lnet/lnet/api-ni.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LNET
  38 #include <linux/log2.h>
  39 #include <lnet/lib-lnet.h>
  40
  41 #define D_LNI D_CONSOLE
  42
  43 lnet_t      the_lnet;                           /* THE state of the network */
  44 EXPORT_SYMBOL(the_lnet);
  45
  46 static char *ip2nets = "";
  47 CFS_MODULE_PARM(ip2nets, "s", charp, 0444,
  48                 "LNET network <- IP table");
  49
  50 static char *networks = "";
  51 CFS_MODULE_PARM(networks, "s", charp, 0444,
  52                 "local networks");
  53
  54 static char *routes = "";
  55 CFS_MODULE_PARM(routes, "s", charp, 0444,
  56                 "routes to non-local networks");
  57
  58 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  59 CFS_MODULE_PARM(rnet_htable_size, "i", int, 0444,
  60                 "size of remote network hash table");
  61
  62 static int lnet_ping(lnet_process_id_t id, int timeout_ms,
  63                      lnet_process_id_t __user *ids, int n_ids);
  64
  65 static char *
  66 lnet_get_routes(void)
  67 {
  68         return routes;
  69 }
  70
  71 static char *
  72 lnet_get_networks(void)
  73 {
  74         char   *nets;
  75         int     rc;
  76
  77         if (*networks != 0 && *ip2nets != 0) {
  78                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
  79                                    "'ip2nets' but not both at once\n");
  80                 return NULL;
  81         }
  82
  83         if (*ip2nets != 0) {
  84                 rc = lnet_parse_ip2nets(&nets, ip2nets);
  85                 return (rc == 0) ? nets : NULL;
  86         }
  87
  88         if (*networks != 0)
  89                 return networks;
  90
  91         return "tcp";
  92 }
  93
  94 static void
  95 lnet_init_locks(void)
  96 {
  97         spin_lock_init(&the_lnet.ln_eq_wait_lock);
  98         init_waitqueue_head(&the_lnet.ln_eq_waitq);
  99         init_waitqueue_head(&the_lnet.ln_rc_waitq);
 100         mutex_init(&the_lnet.ln_lnd_mutex);
 101         mutex_init(&the_lnet.ln_api_mutex);
 102 }
 103
 104 static void
 105 lnet_fini_locks(void)
 106 {
 107 }
 108
 109 static int
 110 lnet_create_remote_nets_table(void)
 111 {
 112         int               i;
 113         struct list_head *hash;
 114
 115         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 116         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 117         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 118         if (hash == NULL) {
 119                 CERROR("Failed to create remote nets hash table\n");
 120                 return -ENOMEM;
 121         }
 122
 123         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 124                 INIT_LIST_HEAD(&hash[i]);
 125         the_lnet.ln_remote_nets_hash = hash;
 126         return 0;
 127 }
 128
 129 static void
 130 lnet_destroy_remote_nets_table(void)
 131 {
 132         int i;
 133
 134         if (the_lnet.ln_remote_nets_hash == NULL)
 135                 return;
 136
 137         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 138                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 139
 140         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 141                     LNET_REMOTE_NETS_HASH_SIZE *
 142                     sizeof(the_lnet.ln_remote_nets_hash[0]));
 143         the_lnet.ln_remote_nets_hash = NULL;
 144 }
 145
 146 static void
 147 lnet_destroy_locks(void)
 148 {
 149         if (the_lnet.ln_res_lock != NULL) {
 150                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 151                 the_lnet.ln_res_lock = NULL;
 152         }
 153
 154         if (the_lnet.ln_net_lock != NULL) {
 155                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 156                 the_lnet.ln_net_lock = NULL;
 157         }
 158
 159         lnet_fini_locks();
 160 }
 161
 162 static int
 163 lnet_create_locks(void)
 164 {
 165         lnet_init_locks();
 166
 167         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 168         if (the_lnet.ln_res_lock == NULL)
 169                 goto failed;
 170
 171         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 172         if (the_lnet.ln_net_lock == NULL)
 173                 goto failed;
 174
 175         return 0;
 176
 177  failed:
 178         lnet_destroy_locks();
 179         return -ENOMEM;
 180 }
 181
 182 static void lnet_assert_wire_constants(void)
 183 {
 184         /* Wire protocol assertions generated by 'wirecheck'
 185          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 186          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 187          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 188
 189         /* Constants... */
 190         CLASSERT (LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 191         CLASSERT (LNET_PROTO_TCP_VERSION_MAJOR == 1);
 192         CLASSERT (LNET_PROTO_TCP_VERSION_MINOR == 0);
 193         CLASSERT (LNET_MSG_ACK == 0);
 194         CLASSERT (LNET_MSG_PUT == 1);
 195         CLASSERT (LNET_MSG_GET == 2);
 196         CLASSERT (LNET_MSG_REPLY == 3);
 197         CLASSERT (LNET_MSG_HELLO == 4);
 198
 199         /* Checks for struct ptl_handle_wire_t */
 200         CLASSERT ((int)sizeof(lnet_handle_wire_t) == 16);
 201         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_interface_cookie) == 0);
 202         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_interface_cookie) == 8);
 203         CLASSERT ((int)offsetof(lnet_handle_wire_t, wh_object_cookie) == 8);
 204         CLASSERT ((int)sizeof(((lnet_handle_wire_t *)0)->wh_object_cookie) == 8);
 205
 206         /* Checks for struct lnet_magicversion_t */
 207         CLASSERT ((int)sizeof(lnet_magicversion_t) == 8);
 208         CLASSERT ((int)offsetof(lnet_magicversion_t, magic) == 0);
 209         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->magic) == 4);
 210         CLASSERT ((int)offsetof(lnet_magicversion_t, version_major) == 4);
 211         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_major) == 2);
 212         CLASSERT ((int)offsetof(lnet_magicversion_t, version_minor) == 6);
 213         CLASSERT ((int)sizeof(((lnet_magicversion_t *)0)->version_minor) == 2);
 214
 215         /* Checks for struct lnet_hdr_t */
 216         CLASSERT ((int)sizeof(lnet_hdr_t) == 72);
 217         CLASSERT ((int)offsetof(lnet_hdr_t, dest_nid) == 0);
 218         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_nid) == 8);
 219         CLASSERT ((int)offsetof(lnet_hdr_t, src_nid) == 8);
 220         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_nid) == 8);
 221         CLASSERT ((int)offsetof(lnet_hdr_t, dest_pid) == 16);
 222         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->dest_pid) == 4);
 223         CLASSERT ((int)offsetof(lnet_hdr_t, src_pid) == 20);
 224         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->src_pid) == 4);
 225         CLASSERT ((int)offsetof(lnet_hdr_t, type) == 24);
 226         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->type) == 4);
 227         CLASSERT ((int)offsetof(lnet_hdr_t, payload_length) == 28);
 228         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->payload_length) == 4);
 229         CLASSERT ((int)offsetof(lnet_hdr_t, msg) == 32);
 230         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg) == 40);
 231
 232         /* Ack */
 233         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.dst_wmd) == 32);
 234         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.dst_wmd) == 16);
 235         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.match_bits) == 48);
 236         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.match_bits) == 8);
 237         CLASSERT ((int)offsetof(lnet_hdr_t, msg.ack.mlength) == 56);
 238         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.ack.mlength) == 4);
 239
 240         /* Put */
 241         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ack_wmd) == 32);
 242         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ack_wmd) == 16);
 243         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.match_bits) == 48);
 244         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.match_bits) == 8);
 245         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.hdr_data) == 56);
 246         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.hdr_data) == 8);
 247         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.ptl_index) == 64);
