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, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38 #include <lnet/lib-lnet.h>
  39 #include <lnet/lib-dlc.h>
  40 #ifdef __KERNEL__
  41 #include <linux/log2.h>
  42 #endif
  43
  44 #ifdef __KERNEL__
  45 #define D_LNI D_CONSOLE
  46 #else
  47 #define D_LNI D_CONFIG
  48 #endif
  49
  50 lnet_t      the_lnet;                           /* THE state of the network */
  51 EXPORT_SYMBOL(the_lnet);
  52
  53 #ifdef __KERNEL__
  54
  55 static char *ip2nets = "";
  56 CFS_MODULE_PARM(ip2nets, "s", charp, 0444,
  57                 "LNET network <- IP table");
  58
  59 static char *networks = "";
  60 CFS_MODULE_PARM(networks, "s", charp, 0444,
  61                 "local networks");
  62
  63 static char *routes = "";
  64 CFS_MODULE_PARM(routes, "s", charp, 0444,
  65                 "routes to non-local networks");
  66
  67 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  68 CFS_MODULE_PARM(rnet_htable_size, "i", int, 0444,
  69                 "size of remote network hash table");
  70
  71 static void lnet_ping_target_fini(void);
  72 static int lnet_ping(lnet_process_id_t id, int timeout_ms,
  73                      lnet_process_id_t *ids, int n_ids);
  74
  75 static char *
  76 lnet_get_routes(void)
  77 {
  78         return routes;
  79 }
  80
  81 static char *
  82 lnet_get_networks(void)
  83 {
  84         char   *nets;
  85         int     rc;
  86
  87         if (*networks != 0 && *ip2nets != 0) {
  88                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  89                                    "'ip2nets' but not both at once\n");
  90                 return NULL;
  91         }
  92
  93         if (*ip2nets != 0) {
  94                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  95                 return (rc == 0) ? nets : NULL;
  96         }
  97
  98         if (*networks != 0)
  99                 return networks;
 100
 101         return "tcp";
 102 }
 103
 104 static void
 105 lnet_init_locks(void)
 106 {
 107         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 108         init_waitqueue_head(&the_lnet.ln_eq_waitq);
 109         mutex_init(&the_lnet.ln_lnd_mutex);
 110         mutex_init(&the_lnet.ln_api_mutex);
 111 }
 112
 113 static void
 114 lnet_fini_locks(void)
 115 {
 116 }
 117
 118 #else
 119
 120 static char *
 121 lnet_get_routes(void)
 122 {
 123         char *str = getenv("LNET_ROUTES");
 124
 125         return (str == NULL) ? "" : str;
 126 }
 127
 128 static char *
 129 lnet_get_networks (void)
 130 {
 131         static char       default_networks[256];
 132         char             *networks = getenv("LNET_NETWORKS");
 133         char             *str;
 134         char             *sep;
 135         int               len;
 136         int               nob;
 137         struct list_head *tmp;
 138
 139         if (networks != NULL)
 140                 return networks;
 141
 142         /* In userland, the default 'networks=' is the list of known net types */
 143         len = sizeof(default_networks);
 144         str = default_networks;
 145         *str = 0;
 146         sep = "";
 147
 148         list_for_each(tmp, &the_lnet.ln_lnds) {
 149                 lnd_t *lnd = list_entry(tmp, lnd_t, lnd_list);
 150
 151                 nob = snprintf(str, len, "%s%s", sep,
 152                                libcfs_lnd2str(lnd->lnd_type));
 153                 if (nob >= len) {
 154                         /* overflowed the string; leave it where it was */
 155                         *str = 0;
 156                         break;
 157                 }
 158                 len -= nob;
 159                 str += nob;
 160                 sep = ",";
 161         }
 162
 163         return default_networks;
 164 }
 165
 166 # ifndef HAVE_LIBPTHREAD
 167
 168 static void lnet_init_locks(void)
 169 {
 170         the_lnet.ln_eq_wait_lock = 0;
 171         the_lnet.ln_lnd_mutex = 0;
 172         the_lnet.ln_api_mutex = 0;
 173 }
 174
 175 static void lnet_fini_locks(void)
 176 {
 177         LASSERT(the_lnet.ln_api_mutex == 0);
 178         LASSERT(the_lnet.ln_lnd_mutex == 0);
 179         LASSERT(the_lnet.ln_eq_wait_lock == 0);
 180 }
 181
 182 # else
 183
 184 static void lnet_init_locks(void)
 185 {
 186         pthread_cond_init(&the_lnet.ln_eq_cond, NULL);
 187         pthread_mutex_init(&the_lnet.ln_eq_wait_lock, NULL);
 188         pthread_mutex_init(&the_lnet.ln_lnd_mutex, NULL);
 189         pthread_mutex_init(&the_lnet.ln_api_mutex, NULL);
 190 }
 191
 192 static void lnet_fini_locks(void)
 193 {
 194         pthread_mutex_destroy(&the_lnet.ln_api_mutex);
 195         pthread_mutex_destroy(&the_lnet.ln_lnd_mutex);
 196         pthread_mutex_destroy(&the_lnet.ln_eq_wait_lock);
 197         pthread_cond_destroy(&the_lnet.ln_eq_cond);
 198 }
 199
 200 # endif
 201 #endif
 202
 203 static int
 204 lnet_create_remote_nets_table(void)
 205 {
 206         int               i;
 207         struct list_head *hash;
 208
 209         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 210         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 211         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 212         if (hash == NULL) {
 213                 CERROR("Failed to create remote nets hash table\n");
 214                 return -ENOMEM;
 215         }
 216
 217         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 218                 INIT_LIST_HEAD(&hash[i]);
 219         the_lnet.ln_remote_nets_hash = hash;
 220         return 0;
 221 }
 222
 223 static void
 224 lnet_destroy_remote_nets_table(void)
 225 {
 226         int i;
 227
 228         if (the_lnet.ln_remote_nets_hash == NULL)
 229                 return;
 230
 231         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 232                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 233
 234         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 235                     LNET_REMOTE_NETS_HASH_SIZE *
 236                     sizeof(the_lnet.ln_remote_nets_hash[0]));
 237         the_lnet.ln_remote_nets_hash = NULL;
 238 }
 239
 240 static void
 241 lnet_destroy_locks(void)
 242 {
 243         if (the_lnet.ln_res_lock != NULL) {
 244                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 245                 the_lnet.ln_res_lock = NULL;
 246         }
 247
 248         if (the_lnet.ln_net_lock != NULL) {
 249                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 250                 the_lnet.ln_net_lock = NULL;
 251         }
 252
 253         lnet_fini_locks();
 254 }
 255
 256 static int
 257 lnet_create_locks(void)
 258 {
 259         lnet_init_locks();
 260
 261         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 262         if (the_lnet.ln_res_lock == NULL)
 263                 goto failed;
 264
 265         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 266         if (the_lnet.ln_net_lock == NULL)
 267                 goto failed;
 268
 269         return 0;
 270
 271  failed:
 272         lnet_destroy_locks();
 273         return -ENOMEM;
 274 }
 275
 276 static void lnet_assert_wire_constants(void)
 277 {
 278         /* Wire protocol assertions generated by 'wirecheck'
 279          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 280          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 281          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 282
 283         /* Constants... */
 284         CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 285         CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
 286         CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
 287         CLASSERT (LNET_MSG_ACK == 0);
 288         CLASSERT (LNET_MSG_PUT == 1);
 289         CLASSERT (LNET_MSG_GET == 2);
 290         CLASSERT (LNET_MSG_REPLY == 3);
 291         CLASSERT (LNET_MSG_HELLO == 4);
 292
 293         /* Checks for struct ptl_handle_wire_t */
 294         CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
 295         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 296         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 297         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 298         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 299
 300         /* Checks for struct lnet_magicversion_t */
 301         CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
 302         CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
 303         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 304         CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
 305         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 306         CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 307         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 308
 309         /* Checks for struct lnet_hdr_t */
 310         CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
 311         CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 312         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 313         CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
 314         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 315         CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 316         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 317         CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
 318         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 319         CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
 320         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 321         CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
 322         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 323         CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
 324         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 325
 326         /* Ack */
 327         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 328         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 329         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 330         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 331         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 332         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 333
 334         /* Put */
 335         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 336         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 337         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 338         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 339         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 340         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 341         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 342         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 343         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 344         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 345
 346         /* Get */
 347         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 348         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 349         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 350         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 351         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 352         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 353         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 354         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 355         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 356         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 357
