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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #define DEBUG_SUBSYSTEM S_LNET
  34 #include <linux/log2.h>
  35 #include <linux/ktime.h>
  36
  37 #include <lnet/lib-lnet.h>
  38
  39 #define D_LNI D_CONSOLE
  40
  41 lnet_t      the_lnet;                           /* THE state of the network */
  42 EXPORT_SYMBOL(the_lnet);
  43
  44 static char *ip2nets = "";
  45 module_param(ip2nets, charp, 0444);
  46 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
  47
  48 static char *networks = "";
  49 module_param(networks, charp, 0444);
  50 MODULE_PARM_DESC(networks, "local networks");
  51
  52 static char *routes = "";
  53 module_param(routes, charp, 0444);
  54 MODULE_PARM_DESC(routes, "routes to non-local networks");
  55
  56 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  57 module_param(rnet_htable_size, int, 0444);
  58 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
  59
  60 static int use_tcp_bonding = false;
  61 module_param(use_tcp_bonding, int, 0444);
  62 MODULE_PARM_DESC(use_tcp_bonding,
  63                  "Set to 1 to use socklnd bonding. 0 to use Multi-Rail");
  64
  65 unsigned int lnet_numa_range = 0;
  66 module_param(lnet_numa_range, uint, 0444);
  67 MODULE_PARM_DESC(lnet_numa_range,
  68                 "NUMA range to consider during Multi-Rail selection");
  69
  70 /*
  71  * This sequence number keeps track of how many times DLC was used to
  72  * update the local NIs. It is incremented when a NI is added or
  73  * removed and checked when sending a message to determine if there is
  74  * a need to re-run the selection algorithm. See lnet_select_pathway()
  75  * for more details on its usage.
  76  */
  77 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
  78
  79 static int lnet_ping(lnet_process_id_t id, signed long timeout,
  80                      lnet_process_id_t __user *ids, int n_ids);
  81
  82 static char *
  83 lnet_get_routes(void)
  84 {
  85         return routes;
  86 }
  87
  88 static char *
  89 lnet_get_networks(void)
  90 {
  91         char   *nets;
  92         int     rc;
  93
  94         if (*networks != 0 && *ip2nets != 0) {
  95                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  96                                    "'ip2nets' but not both at once\n");
  97                 return NULL;
  98         }
  99
 100         if (*ip2nets != 0) {
 101                 rc = lnet_parse_ip2nets(&nets, ip2nets);
 102                 return (rc == 0) ? nets : NULL;
 103         }
 104
 105         if (*networks != 0)
 106                 return networks;
 107
 108         return "tcp";
 109 }
 110
 111 static void
 112 lnet_init_locks(void)
 113 {
 114         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 115         init_waitqueue_head(&the_lnet.ln_eq_waitq);
 116         init_waitqueue_head(&the_lnet.ln_rc_waitq);
 117         mutex_init(&the_lnet.ln_lnd_mutex);
 118         mutex_init(&the_lnet.ln_api_mutex);
 119 }
 120
 121 static void
 122 lnet_fini_locks(void)
 123 {
 124 }
 125
 126 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
 127 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
 128                                             *  MDs kmem_cache */
 129
 130 static int
 131 lnet_descriptor_setup(void)
 132 {
 133         /* create specific kmem_cache for MEs and small MDs (i.e., originally
 134          * allocated in <size-xxx> kmem_cache).
 135          */
 136         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(lnet_me_t),
 137                                             0, 0, NULL);
 138         if (!lnet_mes_cachep)
 139                 return -ENOMEM;
 140
 141         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
 142                                                   LNET_SMALL_MD_SIZE, 0, 0,
 143                                                   NULL);
 144         if (!lnet_small_mds_cachep)
 145                 return -ENOMEM;
 146
 147         return 0;
 148 }
 149
 150 static void
 151 lnet_descriptor_cleanup(void)
 152 {
 153
 154         if (lnet_small_mds_cachep) {
 155                 kmem_cache_destroy(lnet_small_mds_cachep);
 156                 lnet_small_mds_cachep = NULL;
 157         }
 158
 159         if (lnet_mes_cachep) {
 160                 kmem_cache_destroy(lnet_mes_cachep);
 161                 lnet_mes_cachep = NULL;
 162         }
 163 }
 164
 165 static int
 166 lnet_create_remote_nets_table(void)
 167 {
 168         int               i;
 169         struct list_head *hash;
 170
 171         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 172         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 173         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 174         if (hash == NULL) {
 175                 CERROR("Failed to create remote nets hash table\n");
 176                 return -ENOMEM;
 177         }
 178
 179         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 180                 INIT_LIST_HEAD(&hash[i]);
 181         the_lnet.ln_remote_nets_hash = hash;
 182         return 0;
 183 }
 184
 185 static void
 186 lnet_destroy_remote_nets_table(void)
 187 {
 188         int i;
 189
 190         if (the_lnet.ln_remote_nets_hash == NULL)
 191                 return;
 192
 193         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 194                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 195
 196         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 197                     LNET_REMOTE_NETS_HASH_SIZE *
 198                     sizeof(the_lnet.ln_remote_nets_hash[0]));
 199         the_lnet.ln_remote_nets_hash = NULL;
 200 }
 201
 202 static void
 203 lnet_destroy_locks(void)
 204 {
 205         if (the_lnet.ln_res_lock != NULL) {
 206                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 207                 the_lnet.ln_res_lock = NULL;
 208         }
 209
 210         if (the_lnet.ln_net_lock != NULL) {
 211                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 212                 the_lnet.ln_net_lock = NULL;
 213         }
 214
 215         lnet_fini_locks();
 216 }
 217
 218 static int
 219 lnet_create_locks(void)
 220 {
 221         lnet_init_locks();
 222
 223         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 224         if (the_lnet.ln_res_lock == NULL)
 225                 goto failed;
 226
 227         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 228         if (the_lnet.ln_net_lock == NULL)
 229                 goto failed;
 230
 231         return 0;
 232
 233  failed:
 234         lnet_destroy_locks();
 235         return -ENOMEM;
 236 }
 237
 238 static void lnet_assert_wire_constants(void)
 239 {
 240         /* Wire protocol assertions generated by 'wirecheck'
 241          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 242          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 243          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 244
 245         /* Constants... */
 246         CLASSERT(LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 247         CLASSERT(LNET_PROTO_TCP_VERSION_MAJOR == 1);
 248         CLASSERT(LNET_PROTO_TCP_VERSION_MINOR == 0);
 249         CLASSERT(LNET_MSG_ACK == 0);
 250         CLASSERT(LNET_MSG_PUT == 1);
 251         CLASSERT(LNET_MSG_GET == 2);
 252         CLASSERT(LNET_MSG_REPLY == 3);
 253         CLASSERT(LNET_MSG_HELLO == 4);
 254
 255         /* Checks for struct lnet_handle_wire */
 256         CLASSERT((int)sizeof(struct lnet_handle_wire) == 16);
 257         CLASSERT((int)offsetof(struct lnet_handle_wire, wh_interface_cookie) == 0);
 258         CLASSERT((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) == 8);
 259         CLASSERT((int)offsetof(struct lnet_handle_wire, wh_object_cookie) == 8);
 260         CLASSERT((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) == 8);
 261
 262         /* Checks for struct lnet_magicversion_t */
 263         CLASSERT((int)sizeof(lnet_magicversion_t) == 8);
 264         CLASSERT((int)offsetof(lnet_magicversion_t, magic) == 0);
 265         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 266         CLASSERT((int)offsetof(lnet_magicversion_t, version_major) == 4);
 267         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 268         CLASSERT((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 269         CLASSERT((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 270
 271         /* Checks for struct lnet_hdr_t */
 272         CLASSERT((int)sizeof(lnet_hdr_t) == 72);
 273         CLASSERT((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 274         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 275         CLASSERT((int)offsetof(lnet_hdr_t, src_nid) == 8);
 276         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 277         CLASSERT((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 278         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 279         CLASSERT((int)offsetof(lnet_hdr_t, src_pid) == 20);
 280         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 281         CLASSERT((int)offsetof(lnet_hdr_t, type) == 24);
 282         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 283         CLASSERT((int)offsetof(lnet_hdr_t, payload_length) == 28);
 284         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 285         CLASSERT((int)offsetof(lnet_hdr_t, msg) == 32);
 286         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 287
 288         /* Ack */
 289         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 290         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 291         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 292         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 293         CLASSERT((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 294         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 295
 296         /* Put */
 297         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 298         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 299         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 300         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 301         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 302         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 303         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 304         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 305         CLASSERT((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 306         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 307
 308         /* Get */
 309         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 310         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 311         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 312         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 313         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 314         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 315         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 316         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 317         CLASSERT((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 318         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 319
 320         /* Reply */
 321         CLASSERT((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 322         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 323
 324         /* Hello */
 325         CLASSERT((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 326         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 327         CLASSERT((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 328         CLASSERT((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 329 }
 330
 331 static lnd_t *lnet_find_lnd_by_type(__u32 type)
 332 {
 333         lnd_t            *lnd;
 334         struct list_head *tmp;
 335
 336         /* holding lnd mutex */
 337         list_for_each(tmp, &the_lnet.ln_lnds) {
 338                 lnd = list_entry(tmp, lnd_t, lnd_list);
 339
 340                 if (lnd->lnd_type == type)
 341                         return lnd;
 342         }
 343         return NULL;
 344 }
 345
 346 void
 347 lnet_register_lnd (lnd_t *lnd)
 348 {
 349         mutex_lock(&the_lnet.ln_lnd_mutex);
 350
 351         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 352         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 353
 354         list_add_tail(&lnd->lnd_list, &the_lnet.ln_lnds);
 355         lnd->lnd_refcount = 0;
 356
 357         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 358
 359         mutex_unlock(&the_lnet.ln_lnd_mutex);
 360 }
 361 EXPORT_SYMBOL(lnet_register_lnd);
 362
 363 void
 364 lnet_unregister_lnd (lnd_t *lnd)
 365 {
 366         mutex_lock(&the_lnet.ln_lnd_mutex);
 367
 368         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 369         LASSERT(lnd->lnd_refcount == 0);
 370
 371         list_del(&lnd->lnd_list);
 372         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 373
 374         mutex_unlock(&the_lnet.ln_lnd_mutex);
 375 }
 376 EXPORT_SYMBOL(lnet_unregister_lnd);
 377
 378 void
 379 lnet_counters_get(lnet_counters_t *counters)
 380 {
 381         lnet_counters_t *ctr;
 382         int             i;
 383
 384         memset(counters, 0, sizeof(*counters));
 385
 386         lnet_net_lock(LNET_LOCK_EX);
 387
 388         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 389                 counters->msgs_max     += ctr->msgs_max;
