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 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_MASK) != 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         rc = lnet_msg_containers_create();
 742         if (rc != 0)
 743                 goto failed;
 744
 745         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 746                                       LNET_COOKIE_TYPE_EQ, LNET_FL_MAX_EQS,
 747                                       sizeof(lnet_eq_t));
 748         if (rc != 0)
 749                 goto failed;
 750
 751         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME, LNET_FL_MAX_MES,
 752                                           sizeof(lnet_me_t));
 753         if (recs == NULL)
 754                 goto failed;
 755
 756         the_lnet.ln_me_containers = recs;
 757
 758         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD, LNET_FL_MAX_MDS,
 759                                           sizeof(lnet_libmd_t));
 760         if (recs == NULL)
 761                 goto failed;
 762
 763         the_lnet.ln_md_containers = recs;
 764
 765         rc = lnet_portals_create();
 766         if (rc != 0) {
 767                 CERROR("Failed to create portals for LNet: %d\n", rc);
 768                 goto failed;
 769         }
 770
 771         return 0;
 772
 773  failed:
 774         lnet_unprepare();
 775         return rc;
 776 }
 777
 778 int
 779 lnet_unprepare (void)
 780 {
 781         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 782          * have shut down already, so it is safe to unlink and free all
 783          * descriptors, even those that appear committed to a network op (eg MD
 784          * with non-zero pending count) */
 785
 786         lnet_fail_nid(LNET_NID_ANY, 0);
 787
 788         LASSERT(the_lnet.ln_refcount == 0);
 789         LASSERT(cfs_list_empty(&the_lnet.ln_test_peers));
 790         LASSERT(cfs_list_empty(&the_lnet.ln_nis));
 791         LASSERT(cfs_list_empty(&the_lnet.ln_nis_cpt));
 792         LASSERT(cfs_list_empty(&the_lnet.ln_nis_zombie));
 793
 794         lnet_portals_destroy();
 795
 796         if (the_lnet.ln_md_containers != NULL) {
 797                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 798                 the_lnet.ln_md_containers = NULL;
 799         }
 800
 801         if (the_lnet.ln_me_containers != NULL) {
 802                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 803                 the_lnet.ln_me_containers = NULL;
 804         }
 805
 806         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 807
 808         lnet_msg_containers_destroy();
 809         lnet_peer_tables_destroy();
 810         lnet_rtrpools_free();
 811
 812         if (the_lnet.ln_counters != NULL) {
 813                 cfs_percpt_free(the_lnet.ln_counters);
 814                 the_lnet.ln_counters = NULL;
 815         }
 816
 817         return 0;
 818 }
 819
 820 lnet_ni_t  *
 821 lnet_net2ni_locked(__u32 net, int cpt)
 822 {
 823         cfs_list_t      *tmp;
 824         lnet_ni_t       *ni;
 825
 826         LASSERT(cpt != LNET_LOCK_EX);
 827
 828         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 829                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 830
 831                 if (LNET_NIDNET(ni->ni_nid) == net) {
 832                         lnet_ni_addref_locked(ni, cpt);
 833                         return ni;
 834                 }
 835         }
 836
 837         return NULL;
 838 }
 839
 840 lnet_ni_t *
 841 lnet_net2ni(__u32 net)
 842 {
 843         lnet_ni_t *ni;
 844
 845         lnet_net_lock(0);
 846         ni = lnet_net2ni_locked(net, 0);
 847         lnet_net_unlock(0);
 848
 849         return ni;
 850 }
 851 EXPORT_SYMBOL(lnet_net2ni);
 852
 853 static unsigned int
 854 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 855 {
 856         __u64           key = nid;
 857         unsigned int    val;
 858
 859         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 860
 861         if (number == 1)
 862                 return 0;
 863
 864         val = cfs_hash_long(key, LNET_CPT_BITS);
 865         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 866         if (val < number)
 867                 return val;
 868
 869         return (unsigned int)(key + val + (val >> 1)) % number;
 870 }
 871
 872 int
 873 lnet_cpt_of_nid_locked(lnet_nid_t nid)
 874 {
 875         struct lnet_ni *ni;
 876
 877         /* must called with hold of lnet_net_lock */
 878         if (LNET_CPT_NUMBER == 1)
 879                 return 0; /* the only one */
 880
 881         /* take lnet_net_lock(any) would be OK */
 882         if (!cfs_list_empty(&the_lnet.ln_nis_cpt)) {
 883                 cfs_list_for_each_entry(ni, &the_lnet.ln_nis_cpt, ni_cptlist) {
 884                         if (LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid))
 885                                 continue;
 886
 887                         LASSERT(ni->ni_cpts != NULL);
 888                         return ni->ni_cpts[lnet_nid_cpt_hash
 889                                            (nid, ni->ni_ncpts)];
 890                 }
 891         }
 892
 893         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 894 }
 895
 896 int
 897 lnet_cpt_of_nid(lnet_nid_t nid)
 898 {
 899         int     cpt;
 900         int     cpt2;
 901
 902         if (LNET_CPT_NUMBER == 1)
 903                 return 0; /* the only one */
 904
 905         if (cfs_list_empty(&the_lnet.ln_nis_cpt))
 906                 return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 907
 908         cpt = lnet_net_lock_current();
 909         cpt2 = lnet_cpt_of_nid_locked(nid);
 910         lnet_net_unlock(cpt);
 911
 912         return cpt2;
 913 }
 914 EXPORT_SYMBOL(lnet_cpt_of_nid);
 915
 916 int
 917 lnet_islocalnet(__u32 net)
 918 {
 919         struct lnet_ni  *ni;
 920         int             cpt;
 921
 922         cpt = lnet_net_lock_current();
 923
 924         ni = lnet_net2ni_locked(net, cpt);
 925         if (ni != NULL)
 926                 lnet_ni_decref_locked(ni, cpt);
 927
 928         lnet_net_unlock(cpt);
 929
 930         return ni != NULL;
 931 }
 932
 933 lnet_ni_t  *
 934 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 935 {
 936         struct lnet_ni  *ni;
