  *
1535  * \return always 0 for current implementation.
1536  */
1537 int
1538 LNetNIFini()
1539 {
1540         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
1541
1542         LASSERT (the_lnet.ln_init);
1543         LASSERT (the_lnet.ln_refcount > 0);
1544
1545         if (the_lnet.ln_refcount != 1) {
1546                 the_lnet.ln_refcount--;
1547         } else {
1548                 LASSERT (!the_lnet.ln_niinit_self);
1549
1550                 lnet_proc_fini();
1551                 lnet_router_checker_stop();
1552                 lnet_ping_target_fini();
1553
1554                 /* Teardown fns that use my own API functions BEFORE here */
1555                 the_lnet.ln_refcount = 0;
1556
1557                 lnet_acceptor_stop();
1558                 lnet_destroy_routes();
1559                 lnet_shutdown_lndnis();
1560                 lnet_unprepare();
1561         }
1562
1563         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1564         return 0;
1565 }
1566 EXPORT_SYMBOL(LNetNIFini);
1567
1568 /**
1569  * This is an ugly hack to export IOC_LIBCFS_DEBUG_PEER and
1570  * IOC_LIBCFS_PORTALS_COMPATIBILITY commands to users, by tweaking the LNet
1571  * internal ioctl handler.
1572  *
1573  * IOC_LIBCFS_PORTALS_COMPATIBILITY is now deprecated, don't use it.
1574  *
1575  * \param cmd IOC_LIBCFS_DEBUG_PEER to print debugging data about a peer.
1576  * The data will be printed to system console. Don't use it excessively.
1577  * \param arg A pointer to lnet_process_id_t, process ID of the peer.
1578  *
1579  * \return Always return 0 when called by users directly (i.e., not via ioctl).
1580  */
1581 int
1582 LNetCtl(unsigned int cmd, void *arg)
1583 {
1584         struct libcfs_ioctl_data *data = arg;
1585         lnet_process_id_t         id = {0};
1586         lnet_ni_t                *ni;
1587         int                       rc;
1588
1589         LASSERT (the_lnet.ln_init);
1590         LASSERT (the_lnet.ln_refcount > 0);
1591
1592         switch (cmd) {
1593         case IOC_LIBCFS_GET_NI:
1594                 rc = LNetGetId(data->ioc_count, &id);
1595                 data->ioc_nid = id.nid;
1596                 return rc;
1597
1598         case IOC_LIBCFS_FAIL_NID:
1599                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1600
1601         case IOC_LIBCFS_ADD_ROUTE:
1602                 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1603                                     data->ioc_nid);
1604                 return (rc != 0) ? rc : lnet_check_routes();
1605
1606         case IOC_LIBCFS_DEL_ROUTE:
1607                 return lnet_del_route(data->ioc_net, data->ioc_nid);
1608
1609         case IOC_LIBCFS_GET_ROUTE:
1610                 return lnet_get_route(data->ioc_count,
1611                                       &data->ioc_net, &data->ioc_count,
1612                                       &data->ioc_nid, &data->ioc_flags);
1613         case IOC_LIBCFS_NOTIFY_ROUTER:
1614                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1615                                    cfs_time_current() -
1616                                    cfs_time_seconds(cfs_time_current_sec() -
1617                                                     (time_t)data->ioc_u64[0]));
1618
1619         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1620                 /* This can be removed once lustre stops calling it */
1621                 return 0;
1622
1623         case IOC_LIBCFS_LNET_DIST:
1624                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1625                 if (rc < 0 && rc != -EHOSTUNREACH)
1626                         return rc;
1627
1628                 data->ioc_u32[0] = rc;