 248         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.ptl_index) == 4);
 249         CLASSERT ((int)offsetof(lnet_hdr_t, msg.put.offset) == 68);
 250         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.put.offset) == 4);
 251
 252         /* Get */
 253         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.return_wmd) == 32);
 254         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.return_wmd) == 16);
 255         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.match_bits) == 48);
 256         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.match_bits) == 8);
 257         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.ptl_index) == 56);
 258         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.ptl_index) == 4);
 259         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.src_offset) == 60);
 260         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.src_offset) == 4);
 261         CLASSERT ((int)offsetof(lnet_hdr_t, msg.get.sink_length) == 64);
 262         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.get.sink_length) == 4);
 263
 264         /* Reply */
 265         CLASSERT ((int)offsetof(lnet_hdr_t, msg.reply.dst_wmd) == 32);
 266         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.reply.dst_wmd) == 16);
 267
 268         /* Hello */
 269         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.incarnation) == 32);
 270         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.incarnation) == 8);
 271         CLASSERT ((int)offsetof(lnet_hdr_t, msg.hello.type) == 40);
 272         CLASSERT ((int)sizeof(((lnet_hdr_t *)0)->msg.hello.type) == 4);
 273 }
 274
 275 static lnd_t *lnet_find_lnd_by_type(__u32 type)
 276 {
 277         lnd_t            *lnd;
 278         struct list_head *tmp;
 279
 280         /* holding lnd mutex */
 281         list_for_each(tmp, &the_lnet.ln_lnds) {
 282                 lnd = list_entry(tmp, lnd_t, lnd_list);
 283
 284                 if (lnd->lnd_type == type)
 285                         return lnd;
 286         }
 287         return NULL;
 288 }
 289
 290 void
 291 lnet_register_lnd (lnd_t *lnd)
 292 {
 293         mutex_lock(&the_lnet.ln_lnd_mutex);
 294
 295         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 296         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 297
 298         list_add_tail(&lnd->lnd_list, &the_lnet.ln_lnds);
 299         lnd->lnd_refcount = 0;
 300
 301         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 302
 303         mutex_unlock(&the_lnet.ln_lnd_mutex);
 304 }
 305 EXPORT_SYMBOL(lnet_register_lnd);
 306
 307 void
 308 lnet_unregister_lnd (lnd_t *lnd)
 309 {
 310         mutex_lock(&the_lnet.ln_lnd_mutex);
 311
 312         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 313         LASSERT(lnd->lnd_refcount == 0);
 314
 315         list_del(&lnd->lnd_list);
 316         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 317
 318         mutex_unlock(&the_lnet.ln_lnd_mutex);
 319 }
 320 EXPORT_SYMBOL(lnet_unregister_lnd);
 321
 322 void
 323 lnet_counters_get(lnet_counters_t *counters)
 324 {
 325         lnet_counters_t *ctr;
 326         int             i;
 327
 328         memset(counters, 0, sizeof(*counters));
 329
 330         lnet_net_lock(LNET_LOCK_EX);
 331
 332         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 333                 counters->msgs_max     += ctr->msgs_max;
 334                 counters->msgs_alloc   += ctr->msgs_alloc;
 335                 counters->errors       += ctr->errors;
 336                 counters->send_count   += ctr->send_count;
 337                 counters->recv_count   += ctr->recv_count;
 338                 counters->route_count  += ctr->route_count;
 339                 counters->drop_count   += ctr->drop_count;
 340                 counters->send_length  += ctr->send_length;
 341                 counters->recv_length  += ctr->recv_length;
 342                 counters->route_length += ctr->route_length;
 343                 counters->drop_length  += ctr->drop_length;
 344
 345         }
 346         lnet_net_unlock(LNET_LOCK_EX);
 347 }
 348 EXPORT_SYMBOL(lnet_counters_get);
 349
 350 void
 351 lnet_counters_reset(void)
 352 {
 353         lnet_counters_t *counters;
 354         int             i;
 355
 356         lnet_net_lock(LNET_LOCK_EX);
 357
 358         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 359                 memset(counters, 0, sizeof(lnet_counters_t));
 360
 361         lnet_net_unlock(LNET_LOCK_EX);
 362 }
 363
 364 static __u64 lnet_create_interface_cookie(void)
 365 {
 366         /* NB the interface cookie in wire handles guards against delayed
 367          * replies and ACKs appearing valid after reboot. Initialisation time,
 368          * even if it's only implemented to millisecond resolution is probably
 369          * easily good enough. */
 370         struct timeval tv;
 371         __u64          cookie;
 372         do_gettimeofday(&tv);
 373         cookie = tv.tv_sec;
 374         cookie *= 1000000;
 375         cookie += tv.tv_usec;
 376         return cookie;
 377 }
 378
 379 static char *
 380 lnet_res_type2str(int type)
 381 {
 382         switch (type) {
 383         default:
 384                 LBUG();
 385         case LNET_COOKIE_TYPE_MD:
 386                 return "MD";
 387         case LNET_COOKIE_TYPE_ME:
 388                 return "ME";
 389         case LNET_COOKIE_TYPE_EQ:
 390                 return "EQ";
 391         }
 392 }
 393
 394 static void
 395 lnet_res_container_cleanup(struct lnet_res_container *rec)
 396 {
 397         int     count = 0;
 398
 399         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 400                 return;
 401
 402         while (!list_empty(&rec->rec_active)) {
 403                 struct list_head *e = rec->rec_active.next;
 404
 405                 list_del_init(e);
 406                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 407                         lnet_eq_free(list_entry(e, lnet_eq_t, eq_list));
 408
 409                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 410                         lnet_md_free(list_entry(e, lnet_libmd_t, md_list));
 411
 412                 } else { /* NB: Active MEs should be attached on portals */
 413                         LBUG();
 414                 }
 415                 count++;
 416         }
 417
 418         if (count > 0) {
 419                 /* Found alive MD/ME/EQ, user really should unlink/free
 420                  * all of them before finalize LNet, but if someone didn't,
 421                  * we have to recycle garbage for him */
 422                 CERROR("%d active elements on exit of %s container\n",
 423                        count, lnet_res_type2str(rec->rec_type));
 424         }
 425
 426         if (rec->rec_lh_hash != NULL) {
 427                 LIBCFS_FREE(rec->rec_lh_hash,
 428                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 429                 rec->rec_lh_hash = NULL;
 430         }
 431
 432         rec->rec_type = 0; /* mark it as finalized */
 433 }
 434
 435 static int
 436 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
 437 {
 438         int     rc = 0;
 439         int     i;
 440
 441         LASSERT(rec->rec_type == 0);
 442
 443         rec->rec_type = type;
 444         INIT_LIST_HEAD(&rec->rec_active);
 445
 446         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 447
 448         /* Arbitrary choice of hash table size */
 449         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 450                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 451         if (rec->rec_lh_hash == NULL) {
 452                 rc = -ENOMEM;
 453                 goto out;
 454         }
 455
 456         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 457                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 458
 459         return 0;
 460
 461 out:
 462         CERROR("Failed to setup %s resource container\n",
 463                lnet_res_type2str(type));
 464         lnet_res_container_cleanup(rec);
 465         return rc;
 466 }
 467
 468 static void
 469 lnet_res_containers_destroy(struct lnet_res_container **recs)
 470 {
 471         struct lnet_res_container       *rec;
 472         int                             i;
 473
 474         cfs_percpt_for_each(rec, i, recs)
 475                 lnet_res_container_cleanup(rec);