 937         cfs_list_t      *tmp;
 938
 939         LASSERT(cpt != LNET_LOCK_EX);
 940
 941         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 942                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 943
 944                 if (ni->ni_nid == nid) {
 945                         lnet_ni_addref_locked(ni, cpt);
 946                         return ni;
 947                 }
 948         }
 949
 950         return NULL;
 951 }
 952
 953 int
 954 lnet_islocalnid(lnet_nid_t nid)
 955 {
 956         struct lnet_ni  *ni;
 957         int             cpt;
 958
 959         cpt = lnet_net_lock_current();
 960         ni = lnet_nid2ni_locked(nid, cpt);
 961         if (ni != NULL)
 962                 lnet_ni_decref_locked(ni, cpt);
 963         lnet_net_unlock(cpt);
 964
 965         return ni != NULL;
 966 }
 967
 968 int
 969 lnet_count_acceptor_nis (void)
 970 {
 971         /* Return the # of NIs that need the acceptor. */
 972         int             count = 0;
 973 #if defined(__KERNEL__) || defined(HAVE_LIBPTHREAD)
 974         cfs_list_t      *tmp;
 975         struct lnet_ni  *ni;
 976         int             cpt;
 977
 978         cpt = lnet_net_lock_current();
 979         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
 980                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
 981
 982                 if (ni->ni_lnd->lnd_accept != NULL)
 983                         count++;
 984         }
 985
 986         lnet_net_unlock(cpt);
 987
 988 #endif /* defined(__KERNEL__) || defined(HAVE_LIBPTHREAD) */
 989         return count;
 990 }
 991
 992 static int
 993 lnet_ni_tq_credits(lnet_ni_t *ni)
 994 {
 995         int     credits;
 996
 997         LASSERT(ni->ni_ncpts >= 1);
 998
 999         if (ni->ni_ncpts == 1)
1000                 return ni->ni_maxtxcredits;
1001
1002         credits = ni->ni_maxtxcredits / ni->ni_ncpts;
1003         credits = max(credits, 8 * ni->ni_peertxcredits);
1004         credits = min(credits, ni->ni_maxtxcredits);
1005
1006         return credits;
1007 }
1008
1009 void
1010 lnet_shutdown_lndnis (void)
1011 {
1012         int                i;
1013         int                islo;
1014         lnet_ni_t         *ni;
1015
1016         /* NB called holding the global mutex */
1017
1018         /* All quiet on the API front */
1019         LASSERT(!the_lnet.ln_shutdown);
1020         LASSERT(the_lnet.ln_refcount == 0);
1021         LASSERT(cfs_list_empty(&the_lnet.ln_nis_zombie));
1022         LASSERT(cfs_list_empty(&the_lnet.ln_remote_nets));
1023
1024         lnet_net_lock(LNET_LOCK_EX);
1025         the_lnet.ln_shutdown = 1;       /* flag shutdown */
1026
1027         /* Unlink NIs from the global table */
1028         while (!cfs_list_empty(&the_lnet.ln_nis)) {
1029                 ni = cfs_list_entry(the_lnet.ln_nis.next,
1030                                     lnet_ni_t, ni_list);
1031                 /* move it to zombie list and nobody can find it anymore */
1032                 cfs_list_move(&ni->ni_list, &the_lnet.ln_nis_zombie);
1033                 lnet_ni_decref_locked(ni, 0);   /* drop ln_nis' ref */
1034
1035                 if (!cfs_list_empty(&ni->ni_cptlist)) {
1036                         cfs_list_del_init(&ni->ni_cptlist);
1037                         lnet_ni_decref_locked(ni, 0);
1038                 }
1039         }
1040
1041         /* Drop the cached eqwait NI. */
1042         if (the_lnet.ln_eq_waitni != NULL) {
1043                 lnet_ni_decref_locked(the_lnet.ln_eq_waitni, 0);
1044                 the_lnet.ln_eq_waitni = NULL;
1045         }
1046
1047         /* Drop the cached loopback NI. */
1048         if (the_lnet.ln_loni != NULL) {
1049                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1050                 the_lnet.ln_loni = NULL;
1051         }
1052
1053         lnet_net_unlock(LNET_LOCK_EX);
1054
1055         /* Clear lazy portals and drop delayed messages which hold refs
1056          * on their lnet_msg_t::msg_rxpeer */
1057         for (i = 0; i < the_lnet.ln_nportals; i++)
1058                 LNetClearLazyPortal(i);
1059
1060         /* Clear the peer table and wait for all peers to go (they hold refs on
1061          * their NIs) */
1062         lnet_peer_tables_cleanup();
1063
1064         lnet_net_lock(LNET_LOCK_EX);
1065         /* Now wait for the NI's I just nuked to show up on ln_zombie_nis
1066          * and shut them down in guaranteed thread context */
1067         i = 2;
1068         while (!cfs_list_empty(&the_lnet.ln_nis_zombie)) {
1069                 int     *ref;
1070                 int     j;
1071
1072                 ni = cfs_list_entry(the_lnet.ln_nis_zombie.next,
1073                                     lnet_ni_t, ni_list);
1074                 cfs_list_del_init(&ni->ni_list);
1075                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1076                         if (*ref == 0)
1077                                 continue;
1078                         /* still busy, add it back to zombie list */
1079                         cfs_list_add(&ni->ni_list, &the_lnet.ln_nis_zombie);
1080                         break;
1081                 }
1082
1083                 while (!cfs_list_empty(&ni->ni_list)) {
1084                         lnet_net_unlock(LNET_LOCK_EX);
1085                         ++i;
1086                         if ((i & (-i)) == i) {
1087                                 CDEBUG(D_WARNING,
1088                                        "Waiting for zombie LNI %s\n",
1089                                        libcfs_nid2str(ni->ni_nid));
1090                         }
1091                         cfs_pause(cfs_time_seconds(1));
1092                         lnet_net_lock(LNET_LOCK_EX);
1093                         continue;
1094                 }
1095
1096                 ni->ni_lnd->lnd_refcount--;
1097                 lnet_net_unlock(LNET_LOCK_EX);
1098
1099                 islo = ni->ni_lnd->lnd_type == LOLND;
1100
1101                 LASSERT (!cfs_in_interrupt ());
1102                 (ni->ni_lnd->lnd_shutdown)(ni);
1103
1104                 /* can't deref lnd anymore now; it might have unregistered
1105                  * itself...  */
1106
1107                 if (!islo)
1108                         CDEBUG(D_LNI, "Removed LNI %s\n",