 358         /* Reply */
 359         CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 360         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 361
 362         /* Hello */
 363         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 364         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 365         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 366         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 367 }
 368
 369 static lnd_t *
 370 lnet_find_lnd_by_type (int type)
 371 {
 372         lnd_t            *lnd;
 373         struct list_head *tmp;
 374
 375         /* holding lnd mutex */
 376         list_for_each(tmp, &the_lnet.ln_lnds) {
 377                 lnd = list_entry(tmp, lnd_t, lnd_list);
 378
 379                 if ((int)lnd->lnd_type == type)
 380                         return lnd;
 381         }
 382         return NULL;
 383 }
 384
 385 void
 386 lnet_register_lnd (lnd_t *lnd)
 387 {
 388         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
 389
 390         LASSERT(the_lnet.ln_init);
 391         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 392         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 393
 394         list_add_tail(&lnd->lnd_list, &the_lnet.ln_lnds);
 395         lnd->lnd_refcount = 0;
 396
 397         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 398
 399         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
 400 }
 401 EXPORT_SYMBOL(lnet_register_lnd);
 402
 403 void
 404 lnet_unregister_lnd (lnd_t *lnd)
 405 {
 406         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
 407
 408         LASSERT(the_lnet.ln_init);
 409         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 410         LASSERT(lnd->lnd_refcount == 0);
 411
 412         list_del(&lnd->lnd_list);
 413         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 414
 415         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
 416 }
 417 EXPORT_SYMBOL(lnet_unregister_lnd);
 418
 419 void
 420 lnet_counters_get(lnet_counters_t *counters)
 421 {
 422         lnet_counters_t *ctr;
 423         int             i;
 424
 425         memset(counters, 0, sizeof(*counters));
 426
 427         lnet_net_lock(LNET_LOCK_EX);
 428
 429         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 430                 counters->msgs_max     += ctr->msgs_max;
 431                 counters->msgs_alloc   += ctr->msgs_alloc;
 432                 counters->errors       += ctr->errors;
 433                 counters->send_count   += ctr->send_count;
 434                 counters->recv_count   += ctr->recv_count;
 435                 counters->route_count  += ctr->route_count;
 436                 counters->drop_count   += ctr->drop_count;
 437                 counters->send_length  += ctr->send_length;
 438                 counters->recv_length  += ctr->recv_length;
 439                 counters->route_length += ctr->route_length;
 440                 counters->drop_length  += ctr->drop_length;
 441
 442         }
 443         lnet_net_unlock(LNET_LOCK_EX);
 444 }
 445 EXPORT_SYMBOL(lnet_counters_get);
 446
 447 void
 448 lnet_counters_reset(void)
 449 {
 450         lnet_counters_t *counters;
 451         int             i;
 452
 453         lnet_net_lock(LNET_LOCK_EX);
 454
 455         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 456                 memset(counters, 0, sizeof(lnet_counters_t));
 457
 458         lnet_net_unlock(LNET_LOCK_EX);
 459 }
 460 EXPORT_SYMBOL(lnet_counters_reset);
 461
 462 #ifdef LNET_USE_LIB_FREELIST
 463
 464 int
 465 lnet_freelist_init(lnet_freelist_t *fl, int n, int size)
 466 {
 467         char *space;
 468
 469         LASSERT (n > 0);
 470
 471         size += offsetof (lnet_freeobj_t, fo_contents);
 472
 473         LIBCFS_ALLOC(space, n * size);
 474         if (space == NULL)
 475                 return (-ENOMEM);
 476
 477         INIT_LIST_HEAD(&fl->fl_list);
 478         fl->fl_objs = space;
 479         fl->fl_nobjs = n;
 480         fl->fl_objsize = size;
 481
 482         do {
 483                 list_add((struct list_head *)space, &fl->fl_list);
 484                 space += size;
 485         } while (--n != 0);
 486
 487         return 0;
 488 }
 489
 490 void
 491 lnet_freelist_fini(lnet_freelist_t *fl)
 492 {
 493         struct list_head *el;
 494         int               count;
 495
 496         if (fl->fl_nobjs == 0)
 497                 return;
 498
 499         count = 0;
 500         for (el = fl->fl_list.next; el != &fl->fl_list; el = el->next)
 501                 count++;
 502
 503         LASSERT (count == fl->fl_nobjs);
 504
 505         LIBCFS_FREE(fl->fl_objs, fl->fl_nobjs * fl->fl_objsize);
 506         memset (fl, 0, sizeof (*fl));
 507 }
 508
 509 #endif /* LNET_USE_LIB_FREELIST */
 510
 511 static __u64 lnet_create_interface_cookie(void)
 512 {
 513         /* NB the interface cookie in wire handles guards against delayed
 514          * replies and ACKs appearing valid after reboot. Initialisation time,
 515          * even if it's only implemented to millisecond resolution is probably
 516          * easily good enough. */
 517         struct timeval tv;
 518         __u64          cookie;
 519 #ifndef __KERNEL__
 520         int            rc = gettimeofday (&tv, NULL);
 521         LASSERT (rc == 0);
 522 #else
 523         do_gettimeofday(&tv);
 524 #endif
 525         cookie = tv.tv_sec;
 526         cookie *= 1000000;
 527         cookie += tv.tv_usec;
 528         return cookie;
 529 }
 530
 531 static char *
 532 lnet_res_type2str(int type)
 533 {
 534         switch (type) {
 535         default:
 536                 LBUG();
 537         case LNET_COOKIE_TYPE_MD:
 538                 return "MD";
 539         case LNET_COOKIE_TYPE_ME:
 540                 return "ME";
 541         case LNET_COOKIE_TYPE_EQ:
 542                 return "EQ";
 543         }
 544 }
 545
 546 static void
 547 lnet_res_container_cleanup(struct lnet_res_container *rec)
 548 {
 549         int     count = 0;
 550
 551         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 552                 return;
 553
 554         while (!list_empty(&rec->rec_active)) {
 555                 struct list_head *e = rec->rec_active.next;
 556
 557                 list_del_init(e);
 558                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 559                         lnet_eq_free(list_entry(e, lnet_eq_t, eq_list));
 560
 561                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 562                         lnet_md_free(list_entry(e, lnet_libmd_t, md_list));
 563
 564                 } else { /* NB: Active MEs should be attached on portals */
 565                         LBUG();
 566                 }
 567                 count++;
 568         }
 569
 570         if (count > 0) {
 571                 /* Found alive MD/ME/EQ, user really should unlink/free
 572                  * all of them before finalize LNet, but if someone didn't,
 573                  * we have to recycle garbage for him */
 574                 CERROR("%d active elements on exit of %s container\n",
 575                        count, lnet_res_type2str(rec->rec_type));
 576         }
 577
 578 #ifdef LNET_USE_LIB_FREELIST
 579         lnet_freelist_fini(&rec->rec_freelist);
 580 #endif
 581         if (rec->rec_lh_hash != NULL) {
 582                 LIBCFS_FREE(rec->rec_lh_hash,
 583                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 584                 rec->rec_lh_hash = NULL;
 585         }
 586
 587         rec->rec_type = 0; /* mark it as finalized */
 588 }
 589
 590 static int
 591 lnet_res_container_setup(struct lnet_res_container *rec,
 592                          int cpt, int type, int objnum, int objsz)
 593 {
 594         int     rc = 0;
 595         int     i;
 596
 597         LASSERT(rec->rec_type == 0);
 598
 599         rec->rec_type = type;
 600         INIT_LIST_HEAD(&rec->rec_active);
 601
 602 #ifdef LNET_USE_LIB_FREELIST
 603         memset(&rec->rec_freelist, 0, sizeof(rec->rec_freelist));
 604         rc = lnet_freelist_init(&rec->rec_freelist, objnum, objsz);
 605         if (rc != 0)
 606                 goto out;
 607 #endif
 608         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 609
 610         /* Arbitrary choice of hash table size */
 611         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 612                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 613         if (rec->rec_lh_hash == NULL) {
 614                 rc = -ENOMEM;
 615                 goto out;
 616         }
 617
 618         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 619                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 620
 621         return 0;
 622
 623 out:
 624         CERROR("Failed to setup %s resource container\n",
 625                lnet_res_type2str(type));
 626         lnet_res_container_cleanup(rec);
 627         return rc;
 628 }
 629
 630 static void
 631 lnet_res_containers_destroy(struct lnet_res_container **recs)
 632 {
 633         struct lnet_res_container       *rec;
 634         int                             i;
 635
 636         cfs_percpt_for_each(rec, i, recs)
 637                 lnet_res_container_cleanup(rec);
 638
 639         cfs_percpt_free(recs);
 640 }
 641
 642 static struct lnet_res_container **
 643 lnet_res_containers_create(int type, int objnum, int objsz)
 644 {
 645         struct lnet_res_container       **recs;
 646         struct lnet_res_container       *rec;
 647         int                             rc;
 648         int                             i;
 649
 650         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 651         if (recs == NULL) {
 652                 CERROR("Failed to allocate %s resource containers\n",
 653                        lnet_res_type2str(type));
 654                 return NULL;
 655         }
 656
 657         cfs_percpt_for_each(rec, i, recs) {
 658                 rc = lnet_res_container_setup(rec, i, type, objnum, objsz);
 659                 if (rc != 0) {
 660                         lnet_res_containers_destroy(recs);
 661                         return NULL;
 662                 }
 663         }
 664
 665         return recs;
 666 }
 667
 668 lnet_libhandle_t *
 669 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 670 {
 671         /* ALWAYS called with lnet_res_lock held */
 672         struct list_head        *head;
 673         lnet_libhandle_t        *lh;
 674         unsigned int            hash;
 675
 676         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
 677                 return NULL;
 678
 679         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 680         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 681
 682         list_for_each_entry(lh, head, lh_hash_chain) {
 683                 if (lh->lh_cookie == cookie)
 684                         return lh;
 685         }
 686
 687         return NULL;
 688 }
 689
 690 void
 691 lnet_res_lh_initialize(struct lnet_res_container *rec, lnet_libhandle_t *lh)
 692 {
 693         /* ALWAYS called with lnet_res_lock held */
 694         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 695         unsigned int    hash;
 696
 697         lh->lh_cookie = rec->rec_lh_cookie;
 698         rec->rec_lh_cookie += 1 << ibits;
 699
 700         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 701
 702         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 703 }
 704
 705 #ifndef __KERNEL__
 706 /**
 707  * Reserved API - do not use.