 476
 477         cfs_percpt_free(recs);
 478 }
 479
 480 static struct lnet_res_container **
 481 lnet_res_containers_create(int type)
 482 {
 483         struct lnet_res_container       **recs;
 484         struct lnet_res_container       *rec;
 485         int                             rc;
 486         int                             i;
 487
 488         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 489         if (recs == NULL) {
 490                 CERROR("Failed to allocate %s resource containers\n",
 491                        lnet_res_type2str(type));
 492                 return NULL;
 493         }
 494
 495         cfs_percpt_for_each(rec, i, recs) {
 496                 rc = lnet_res_container_setup(rec, i, type);
 497                 if (rc != 0) {
 498                         lnet_res_containers_destroy(recs);
 499                         return NULL;
 500                 }
 501         }
 502
 503         return recs;
 504 }
 505
 506 lnet_libhandle_t *
 507 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 508 {
 509         /* ALWAYS called with lnet_res_lock held */
 510         struct list_head        *head;
 511         lnet_libhandle_t        *lh;
 512         unsigned int            hash;
 513
 514         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
 515                 return NULL;
 516
 517         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 518         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 519
 520         list_for_each_entry(lh, head, lh_hash_chain) {
 521                 if (lh->lh_cookie == cookie)
 522                         return lh;
 523         }
 524
 525         return NULL;
 526 }
 527
 528 void
 529 lnet_res_lh_initialize(struct lnet_res_container *rec, lnet_libhandle_t *lh)
 530 {
 531         /* ALWAYS called with lnet_res_lock held */
 532         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 533         unsigned int    hash;
 534
 535         lh->lh_cookie = rec->rec_lh_cookie;
 536         rec->rec_lh_cookie += 1 << ibits;
 537
 538         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 539
 540         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 541 }
 542
 543 static int lnet_unprepare(void);
 544
 545 static int
 546 lnet_prepare(lnet_pid_t requested_pid)
 547 {
 548         /* Prepare to bring up the network */
 549         struct lnet_res_container **recs;
 550         int                       rc = 0;
 551
 552         if (requested_pid == LNET_PID_ANY) {
 553                 /* Don't instantiate LNET just for me */
 554                 return -ENETDOWN;
 555         }
 556
 557         LASSERT(the_lnet.ln_refcount == 0);
 558
 559         the_lnet.ln_routing = 0;
 560
 561         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
 562         the_lnet.ln_pid = requested_pid;
 563
 564         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
 565         INIT_LIST_HEAD(&the_lnet.ln_nis);
 566         INIT_LIST_HEAD(&the_lnet.ln_nis_cpt);
 567         INIT_LIST_HEAD(&the_lnet.ln_nis_zombie);
 568         INIT_LIST_HEAD(&the_lnet.ln_routers);
 569         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
 570         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
 571
 572         rc = lnet_create_remote_nets_table();
 573         if (rc != 0)
 574                 goto failed;
 575
 576         the_lnet.ln_interface_cookie = lnet_create_interface_cookie();
 577
 578         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
 579                                                 sizeof(lnet_counters_t));
 580         if (the_lnet.ln_counters == NULL) {
 581                 CERROR("Failed to allocate counters for LNet\n");
 582                 rc = -ENOMEM;
 583                 goto failed;
 584         }
 585
 586         rc = lnet_peer_tables_create();
 587         if (rc != 0)
 588                 goto failed;
 589
 590         rc = lnet_msg_containers_create();
 591         if (rc != 0)
 592                 goto failed;
 593
 594         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
 595                                       LNET_COOKIE_TYPE_EQ);
 596         if (rc != 0)
 597                 goto failed;
 598
 599         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME);
 600         if (recs == NULL) {
 601                 rc = -ENOMEM;
 602                 goto failed;
 603         }
 604
 605         the_lnet.ln_me_containers = recs;
 606
 607         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
 608         if (recs == NULL) {
 609                 rc = -ENOMEM;
 610                 goto failed;
 611         }
 612
 613         the_lnet.ln_md_containers = recs;
 614
 615         rc = lnet_portals_create();
 616         if (rc != 0) {
 617                 CERROR("Failed to create portals for LNet: %d\n", rc);
 618                 goto failed;
 619         }
 620
 621         return 0;
 622
 623  failed:
 624         lnet_unprepare();
 625         return rc;
 626 }
 627
 628 static int
 629 lnet_unprepare (void)
 630 {
 631         /* NB no LNET_LOCK since this is the last reference.  All LND instances
 632          * have shut down already, so it is safe to unlink and free all
 633          * descriptors, even those that appear committed to a network op (eg MD
 634          * with non-zero pending count) */
 635
 636         lnet_fail_nid(LNET_NID_ANY, 0);
 637
 638         LASSERT(the_lnet.ln_refcount == 0);
 639         LASSERT(list_empty(&the_lnet.ln_test_peers));
 640         LASSERT(list_empty(&the_lnet.ln_nis));
 641         LASSERT(list_empty(&the_lnet.ln_nis_cpt));
 642         LASSERT(list_empty(&the_lnet.ln_nis_zombie));
 643
 644         lnet_portals_destroy();
 645
 646         if (the_lnet.ln_md_containers != NULL) {
 647                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
 648                 the_lnet.ln_md_containers = NULL;
 649         }
 650
 651         if (the_lnet.ln_me_containers != NULL) {
 652                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
 653                 the_lnet.ln_me_containers = NULL;
 654         }
 655
 656         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
 657
 658         lnet_msg_containers_destroy();
 659         lnet_peer_tables_destroy();
 660         lnet_rtrpools_free(0);
 661
 662         if (the_lnet.ln_counters != NULL) {
 663                 cfs_percpt_free(the_lnet.ln_counters);
 664                 the_lnet.ln_counters = NULL;
 665         }
 666         lnet_destroy_remote_nets_table();
 667
 668         return 0;
 669 }
 670
 671 lnet_ni_t  *
 672 lnet_net2ni_locked(__u32 net, int cpt)
 673 {
 674         struct list_head *tmp;
 675         lnet_ni_t        *ni;
 676
 677         LASSERT(cpt != LNET_LOCK_EX);
 678
 679         list_for_each(tmp, &the_lnet.ln_nis) {
 680                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 681
 682                 if (LNET_NIDNET(ni->ni_nid) == net) {
 683                         lnet_ni_addref_locked(ni, cpt);
 684                         return ni;
 685                 }
 686         }
 687
 688         return NULL;
 689 }
 690
 691 lnet_ni_t *
 692 lnet_net2ni(__u32 net)
 693 {
 694         lnet_ni_t *ni;
 695
 696         lnet_net_lock(0);
 697         ni = lnet_net2ni_locked(net, 0);
 698         lnet_net_unlock(0);
 699
 700         return ni;
 701 }
 702 EXPORT_SYMBOL(lnet_net2ni);
 703
 704 static unsigned int
 705 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
 706 {
 707         __u64           key = nid;
 708         unsigned int    val;
 709
 710         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
 711
 712         if (number == 1)
 713                 return 0;
 714
 715         val = hash_long(key, LNET_CPT_BITS);
 716         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
 717         if (val < number)
 718                 return val;
 719
 720         return (unsigned int)(key + val + (val >> 1)) % number;
 721 }
 722
 723 int
 724 lnet_cpt_of_nid_locked(lnet_nid_t nid)
 725 {
 726         struct lnet_ni *ni;
 727
 728         /* must called with hold of lnet_net_lock */
 729         if (LNET_CPT_NUMBER == 1)
 730                 return 0; /* the only one */
 731
 732         /* take lnet_net_lock(any) would be OK */
 733         if (!list_empty(&the_lnet.ln_nis_cpt)) {
 734                 list_for_each_entry(ni, &the_lnet.ln_nis_cpt, ni_cptlist) {
 735                         if (LNET_NIDNET(ni->ni_nid) != LNET_NIDNET(nid))