1109                                libcfs_nid2str(ni->ni_nid));
1110
1111                 lnet_ni_free(ni);
1112                 lnet_net_lock(LNET_LOCK_EX);
1113         }
1114
1115         the_lnet.ln_shutdown = 0;
1116         lnet_net_unlock(LNET_LOCK_EX);
1117
1118         if (the_lnet.ln_network_tokens != NULL) {
1119                 LIBCFS_FREE(the_lnet.ln_network_tokens,
1120                             the_lnet.ln_network_tokens_nob);
1121                 the_lnet.ln_network_tokens = NULL;
1122         }
1123 }
1124
1125 int
1126 lnet_startup_lndnis (void)
1127 {
1128         lnd_t                   *lnd;
1129         struct lnet_ni          *ni;
1130         struct lnet_tx_queue    *tq;
1131         cfs_list_t              nilist;
1132         int                     i;
1133         int                rc = 0;
1134         int                lnd_type;
1135         int                nicount = 0;
1136         char              *nets = lnet_get_networks();
1137
1138         CFS_INIT_LIST_HEAD(&nilist);
1139
1140         if (nets == NULL)
1141                 goto failed;
1142
1143         rc = lnet_parse_networks(&nilist, nets);
1144         if (rc != 0)
1145                 goto failed;
1146
1147         while (!cfs_list_empty(&nilist)) {
1148                 ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
1149                 lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
1150
1151                 LASSERT (libcfs_isknown_lnd(lnd_type));
1152
1153                 if (lnd_type == CIBLND    ||
1154                     lnd_type == OPENIBLND ||
1155                     lnd_type == IIBLND    ||
1156                     lnd_type == VIBLND) {
1157                         CERROR("LND %s obsoleted\n",
1158                                libcfs_lnd2str(lnd_type));
1159                         goto failed;
1160                 }
1161
1162                 LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
1163                 lnd = lnet_find_lnd_by_type(lnd_type);
1164
1165 #ifdef __KERNEL__
1166                 if (lnd == NULL) {
1167                         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1168                         rc = cfs_request_module("%s",
1169                                                 libcfs_lnd2modname(lnd_type));
1170                         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
1171
1172                         lnd = lnet_find_lnd_by_type(lnd_type);
1173                         if (lnd == NULL) {
1174                                 LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1175                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1176                                        libcfs_lnd2str(lnd_type),
1177                                        libcfs_lnd2modname(lnd_type), rc);
1178 #ifndef HAVE_MODULE_LOADING_SUPPORT
1179                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1180                                          "compiled with kernel module "
1181                                          "loading support.");
1182 #endif
1183                                 goto failed;
1184                         }
1185                 }
1186 #else
1187                 if (lnd == NULL) {
1188                         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1189                         CERROR("LND %s not supported\n",
1190                                libcfs_lnd2str(lnd_type));
1191                         goto failed;
1192                 }
1193 #endif
1194
1195                 lnet_net_lock(LNET_LOCK_EX);
1196                 lnd->lnd_refcount++;
1197                 lnet_net_unlock(LNET_LOCK_EX);
1198
1199                 ni->ni_lnd = lnd;
1200
1201                 rc = (lnd->lnd_startup)(ni);
1202
1203                 LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1204
1205                 if (rc != 0) {
1206                         LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s"
1207                                            "\n",
1208                                            rc, libcfs_lnd2str(lnd->lnd_type));
1209                         lnet_net_lock(LNET_LOCK_EX);
1210                         lnd->lnd_refcount--;
1211                         lnet_net_unlock(LNET_LOCK_EX);
1212                         goto failed;
1213                 }
1214
1215                 LASSERT (ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1216
1217                 cfs_list_del(&ni->ni_list);
1218
1219                 lnet_net_lock(LNET_LOCK_EX);
1220                 /* refcount for ln_nis */
1221                 lnet_ni_addref_locked(ni, 0);
1222                 cfs_list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1223                 if (ni->ni_cpts != NULL) {
1224                         cfs_list_add_tail(&ni->ni_cptlist,
1225                                           &the_lnet.ln_nis_cpt);
1226                         lnet_ni_addref_locked(ni, 0);
1227                 }
1228
1229                 lnet_net_unlock(LNET_LOCK_EX);
1230
1231                 if (lnd->lnd_type == LOLND) {
1232                         lnet_ni_addref(ni);
1233                         LASSERT (the_lnet.ln_loni == NULL);
1234                         the_lnet.ln_loni = ni;
1235                         continue;
1236                 }
1237
1238 #ifndef __KERNEL__
1239                 if (lnd->lnd_wait != NULL) {
1240                         if (the_lnet.ln_eq_waitni == NULL) {
1241                                 lnet_ni_addref(ni);
1242                                 the_lnet.ln_eq_waitni = ni;
1243                         }
1244                 } else {
1245 # ifndef HAVE_LIBPTHREAD
1246                         LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1247                                            "single-threaded runtime\n",
1248                                            libcfs_lnd2str(lnd_type));
1249                         goto failed;
1250 # endif
1251                 }
1252 #endif
1253                 if (ni->ni_peertxcredits == 0 ||
1254                     ni->ni_maxtxcredits == 0) {
1255                         LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1256                                            libcfs_lnd2str(lnd->lnd_type),
1257                                            ni->ni_peertxcredits == 0 ?