 708  * Temporary workaround to allow uOSS and test programs force server
 709  * mode in userspace. See comments near ln_server_mode_flag in
 710  * lnet/lib-types.h */
 711
 712 void
 713 lnet_server_mode() {
 714         the_lnet.ln_server_mode_flag = 1;
 715 }
 716 #endif
 717
 718 static int lnet_unprepare(void);
 719
 720 static int
 721 lnet_prepare(lnet_pid_t requested_pid)
 722 {
 723         /* Prepare to bring up the network */
 724         struct lnet_res_container **recs;
 725         int                       rc = 0;
 726
 727         if (requested_pid == LNET_PID_ANY) {
 728                 /* Don't instantiate LNET just for me */
 729                 return -ENETDOWN;
 730         }
 731
 732         LASSERT (the_lnet.ln_refcount == 0);
 733
 734         the_lnet.ln_routing = 0;
 735
 736 #ifdef __KERNEL__
 737         LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 738         the_lnet.ln_pid = requested_pid;
 739 #else
 740         if (the_lnet.ln_server_mode_flag) {/* server case (uOSS) */
 741                 LASSERT ((requested_pid & LNET_PID_USERFLAG) == 0);
 742
 743                 if (current_uid() != 0) /* Only root can run user-space server */
 744                         return -EPERM;
 745                 the_lnet.ln_pid = requested_pid;
 746
 747         } else {/* client case (liblustre) */
 748
 749                 /* My PID must be unique on this node and flag I'm userspace */
 750                 the_lnet.ln_pid = getpid() | LNET_PID_USERFLAG;
 751         }
 752 #endif
 753
 754         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
 755         INIT_LIST_HEAD(&the_lnet.ln_nis);
 756         INIT_LIST_HEAD(&the_lnet.ln_nis_cpt);
 757         INIT_LIST_HEAD(&the_lnet.ln_nis_zombie);
 758         INIT_LIST_HEAD(&the_lnet.ln_routers);
 759         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
 760         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
 761
 762         rc = lnet_create_remote_nets_table();
 763         if (rc != 0)
 764                 goto failed;
 765
 766         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 767
 768         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
 769                                                 sizeof(lnet_counters_t));
 770         if (the_lnet.ln_counters == NULL) {
 771                 CERROR("Failed to allocate counters for LNet\n");
 772                 rc = -ENOMEM;
 773                 goto failed;
 774         }
 775
 776         rc = lnet_peer_tables_create();
 777         if (rc != 0)
 778                 goto failed;
 779
 780         rc = lnet_msg_containers_create();
 781         if (rc != 0)
 782                 goto failed;
 783
 784         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 785                                       LNET_COOKIE_TYPE_EQ, LNET_FL_MAX_EQS,
 786                                       sizeof(lnet_eq_t));
 787         if (rc != 0)
 788                 goto failed;
 789
 790         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME, LNET_FL_MAX_MES,
 791                                           sizeof(lnet_me_t));
 792         if (recs == NULL)
 793                 goto failed;
 794
 795         the_lnet.ln_me_containers = recs;
 796
 797         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD, LNET_FL_MAX_MDS,
 798                                           sizeof(lnet_libmd_t));
 799         if (recs == NULL)
 800                 goto failed;
 801
 802         the_lnet.ln_md_containers = recs;
 803
 804         rc = lnet_portals_create();
 805         if (rc != 0) {
 806                 CERROR("Failed to create portals for LNet: %d\n", rc);
 807                 goto failed;
 808         }
 809
 810         return 0;
 811
 812  failed:
 813         lnet_unprepare();
 814         return rc;
 815 }
 816
 817 static int
 818 lnet_unprepare (void)
 819 {
 820         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 821          * have shut down already, so it is safe to unlink and free all
 822          * descriptors, even those that appear committed to a network op (eg MD
 823          * with non-zero pending count) */
 824
 825         lnet_fail_nid(LNET_NID_ANY, 0);
 826
 827         LASSERT(the_lnet.ln_refcount == 0);
 828         LASSERT(list_empty(&the_lnet.ln_test_peers));
 829         LASSERT(list_empty(&the_lnet.ln_nis));
 830         LASSERT(list_empty(&the_lnet.ln_nis_cpt));
 831         LASSERT(list_empty(&the_lnet.ln_nis_zombie));
 832
 833         lnet_portals_destroy();
 834
 835         if (the_lnet.ln_md_containers != NULL) {
 836                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 837                 the_lnet.ln_md_containers = NULL;
 838         }
 839
 840         if (the_lnet.ln_me_containers != NULL) {
 841                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 842                 the_lnet.ln_me_containers = NULL;
 843         }
 844
 845         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 846
 847         lnet_msg_containers_destroy();
 848         lnet_peer_tables_destroy();
 849         lnet_rtrpools_free(0);
 850
 851         if (the_lnet.ln_counters != NULL) {
 852                 cfs_percpt_free(the_lnet.ln_counters);
 853                 the_lnet.ln_counters = NULL;
 854         }
 855         lnet_destroy_remote_nets_table();
 856
 857         return 0;
 858 }
 859
 860 lnet_ni_t  *
 861 lnet_net2ni_locked(__u32 net, int cpt)
 862 {
 863         struct list_head *tmp;
 864         lnet_ni_t        *ni;
 865
 866         LASSERT(cpt != LNET_LOCK_EX);
 867
 868         list_for_each(tmp, &the_lnet.ln_nis) {
 869                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 870
 871                 if (LNET_NIDNET(ni->ni_nid) == net) {
 872                         lnet_ni_addref_locked(ni, cpt);
 873                         return ni;
 874                 }
 875         }
 876
 877         return NULL;
 878 }
 879
 880 lnet_ni_t *
 881 lnet_net2ni(__u32 net)
 882 {
 883         lnet_ni_t *ni;
 884
 885         lnet_net_lock(0);
 886         ni = lnet_net2ni_locked(net, 0);
 887         lnet_net_unlock(0);
 888
 889         return ni;
 890 }
 891 EXPORT_SYMBOL(lnet_net2ni);
 892
 893 static unsigned int
 894 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 895 {
 896         __u64           key = nid;
 897         unsigned int    val;
 898
 899         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 900
 901         if (number == 1)
 902                 return 0;
 903
 904         val = hash_long(key, LNET_CPT_BITS);
 905         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 906         if (val < number)
 907                 return val;
 908
 909         return (unsigned int)(key + val + (val >> 1)) % number;
 910 }
 911
 912 int
 913 lnet_cpt_of_nid_locked(lnet_nid_t nid)
 914 {
 915         struct lnet_ni *ni;
 916
 917         /* must called with hold of lnet_net_lock */
 918         if (LNET_CPT_NUMBER == 1)
 919                 return 0; /* the only one */
 920
 921         /* take lnet_net_lock(any) would be OK */
 922         if (!list_empty(&the_lnet.ln_nis_cpt)) {
 923                 list_for_each_entry(ni, &the_lnet.ln_nis_cpt, ni_cptlist) {
 924                         if (LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid))
 925                                 continue;
 926
 927                         LASSERT(ni->ni_cpts != NULL);
 928                         return ni->ni_cpts[lnet_nid_cpt_hash
 929                                            (nid, ni->ni_ncpts)];
 930                 }
 931         }
 932
 933         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 934 }
 935
 936 int
 937 lnet_cpt_of_nid(lnet_nid_t nid)
 938 {
 939         int     cpt;
 940         int     cpt2;
 941
 942         if (LNET_CPT_NUMBER == 1)
 943                 return 0; /* the only one */
 944
 945         if (list_empty(&the_lnet.ln_nis_cpt))
 946                 return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 947
 948         cpt = lnet_net_lock_current();
 949         cpt2 = lnet_cpt_of_nid_locked(nid);
 950         lnet_net_unlock(cpt);
 951
 952         return cpt2;
 953 }
 954 EXPORT_SYMBOL(lnet_cpt_of_nid);
 955
 956 int
 957 lnet_islocalnet(__u32 net)
 958 {
 959         struct lnet_ni  *ni;
 960         int             cpt;
 961
 962         cpt = lnet_net_lock_current();
 963
 964         ni = lnet_net2ni_locked(net, cpt);
 965         if (ni != NULL)
 966                 lnet_ni_decref_locked(ni, cpt);
 967
 968         lnet_net_unlock(cpt);
 969
 970         return ni != NULL;
 971 }
 972
 973 lnet_ni_t  *
 974 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 975 {
 976         struct lnet_ni   *ni;
 977         struct list_head *tmp;
 978
 979         LASSERT(cpt != LNET_LOCK_EX);
 980
 981         list_for_each(tmp, &the_lnet.ln_nis) {
 982                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 983
 984                 if (ni->ni_nid == nid) {
 985                         lnet_ni_addref_locked(ni, cpt);
 986                         return ni;
 987                 }
 988         }
 989
 990         return NULL;
 991 }
 992
 993 int
 994 lnet_islocalnid(lnet_nid_t nid)
 995 {
 996         struct lnet_ni  *ni;
 997         int             cpt;
 998
 999         cpt = lnet_net_lock_current();
1000         ni = lnet_nid2ni_locked(nid, cpt);
1001         if (ni != NULL)
1002                 lnet_ni_decref_locked(ni, cpt);
1003         lnet_net_unlock(cpt);
1004
1005         return ni != NULL;
1006 }
1007
1008 int
1009 lnet_count_acceptor_nis (void)
1010 {
1011         /* Return the # of NIs that need the acceptor. */
1012         int              count = 0;
1013 #if defined(__KERNEL__) || defined(HAVE_LIBPTHREAD)
1014         struct list_head *tmp;
1015         struct lnet_ni   *ni;
1016         int              cpt;
1017
1018         cpt = lnet_net_lock_current();
1019         list_for_each(tmp, &the_lnet.ln_nis) {
1020                 ni = list_entry(tmp, lnet_ni_t, ni_list);
1021
1022                 if (ni->ni_lnd->lnd_accept != NULL)
1023                         count++;
1024         }
1025
1026         lnet_net_unlock(cpt);
1027
1028 #endif /* defined(__KERNEL__) || defined(HAVE_LIBPTHREAD) */
1029         return count;
1030 }
1031
1032 static lnet_ping_info_t *
1033 lnet_ping_info_create(int num_ni)
1034 {
1035         lnet_ping_info_t *ping_info;
1036         unsigned int     infosz;
1037
1038         infosz = offsetof(lnet_ping_info_t, pi_ni[num_ni]);
1039         LIBCFS_ALLOC(ping_info, infosz);