1629                 return 0;
1630
1631         case IOC_LIBCFS_TESTPROTOCOMPAT:
1632                 lnet_net_lock(LNET_LOCK_EX);
1633                 the_lnet.ln_testprotocompat = data->ioc_flags;
1634                 lnet_net_unlock(LNET_LOCK_EX);
1635                 return 0;
1636
1637         case IOC_LIBCFS_PING:
1638                 id.nid = data->ioc_nid;
1639                 id.pid = data->ioc_u32[0];
1640                 rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
1641                                (lnet_process_id_t *)data->ioc_pbuf1,
1642                                data->ioc_plen1/sizeof(lnet_process_id_t));
1643                 if (rc < 0)
1644                         return rc;
1645                 data->ioc_count = rc;
1646                 return 0;
1647
1648         case IOC_LIBCFS_DEBUG_PEER: {
1649                 /* CAVEAT EMPTOR: this one designed for calling directly; not
1650                  * via an ioctl */
1651                 id = *((lnet_process_id_t *) arg);
1652
1653                 lnet_debug_peer(id.nid);
1654
1655                 ni = lnet_net2ni(LNET_NIDNET(id.nid));
1656                 if (ni == NULL) {
1657                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(id));
1658                 } else {
1659                         if (ni->ni_lnd->lnd_ctl == NULL) {
1660                                 CDEBUG(D_WARNING, "No ctl for %s\n",
1661                                        libcfs_id2str(id));
1662                         } else {
1663                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1664                         }
1665
1666                         lnet_ni_decref(ni);
1667                 }
1668                 return 0;
1669         }
1670
1671         default:
1672                 ni = lnet_net2ni(data->ioc_net);
1673                 if (ni == NULL)
1674                         return -EINVAL;
1675
1676                 if (ni->ni_lnd->lnd_ctl == NULL)
1677                         rc = -EINVAL;
1678                 else
1679                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1680
1681                 lnet_ni_decref(ni);
1682                 return rc;
1683         }
1684         /* not reached */
1685 }
1686 EXPORT_SYMBOL(LNetCtl);
1687
1688 /**
1689  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
1690  * all interfaces share a same PID, as requested by LNetNIInit().
1691  *
1692  * \param index Index of the interface to look up.
1693  * \param id On successful return, this location will hold the
1694  * lnet_process_id_t ID of the interface.
1695  *
1696  * \retval 0 If an interface exists at \a index.
1697  * \retval -ENOENT If no interface has been found.
1698  */
1699 int
1700 LNetGetId(unsigned int index, lnet_process_id_t *id)
1701 {
1702         struct lnet_ni  *ni;
1703         cfs_list_t      *tmp;
1704         int             cpt;
1705         int             rc = -ENOENT;
1706
1707         LASSERT(the_lnet.ln_init);
1708         LASSERT(the_lnet.ln_refcount > 0);
1709
1710         cpt = lnet_net_lock_current();
1711
1712         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
1713                 if (index-- != 0)
1714                         continue;
1715
1716                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
1717
1718                 id->nid = ni->ni_nid;
1719                 id->pid = the_lnet.ln_pid;
1720                 rc = 0;
1721                 break;
1722         }
1723
1724         lnet_net_unlock(cpt);
1725         return rc;
1726 }
1727 EXPORT_SYMBOL(LNetGetId);
1728
1729 /**
1730  * Print a string representation of handle \a h into buffer \a str of
1731  * \a len bytes.