1258                                            "" : "per-peer ");
1259                         goto failed;
1260                 }
1261
1262                 cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1263                         tq->tq_credits_min =
1264                         tq->tq_credits_max =
1265                         tq->tq_credits = lnet_ni_tq_credits(ni);
1266                 }
1267
1268                 CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1269                        libcfs_nid2str(ni->ni_nid), ni->ni_peertxcredits,
1270                        lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1271                        ni->ni_peerrtrcredits, ni->ni_peertimeout);
1272
1273                 nicount++;
1274         }
1275
1276         if (the_lnet.ln_eq_waitni != NULL && nicount > 1) {
1277                 lnd_type = the_lnet.ln_eq_waitni->ni_lnd->lnd_type;
1278                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1279                                    "\n",
1280                                    libcfs_lnd2str(lnd_type));
1281                 goto failed;
1282         }
1283
1284         return 0;
1285
1286  failed:
1287         lnet_shutdown_lndnis();
1288
1289         while (!cfs_list_empty(&nilist)) {
1290                 ni = cfs_list_entry(nilist.next, lnet_ni_t, ni_list);
1291                 cfs_list_del(&ni->ni_list);
1292                 lnet_ni_free(ni);
1293         }
1294
1295         return -ENETDOWN;
1296 }
1297
1298 /**
1299  * Initialize LNet library.
1300  *
1301  * Only userspace program needs to call this function - it's automatically
1302  * called in the kernel at module loading time. Caller has to call LNetFini()
1303  * after a call to LNetInit(), if and only if the latter returned 0. It must
1304  * be called exactly once.
1305  *
1306  * \return 0 on success, and -ve on failures.
1307  */
1308 int
1309 LNetInit(void)
1310 {
1311         int     rc;
1312
1313         lnet_assert_wire_constants();
1314         LASSERT(!the_lnet.ln_init);
1315
1316         memset(&the_lnet, 0, sizeof(the_lnet));
1317
1318         /* refer to global cfs_cpt_table for now */
1319         the_lnet.ln_cpt_table   = cfs_cpt_table;
1320         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1321
1322         LASSERT(the_lnet.ln_cpt_number > 0);
1323         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1324                 /* we are under risk of consuming all lh_cookie */
1325                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
1326                        "please change setting of CPT-table and retry\n",
1327                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1328                 return -1;
1329         }
1330
1331         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1332                 the_lnet.ln_cpt_bits++;
1333
1334         rc = lnet_create_locks();
1335         if (rc != 0) {
1336                 CERROR("Can't create LNet global locks: %d\n", rc);
1337                 return -1;
1338         }
1339
1340         the_lnet.ln_refcount = 0;
1341         the_lnet.ln_init = 1;
1342         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1343         CFS_INIT_LIST_HEAD(&the_lnet.ln_lnds);
1344         CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1345         CFS_INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1346
1347 #ifdef __KERNEL__
1348         /* All LNDs apart from the LOLND are in separate modules.  They
1349          * register themselves when their module loads, and unregister
1350          * themselves when their module is unloaded. */
1351 #else
1352         /* Register LNDs
1353          * NB the order here determines default 'networks=' order */
1354 # ifdef CRAY_XT3
1355         LNET_REGISTER_ULND(the_ptllnd);
1356 # endif
1357 # ifdef HAVE_LIBPTHREAD
1358         LNET_REGISTER_ULND(the_tcplnd);
1359 # endif
1360 #endif
1361         lnet_register_lnd(&the_lolnd);
1362         return 0;
1363 }
1364
1365 /**
1366  * Finalize LNet library.
1367  *
1368  * Only userspace program needs to call this function. It can be called
1369  * at most once.
1370  *
1371  * \pre LNetInit() called with success.
1372  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1373  */
1374 void
1375 LNetFini(void)
1376 {
1377         LASSERT(the_lnet.ln_init);
1378         LASSERT(the_lnet.ln_refcount == 0);
1379
1380         while (!cfs_list_empty(&the_lnet.ln_lnds))
1381                 lnet_unregister_lnd(cfs_list_entry(the_lnet.ln_lnds.next,
1382                                                    lnd_t, lnd_list));
1383         lnet_destroy_locks();
1384
1385         the_lnet.ln_init = 0;
1386 }
1387
1388 /**
1389  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1390  *
1391  * Userspace program should call this after a successful call to LNetInit().
1392  * Users must call this function at least once before any other functions.
1393  * For each successful call there must be a corresponding call to
1394  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1395  * ignored.
1396  *
1397  * The PID used by LNet may be different from the one requested.
1398  * See LNetGetId().