1040         if (ping_info == NULL) {
1041                 CERROR("Can't allocate ping info[%d]\n", num_ni);
1042                 return NULL;
1043         }
1044
1045         ping_info->pi_nnis = num_ni;
1046         ping_info->pi_pid = the_lnet.ln_pid;
1047         ping_info->pi_magic = LNET_PROTO_PING_MAGIC;
1048         ping_info->pi_features = LNET_PING_FEAT_NI_STATUS;
1049
1050         return ping_info;
1051 }
1052
1053 static inline int
1054 lnet_get_ni_count(void)
1055 {
1056         struct lnet_ni *ni;
1057         int            count = 0;
1058
1059         lnet_net_lock(0);
1060
1061         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list)
1062                 count++;
1063
1064         lnet_net_unlock(0);
1065
1066         return count;
1067 }
1068
1069 static inline void
1070 lnet_ping_info_free(lnet_ping_info_t *pinfo)
1071 {
1072         LIBCFS_FREE(pinfo,
1073                     offsetof(lnet_ping_info_t,
1074                              pi_ni[pinfo->pi_nnis]));
1075 }
1076
1077 static void
1078 lnet_ping_info_destroy(void)
1079 {
1080         struct lnet_ni  *ni;
1081
1082         lnet_net_lock(LNET_LOCK_EX);
1083
1084         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
1085                 lnet_ni_lock(ni);
1086                 ni->ni_status = NULL;
1087                 lnet_ni_unlock(ni);
1088         }
1089
1090         lnet_ping_info_free(the_lnet.ln_ping_info);
1091         the_lnet.ln_ping_info = NULL;
1092
1093         lnet_net_unlock(LNET_LOCK_EX);
1094 }
1095
1096 static void
1097 lnet_ping_event_handler(lnet_event_t *event)
1098 {
1099         lnet_ping_info_t *pinfo = event->md.user_ptr;
1100
1101         if (event->unlinked)
1102                 pinfo->pi_features = LNET_PING_FEAT_INVAL;
1103 }
1104
1105 static int
1106 lnet_ping_info_setup(lnet_ping_info_t **ppinfo, lnet_handle_md_t *md_handle,
1107                      int ni_count, bool set_eq)
1108 {
1109         lnet_handle_me_t  me_handle;
1110         lnet_process_id_t id = {LNET_NID_ANY, LNET_PID_ANY};
1111         lnet_md_t         md = {0};
1112         int               rc, rc2;
1113
1114         if (set_eq) {
1115                 rc = LNetEQAlloc(0, lnet_ping_event_handler,
1116                                  &the_lnet.ln_ping_target_eq);
1117                 if (rc != 0) {
1118                         CERROR("Can't allocate ping EQ: %d\n", rc);
1119                         return rc;
1120                 }
1121         }
1122
1123         *ppinfo = lnet_ping_info_create(ni_count);
1124         if (*ppinfo == NULL) {
1125                 rc = -ENOMEM;
1126                 goto failed_0;
1127         }
1128
1129         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1130                           LNET_PROTO_PING_MATCHBITS, 0,
1131                           LNET_UNLINK, LNET_INS_AFTER,
1132                           &me_handle);
1133         if (rc != 0) {
1134                 CERROR("Can't create ping ME: %d\n", rc);
1135                 goto failed_1;
1136         }
1137
1138         /* initialize md content */
1139         md.start     = *ppinfo;
1140         md.length    = offsetof(lnet_ping_info_t,
1141                                 pi_ni[(*ppinfo)->pi_nnis]);
1142         md.threshold = LNET_MD_THRESH_INF;
1143         md.max_size  = 0;
1144         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1145                        LNET_MD_MANAGE_REMOTE;
1146         md.user_ptr  = NULL;
1147         md.eq_handle = the_lnet.ln_ping_target_eq;
1148         md.user_ptr = *ppinfo;
1149
1150         rc = LNetMDAttach(me_handle, md, LNET_RETAIN, md_handle);
1151         if (rc != 0) {
1152                 CERROR("Can't attach ping MD: %d\n", rc);
1153                 goto failed_2;
1154         }
1155
1156         return 0;
1157
1158 failed_2:
1159         rc2 = LNetMEUnlink(me_handle);
1160         LASSERT(rc2 == 0);
1161 failed_1:
1162         lnet_ping_info_free(*ppinfo);
1163         *ppinfo = NULL;
1164 failed_0:
1165         if (set_eq)
1166                 LNetEQFree(the_lnet.ln_ping_target_eq);
1167         return rc;
1168 }
1169
1170 static void
1171 lnet_ping_md_unlink(lnet_ping_info_t *pinfo, lnet_handle_md_t *md_handle)
1172 {
1173         sigset_t        blocked = cfs_block_allsigs();
1174
1175         LNetMDUnlink(*md_handle);
1176         LNetInvalidateHandle(md_handle);
1177
1178         /* NB md could be busy; this just starts the unlink */
1179         while (pinfo->pi_features != LNET_PING_FEAT_INVAL) {
1180                 CDEBUG(D_NET, "Still waiting for ping MD to unlink\n");
1181                 cfs_pause(cfs_time_seconds(1));
1182         }
1183
1184         cfs_restore_sigs(blocked);
1185 }
1186
1187 static void
1188 lnet_ping_info_install_locked(lnet_ping_info_t *ping_info)
1189 {
1190         int                     i;
1191         lnet_ni_t               *ni;
1192         lnet_ni_status_t        *ns;
1193
1194         i = 0;
1195         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
1196                 LASSERT(i < ping_info->pi_nnis);
1197
1198                 ns = &ping_info->pi_ni[i];
1199
1200                 ns->ns_nid = ni->ni_nid;
1201
1202                 lnet_ni_lock(ni);
1203                 ns->ns_status = (ni->ni_status != NULL) ?
1204                                 ni->ni_status->ns_status : LNET_NI_STATUS_UP;
1205                 ni->ni_status = ns;
1206                 lnet_ni_unlock(ni);
1207
1208                 i++;
1209         }
1210 }
1211
1212 static void
1213 lnet_ping_target_update(lnet_ping_info_t *pinfo, lnet_handle_md_t md_handle)
1214 {
1215         lnet_ping_info_t *old_pinfo = NULL;
1216         lnet_handle_md_t old_md;
1217
1218         /* switch the NIs to point to the new ping info created */
1219         lnet_net_lock(LNET_LOCK_EX);
1220
1221         if (!the_lnet.ln_routing)
1222                 pinfo->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1223         lnet_ping_info_install_locked(pinfo);
1224
1225         if (the_lnet.ln_ping_info != NULL) {
1226                 old_pinfo = the_lnet.ln_ping_info;
1227                 old_md = the_lnet.ln_ping_target_md;
1228         }
1229         the_lnet.ln_ping_target_md = md_handle;
1230         the_lnet.ln_ping_info = pinfo;
1231
1232         lnet_net_unlock(LNET_LOCK_EX);
1233
1234         if (old_pinfo != NULL) {
1235                 /* unlink the old ping info */
1236                 lnet_ping_md_unlink(old_pinfo, &old_md);
1237                 lnet_ping_info_free(old_pinfo);
1238         }
1239 }
1240
1241 static void
1242 lnet_ping_target_fini(void)
1243 {
1244         int             rc;
1245
1246         lnet_ping_md_unlink(the_lnet.ln_ping_info,
1247                             &the_lnet.ln_ping_target_md);
1248
1249         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1250         LASSERT(rc == 0);
1251
1252         lnet_ping_info_destroy();
1253 }
1254
1255 static int
1256 lnet_ni_tq_credits(lnet_ni_t *ni)
1257 {
1258         int     credits;
1259
1260         LASSERT(ni->ni_ncpts >= 1);
1261
1262         if (ni->ni_ncpts == 1)
1263                 return ni->ni_maxtxcredits;
1264
1265         credits = ni->ni_maxtxcredits / ni->ni_ncpts;
1266         credits = max(credits, 8 * ni->ni_peertxcredits);
1267         credits = min(credits, ni->ni_maxtxcredits);
1268
1269         return credits;
1270 }
1271
1272 static void
1273 lnet_ni_unlink_locked(lnet_ni_t *ni)
1274 {
1275         if (!list_empty(&ni->ni_cptlist)) {
1276                 list_del_init(&ni->ni_cptlist);
1277                 lnet_ni_decref_locked(ni, 0);
1278         }
1279
1280         /* move it to zombie list and nobody can find it anymore */
1281         LASSERT(!list_empty(&ni->ni_list));
1282         list_move(&ni->ni_list, &the_lnet.ln_nis_zombie);
1283         lnet_ni_decref_locked(ni, 0);   /* drop ln_nis' ref */
1284 }
1285
1286 static void
1287 lnet_clear_zombies_nis_locked(void)
1288 {
1289         int             i;
1290         int             islo;
1291         lnet_ni_t       *ni;
1292
1293         /* Now wait for the NI's I just nuked to show up on ln_zombie_nis
1294          * and shut them down in guaranteed thread context */
1295         i = 2;
1296         while (!list_empty(&the_lnet.ln_nis_zombie)) {
1297                 int     *ref;
1298                 int     j;
1299
1300                 ni = list_entry(the_lnet.ln_nis_zombie.next,
1301                                 lnet_ni_t, ni_list);
1302                 list_del_init(&ni->ni_list);
1303                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1304                         if (*ref == 0)
1305                                 continue;
1306                         /* still busy, add it back to zombie list */
1307                         list_add(&ni->ni_list, &the_lnet.ln_nis_zombie);
1308                         break;
1309                 }
1310
1311                 if (!list_empty(&ni->ni_list)) {
1312                         lnet_net_unlock(LNET_LOCK_EX);
1313                         ++i;
1314                         if ((i & (-i)) == i) {
1315                                 CDEBUG(D_WARNING,
1316                                        "Waiting for zombie LNI %s\n",
1317                                        libcfs_nid2str(ni->ni_nid));
1318                         }
1319                         cfs_pause(cfs_time_seconds(1));
1320                         lnet_net_lock(LNET_LOCK_EX);
1321                         continue;
1322                 }
1323
1324                 ni->ni_lnd->lnd_refcount--;
1325                 lnet_net_unlock(LNET_LOCK_EX);
1326
1327                 islo = ni->ni_lnd->lnd_type == LOLND;
1328
1329                 LASSERT(!in_interrupt());
1330                 (ni->ni_lnd->lnd_shutdown)(ni);
1331
1332                 /* can't deref lnd anymore now; it might have unregistered
1333                  * itself...  */
1334
1335                 if (!islo)
1336                         CDEBUG(D_LNI, "Removed LNI %s\n",
1337                               libcfs_nid2str(ni->ni_nid));
1338
1339                 lnet_ni_free(ni);
1340                 i = 2;
1341                 lnet_net_lock(LNET_LOCK_EX);
1342         }
1343 }
1344
1345 static void
1346 lnet_shutdown_lndnis(void)
1347 {
1348         int             i;
1349         lnet_ni_t       *ni;
1350
1351         /* NB called holding the global mutex */
1352
1353         /* All quiet on the API front */
1354         LASSERT(!the_lnet.ln_shutdown);
1355         LASSERT(the_lnet.ln_refcount == 0);
1356         LASSERT(list_empty(&the_lnet.ln_nis_zombie));