1732  */
1733 void
1734 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1735 {
1736         snprintf(str, len, LPX64, h.cookie);
1737 }
1738 EXPORT_SYMBOL(LNetSnprintHandle);
1739
1740 static int
1741 lnet_create_ping_info(void)
1742 {
1743         int               i;
1744         int               n;
1745         int               rc;
1746         unsigned int      infosz;
1747         lnet_ni_t        *ni;
1748         lnet_process_id_t id;
1749         lnet_ping_info_t *pinfo;
1750
1751         for (n = 0; ; n++) {
1752                 rc = LNetGetId(n, &id);
1753                 if (rc == -ENOENT)
1754                         break;
1755
1756                 LASSERT (rc == 0);
1757         }
1758
1759         infosz = offsetof(lnet_ping_info_t, pi_ni[n]);
1760         LIBCFS_ALLOC(pinfo, infosz);
1761         if (pinfo == NULL) {
1762                 CERROR("Can't allocate ping info[%d]\n", n);
1763                 return -ENOMEM;
1764         }
1765
1766         pinfo->pi_nnis    = n;
1767         pinfo->pi_pid     = the_lnet.ln_pid;
1768         pinfo->pi_magic   = LNET_PROTO_PING_MAGIC;
1769         pinfo->pi_features = LNET_PING_FEAT_NI_STATUS;
1770
1771         for (i = 0; i < n; i++) {
1772                 lnet_ni_status_t *ns = &pinfo->pi_ni[i];
1773
1774                 rc = LNetGetId(i, &id);
1775                 LASSERT (rc == 0);
1776
1777                 ns->ns_nid    = id.nid;
1778                 ns->ns_status = LNET_NI_STATUS_UP;
1779
1780                 lnet_net_lock(0);
1781
1782                 ni = lnet_nid2ni_locked(id.nid, 0);
1783                 LASSERT(ni != NULL);
1784
1785                 lnet_ni_lock(ni);
1786                 LASSERT(ni->ni_status == NULL);
1787                 ni->ni_status = ns;
1788                 lnet_ni_unlock(ni);
1789
1790                 lnet_ni_decref_locked(ni, 0);
1791                 lnet_net_unlock(0);
1792         }
1793
1794         the_lnet.ln_ping_info = pinfo;
1795         return 0;
1796 }
1797
1798 static void
1799 lnet_destroy_ping_info(void)
1800 {
1801         struct lnet_ni  *ni;
1802
1803         lnet_net_lock(0);
1804
1805         cfs_list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
1806                 lnet_ni_lock(ni);
1807                 ni->ni_status = NULL;
1808                 lnet_ni_unlock(ni);
1809         }
1810
1811         lnet_net_unlock(0);
1812
1813         LIBCFS_FREE(the_lnet.ln_ping_info,
1814                     offsetof(lnet_ping_info_t,
1815                              pi_ni[the_lnet.ln_ping_info->pi_nnis]));
1816         the_lnet.ln_ping_info = NULL;
1817         return;
1818 }
1819
1820 int
1821 lnet_ping_target_init(void)
1822 {
1823         lnet_md_t         md = {0};
1824         lnet_handle_me_t  meh;
1825         lnet_process_id_t id;
1826         int               rc;
1827         int               rc2;
1828         int               infosz;
1829
1830         rc = lnet_create_ping_info();
1831         if (rc != 0)
1832                 return rc;
1833
1834         /* We can have a tiny EQ since we only need to see the unlink event on
1835          * teardown, which by definition is the last one! */
1836         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1837         if (rc != 0) {
1838                 CERROR("Can't allocate ping EQ: %d\n", rc);
1839                 goto failed_0;
1840         }
1841
1842         memset(&id, 0, sizeof(lnet_process_id_t));
1843         id.nid = LNET_NID_ANY;
1844         id.pid = LNET_PID_ANY;
1845
1846         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1847                           LNET_PROTO_PING_MATCHBITS, 0,
1848                           LNET_UNLINK, LNET_INS_AFTER,
1849                           &meh);
1850         if (rc != 0) {
1851                 CERROR("Can't create ping ME: %d\n", rc);
1852                 goto failed_1;
1853         }
1854
1855         /* initialize md content */
1856         infosz = offsetof(lnet_ping_info_t,
1857                           pi_ni[the_lnet.ln_ping_info->pi_nnis]);
1858         md.start     = the_lnet.ln_ping_info;
1859         md.length    = infosz;
1860         md.threshold = LNET_MD_THRESH_INF;
1861         md.max_size  = 0;
1862         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1863                        LNET_MD_MANAGE_REMOTE;
1864         md.user_ptr  = NULL;
1865         md.eq_handle = the_lnet.ln_ping_target_eq;
1866
1867         rc = LNetMDAttach(meh, md,
