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, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38 #include <lnet/lib-lnet.h>
  39
  40 #ifdef __KERNEL__
  41 #define D_LNI D_CONSOLE
  42 #else
  43 #define D_LNI D_CONFIG
  44 #endif
  45
  46 lnet_t      the_lnet;                           /* THE state of the network */
  47
  48 #ifdef __KERNEL__
  49
  50 static char *ip2nets = "";
  51 CFS_MODULE_PARM(ip2nets, "s", charp, 0444,
  52                 "LNET network <- IP table");
  53
  54 static char *networks = "";
  55 CFS_MODULE_PARM(networks, "s", charp, 0444,
  56                 "local networks");
  57
  58 static char *routes = "";
  59 CFS_MODULE_PARM(routes, "s", charp, 0444,
  60                 "routes to non-local networks");
  61
  62 char *
  63 lnet_get_routes(void)
  64 {
  65         return routes;
  66 }
  67
  68 char *
  69 lnet_get_networks(void)
  70 {
  71         char   *nets;
  72         int     rc;
  73
  74         if (*networks != 0 && *ip2nets != 0) {
  75                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  76                                    "'ip2nets' but not both at once\n");
  77                 return NULL;
  78         }
  79
  80         if (*ip2nets != 0) {
  81                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  82                 return (rc == 0) ? nets : NULL;
  83         }
  84
  85         if (*networks != 0)
  86                 return networks;
  87
  88         return "tcp";
  89 }
  90
  91 void
  92 lnet_init_locks(void)
  93 {
  94         cfs_spin_lock_init(&the_lnet.ln_eq_wait_lock);
  95         cfs_waitq_init(&the_lnet.ln_eq_waitq);
  96         cfs_mutex_init(&the_lnet.ln_lnd_mutex);
  97         cfs_mutex_init(&the_lnet.ln_api_mutex);
  98 }
  99
 100 void
 101 lnet_fini_locks(void)
 102 {
 103 }
 104
 105 #else
 106
 107 char *
 108 lnet_get_routes(void)
 109 {
 110         char *str = getenv("LNET_ROUTES");
 111
 112         return (str == NULL) ? "" : str;
 113 }
 114
 115 char *
 116 lnet_get_networks (void)
 117 {
 118         static char       default_networks[256];
 119         char             *networks = getenv ("LNET_NETWORKS");
 120         char             *ip2nets  = getenv ("LNET_IP2NETS");
 121         char             *str;
 122         char             *sep;
 123         int               len;
 124         int               nob;
 125         int               rc;
 126         cfs_list_t       *tmp;
 127
 128 #ifdef NOT_YET
 129         if (networks != NULL && ip2nets != NULL) {
 130                 LCONSOLE_ERROR_MSG(0x103, "Please set EITHER 'LNET_NETWORKS' or"
 131                                    " 'LNET_IP2NETS' but not both at once\n");
 132                 return NULL;
 133         }
 134
 135         if (ip2nets != NULL) {
 136                 rc = lnet_parse_ip2nets(&networks, ip2nets);
 137                 return (rc == 0) ? networks : NULL;
 138         }
 139 #else
 140         SET_BUT_UNUSED(ip2nets);
 141         SET_BUT_UNUSED(rc);
 142 #endif
 143         if (networks != NULL)
 144                 return networks;
 145
 146         /* In userland, the default 'networks=' is the list of known net types */
 147
 148         len = sizeof(default_networks);
 149         str = default_networks;
 150         *str = 0;
 151         sep = "";
 152
 153         cfs_list_for_each (tmp, &the_lnet.ln_lnds) {
 154                 lnd_t *lnd = cfs_list_entry(tmp, lnd_t, lnd_list);
 155
 156                 nob = snprintf(str, len, "%s%s", sep,
 157                                libcfs_lnd2str(lnd->lnd_type));
 158                 len -= nob;
 159                 if (len < 0) {
 160                         /* overflowed the string; leave it where it was */
 161                         *str = 0;
 162                         break;
 163                 }
 164
 165                 str += nob;
 166                 sep = ",";
 167         }
 168
 169         return default_networks;
 170 }
 171
 172 # ifndef HAVE_LIBPTHREAD
 173
 174 void lnet_init_locks(void)
 175 {
 176         the_lnet.ln_eq_wait_lock = 0;
 177         the_lnet.ln_lnd_mutex = 0;
 178         the_lnet.ln_api_mutex = 0;
 179 }
 180
 181 void lnet_fini_locks(void)
 182 {
 183         LASSERT(the_lnet.ln_api_mutex == 0);
 184         LASSERT(the_lnet.ln_lnd_mutex == 0);
 185         LASSERT(the_lnet.ln_eq_wait_lock == 0);
 186 }
 187
 188 # else
 189
 190 void lnet_init_locks(void)
 191 {
 192         pthread_cond_init(&the_lnet.ln_eq_cond, NULL);
 193         pthread_mutex_init(&the_lnet.ln_eq_wait_lock, NULL);
 194         pthread_mutex_init(&the_lnet.ln_lnd_mutex, NULL);
 195         pthread_mutex_init(&the_lnet.ln_api_mutex, NULL);
 196 }
 197
 198 void lnet_fini_locks(void)
 199 {
 200         pthread_mutex_destroy(&the_lnet.ln_api_mutex);
 201         pthread_mutex_destroy(&the_lnet.ln_lnd_mutex);
 202         pthread_mutex_destroy(&the_lnet.ln_eq_wait_lock);
 203         pthread_cond_destroy(&the_lnet.ln_eq_cond);
 204 }
 205
 206 # endif
 207 #endif
 208
 209 static int
 210 lnet_create_locks(void)
 211 {
 212         lnet_init_locks();
 213
 214         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 215         if (the_lnet.ln_res_lock == NULL)
 216                 goto failed;
 217
 218         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 219         if (the_lnet.ln_net_lock == NULL)
 220                 goto failed;
 221
 222         return 0;
 223
 224  failed:
 225         lnet_fini_locks();
 226         return -ENOMEM;
 227 }
 228
 229 static void
 230 lnet_destroy_locks(void)
 231 {
 232         if (the_lnet.ln_res_lock != NULL) {
 233                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 234                 the_lnet.ln_res_lock = NULL;
 235         }
 236
 237         if (the_lnet.ln_net_lock != NULL) {
 238                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 239                 the_lnet.ln_net_lock = NULL;
 240         }
 241
 242         lnet_fini_locks();
 243 }
 244
 245 void lnet_assert_wire_constants (void)
 246 {
 247         /* Wire protocol assertions generated by 'wirecheck'
 248          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 249          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 250          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 251
 252         /* Constants... */
 253         CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 254         CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
 255         CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
 256         CLASSERT (LNET_MSG_ACK == 0);
 257         CLASSERT (LNET_MSG_PUT == 1);
 258         CLASSERT (LNET_MSG_GET == 2);
 259         CLASSERT (LNET_MSG_REPLY == 3);
 260         CLASSERT (LNET_MSG_HELLO == 4);
 261
 262         /* Checks for struct ptl_handle_wire_t */
 263         CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
 264         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 265         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 266         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 267         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 268
 269         /* Checks for struct lnet_magicversion_t */
 270         CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
 271         CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
 272         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 273         CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
 274         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 275         CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 276         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 277
 278         /* Checks for struct lnet_hdr_t */
 279         CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
 280         CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 281         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 282         CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
 283         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 284         CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 285         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 286         CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
 287         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 288         CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
 289         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 290         CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
 291         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 292         CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
 293         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 294
 295         /* Ack */
 296         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 297         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 298         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 299         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 300         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 301         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 302