1357
1358         lnet_net_lock(LNET_LOCK_EX);
1359         the_lnet.ln_shutdown = 1;       /* flag shutdown */
1360
1361         /* Unlink NIs from the global table */
1362         while (!list_empty(&the_lnet.ln_nis)) {
1363                 ni = list_entry(the_lnet.ln_nis.next,
1364                                 lnet_ni_t, ni_list);
1365                 lnet_ni_unlink_locked(ni);
1366         }
1367
1368         /* Drop the cached eqwait NI. */
1369         if (the_lnet.ln_eq_waitni != NULL) {
1370                 lnet_ni_decref_locked(the_lnet.ln_eq_waitni, 0);
1371                 the_lnet.ln_eq_waitni = NULL;
1372         }
1373
1374         /* Drop the cached loopback NI. */
1375         if (the_lnet.ln_loni != NULL) {
1376                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1377                 the_lnet.ln_loni = NULL;
1378         }
1379
1380         lnet_net_unlock(LNET_LOCK_EX);
1381
1382         /* Clear lazy portals and drop delayed messages which hold refs
1383          * on their lnet_msg_t::msg_rxpeer */
1384         for (i = 0; i < the_lnet.ln_nportals; i++)
1385                 LNetClearLazyPortal(i);
1386
1387         /* Clear the peer table and wait for all peers to go (they hold refs on
1388          * their NIs) */
1389         lnet_peer_tables_cleanup(NULL);
1390
1391         lnet_net_lock(LNET_LOCK_EX);
1392
1393         lnet_clear_zombies_nis_locked();
1394         the_lnet.ln_shutdown = 0;
1395         lnet_net_unlock(LNET_LOCK_EX);
1396 }
1397
1398 /* shutdown down the NI and release refcount */
1399 static void
1400 lnet_shutdown_lndni(struct lnet_ni *ni)
1401 {
1402         lnet_net_lock(LNET_LOCK_EX);
1403         lnet_ni_unlink_locked(ni);
1404         lnet_net_unlock(LNET_LOCK_EX);
1405
1406         /* Do peer table cleanup for this ni */
1407         lnet_peer_tables_cleanup(ni);
1408
1409         lnet_net_lock(LNET_LOCK_EX);
1410         lnet_clear_zombies_nis_locked();
1411         lnet_net_unlock(LNET_LOCK_EX);
1412 }
1413
1414 static int
1415 lnet_startup_lndni(struct lnet_ni *ni, __s32 peer_timeout,
1416                    __s32 peer_cr, __s32 peer_buf_cr, __s32 credits)
1417 {
1418         int                     rc = 0;
1419         int                     lnd_type;
1420         lnd_t                   *lnd;
1421         struct lnet_tx_queue    *tq;
1422         int                     i;
1423
1424         lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
1425
1426         LASSERT(libcfs_isknown_lnd(lnd_type));
1427
1428         if (lnd_type == CIBLND || lnd_type == OPENIBLND ||
1429             lnd_type == IIBLND || lnd_type == VIBLND) {
1430                 CERROR("LND %s obsoleted\n", libcfs_lnd2str(lnd_type));
1431                 goto failed0;
1432         }
1433
1434         /* Make sure this new NI is unique. */
1435         lnet_net_lock(LNET_LOCK_EX);
1436         if (!lnet_net_unique(LNET_NIDNET(ni->ni_nid), &the_lnet.ln_nis)) {
1437                 if (lnd_type == LOLND) {
1438                         lnet_net_unlock(LNET_LOCK_EX);
1439                         lnet_ni_free(ni);
1440                         return 0;
1441                 }
1442                 lnet_net_unlock(LNET_LOCK_EX);
1443
1444                 CERROR("Net %s is not unique\n",
1445                        libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
1446                 goto failed0;
1447         }
1448         lnet_net_unlock(LNET_LOCK_EX);
1449
1450         LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
1451         lnd = lnet_find_lnd_by_type(lnd_type);
1452
1453 #ifdef __KERNEL__
1454         if (lnd == NULL) {
1455                 LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1456                 rc = request_module("%s",
1457                                     libcfs_lnd2modname(lnd_type));
1458                 LNET_MUTEX_LOCK(&the_lnet.ln_lnd_mutex);
1459
1460                 lnd = lnet_find_lnd_by_type(lnd_type);
1461                 if (lnd == NULL) {
1462                         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1463                         CERROR("Can't load LND %s, module %s, rc=%d\n",
1464                                libcfs_lnd2str(lnd_type),
1465                                libcfs_lnd2modname(lnd_type), rc);
1466 #ifndef HAVE_MODULE_LOADING_SUPPORT
1467                         LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1468                                            "compiled with kernel module "
1469                                            "loading support.");
1470 #endif
1471                         goto failed0;
1472                 }
1473         }
1474 #else
1475         if (lnd == NULL) {
1476                 LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1477                 CERROR("LND %s not supported\n",
1478                         libcfs_lnd2str(lnd_type));
1479                 goto failed0;
1480         }
1481 #endif
1482
1483         lnet_net_lock(LNET_LOCK_EX);
1484         lnd->lnd_refcount++;
1485         lnet_net_unlock(LNET_LOCK_EX);
1486
1487         ni->ni_lnd = lnd;
1488
1489         rc = (lnd->lnd_startup)(ni);
1490
1491         LNET_MUTEX_UNLOCK(&the_lnet.ln_lnd_mutex);
1492
1493         if (rc != 0) {
1494                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
1495                                    rc, libcfs_lnd2str(lnd->lnd_type));
1496                 lnet_net_lock(LNET_LOCK_EX);
1497                 lnd->lnd_refcount--;
1498                 lnet_net_unlock(LNET_LOCK_EX);
1499                 goto failed0;
1500         }
1501
1502         /* If given some LND tunable parameters, parse those now to
1503          * override the values in the NI structure. */
1504         if (peer_buf_cr >= 0)
1505                 ni->ni_peerrtrcredits = peer_buf_cr;
1506         if (peer_timeout >= 0)
1507                 ni->ni_peertimeout = peer_timeout;
1508         /*
1509          * TODO
1510          * Note: For now, don't allow the user to change
1511          * peertxcredits as this number is used in the
1512          * IB LND to control queue depth.
1513          * if (peer_cr != -1)
1514          *      ni->ni_peertxcredits = peer_cr;
1515          */
1516         if (credits >= 0)
1517                 ni->ni_maxtxcredits = credits;
1518
1519         LASSERT(ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1520
1521         lnet_net_lock(LNET_LOCK_EX);
1522         /* refcount for ln_nis */
1523         lnet_ni_addref_locked(ni, 0);
1524         list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1525         if (ni->ni_cpts != NULL) {
1526                 lnet_ni_addref_locked(ni, 0);
1527                 list_add_tail(&ni->ni_cptlist, &the_lnet.ln_nis_cpt);
1528         }
1529
1530         lnet_net_unlock(LNET_LOCK_EX);
1531
1532         if (lnd->lnd_type == LOLND) {
1533                 lnet_ni_addref(ni);
1534                 LASSERT(the_lnet.ln_loni == NULL);
1535                 the_lnet.ln_loni = ni;
1536                 return 0;
1537         }
1538
1539 #ifndef __KERNEL__
1540         if (lnd->lnd_wait != NULL) {
1541                 if (the_lnet.ln_eq_waitni == NULL) {
1542                         lnet_ni_addref(ni);
1543                         the_lnet.ln_eq_waitni = ni;
1544                 }
1545         } else {
1546 # ifndef HAVE_LIBPTHREAD
1547                 LCONSOLE_ERROR_MSG(0x106, "LND %s not supported in a "
1548                                         "single-threaded runtime\n",
1549                                         libcfs_lnd2str(lnd_type));
1550                 /* shutdown the NI since if we get here then it must've already
1551                  * been started
1552                  */
1553                 lnet_shutdown_lndni(ni);
1554                 return -EINVAL;
1555 # endif
1556         }
1557 #endif
1558         if (ni->ni_peertxcredits == 0 || ni->ni_maxtxcredits == 0) {
1559                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1560                                    libcfs_lnd2str(lnd->lnd_type),
1561                                    ni->ni_peertxcredits == 0 ?
1562                                         "" : "per-peer ");
1563                 /* shutdown the NI since if we get here then it must've already
1564                  * been started
1565                  */
1566                 lnet_shutdown_lndni(ni);
1567                 return -EINVAL;
1568         }
1569
1570         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1571                 tq->tq_credits_min =
1572                 tq->tq_credits_max =
1573                 tq->tq_credits = lnet_ni_tq_credits(ni);
1574         }
1575
1576         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1577                 libcfs_nid2str(ni->ni_nid), ni->ni_peertxcredits,
1578                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1579                 ni->ni_peerrtrcredits, ni->ni_peertimeout);
1580
1581         return 0;
1582 failed0:
1583         lnet_ni_free(ni);
1584         return -EINVAL;
1585 }
1586
1587 static int
1588 lnet_startup_lndnis(struct list_head *nilist)
1589 {
1590         struct lnet_ni          *ni;
1591         int                     rc;
1592         int                     lnd_type;
1593         int                     ni_count = 0;
1594
1595         while (!list_empty(nilist)) {
1596                 ni = list_entry(nilist->next, lnet_ni_t, ni_list);
1597                 list_del(&ni->ni_list);
1598                 rc = lnet_startup_lndni(ni, -1, -1, -1, -1);
1599
1600                 if (rc < 0)
1601                         goto failed;
1602
1603                 ni_count++;
1604         }
1605
1606         if (the_lnet.ln_eq_waitni != NULL && ni_count > 1) {
1607                 lnd_type = the_lnet.ln_eq_waitni->ni_lnd->lnd_type;
1608                 LCONSOLE_ERROR_MSG(0x109, "LND %s can only run single-network"
1609                                    "\n",
1610                                    libcfs_lnd2str(lnd_type));
1611                 rc = -EINVAL;
1612                 goto failed;
1613         }
1614
1615         return ni_count;
1616 failed:
1617         lnet_shutdown_lndnis();
1618
1619         return rc;
1620 }
1621
1622 /**
1623  * Initialize LNet library.
1624  *
1625  * Only userspace program needs to call this function - it's automatically
1626  * called in the kernel at module loading time. Caller has to call LNetFini()
1627  * after a call to LNetInit(), if and only if the latter returned 0. It must
1628  * be called exactly once.
1629  *
1630  * \return 0 on success, and -ve on failures.