1868                           LNET_RETAIN,
1869                           &the_lnet.ln_ping_target_md);
1870         if (rc != 0) {
1871                 CERROR("Can't attach ping MD: %d\n", rc);
1872                 goto failed_2;
1873         }
1874
1875         return 0;
1876
1877  failed_2:
1878         rc2 = LNetMEUnlink(meh);
1879         LASSERT (rc2 == 0);
1880  failed_1:
1881         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1882         LASSERT (rc2 == 0);
1883  failed_0:
1884         lnet_destroy_ping_info();
1885         return rc;
1886 }
1887
1888 void
1889 lnet_ping_target_fini(void)
1890 {
1891         lnet_event_t    event;
1892         int             rc;
1893         int             which;
1894         int             timeout_ms = 1000;
1895         cfs_sigset_t    blocked = cfs_block_allsigs();
1896
1897         LNetMDUnlink(the_lnet.ln_ping_target_md);
1898         /* NB md could be busy; this just starts the unlink */
1899
1900         for (;;) {
1901                 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1902                                 timeout_ms, &event, &which);
1903
1904                 /* I expect overflow... */
1905                 LASSERT (rc >= 0 || rc == -EOVERFLOW);
1906
1907                 if (rc == 0) {
1908                         /* timed out: provide a diagnostic */
1909                         CWARN("Still waiting for ping MD to unlink\n");
1910                         timeout_ms *= 2;
1911                         continue;
1912                 }
1913
1914                 /* Got a valid event */
1915                 if (event.unlinked)
1916                         break;
1917         }
1918
1919         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1920         LASSERT (rc == 0);
1921         lnet_destroy_ping_info();
1922         cfs_restore_sigs(blocked);
1923 }
1924
1925 int
1926 lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1927 {
1928         lnet_handle_eq_t     eqh;
1929         lnet_handle_md_t     mdh;
1930         lnet_event_t         event;
1931         lnet_md_t            md = {0};
1932         int                  which;
1933         int                  unlinked = 0;
1934         int                  replied = 0;
1935         const int            a_long_time = 60000; /* mS */
1936         int                  infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
1937         lnet_ping_info_t    *info;
1938         lnet_process_id_t    tmpid;
1939         int                  i;
1940         int                  nob;
1941         int                  rc;
1942         int                  rc2;
1943         cfs_sigset_t         blocked;
1944
1945         if (n_ids <= 0 ||
1946             id.nid == LNET_NID_ANY ||
1947             timeout_ms > 500000 ||              /* arbitrary limit! */
1948             n_ids > 20)                         /* arbitrary limit! */
1949                 return -EINVAL;
1950
1951         if (id.pid == LNET_PID_ANY)
1952                 id.pid = LUSTRE_SRV_LNET_PID;
1953
1954         LIBCFS_ALLOC(info, infosz);
1955         if (info == NULL)
1956                 return -ENOMEM;
1957
1958         /* NB 2 events max (including any unlink event) */
1959         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1960         if (rc != 0) {
1961                 CERROR("Can't allocate EQ: %d\n", rc);
1962                 goto out_0;
1963         }
1964
1965         /* initialize md content */
1966         md.start     = info;
1967         md.length    = infosz;
1968         md.threshold = 2; /*GET/REPLY*/
1969         md.max_size  = 0;
1970         md.options   = LNET_MD_TRUNCATE;
1971         md.user_ptr  = NULL;
1972         md.eq_handle = eqh;
1973
1974         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
1975         if (rc != 0) {
1976                 CERROR("Can't bind MD: %d\n", rc);
1977                 goto out_1;
1978         }
1979
1980         rc = LNetGet(LNET_NID_ANY, mdh, id,
1981                      LNET_RESERVED_PORTAL,
1982                      LNET_PROTO_PING_MATCHBITS, 0);
1983
1984         if (rc != 0) {
1985                 /* Don't CERROR; this could be deliberate! */
1986
1987                 rc2 = LNetMDUnlink(mdh);
1988                 LASSERT (rc2 == 0);
1989
1990                 /* NB must wait for the UNLINK event below... */
1991                 unlinked = 1;
1992                 timeout_ms = a_long_time;
1993         }
1994
1995         do {
1996                 /* MUST block for unlink to complete */