 303         /* Put */
 304         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 305         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 306         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 307         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 308         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 309         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 310         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 311         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 312         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 313         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 314
 315         /* Get */
 316         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 317         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 318         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 319         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 320         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 321         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 322         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 323         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 324         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 325         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 326
 327         /* Reply */
 328         CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 329         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 330
 331         /* Hello */
 332         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 333         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 334         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 335         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 336 }
 337
 338 lnd_t *
 339 lnet_find_lnd_by_type (int type)
 340 {
 341         lnd_t              *lnd;
 342         cfs_list_t         *tmp;
 343
 344         /* holding lnd mutex */
 345         cfs_list_for_each (tmp, &the_lnet.ln_lnds) {
 346                 lnd = cfs_list_entry(tmp, lnd_t, lnd_list);
 347
 348                 if ((int)lnd->lnd_type == type)
 349                         return lnd;
 350         }
 351
 352         return NULL;
 353 }
 354
 355 void
 356 lnet_register_lnd (lnd_t *lnd)
 357 {
 358         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
 359
 360         LASSERT (the_lnet.ln_init);
 361         LASSERT (libcfs_isknown_lnd(lnd->lnd_type));
 362         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 363
 364         cfs_list_add_tail (&lnd->lnd_list, &the_lnet.ln_lnds);
 365         lnd->lnd_refcount = 0;
 366
 367         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 368
 369         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
 370 }
 371
 372 void
 373 lnet_unregister_lnd (lnd_t *lnd)
 374 {
 375         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
 376
 377         LASSERT (the_lnet.ln_init);
 378         LASSERT (lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 379         LASSERT (lnd->lnd_refcount == 0);
 380
 381         cfs_list_del (&lnd->lnd_list);
 382         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 383
 384         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
 385 }
 386
 387 void
 388 lnet_counters_get(lnet_counters_t *counters)
 389 {
 390         lnet_counters_t *ctr;
 391         int             i;
 392
 393         memset(counters, 0, sizeof(*counters));
 394
 395         lnet_net_lock(LNET_LOCK_EX);
 396
 397         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 398                 counters->msgs_max     += ctr->msgs_max;
 399                 counters->msgs_alloc   += ctr->msgs_alloc;
 400                 counters->errors       += ctr->errors;
 401                 counters->send_count   += ctr->send_count;
 402                 counters->recv_count   += ctr->recv_count;
 403                 counters->route_count  += ctr->route_count;
 404                 counters->drop_length  += ctr->drop_length;
 405                 counters->send_length  += ctr->send_length;
 406                 counters->recv_length  += ctr->recv_length;
 407                 counters->route_length += ctr->route_length;
 408                 counters->drop_length  += ctr->drop_length;
 409
 410         }
 411         lnet_net_unlock(LNET_LOCK_EX);
 412 }
 413 EXPORT_SYMBOL(lnet_counters_get);
 414
 415 void
 416 lnet_counters_reset(void)
 417 {
 418         lnet_counters_t *counters;
 419         int             i;
 420
 421         lnet_net_lock(LNET_LOCK_EX);
 422
 423         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 424                 memset(counters, 0, sizeof(lnet_counters_t));
 425
 426         lnet_net_unlock(LNET_LOCK_EX);
 427 }
 428 EXPORT_SYMBOL(lnet_counters_reset);
 429
 430 #ifdef LNET_USE_LIB_FREELIST
 431
 432 int
 433 lnet_freelist_init (lnet_freelist_t *fl, int n, int size)
 434 {
 435         char *space;
 436
 437         LASSERT (n > 0);
 438
 439         size += offsetof (lnet_freeobj_t, fo_contents);
 440
 441         LIBCFS_ALLOC(space, n * size);
 442         if (space == NULL)
 443                 return (-ENOMEM);
 444
 445         CFS_INIT_LIST_HEAD (&fl->fl_list);
 446         fl->fl_objs = space;
 447         fl->fl_nobjs = n;
 448         fl->fl_objsize = size;
 449
 450         do
 451         {
 452                 memset (space, 0, size);
 453                 cfs_list_add ((cfs_list_t *)space, &fl->fl_list);
 454                 space += size;
 455         } while (--n != 0);
 456
 457         return (0);
 458 }
 459
 460 void
 461 lnet_freelist_fini (lnet_freelist_t *fl)
 462 {
 463         cfs_list_t       *el;
 464         int               count;
 465
 466         if (fl->fl_nobjs == 0)
 467                 return;
 468
 469         count = 0;
 470         for (el = fl->fl_list.next; el != &fl->fl_list; el = el->next)
 471                 count++;
 472
 473         LASSERT (count == fl->fl_nobjs);
 474
 475         LIBCFS_FREE(fl->fl_objs, fl->fl_nobjs * fl->fl_objsize);
 476         memset (fl, 0, sizeof (*fl));
 477 }
 478
 479 #endif /* LNET_USE_LIB_FREELIST */
 480
 481 __u64
 482 lnet_create_interface_cookie (void)
 483 {
 484         /* NB the interface cookie in wire handles guards against delayed
 485          * replies and ACKs appearing valid after reboot. Initialisation time,
 486          * even if it's only implemented to millisecond resolution is probably
 487          * easily good enough. */
 488         struct timeval tv;
 489         __u64          cookie;
 490 #ifndef __KERNEL__
 491         int            rc = gettimeofday (&tv, NULL);
 492         LASSERT (rc == 0);
 493 #else
 494         cfs_gettimeofday(&tv);
 495 #endif
 496         cookie = tv.tv_sec;
 497         cookie *= 1000000;
 498         cookie += tv.tv_usec;
 499         return cookie;
 500 }
 501
 502 static char *
 503 lnet_res_type2str(int type)
 504 {
 505         switch (type) {
 506         default:
 507                 LBUG();
 508         case LNET_COOKIE_TYPE_MD:
 509                 return "MD";
 510         case LNET_COOKIE_TYPE_ME:
 511                 return "ME";
 512         case LNET_COOKIE_TYPE_EQ:
 513                 return "EQ";
 514         }
 515 }
 516
 517 void
 518 lnet_res_container_cleanup(struct lnet_res_container *rec)
 519 {
 520         int     count = 0;
 521
 522         if (rec->rec_type == 0) /* not set yet, it's a uninitialized */
 523                 return;
 524
 525         while (!cfs_list_empty(&rec->rec_active)) {
 526                 cfs_list_t *e = rec->rec_active.next;
 527
 528                 cfs_list_del_init(e);
 529                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 530                         lnet_eq_free(cfs_list_entry(e, lnet_eq_t, eq_list));
 531
 532                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 533                         lnet_md_free(cfs_list_entry(e, lnet_libmd_t, md_list));
 534
 535                 } else { /* NB: Active MEs should be attached on portals */
 536                         LBUG();
 537                 }
 538                 count++;
 539         }
 540
 541         if (count > 0) {
 542                 /* Found alive MD/ME/EQ, user really should unlink/free
 543                  * all of them before finalize LNet, but if someone didn't,
 544                  * we have to recycle garbage for him */
 545                 CERROR("%d active elements on exit of %s container\n",
 546                        count, lnet_res_type2str(rec->rec_type));
 547         }
 548
 549 #ifdef LNET_USE_LIB_FREELIST
 550         lnet_freelist_fini(&rec->rec_freelist);
 551 #endif
 552         if (rec->rec_lh_hash != NULL) {
 553                 LIBCFS_FREE(rec->rec_lh_hash,