1631  */
1632 int
1633 LNetInit(void)
1634 {
1635         int     rc;
1636
1637         lnet_assert_wire_constants();
1638         LASSERT(!the_lnet.ln_init);
1639
1640         memset(&the_lnet, 0, sizeof(the_lnet));
1641
1642         /* refer to global cfs_cpt_table for now */
1643         the_lnet.ln_cpt_table   = cfs_cpt_table;
1644         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1645
1646         LASSERT(the_lnet.ln_cpt_number > 0);
1647         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1648                 /* we are under risk of consuming all lh_cookie */
1649                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
1650                        "please change setting of CPT-table and retry\n",
1651                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1652                 return -1;
1653         }
1654
1655         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1656                 the_lnet.ln_cpt_bits++;
1657
1658         rc = lnet_create_locks();
1659         if (rc != 0) {
1660                 CERROR("Can't create LNet global locks: %d\n", rc);
1661                 return -1;
1662         }
1663
1664         the_lnet.ln_refcount = 0;
1665         the_lnet.ln_init = 1;
1666         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1667         INIT_LIST_HEAD(&the_lnet.ln_lnds);
1668         INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1669         INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1670
1671 #ifdef __KERNEL__
1672         /* The hash table size is the number of bits it takes to express the set
1673          * ln_num_routes, minus 1 (better to under estimate than over so we
1674          * don't waste memory). */
1675         if (rnet_htable_size <= 0)
1676                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
1677         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
1678                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
1679         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
1680                                            order_base_2(rnet_htable_size) - 1);
1681
1682         /* All LNDs apart from the LOLND are in separate modules.  They
1683          * register themselves when their module loads, and unregister
1684          * themselves when their module is unloaded. */
1685 #else
1686         the_lnet.ln_remote_nets_hbits = 8;
1687
1688         /* Register LNDs
1689          * NB the order here determines default 'networks=' order */
1690 # ifdef HAVE_LIBPTHREAD
1691         LNET_REGISTER_ULND(the_tcplnd);
1692 # endif
1693 #endif
1694         lnet_register_lnd(&the_lolnd);
1695         return 0;
1696 }
1697 EXPORT_SYMBOL(LNetInit);
1698
1699 /**
1700  * Finalize LNet library.
1701  *
1702  * Only userspace program needs to call this function. It can be called
1703  * at most once.
1704  *
1705  * \pre LNetInit() called with success.
1706  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1707  */
1708 void
1709 LNetFini(void)
1710 {
1711         LASSERT(the_lnet.ln_init);
1712         LASSERT(the_lnet.ln_refcount == 0);
1713
1714         while (!list_empty(&the_lnet.ln_lnds))
1715                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1716                                                lnd_t, lnd_list));
1717         lnet_destroy_locks();
1718
1719         the_lnet.ln_init = 0;
1720 }
1721 EXPORT_SYMBOL(LNetFini);
1722
1723 /**
1724  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1725  *
1726  * Userspace program should call this after a successful call to LNetInit().
1727  * Users must call this function at least once before any other functions.
1728  * For each successful call there must be a corresponding call to
1729  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1730  * ignored.
1731  *
1732  * The PID used by LNet may be different from the one requested.
1733  * See LNetGetId().
1734  *
1735  * \param requested_pid PID requested by the caller.
1736  *
1737  * \return >= 0 on success, and < 0 error code on failures.
1738  */
1739 int
1740 LNetNIInit(lnet_pid_t requested_pid)
1741 {
1742         int                     im_a_router = 0;
1743         int                     rc, rc2;
1744         int                     ni_count;
1745         lnet_ping_info_t        *pinfo;
1746         lnet_handle_md_t        md_handle;
1747         struct list_head        net_head;
1748
1749         INIT_LIST_HEAD(&net_head);
1750
1751         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
1752
1753         LASSERT(the_lnet.ln_init);
1754         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1755
1756         if (the_lnet.ln_refcount > 0) {
1757                 rc = the_lnet.ln_refcount++;
1758                 LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1759                 return rc;
1760         }
1761
1762         rc = lnet_prepare(requested_pid);
1763         if (rc != 0) {
1764                 LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1765                 return rc;
1766         }
1767
1768         /* Add in the loopback network */
1769         if (lnet_ni_alloc(LNET_MKNET(LOLND, 0), NULL, &net_head) == NULL) {
1770                 rc = -ENOMEM;
1771                 goto failed0;
1772         }
1773
1774         /* If LNet is being initialized via DLC it is possible
1775          * that the user requests not to load module parameters (ones which
1776          * are supported by DLC) on initialization.  Therefore, make sure not
1777          * to load networks, routes and forwarding from module parameters
1778          * in this case.  On cleanup in case of failure only clean up
1779          * routes if it has been loaded */
1780         if (!the_lnet.ln_nis_from_mod_params) {
1781                 rc = lnet_parse_networks(&net_head,
1782                                          lnet_get_networks());
1783                 if (rc < 0)
1784                         goto failed0;
1785         }
1786
1787         ni_count = lnet_startup_lndnis(&net_head);
1788         if (ni_count < 0) {
1789                 rc = ni_count;
1790                 goto failed0;
1791         }
1792
1793         if (!the_lnet.ln_nis_from_mod_params) {
1794                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1795                 if (rc != 0)
1796                         goto failed1;
1797
1798                 rc = lnet_check_routes();
1799                 if (rc != 0)
1800                         goto failed2;
1801
1802                 rc = lnet_rtrpools_alloc(im_a_router);
1803                 if (rc != 0)
1804                         goto failed2;
1805         }
1806
1807         rc = lnet_acceptor_start();
1808         if (rc != 0)
1809                 goto failed2;
1810         the_lnet.ln_refcount = 1;
1811         /* Now I may use my own API functions... */
1812
1813         rc = lnet_ping_info_setup(&pinfo, &md_handle, ni_count, true);
1814         if (rc != 0)
1815                 goto failed3;
1816
1817         lnet_ping_target_update(pinfo, md_handle);
1818
1819         rc = lnet_router_checker_start();
1820         if (rc != 0)
1821                 goto failed4;
1822
1823         lnet_fault_init();
1824         lnet_proc_init();
1825
1826         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1827
1828         return 0;
1829
1830 failed4:
1831         lnet_ping_md_unlink(pinfo, &md_handle);
1832         lnet_ping_info_free(pinfo);
1833         rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1834         LASSERT(rc2 == 0);
1835 failed3:
1836         the_lnet.ln_refcount = 0;
1837         lnet_acceptor_stop();
1838 failed2:
1839         if (!the_lnet.ln_nis_from_mod_params)
1840                 lnet_destroy_routes();
1841 failed1:
1842         lnet_shutdown_lndnis();
1843 failed0:
1844         lnet_unprepare();
1845         LASSERT(rc < 0);
1846         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1847         while (!list_empty(&net_head)) {
1848                 struct lnet_ni *ni;
1849                 ni = list_entry(net_head.next, struct lnet_ni, ni_list);
1850                 list_del_init(&ni->ni_list);
1851                 lnet_ni_free(ni);
1852         }
1853         return rc;
1854 }
1855 EXPORT_SYMBOL(LNetNIInit);
1856
1857 /**
1858  * Stop LNet interfaces, routing, and forwarding.
1859  *
1860  * Users must call this function once for each successful call to LNetNIInit().
1861  * Once the LNetNIFini() operation has been started, the results of pending
1862  * API operations are undefined.
1863  *
1864  * \return always 0 for current implementation.
1865  */
1866 int
1867 LNetNIFini()
1868 {
1869         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
1870
1871         LASSERT (the_lnet.ln_init);
1872         LASSERT (the_lnet.ln_refcount > 0);
1873
1874         if (the_lnet.ln_refcount != 1) {
1875                 the_lnet.ln_refcount--;
1876         } else {
1877                 LASSERT(!the_lnet.ln_niinit_self);
1878
1879                 lnet_fault_fini();
1880
1881                 lnet_proc_fini();
1882                 lnet_router_checker_stop();
1883                 lnet_ping_target_fini();
1884
1885                 /* Teardown fns that use my own API functions BEFORE here */
1886                 the_lnet.ln_refcount = 0;
1887
1888                 lnet_acceptor_stop();
1889                 lnet_destroy_routes();
1890                 lnet_shutdown_lndnis();
1891                 lnet_unprepare();
1892         }
1893
1894         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
1895         return 0;
1896 }
1897 EXPORT_SYMBOL(LNetNIFini);
1898
1899 /**
1900  * Grabs the ni data from the ni structure and fills the out
1901  * parameters
1902  *
1903  * \param[in] ni network        interface structure
1904  * \param[out] cpt_count        the number of cpts the ni is on
1905  * \param[out] nid              Network Interface ID
1906  * \param[out] peer_timeout     NI peer timeout
1907  * \param[out] peer_tx_crdits   NI peer transmit credits
1908  * \param[out] peer_rtr_credits NI peer router credits
1909  * \param[out] max_tx_credits   NI max transmit credit
1910  * \param[out] net_config       Network configuration
1911  */
1912 static void
1913 lnet_fill_ni_info(struct lnet_ni *ni, __u32 *cpt_count, __u64 *nid,
1914                   int *peer_timeout, int *peer_tx_credits,
1915                   int *peer_rtr_credits, int *max_tx_credits,
1916                   struct lnet_ioctl_net_config *net_config)
1917 {
1918         int i;
1919
1920         if (ni == NULL)
1921                 return;
1922
1923         if (net_config == NULL)
1924                 return;
1925
1926         CLASSERT(ARRAY_SIZE(ni->ni_interfaces) ==
1927                  ARRAY_SIZE(net_config->ni_interfaces));
1928
1929         if (ni->ni_interfaces[0] != NULL) {
1930                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
1931                         if (ni->ni_interfaces[i] != NULL) {
1932                                 strncpy(net_config->ni_interfaces[i],
1933                                         ni->ni_interfaces[i],
1934                                         sizeof(net_config->ni_interfaces[i]));
1935                         }
1936                 }
1937         }
1938
1939         *nid = ni->ni_nid;
1940         *peer_timeout = ni->ni_peertimeout;
1941         *peer_tx_credits = ni->ni_peertxcredits;
1942         *peer_rtr_credits = ni->ni_peerrtrcredits;
