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