 736                                 continue;
 737
 738                         LASSERT(ni->ni_cpts != NULL);
 739                         return ni->ni_cpts[lnet_nid_cpt_hash
 740                                            (nid, ni->ni_ncpts)];
 741                 }
 742         }
 743
 744         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 745 }
 746
 747 int
 748 lnet_cpt_of_nid(lnet_nid_t nid)
 749 {
 750         int     cpt;
 751         int     cpt2;
 752
 753         if (LNET_CPT_NUMBER == 1)
 754                 return 0; /* the only one */
 755
 756         if (list_empty(&the_lnet.ln_nis_cpt))
 757                 return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
 758
 759         cpt = lnet_net_lock_current();
 760         cpt2 = lnet_cpt_of_nid_locked(nid);
 761         lnet_net_unlock(cpt);
 762
 763         return cpt2;
 764 }
 765 EXPORT_SYMBOL(lnet_cpt_of_nid);
 766
 767 int
 768 lnet_islocalnet(__u32 net)
 769 {
 770         struct lnet_ni  *ni;
 771         int             cpt;
 772
 773         cpt = lnet_net_lock_current();
 774
 775         ni = lnet_net2ni_locked(net, cpt);
 776         if (ni != NULL)
 777                 lnet_ni_decref_locked(ni, cpt);
 778
 779         lnet_net_unlock(cpt);
 780
 781         return ni != NULL;
 782 }
 783
 784 lnet_ni_t  *
 785 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
 786 {
 787         struct lnet_ni   *ni;
 788         struct list_head *tmp;
 789
 790         LASSERT(cpt != LNET_LOCK_EX);
 791
 792         list_for_each(tmp, &the_lnet.ln_nis) {
 793                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 794
 795                 if (ni->ni_nid == nid) {
 796                         lnet_ni_addref_locked(ni, cpt);
 797                         return ni;
 798                 }
 799         }
 800
 801         return NULL;
 802 }
 803
 804 int
 805 lnet_islocalnid(lnet_nid_t nid)
 806 {
 807         struct lnet_ni  *ni;
 808         int             cpt;
 809
 810         cpt = lnet_net_lock_current();
 811         ni = lnet_nid2ni_locked(nid, cpt);
 812         if (ni != NULL)
 813                 lnet_ni_decref_locked(ni, cpt);
 814         lnet_net_unlock(cpt);
 815
 816         return ni != NULL;
 817 }
 818
 819 int
 820 lnet_count_acceptor_nis (void)
 821 {
 822         /* Return the # of NIs that need the acceptor. */
 823         int              count = 0;
 824         struct list_head *tmp;
 825         struct lnet_ni   *ni;
 826         int              cpt;
 827
 828         cpt = lnet_net_lock_current();
 829         list_for_each(tmp, &the_lnet.ln_nis) {
 830                 ni = list_entry(tmp, lnet_ni_t, ni_list);
 831
 832                 if (ni->ni_lnd->lnd_accept != NULL)
 833                         count++;
 834         }
 835
 836         lnet_net_unlock(cpt);
 837
 838         return count;
 839 }
 840
 841 static lnet_ping_info_t *
 842 lnet_ping_info_create(int num_ni)
 843 {
 844         lnet_ping_info_t *ping_info;
 845         unsigned int     infosz;
 846
 847         infosz = offsetof(lnet_ping_info_t, pi_ni[num_ni]);
 848         LIBCFS_ALLOC(ping_info, infosz);
 849         if (ping_info == NULL) {
 850                 CERROR("Can't allocate ping info[%d]\n", num_ni);
 851                 return NULL;
 852         }
 853
 854         ping_info->pi_nnis = num_ni;
 855         ping_info->pi_pid = the_lnet.ln_pid;
 856         ping_info->pi_magic = LNET_PROTO_PING_MAGIC;
 857         ping_info->pi_features = LNET_PING_FEAT_NI_STATUS;
 858
 859         return ping_info;
 860 }
 861
 862 static inline int
 863 lnet_get_ni_count(void)
 864 {
 865         struct lnet_ni *ni;
 866         int            count = 0;
 867
 868         lnet_net_lock(0);
 869
 870         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list)
 871                 count++;
 872
 873         lnet_net_unlock(0);
 874
 875         return count;
 876 }
 877
 878 static inline void
 879 lnet_ping_info_free(lnet_ping_info_t *pinfo)
 880 {
 881         LIBCFS_FREE(pinfo,
 882                     offsetof(lnet_ping_info_t,
 883                              pi_ni[pinfo->pi_nnis]));
 884 }
 885
 886 static void
 887 lnet_ping_info_destroy(void)
 888 {
 889         struct lnet_ni  *ni;
 890
 891         lnet_net_lock(LNET_LOCK_EX);
 892
 893         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
 894                 lnet_ni_lock(ni);
 895                 ni->ni_status = NULL;
 896                 lnet_ni_unlock(ni);
 897         }
 898
 899         lnet_ping_info_free(the_lnet.ln_ping_info);
 900         the_lnet.ln_ping_info = NULL;
 901
 902         lnet_net_unlock(LNET_LOCK_EX);
 903 }
 904
 905 static void
 906 lnet_ping_event_handler(lnet_event_t *event)
 907 {
 908         lnet_ping_info_t *pinfo = event->md.user_ptr;
 909
 910         if (event->unlinked)
 911                 pinfo->pi_features = LNET_PING_FEAT_INVAL;
 912 }
 913
 914 static int
 915 lnet_ping_info_setup(lnet_ping_info_t **ppinfo, lnet_handle_md_t *md_handle,
 916                      int ni_count, bool set_eq)
 917 {
 918         lnet_handle_me_t  me_handle;
 919         lnet_process_id_t id = {LNET_NID_ANY, LNET_PID_ANY};
 920         lnet_md_t         md = {NULL};
 921         int               rc, rc2;
 922
 923         if (set_eq) {
 924                 rc = LNetEQAlloc(0, lnet_ping_event_handler,
 925                                  &the_lnet.ln_ping_target_eq);
 926                 if (rc != 0) {
 927                         CERROR("Can't allocate ping EQ: %d\n", rc);
 928                         return rc;
 929                 }
 930         }
 931
 932         *ppinfo = lnet_ping_info_create(ni_count);
 933         if (*ppinfo == NULL) {
 934                 rc = -ENOMEM;
 935                 goto failed_0;
 936         }
 937
 938         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
 939                           LNET_PROTO_PING_MATCHBITS, 0,
 940                           LNET_UNLINK, LNET_INS_AFTER,
 941                           &me_handle);
 942         if (rc != 0) {
 943                 CERROR("Can't create ping ME: %d\n", rc);
 944                 goto failed_1;
 945         }
 946
 947         /* initialize md content */
 948         md.start     = *ppinfo;
 949         md.length    = offsetof(lnet_ping_info_t,
 950                                 pi_ni[(*ppinfo)->pi_nnis]);
 951         md.threshold = LNET_MD_THRESH_INF;
 952         md.max_size  = 0;
 953         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
 954                        LNET_MD_MANAGE_REMOTE;
 955         md.user_ptr  = NULL;
 956         md.eq_handle = the_lnet.ln_ping_target_eq;
 957         md.user_ptr = *ppinfo;
 958
 959         rc = LNetMDAttach(me_handle, md, LNET_RETAIN, md_handle);
 960         if (rc != 0) {
 961                 CERROR("Can't attach ping MD: %d\n", rc);
 962                 goto failed_2;
 963         }
 964
 965         return 0;
 966
 967 failed_2:
 968         rc2 = LNetMEUnlink(me_handle);
 969         LASSERT(rc2 == 0);
 970 failed_1:
 971         lnet_ping_info_free(*ppinfo);
 972         *ppinfo = NULL;
 973 failed_0:
 974         if (set_eq)
 975                 LNetEQFree(the_lnet.ln_ping_target_eq);
 976         return rc;
 977 }
 978
 979 static void
 980 lnet_ping_md_unlink(lnet_ping_info_t *pinfo, lnet_handle_md_t *md_handle)
 981 {
 982         sigset_t        blocked = cfs_block_allsigs();
 983
 984         LNetMDUnlink(*md_handle);
 985         LNetInvalidateHandle(md_handle);
 986
 987         /* NB md could be busy; this just starts the unlink */
 988         while (pinfo->pi_features != LNET_PING_FEAT_INVAL) {
 989                 CDEBUG(D_NET, "Still waiting for ping MD to unlink\n");
 990                 set_current_state(TASK_UNINTERRUPTIBLE);
 991                 schedule_timeout(cfs_time_seconds(1));
 992         }
 993
 994         cfs_restore_sigs(blocked);
 995 }
 996
 997 static void
 998 lnet_ping_info_install_locked(lnet_ping_info_t *ping_info)
 999 {
1000         int                     i;
1001         lnet_ni_t               *ni;
1002         lnet_ni_status_t        *ns;
1003
1004         i = 0;
1005         list_for_each_entry(ni, &the_lnet.ln_nis, ni_list) {
1006                 LASSERT(i < ping_info->pi_nnis);
1007
1008                 ns = &ping_info->pi_ni[i];
1009
1010                 ns->ns_nid = ni->ni_nid;
1011
1012                 lnet_ni_lock(ni);
1013                 ns->ns_status = (ni->ni_status != NULL) ?