1943         *max_tx_credits = ni->ni_maxtxcredits;
1944
1945         net_config->ni_status = ni->ni_status->ns_status;
1946
1947         for (i = 0;
1948              ni->ni_cpts != NULL && i < ni->ni_ncpts &&
1949              i < LNET_MAX_SHOW_NUM_CPT;
1950              i++)
1951                 net_config->ni_cpts[i] = ni->ni_cpts[i];
1952
1953         *cpt_count = ni->ni_ncpts;
1954 }
1955
1956 int
1957 lnet_get_net_config(int idx, __u32 *cpt_count, __u64 *nid, int *peer_timeout,
1958                     int *peer_tx_credits, int *peer_rtr_credits,
1959                     int *max_tx_credits,
1960                     struct lnet_ioctl_net_config *net_config)
1961 {
1962         struct lnet_ni          *ni;
1963         struct list_head        *tmp;
1964         int                     cpt;
1965         int                     rc = -ENOENT;
1966
1967         cpt = lnet_net_lock_current();
1968
1969         list_for_each(tmp, &the_lnet.ln_nis) {
1970                 ni = list_entry(tmp, lnet_ni_t, ni_list);
1971                 if (idx-- == 0) {
1972                         rc = 0;
1973                         lnet_ni_lock(ni);
1974                         lnet_fill_ni_info(ni, cpt_count, nid, peer_timeout,
1975                                           peer_tx_credits, peer_rtr_credits,
1976                                           max_tx_credits, net_config);
1977                         lnet_ni_unlock(ni);
1978                         break;
1979                 }
1980         }
1981
1982         lnet_net_unlock(cpt);
1983         return rc;
1984 }
1985
1986 int
1987 lnet_dyn_add_ni(lnet_pid_t requested_pid, char *nets,
1988                 __s32 peer_timeout, __s32 peer_cr, __s32 peer_buf_cr,
1989                 __s32 credits)
1990 {
1991         lnet_ping_info_t        *pinfo;
1992         lnet_handle_md_t        md_handle;
1993         struct lnet_ni          *ni;
1994         struct list_head        net_head;
1995         int                     rc;
1996
1997         INIT_LIST_HEAD(&net_head);
1998
1999         /* Create a ni structure for the network string */
2000         rc = lnet_parse_networks(&net_head, nets);
2001         if (rc <= 0)
2002                 return rc == 0 ? -EINVAL : rc;
2003
2004         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
2005
2006         if (rc > 1) {
2007                 rc = -EINVAL; /* only add one interface per call */
2008                 goto failed0;
2009         }
2010
2011         rc = lnet_ping_info_setup(&pinfo, &md_handle, 1 + lnet_get_ni_count(),
2012                                   false);
2013         if (rc != 0)
2014                 goto failed0;
2015
2016         ni = list_entry(net_head.next, struct lnet_ni, ni_list);
2017         list_del_init(&ni->ni_list);
2018
2019         rc = lnet_startup_lndni(ni, peer_timeout, peer_cr,
2020                                 peer_buf_cr, credits);
2021         if (rc != 0)
2022                 goto failed1;
2023
2024         lnet_ping_target_update(pinfo, md_handle);
2025         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
2026
2027         return 0;
2028
2029 failed1:
2030         lnet_ping_md_unlink(pinfo, &md_handle);
2031         lnet_ping_info_free(pinfo);
2032 failed0:
2033         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
2034         while (!list_empty(&net_head)) {
2035                 ni = list_entry(net_head.next, struct lnet_ni, ni_list);
2036                 list_del_init(&ni->ni_list);
2037                 lnet_ni_free(ni);
2038         }
2039         return rc;
2040 }
2041
2042 int
2043 lnet_dyn_del_ni(__u32 net)
2044 {
2045         lnet_ni_t        *ni;
2046         lnet_ping_info_t *pinfo;
2047         lnet_handle_md_t  md_handle;
2048         int               rc;
2049
2050         /* don't allow userspace to shutdown the LOLND */
2051         if (LNET_NETTYP(net) == LOLND)
2052                 return -EINVAL;
2053
2054         LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
2055         /* create and link a new ping info, before removing the old one */
2056         rc = lnet_ping_info_setup(&pinfo, &md_handle,
2057                                   lnet_get_ni_count() - 1, false);
2058         if (rc != 0)
2059                 goto out;
2060
2061         ni = lnet_net2ni(net);
2062         if (ni == NULL) {
2063                 rc = -EINVAL;
2064                 goto failed;
2065         }
2066
2067         /* decrement the reference counter taken by lnet_net2ni() */
2068         lnet_ni_decref_locked(ni, 0);
2069
2070         lnet_shutdown_lndni(ni);
2071         lnet_ping_target_update(pinfo, md_handle);
2072         goto out;
2073 failed:
2074         lnet_ping_md_unlink(pinfo, &md_handle);
2075         lnet_ping_info_free(pinfo);
2076 out:
2077         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
2078
2079         return rc;
2080 }
2081
2082 /**
2083  * This is an ugly hack to export IOC_LIBCFS_DEBUG_PEER and
2084  * IOC_LIBCFS_PORTALS_COMPATIBILITY commands to users, by tweaking the LNet
2085  * internal ioctl handler.
2086  *
2087  * IOC_LIBCFS_PORTALS_COMPATIBILITY is now deprecated, don't use it.
2088  *
2089  * \param cmd IOC_LIBCFS_DEBUG_PEER to print debugging data about a peer.
2090  * The data will be printed to system console. Don't use it excessively.
2091  * \param arg A pointer to lnet_process_id_t, process ID of the peer.
2092  *
2093  * \return Always return 0 when called by users directly (i.e., not via ioctl).
2094  */
2095 int
2096 LNetCtl(unsigned int cmd, void *arg)
2097 {
2098         struct libcfs_ioctl_data *data = arg;
2099         struct lnet_ioctl_config_data *config;
2100         lnet_process_id_t         id = {0};
2101         lnet_ni_t                *ni;
2102         int                       rc;
2103
2104         CLASSERT(LIBCFS_IOC_DATA_MAX >= sizeof(struct lnet_ioctl_net_config) +
2105                                         sizeof(struct lnet_ioctl_config_data));
2106         LASSERT(the_lnet.ln_init);
2107
2108         switch (cmd) {
2109         case IOC_LIBCFS_GET_NI:
2110                 rc = LNetGetId(data->ioc_count, &id);
2111                 data->ioc_nid = id.nid;
2112                 return rc;
2113
2114         case IOC_LIBCFS_FAIL_NID:
2115                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
2116
2117         case IOC_LIBCFS_ADD_ROUTE:
2118                 config = arg;
2119
2120                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2121                         return -EINVAL;
2122
2123                 LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
2124                 rc = lnet_add_route(config->cfg_net,
2125                                     config->cfg_config_u.cfg_route.rtr_hop,
2126                                     config->cfg_nid,
2127                                     config->cfg_config_u.cfg_route.
2128                                         rtr_priority);
2129                 LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
2130                 return (rc != 0) ? rc : lnet_check_routes();
2131
2132         case IOC_LIBCFS_DEL_ROUTE:
2133                 config = arg;
2134
2135                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2136                         return -EINVAL;
2137
2138                 LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
2139                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
2140                 LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
2141                 return rc;
2142
2143         case IOC_LIBCFS_GET_ROUTE:
2144                 config = arg;
2145
2146                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2147                         return -EINVAL;
2148
2149                 return lnet_get_route(config->cfg_count,
2150                                       &config->cfg_net,
2151                                       &config->cfg_config_u.cfg_route.rtr_hop,
2152                                       &config->cfg_nid,
2153                                       &config->cfg_config_u.cfg_route.rtr_flags,
2154                                       &config->cfg_config_u.cfg_route.
2155                                         rtr_priority);
2156
2157         case IOC_LIBCFS_GET_NET: {
2158                 struct lnet_ioctl_net_config *net_config;
2159                 size_t total = sizeof(*config) + sizeof(*net_config);
2160
2161                 config = arg;
2162
2163                 if (config->cfg_hdr.ioc_len < total)
2164                         return -EINVAL;
2165
2166                 net_config = (struct lnet_ioctl_net_config *)
2167                         config->cfg_bulk;
2168                 if (config == NULL || net_config == NULL)
2169                         return -1;
2170
2171                 return lnet_get_net_config(config->cfg_count,
2172                                            &config->cfg_ncpts,
2173                                            &config->cfg_nid,
2174                                            &config->cfg_config_u.
2175                                                 cfg_net.net_peer_timeout,
2176                                            &config->cfg_config_u.cfg_net.
2177                                                 net_peer_tx_credits,
2178                                            &config->cfg_config_u.cfg_net.
2179                                                 net_peer_rtr_credits,
2180                                            &config->cfg_config_u.cfg_net.
2181                                                 net_max_tx_credits,
2182                                            net_config);
2183         }
2184
2185         case IOC_LIBCFS_GET_LNET_STATS:
2186         {
2187                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
2188
2189                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
2190                         return -EINVAL;
2191
2192                 lnet_counters_get(&lnet_stats->st_cntrs);
2193                 return 0;
2194         }
2195
2196 #if defined(__KERNEL__) && defined(LNET_ROUTER)
2197         case IOC_LIBCFS_CONFIG_RTR:
2198                 config = arg;
2199
2200                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2201                         return -EINVAL;
2202
2203                 LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
2204                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
2205                         rc = lnet_rtrpools_enable();
2206                         LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
2207                         return rc;
2208                 }
2209                 lnet_rtrpools_disable();
2210                 LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
2211                 return 0;
2212
2213         case IOC_LIBCFS_ADD_BUF:
2214                 config = arg;
2215
2216                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2217                         return -EINVAL;
2218
2219                 LNET_MUTEX_LOCK(&the_lnet.ln_api_mutex);
2220                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
2221                                                 buf_tiny,
2222                                           config->cfg_config_u.cfg_buffers.
2223                                                 buf_small,
2224                                           config->cfg_config_u.cfg_buffers.