1997                 if (unlinked)
1998                         blocked = cfs_block_allsigs();
1999
2000                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
2001
2002                 if (unlinked)
2003                         cfs_restore_sigs(blocked);
2004
2005                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
2006                        (rc2 <= 0) ? -1 : event.type,
2007                        (rc2 <= 0) ? -1 : event.status,
2008                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
2009
2010                 LASSERT (rc2 != -EOVERFLOW);     /* can't miss anything */
2011
2012                 if (rc2 <= 0 || event.status != 0) {
2013                         /* timeout or error */
2014                         if (!replied && rc == 0)
2015                                 rc = (rc2 < 0) ? rc2 :
2016                                      (rc2 == 0) ? -ETIMEDOUT :
2017                                      event.status;
2018
2019                         if (!unlinked) {
2020                                 /* Ensure completion in finite time... */
2021                                 LNetMDUnlink(mdh);
2022                                 /* No assertion (racing with network) */
2023                                 unlinked = 1;
2024                                 timeout_ms = a_long_time;
2025                         } else if (rc2 == 0) {
2026                                 /* timed out waiting for unlink */
2027                                 CWARN("ping %s: late network completion\n",
2028                                       libcfs_id2str(id));
2029                         }
2030                 } else if (event.type == LNET_EVENT_REPLY) {
2031                         replied = 1;
2032                         rc = event.mlength;
2033                 }
2034
2035         } while (rc2 <= 0 || !event.unlinked);
2036
2037         if (!replied) {
2038                 if (rc >= 0)
2039                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
2040                               libcfs_id2str(id));
2041                 rc = -EIO;
2042                 goto out_1;
2043         }
2044
2045         nob = rc;
2046         LASSERT (nob >= 0 && nob <= infosz);
2047
2048         rc = -EPROTO;                           /* if I can't parse... */
2049
2050         if (nob < 8) {
2051                 /* can't check magic/version */
2052                 CERROR("%s: ping info too short %d\n",
2053                        libcfs_id2str(id), nob);
2054                 goto out_1;
2055         }
2056
2057         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
2058                 lnet_swap_pinginfo(info);
2059         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
2060                 CERROR("%s: Unexpected magic %08x\n",
2061                        libcfs_id2str(id), info->pi_magic);
2062                 goto out_1;
2063         }
2064
2065         if ((info->pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
2066                 CERROR("%s: ping w/o NI status: 0x%x\n",
2067                        libcfs_id2str(id), info->pi_features);
2068                 goto out_1;
2069         }
2070
2071         if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
2072                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
2073                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
2074                 goto out_1;
2075         }
2076
2077         if (info->pi_nnis < n_ids)
2078                 n_ids = info->pi_nnis;
2079
2080         if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
2081                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
2082                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
2083                 goto out_1;
2084         }
2085
2086         rc = -EFAULT;                           /* If I SEGV... */
2087
2088         for (i = 0; i < n_ids; i++) {
2089                 tmpid.pid = info->pi_pid;
2090                 tmpid.nid = info->pi_ni[i].ns_nid;
2091                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
2092                         goto out_1;
2093         }
2094         rc = info->pi_nnis;
2095
2096  out_1:
2097         rc2 = LNetEQFree(eqh);
2098         if (rc2 != 0)
2099                 CERROR("rc2 %d\n", rc2);
2100         LASSERT (rc2 == 0);
2101
2102  out_0:
2103         LIBCFS_FREE(info, infosz);
2104         return rc;
2105 }