1014                                 ni->ni_status->ns_status : LNET_NI_STATUS_UP;
1015                 ni->ni_status = ns;
1016                 lnet_ni_unlock(ni);
1017
1018                 i++;
1019         }
1020 }
1021
1022 static void
1023 lnet_ping_target_update(lnet_ping_info_t *pinfo, lnet_handle_md_t md_handle)
1024 {
1025         lnet_ping_info_t *old_pinfo = NULL;
1026         lnet_handle_md_t old_md;
1027
1028         /* switch the NIs to point to the new ping info created */
1029         lnet_net_lock(LNET_LOCK_EX);
1030
1031         if (!the_lnet.ln_routing)
1032                 pinfo->pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1033         lnet_ping_info_install_locked(pinfo);
1034
1035         if (the_lnet.ln_ping_info != NULL) {
1036                 old_pinfo = the_lnet.ln_ping_info;
1037                 old_md = the_lnet.ln_ping_target_md;
1038         }
1039         the_lnet.ln_ping_target_md = md_handle;
1040         the_lnet.ln_ping_info = pinfo;
1041
1042         lnet_net_unlock(LNET_LOCK_EX);
1043
1044         if (old_pinfo != NULL) {
1045                 /* unlink the old ping info */
1046                 lnet_ping_md_unlink(old_pinfo, &old_md);
1047                 lnet_ping_info_free(old_pinfo);
1048         }
1049 }
1050
1051 static void
1052 lnet_ping_target_fini(void)
1053 {
1054         int             rc;
1055
1056         lnet_ping_md_unlink(the_lnet.ln_ping_info,
1057                             &the_lnet.ln_ping_target_md);
1058
1059         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1060         LASSERT(rc == 0);
1061
1062         lnet_ping_info_destroy();
1063 }
1064
1065 static int
1066 lnet_ni_tq_credits(lnet_ni_t *ni)
1067 {
1068         int     credits;
1069
1070         LASSERT(ni->ni_ncpts >= 1);
1071
1072         if (ni->ni_ncpts == 1)
1073                 return ni->ni_maxtxcredits;
1074
1075         credits = ni->ni_maxtxcredits / ni->ni_ncpts;
1076         credits = max(credits, 8 * ni->ni_peertxcredits);
1077         credits = min(credits, ni->ni_maxtxcredits);
1078
1079         return credits;
1080 }
1081
1082 static void
1083 lnet_ni_unlink_locked(lnet_ni_t *ni)
1084 {
1085         if (!list_empty(&ni->ni_cptlist)) {
1086                 list_del_init(&ni->ni_cptlist);
1087                 lnet_ni_decref_locked(ni, 0);
1088         }
1089
1090         /* move it to zombie list and nobody can find it anymore */
1091         LASSERT(!list_empty(&ni->ni_list));
1092         list_move(&ni->ni_list, &the_lnet.ln_nis_zombie);
1093         lnet_ni_decref_locked(ni, 0);   /* drop ln_nis' ref */
1094 }
1095
1096 static void
1097 lnet_clear_zombies_nis_locked(void)
1098 {
1099         int             i;
1100         int             islo;
1101         lnet_ni_t       *ni;
1102
1103         /* Now wait for the NI's I just nuked to show up on ln_zombie_nis
1104          * and shut them down in guaranteed thread context */
1105         i = 2;
1106         while (!list_empty(&the_lnet.ln_nis_zombie)) {
1107                 int     *ref;
1108                 int     j;
1109
1110                 ni = list_entry(the_lnet.ln_nis_zombie.next,
1111                                 lnet_ni_t, ni_list);
1112                 list_del_init(&ni->ni_list);
1113                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1114                         if (*ref == 0)
1115                                 continue;
1116                         /* still busy, add it back to zombie list */
1117                         list_add(&ni->ni_list, &the_lnet.ln_nis_zombie);
1118                         break;
1119                 }
1120
1121                 if (!list_empty(&ni->ni_list)) {
1122                         lnet_net_unlock(LNET_LOCK_EX);
1123                         ++i;
1124                         if ((i & (-i)) == i) {
1125                                 CDEBUG(D_WARNING,
1126                                        "Waiting for zombie LNI %s\n",
1127                                        libcfs_nid2str(ni->ni_nid));
1128                         }
1129                         set_current_state(TASK_UNINTERRUPTIBLE);
1130                         schedule_timeout(cfs_time_seconds(1));
1131                         lnet_net_lock(LNET_LOCK_EX);
1132                         continue;
1133                 }
1134
1135                 ni->ni_lnd->lnd_refcount--;
1136                 lnet_net_unlock(LNET_LOCK_EX);
1137
1138                 islo = ni->ni_lnd->lnd_type == LOLND;
1139
1140                 LASSERT(!in_interrupt());
1141                 (ni->ni_lnd->lnd_shutdown)(ni);
1142
1143                 /* can't deref lnd anymore now; it might have unregistered
1144                  * itself...  */
1145
1146                 if (!islo)
1147                         CDEBUG(D_LNI, "Removed LNI %s\n",
1148                               libcfs_nid2str(ni->ni_nid));
1149
1150                 lnet_ni_free(ni);
1151                 i = 2;
1152                 lnet_net_lock(LNET_LOCK_EX);
1153         }
1154 }
1155
1156 static void
1157 lnet_shutdown_lndnis(void)
1158 {
1159         int             i;
1160         lnet_ni_t       *ni;
1161
1162         /* NB called holding the global mutex */
1163
1164         /* All quiet on the API front */
1165         LASSERT(!the_lnet.ln_shutdown);
1166         LASSERT(the_lnet.ln_refcount == 0);
1167         LASSERT(list_empty(&the_lnet.ln_nis_zombie));
1168
1169         lnet_net_lock(LNET_LOCK_EX);
1170         the_lnet.ln_shutdown = 1;       /* flag shutdown */
1171
1172         /* Unlink NIs from the global table */
1173         while (!list_empty(&the_lnet.ln_nis)) {
1174                 ni = list_entry(the_lnet.ln_nis.next,
1175                                 lnet_ni_t, ni_list);
1176                 lnet_ni_unlink_locked(ni);
1177         }
1178
1179         /* Drop the cached loopback NI. */
1180         if (the_lnet.ln_loni != NULL) {
1181                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1182                 the_lnet.ln_loni = NULL;