2225                                                 buf_large);
2226                 LNET_MUTEX_UNLOCK(&the_lnet.ln_api_mutex);
2227                 return rc;
2228 #endif
2229
2230         case IOC_LIBCFS_GET_BUF: {
2231                 struct lnet_ioctl_pool_cfg *pool_cfg;
2232                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
2233
2234                 config = arg;
2235
2236                 if (config->cfg_hdr.ioc_len < total)
2237                         return -EINVAL;
2238
2239                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
2240                 return lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
2241         }
2242
2243         case IOC_LIBCFS_GET_PEER_INFO: {
2244                 struct lnet_ioctl_peer *peer_info = arg;
2245
2246                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
2247                         return -EINVAL;
2248
2249                 return lnet_get_peer_info(
2250                    peer_info->pr_count,
2251                    &peer_info->pr_nid,
2252                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
2253                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
2254                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
2255                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
2256                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
2257                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
2258                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_rtr_credits,
2259                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
2260         }
2261
2262         case IOC_LIBCFS_NOTIFY_ROUTER:
2263                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
2264                                    cfs_time_current() -
2265                                    cfs_time_seconds(cfs_time_current_sec() -
2266                                                     (time_t)data->ioc_u64[0]));
2267
2268         case IOC_LIBCFS_PORTALS_COMPATIBILITY:
2269                 /* This can be removed once lustre stops calling it */
2270                 return 0;
2271
2272         case IOC_LIBCFS_LNET_DIST:
2273                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
2274                 if (rc < 0 && rc != -EHOSTUNREACH)
2275                         return rc;
2276
2277                 data->ioc_u32[0] = rc;
2278                 return 0;
2279
2280         case IOC_LIBCFS_TESTPROTOCOMPAT:
2281                 lnet_net_lock(LNET_LOCK_EX);
2282                 the_lnet.ln_testprotocompat = data->ioc_flags;
2283                 lnet_net_unlock(LNET_LOCK_EX);
2284                 return 0;
2285
2286         case IOC_LIBCFS_LNET_FAULT:
2287                 return lnet_fault_ctl(data->ioc_flags, data);
2288
2289         case IOC_LIBCFS_PING:
2290                 id.nid = data->ioc_nid;
2291                 id.pid = data->ioc_u32[0];
2292                 rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
2293                                (lnet_process_id_t __user *)data->ioc_pbuf1,
2294                                data->ioc_plen1/sizeof(lnet_process_id_t));
2295                 if (rc < 0)
2296                         return rc;
2297                 data->ioc_count = rc;
2298                 return 0;
2299
2300         case IOC_LIBCFS_DEBUG_PEER: {
2301                 /* CAVEAT EMPTOR: this one designed for calling directly; not
2302                  * via an ioctl */
2303                 id = *((lnet_process_id_t *) arg);
2304
2305                 lnet_debug_peer(id.nid);
2306
2307                 ni = lnet_net2ni(LNET_NIDNET(id.nid));
2308                 if (ni == NULL) {
2309                         CDEBUG(D_WARNING, "No NI for %s\n", libcfs_id2str(id));
2310                 } else {
2311                         if (ni->ni_lnd->lnd_ctl == NULL) {
2312                                 CDEBUG(D_WARNING, "No ctl for %s\n",
2313                                        libcfs_id2str(id));
2314                         } else {
2315                                 (void)ni->ni_lnd->lnd_ctl(ni, cmd, arg);
2316                         }
2317
2318                         lnet_ni_decref(ni);
2319                 }
2320                 return 0;
2321         }
2322
2323         default:
2324                 ni = lnet_net2ni(data->ioc_net);
2325                 if (ni == NULL)
2326                         return -EINVAL;
2327
2328                 if (ni->ni_lnd->lnd_ctl == NULL)
2329                         rc = -EINVAL;
2330                 else
2331                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
2332
2333                 lnet_ni_decref(ni);
2334                 return rc;
2335         }
2336         /* not reached */
2337 }
2338 EXPORT_SYMBOL(LNetCtl);
2339
2340 /**
2341  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
2342  * all interfaces share a same PID, as requested by LNetNIInit().
2343  *
2344  * \param index Index of the interface to look up.
2345  * \param id On successful return, this location will hold the
2346  * lnet_process_id_t ID of the interface.
2347  *
2348  * \retval 0 If an interface exists at \a index.
2349  * \retval -ENOENT If no interface has been found.
2350  */
2351 int
2352 LNetGetId(unsigned int index, lnet_process_id_t *id)
2353 {
2354         struct lnet_ni   *ni;
2355         struct list_head *tmp;
2356         int               cpt;
2357         int               rc = -ENOENT;
2358
2359         LASSERT(the_lnet.ln_init);
2360         LASSERT(the_lnet.ln_refcount > 0);
2361
2362         cpt = lnet_net_lock_current();
2363
2364         list_for_each(tmp, &the_lnet.ln_nis) {
2365                 if (index-- != 0)
2366                         continue;
2367
2368                 ni = list_entry(tmp, lnet_ni_t, ni_list);
2369
2370                 id->nid = ni->ni_nid;
2371                 id->pid = the_lnet.ln_pid;
2372                 rc = 0;
2373                 break;
2374         }
2375
2376         lnet_net_unlock(cpt);
2377         return rc;
2378 }
2379 EXPORT_SYMBOL(LNetGetId);
2380
2381 /**
2382  * Print a string representation of handle \a h into buffer \a str of
2383  * \a len bytes.
2384  */
2385 void
2386 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
2387 {
2388         snprintf(str, len, LPX64, h.cookie);
2389 }
2390 EXPORT_SYMBOL(LNetSnprintHandle);
2391
2392 static int
2393 lnet_ping(lnet_process_id_t id, int timeout_ms, lnet_process_id_t __user *ids,
2394           int n_ids)
2395 {
2396         lnet_handle_eq_t     eqh;
2397         lnet_handle_md_t     mdh;
2398         lnet_event_t         event;
2399         lnet_md_t            md = {0};
2400         int                  which;
2401         int                  unlinked = 0;
2402         int                  replied = 0;
2403         const int            a_long_time = 60000; /* mS */
2404         int                  infosz;
2405         lnet_ping_info_t    *info;
2406         lnet_process_id_t    tmpid;
2407         int                  i;
2408         int                  nob;
2409         int                  rc;
2410         int                  rc2;
2411         sigset_t         blocked;
2412
2413         infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
2414
2415         if (n_ids <= 0 ||
2416             id.nid == LNET_NID_ANY ||
2417             timeout_ms > 500000 ||              /* arbitrary limit! */
2418             n_ids > 20)                         /* arbitrary limit! */
2419                 return -EINVAL;
2420
2421         if (id.pid == LNET_PID_ANY)
2422                 id.pid = LNET_PID_LUSTRE;
2423
2424         LIBCFS_ALLOC(info, infosz);
2425         if (info == NULL)
2426                 return -ENOMEM;
2427
2428         /* NB 2 events max (including any unlink event) */
2429         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
2430         if (rc != 0) {
2431                 CERROR("Can't allocate EQ: %d\n", rc);
2432                 goto out_0;
2433         }
2434
2435         /* initialize md content */
2436         md.start     = info;
2437         md.length    = infosz;
2438         md.threshold = 2; /*GET/REPLY*/
2439         md.max_size  = 0;
2440         md.options   = LNET_MD_TRUNCATE;
2441         md.user_ptr  = NULL;
2442         md.eq_handle = eqh;
2443
2444         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
2445         if (rc != 0) {
2446                 CERROR("Can't bind MD: %d\n", rc);
2447                 goto out_1;
2448         }
2449
2450         rc = LNetGet(LNET_NID_ANY, mdh, id,
2451                      LNET_RESERVED_PORTAL,
2452                      LNET_PROTO_PING_MATCHBITS, 0);
2453
2454         if (rc != 0) {
2455                 /* Don't CERROR; this could be deliberate! */
2456
2457                 rc2 = LNetMDUnlink(mdh);
2458                 LASSERT(rc2 == 0);
2459
2460                 /* NB must wait for the UNLINK event below... */
2461                 unlinked = 1;
2462                 timeout_ms = a_long_time;
2463         }
2464
2465         do {
2466                 /* MUST block for unlink to complete */
2467                 if (unlinked)
2468                         blocked = cfs_block_allsigs();
2469
2470                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
2471
2472                 if (unlinked)
2473                         cfs_restore_sigs(blocked);
2474
2475                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
2476                        (rc2 <= 0) ? -1 : event.type,
2477                        (rc2 <= 0) ? -1 : event.status,
2478                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
2479
2480                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
2481
2482                 if (rc2 <= 0 || event.status != 0) {
2483                         /* timeout or error */
2484                         if (!replied && rc == 0)
2485                                 rc = (rc2 < 0) ? rc2 :
2486                                      (rc2 == 0) ? -ETIMEDOUT :
2487                                      event.status;
2488
2489                         if (!unlinked) {
2490                                 /* Ensure completion in finite time... */
2491                                 LNetMDUnlink(mdh);
2492                                 /* No assertion (racing with network) */
2493                                 unlinked = 1;
2494                                 timeout_ms = a_long_time;
2495                         } else if (rc2 == 0) {
2496                                 /* timed out waiting for unlink */
2497                                 CWARN("ping %s: late network completion\n",
2498                                       libcfs_id2str(id));
2499                         }
2500                 } else if (event.type == LNET_EVENT_REPLY) {
2501                         replied = 1;
2502                         rc = event.mlength;
2503                 }
2504
2505         } while (rc2 <= 0 || !event.unlinked);
2506
2507         if (!replied) {
2508                 if (rc >= 0)
2509                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
2510                               libcfs_id2str(id));
2511                 rc = -EIO;
2512                 goto out_1;
2513         }
2514
2515         nob = rc;
2516         LASSERT(nob >= 0 && nob <= infosz);
2517
2518         rc = -EPROTO;                           /* if I can't parse... */
2519
2520         if (nob < 8) {
2521                 /* can't check magic/version */
2522                 CERROR("%s: ping info too short %d\n",
2523                        libcfs_id2str(id), nob);
2524                 goto out_1;
2525         }
2526
2527         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
2528                 lnet_swap_pinginfo(info);
2529         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
2530                 CERROR("%s: Unexpected magic %08x\n",
2531                        libcfs_id2str(id), info->pi_magic);
2532                 goto out_1;
2533         }
2534
2535         if ((info->pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
2536                 CERROR("%s: ping w/o NI status: 0x%x\n",
2537                        libcfs_id2str(id), info->pi_features);
2538                 goto out_1;
2539         }
2540
2541         if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
2542                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
2543                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
2544                 goto out_1;
2545         }
2546
2547         if (info->pi_nnis < n_ids)
2548                 n_ids = info->pi_nnis;
2549
2550         if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
2551                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
2552                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
2553                 goto out_1;
2554         }
2555
2556         rc = -EFAULT;                           /* If I SEGV... */
2557
2558         for (i = 0; i < n_ids; i++) {
2559                 tmpid.pid = info->pi_pid;
2560                 tmpid.nid = info->pi_ni[i].ns_nid;
2561                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
2562                         goto out_1;
2563         }
2564         rc = info->pi_nnis;
2565
2566  out_1:
2567         rc2 = LNetEQFree(eqh);
2568         if (rc2 != 0)
2569                 CERROR("rc2 %d\n", rc2);
2570         LASSERT(rc2 == 0);
2571
2572  out_0:
2573         LIBCFS_FREE(info, infosz);
2574         return rc;
2575 }