 554                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 555                 rec->rec_lh_hash = NULL;
 556         }
 557
 558         rec->rec_type = 0; /* mark it as finalized */
 559 }
 560
 561 int
 562 lnet_res_container_setup(struct lnet_res_container *rec,
 563                          int cpt, int type, int objnum, int objsz)
 564 {
 565         int     rc = 0;
 566         int     i;
 567
 568         LASSERT(rec->rec_type == 0);
 569
 570         rec->rec_type = type;
 571         CFS_INIT_LIST_HEAD(&rec->rec_active);
 572
 573 #ifdef LNET_USE_LIB_FREELIST
 574         memset(&rec->rec_freelist, 0, sizeof(rec->rec_freelist));
 575         rc = lnet_freelist_init(&rec->rec_freelist, objnum, objsz);
 576         if (rc != 0)
 577                 goto out;
 578 #endif
 579         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 580
 581         /* Arbitrary choice of hash table size */
 582         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 583                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 584         if (rec->rec_lh_hash == NULL) {
 585                 rc = -ENOMEM;
 586                 goto out;
 587         }
 588
 589         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 590                 CFS_INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 591
 592         return 0;
 593
 594 out:
 595         CERROR("Failed to setup %s resource container\n",
 596                lnet_res_type2str(type));
 597         lnet_res_container_cleanup(rec);
 598         return rc;
 599 }
 600
 601 static void
 602 lnet_res_containers_destroy(struct lnet_res_container **recs)
 603 {
 604         struct lnet_res_container       *rec;
 605         int                             i;
 606
 607         cfs_percpt_for_each(rec, i, recs)
 608                 lnet_res_container_cleanup(rec);
 609
 610         cfs_percpt_free(recs);
 611 }
 612
 613 static struct lnet_res_container **
 614 lnet_res_containers_create(int type, int objnum, int objsz)
 615 {
 616         struct lnet_res_container       **recs;
 617         struct lnet_res_container       *rec;
 618         int                             rc;
 619         int                             i;
 620
 621         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 622         if (recs == NULL) {
 623                 CERROR("Failed to allocate %s resource containers\n",
 624                        lnet_res_type2str(type));
 625                 return NULL;
 626         }
 627
 628         cfs_percpt_for_each(rec, i, recs) {
 629                 rc = lnet_res_container_setup(rec, i, type, objnum, objsz);
 630                 if (rc != 0) {
 631                         lnet_res_containers_destroy(recs);
 632                         return NULL;
 633                 }
 634         }
 635
 636         return recs;
 637 }
 638
 639 lnet_libhandle_t *
 640 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 641 {
 642         /* ALWAYS called with lnet_res_lock held */
 643         cfs_list_t              *head;
 644         lnet_libhandle_t        *lh;
 645         unsigned int            hash;
 646
 647         if ((cookie & (LNET_COOKIE_TYPES - 1)) != rec->rec_type)
 648                 return NULL;
 649
 650         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 651         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 652
 653         cfs_list_for_each_entry(lh, head, lh_hash_chain) {
 654                 if (lh->lh_cookie == cookie)
 655                         return lh;
 656         }
 657
 658         return NULL;
 659 }
 660
 661 void
 662 lnet_res_lh_initialize(struct lnet_res_container *rec, lnet_libhandle_t *lh)
 663 {
 664         /* ALWAYS called with lnet_res_lock held */
 665         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 666         unsigned int    hash;
 667
 668         lh->lh_cookie = rec->rec_lh_cookie;
 669         rec->rec_lh_cookie += 1 << ibits;
 670
 671         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 672
 673         cfs_list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 674 }
 675
 676 #ifndef __KERNEL__
 677 /**
 678  * Reserved API - do not use.
 679  * Temporary workaround to allow uOSS and test programs force server
 680  * mode in userspace. See comments near ln_server_mode_flag in
 681  * lnet/lib-types.h */
 682
 683 void
 684 lnet_server_mode() {
 685         the_lnet.ln_server_mode_flag = 1;
 686 }
 687 #endif
 688
 689 int lnet_unprepare(void);
 690
 691 int
 692 lnet_prepare(lnet_pid_t requested_pid)
 693 {
 694         /* Prepare to bring up the network */
 695         struct lnet_res_container **recs;
 696         int                       rc = 0;
 697
 698         LASSERT (the_lnet.ln_refcount == 0);
 699
 700         the_lnet.ln_routing = 0;
 701
 702 #ifdef __KERNEL__
 703         LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 704         the_lnet.ln_pid = requested_pid;
 705 #else
 706         if (the_lnet.ln_server_mode_flag) {/* server case (uOSS) */
 707                 LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 708
 709                 if (cfs_curproc_uid())/* Only root can run user-space server */
 710                         return -EPERM;
 711                 the_lnet.ln_pid = requested_pid;
 712
 713         } else {/* client case (liblustre) */
 714
 715                 /* My PID must be unique on this node and flag I'm userspace */
 716                 the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
 717         }
 718 #endif
 719
 720         CFS_INIT_LIST_HEAD(&the_lnet.ln_test_peers);
 721         CFS_INIT_LIST_HEAD(&the_lnet.ln_nis);
 722         CFS_INIT_LIST_HEAD(&the_lnet.ln_nis_cpt);
 723         CFS_INIT_LIST_HEAD(&the_lnet.ln_nis_zombie);
 724         CFS_INIT_LIST_HEAD(&the_lnet.ln_remote_nets);
 725         CFS_INIT_LIST_HEAD(&the_lnet.ln_routers);
 726
 727         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 728
 729         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
 730                                                 sizeof(lnet_counters_t));
 731         if (the_lnet.ln_counters == NULL) {
 732                 CERROR("Failed to allocate counters for LNet\n");
 733                 rc = -ENOMEM;
 734                 goto failed;
 735         }
 736
 737         rc = lnet_peer_tables_create();
 738         if (rc != 0)
 739                 goto failed;
 740
 741         /* NB: we will have instance of message container per CPT soon */
 742         rc = lnet_msg_containers_create();
 743         if (rc != 0)
 744                 goto failed;
 745
 746         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 747                                       LNET_COOKIE_TYPE_EQ, LNET_FL_MAX_EQS,
 748                                       sizeof(lnet_eq_t));
 749         if (rc != 0)
 750                 goto failed;
 751
 752         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME, LNET_FL_MAX_MES,
 753                                           sizeof(lnet_me_t));
 754         if (recs == NULL)
 755                 goto failed;
 756
 757         the_lnet.ln_me_containers = recs;
 758
 759         /* NB: we will have instance of MD container per CPT soon */
 760         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD, LNET_FL_MAX_MDS,
 761                                           sizeof(lnet_libmd_t));
 762         if (recs == NULL)
 763                 goto failed;
 764
 765         the_lnet.ln_md_containers = recs;
 766
 767         rc = lnet_portals_create();
 768         if (rc != 0) {
 769                 CERROR("Failed to create portals for LNet: %d\n", rc);
 770                 goto failed;
 771         }
 772
 773         return 0;
 774
 775  failed:
 776         lnet_unprepare();
 777         return rc;
 778 }
 779
 780 int
 781 lnet_unprepare (void)
 782 {
 783         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 784          * have shut down already, so it is safe to unlink and free all
 785          * descriptors, even those that appear committed to a network op (eg MD
 786          * with non-zero pending count) */
 787
 788         lnet_fail_nid(LNET_NID_ANY, 0);
 789
 790         LASSERT(the_lnet.ln_refcount == 0);
 791         LASSERT(cfs_list_empty(&the_lnet.ln_test_peers));
 792         LASSERT(cfs_list_empty(&the_lnet.ln_nis));
 793         LASSERT(cfs_list_empty(&the_lnet.ln_nis_cpt));
 794         LASSERT(cfs_list_empty(&the_lnet.ln_nis_zombie));
 795
 796         lnet_portals_destroy();
 797
 798         if (the_lnet.ln_md_containers != NULL) {
 799                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 800                 the_lnet.ln_md_containers = NULL;
 801         }
 802
 803         if (the_lnet.ln_me_containers != NULL) {
 804                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 805                 the_lnet.ln_me_containers = NULL;
 806         }
 807
 808         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 809