 390                 counters->msgs_alloc   += ctr->msgs_alloc;
 391                 counters->errors       += ctr->errors;
 392                 counters->send_count   += ctr->send_count;
 393                 counters->recv_count   += ctr->recv_count;
 394                 counters->route_count  += ctr->route_count;
 395                 counters->drop_count   += ctr->drop_count;
 396                 counters->send_length  += ctr->send_length;
 397                 counters->recv_length  += ctr->recv_length;
 398                 counters->route_length += ctr->route_length;
 399                 counters->drop_length  += ctr->drop_length;
 400
 401         }
 402         lnet_net_unlock(LNET_LOCK_EX);
 403 }
 404 EXPORT_SYMBOL(lnet_counters_get);
 405
 406 void
 407 lnet_counters_reset(void)
 408 {
 409         lnet_counters_t *counters;
 410         int             i;
 411
 412         lnet_net_lock(LNET_LOCK_EX);
 413
 414         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 415                 memset(counters, 0, sizeof(lnet_counters_t));
 416
 417         lnet_net_unlock(LNET_LOCK_EX);
 418 }
 419
 420 static char *
 421 lnet_res_type2str(int type)
 422 {
 423         switch (type) {
 424         default:
 425                 LBUG();
 426         case LNET_COOKIE_TYPE_MD:
 427                 return "MD";
 428         case LNET_COOKIE_TYPE_ME:
 429                 return "ME";
 430         case LNET_COOKIE_TYPE_EQ:
 431                 return "EQ";
 432         }
 433 }
 434
 435 static void
 436 lnet_res_container_cleanup(struct lnet_res_container *rec)
 437 {
 438         int     count = 0;
 439
 440         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 441                 return;
 442
 443         while (!list_empty(&rec->rec_active)) {
 444                 struct list_head *e = rec->rec_active.next;
 445
 446                 list_del_init(e);
 447                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 448                         lnet_eq_free(list_entry(e, lnet_eq_t, eq_list));
 449
 450                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 451                         lnet_md_free(list_entry(e, lnet_libmd_t, md_list));
 452
 453                 } else { /* NB: Active MEs should be attached on portals */
 454                         LBUG();
 455                 }
 456                 count++;
 457         }
 458
 459         if (count > 0) {
 460                 /* Found alive MD/ME/EQ, user really should unlink/free
 461                  * all of them before finalize LNet, but if someone didn't,
 462                  * we have to recycle garbage for him */
 463                 CERROR("%d active elements on exit of %s container\n",
 464                        count, lnet_res_type2str(rec->rec_type));
 465         }
 466
 467         if (rec->rec_lh_hash != NULL) {
 468                 LIBCFS_FREE(rec->rec_lh_hash,
 469                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 470                 rec->rec_lh_hash = NULL;
 471         }
 472
 473         rec->rec_type = 0; /* mark it as finalized */
 474 }
 475
 476 static int
 477 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
 478 {
 479         int     rc = 0;
 480         int     i;
 481
 482         LASSERT(rec->rec_type == 0);
 483
 484         rec->rec_type = type;
 485         INIT_LIST_HEAD(&rec->rec_active);
 486
 487         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 488
 489         /* Arbitrary choice of hash table size */
 490         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 491                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 492         if (rec->rec_lh_hash == NULL) {
 493                 rc = -ENOMEM;
 494                 goto out;
 495         }
 496
 497         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 498                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 499
 500         return 0;
 501
 502 out:
 503         CERROR("Failed to setup %s resource container\n",
 504                lnet_res_type2str(type));
 505         lnet_res_container_cleanup(rec);
 506         return rc;
 507 }
 508
 509 static void
 510 lnet_res_containers_destroy(struct lnet_res_container **recs)
 511 {
 512         struct lnet_res_container       *rec;
 513         int                             i;
 514
 515         cfs_percpt_for_each(rec, i, recs)
 516                 lnet_res_container_cleanup(rec);
 517
 518         cfs_percpt_free(recs);
 519 }
 520
 521 static struct lnet_res_container **
 522 lnet_res_containers_create(int type)
 523 {
 524         struct lnet_res_container       **recs;
 525         struct lnet_res_container       *rec;
 526         int                             rc;
 527         int                             i;
 528
 529         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 530         if (recs == NULL) {
 531                 CERROR("Failed to allocate %s resource containers\n",
 532                        lnet_res_type2str(type));
 533                 return NULL;
 534         }
 535
 536         cfs_percpt_for_each(rec, i, recs) {
 537                 rc = lnet_res_container_setup(rec, i, type);
 538                 if (rc != 0) {
 539                         lnet_res_containers_destroy(recs);
 540                         return NULL;
 541                 }
 542         }
 543
 544         return recs;
 545 }
 546
 547 lnet_libhandle_t *
 548 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 549 {
 550         /* ALWAYS called with lnet_res_lock held */
 551         struct list_head        *head;
 552         lnet_libhandle_t        *lh;
 553         unsigned int            hash;
 554
 555         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
 556                 return NULL;
 557
 558         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 559         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 560
 561         list_for_each_entry(lh, head, lh_hash_chain) {
 562                 if (lh->lh_cookie == cookie)
 563                         return lh;
 564         }
 565
 566         return NULL;
 567 }
 568
 569 void
 570 lnet_res_lh_initialize(struct lnet_res_container *rec, lnet_libhandle_t *lh)
 571 {
 572         /* ALWAYS called with lnet_res_lock held */
 573         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 574         unsigned int    hash;
 575
 576         lh->lh_cookie = rec->rec_lh_cookie;
 577         rec->rec_lh_cookie += 1 << ibits;
 578
 579         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 580
 581         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 582 }
 583
 584 static int lnet_unprepare(void);
 585
 586 static int
 587 lnet_prepare(lnet_pid_t requested_pid)
 588 {
 589         /* Prepare to bring up the network */
 590         struct lnet_res_container **recs;
 591         int                       rc = 0;
 592
 593         if (requested_pid == LNET_PID_ANY) {
 594                 /* Don't instantiate LNET just for me */
 595                 return -ENETDOWN;
 596         }
 597
 598         LASSERT(the_lnet.ln_refcount == 0);
 599
 600         the_lnet.ln_routing = 0;
 601
 602         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
 603         the_lnet.ln_pid = requested_pid;
 604
 605         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
 606         INIT_LIST_HEAD(&the_lnet.ln_peers);
 607         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
 608         INIT_LIST_HEAD(&the_lnet.ln_nets);
 609         INIT_LIST_HEAD(&the_lnet.ln_routers);
 610         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
 611         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
 612
 613         rc = lnet_descriptor_setup();
 614         if (rc != 0)
 615                 goto failed;
 616
 617         rc = lnet_create_remote_nets_table();
 618         if (rc != 0)
 619                 goto failed;
 620
 621         /*
 622          * NB the interface cookie in wire handles guards against delayed
 623          * replies and ACKs appearing valid after reboot.
 624          */
 625         the_lnet.ln_interface_cookie = ktime_get_real_ns();
 626
 627         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
 628                                                 sizeof(lnet_counters_t));
 629         if (the_lnet.ln_counters == NULL) {
 630                 CERROR("Failed to allocate counters for LNet\n");
 631                 rc = -ENOMEM;
 632                 goto failed;
 633         }
 634
 635         rc = lnet_peer_tables_create();
 636         if (rc != 0)
 637                 goto failed;
 638
 639         rc = lnet_msg_containers_create();
 640         if (rc != 0)
 641                 goto failed;
 642
 643         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 644                                       LNET_COOKIE_TYPE_EQ);
 645         if (rc != 0)
 646                 goto failed;
 647
 648         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME);
 649         if (recs == NULL) {
 650                 rc = -ENOMEM;
 651                 goto failed;
 652         }
 653
 654         the_lnet.ln_me_containers = recs;
 655
 656         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
 657         if (recs == NULL) {
 658                 rc = -ENOMEM;
 659                 goto failed;
 660         }
 661
 662         the_lnet.ln_md_containers = recs;
 663
 664         rc = lnet_portals_create();
 665         if (rc != 0) {
 666                 CERROR("Failed to create portals for LNet: %d\n", rc);
 667                 goto failed;
 668         }
 669
 670         return 0;
 671
 672  failed:
 673         lnet_unprepare();
 674         return rc;
 675 }
 676
 677 static int
 678 lnet_unprepare (void)
 679 {
 680         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 681          * have shut down already, so it is safe to unlink and free all
 682          * descriptors, even those that appear committed to a network op (eg MD
 683          * with non-zero pending count) */
 684
 685         lnet_fail_nid(LNET_NID_ANY, 0);
 686
 687         LASSERT(the_lnet.ln_refcount == 0);
 688         LASSERT(list_empty(&the_lnet.ln_test_peers));
 689         LASSERT(list_empty(&the_lnet.ln_nets));
 690
 691         lnet_portals_destroy();
 692
 693         if (the_lnet.ln_md_containers != NULL) {
 694                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 695                 the_lnet.ln_md_containers = NULL;
 696         }
 697
 698         if (the_lnet.ln_me_containers != NULL) {
 699                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 700                 the_lnet.ln_me_containers = NULL;
 701         }
 702
 703         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 704
 705         lnet_msg_containers_destroy();
 706         lnet_peer_uninit();
 707         lnet_rtrpools_free(0);
 708
 709         if (the_lnet.ln_counters != NULL) {
 710                 cfs_percpt_free(the_lnet.ln_counters);
 711                 the_lnet.ln_counters = NULL;
 712         }
 713         lnet_destroy_remote_nets_table();
 714         lnet_descriptor_cleanup();
 715
 716         return 0;
 717 }
 718
 719 lnet_ni_t  *
 720 lnet_net2ni_locked(__u32 net_id, int cpt)
 721 {
 722         struct lnet_ni   *ni;
 723         struct lnet_net  *net;
 724
 725         LASSERT(cpt != LNET_LOCK_EX);
 726
 727         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 728                 if (net->net_id == net_id) {
 729                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
 730                                         ni_netlist);
 731                         return ni;
 732                 }
 733         }
 734
 735         return NULL;
 736 }
 737
 738 lnet_ni_t *
 739 lnet_net2ni_addref(__u32 net)
 740 {
 741         lnet_ni_t *ni;
 742
 743         lnet_net_lock(0);
 744         ni = lnet_net2ni_locked(net, 0);
 745         if (ni)
 746                 lnet_ni_addref_locked(ni, 0);
 747         lnet_net_unlock(0);
 748
 749         return ni;
 750 }
 751 EXPORT_SYMBOL(lnet_net2ni_addref);
 752
 753 struct lnet_net *
 754 lnet_get_net_locked(__u32 net_id)
 755 {
 756         struct lnet_net  *net;
 757
 758         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 759                 if (net->net_id == net_id)
 760                         return net;
 761         }
 762
 763         return NULL;
 764 }
 765
 766 unsigned int
 767 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 768 {
 769         __u64           key = nid;
 770         unsigned int    val;
 771
 772         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 773
 774         if (number == 1)
 775                 return 0;
 776
 777         val = hash_long(key, LNET_CPT_BITS);
 778         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 779         if (val < number)
 780                 return val;
 781
 782         return (unsigned int)(key + val + (val >> 1)) % number;
 783 }
 784
 785 int
 786 lnet_cpt_of_nid_locked(lnet_nid_t nid, struct lnet_ni *ni)
 787 {
 788         struct lnet_net *net;
 789
 790         /* must called with hold of lnet_net_lock */
 791         if (LNET_CPT_NUMBER == 1)
 792                 return 0; /* the only one */
 793
 794         /*
 795          * If NI is provided then use the CPT identified in the NI cpt
 796          * list if one exists. If one doesn't exist, then that NI is
 797          * associated with all CPTs and it follows that the net it belongs
 798          * to is implicitly associated with all CPTs, so just hash the nid
 799          * and return that.