1399  *
1400  * \param requested_pid PID requested by the caller.
1401  *
1402  * \return >= 0 on success, and < 0 error code on failures.
1403  */
1404 int
1405 LNetNIInit(lnet_pid_t requested_pid)
1406 {
1407         int         im_a_router = 0;
1408         int         rc;
1409
1410         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
1411
1412         LASSERT (the_lnet.ln_init);
1413         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1414
1415         if (the_lnet.ln_refcount > 0) {
1416                 rc = the_lnet.ln_refcount++;
1417                 goto out;
1418         }
1419
1420         lnet_get_tunables();
1421
1422         if (requested_pid == LNET_PID_ANY) {
1423                 /* Don't instantiate LNET just for me */
1424                 rc = -ENETDOWN;
1425                 goto failed0;
1426         }
1427
1428         rc = lnet_prepare(requested_pid);
1429         if (rc != 0)
1430                 goto failed0;
1431
1432         rc = lnet_startup_lndnis();
1433         if (rc != 0)
1434                 goto failed1;
1435
1436         rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1437         if (rc != 0)
1438                 goto failed2;
1439
1440         rc = lnet_check_routes();
1441         if (rc != 0)
1442                 goto failed2;
1443
1444         rc = lnet_rtrpools_alloc(im_a_router);
1445         if (rc != 0)
1446                 goto failed2;
1447
1448         rc = lnet_acceptor_start();
1449         if (rc != 0)
1450                 goto failed2;
1451
1452         the_lnet.ln_refcount = 1;
1453         /* Now I may use my own API functions... */
1454
1455         /* NB router checker needs the_lnet.ln_ping_info in
1456          * lnet_router_checker -> lnet_update_ni_status_locked */
1457         rc = lnet_ping_target_init();
1458         if (rc != 0)
1459                 goto failed3;
1460
1461         rc = lnet_router_checker_start();
1462         if (rc != 0)
1463                 goto failed4;
1464
1465         lnet_proc_init();
1466         goto out;
1467
1468  failed4:
1469         lnet_ping_target_fini();
1470  failed3:
1471         the_lnet.ln_refcount = 0;
1472         lnet_acceptor_stop();
1473  failed2:
1474         lnet_destroy_routes();
1475         lnet_shutdown_lndnis();
1476  failed1:
1477         lnet_unprepare();
1478  failed0:
1479         LASSERT (rc < 0);
1480  out:
1481         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1482         return rc;
1483 }
1484
1485 /**
1486  * Stop LNet interfaces, routing, and forwarding.
1487  *
1488  * Users must call this function once for each successful call to LNetNIInit().
1489  * Once the LNetNIFini() operation has been started, the results of pending
1490  * API operations are undefined.
1491  *
1492  * \return always 0 for current implementation.
1493  */
1494 int
1495 LNetNIFini()
1496 {
1497         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
1498
1499         LASSERT (the_lnet.ln_init);
1500         LASSERT (the_lnet.ln_refcount > 0);
1501
1502         if (the_lnet.ln_refcount != 1) {
1503                 the_lnet.ln_refcount--;
1504         } else {
1505                 LASSERT (!the_lnet.ln_niinit_self);
1506
1507                 lnet_proc_fini();
1508                 lnet_router_checker_stop();
1509                 lnet_ping_target_fini();
1510
1511                 /* Teardown fns that use my own API functions BEFORE here */
1512                 the_lnet.ln_refcount = 0;
1513
1514                 lnet_acceptor_stop();
1515                 lnet_destroy_routes();
1516                 lnet_shutdown_lndnis();
1517                 lnet_unprepare();
1518         }
1519
1520         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1521         return 0;
1522 }
1523
1524 /**
1525  * This is an ugly hack to export IOC_LIBCFS_DEBUG_PEER and
1526  * IOC_LIBCFS_PORTALS_COMPATIBILITY commands to users, by tweaking the LNet
1527  * internal ioctl handler.
1528  *
1529  * IOC_LIBCFS_PORTALS_COMPATIBILITY is now deprecated, don't use it.
1530  *
1531  * \param cmd IOC_LIBCFS_DEBUG_PEER to print debugging data about a peer.
1532  * The data will be printed to system console. Don't use it excessively.
1533  * \param arg A pointer to lnet_process_id_t, process ID of the peer.
1534  *
1535  * \return Always return 0 when called by users directly (i.e., not via ioctl).