 810         lnet_msg_containers_destroy();
 811         lnet_peer_tables_destroy();
 812         lnet_rtrpools_free();
 813
 814         if (the_lnet.ln_counters != NULL) {
 815                 cfs_percpt_free(the_lnet.ln_counters);
 816                 the_lnet.ln_counters = NULL;
 817         }
 818
 819         return 0;
 820 }
 821
 822 lnet_ni_t  *
 823 lnet_net2ni_locked(__u32 net, int cpt)
 824 {
 825         cfs_list_t      *tmp;
 826         lnet_ni_t       *ni;
 827
 828         LASSERT(cpt != LNET_LOCK_EX);
 829
 830         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 831                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 832
 833                 if (LNET_NIDNET(ni->ni_nid) == net) {
 834                         lnet_ni_addref_locked(ni, cpt);
 835                         return ni;
 836                 }
 837         }
 838
 839         return NULL;
 840 }
 841
 842 static unsigned int
 843 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 844 {
 845         __u64           key = nid;
 846         unsigned int    val;
 847
 848         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 849
 850         if (number == 1)
 851                 return 0;
 852
 853         val = cfs_hash_long(key, LNET_CPT_BITS);
 854         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 855         if (val < number)
 856                 return val;
 857
 858         return (unsigned int)((key + val + (val >> 1)) % number);
 859 }
 860
 861 int
 862 lnet_cpt_of_nid_locked(lnet_nid_t nid)
 863 {
 864         struct lnet_ni *ni;
 865
 866         /* must called with hold of lnet_net_lock */
 867         if (LNET_CPT_NUMBER == 1)
 868                 return 0; /* the only one */
 869
 870         /* take lnet_net_lock(any) would be OK */
 871         if (!cfs_list_empty(&the_lnet.ln_nis_cpt)) {
 872                 cfs_list_for_each_entry(ni, &the_lnet.ln_nis_cpt, ni_cptlist) {
 873                         if (LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid))
 874                                 continue;
 875
 876                         LASSERT(ni->ni_cpts != NULL);
 877                         return ni->ni_cpts[lnet_nid_cpt_hash
 878                                            (nid, ni->ni_ncpts)];
 879                 }
 880         }
 881
 882         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 883 }
 884
 885 int
 886 lnet_cpt_of_nid(lnet_nid_t nid)
 887 {
 888         int     cpt;
 889         int     cpt2;
 890
 891         if (LNET_CPT_NUMBER == 1)
 892                 return 0; /* the only one */
 893
 894         if (cfs_list_empty(&the_lnet.ln_nis_cpt))
 895                 return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 896
 897         cpt = lnet_net_lock_current();
 898         cpt2 = lnet_cpt_of_nid_locked(nid);
 899         lnet_net_unlock(cpt);
 900
 901         return cpt2;
 902 }
 903 EXPORT_SYMBOL(lnet_cpt_of_nid);
 904
 905 int
 906 lnet_islocalnet(__u32 net)
 907 {
 908         struct lnet_ni  *ni;
 909         int             cpt;
 910
 911         cpt = lnet_net_lock_current();
 912
 913         ni = lnet_net2ni_locked(net, cpt);
 914         if (ni != NULL)
 915                 lnet_ni_decref_locked(ni, cpt);
 916
 917         lnet_net_unlock(cpt);
 918
 919         return ni != NULL;
 920 }
 921
 922 lnet_ni_t  *
 923 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 924 {
 925         struct lnet_ni  *ni;
 926         cfs_list_t      *tmp;
 927
 928         LASSERT(cpt != LNET_LOCK_EX);
 929
 930         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 931                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 932
 933                 if (ni->ni_nid == nid) {
 934                         lnet_ni_addref_locked(ni, cpt);
 935                         return ni;
 936                 }
 937         }
 938
 939         return NULL;
 940 }
 941
 942 int
 943 lnet_islocalnid(lnet_nid_t nid)
 944 {
 945         struct lnet_ni  *ni;
 946         int             cpt;
 947
 948         cpt = lnet_net_lock_current();
 949         ni = lnet_nid2ni_locked(nid, cpt);
 950         if (ni != NULL)
 951                 lnet_ni_decref_locked(ni, cpt);
 952         lnet_net_unlock(cpt);
 953
 954         return ni != NULL;
 955 }
 956
 957 int
 958 lnet_count_acceptor_nis (void)
 959 {
 960         /* Return the # of NIs that need the acceptor. */
 961         int             count = 0;
 962 #if defined(__KERNEL__) || defined(HAVE_LIBPTHREAD)
 963         cfs_list_t      *tmp;
 964         struct lnet_ni  *ni;
 965         int             cpt;
 966
 967         cpt = lnet_net_lock_current();
 968         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 969                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 970
 971                 if (ni->ni_lnd->lnd_accept != NULL)
 972                         count++;
 973         }
 974
 975         lnet_net_unlock(cpt);
 976
 977 #endif /* defined(__KERNEL__) || defined(HAVE_LIBPTHREAD) */
 978         return count;
 979 }
 980
 981 static int
 982 lnet_ni_tq_credits(lnet_ni_t *ni)
 983 {
 984         int     credits;
 985
 986         LASSERT(ni->ni_ncpts >= 1);
 987
 988         if (ni->ni_ncpts == 1)
 989                 return ni->ni_maxtxcredits;
 990
 991         credits = ni->ni_maxtxcredits / ni->ni_ncpts;
 992         credits = max(credits, 8 * ni->ni_peertxcredits);
 993         credits = min(credits, ni->ni_maxtxcredits);
 994
 995         return credits;
 996 }
 997
 998 void
 999 lnet_shutdown_lndnis (void)
1000 {
1001         int                i;
1002         int                islo;
1003         lnet_ni_t         *ni;
1004
1005         /* NB called holding the global mutex */
1006
1007         /* All quiet on the API front */
1008         LASSERT(!the_lnet.ln_shutdown);
1009         LASSERT(the_lnet.ln_refcount == 0);
1010         LASSERT(cfs_list_empty(&the_lnet.ln_nis_zombie));
1011         LASSERT(cfs_list_empty(&the_lnet.ln_remote_nets));
1012
1013         lnet_net_lock(LNET_LOCK_EX);
1014         the_lnet.ln_shutdown = 1;       /* flag shutdown */
1015
1016         /* Unlink NIs from the global table */
1017         while (!cfs_list_empty(&the_lnet.ln_nis)) {
1018                 ni = cfs_list_entry(the_lnet.ln_nis.next,
1019                                     lnet_ni_t, ni_list);
1020                 /* move it to zombie list and nobody can find it anymore */
1021                 cfs_list_move(&ni->ni_list, &the_lnet.ln_nis_zombie);
1022                 lnet_ni_decref_locked(ni, 0);   /* drop ln_nis' ref */
1023
1024                 if (!cfs_list_empty(&ni->ni_cptlist)) {
1025                         cfs_list_del_init(&ni->ni_cptlist);
1026                         lnet_ni_decref_locked(ni, 0);
1027                 }
1028         }
1029
1030         /* Drop the cached eqwait NI. */
1031         if (the_lnet.ln_eq_waitni != NULL) {
1032                 lnet_ni_decref_locked(the_lnet.ln_eq_waitni, 0);
1033                 the_lnet.ln_eq_waitni = NULL;
1034         }
1035
1036         /* Drop the cached loopback NI. */
1037         if (the_lnet.ln_loni != NULL) {
1038                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1039                 the_lnet.ln_loni = NULL;
1040         }
1041
1042         lnet_net_unlock(LNET_LOCK_EX);
1043
1044         /* Clear lazy portals and drop delayed messages which hold refs
1045          * on their lnet_msg_t::msg_rxpeer */
1046         for (i = 0; i < the_lnet.ln_nportals; i++)
1047                 LNetClearLazyPortal(i);
1048
1049         /* Clear the peer table and wait for all peers to go (they hold refs on
1050          * their NIs) */
1051         lnet_peer_tables_cleanup();
1052
1053         lnet_net_lock(LNET_LOCK_EX);
1054         /* Now wait for the NI's I just nuked to show up on ln_zombie_nis
1055          * and shut them down in guaranteed thread context */
1056         i = 2;
1057         while (!cfs_list_empty(&the_lnet.ln_nis_zombie)) {
1058                 int     *ref;
1059                 int     j;
1060
1061                 ni = cfs_list_entry(the_lnet.ln_nis_zombie.next,
1062                                     lnet_ni_t, ni_list);
1063                 cfs_list_del_init(&ni->ni_list);
1064                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1065                         if (*ref == 0)
1066                                 continue;
1067                         /* still busy, add it back to zombie list */
1068                         cfs_list_add(&ni->ni_list, &the_lnet.ln_nis_zombie);
1069                         break;
1070                 }
1071
1072                 while (!cfs_list_empty(&ni->ni_list)) {
1073                         lnet_net_unlock(LNET_LOCK_EX);
1074                         ++i;
1075                         if ((i & (-i)) == i) {
1076                                 CDEBUG(D_WARNING,
1077                                        "Waiting for zombie LNI %s\n",