1183         }
1184
1185         lnet_net_unlock(LNET_LOCK_EX);
1186
1187         /* Clear lazy portals and drop delayed messages which hold refs
1188          * on their lnet_msg_t::msg_rxpeer */
1189         for (i = 0; i < the_lnet.ln_nportals; i++)
1190                 LNetClearLazyPortal(i);
1191
1192         /* Clear the peer table and wait for all peers to go (they hold refs on
1193          * their NIs) */
1194         lnet_peer_tables_cleanup(NULL);
1195
1196         lnet_net_lock(LNET_LOCK_EX);
1197
1198         lnet_clear_zombies_nis_locked();
1199         the_lnet.ln_shutdown = 0;
1200         lnet_net_unlock(LNET_LOCK_EX);
1201 }
1202
1203 /* shutdown down the NI and release refcount */
1204 static void
1205 lnet_shutdown_lndni(struct lnet_ni *ni)
1206 {
1207         int i;
1208
1209         lnet_net_lock(LNET_LOCK_EX);
1210         lnet_ni_unlink_locked(ni);
1211         lnet_net_unlock(LNET_LOCK_EX);
1212
1213         /* clear messages for this NI on the lazy portal */
1214         for (i = 0; i < the_lnet.ln_nportals; i++)
1215                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
1216
1217         /* Do peer table cleanup for this ni */
1218         lnet_peer_tables_cleanup(ni);
1219
1220         lnet_net_lock(LNET_LOCK_EX);
1221         lnet_clear_zombies_nis_locked();
1222         lnet_net_unlock(LNET_LOCK_EX);
1223 }
1224
1225 static int
1226 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_ioctl_config_data *conf)
1227 {
1228         struct lnet_ioctl_config_lnd_tunables *lnd_tunables = NULL;
1229         int                     rc = -EINVAL;
1230         __u32                   lnd_type;
1231         lnd_t                   *lnd;
1232         struct lnet_tx_queue    *tq;
1233         int                     i;
1234
1235         lnd_type = LNET_NETTYP(LNET_NIDNET(ni->ni_nid));
1236
1237         LASSERT(libcfs_isknown_lnd(lnd_type));
1238
1239         if (lnd_type == CIBLND || lnd_type == OPENIBLND ||
1240             lnd_type == IIBLND || lnd_type == VIBLND) {
1241                 CERROR("LND %s obsoleted\n", libcfs_lnd2str(lnd_type));
1242                 goto failed0;
1243         }
1244
1245         /* Make sure this new NI is unique. */
1246         lnet_net_lock(LNET_LOCK_EX);
1247         rc = lnet_net_unique(LNET_NIDNET(ni->ni_nid), &the_lnet.ln_nis);
1248         lnet_net_unlock(LNET_LOCK_EX);
1249
1250         if (!rc) {
1251                 if (lnd_type == LOLND) {
1252                         lnet_ni_free(ni);
1253                         return 0;
1254                 }
1255
1256                 CERROR("Net %s is not unique\n",
1257                        libcfs_net2str(LNET_NIDNET(ni->ni_nid)));
1258
1259                 rc = -EEXIST;
1260                 goto failed0;
1261         }
1262
1263         mutex_lock(&the_lnet.ln_lnd_mutex);
1264         lnd = lnet_find_lnd_by_type(lnd_type);
1265
1266         if (lnd == NULL) {
1267                 mutex_unlock(&the_lnet.ln_lnd_mutex);
1268                 rc = request_module("%s", libcfs_lnd2modname(lnd_type));
1269                 mutex_lock(&the_lnet.ln_lnd_mutex);
1270
1271                 lnd = lnet_find_lnd_by_type(lnd_type);
1272                 if (lnd == NULL) {
1273                         mutex_unlock(&the_lnet.ln_lnd_mutex);
1274                         CERROR("Can't load LND %s, module %s, rc=%d\n",
1275                                libcfs_lnd2str(lnd_type),
1276                                libcfs_lnd2modname(lnd_type), rc);
1277 #ifndef HAVE_MODULE_LOADING_SUPPORT
1278                         LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
1279                                            "compiled with kernel module "
1280                                            "loading support.");
1281 #endif
1282                         rc = -EINVAL;
1283                         goto failed0;
1284                 }
1285         }
1286
1287         lnet_net_lock(LNET_LOCK_EX);
1288         lnd->lnd_refcount++;
1289         lnet_net_unlock(LNET_LOCK_EX);
1290
1291         ni->ni_lnd = lnd;
1292
1293         if (conf && conf->cfg_hdr.ioc_len > sizeof(*conf))
1294                 lnd_tunables = (struct lnet_ioctl_config_lnd_tunables *)conf->cfg_bulk;
1295
1296         if (lnd_tunables != NULL) {
1297                 LIBCFS_ALLOC(ni->ni_lnd_tunables,
1298                              sizeof(*ni->ni_lnd_tunables));
1299                 if (ni->ni_lnd_tunables == NULL) {
1300                         mutex_unlock(&the_lnet.ln_lnd_mutex);
1301                         rc = -ENOMEM;
1302                         goto failed0;
1303                 }
1304                 memcpy(ni->ni_lnd_tunables, lnd_tunables,
1305                        sizeof(*ni->ni_lnd_tunables));
1306         }
1307
1308         rc = (lnd->lnd_startup)(ni);
1309
1310         mutex_unlock(&the_lnet.ln_lnd_mutex);
1311
1312         if (rc != 0) {
1313                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
1314                                    rc, libcfs_lnd2str(lnd->lnd_type));
1315                 lnet_net_lock(LNET_LOCK_EX);
1316                 lnd->lnd_refcount--;
1317                 lnet_net_unlock(LNET_LOCK_EX);
1318                 goto failed0;
1319         }
1320
1321         /* If given some LND tunable parameters, parse those now to
1322          * override the values in the NI structure. */
1323         if (conf && conf->cfg_config_u.cfg_net.net_peer_rtr_credits >= 0) {
1324                 ni->ni_peerrtrcredits =
1325                         conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
1326         }
1327         if (conf && conf->cfg_config_u.cfg_net.net_peer_timeout >= 0) {
1328                 ni->ni_peertimeout =
1329                         conf->cfg_config_u.cfg_net.net_peer_timeout;
1330         }
1331
1332         /*
1333          * TODO
1334          * Note: For now, don't allow the user to change
1335          * peertxcredits as this number is used in the
1336          * IB LND to control queue depth.