 800          */
 801         if (ni != NULL) {
 802                 if (ni->ni_cpts != NULL)
 803                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
 804                                                              ni->ni_ncpts)];
 805                 else
 806                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 807         }
 808
 809         /* no NI provided so look at the net */
 810         net = lnet_get_net_locked(LNET_NIDNET(nid));
 811
 812         if (net != NULL && net->net_cpts != NULL) {
 813                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
 814         }
 815
 816         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 817 }
 818
 819 int
 820 lnet_cpt_of_nid(lnet_nid_t nid, struct lnet_ni *ni)
 821 {
 822         int     cpt;
 823         int     cpt2;
 824
 825         if (LNET_CPT_NUMBER == 1)
 826                 return 0; /* the only one */
 827
 828         cpt = lnet_net_lock_current();
 829
 830         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
 831
 832         lnet_net_unlock(cpt);
 833
 834         return cpt2;
 835 }
 836 EXPORT_SYMBOL(lnet_cpt_of_nid);
 837
 838 int
 839 lnet_islocalnet(__u32 net_id)
 840 {
 841         struct lnet_net *net;
 842         int             cpt;
 843         bool            local;
 844
 845         cpt = lnet_net_lock_current();
 846
 847         net = lnet_get_net_locked(net_id);
 848
 849         local = net != NULL;
 850
 851         lnet_net_unlock(cpt);
 852
 853         return local;
 854 }
 855
 856 bool
 857 lnet_is_ni_healthy_locked(struct lnet_ni *ni)
 858 {
 859         if (ni->ni_state == LNET_NI_STATE_ACTIVE ||
 860             ni->ni_state == LNET_NI_STATE_DEGRADED)
 861                 return true;
 862
 863         return false;
 864 }
 865
 866 lnet_ni_t  *
 867 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 868 {
 869         struct lnet_net  *net;
 870         struct lnet_ni   *ni;
 871
 872         LASSERT(cpt != LNET_LOCK_EX);
 873
 874         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 875                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
 876                         if (ni->ni_nid == nid)
 877                                 return ni;
 878                 }
 879         }
 880
 881         return NULL;
 882 }
 883
 884 lnet_ni_t *
 885 lnet_nid2ni_addref(lnet_nid_t nid)
 886 {
 887         lnet_ni_t *ni;
 888
 889         lnet_net_lock(0);
 890         ni = lnet_nid2ni_locked(nid, 0);
 891         if (ni)
 892                 lnet_ni_addref_locked(ni, 0);
 893         lnet_net_unlock(0);
 894
 895         return ni;
 896 }
 897 EXPORT_SYMBOL(lnet_nid2ni_addref);
 898
 899 int
 900 lnet_islocalnid(lnet_nid_t nid)
 901 {
 902         struct lnet_ni  *ni;
 903         int             cpt;
 904
 905         cpt = lnet_net_lock_current();
 906         ni = lnet_nid2ni_locked(nid, cpt);
 907         lnet_net_unlock(cpt);
 908
 909         return ni != NULL;
 910 }
 911
 912 int
 913 lnet_count_acceptor_nets(void)
 914 {
 915         /* Return the # of NIs that need the acceptor. */
 916         int              count = 0;
 917         struct lnet_net  *net;
 918         int              cpt;
 919
 920         cpt = lnet_net_lock_current();
 921         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 922                 /* all socklnd type networks should have the acceptor
 923                  * thread started */
 924                 if (net->net_lnd->lnd_accept != NULL)
 925                         count++;
 926         }
 927
 928         lnet_net_unlock(cpt);
 929
 930         return count;
 931 }
 932
 933 static struct lnet_ping_info *
 934 lnet_ping_info_create(int num_ni)
 935 {
 936         struct lnet_ping_info *ping_info;
 937         unsigned int     infosz;
 938
 939         infosz = offsetof(struct lnet_ping_info, pi_ni[num_ni]);
 940         LIBCFS_ALLOC(ping_info, infosz);
 941         if (ping_info == NULL) {
 942                 CERROR("Can't allocate ping info[%d]\n", num_ni);
 943                 return NULL;
 944         }
 945
 946         ping_info->pi_nnis = num_ni;
 947         ping_info->pi_pid = the_lnet.ln_pid;
 948         ping_info->pi_magic = LNET_PROTO_PING_MAGIC;
 949         ping_info->pi_features = LNET_PING_FEAT_NI_STATUS;
 950
 951         return ping_info;
 952 }
 953
 954 static inline int
 955 lnet_get_net_ni_count_locked(struct lnet_net *net)
 956 {
 957         struct lnet_ni  *ni;
 958         int             count = 0;
 959
 960         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
 961                 count++;
 962
 963         return count;
 964 }
 965
 966 static inline int
 967 lnet_get_net_ni_count_pre(struct lnet_net *net)
 968 {
 969         struct lnet_ni  *ni;
 970         int             count = 0;
 971
 972         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
 973                 count++;
 974
 975         return count;
 976 }
 977
 978 static inline int
 979 lnet_get_ni_count(void)
 980 {
 981         struct lnet_ni  *ni;
 982         struct lnet_net *net;
 983         int             count = 0;
 984
 985         lnet_net_lock(0);
 986
 987         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 988                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
 989                         count++;
 990         }
 991
 992         lnet_net_unlock(0);
 993
 994         return count;
 995 }
 996
 997 static inline void
 998 lnet_ping_info_free(struct lnet_ping_info *pinfo)
 999 {
1000         LIBCFS_FREE(pinfo,
1001                     offsetof(struct lnet_ping_info,
1002                              pi_ni[pinfo->pi_nnis]));
1003 }
1004
1005 static void
1006 lnet_ping_info_destroy(void)
1007 {
1008         struct lnet_net *net;
1009         struct lnet_ni  *ni;
1010
1011         lnet_net_lock(LNET_LOCK_EX);
1012
1013         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1014                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1015                         lnet_ni_lock(ni);
1016                         ni->ni_status = NULL;
1017                         lnet_ni_unlock(ni);
1018                 }
1019         }
1020
1021         lnet_ping_info_free(the_lnet.ln_ping_info);
1022         the_lnet.ln_ping_info = NULL;
1023
1024         lnet_net_unlock(LNET_LOCK_EX);
1025 }
1026
1027 static void
1028 lnet_ping_event_handler(lnet_event_t *event)
1029 {
1030         struct lnet_ping_info *pinfo = event->md.user_ptr;
1031
1032         if (event->unlinked)
1033                 pinfo->pi_features = LNET_PING_FEAT_INVAL;
1034 }
1035
1036 static int
1037 lnet_ping_info_setup(struct lnet_ping_info **ppinfo, lnet_handle_md_t *md_handle,
1038                      int ni_count, bool set_eq)
1039 {
1040         lnet_handle_me_t  me_handle;
1041         lnet_process_id_t id = {LNET_NID_ANY, LNET_PID_ANY};
1042         lnet_md_t         md = {NULL};
1043         int               rc, rc2;
1044
1045         if (set_eq) {
1046                 rc = LNetEQAlloc(0, lnet_ping_event_handler,
1047                                  &the_lnet.ln_ping_target_eq);
1048                 if (rc != 0) {
1049                         CERROR("Can't allocate ping EQ: %d\n", rc);
1050                         return rc;
1051                 }
1052         }
1053
1054         *ppinfo = lnet_ping_info_create(ni_count);
1055         if (*ppinfo == NULL) {
1056                 rc = -ENOMEM;
1057                 goto failed_0;
1058         }
1059
1060         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1061                           LNET_PROTO_PING_MATCHBITS, 0,
1062                           LNET_UNLINK, LNET_INS_AFTER,
1063                           &me_handle);
1064         if (rc != 0) {
1065                 CERROR("Can't create ping ME: %d\n", rc);
1066                 goto failed_1;
1067         }
1068
1069         /* initialize md content */
1070         md.start     = *ppinfo;
1071         md.length    = offsetof(struct lnet_ping_info,
1072                                 pi_ni[(*ppinfo)->pi_nnis]);
1073         md.threshold = LNET_MD_THRESH_INF;
1074         md.max_size  = 0;
1075         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1076                        LNET_MD_MANAGE_REMOTE;
1077         md.user_ptr  = NULL;
1078         md.eq_handle = the_lnet.ln_ping_target_eq;
1079         md.user_ptr = *ppinfo;
1080
1081         rc = LNetMDAttach(me_handle, md, LNET_RETAIN, md_handle);
1082         if (rc != 0) {
1083                 CERROR("Can't attach ping MD: %d\n", rc);
1084                 goto failed_2;
1085         }
1086
1087         return 0;
1088
1089 failed_2:
1090         rc2 = LNetMEUnlink(me_handle);
1091         LASSERT(rc2 == 0);
1092 failed_1:
1093         lnet_ping_info_free(*ppinfo);
1094         *ppinfo = NULL;
1095 failed_0:
1096         if (set_eq)
1097                 LNetEQFree(the_lnet.ln_ping_target_eq);
1098         return rc;
1099 }
1100
1101 static void
1102 lnet_ping_md_unlink(struct lnet_ping_info *pinfo, lnet_handle_md_t *md_handle)
1103 {
1104         sigset_t        blocked = cfs_block_allsigs();
1105
1106         LNetMDUnlink(*md_handle);
1107         LNetInvalidateHandle(md_handle);
1108
1109         /* NB md could be busy; this just starts the unlink */
1110         while (pinfo->pi_features != LNET_PING_FEAT_INVAL) {
1111                 CDEBUG(D_NET, "Still waiting for ping MD to unlink\n");
1112                 set_current_state(TASK_UNINTERRUPTIBLE);
1113                 schedule_timeout(cfs_time_seconds(1));
1114         }
1115
1116         cfs_restore_sigs(blocked);
1117 }
1118
1119 static void
1120 lnet_ping_info_install_locked(struct lnet_ping_info *ping_info)
1121 {
1122         int                     i;
1123         struct lnet_ni          *ni;
1124         struct lnet_net         *net;
1125         struct lnet_ni_status *ns;
1126
1127         i = 0;
1128         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1129                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1130                         LASSERT(i < ping_info->pi_nnis);
1131
1132                         ns = &ping_info->pi_ni[i];
1133
1134                         ns->ns_nid = ni->ni_nid;
1135
1136                         lnet_ni_lock(ni);
1137                         ns->ns_status = (ni->ni_status != NULL) ?
1138                                         ni->ni_status->ns_status :
1139                                                 LNET_NI_STATUS_UP;
1140                         ni->ni_status = ns;
1141                         lnet_ni_unlock(ni);
1142
1143                         i++;
1144                 }
1145
1146         }
1147 }
1148
1149 static void
1150 lnet_ping_target_update(struct lnet_ping_info *pinfo, lnet_handle_md_t md_handle)
1151 {
1152         struct lnet_ping_info *old_pinfo = NULL;
1153         lnet_handle_md_t old_md;
1154
1155         /* switch the NIs to point to the new ping info created */
1156         lnet_net_lock(LNET_LOCK_EX);
1157
1158         if (!the_lnet.ln_routing)
1159                 pinfo->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1160         lnet_ping_info_install_locked(pinfo);
1161
1162         if (the_lnet.ln_ping_info != NULL) {
1163                 old_pinfo = the_lnet.ln_ping_info;
1164                 old_md = the_lnet.ln_ping_target_md;
1165         }
1166         the_lnet.ln_ping_target_md = md_handle;
1167         the_lnet.ln_ping_info = pinfo;
1168
1169         lnet_net_unlock(LNET_LOCK_EX);
1170
1171         if (old_pinfo != NULL) {
1172                 /* unlink the old ping info */
1173                 lnet_ping_md_unlink(old_pinfo, &old_md);
1174                 lnet_ping_info_free(old_pinfo);
1175         }
1176 }
1177
1178 static void
1179 lnet_ping_target_fini(void)
1180 {
1181         int             rc;
1182
1183         lnet_ping_md_unlink(the_lnet.ln_ping_info,
1184                             &the_lnet.ln_ping_target_md);
1185
1186         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1187         LASSERT(rc == 0);
1188
1189         lnet_ping_info_destroy();
1190 }
1191
1192 static int
1193 lnet_ni_tq_credits(lnet_ni_t *ni)
1194 {
1195         int     credits;
1196
1197         LASSERT(ni->ni_ncpts >= 1);
1198
1199         if (ni->ni_ncpts == 1)
1200                 return ni->ni_net->net_tunables.lct_max_tx_credits;
1201
1202         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
1203         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
1204         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
1205
1206         return credits;
1207 }
1208
1209 static void
1210 lnet_ni_unlink_locked(lnet_ni_t *ni)
1211 {
1212         if (!list_empty(&ni->ni_cptlist)) {
1213                 list_del_init(&ni->ni_cptlist);
1214                 lnet_ni_decref_locked(ni, 0);
1215         }
1216
1217         /* move it to zombie list and nobody can find it anymore */
1218         LASSERT(!list_empty(&ni->ni_netlist));
1219         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
1220         lnet_ni_decref_locked(ni, 0);
1221 }
1222
1223 static void
1224 lnet_clear_zombies_nis_locked(struct lnet_net *net)
1225 {
1226         int             i;
1227         int             islo;
1228         lnet_ni_t       *ni;
1229         struct list_head *zombie_list = &net->net_ni_zombie;
1230
1231         /*
1232          * Now wait for the NIs I just nuked to show up on the zombie
1233          * list and shut them down in guaranteed thread context
1234          */
1235         i = 2;
1236         while (!list_empty(zombie_list)) {
1237                 int     *ref;
1238                 int     j;
1239
1240                 ni = list_entry(zombie_list->next,
1241                                 lnet_ni_t, ni_netlist);
1242                 list_del_init(&ni->ni_netlist);
1243                 /* the ni should be in deleting state. If it's not it's
1244                  * a bug */
1245                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
1246                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1247                         if (*ref == 0)
1248                                 continue;
1249                         /* still busy, add it back to zombie list */
1250                         list_add(&ni->ni_netlist, zombie_list);
1251                         break;
1252                 }
1253
1254                 if (!list_empty(&ni->ni_netlist)) {
1255                         lnet_net_unlock(LNET_LOCK_EX);
1256                         ++i;
1257                         if ((i & (-i)) == i) {
1258                                 CDEBUG(D_WARNING,
1259                                        "Waiting for zombie LNI %s\n",
1260                                        libcfs_nid2str(ni->ni_nid));
1261                         }
1262                         set_current_state(TASK_UNINTERRUPTIBLE);
1263                         schedule_timeout(cfs_time_seconds(1));
1264                         lnet_net_lock(LNET_LOCK_EX);
1265                         continue;
1266                 }
1267
1268                 lnet_net_unlock(LNET_LOCK_EX);
1269
1270                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
1271
1272                 LASSERT(!in_interrupt());
1273                 (net->net_lnd->lnd_shutdown)(ni);
1274
1275                 if (!islo)
1276                         CDEBUG(D_LNI, "Removed LNI %s\n",
1277                               libcfs_nid2str(ni->ni_nid));
1278
1279                 lnet_ni_free(ni);
1280                 i = 2;