1536  */
1537 int
1538 LNetCtl(unsigned int cmd, void *arg)
1539 {
1540         struct libcfs_ioctl_data *data = arg;
1541         lnet_process_id_t         id = {0};
1542         lnet_ni_t                *ni;
1543         int                       rc;
1544
1545         LASSERT (the_lnet.ln_init);
1546         LASSERT (the_lnet.ln_refcount > 0);
1547
1548         switch (cmd) {
1549         case IOC_LIBCFS_GET_NI:
1550                 rc = LNetGetId(data->ioc_count, &id);
1551                 data->ioc_nid = id.nid;
1552                 return rc;
1553
1554         case IOC_LIBCFS_FAIL_NID:
1555                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1556
1557         case IOC_LIBCFS_ADD_ROUTE:
1558                 rc = lnet_add_route(data->ioc_net, data->ioc_count,
1559                                     data->ioc_nid);
1560                 return (rc != 0) ? rc : lnet_check_routes();
1561
1562         case IOC_LIBCFS_DEL_ROUTE:
1563                 return lnet_del_route(data->ioc_net, data->ioc_nid);
1564
1565         case IOC_LIBCFS_GET_ROUTE:
1566                 return lnet_get_route(data->ioc_count,
1567                                       &data->ioc_net, &data->ioc_count,
1568                                       &data->ioc_nid, &data->ioc_flags);
1569         case IOC_LIBCFS_NOTIFY_ROUTER:
1570                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
1571                                    cfs_time_current() -
1572                                    cfs_time_seconds(cfs_time_current_sec() -
1573                                                     (time_t)data->ioc_u64[0]));
1574
1575         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
1576                 /* This can be removed once lustre stops calling it */
1577                 return 0;
1578
1579         case IOC_LIBCFS_LNET_DIST:
1580                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
1581                 if (rc < 0 && rc != -EHOSTUNREACH)
1582                         return rc;
1583
1584                 data->ioc_u32[0] = rc;
1585                 return 0;
1586
1587         case IOC_LIBCFS_TESTPROTOCOMPAT:
1588                 lnet_net_lock(LNET_LOCK_EX);
1589                 the_lnet.ln_testprotocompat = data->ioc_flags;
1590                 lnet_net_unlock(LNET_LOCK_EX);
1591                 return 0;
1592
1593         case IOC_LIBCFS_PING:
1594                 id.nid = data->ioc_nid;
1595                 id.pid = data->ioc_u32[0];
1596                 rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
1597                                (lnet_process_id_t *)data->ioc_pbuf1,
1598                                data->ioc_plen1/sizeof(lnet_process_id_t));
1599                 if (rc < 0)
1600                         return rc;
1601                 data->ioc_count = rc;
1602                 return 0;
1603
1604         case IOC_LIBCFS_DEBUG_PEER: {
1605                 /* CAVEAT EMPTOR: this one designed for calling directly; not
1606                  * via an ioctl */
1607                 id = *((lnet_process_id_t *) arg);
1608
1609                 lnet_debug_peer(id.nid);
1610
1611                 ni = lnet_net2ni(LNET_NIDNET(id.nid));
1612                 if (ni == NULL) {
1613                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(id));
1614                 } else {
1615                         if (ni->ni_lnd->lnd_ctl == NULL) {
1616                                 CDEBUG(D_WARNING, "No ctl for %s\n",
1617                                        libcfs_id2str(id));
1618                         } else {
1619                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1620                         }
1621
1622                         lnet_ni_decref(ni);
1623                 }
1624                 return 0;
1625         }
1626
1627         default:
1628                 ni = lnet_net2ni(data->ioc_net);
1629                 if (ni == NULL)
1630                         return -EINVAL;
1631
1632                 if (ni->ni_lnd->lnd_ctl == NULL)
1633                         rc = -EINVAL;
1634                 else
1635                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
1636
1637                 lnet_ni_decref(ni);
1638                 return rc;
1639         }
1640         /* not reached */
1641 }
1642
1643 /**
1644  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
1645  * all interfaces share a same PID, as requested by LNetNIInit().
1646  *
1647  * \param index Index of the interface to look up.
1648  * \param id On successful return, this location will hold the
1649  * lnet_process_id_t ID of the interface.
1650  *
1651  * \retval 0 If an interface exists at \a index.
1652  * \retval -ENOENT If no interface has been found.
1653  */
1654 int
1655 LNetGetId(unsigned int index, lnet_process_id_t *id)
1656 {
1657         struct lnet_ni  *ni;
1658         cfs_list_t      *tmp;
1659         int             cpt;
1660         int             rc = -ENOENT;
1661
1662         LASSERT(the_lnet.ln_init);
1663         LASSERT(the_lnet.ln_refcount > 0);
1664
1665         cpt = lnet_net_lock_current();
1666
1667         cfs_list_for_each(tmp, &the_lnet.ln_nis) {
1668                 if (index-- != 0)
1669                         continue;
1670
1671                 ni = cfs_list_entry(tmp, lnet_ni_t, ni_list);
1672
1673                 id->nid = ni->ni_nid;
1674                 id->pid = the_lnet.ln_pid;
1675                 rc = 0;
1676                 break;
1677         }
1678
1679         lnet_net_unlock(cpt);
1680         return rc;
1681 }
1682
1683 /**
1684  * Print a string representation of handle \a h into buffer \a str of
1685  * \a len bytes.