1078                                        libcfs_nid2str(ni->ni_nid));
1079                         }
1080                         cfs_pause(cfs_time_seconds(1));
1081                         lnet_net_lock(LNET_LOCK_EX);
1082                         continue;
1083                 }
1084
1085                 ni->ni_lnd->lnd_refcount--;
1086                 lnet_net_unlock(LNET_LOCK_EX);
1087
1088                 islo = ni->ni_lnd->lnd_type == LOLND;
1089
1090                 LASSERT (!cfs_in_interrupt ());
1091                 (ni->ni_lnd->lnd_shutdown)(ni);
1092
1093                 /* can't deref lnd anymore now; it might have unregistered
1094                  * itself...  */
1095
1096                 if (!islo)
1097                         CDEBUG(D_LNI, "Removed LNI %s\n",
1098                                libcfs_nid2str(ni->ni_nid));
1099
1100                 lnet_ni_free(ni);
1101                 lnet_net_lock(LNET_LOCK_EX);
1102         }
1103
1104         the_lnet.ln_shutdown = 0;
1105         lnet_net_unlock(LNET_LOCK_EX);
1106
1107         if (the_lnet.ln_network_tokens != NULL) {
1108                 LIBCFS_FREE(the_lnet.ln_network_tokens,
1109                             the_lnet.ln_network_tokens_nob);
1110                 the_lnet.ln_network_tokens = NULL;
1111         }
1112 }
1113
1114 int
1115 lnet_startup_lndnis (void)
1116 {
1117         lnd_t                   *lnd;
1118         struct lnet_ni          *ni;
1119         struct lnet_tx_queue    *tq;
1120         cfs_list_t              nilist;
1121         int                     i;
1122         int                rc = 0;
1123         int                lnd_type;
1124         int                nicount = 0;
1125         char              *nets = lnet_get_networks();
1126
1127         CFS_INIT_LIST_HEAD(&nilist);
1128
1129         if (nets == NULL)
1130                 goto failed;
1131
1132         rc = lnet_parse_networks(&nilist, nets);
1133         if (rc != 0)
1134                 goto failed;
1135
1136         while (!cfs_list_empty(&nilist)) {
1137                 ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
1138                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
1139
1140                 LASSERT (libcfs_isknown_lnd(lnd_type));
1141
1142                 if (lnd_type == CIBLND    ||
1143                     lnd_type == OPENIBLND ||
1144                     lnd_type == IIBLND    ||
1145                     lnd_type == VIBLND) {
1146                         CERROR("LND %s obsoleted\n",
1147                                libcfs_lnd2str(lnd_type));
1148                         goto failed;
1149                 }
1150
1151                 LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
1152                 lnd = lnet_find_lnd_by_type(lnd_type);
1153
1154 #ifdef __KERNEL__
1155                 if (lnd == NULL) {
1156                         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1157                         rc = cfs_request_module("%s",
1158                                                 libcfs_lnd2modname(lnd_type));
1159                         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
1160
1161                         lnd = lnet_find_lnd_by_type(lnd_type);
1162                         if (lnd == NULL) {
1163                                 LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1164                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1165                                        libcfs_lnd2str(lnd_type),
1166                                        libcfs_lnd2modname(lnd_type), rc);
1167 #ifndef HAVE_MODULE_LOADING_SUPPORT
1168                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1169                                          "compiled with kernel module "
1170                                          "loading support.");
1171 #endif
1172                                 goto failed;
1173                         }
1174                 }
1175 #else
1176                 if (lnd == NULL) {
1177                         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1178                         CERROR("LND %s not supported\n",
1179                                libcfs_lnd2str(lnd_type));
1180                         goto failed;
1181                 }
1182 #endif
1183
1184                 lnet_net_lock(LNET_LOCK_EX);
1185                 lnd->lnd_refcount++;
1186                 lnet_net_unlock(LNET_LOCK_EX);
1187
1188                 ni->ni_lnd = lnd;
1189
1190                 rc = (lnd->lnd_startup)(ni);
1191
1192                 LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1193
1194                 if (rc != 0) {
1195                         LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
1196                                            "\n",
1197                                            rc, libcfs_lnd2str(lnd->lnd_type));
1198                         lnet_net_lock(LNET_LOCK_EX);
1199                         lnd->lnd_refcount--;
1200                         lnet_net_unlock(LNET_LOCK_EX);
1201                         goto failed;
1202                 }
1203
1204                 LASSERT (ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1205
1206                 cfs_list_del(&ni->ni_list);
1207
1208                 lnet_net_lock(LNET_LOCK_EX);
1209                 /* refcount for ln_nis */
1210                 lnet_ni_addref_locked(ni, 0);
1211                 cfs_list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1212                 if (ni->ni_cpts != NULL) {
1213                         cfs_list_add_tail(&ni->ni_cptlist,
1214                                           &the_lnet.ln_nis_cpt);
1215                         lnet_ni_addref_locked(ni, 0);
1216                 }
1217
1218                 lnet_net_unlock(LNET_LOCK_EX);
1219
1220                 if (lnd->lnd_type == LOLND) {
1221                         lnet_ni_addref(ni);
1222                         LASSERT (the_lnet.ln_loni == NULL);
1223                         the_lnet.ln_loni = ni;
1224                         continue;
1225                 }
1226
1227 #ifndef __KERNEL__
1228                 if (lnd->lnd_wait != NULL) {
1229                         if (the_lnet.ln_eq_waitni == NULL) {
1230                                 lnet_ni_addref(ni);
1231                                 the_lnet.ln_eq_waitni = ni;
1232                         }
1233                 } else {
1234 # ifndef HAVE_LIBPTHREAD
1235                         LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1236                                            "single-threaded runtime\n",
1237                                            libcfs_lnd2str(lnd_type));
1238                         goto failed;
1239 # endif
1240                 }
1241 #endif
1242                 if (ni->ni_peertxcredits == 0 ||
1243                     ni->ni_maxtxcredits == 0) {
1244                         LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1245                                            libcfs_lnd2str(lnd->lnd_type),
1246                                            ni->ni_peertxcredits == 0 ?
1247                                            "" : "per-peer ");
1248                         goto failed;
1249                 }
1250
1251                 cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1252                         tq->tq_credits_min =
1253                         tq->tq_credits_max =
1254                         tq->tq_credits = lnet_ni_tq_credits(ni);
1255                 }
1256
1257                 CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1258                        libcfs_nid2str(ni->ni_nid), ni->ni_peertxcredits,
1259                        lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1260                        ni->ni_peerrtrcredits, ni->ni_peertimeout);
1261
1262                 nicount++;
1263         }
1264
1265         if (the_lnet.ln_eq_waitni != NULL && nicount > 1) {
1266                 lnd_type = the_lnet.ln_eq_waitni->ni_lnd->lnd_type;
1267                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1268                                    "\n",
1269                                    libcfs_lnd2str(lnd_type));
1270                 goto failed;
1271         }
1272
1273         return 0;
1274
1275  failed:
1276         lnet_shutdown_lndnis();
1277
1278         while (!cfs_list_empty(&nilist)) {
1279                 ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
1280                 cfs_list_del(&ni->ni_list);
1281                 lnet_ni_free(ni);
1282         }
1283
1284         return -ENETDOWN;
1285 }
1286
1287 /**
1288  * Initialize LNet library.
1289  *
1290  * Only userspace program needs to call this function - it's automatically
1291  * called in the kernel at module loading time. Caller has to call LNetFini()
1292  * after a call to LNetInit(), if and only if the latter returned 0. It must
1293  * be called exactly once.
1294  *
1295  * \return 0 on success, and -ve on failures.