1281                 lnet_net_lock(LNET_LOCK_EX);
1282         }
1283 }
1284
1285 /* shutdown down the NI and release refcount */
1286 static void
1287 lnet_shutdown_lndni(struct lnet_ni *ni)
1288 {
1289         int i;
1290         struct lnet_net *net = ni->ni_net;
1291
1292         lnet_net_lock(LNET_LOCK_EX);
1293         ni->ni_state = LNET_NI_STATE_DELETING;
1294         lnet_ni_unlink_locked(ni);
1295         lnet_incr_dlc_seq();
1296         lnet_net_unlock(LNET_LOCK_EX);
1297
1298         /* clear messages for this NI on the lazy portal */
1299         for (i = 0; i < the_lnet.ln_nportals; i++)
1300                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
1301
1302         lnet_net_lock(LNET_LOCK_EX);
1303         lnet_clear_zombies_nis_locked(net);
1304         lnet_net_unlock(LNET_LOCK_EX);
1305 }
1306
1307 static void
1308 lnet_shutdown_lndnet(struct lnet_net *net)
1309 {
1310         struct lnet_ni *ni;
1311
1312         lnet_net_lock(LNET_LOCK_EX);
1313
1314         net->net_state = LNET_NET_STATE_DELETING;
1315
1316         list_del_init(&net->net_list);
1317
1318         while (!list_empty(&net->net_ni_list)) {
1319                 ni = list_entry(net->net_ni_list.next,
1320                                 lnet_ni_t, ni_netlist);
1321                 lnet_net_unlock(LNET_LOCK_EX);
1322                 lnet_shutdown_lndni(ni);
1323                 lnet_net_lock(LNET_LOCK_EX);
1324         }
1325
1326         lnet_net_unlock(LNET_LOCK_EX);
1327
1328         /* Do peer table cleanup for this net */
1329         lnet_peer_tables_cleanup(net);
1330
1331         lnet_net_lock(LNET_LOCK_EX);
1332         /*
1333          * decrement ref count on lnd only when the entire network goes
1334          * away
1335          */
1336         net->net_lnd->lnd_refcount--;
1337
1338         lnet_net_unlock(LNET_LOCK_EX);
1339
1340         lnet_net_free(net);
1341 }
1342
1343 static void
1344 lnet_shutdown_lndnets(void)
1345 {
1346         struct lnet_net *net;
1347
1348         /* NB called holding the global mutex */
1349
1350         /* All quiet on the API front */
1351         LASSERT(!the_lnet.ln_shutdown);
1352         LASSERT(the_lnet.ln_refcount == 0);
1353
1354         lnet_net_lock(LNET_LOCK_EX);
1355         the_lnet.ln_shutdown = 1;       /* flag shutdown */
1356
1357         while (!list_empty(&the_lnet.ln_nets)) {
1358                 /*
1359                  * move the nets to the zombie list to avoid them being
1360                  * picked up for new work. LONET is also included in the
1361                  * Nets that will be moved to the zombie list
1362                  */
1363                 net = list_entry(the_lnet.ln_nets.next,
1364                                  struct lnet_net, net_list);
1365                 list_move(&net->net_list, &the_lnet.ln_net_zombie);
1366         }
1367
1368         /* Drop the cached loopback Net. */
1369         if (the_lnet.ln_loni != NULL) {
1370                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1371                 the_lnet.ln_loni = NULL;
1372         }
1373         lnet_net_unlock(LNET_LOCK_EX);
1374
1375         /* iterate through the net zombie list and delete each net */
1376         while (!list_empty(&the_lnet.ln_net_zombie)) {
1377                 net = list_entry(the_lnet.ln_net_zombie.next,
1378                                  struct lnet_net, net_list);
1379                 lnet_shutdown_lndnet(net);
1380         }
1381
1382         lnet_net_lock(LNET_LOCK_EX);
1383         the_lnet.ln_shutdown = 0;
1384         lnet_net_unlock(LNET_LOCK_EX);
1385 }
1386
1387 static int
1388 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
1389 {
1390         int                     rc = -EINVAL;
1391         struct lnet_tx_queue    *tq;
1392         int                     i;
1393         struct lnet_net         *net = ni->ni_net;
1394
1395         mutex_lock(&the_lnet.ln_lnd_mutex);
1396
1397         if (tun) {
1398                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
1399                 ni->ni_lnd_tunables_set = true;
1400         }
1401
1402         rc = (net->net_lnd->lnd_startup)(ni);
1403
1404         mutex_unlock(&the_lnet.ln_lnd_mutex);
1405
1406         if (rc != 0) {
1407                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
1408                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
1409                 lnet_net_lock(LNET_LOCK_EX);
1410                 net->net_lnd->lnd_refcount--;
1411                 lnet_net_unlock(LNET_LOCK_EX);
1412                 goto failed0;
1413         }
1414
1415         ni->ni_state = LNET_NI_STATE_ACTIVE;
1416
1417         /* We keep a reference on the loopback net through the loopback NI */
1418         if (net->net_lnd->lnd_type == LOLND) {
1419                 lnet_ni_addref(ni);
1420                 LASSERT(the_lnet.ln_loni == NULL);
1421                 the_lnet.ln_loni = ni;
1422                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
1423                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
1424                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
1425                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
1426                 return 0;
1427         }
1428
1429         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
1430             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
1431                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1432                                    libcfs_lnd2str(net->net_lnd->lnd_type),
1433                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
1434                                         "" : "per-peer ");
1435                 /* shutdown the NI since if we get here then it must've already
1436                  * been started
1437                  */
1438                 lnet_shutdown_lndni(ni);
1439                 return -EINVAL;
1440         }
1441
1442         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1443                 tq->tq_credits_min =
1444                 tq->tq_credits_max =
1445                 tq->tq_credits = lnet_ni_tq_credits(ni);
1446         }
1447
1448         atomic_set(&ni->ni_tx_credits,
1449                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
1450
1451         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1452                 libcfs_nid2str(ni->ni_nid),
1453                 ni->ni_net->net_tunables.lct_peer_tx_credits,
1454                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1455                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
1456                 ni->ni_net->net_tunables.lct_peer_timeout);
1457
1458         return 0;
1459 failed0:
1460         lnet_ni_free(ni);
1461         return rc;
1462 }
1463
1464 static int
1465 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
1466 {
1467         struct lnet_ni          *ni;
1468         struct lnet_net         *net_l = NULL;
1469         struct list_head        local_ni_list;
1470         int                     rc;
1471         int                     ni_count = 0;
1472         __u32                   lnd_type;
1473         lnd_t                   *lnd;
1474         int                     peer_timeout =
1475                 net->net_tunables.lct_peer_timeout;
1476         int                     maxtxcredits =
1477                 net->net_tunables.lct_max_tx_credits;
1478         int                     peerrtrcredits =
1479                 net->net_tunables.lct_peer_rtr_credits;
1480
1481         INIT_LIST_HEAD(&local_ni_list);
1482
1483         /*
1484          * make sure that this net is unique. If it isn't then
1485          * we are adding interfaces to an already existing network, and
1486          * 'net' is just a convenient way to pass in the list.
1487          * if it is unique we need to find the LND and load it if
1488          * necessary.
1489          */
1490         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
1491                 lnd_type = LNET_NETTYP(net->net_id);
1492
1493                 LASSERT(libcfs_isknown_lnd(lnd_type));
1494
1495                 if (lnd_type == CIBLND || lnd_type == OPENIBLND ||
1496                     lnd_type == IIBLND || lnd_type == VIBLND) {
1497                         CERROR("LND %s obsoleted\n", libcfs_lnd2str(lnd_type));
1498                         rc = -EINVAL;
1499                         goto failed0;
1500                 }
1501
1502                 mutex_lock(&the_lnet.ln_lnd_mutex);
1503                 lnd = lnet_find_lnd_by_type(lnd_type);
1504
1505                 if (lnd == NULL) {
1506                         mutex_unlock(&the_lnet.ln_lnd_mutex);
1507                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
1508                         mutex_lock(&the_lnet.ln_lnd_mutex);
1509
1510                         lnd = lnet_find_lnd_by_type(lnd_type);
1511                         if (lnd == NULL) {
1512                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
1513                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
1514                                 libcfs_lnd2str(lnd_type),
1515                                 libcfs_lnd2modname(lnd_type), rc);
1516 #ifndef HAVE_MODULE_LOADING_SUPPORT
1517                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1518                                                 "compiled with kernel module "
1519                                                 "loading support.");
1520 #endif
1521                                 rc = -EINVAL;
1522                                 goto failed0;
1523                         }
1524                 }
1525
1526                 lnet_net_lock(LNET_LOCK_EX);
1527                 lnd->lnd_refcount++;
1528                 lnet_net_unlock(LNET_LOCK_EX);
1529
1530                 net->net_lnd = lnd;
1531
1532                 mutex_unlock(&the_lnet.ln_lnd_mutex);
1533
1534                 net_l = net;
1535         }
1536
1537         /*
1538          * net_l: if the network being added is unique then net_l
1539          *        will point to that network
1540          *        if the network being added is not unique then
1541          *        net_l points to the existing network.
1542          *
1543          * When we enter the loop below, we'll pick NIs off he
1544          * network beign added and start them up, then add them to
1545          * a local ni list. Once we've successfully started all
1546          * the NIs then we join the local NI list (of started up
1547          * networks) with the net_l->net_ni_list, which should
1548          * point to the correct network to add the new ni list to
1549          *
1550          * If any of the new NIs fail to start up, then we want to
1551          * iterate through the local ni list, which should include
1552          * any NIs which were successfully started up, and shut
1553          * them down.
1554          *
1555          * After than we want to delete the network being added,
1556          * to avoid a memory leak.
1557          */
1558
1559         /*
1560          * When a network uses TCP bonding then all its interfaces
1561          * must be specified when the network is first defined: the
1562          * TCP bonding code doesn't allow for interfaces to be added
1563          * or removed.
1564          */
1565         if (net_l != net && net_l != NULL && use_tcp_bonding &&
1566             LNET_NETTYP(net_l->net_id) == SOCKLND) {
1567                 rc = -EINVAL;
1568                 goto failed0;
1569         }
1570
1571         while (!list_empty(&net->net_ni_added)) {
1572                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
1573                                 ni_netlist);
1574                 list_del_init(&ni->ni_netlist);
1575
1576                 /* make sure that the the NI we're about to start
1577                  * up is actually unique. if it's not fail. */
1578                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
1579                                         ni->ni_interfaces[0])) {
1580                         rc = -EINVAL;
1581                         goto failed1;
1582                 }
1583
1584                 /* adjust the pointer the parent network, just in case it
1585                  * the net is a duplicate */
1586                 ni->ni_net = net_l;
1587
1588                 rc = lnet_startup_lndni(ni, tun);
1589
1590                 LASSERT(ni->ni_net->net_tunables.lct_peer_timeout <= 0 ||
1591                         ni->ni_net->net_lnd->lnd_query != NULL);
1592
1593                 if (rc < 0)
1594                         goto failed1;
1595
1596                 lnet_ni_addref(ni);
1597                 list_add_tail(&ni->ni_netlist, &local_ni_list);
1598
1599                 ni_count++;
1600         }
1601
1602         lnet_net_lock(LNET_LOCK_EX);
1603         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
1604         lnet_incr_dlc_seq();
1605         lnet_net_unlock(LNET_LOCK_EX);
1606
1607         /* if the network is not unique then we don't want to keep
1608          * it around after we're done. Free it. Otherwise add that
1609          * net to the global the_lnet.ln_nets */
1610         if (net_l != net && net_l != NULL) {
1611                 /*
1612                  * TODO - note. currently the tunables can not be updated
1613                  * once added
1614                  */
1615                 lnet_net_free(net);
1616         } else {
1617                 net->net_state = LNET_NET_STATE_ACTIVE;
1618                 /*
1619                  * restore tunables after it has been overwitten by the
1620                  * lnd
1621                  */
1622                 if (peer_timeout != -1)
1623                         net->net_tunables.lct_peer_timeout = peer_timeout;
1624                 if (maxtxcredits != -1)
1625                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
1626                 if (peerrtrcredits != -1)
1627                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
1628
1629                 lnet_net_lock(LNET_LOCK_EX);
1630                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
1631                 lnet_net_unlock(LNET_LOCK_EX);
1632         }
1633
1634         return ni_count;
1635
1636 failed1:
1637         /*
1638          * shutdown the new NIs that are being started up
1639          * free the NET being started
1640          */
1641         while (!list_empty(&local_ni_list)) {
1642                 ni = list_entry(local_ni_list.next, struct lnet_ni,
1643                                 ni_netlist);
1644
1645                 lnet_shutdown_lndni(ni);
1646         }
1647
1648 failed0:
1649         lnet_net_free(net);
1650
1651         return rc;
1652 }
1653
1654 static int
1655 lnet_startup_lndnets(struct list_head *netlist)
1656 {
1657         struct lnet_net         *net;
1658         int                     rc;
1659         int                     ni_count = 0;
1660
1661         while (!list_empty(netlist)) {
1662                 net = list_entry(netlist->next, struct lnet_net, net_list);
1663                 list_del_init(&net->net_list);
1664
1665                 rc = lnet_startup_lndnet(net, NULL);
1666
1667                 if (rc < 0)
1668                         goto failed;
1669
1670                 ni_count += rc;
1671         }
1672
1673         return ni_count;
1674 failed:
1675         lnet_shutdown_lndnets();
1676
1677         return rc;
1678 }
1679
1680 /**
1681  * Initialize LNet library.