1686  */
1687 void
1688 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
1689 {
1690         snprintf(str, len, LPX64, h.cookie);
1691 }
1692
1693 static int
1694 lnet_create_ping_info(void)
1695 {
1696         int               i;
1697         int               n;
1698         int               rc;
1699         unsigned int      infosz;
1700         lnet_ni_t        *ni;
1701         lnet_process_id_t id;
1702         lnet_ping_info_t *pinfo;
1703
1704         for (n = 0; ; n++) {
1705                 rc = LNetGetId(n, &id);
1706                 if (rc == -ENOENT)
1707                         break;
1708
1709                 LASSERT (rc == 0);
1710         }
1711
1712         infosz = offsetof(lnet_ping_info_t, pi_ni[n]);
1713         LIBCFS_ALLOC(pinfo, infosz);
1714         if (pinfo == NULL) {
1715                 CERROR("Can't allocate ping info[%d]\n", n);
1716                 return -ENOMEM;
1717         }
1718
1719         pinfo->pi_nnis    = n;
1720         pinfo->pi_pid     = the_lnet.ln_pid;
1721         pinfo->pi_magic   = LNET_PROTO_PING_MAGIC;
1722         pinfo->pi_features = LNET_PING_FEAT_NI_STATUS;
1723
1724         for (i = 0; i < n; i++) {
1725                 lnet_ni_status_t *ns = &pinfo->pi_ni[i];
1726
1727                 rc = LNetGetId(i, &id);
1728                 LASSERT (rc == 0);
1729
1730                 ns->ns_nid    = id.nid;
1731                 ns->ns_status = LNET_NI_STATUS_UP;
1732
1733                 lnet_net_lock(0);
1734
1735                 ni = lnet_nid2ni_locked(id.nid, 0);
1736                 LASSERT(ni != NULL);
1737
1738                 lnet_ni_lock(ni);
1739                 LASSERT(ni->ni_status == NULL);
1740                 ni->ni_status = ns;
1741                 lnet_ni_unlock(ni);
1742
1743                 lnet_ni_decref_locked(ni, 0);
1744                 lnet_net_unlock(0);
1745         }
1746
1747         the_lnet.ln_ping_info = pinfo;
1748         return 0;
1749 }
1750
1751 static void
1752 lnet_destroy_ping_info(void)
1753 {
1754         struct lnet_ni  *ni;
1755
1756         lnet_net_lock(0);
1757
1758         cfs_list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
1759                 lnet_ni_lock(ni);
1760                 ni->ni_status = NULL;
1761                 lnet_ni_unlock(ni);
1762         }
1763
1764         lnet_net_unlock(0);
1765
1766         LIBCFS_FREE(the_lnet.ln_ping_info,
1767                     offsetof(lnet_ping_info_t,
1768                              pi_ni[the_lnet.ln_ping_info->pi_nnis]));
1769         the_lnet.ln_ping_info = NULL;
1770         return;
1771 }
1772
1773 int
1774 lnet_ping_target_init(void)
1775 {
1776         lnet_md_t         md = {0};
1777         lnet_handle_me_t  meh;
1778         lnet_process_id_t id;
1779         int               rc;
1780         int               rc2;
1781         int               infosz;
1782
1783         rc = lnet_create_ping_info();
1784         if (rc != 0)
1785                 return rc;
1786
1787         /* We can have a tiny EQ since we only need to see the unlink event on
1788          * teardown, which by definition is the last one! */
1789         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &the_lnet.ln_ping_target_eq);
1790         if (rc != 0) {
1791                 CERROR("Can't allocate ping EQ: %d\n", rc);
1792                 goto failed_0;
1793         }
1794
1795         memset(&id, 0, sizeof(lnet_process_id_t));
1796         id.nid = LNET_NID_ANY;
1797         id.pid = LNET_PID_ANY;
1798
1799         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1800                           LNET_PROTO_PING_MATCHBITS, 0,
1801                           LNET_UNLINK, LNET_INS_AFTER,
1802                           &meh);
1803         if (rc != 0) {
1804                 CERROR("Can't create ping ME: %d\n", rc);
1805                 goto failed_1;
1806         }
1807
1808         /* initialize md content */
1809         infosz = offsetof(lnet_ping_info_t,
1810                           pi_ni[the_lnet.ln_ping_info->pi_nnis]);
1811         md.start     = the_lnet.ln_ping_info;
1812         md.length    = infosz;
1813         md.threshold = LNET_MD_THRESH_INF;
1814         md.max_size  = 0;
1815         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1816                        LNET_MD_MANAGE_REMOTE;
1817         md.user_ptr  = NULL;
1818         md.eq_handle = the_lnet.ln_ping_target_eq;
1819
1820         rc = LNetMDAttach(meh, md,
1821                           LNET_RETAIN,
1822                           &the_lnet.ln_ping_target_md);
1823         if (rc != 0) {
1824                 CERROR("Can't attach ping MD: %d\n", rc);
1825                 goto failed_2;
1826         }
1827
1828         return 0;
1829
1830  failed_2:
1831         rc2 = LNetMEUnlink(meh);
1832         LASSERT (rc2 == 0);
1833  failed_1:
1834         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1835         LASSERT (rc2 == 0);
1836  failed_0:
1837         lnet_destroy_ping_info();
1838         return rc;
1839 }
1840
1841 void
1842 lnet_ping_target_fini(void)
1843 {
1844         lnet_event_t    event;
1845         int             rc;
1846         int             which;
1847         int             timeout_ms = 1000;
1848         cfs_sigset_t    blocked = cfs_block_allsigs();
1849
1850         LNetMDUnlink(the_lnet.ln_ping_target_md);
1851         /* NB md could be busy; this just starts the unlink */
1852
1853         for (;;) {
1854                 rc = LNetEQPoll(&the_lnet.ln_ping_target_eq, 1,
1855                                 timeout_ms, &event, &which);
1856
1857                 /* I expect overflow... */
1858                 LASSERT (rc >= 0 || rc == -EOVERFLOW);
1859
1860                 if (rc == 0) {
1861                         /* timed out: provide a diagnostic */
1862                         CWARN("Still waiting for ping MD to unlink\n");
1863                         timeout_ms *= 2;
1864                         continue;
1865                 }
1866
1867                 /* Got a valid event */
1868                 if (event.unlinked)
1869                         break;
1870         }
1871
1872         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1873         LASSERT (rc == 0);
1874         lnet_destroy_ping_info();
1875         cfs_restore_sigs(blocked);
1876 }
1877
1878 int
1879 lnet_ping (lnet_process_id_t id, int timeout_ms, lnet_process_id_t *ids, int n_ids)
1880 {
1881         lnet_handle_eq_t     eqh;
1882         lnet_handle_md_t     mdh;
1883         lnet_event_t         event;
1884         lnet_md_t            md = {0};
1885         int                  which;
1886         int                  unlinked = 0;
1887         int                  replied = 0;
1888         const int            a_long_time = 60000; /* mS */