1337          *
1338          * if (conf && conf->cfg_config_u.cfg_net.net_peer_tx_credits != -1)
1339          *      ni->ni_peertxcredits =
1340          *              conf->cfg_config_u.cfg_net.net_peer_tx_credits;
1341          */
1342         if (conf && conf->cfg_config_u.cfg_net.net_max_tx_credits >= 0) {
1343                 ni->ni_maxtxcredits =
1344                         conf->cfg_config_u.cfg_net.net_max_tx_credits;
1345         }
1346
1347         LASSERT(ni->ni_peertimeout <= 0 || lnd->lnd_query != NULL);
1348
1349         lnet_net_lock(LNET_LOCK_EX);
1350         /* refcount for ln_nis */
1351         lnet_ni_addref_locked(ni, 0);
1352         list_add_tail(&ni->ni_list, &the_lnet.ln_nis);
1353         if (ni->ni_cpts != NULL) {
1354                 lnet_ni_addref_locked(ni, 0);
1355                 list_add_tail(&ni->ni_cptlist, &the_lnet.ln_nis_cpt);
1356         }
1357
1358         lnet_net_unlock(LNET_LOCK_EX);
1359
1360         if (lnd->lnd_type == LOLND) {
1361                 lnet_ni_addref(ni);
1362                 LASSERT(the_lnet.ln_loni == NULL);
1363                 the_lnet.ln_loni = ni;
1364                 return 0;
1365         }
1366
1367         if (ni->ni_peertxcredits == 0 || ni->ni_maxtxcredits == 0) {
1368                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
1369                                    libcfs_lnd2str(lnd->lnd_type),
1370                                    ni->ni_peertxcredits == 0 ?
1371                                         "" : "per-peer ");
1372                 /* shutdown the NI since if we get here then it must've already
1373                  * been started
1374                  */
1375                 lnet_shutdown_lndni(ni);
1376                 return -EINVAL;
1377         }
1378
1379         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
1380                 tq->tq_credits_min =
1381                 tq->tq_credits_max =
1382                 tq->tq_credits = lnet_ni_tq_credits(ni);
1383         }
1384
1385         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
1386                 libcfs_nid2str(ni->ni_nid), ni->ni_peertxcredits,
1387                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
1388                 ni->ni_peerrtrcredits, ni->ni_peertimeout);
1389
1390         return 0;
1391 failed0:
1392         lnet_ni_free(ni);
1393         return rc;
1394 }
1395
1396 static int
1397 lnet_startup_lndnis(struct list_head *nilist)
1398 {
1399         struct lnet_ni          *ni;
1400         int                     rc;
1401         int                     ni_count = 0;
1402
1403         while (!list_empty(nilist)) {
1404                 ni = list_entry(nilist->next, lnet_ni_t, ni_list);
1405                 list_del(&ni->ni_list);
1406                 rc = lnet_startup_lndni(ni, NULL);
1407
1408                 if (rc < 0)
1409                         goto failed;
1410
1411                 ni_count++;
1412         }
1413
1414         return ni_count;
1415 failed:
1416         lnet_shutdown_lndnis();
1417
1418         return rc;
1419 }
1420
1421 /**
1422  * Initialize LNet library.
1423  *
1424  * Automatically called at module loading time. Caller has to call
1425  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
1426  * latter returned 0. It must be called exactly once.
1427  *
1428  * \retval 0 on success
1429  * \retval -ve on failures.
1430  */
1431 int lnet_lib_init(void)
1432 {
1433         int rc;
1434
1435         lnet_assert_wire_constants();
1436
1437         memset(&the_lnet, 0, sizeof(the_lnet));
1438
1439         /* refer to global cfs_cpt_table for now */
1440         the_lnet.ln_cpt_table   = cfs_cpt_table;
1441         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
1442
1443         LASSERT(the_lnet.ln_cpt_number > 0);
1444         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
1445                 /* we are under risk of consuming all lh_cookie */
1446                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
1447                        "please change setting of CPT-table and retry\n",
1448                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
1449                 return -E2BIG;
1450         }
1451
1452         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
1453                 the_lnet.ln_cpt_bits++;
1454
1455         rc = lnet_create_locks();
1456         if (rc != 0) {
1457                 CERROR("Can't create LNet global locks: %d\n", rc);
1458                 return rc;
1459         }
1460
1461         the_lnet.ln_refcount = 0;
1462         LNetInvalidateHandle(&the_lnet.ln_rc_eqh);
1463         INIT_LIST_HEAD(&the_lnet.ln_lnds);
1464         INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
1465         INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
1466
1467         /* The hash table size is the number of bits it takes to express the set
1468          * ln_num_routes, minus 1 (better to under estimate than over so we
1469          * don't waste memory). */
1470         if (rnet_htable_size <= 0)
1471                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
1472         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
1473                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
1474         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
1475                                            order_base_2(rnet_htable_size) - 1);
1476
1477         /* All LNDs apart from the LOLND are in separate modules.  They
1478          * register themselves when their module loads, and unregister
1479          * themselves when their module is unloaded. */
1480         lnet_register_lnd(&the_lolnd);
1481         return 0;
1482 }
1483
1484 /**
1485  * Finalize LNet library.
1486  *
1487  * \pre lnet_lib_init() called with success.
1488  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
1489  */
1490 void lnet_lib_exit(void)
1491 {
1492         LASSERT(the_lnet.ln_refcount == 0);
1493
1494         while (!list_empty(&the_lnet.ln_lnds))
1495                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
1496                                                lnd_t, lnd_list));
1497         lnet_destroy_locks();
1498 }
1499
1500 /**
1501  * Set LNet PID and start LNet interfaces, routing, and forwarding.
1502  *
1503  * Users must call this function at least once before any other functions.
1504  * For each successful call there must be a corresponding call to
1505  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
1506  * ignored.
1507  *
1508  * The PID used by LNet may be different from the one requested.
1509  * See LNetGetId().
1510  *
1511  * \param requested_pid PID requested by the caller.
1512  *
1513  * \return >= 0 on success, and < 0 error code on failures.