1682  *
1683  * Automatically called at module loading time. Caller has to call
1684  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
1685  * latter returned 0. It must be called exactly once.
1686  *
1687  * \retval 0 on success
1688  * \retval -ve on failures.
1689  */
1690 int lnet_lib_init(void)
1691 {
1692         int rc;
1693
1694         lnet_assert_wire_constants();
1695
1696         memset(&the_lnet, 0, sizeof(the_lnet));
1697
1698         /* refer to global cfs_cpt_table for now */
1699         the_lnet.ln_cpt_table   = cfs_cpt_table;
1700         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1701
1702         LASSERT(the_lnet.ln_cpt_number > 0);
1703         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1704                 /* we are under risk of consuming all lh_cookie */
1705                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
1706                        "please change setting of CPT-table and retry\n",
1707                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1708                 return -E2BIG;
1709         }
1710
1711         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1712                 the_lnet.ln_cpt_bits++;
1713
1714         rc = lnet_create_locks();
1715         if (rc != 0) {
1716                 CERROR("Can't create LNet global locks: %d\n", rc);
1717                 return rc;
1718         }
1719
1720         the_lnet.ln_refcount = 0;
1721         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1722         INIT_LIST_HEAD(&the_lnet.ln_lnds);
1723         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
1724         INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1725         INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1726
1727         /* The hash table size is the number of bits it takes to express the set
1728          * ln_num_routes, minus 1 (better to under estimate than over so we
1729          * don't waste memory). */
1730         if (rnet_htable_size <= 0)
1731                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
1732         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
1733                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
1734         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
1735                                            order_base_2(rnet_htable_size) - 1);
1736
1737         /* All LNDs apart from the LOLND are in separate modules.  They
1738          * register themselves when their module loads, and unregister
1739          * themselves when their module is unloaded. */
1740         lnet_register_lnd(&the_lolnd);
1741         return 0;
1742 }
1743
1744 /**
1745  * Finalize LNet library.
1746  *
1747  * \pre lnet_lib_init() called with success.
1748  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1749  */
1750 void lnet_lib_exit(void)
1751 {
1752         LASSERT(the_lnet.ln_refcount == 0);
1753
1754         while (!list_empty(&the_lnet.ln_lnds))
1755                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1756                                                lnd_t, lnd_list));
1757         lnet_destroy_locks();
1758 }
1759
1760 /**
1761  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1762  *
1763  * Users must call this function at least once before any other functions.
1764  * For each successful call there must be a corresponding call to
1765  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1766  * ignored.
1767  *
1768  * The PID used by LNet may be different from the one requested.
1769  * See LNetGetId().
1770  *
1771  * \param requested_pid PID requested by the caller.
1772  *
1773  * \return >= 0 on success, and < 0 error code on failures.
1774  */
1775 int
1776 LNetNIInit(lnet_pid_t requested_pid)
1777 {
1778         int                     im_a_router = 0;
1779         int                     rc;
1780         int                     ni_count;
1781         struct lnet_ping_info   *pinfo;
1782         lnet_handle_md_t        md_handle;
1783         struct list_head        net_head;
1784         struct lnet_net         *net;
1785
1786         INIT_LIST_HEAD(&net_head);
1787
1788         mutex_lock(&the_lnet.ln_api_mutex);
1789
1790         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1791
1792         if (the_lnet.ln_refcount > 0) {
1793                 rc = the_lnet.ln_refcount++;
1794                 mutex_unlock(&the_lnet.ln_api_mutex);
1795                 return rc;
1796         }
1797
1798         rc = lnet_prepare(requested_pid);
1799         if (rc != 0) {
1800                 mutex_unlock(&the_lnet.ln_api_mutex);
1801                 return rc;
1802         }
1803
1804         /* create a network for Loopback network */
1805         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
1806         if (net == NULL) {
1807                 rc = -ENOMEM;
1808                 goto err_empty_list;
1809         }
1810
1811         /* Add in the loopback NI */
1812         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
1813                 rc = -ENOMEM;
1814                 goto err_empty_list;
1815         }
1816
1817         /* If LNet is being initialized via DLC it is possible
1818          * that the user requests not to load module parameters (ones which
1819          * are supported by DLC) on initialization.  Therefore, make sure not
1820          * to load networks, routes and forwarding from module parameters
1821          * in this case.  On cleanup in case of failure only clean up
1822          * routes if it has been loaded */
1823         if (!the_lnet.ln_nis_from_mod_params) {
1824                 rc = lnet_parse_networks(&net_head, lnet_get_networks(),
1825                                          use_tcp_bonding);
1826                 if (rc < 0)
1827                         goto err_empty_list;
1828         }
1829
1830         ni_count = lnet_startup_lndnets(&net_head);
1831         if (ni_count < 0) {
1832                 rc = ni_count;
1833                 goto err_empty_list;
1834         }
1835
1836         if (!the_lnet.ln_nis_from_mod_params) {
1837                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1838                 if (rc != 0)
1839                         goto err_shutdown_lndnis;
1840
1841                 rc = lnet_check_routes();
1842                 if (rc != 0)
1843                         goto err_destroy_routes;
1844
1845                 rc = lnet_rtrpools_alloc(im_a_router);
1846                 if (rc != 0)
1847                         goto err_destroy_routes;
1848         }
1849
1850         rc = lnet_acceptor_start();
1851         if (rc != 0)
1852                 goto err_destroy_routes;
1853
1854         the_lnet.ln_refcount = 1;
1855         /* Now I may use my own API functions... */
1856
1857         rc = lnet_ping_info_setup(&pinfo, &md_handle, ni_count, true);
1858         if (rc != 0)
1859                 goto err_acceptor_stop;
1860
1861         lnet_ping_target_update(pinfo, md_handle);
1862
1863         rc = lnet_router_checker_start();
1864         if (rc != 0)
1865                 goto err_stop_ping;
1866
1867         lnet_fault_init();
1868         lnet_proc_init();
1869
1870         mutex_unlock(&the_lnet.ln_api_mutex);
1871
1872         return 0;
1873
1874 err_stop_ping:
1875         lnet_ping_target_fini();
1876 err_acceptor_stop:
1877         the_lnet.ln_refcount = 0;
1878         lnet_acceptor_stop();
1879 err_destroy_routes:
1880         if (!the_lnet.ln_nis_from_mod_params)
1881                 lnet_destroy_routes();
1882 err_shutdown_lndnis:
1883         lnet_shutdown_lndnets();
1884 err_empty_list:
1885         lnet_unprepare();
1886         LASSERT(rc < 0);
1887         mutex_unlock(&the_lnet.ln_api_mutex);
1888         while (!list_empty(&net_head)) {
1889                 struct lnet_net *net;
1890
1891                 net = list_entry(net_head.next, struct lnet_net, net_list);
1892                 list_del_init(&net->net_list);
1893                 lnet_net_free(net);
1894         }
1895         return rc;
1896 }
1897 EXPORT_SYMBOL(LNetNIInit);
1898
1899 /**
1900  * Stop LNet interfaces, routing, and forwarding.
1901  *
1902  * Users must call this function once for each successful call to LNetNIInit().
1903  * Once the LNetNIFini() operation has been started, the results of pending
1904  * API operations are undefined.
1905  *
1906  * \return always 0 for current implementation.
1907  */
1908 int
1909 LNetNIFini()
1910 {
1911         mutex_lock(&the_lnet.ln_api_mutex);
1912
1913         LASSERT(the_lnet.ln_refcount > 0);
1914
1915         if (the_lnet.ln_refcount != 1) {
1916                 the_lnet.ln_refcount--;
1917         } else {
1918                 LASSERT(!the_lnet.ln_niinit_self);
1919
1920                 lnet_fault_fini();
1921
1922                 lnet_proc_fini();
1923                 lnet_router_checker_stop();
1924                 lnet_ping_target_fini();
1925
1926                 /* Teardown fns that use my own API functions BEFORE here */
1927                 the_lnet.ln_refcount = 0;
1928
1929                 lnet_acceptor_stop();
1930                 lnet_destroy_routes();
1931                 lnet_shutdown_lndnets();
1932                 lnet_unprepare();
1933         }
1934
1935         mutex_unlock(&the_lnet.ln_api_mutex);
1936         return 0;
1937 }
1938 EXPORT_SYMBOL(LNetNIFini);
1939
1940
1941 static int lnet_handle_dbg_task(struct lnet_ioctl_dbg *dbg,
1942                                 struct lnet_dbg_task_info *dbg_info)
1943 {
1944         switch (dbg->dbg_task) {
1945         case LNET_DBG_INCR_DLC_SEQ:
1946                 lnet_incr_dlc_seq();
1947         }
1948
1949         return 0;
1950 }
1951 /**
1952  * Grabs the ni data from the ni structure and fills the out
1953  * parameters
1954  *
1955  * \param[in] ni network        interface structure
1956  * \param[out] cfg_ni           NI config information
1957  * \param[out] tun              network and LND tunables
1958  */
1959 static void
1960 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
1961                    struct lnet_ioctl_config_lnd_tunables *tun,
1962                    struct lnet_ioctl_element_stats *stats,
1963                    __u32 tun_size)
1964 {
1965         size_t min_size = 0;
1966         int i;
1967
1968         if (!ni || !cfg_ni || !tun)
1969                 return;
1970
1971         if (ni->ni_interfaces[0] != NULL) {
1972                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
1973                         if (ni->ni_interfaces[i] != NULL) {
1974                                 strncpy(cfg_ni->lic_ni_intf[i],
1975                                         ni->ni_interfaces[i],
1976                                         sizeof(cfg_ni->lic_ni_intf[i]));
1977                         }
1978                 }
1979         }
1980
1981         cfg_ni->lic_nid = ni->ni_nid;
1982         cfg_ni->lic_status = ni->ni_status->ns_status;
1983         cfg_ni->lic_tcp_bonding = use_tcp_bonding;
1984         cfg_ni->lic_dev_cpt = ni->dev_cpt;
1985
1986         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
1987
1988         if (stats) {
1989                 stats->send_count = atomic_read(&ni->ni_stats.send_count);
1990                 stats->recv_count = atomic_read(&ni->ni_stats.recv_count);
1991         }
1992
1993         /*
1994          * tun->lt_tun will always be present, but in order to be
1995          * backwards compatible, we need to deal with the cases when
1996          * tun->lt_tun is smaller than what the kernel has, because it
1997          * comes from an older version of a userspace program, then we'll
1998          * need to copy as much information as we have available space.
1999          */
2000         min_size = tun_size - sizeof(tun->lt_cmn);
2001         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
2002
2003         /* copy over the cpts */
2004         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
2005             ni->ni_cpts == NULL)  {
2006                 for (i = 0; i < ni->ni_ncpts; i++)
2007                         cfg_ni->lic_cpts[i] = i;
2008         } else {
2009                 for (i = 0;
2010                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
2011                      i < LNET_MAX_SHOW_NUM_CPT;
2012                      i++)
2013                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
2014         }
2015         cfg_ni->lic_ncpts = ni->ni_ncpts;
2016 }
2017
2018 /**
2019  * NOTE: This is a legacy function left in the code to be backwards
2020  * compatible with older userspace programs. It should eventually be
2021  * removed.
2022  *
2023  * Grabs the ni data from the ni structure and fills the out
2024  * parameters
2025  *
2026  * \param[in] ni network        interface structure
2027  * \param[out] config           config information
2028  */
2029 static void
2030 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
2031                          struct lnet_ioctl_config_data *config)
2032 {
2033         struct lnet_ioctl_net_config *net_config;
2034         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
2035         size_t min_size, tunable_size = 0;
2036         int i;
2037
2038         if (!ni || !config)
2039                 return;
2040
2041         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
2042         if (!net_config)
2043                 return;
2044
2045         BUILD_BUG_ON(ARRAY_SIZE(ni->ni_interfaces) !=
2046                      ARRAY_SIZE(net_config->ni_interfaces));
2047
2048         for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2049                 if (!ni->ni_interfaces[i])
2050                         break;
2051
2052                 strncpy(net_config->ni_interfaces[i],
2053                         ni->ni_interfaces[i],
2054                         sizeof(net_config->ni_interfaces[i]));
2055         }
2056
2057         config->cfg_nid = ni->ni_nid;
2058         config->cfg_config_u.cfg_net.net_peer_timeout =
2059                 ni->ni_net->net_tunables.lct_peer_timeout;
2060         config->cfg_config_u.cfg_net.net_max_tx_credits =
2061                 ni->ni_net->net_tunables.lct_max_tx_credits;
2062         config->cfg_config_u.cfg_net.net_peer_tx_credits =
2063                 ni->ni_net->net_tunables.lct_peer_tx_credits;
2064         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
2065                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
2066
2067         net_config->ni_status = ni->ni_status->ns_status;
2068
2069         if (ni->ni_cpts) {
2070                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
2071
2072                 for (i = 0; i < num_cpts; i++)
2073                         net_config->ni_cpts[i] = ni->ni_cpts[i];
2074
2075                 config->cfg_ncpts = num_cpts;
2076         }
2077
2078         /*
2079          * See if user land tools sent in a newer and larger version
2080          * of struct lnet_tunables than what the kernel uses.