1889         int                  infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
1890         lnet_ping_info_t    *info;
1891         lnet_process_id_t    tmpid;
1892         int                  i;
1893         int                  nob;
1894         int                  rc;
1895         int                  rc2;
1896         cfs_sigset_t         blocked;
1897
1898         if (n_ids <= 0 ||
1899             id.nid == LNET_NID_ANY ||
1900             timeout_ms > 500000 ||              /* arbitrary limit! */
1901             n_ids > 20)                         /* arbitrary limit! */
1902                 return -EINVAL;
1903
1904         if (id.pid == LNET_PID_ANY)
1905                 id.pid = LUSTRE_SRV_LNET_PID;
1906
1907         LIBCFS_ALLOC(info, infosz);
1908         if (info == NULL)
1909                 return -ENOMEM;
1910
1911         /* NB 2 events max (including any unlink event) */
1912         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
1913         if (rc != 0) {
1914                 CERROR("Can't allocate EQ: %d\n", rc);
1915                 goto out_0;
1916         }
1917
1918         /* initialize md content */
1919         md.start     = info;
1920         md.length    = infosz;
1921         md.threshold = 2; /*GET/REPLY*/
1922         md.max_size  = 0;
1923         md.options   = LNET_MD_TRUNCATE;
1924         md.user_ptr  = NULL;
1925         md.eq_handle = eqh;
1926
1927         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
1928         if (rc != 0) {
1929                 CERROR("Can't bind MD: %d\n", rc);
1930                 goto out_1;
1931         }
1932
1933         rc = LNetGet(LNET_NID_ANY, mdh, id,
1934                      LNET_RESERVED_PORTAL,
1935                      LNET_PROTO_PING_MATCHBITS, 0);
1936
1937         if (rc != 0) {
1938                 /* Don't CERROR; this could be deliberate! */
1939
1940                 rc2 = LNetMDUnlink(mdh);
1941                 LASSERT (rc2 == 0);
1942
1943                 /* NB must wait for the UNLINK event below... */
1944                 unlinked = 1;
1945                 timeout_ms = a_long_time;
1946         }
1947
1948         do {
1949                 /* MUST block for unlink to complete */
1950                 if (unlinked)
1951                         blocked = cfs_block_allsigs();
1952
1953                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
1954
1955                 if (unlinked)
1956                         cfs_restore_sigs(blocked);
1957
1958                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
1959                        (rc2 <= 0) ? -1 : event.type,
1960                        (rc2 <= 0) ? -1 : event.status,
1961                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
1962
1963                 LASSERT (rc2 != -EOVERFLOW);     /* can't miss anything */
1964
1965                 if (rc2 <= 0 || event.status != 0) {
1966                         /* timeout or error */
1967                         if (!replied && rc == 0)
1968                                 rc = (rc2 < 0) ? rc2 :
1969                                      (rc2 == 0) ? -ETIMEDOUT :
1970                                      event.status;
1971
1972                         if (!unlinked) {
1973                                 /* Ensure completion in finite time... */
1974                                 LNetMDUnlink(mdh);
1975                                 /* No assertion (racing with network) */
1976                                 unlinked = 1;
1977                                 timeout_ms = a_long_time;
1978                         } else if (rc2 == 0) {
1979                                 /* timed out waiting for unlink */
1980                                 CWARN("ping %s: late network completion\n",
1981                                       libcfs_id2str(id));
1982                         }
1983                 } else if (event.type == LNET_EVENT_REPLY) {
1984                         replied = 1;
1985                         rc = event.mlength;
1986                 }
1987
1988         } while (rc2 <= 0 || !event.unlinked);
1989
1990         if (!replied) {
1991                 if (rc >= 0)
1992                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
1993                               libcfs_id2str(id));
1994                 rc = -EIO;
1995                 goto out_1;
1996         }
1997
1998         nob = rc;
1999         LASSERT (nob >= 0 && nob <= infosz);
2000
2001         rc = -EPROTO;                           /* if I can't parse... */
2002
2003         if (nob < 8) {
2004                 /* can't check magic/version */
2005                 CERROR("%s: ping info too short %d\n",
2006                        libcfs_id2str(id), nob);
2007                 goto out_1;
2008         }
2009
2010         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
2011                 lnet_swap_pinginfo(info);
2012         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
2013                 CERROR("%s: Unexpected magic %08x\n",
2014                        libcfs_id2str(id), info->pi_magic);
2015                 goto out_1;
2016         }
2017
2018         if ((info->pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
2019                 CERROR("%s: ping w/o NI status: 0x%x\n",
2020                        libcfs_id2str(id), info->pi_features);
2021                 goto out_1;
2022         }
2023
2024         if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
2025                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
2026                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
2027                 goto out_1;
2028         }
2029
2030         if (info->pi_nnis < n_ids)
2031                 n_ids = info->pi_nnis;
2032
2033         if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
2034                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
2035                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
2036                 goto out_1;
2037         }
2038
2039         rc = -EFAULT;                           /* If I SEGV... */
2040
2041         for (i = 0; i < n_ids; i++) {
2042                 tmpid.pid = info->pi_pid;
2043                 tmpid.nid = info->pi_ni[i].ns_nid;
2044 #ifdef __KERNEL__
2045                 if (cfs_copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
2046                         goto out_1;
2047 #else
2048                 ids[i] = tmpid;
2049 #endif
2050         }
2051         rc = info->pi_nnis;
2052
2053  out_1:
2054         rc2 = LNetEQFree(eqh);
2055         if (rc2 != 0)
2056                 CERROR("rc2 %d\n", rc2);
2057         LASSERT (rc2 == 0);
2058
2059  out_0:
2060         LIBCFS_FREE(info, infosz);
2061         return rc;
2062 }