1296  */
1297 int
1298 LNetInit(void)
1299 {
1300         int     rc;
1301
1302         lnet_assert_wire_constants ();
1303         LASSERT (!the_lnet.ln_init);
1304
1305         memset(&the_lnet, 0, sizeof(the_lnet));
1306
1307         /* refer to global cfs_cpt_table for now */
1308         the_lnet.ln_cpt_table   = cfs_cpt_table;
1309         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1310
1311         LASSERT(the_lnet.ln_cpt_number > 0);
1312         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1313                 /* we are under risk of consuming all lh_cookie */
1314                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
1315                        "please change setting of CPT-table and retry\n",
1316                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1317                 return -1;
1318         }
1319
1320         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1321                 the_lnet.ln_cpt_bits++;
1322
1323         rc = lnet_create_locks();
1324         if (rc != 0) {
1325                 CERROR("Can't create LNet global locks: %d\n", rc);
1326                 return -1;
1327         }
1328
1329         the_lnet.ln_refcount = 0;
1330         the_lnet.ln_init = 1;
1331         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1332         CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
1333         CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1334         CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1335
1336 #ifdef __KERNEL__
1337         /* All LNDs apart from the LOLND are in separate modules.  They
1338          * register themselves when their module loads, and unregister
1339          * themselves when their module is unloaded. */
1340 #else
1341         /* Register LNDs
1342          * NB the order here determines default 'networks=' order */
1343 # ifdef CRAY_XT3
1344         LNET_REGISTER_ULND(the_ptllnd);
1345 # endif
1346 # ifdef HAVE_LIBPTHREAD
1347         LNET_REGISTER_ULND(the_tcplnd);
1348 # endif
1349 #endif
1350         lnet_register_lnd(&the_lolnd);
1351         return 0;
1352 }
1353
1354 /**
1355  * Finalize LNet library.
1356  *
1357  * Only userspace program needs to call this function. It can be called
1358  * at most once.
1359  *
1360  * \pre LNetInit() called with success.
1361  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1362  */
1363 void
1364 LNetFini(void)
1365 {
1366         LASSERT(the_lnet.ln_init);
1367         LASSERT(the_lnet.ln_refcount == 0);
1368
1369         while (!cfs_list_empty(&the_lnet.ln_lnds))
1370                 lnet_unregister_lnd(cfs_list_entry(the_lnet.ln_lnds.next,
1371                                                    lnd_t, lnd_list));
1372         lnet_destroy_locks();
1373
1374         the_lnet.ln_init = 0;
1375 }
1376
1377 /**
1378  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1379  *
1380  * Userspace program should call this after a successful call to LNetInit().
1381  * Users must call this function at least once before any other functions.
1382  * For each successful call there must be a corresponding call to
1383  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1384  * ignored.
1385  *
1386  * The PID used by LNet may be different from the one requested.
1387  * See LNetGetId().
1388  *
1389  * \param requested_pid PID requested by the caller.
1390  *
1391  * \return >= 0 on success, and < 0 error code on failures.
1392  */
1393 int
1394 LNetNIInit(lnet_pid_t requested_pid)
1395 {
1396         int         im_a_router = 0;
1397         int         rc;
1398
1399         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
1400
1401         LASSERT (the_lnet.ln_init);
1402         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1403
1404         if (the_lnet.ln_refcount > 0) {
1405                 rc = the_lnet.ln_refcount++;
1406                 goto out;
1407         }
1408
1409         lnet_get_tunables();
1410
1411         if (requested_pid == LNET_PID_ANY) {
1412                 /* Don't instantiate LNET just for me */
1413                 rc = -ENETDOWN;
1414                 goto failed0;
1415         }
1416
1417         rc = lnet_prepare(requested_pid);
1418         if (rc != 0)
1419                 goto failed0;
1420
1421         rc = lnet_startup_lndnis();
1422         if (rc != 0)
1423                 goto failed1;
1424
1425         rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1426         if (rc != 0)
1427                 goto failed2;
1428
1429         rc = lnet_check_routes();
1430         if (rc != 0)
1431                 goto failed2;
1432
1433         rc = lnet_rtrpools_alloc(im_a_router);
1434         if (rc != 0)
1435                 goto failed2;
1436
1437         rc = lnet_acceptor_start();
1438         if (rc != 0)
1439                 goto failed2;
1440
1441         the_lnet.ln_refcount = 1;
1442         /* Now I may use my own API functions... */
1443
1444         /* NB router checker needs the_lnet.ln_ping_info in
1445          * lnet_router_checker -> lnet_update_ni_status_locked */
1446         rc = lnet_ping_target_init();
1447         if (rc != 0)
1448                 goto failed3;
1449
1450         rc = lnet_router_checker_start();
1451         if (rc != 0)
1452                 goto failed4;
1453
1454         lnet_proc_init();
1455         goto out;
1456
1457  failed4:
1458         lnet_ping_target_fini();
1459  failed3:
1460         the_lnet.ln_refcount = 0;
1461         lnet_acceptor_stop();
1462  failed2:
1463         lnet_destroy_routes();
1464         lnet_shutdown_lndnis();
1465  failed1:
1466         lnet_unprepare();
1467  failed0:
1468         LASSERT (rc < 0);
1469  out:
1470         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1471         return rc;
1472 }
1473
1474 /**
1475  * Stop LNet interfaces, routing, and forwarding.
1476  *
1477  * Users must call this function once for each successful call to LNetNIInit().
1478  * Once the LNetNIFini() operation has been started, the results of pending
1479  * API operations are undefined.
1480  *
1481  * \return always 0 for current implementation.
1482  */
1483 int
1484 LNetNIFini()
1485 {
1486         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
1487
1488         LASSERT (the_lnet.ln_init);
1489         LASSERT (the_lnet.ln_refcount > 0);
1490
1491         if (the_lnet.ln_refcount != 1) {
1492                 the_lnet.ln_refcount--;
1493         } else {
1494                 LASSERT (!the_lnet.ln_niinit_self);
1495
1496                 lnet_proc_fini();
1497                 lnet_router_checker_stop();
1498                 lnet_ping_target_fini();
1499
1500                 /* Teardown fns that use my own API functions BEFORE here */
1501                 the_lnet.ln_refcount = 0;
1502
1503                 lnet_acceptor_stop();
1504                 lnet_destroy_routes();
1505                 lnet_shutdown_lndnis();
1506                 lnet_unprepare();
1507         }
1508
1509         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1510         return 0;
1511 }
1512
1513 /**
1514  * This is an ugly hack to export IOC_LIBCFS_DEBUG_PEER and
1515  * IOC_LIBCFS_PORTALS_COMPATIBILITY commands to users, by tweaking the LNet
1516  * internal ioctl handler.
1517  *
1518  * IOC_LIBCFS_PORTALS_COMPATIBILITY is now deprecated, don't use it.
1519  *
1520  * \param cmd IOC_LIBCFS_DEBUG_PEER to print debugging data about a peer.
1521  * The data will be printed to system console. Don't use it excessively.
1522  * \param arg A pointer to lnet_process_id_t, process ID of the peer.
1523  *
1524  * \return Always return 0 when called by users directly (i.e., not via ioctl).