2081          */
2082         min_size = sizeof(*config) + sizeof(*net_config);
2083
2084         if (config->cfg_hdr.ioc_len > min_size)
2085                 tunable_size = config->cfg_hdr.ioc_len - min_size;
2086
2087         /* Don't copy too much data to user space */
2088         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
2089         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
2090
2091         if (lnd_cfg && min_size) {
2092                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
2093                 config->cfg_config_u.cfg_net.net_interface_count = 1;
2094
2095                 /* Tell user land that kernel side has less data */
2096                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
2097                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
2098                         config->cfg_hdr.ioc_len -= min_size;
2099                 }
2100         }
2101 }
2102
2103 struct lnet_ni *
2104 lnet_get_ni_idx_locked(int idx)
2105 {
2106         struct lnet_ni          *ni;
2107         struct lnet_net         *net;
2108
2109         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2110                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2111                         if (idx-- == 0)
2112                                 return ni;
2113                 }
2114         }
2115
2116         return NULL;
2117 }
2118
2119 struct lnet_ni *
2120 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
2121 {
2122         struct lnet_ni          *ni;
2123         struct lnet_net         *net = mynet;
2124
2125         if (prev == NULL) {
2126                 if (net == NULL)
2127                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
2128                                         net_list);
2129                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2130                                 ni_netlist);
2131
2132                 return ni;
2133         }
2134
2135         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
2136                 /* if you reached the end of the ni list and the net is
2137                  * specified, then there are no more nis in that net */
2138                 if (net != NULL)
2139                         return NULL;
2140
2141                 /* we reached the end of this net ni list. move to the
2142                  * next net */
2143                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
2144                         /* no more nets and no more NIs. */
2145                         return NULL;
2146
2147                 /* get the next net */
2148                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
2149                                  net_list);
2150                 /* get the ni on it */
2151                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2152                                 ni_netlist);
2153
2154                 return ni;
2155         }
2156
2157         /* there are more nis left */
2158         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
2159
2160         return ni;
2161 }
2162
2163 int
2164 lnet_get_net_config(struct lnet_ioctl_config_data *config)
2165 {
2166         struct lnet_ni *ni;
2167         int cpt;
2168         int rc = -ENOENT;
2169         int idx = config->cfg_count;
2170
2171         cpt = lnet_net_lock_current();
2172
2173         ni = lnet_get_ni_idx_locked(idx);
2174
2175         if (ni != NULL) {
2176                 rc = 0;
2177                 lnet_ni_lock(ni);
2178                 lnet_fill_ni_info_legacy(ni, config);
2179                 lnet_ni_unlock(ni);
2180         }
2181
2182         lnet_net_unlock(cpt);
2183         return rc;
2184 }
2185
2186 int
2187 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
2188                    struct lnet_ioctl_config_lnd_tunables *tun,
2189                    struct lnet_ioctl_element_stats *stats,
2190                    __u32 tun_size)
2191 {
2192         struct lnet_ni          *ni;
2193         int                     cpt;
2194         int                     rc = -ENOENT;
2195
2196         if (!cfg_ni || !tun || !stats)
2197                 return -EINVAL;
2198
2199         cpt = lnet_net_lock_current();
2200
2201         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
2202
2203         if (ni) {
2204                 rc = 0;
2205                 lnet_ni_lock(ni);
2206                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
2207                 lnet_ni_unlock(ni);
2208         }
2209
2210         lnet_net_unlock(cpt);
2211         return rc;
2212 }
2213
2214 static int lnet_add_net_common(struct lnet_net *net,
2215                                struct lnet_ioctl_config_lnd_tunables *tun)
2216 {
2217         struct lnet_net         *netl = NULL;
2218         __u32                   net_id;
2219         lnet_ping_info_t        *pinfo;
2220         lnet_handle_md_t        md_handle;
2221         int                     rc;
2222         lnet_remotenet_t        *rnet;
2223         int                     net_ni_count;
2224         int                     num_acceptor_nets;
2225
2226         lnet_net_lock(LNET_LOCK_EX);
2227         rnet = lnet_find_rnet_locked(net->net_id);
2228         lnet_net_unlock(LNET_LOCK_EX);
2229         /*
2230          * make sure that the net added doesn't invalidate the current
2231          * configuration LNet is keeping
2232          */
2233         if (rnet) {
2234                 CERROR("Adding net %s will invalidate routing configuration\n",
2235                        libcfs_net2str(net->net_id));
2236                 rc = -EUSERS;
2237                 goto failed1;
2238         }
2239
2240         /*
2241          * make sure you calculate the correct number of slots in the ping
2242          * info. Since the ping info is a flattened list of all the NIs,
2243          * we should allocate enough slots to accomodate the number of NIs
2244          * which will be added.
2245          *
2246          * since ni hasn't been configured yet, use
2247          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
2248          */
2249         net_ni_count = lnet_get_net_ni_count_pre(net);
2250
2251         rc = lnet_ping_info_setup(&pinfo, &md_handle,
2252                                   net_ni_count + lnet_get_ni_count(),
2253                                   false);
2254         if (rc < 0)
2255                 goto failed1;
2256
2257         if (tun)
2258                 memcpy(&net->net_tunables,
2259                        &tun->lt_cmn, sizeof(net->net_tunables));
2260         else
2261                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
2262
2263         /*
2264          * before starting this network get a count of the current TCP
2265          * networks which require the acceptor thread running. If that
2266          * count is == 0 before we start up this network, then we'd want to
2267          * start up the acceptor thread after starting up this network
2268          */
2269         num_acceptor_nets = lnet_count_acceptor_nets();
2270
2271         net_id = net->net_id;
2272
2273         rc = lnet_startup_lndnet(net,
2274                                  (tun) ? &tun->lt_tun : NULL);
2275         if (rc < 0)
2276                 goto failed;
2277
2278         lnet_net_lock(LNET_LOCK_EX);
2279         netl = lnet_get_net_locked(net_id);
2280         lnet_net_unlock(LNET_LOCK_EX);
2281
2282         LASSERT(netl);
2283
2284         /*
2285          * Start the acceptor thread if this is the first network
2286          * being added that requires the thread.
2287          */
2288         if (netl->net_lnd->lnd_accept &&
2289             num_acceptor_nets == 0)
2290         {
2291                 rc = lnet_acceptor_start();
2292                 if (rc < 0) {
2293                         /* shutdown the net that we just started */
2294                         CERROR("Failed to start up acceptor thread\n");
2295                         lnet_shutdown_lndnet(net);
2296                         goto failed;
2297                 }
2298         }
2299
2300         lnet_net_lock(LNET_LOCK_EX);
2301         lnet_peer_net_added(netl);
2302         lnet_net_unlock(LNET_LOCK_EX);
2303
2304         lnet_ping_target_update(pinfo, md_handle);
2305
2306         return 0;
2307
2308 failed:
2309         lnet_ping_md_unlink(pinfo, &md_handle);
2310         lnet_ping_info_free(pinfo);
2311 failed1:
2312         lnet_net_free(net);
2313         return rc;
2314 }
2315
2316 static int lnet_handle_legacy_ip2nets(char *ip2nets,
2317                                       struct lnet_ioctl_config_lnd_tunables *tun)
2318 {
2319         struct lnet_net *net;
2320         char *nets;
2321         int rc;
2322         struct list_head net_head;
2323
2324         INIT_LIST_HEAD(&net_head);
2325
2326         rc = lnet_parse_ip2nets(&nets, ip2nets);
2327         if (rc < 0)
2328                 return rc;
2329
2330         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
2331         if (rc < 0)
2332                 return rc;
2333
2334         mutex_lock(&the_lnet.ln_api_mutex);
2335         while (!list_empty(&net_head)) {
2336                 net = list_entry(net_head.next, struct lnet_net, net_list);
2337                 list_del_init(&net->net_list);
2338                 rc = lnet_add_net_common(net, tun);
2339                 if (rc < 0)
2340                         goto out;
2341         }
2342
2343 out:
2344         mutex_unlock(&the_lnet.ln_api_mutex);
2345
2346         while (!list_empty(&net_head)) {
2347                 net = list_entry(net_head.next, struct lnet_net, net_list);
2348                 list_del_init(&net->net_list);
2349                 lnet_net_free(net);
2350         }
2351         return rc;
2352 }
2353
2354 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
2355 {
2356         struct lnet_net *net;
2357         struct lnet_ni *ni;
2358         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
2359         int rc, i;
2360         __u32 net_id;
2361
2362         /* get the tunables if they are available */
2363         if (conf->lic_cfg_hdr.ioc_len >=
2364             sizeof(*conf) + sizeof(*tun))
2365                 tun = (struct lnet_ioctl_config_lnd_tunables *)
2366                         conf->lic_bulk;
2367
2368         /* handle legacy ip2nets from DLC */
2369         if (conf->lic_legacy_ip2nets[0] != '\0')
2370                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
2371                                                   tun);
2372
2373         net_id = LNET_NIDNET(conf->lic_nid);
2374
2375         net = lnet_net_alloc(net_id, NULL);
2376         if (!net)
2377                 return -ENOMEM;
2378
2379         for (i = 0; i < conf->lic_ncpts; i++) {
2380                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
2381                         return -EINVAL;
2382         }
2383
2384         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
2385                                        conf->lic_ni_intf[0]);
2386         if (!ni)
2387                 return -ENOMEM;
2388
2389         mutex_lock(&the_lnet.ln_api_mutex);
2390
2391         rc = lnet_add_net_common(net, tun);
2392
2393         mutex_unlock(&the_lnet.ln_api_mutex);
2394
2395         return rc;
2396 }
2397
2398 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
2399 {
2400         struct lnet_net  *net;
2401         struct lnet_ni *ni;
2402         __u32 net_id = LNET_NIDNET(conf->lic_nid);
2403         lnet_ping_info_t *pinfo;
2404         lnet_handle_md_t  md_handle;
2405         int               rc;
2406         int               net_count;
2407         __u32             addr;
2408
2409         /* don't allow userspace to shutdown the LOLND */
2410         if (LNET_NETTYP(net_id) == LOLND)
2411                 return -EINVAL;
2412
2413         mutex_lock(&the_lnet.ln_api_mutex);
2414
2415         lnet_net_lock(0);
2416
2417         net = lnet_get_net_locked(net_id);
2418         if (!net) {
2419                 CERROR("net %s not found\n",
2420                        libcfs_net2str(net_id));
2421                 rc = -ENOENT;
2422                 goto net_unlock;
2423         }
2424
2425         addr = LNET_NIDADDR(conf->lic_nid);
2426         if (addr == 0) {
2427                 /* remove the entire net */
2428                 net_count = lnet_get_net_ni_count_locked(net);
2429
2430                 lnet_net_unlock(0);
2431
2432                 /* create and link a new ping info, before removing the old one */
2433                 rc = lnet_ping_info_setup(&pinfo, &md_handle,
2434                                         lnet_get_ni_count() - net_count,
2435                                         false);
2436                 if (rc != 0)
2437                         goto out;
2438
2439                 lnet_shutdown_lndnet(net);
2440
2441                 if (lnet_count_acceptor_nets() == 0)
2442                         lnet_acceptor_stop();
2443
2444                 lnet_ping_target_update(pinfo, md_handle);
2445
2446                 goto out;
2447         }
2448
2449         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
2450         if (!ni) {
2451                 CERROR("nid %s not found \n",
2452                        libcfs_nid2str(conf->lic_nid));
2453                 rc = -ENOENT;
2454                 goto net_unlock;
2455         }
2456
2457         net_count = lnet_get_net_ni_count_locked(net);
2458
2459         lnet_net_unlock(0);
2460
2461         /* create and link a new ping info, before removing the old one */
2462         rc = lnet_ping_info_setup(&pinfo, &md_handle,
2463                                   lnet_get_ni_count() - 1, false);
2464         if (rc != 0)
2465                 goto out;
2466
2467         lnet_shutdown_lndni(ni);
2468
2469         if (lnet_count_acceptor_nets() == 0)
2470                 lnet_acceptor_stop();
2471
2472         lnet_ping_target_update(pinfo, md_handle);
2473
2474         /* check if the net is empty and remove it if it is */
2475         if (net_count == 1)
2476                 lnet_shutdown_lndnet(net);
2477
2478         goto out;
2479
2480 net_unlock:
2481         lnet_net_unlock(0);
2482 out:
2483         mutex_unlock(&the_lnet.ln_api_mutex);
2484
2485         return rc;
2486 }
2487
2488 /*
2489  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
2490  * They are only expected to be called for unique networks.
2491  * That can be as a result of older DLC library
2492  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
2493  */
2494 int
2495 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
2496 {
2497         struct lnet_net         *net;
2498         struct list_head        net_head;
2499         int                     rc;
2500         struct lnet_ioctl_config_lnd_tunables tun;
2501         char *nets = conf->cfg_config_u.cfg_net.net_intf;
2502
2503         INIT_LIST_HEAD(&net_head);
2504
2505         /* Create a net/ni structures for the network string */
2506         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
2507         if (rc <= 0)
2508                 return rc == 0 ? -EINVAL : rc;
2509
2510         mutex_lock(&the_lnet.ln_api_mutex);
2511
2512         if (rc > 1) {
2513                 rc = -EINVAL; /* only add one network per call */
2514                 goto failed;
2515         }
2516
2517         net = list_entry(net_head.next, struct lnet_net, net_list);
2518         list_del_init(&net->net_list);
2519
2520         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
2521
2522         memset(&tun, sizeof(tun), 0);
2523
2524         tun.lt_cmn.lct_peer_timeout =
2525           conf->cfg_config_u.cfg_net.net_peer_timeout;
2526         tun.lt_cmn.lct_peer_tx_credits =
2527           conf->cfg_config_u.cfg_net.net_peer_tx_credits;
2528         tun.lt_cmn.lct_peer_rtr_credits =
2529           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
2530         tun.lt_cmn.lct_max_tx_credits =
2531           conf->cfg_config_u.cfg_net.net_max_tx_credits;
2532
2533         rc = lnet_add_net_common(net, &tun);
2534         if (rc != 0)
2535                 goto failed;
2536
2537         return 0;
2538
2539 failed:
2540         mutex_unlock(&the_lnet.ln_api_mutex);
2541         while (!list_empty(&net_head)) {
2542                 net = list_entry(net_head.next, struct lnet_net, net_list);
2543                 list_del_init(&net->net_list);
2544                 lnet_net_free(net);
2545         }
2546         return rc;
2547 }
2548
2549 int
2550 lnet_dyn_del_net(__u32 net_id)
2551 {
2552         struct lnet_net  *net;
2553         struct lnet_ping_info *pinfo;
2554         lnet_handle_md_t  md_handle;
2555         int               rc;
2556         int               net_ni_count;
2557
2558         /* don't allow userspace to shutdown the LOLND */
2559         if (LNET_NETTYP(net_id) == LOLND)
2560                 return -EINVAL;
2561
2562         mutex_lock(&the_lnet.ln_api_mutex);
2563
2564         lnet_net_lock(0);
2565
2566         net = lnet_get_net_locked(net_id);
2567         if (net == NULL) {
2568                 rc = -EINVAL;
2569                 goto out;
2570         }
2571
2572         net_ni_count = lnet_get_net_ni_count_locked(net);
2573
2574         lnet_net_unlock(0);
2575
2576         /* create and link a new ping info, before removing the old one */
2577         rc = lnet_ping_info_setup(&pinfo, &md_handle,
2578                                   lnet_get_ni_count() - net_ni_count, false);
2579         if (rc != 0)
2580                 goto out;
2581
2582         lnet_shutdown_lndnet(net);
2583
2584         if (lnet_count_acceptor_nets() == 0)
2585                 lnet_acceptor_stop();
2586
2587         lnet_ping_target_update(pinfo, md_handle);
2588
2589 out:
2590         mutex_unlock(&the_lnet.ln_api_mutex);
2591
2592         return rc;
2593 }
2594
2595 void lnet_incr_dlc_seq(void)
2596 {
2597         atomic_inc(&lnet_dlc_seq_no);
2598 }
2599
2600 __u32 lnet_get_dlc_seq_locked(void)
2601 {
2602         return atomic_read(&lnet_dlc_seq_no);
2603 }
2604
2605 /**
2606  * LNet ioctl handler.
2607  *
2608  */
2609 int
2610 LNetCtl(unsigned int cmd, void *arg)
2611 {
2612         struct libcfs_ioctl_data *data = arg;
2613         struct lnet_ioctl_config_data *config;
2614         lnet_process_id_t         id = {0};
2615         lnet_ni_t                *ni;
2616         int                       rc;
2617
2618         BUILD_BUG_ON(sizeof(struct lnet_ioctl_net_config) +
2619                      sizeof(struct lnet_ioctl_config_data) > LIBCFS_IOC_DATA_MAX);
2620
2621         switch (cmd) {
2622         case IOC_LIBCFS_GET_NI:
2623                 rc = LNetGetId(data->ioc_count, &id);
2624                 data->ioc_nid = id.nid;
2625                 return rc;
2626
2627         case IOC_LIBCFS_FAIL_NID:
2628                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
2629
2630         case IOC_LIBCFS_ADD_ROUTE:
2631                 config = arg;
2632
2633                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2634                         return -EINVAL;
2635
2636                 mutex_lock(&the_lnet.ln_api_mutex);
2637                 rc = lnet_add_route(config->cfg_net,
2638                                     config->cfg_config_u.cfg_route.rtr_hop,
2639                                     config->cfg_nid,
2640                                     config->cfg_config_u.cfg_route.
2641                                         rtr_priority);
2642                 if (rc == 0) {
2643                         rc = lnet_check_routes();
2644                         if (rc != 0)
2645                                 lnet_del_route(config->cfg_net,
2646                                                config->cfg_nid);
2647                 }
2648                 mutex_unlock(&the_lnet.ln_api_mutex);
2649                 return rc;
2650
2651         case IOC_LIBCFS_DEL_ROUTE:
2652                 config = arg;
2653
2654                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2655                         return -EINVAL;
2656
2657                 mutex_lock(&the_lnet.ln_api_mutex);
2658                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
2659                 mutex_unlock(&the_lnet.ln_api_mutex);
2660                 return rc;
2661
2662         case IOC_LIBCFS_GET_ROUTE:
2663                 config = arg;
2664
2665                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2666                         return -EINVAL;
2667
2668                 mutex_lock(&the_lnet.ln_api_mutex);
2669                 rc = lnet_get_route(config->cfg_count,
2670                                     &config->cfg_net,
2671                                     &config->cfg_config_u.cfg_route.rtr_hop,
2672                                     &config->cfg_nid,
2673                                     &config->cfg_config_u.cfg_route.rtr_flags,
2674                                     &config->cfg_config_u.cfg_route.
2675                                         rtr_priority);
2676                 mutex_unlock(&the_lnet.ln_api_mutex);
2677                 return rc;
2678
2679         case IOC_LIBCFS_GET_LOCAL_NI: {
2680                 struct lnet_ioctl_config_ni *cfg_ni;
2681                 struct lnet_ioctl_config_lnd_tunables *tun = NULL;
2682                 struct lnet_ioctl_element_stats *stats;
2683                 __u32 tun_size;
2684
2685                 cfg_ni = arg;
2686                 /* get the tunables if they are available */
2687                 if (cfg_ni->lic_cfg_hdr.ioc_len <
2688                     sizeof(*cfg_ni) + sizeof(*stats)+ sizeof(*tun))
2689                         return -EINVAL;
2690
2691                 stats = (struct lnet_ioctl_element_stats *)
2692                         cfg_ni->lic_bulk;
2693                 tun = (struct lnet_ioctl_config_lnd_tunables *)
2694                                 (cfg_ni->lic_bulk + sizeof(*stats));
2695
2696                 tun_size = cfg_ni->lic_cfg_hdr.ioc_len - sizeof(*cfg_ni) -
2697                         sizeof(*stats);
2698
2699                 mutex_lock(&the_lnet.ln_api_mutex);
2700                 rc = lnet_get_ni_config(cfg_ni, tun, stats, tun_size);
2701                 mutex_unlock(&the_lnet.ln_api_mutex);
2702                 return rc;
2703         }
2704
2705         case IOC_LIBCFS_GET_NET: {
2706                 size_t total = sizeof(*config) +
2707                                sizeof(struct lnet_ioctl_net_config);
2708                 config = arg;
2709
2710                 if (config->cfg_hdr.ioc_len < total)
2711                         return -EINVAL;
2712
2713                 mutex_lock(&the_lnet.ln_api_mutex);
2714                 rc = lnet_get_net_config(config);
2715                 mutex_unlock(&the_lnet.ln_api_mutex);
2716                 return rc;
2717         }
2718
2719         case IOC_LIBCFS_GET_LNET_STATS:
2720         {
2721                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
2722
2723                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
2724                         return -EINVAL;
2725
2726                 mutex_lock(&the_lnet.ln_api_mutex);
2727                 lnet_counters_get(&lnet_stats->st_cntrs);
2728                 mutex_unlock(&the_lnet.ln_api_mutex);
2729                 return 0;
2730         }
2731
2732         case IOC_LIBCFS_CONFIG_RTR:
2733                 config = arg;
2734
2735                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2736                         return -EINVAL;
2737
2738                 mutex_lock(&the_lnet.ln_api_mutex);
2739                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
2740                         rc = lnet_rtrpools_enable();
2741                         mutex_unlock(&the_lnet.ln_api_mutex);
2742                         return rc;
2743                 }
2744                 lnet_rtrpools_disable();
2745                 mutex_unlock(&the_lnet.ln_api_mutex);
2746                 return 0;
2747
2748         case IOC_LIBCFS_ADD_BUF:
2749                 config = arg;
2750
2751                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2752                         return -EINVAL;
2753
2754                 mutex_lock(&the_lnet.ln_api_mutex);
2755                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
2756                                                 buf_tiny,
2757                                           config->cfg_config_u.cfg_buffers.
2758                                                 buf_small,
2759                                           config->cfg_config_u.cfg_buffers.