1525  */
1526 int
1527 LNetCtl(unsigned int cmd, void *arg)
1528 {
1529         struct libcfs_ioctl_data *data = arg;
1530         lnet_process_id_t         id = {0};
1531         lnet_ni_t                *ni;
1532         int                       rc;
1533
1534         LASSERT (the_lnet.ln_init);
1535         LASSERT (the_lnet.ln_refcount > 0);
1536
1537         switch (cmd) {
1538         case IOC_LIBCFS_GET_NI:
1539                 rc = LNetGetId(data->ioc_count, &id);
1540                 data->ioc_nid = id.nid;
1541                 return rc;
1542
1543         case IOC_LIBCFS_FAIL_NID:
1544                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1545
1546         case IOC_LIBCFS_ADD_ROUTE:
1547                 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1548                                     data->ioc_nid);
1549                 return (rc != 0) ? rc : lnet_check_routes();
1550
1551         case IOC_LIBCFS_DEL_ROUTE:
1552                 return lnet_del_route(data->ioc_net, data->ioc_nid);
1553
1554         case IOC_LIBCFS_GET_ROUTE:
1555                 return lnet_get_route(data->ioc_count,
1556                                       &data->ioc_net, &data->ioc_count,
1557                                       &data->ioc_nid, &data->ioc_flags);
1558         case IOC_LIBCFS_NOTIFY_ROUTER:
1559                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1560                                    cfs_time_current() -
1561                                    cfs_time_seconds(cfs_time_current_sec() -
1562                                                     (time_t)data->ioc_u64[0]));
1563
1564         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1565                 /* This can be removed once lustre stops calling it */
1566                 return 0;
1567
1568         case IOC_LIBCFS_LNET_DIST:
1569                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1570                 if (rc < 0 && rc != -EHOSTUNREACH)
1571                         return rc;
1572
1573                 data->ioc_u32[0] = rc;
1574                 return 0;
1575
1576         case IOC_LIBCFS_TESTPROTOCOMPAT:
1577                 lnet_net_lock(LNET_LOCK_EX);
1578                 the_lnet.ln_testprotocompat = data->ioc_flags;
1579                 lnet_net_unlock(LNET_LOCK_EX);
1580                 return 0;
1581
1582         case IOC_LIBCFS_PING:
1583                 id.nid = data->ioc_nid;
1584                 id.pid = data->ioc_u32[0];
1585                 rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
1586                                (lnet_process_id_t *)data->ioc_pbuf1,
1587                                data->ioc_plen1/sizeof(lnet_process_id_t));
1588                 if (rc < 0)
1589                         return rc;
1590                 data->ioc_count = rc;
1591                 return 0;
1592
1593         case IOC_LIBCFS_DEBUG_PEER: {
1594                 /* CAVEAT EMPTOR: this one designed for calling directly; not
1595                  * via an ioctl */
1596                 id = *((lnet_process_id_t *) arg);
1597
1598                 lnet_debug_peer(id.nid);
1599
1600                 ni = lnet_net2ni(LNET_NIDNET(id.nid));
1601                 if (ni == NULL) {
1602                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(id));
1603                 } else {
1604                         if (ni->ni_lnd->lnd_ctl == NULL) {
1605                                 CDEBUG(D_WARNING, "No ctl for %s\n",
1606                                        libcfs_id2str(id));
1607                         } else {
1608                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1609                         }
1610
1611                         lnet_ni_decref(ni);
1612                 }
1613                 return 0;
1614         }
1615
1616         default:
1617                 ni = lnet_net2ni(data->ioc_net);
1618                 if (ni == NULL)
1619                         return -EINVAL;
1620
1621                 if (ni->ni_lnd->lnd_ctl == NULL)
1622                         rc = -EINVAL;
1623                 else
1624                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1625
1626                 lnet_ni_decref(ni);
1627                 return rc;
1628         }
1629         /* not reached */
1630 }
1631
1632 /**
1633  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
1634  * all interfaces share a same PID, as requested by LNetNIInit().
1635  *
1636  * \param index Index of the interface to look up.
1637  * \param id On successful return, this location will hold the
1638  * lnet_process_id_t ID of the interface.
1639  *
1640  * \retval 0 If an interface exists at \a index.
1641  * \retval -ENOENT If no interface has been found.
1642  */
1643 int
1644 LNetGetId(unsigned int index, lnet_process_id_t *id)
1645 {
1646         struct lnet_ni  *ni;
1647         cfs_list_t      *tmp;
1648         int             cpt;
1649         int             rc = -ENOENT;
1650
1651         LASSERT(the_lnet.ln_init);
1652         LASSERT(the_lnet.ln_refcount > 0);
1653
1654         cpt = lnet_net_lock_current();
1655
1656         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
1657                 if (index-- != 0)
1658                         continue;
1659
1660                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
1661
1662                 id->nid = ni->ni_nid;
1663                 id->pid = the_lnet.ln_pid;
1664                 rc = 0;
1665                 break;
1666         }
1667
1668         lnet_net_unlock(cpt);
1669         return rc;
1670 }
1671
1672 /**
1673  * Print a string representation of handle \a h into buffer \a str of
1674  * \a len bytes.
1675  */
1676 void
1677 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1678 {
1679         snprintf(str, len, LPX64, h.cookie);
1680 }
1681
1682 static int
1683 lnet_create_ping_info(void)
1684 {
1685         int               i;
1686         int               n;
1687         int               rc;
1688         unsigned int      infosz;
1689         lnet_ni_t        *ni;
1690         lnet_process_id_t id;
1691         lnet_ping_info_t *pinfo;
1692
1693         for (n = 0; ; n++) {
1694                 rc = LNetGetId(n, &id);
1695                 if (rc == -ENOENT)
1696                         break;
1697
1698                 LASSERT (rc == 0);
1699         }
1700
1701         infosz = offsetof(lnet_ping_info_t, pi_ni[n]);
1702         LIBCFS_ALLOC(pinfo, infosz);
1703         if (pinfo == NULL) {
1704                 CERROR("Can't allocate ping info[%d]\n", n);
1705                 return -ENOMEM;
1706         }
1707
1708         pinfo->pi_nnis    = n;
1709         pinfo->pi_pid     = the_lnet.ln_pid;
1710         pinfo->pi_magic   = LNET_PROTO_PING_MAGIC;
1711         pinfo->pi_version = LNET_PROTO_PING_VERSION;
1712
1713         for (i = 0; i < n; i++) {
1714                 lnet_ni_status_t *ns = &pinfo->pi_ni[i];
1715
1716                 rc = LNetGetId(i, &id);
1717                 LASSERT (rc == 0);
1718
1719                 ns->ns_nid    = id.nid;
1720                 ns->ns_status = LNET_NI_STATUS_UP;
1721
1722                 lnet_net_lock(0);
1723
1724                 ni = lnet_nid2ni_locked(id.nid, 0);
1725                 LASSERT(ni != NULL);
1726
1727                 lnet_ni_lock(ni);
1728                 LASSERT(ni->ni_status == NULL);
1729                 ni->ni_status = ns;
1730                 lnet_ni_unlock(ni);
1731
1732                 lnet_ni_decref_locked(ni, 0);
1733                 lnet_net_unlock(0);
1734         }
1735
1736         the_lnet.ln_ping_info = pinfo;
1737         return 0;
1738 }
1739
1740 static void
1741 lnet_destroy_ping_info(void)
1742 {
1743         struct lnet_ni  *ni;
1744
1745         lnet_net_lock(0);
1746
1747         cfs_list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
1748                 lnet_ni_lock(ni);
1749                 ni->ni_status = NULL;
1750                 lnet_ni_unlock(ni);
1751         }
1752
1753         lnet_net_unlock(0);
1754
1755         LIBCFS_FREE(the_lnet.ln_ping_info,
1756                     offsetof(lnet_ping_info_t,
1757                              pi_ni[the_lnet.ln_ping_info->pi_nnis]));
1758         the_lnet.ln_ping_info = NULL;
1759         return;
1760 }
1761
1762 int
1763 lnet_ping_target_init(void)
1764 {
1765         lnet_md_t         md = {0};
1766         lnet_handle_me_t  meh;
1767         lnet_process_id_t id;
1768         int               rc;
1769         int               rc2;
1770         int               infosz;
1771
1772         rc = lnet_create_ping_info();
1773         if (rc != 0)
1774                 return rc;
1775
1776         /* We can have a tiny EQ since we only need to see the unlink event on
1777          * teardown, which by definition is the last one! */
1778         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1779         if (rc != 0) {
1780                 CERROR("Can't allocate ping EQ: %d\n", rc);
1781                 goto failed_0;
1782         }
1783
1784         memset(&id, 0, sizeof(lnet_process_id_t));
1785         id.nid = LNET_NID_ANY;
1786         id.pid = LNET_PID_ANY;
1787
1788         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1789                           LNET_PROTO_PING_MATCHBITS, 0,
1790                           LNET_UNLINK, LNET_INS_AFTER,
1791                           &meh);
1792         if (rc != 0) {
1793                 CERROR("Can't create ping ME: %d\n", rc);
1794                 goto failed_1;
1795         }
1796
1797         /* initialize md content */