2760                                                 buf_large);
2761                 mutex_unlock(&the_lnet.ln_api_mutex);
2762                 return rc;
2763
2764         case IOC_LIBCFS_SET_NUMA_RANGE: {
2765                 struct lnet_ioctl_numa_range *numa;
2766                 numa = arg;
2767                 if (numa->nr_hdr.ioc_len != sizeof(*numa))
2768                         return -EINVAL;
2769                 mutex_lock(&the_lnet.ln_api_mutex);
2770                 lnet_numa_range = numa->nr_range;
2771                 mutex_unlock(&the_lnet.ln_api_mutex);
2772                 return 0;
2773         }
2774
2775         case IOC_LIBCFS_GET_NUMA_RANGE: {
2776                 struct lnet_ioctl_numa_range *numa;
2777                 numa = arg;
2778                 if (numa->nr_hdr.ioc_len != sizeof(*numa))
2779                         return -EINVAL;
2780                 numa->nr_range = lnet_numa_range;
2781                 return 0;
2782         }
2783
2784         case IOC_LIBCFS_GET_BUF: {
2785                 struct lnet_ioctl_pool_cfg *pool_cfg;
2786                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
2787
2788                 config = arg;
2789
2790                 if (config->cfg_hdr.ioc_len < total)
2791                         return -EINVAL;
2792
2793                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
2794
2795                 mutex_lock(&the_lnet.ln_api_mutex);
2796                 rc = lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
2797                 mutex_unlock(&the_lnet.ln_api_mutex);
2798                 return rc;
2799         }
2800
2801         case IOC_LIBCFS_ADD_PEER_NI: {
2802                 struct lnet_ioctl_peer_cfg *cfg = arg;
2803
2804                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
2805                         return -EINVAL;
2806
2807                 mutex_lock(&the_lnet.ln_api_mutex);
2808                 rc = lnet_add_peer_ni_to_peer(cfg->prcfg_key_nid,
2809                                               cfg->prcfg_cfg_nid,
2810                                               cfg->prcfg_mr);
2811                 mutex_unlock(&the_lnet.ln_api_mutex);
2812                 return rc;
2813         }
2814
2815         case IOC_LIBCFS_DEL_PEER_NI: {
2816                 struct lnet_ioctl_peer_cfg *cfg = arg;
2817
2818                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
2819                         return -EINVAL;
2820
2821                 mutex_lock(&the_lnet.ln_api_mutex);
2822                 rc = lnet_del_peer_ni_from_peer(cfg->prcfg_key_nid,
2823                                                 cfg->prcfg_cfg_nid);
2824                 mutex_unlock(&the_lnet.ln_api_mutex);
2825                 return rc;
2826         }
2827
2828         case IOC_LIBCFS_GET_PEER_INFO: {
2829                 struct lnet_ioctl_peer *peer_info = arg;
2830
2831                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
2832                         return -EINVAL;
2833
2834                 mutex_lock(&the_lnet.ln_api_mutex);
2835                 rc = lnet_get_peer_ni_info(
2836                    peer_info->pr_count,
2837                    &peer_info->pr_nid,
2838                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
2839                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
2840                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
2841                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
2842                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
2843                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
2844                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_tx_credits,
2845                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
2846                 mutex_unlock(&the_lnet.ln_api_mutex);
2847                 return rc;
2848         }
2849
2850         case IOC_LIBCFS_GET_PEER_NI: {
2851                 struct lnet_ioctl_peer_cfg *cfg = arg;
2852                 struct lnet_peer_ni_credit_info *lpni_cri;
2853                 struct lnet_ioctl_element_stats *lpni_stats;
2854                 size_t total = sizeof(*cfg) + sizeof(*lpni_cri) +
2855                                sizeof(*lpni_stats);
2856
2857                 if (cfg->prcfg_hdr.ioc_len < total)
2858                         return -EINVAL;
2859
2860                 lpni_cri = (struct lnet_peer_ni_credit_info*) cfg->prcfg_bulk;
2861                 lpni_stats = (struct lnet_ioctl_element_stats *)
2862                              (cfg->prcfg_bulk + sizeof(*lpni_cri));
2863
2864                 mutex_lock(&the_lnet.ln_api_mutex);
2865                 rc = lnet_get_peer_info(cfg->prcfg_idx, &cfg->prcfg_key_nid,
2866                                         &cfg->prcfg_cfg_nid, &cfg->prcfg_mr,
2867                                         lpni_cri, lpni_stats);
2868                 mutex_unlock(&the_lnet.ln_api_mutex);
2869                 return rc;
2870         }
2871
2872         case IOC_LIBCFS_NOTIFY_ROUTER: {
2873                 unsigned long jiffies_passed;
2874
2875                 jiffies_passed = ktime_get_real_seconds() - data->ioc_u64[0];
2876                 jiffies_passed = cfs_time_seconds(jiffies_passed);
2877
2878                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
2879                                    jiffies - jiffies_passed);
2880         }
2881
2882         case IOC_LIBCFS_LNET_DIST:
2883                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
2884                 if (rc < 0 && rc != -EHOSTUNREACH)
2885                         return rc;
2886
2887                 data->ioc_u32[0] = rc;
2888                 return 0;
2889
2890         case IOC_LIBCFS_TESTPROTOCOMPAT:
2891                 lnet_net_lock(LNET_LOCK_EX);
2892                 the_lnet.ln_testprotocompat = data->ioc_flags;
2893                 lnet_net_unlock(LNET_LOCK_EX);
2894                 return 0;
2895
2896         case IOC_LIBCFS_LNET_FAULT:
2897                 return lnet_fault_ctl(data->ioc_flags, data);
2898
2899         case IOC_LIBCFS_PING: {
2900                 signed long timeout;
2901
2902                 id.nid = data->ioc_nid;
2903                 id.pid = data->ioc_u32[0];
2904
2905                 /* Don't block longer than 2 minutes */
2906                 if (data->ioc_u32[1] > 120 * MSEC_PER_SEC)
2907                         return -EINVAL;
2908
2909                 /* If timestamp is negative then disable timeout */
2910                 if ((s32)data->ioc_u32[1] < 0)
2911                         timeout = MAX_SCHEDULE_TIMEOUT;
2912                 else
2913                         timeout = msecs_to_jiffies(data->ioc_u32[1]);
2914
2915                 rc = lnet_ping(id, timeout, data->ioc_pbuf1,
2916                                data->ioc_plen1 / sizeof(lnet_process_id_t));
2917                 if (rc < 0)
2918                         return rc;
2919                 data->ioc_count = rc;
2920                 return 0;
2921         }
2922
2923         case IOC_LIBCFS_DBG: {
2924                 struct lnet_ioctl_dbg *dbg = arg;
2925                 struct lnet_dbg_task_info *dbg_info;
2926                 size_t total = sizeof(*dbg) + sizeof(*dbg_info);
2927
2928                 if (dbg->dbg_hdr.ioc_len < total)
2929                         return -EINVAL;
2930
2931                 dbg_info = (struct lnet_dbg_task_info*) dbg->dbg_bulk;
2932
2933                 return lnet_handle_dbg_task(dbg, dbg_info);
2934         }
2935
2936         default:
2937                 ni = lnet_net2ni_addref(data->ioc_net);
2938                 if (ni == NULL)
2939                         return -EINVAL;
2940
2941                 if (ni->ni_net->net_lnd->lnd_ctl == NULL)
2942                         rc = -EINVAL;
2943                 else
2944                         rc = ni->ni_net->net_lnd->lnd_ctl(ni, cmd, arg);
2945
2946                 lnet_ni_decref(ni);
2947                 return rc;
2948         }
2949         /* not reached */
2950 }
2951 EXPORT_SYMBOL(LNetCtl);
2952
2953 void LNetDebugPeer(lnet_process_id_t id)
2954 {
2955         lnet_debug_peer(id.nid);
2956 }
2957 EXPORT_SYMBOL(LNetDebugPeer);
2958
2959 /**
2960  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
2961  * all interfaces share a same PID, as requested by LNetNIInit().
2962  *
2963  * \param index Index of the interface to look up.
2964  * \param id On successful return, this location will hold the
2965  * lnet_process_id_t ID of the interface.
2966  *
2967  * \retval 0 If an interface exists at \a index.
2968  * \retval -ENOENT If no interface has been found.
2969  */
2970 int
2971 LNetGetId(unsigned int index, lnet_process_id_t *id)
2972 {
2973         struct lnet_ni   *ni;
2974         struct lnet_net  *net;
2975         int               cpt;
2976         int               rc = -ENOENT;
2977
2978         LASSERT(the_lnet.ln_refcount > 0);
2979
2980         cpt = lnet_net_lock_current();
2981
2982         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2983                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2984                         if (index-- != 0)
2985                                 continue;
2986
2987                         id->nid = ni->ni_nid;
2988                         id->pid = the_lnet.ln_pid;
2989                         rc = 0;
2990                         break;
2991                 }
2992         }
2993
2994         lnet_net_unlock(cpt);
2995         return rc;
2996 }
2997 EXPORT_SYMBOL(LNetGetId);
2998
2999 /**
3000  * Print a string representation of handle \a h into buffer \a str of
3001  * \a len bytes.
3002  */
3003 void
3004 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
3005 {
3006         snprintf(str, len, "%#llx", h.cookie);
3007 }
3008 EXPORT_SYMBOL(LNetSnprintHandle);
3009
3010 static int lnet_ping(lnet_process_id_t id, signed long timeout,
3011                      lnet_process_id_t __user *ids, int n_ids)
3012 {
3013         lnet_handle_eq_t     eqh;
3014         lnet_handle_md_t     mdh;
3015         lnet_event_t         event;
3016         lnet_md_t            md = { NULL };
3017         int                  which;
3018         int                  unlinked = 0;
3019         int                  replied = 0;
3020         const signed long a_long_time = msecs_to_jiffies(60 * MSEC_PER_SEC);
3021         int                  infosz;
3022         struct lnet_ping_info    *info;
3023         lnet_process_id_t    tmpid;
3024         int                  i;
3025         int                  nob;
3026         int                  rc;
3027         int                  rc2;
3028         sigset_t         blocked;
3029
3030         infosz = offsetof(struct lnet_ping_info, pi_ni[n_ids]);
3031
3032         /* n_ids limit is arbitrary */
3033         if (n_ids <= 0 || n_ids > 20 || id.nid == LNET_NID_ANY)
3034                 return -EINVAL;
3035
3036         if (id.pid == LNET_PID_ANY)
3037                 id.pid = LNET_PID_LUSTRE;
3038
3039         LIBCFS_ALLOC(info, infosz);
3040         if (info == NULL)
3041                 return -ENOMEM;
3042
3043         /* NB 2 events max (including any unlink event) */
3044         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
3045         if (rc != 0) {
3046                 CERROR("Can't allocate EQ: %d\n", rc);
3047                 goto out_0;
3048         }
3049
3050         /* initialize md content */
3051         md.start     = info;
3052         md.length    = infosz;
3053         md.threshold = 2; /*GET/REPLY*/
3054         md.max_size  = 0;
3055         md.options   = LNET_MD_TRUNCATE;
3056         md.user_ptr  = NULL;
3057         md.eq_handle = eqh;
3058
3059         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
3060         if (rc != 0) {
3061                 CERROR("Can't bind MD: %d\n", rc);
3062                 goto out_1;
3063         }
3064
3065         rc = LNetGet(LNET_NID_ANY, mdh, id,
3066                      LNET_RESERVED_PORTAL,
3067                      LNET_PROTO_PING_MATCHBITS, 0);
3068
3069         if (rc != 0) {
3070                 /* Don't CERROR; this could be deliberate! */
3071
3072                 rc2 = LNetMDUnlink(mdh);
3073                 LASSERT(rc2 == 0);
3074
3075                 /* NB must wait for the UNLINK event below... */
3076                 unlinked = 1;
3077                 timeout = a_long_time;
3078         }
3079
3080         do {
3081                 /* MUST block for unlink to complete */
3082                 if (unlinked)
3083                         blocked = cfs_block_allsigs();
3084
3085                 rc2 = LNetEQPoll(&eqh, 1, timeout, &event, &which);
3086
3087                 if (unlinked)
3088                         cfs_restore_sigs(blocked);
3089
3090                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
3091                        (rc2 <= 0) ? -1 : event.type,
3092                        (rc2 <= 0) ? -1 : event.status,
3093                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
3094
3095                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
3096
3097                 if (rc2 <= 0 || event.status != 0) {
3098                         /* timeout or error */
3099                         if (!replied && rc == 0)
3100                                 rc = (rc2 < 0) ? rc2 :
3101                                      (rc2 == 0) ? -ETIMEDOUT :
3102                                      event.status;
3103
3104                         if (!unlinked) {
3105                                 /* Ensure completion in finite time... */
3106                                 LNetMDUnlink(mdh);
3107                                 /* No assertion (racing with network) */
3108                                 unlinked = 1;
3109                                 timeout = a_long_time;
3110                         } else if (rc2 == 0) {
3111                                 /* timed out waiting for unlink */
3112                                 CWARN("ping %s: late network completion\n",
3113                                       libcfs_id2str(id));
3114                         }
3115                 } else if (event.type == LNET_EVENT_REPLY) {
3116                         replied = 1;
3117                         rc = event.mlength;
3118                 }
3119
3120         } while (rc2 <= 0 || !event.unlinked);
3121
3122         if (!replied) {
3123                 if (rc >= 0)
3124                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
3125                               libcfs_id2str(id));
3126                 rc = -EIO;
3127                 goto out_1;
3128         }
3129
3130         nob = rc;
3131         LASSERT(nob >= 0 && nob <= infosz);
3132
3133         rc = -EPROTO;                           /* if I can't parse... */
3134
3135         if (nob < 8) {
3136                 /* can't check magic/version */
3137                 CERROR("%s: ping info too short %d\n",
3138                        libcfs_id2str(id), nob);
3139                 goto out_1;
3140         }
3141
3142         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
3143                 lnet_swap_pinginfo(info);
3144         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
3145                 CERROR("%s: Unexpected magic %08x\n",
3146                        libcfs_id2str(id), info->pi_magic);
3147                 goto out_1;
3148         }
3149
3150         if ((info->pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
3151                 CERROR("%s: ping w/o NI status: 0x%x\n",
3152                        libcfs_id2str(id), info->pi_features);
3153                 goto out_1;
3154         }
3155
3156         if (nob < offsetof(struct lnet_ping_info, pi_ni[0])) {
3157                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
3158                        nob, (int)offsetof(struct lnet_ping_info, pi_ni[0]));
3159                 goto out_1;
3160         }
3161
3162         if (info->pi_nnis < n_ids)
3163                 n_ids = info->pi_nnis;
3164
3165         if (nob < offsetof(struct lnet_ping_info, pi_ni[n_ids])) {
3166                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
3167                        nob, (int)offsetof(struct lnet_ping_info, pi_ni[n_ids]));
3168                 goto out_1;
3169         }
3170
3171         rc = -EFAULT;                           /* If I SEGV... */
3172
3173         memset(&tmpid, 0, sizeof(tmpid));
3174         for (i = 0; i < n_ids; i++) {
3175                 tmpid.pid = info->pi_pid;
3176                 tmpid.nid = info->pi_ni[i].ns_nid;
3177                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
3178                         goto out_1;
3179         }
3180         rc = info->pi_nnis;
3181
3182  out_1:
3183         rc2 = LNetEQFree(eqh);
3184         if (rc2 != 0)
3185                 CERROR("rc2 %d\n", rc2);
3186         LASSERT(rc2 == 0);
3187
3188  out_0:
3189         LIBCFS_FREE(info, infosz);
3190         return rc;
3191 }