1798         infosz = offsetof(lnet_ping_info_t,
1799                           pi_ni[the_lnet.ln_ping_info->pi_nnis]);
1800         md.start     = the_lnet.ln_ping_info;
1801         md.length    = infosz;
1802         md.threshold = LNET_MD_THRESH_INF;
1803         md.max_size  = 0;
1804         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1805                        LNET_MD_MANAGE_REMOTE;
1806         md.user_ptr  = NULL;
1807         md.eq_handle = the_lnet.ln_ping_target_eq;
1808
1809         rc = LNetMDAttach(meh, md,
1810                           LNET_RETAIN,
1811                           &the_lnet.ln_ping_target_md);
1812         if (rc != 0) {
1813                 CERROR("Can't attach ping MD: %d\n", rc);
1814                 goto failed_2;
1815         }
1816
1817         return 0;
1818
1819  failed_2:
1820         rc2 = LNetMEUnlink(meh);
1821         LASSERT (rc2 == 0);
1822  failed_1:
1823         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1824         LASSERT (rc2 == 0);
1825  failed_0:
1826         lnet_destroy_ping_info();
1827         return rc;
1828 }
1829
1830 void
1831 lnet_ping_target_fini(void)
1832 {
1833         lnet_event_t    event;
1834         int             rc;
1835         int             which;
1836         int             timeout_ms = 1000;
1837         cfs_sigset_t    blocked = cfs_block_allsigs();
1838
1839         LNetMDUnlink(the_lnet.ln_ping_target_md);
1840         /* NB md could be busy; this just starts the unlink */
1841
1842         for (;;) {
1843                 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1844                                 timeout_ms, &event, &which);
1845
1846                 /* I expect overflow... */
1847                 LASSERT (rc >= 0 || rc == -EOVERFLOW);
1848
1849                 if (rc == 0) {
1850                         /* timed out: provide a diagnostic */
1851                         CWARN("Still waiting for ping MD to unlink\n");
1852                         timeout_ms *= 2;
1853                         continue;
1854                 }
1855
1856                 /* Got a valid event */
1857                 if (event.unlinked)
1858                         break;
1859         }
1860
1861         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1862         LASSERT (rc == 0);
1863         lnet_destroy_ping_info();
1864         cfs_restore_sigs(blocked);
1865 }
1866
1867 int
1868 lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1869 {
1870         lnet_handle_eq_t     eqh;
1871         lnet_handle_md_t     mdh;
1872         lnet_event_t         event;
1873         lnet_md_t            md = {0};
1874         int                  which;
1875         int                  unlinked = 0;
1876         int                  replied = 0;
1877         const int            a_long_time = 60000; /* mS */
1878         int                  infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
1879         lnet_ping_info_t    *info;
1880         lnet_process_id_t    tmpid;
1881         int                  i;
1882         int                  nob;
1883         int                  rc;
1884         int                  rc2;
1885         cfs_sigset_t         blocked;
1886
1887         if (n_ids <= 0 ||
1888             id.nid == LNET_NID_ANY ||
1889             timeout_ms > 500000 ||              /* arbitrary limit! */
1890             n_ids > 20)                         /* arbitrary limit! */
1891                 return -EINVAL;
1892
1893         if (id.pid == LNET_PID_ANY)
1894                 id.pid = LUSTRE_SRV_LNET_PID;
1895
1896         LIBCFS_ALLOC(info, infosz);
1897         if (info == NULL)
1898                 return -ENOMEM;
1899
1900         /* NB 2 events max (including any unlink event) */
1901         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1902         if (rc != 0) {
1903                 CERROR("Can't allocate EQ: %d\n", rc);
1904                 goto out_0;
1905         }
1906
1907         /* initialize md content */
1908         md.start     = info;
1909         md.length    = infosz;
1910         md.threshold = 2; /*GET/REPLY*/
1911         md.max_size  = 0;
1912         md.options   = LNET_MD_TRUNCATE;
1913         md.user_ptr  = NULL;
1914         md.eq_handle = eqh;
1915
1916         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
1917         if (rc != 0) {
1918                 CERROR("Can't bind MD: %d\n", rc);
1919                 goto out_1;
1920         }
1921
1922         rc = LNetGet(LNET_NID_ANY, mdh, id,
1923                      LNET_RESERVED_PORTAL,
1924                      LNET_PROTO_PING_MATCHBITS, 0);
1925
1926         if (rc != 0) {
1927                 /* Don't CERROR; this could be deliberate! */
1928
1929                 rc2 = LNetMDUnlink(mdh);
1930                 LASSERT (rc2 == 0);
1931
1932                 /* NB must wait for the UNLINK event below... */
1933                 unlinked = 1;
1934                 timeout_ms = a_long_time;
1935         }
1936
1937         do {
1938                 /* MUST block for unlink to complete */
1939                 if (unlinked)
1940                         blocked = cfs_block_allsigs();
1941
1942                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
1943
1944                 if (unlinked)
1945                         cfs_restore_sigs(blocked);
1946
1947                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
1948                        (rc2 <= 0) ? -1 : event.type,
1949                        (rc2 <= 0) ? -1 : event.status,
1950                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
1951
1952                 LASSERT (rc2 != -EOVERFLOW);     /* can't miss anything */
1953
1954                 if (rc2 <= 0 || event.status != 0) {
1955                         /* timeout or error */
1956                         if (!replied && rc == 0)
1957                                 rc = (rc2 < 0) ? rc2 :
1958                                      (rc2 == 0) ? -ETIMEDOUT :
1959                                      event.status;
1960
1961                         if (!unlinked) {
1962                                 /* Ensure completion in finite time... */
1963                                 LNetMDUnlink(mdh);
1964                                 /* No assertion (racing with network) */
1965                                 unlinked = 1;
1966                                 timeout_ms = a_long_time;
1967                         } else if (rc2 == 0) {
1968                                 /* timed out waiting for unlink */
1969                                 CWARN("ping %s: late network completion\n",
1970                                       libcfs_id2str(id));
1971                         }
1972                 } else if (event.type == LNET_EVENT_REPLY) {
1973                         replied = 1;
1974                         rc = event.mlength;
1975                 }
1976
1977         } while (rc2 <= 0 || !event.unlinked);
1978
1979         if (!replied) {
1980                 if (rc >= 0)
1981                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
1982                               libcfs_id2str(id));
1983                 rc = -EIO;
1984                 goto out_1;
1985         }
1986
1987         nob = rc;
1988         LASSERT (nob >= 0 && nob <= infosz);
1989
1990         rc = -EPROTO;                           /* if I can't parse... */
1991
1992         if (nob < 8) {
1993                 /* can't check magic/version */
1994                 CERROR("%s: ping info too short %d\n",
1995                        libcfs_id2str(id), nob);
1996                 goto out_1;
1997         }
1998
1999         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
2000                 lnet_swap_pinginfo(info);
2001         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
2002                 CERROR("%s: Unexpected magic %08x\n",
2003                        libcfs_id2str(id), info->pi_magic);
2004                 goto out_1;
2005         }
2006
2007         if (info->pi_version != LNET_PROTO_PING_VERSION) {
2008                 CERROR("%s: Unexpected version 0x%x\n",
2009                        libcfs_id2str(id), info->pi_version);
2010                 goto out_1;
2011         }
2012
2013         if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
2014                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
2015                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
2016                 goto out_1;
2017         }
2018
2019         if (info->pi_nnis < n_ids)
2020                 n_ids = info->pi_nnis;
2021
2022         if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
2023                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
2024                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
2025                 goto out_1;
2026         }
2027
2028         rc = -EFAULT;                           /* If I SEGV... */
2029
2030         for (i = 0; i < n_ids; i++) {
2031                 tmpid.pid = info->pi_pid;
2032                 tmpid.nid = info->pi_ni[i].ns_nid;
2033 #ifdef __KERNEL__
2034                 if (cfs_copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
2035                         goto out_1;
2036 #else
2037                 ids[i] = tmpid;
2038 #endif
2039         }
2040         rc = info->pi_nnis;
2041
2042  out_1:
2043         rc2 = LNetEQFree(eqh);
2044         if (rc2 != 0)
2045                 CERROR("rc2 %d\n", rc2);
2046         LASSERT (rc2 == 0);
2047
2048  out_0:
2049         LIBCFS_FREE(info, infosz);
2050         return rc;
2051 }