1514  */
1515 int
1516 LNetNIInit(lnet_pid_t requested_pid)
1517 {
1518         int                     im_a_router = 0;
1519         int                     rc;
1520         int                     ni_count;
1521         lnet_ping_info_t        *pinfo;
1522         lnet_handle_md_t        md_handle;
1523         struct list_head        net_head;
1524
1525         INIT_LIST_HEAD(&net_head);
1526
1527         mutex_lock(&the_lnet.ln_api_mutex);
1528
1529         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
1530
1531         if (the_lnet.ln_refcount > 0) {
1532                 rc = the_lnet.ln_refcount++;
1533                 mutex_unlock(&the_lnet.ln_api_mutex);
1534                 return rc;
1535         }
1536
1537         rc = lnet_prepare(requested_pid);
1538         if (rc != 0) {
1539                 mutex_unlock(&the_lnet.ln_api_mutex);
1540                 return rc;
1541         }
1542
1543         /* Add in the loopback network */
1544         if (lnet_ni_alloc(LNET_MKNET(LOLND, 0), NULL, &net_head) == NULL) {
1545                 rc = -ENOMEM;
1546                 goto failed0;
1547         }
1548
1549         /* If LNet is being initialized via DLC it is possible
1550          * that the user requests not to load module parameters (ones which
1551          * are supported by DLC) on initialization.  Therefore, make sure not
1552          * to load networks, routes and forwarding from module parameters
1553          * in this case.  On cleanup in case of failure only clean up
1554          * routes if it has been loaded */
1555         if (!the_lnet.ln_nis_from_mod_params) {
1556                 rc = lnet_parse_networks(&net_head,
1557                                          lnet_get_networks());
1558                 if (rc < 0)
1559                         goto failed0;
1560         }
1561
1562         ni_count = lnet_startup_lndnis(&net_head);
1563         if (ni_count < 0) {
1564                 rc = ni_count;
1565                 goto failed0;
1566         }
1567
1568         if (!the_lnet.ln_nis_from_mod_params) {
1569                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
1570                 if (rc != 0)
1571                         goto failed1;
1572
1573                 rc = lnet_check_routes();
1574                 if (rc != 0)
1575                         goto failed2;
1576
1577                 rc = lnet_rtrpools_alloc(im_a_router);
1578                 if (rc != 0)
1579                         goto failed2;
1580         }
1581
1582         rc = lnet_acceptor_start();
1583         if (rc != 0)
1584                 goto failed2;
1585         the_lnet.ln_refcount = 1;
1586         /* Now I may use my own API functions... */
1587
1588         rc = lnet_ping_info_setup(&pinfo, &md_handle, ni_count, true);
1589         if (rc != 0)
1590                 goto failed3;
1591
1592         lnet_ping_target_update(pinfo, md_handle);
1593
1594         rc = lnet_router_checker_start();
1595         if (rc != 0)
1596                 goto failed4;
1597
1598         lnet_fault_init();
1599         lnet_proc_init();
1600
1601         mutex_unlock(&the_lnet.ln_api_mutex);
1602
1603         return 0;
1604
1605 failed4:
1606         lnet_ping_target_fini();
1607 failed3:
1608         the_lnet.ln_refcount = 0;
1609         lnet_acceptor_stop();
1610 failed2:
1611         if (!the_lnet.ln_nis_from_mod_params)
1612                 lnet_destroy_routes();
1613 failed1:
1614         lnet_shutdown_lndnis();
1615 failed0:
1616         lnet_unprepare();
1617         LASSERT(rc < 0);
1618         mutex_unlock(&the_lnet.ln_api_mutex);
1619         while (!list_empty(&net_head)) {
1620                 struct lnet_ni *ni;
1621                 ni = list_entry(net_head.next, struct lnet_ni, ni_list);
1622                 list_del_init(&ni->ni_list);
1623                 lnet_ni_free(ni);
1624         }
1625         return rc;
1626 }
1627 EXPORT_SYMBOL(LNetNIInit);
1628
1629 /**
1630  * Stop LNet interfaces, routing, and forwarding.
1631  *
1632  * Users must call this function once for each successful call to LNetNIInit().
1633  * Once the LNetNIFini() operation has been started, the results of pending
1634  * API operations are undefined.
1635  *
1636  * \return always 0 for current implementation.
1637  */
1638 int
1639 LNetNIFini()
1640 {
1641         mutex_lock(&the_lnet.ln_api_mutex);
1642
1643         LASSERT (the_lnet.ln_refcount > 0);
1644
1645         if (the_lnet.ln_refcount != 1) {
1646                 the_lnet.ln_refcount--;
1647         } else {
1648                 LASSERT(!the_lnet.ln_niinit_self);
1649
1650                 lnet_fault_fini();
1651
1652                 lnet_proc_fini();
1653                 lnet_router_checker_stop();
1654                 lnet_ping_target_fini();
1655
1656                 /* Teardown fns that use my own API functions BEFORE here */
1657                 the_lnet.ln_refcount = 0;
1658
1659                 lnet_acceptor_stop();
1660                 lnet_destroy_routes();
1661                 lnet_shutdown_lndnis();
1662                 lnet_unprepare();
1663         }
1664
1665         mutex_unlock(&the_lnet.ln_api_mutex);
1666         return 0;
1667 }
1668 EXPORT_SYMBOL(LNetNIFini);
1669
1670 /**
1671  * Grabs the ni data from the ni structure and fills the out
1672  * parameters
1673  *
1674  * \param[in] ni network        interface structure
1675  * \param[out] cpt_count        the number of cpts the ni is on
1676  * \param[out] nid              Network Interface ID
1677  * \param[out] peer_timeout     NI peer timeout
1678  * \param[out] peer_tx_crdits   NI peer transmit credits
1679  * \param[out] peer_rtr_credits NI peer router credits
1680  * \param[out] max_tx_credits   NI max transmit credit
1681  * \param[out] net_config       Network configuration
1682  */
1683 static void
1684 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_data *config)
1685 {
1686         struct lnet_ioctl_net_config *net_config;
1687         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
1688         size_t min_size, tunable_size = 0;
1689         int i;
1690
1691         if (!ni || !config)
1692                 return;
1693
1694         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
1695         if (!net_config)
1696                 return;
1697
1698         CLASSERT(ARRAY_SIZE(ni->ni_interfaces) ==
1699                  ARRAY_SIZE(net_config->ni_interfaces));
1700
1701         if (ni->ni_interfaces[0] != NULL) {
1702                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
1703                         if (ni->ni_interfaces[i] != NULL) {
1704                                 strncpy(net_config->ni_interfaces[i],
1705                                         ni->ni_interfaces[i],
1706                                         sizeof(net_config->ni_interfaces[i]));
1707                         }
1708                 }
1709         }
1710
1711         config->cfg_nid = ni->ni_nid;
1712         config->cfg_config_u.cfg_net.net_peer_timeout = ni->ni_peertimeout;
1713         config->cfg_config_u.cfg_net.net_max_tx_credits = ni->ni_maxtxcredits;
1714         config->cfg_config_u.cfg_net.net_peer_tx_credits = ni->ni_peertxcredits;
1715         config->cfg_config_u.cfg_net.net_peer_rtr_credits = ni->ni_peerrtrcredits;
1716
1717         net_config->ni_status = ni->ni_status->ns_status;
1718
1719         for (i = 0;
1720              ni->ni_cpts != NULL && i < ni->ni_ncpts &&
1721              i < LNET_MAX_SHOW_NUM_CPT;
1722              i++)
1723                 net_config->ni_cpts[i] = ni->ni_cpts[i];
1724
1725         config->cfg_ncpts = ni->ni_ncpts;
1726
1727         /*
1728          * See if user land tools sent in a newer and larger version
1729          * of struct lnet_tunables than what the kernel uses.
1730          */
1731         min_size = sizeof(*config) + sizeof(*net_config);
1732
1733         if (config->cfg_hdr.ioc_len > min_size)
1734                 tunable_size = config->cfg_hdr.ioc_len - min_size;
1735
1736         /* Don't copy to much data to user space */
1737         min_size = min(tunable_size, sizeof(*ni->ni_lnd_tunables));
1738         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
1739
1740         if (ni->ni_lnd_tunables && lnd_cfg && min_size) {
1741                 memcpy(lnd_cfg, ni->ni_lnd_tunables, min_size);
1742                 config->cfg_config_u.cfg_net.net_interface_count = 1;
1743
1744                 /* Tell user land that kernel side has less data */
1745                 if (tunable_size > sizeof(*ni->ni_lnd_tunables)) {
1746                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
1747                         config->cfg_hdr.ioc_len -= min_size;
1748                 }
1749         }
1750 }
1751
1752 static int
1753 lnet_get_net_config(struct lnet_ioctl_config_data *config)
1754 {
1755         struct lnet_ni *ni;
1756         struct list_head *tmp;
1757         int idx = config->cfg_count;
1758         int rc = -ENOENT;
1759         int cpt;
1760
1761         if (unlikely(!config->cfg_bulk))
1762                 return -EINVAL;
1763
1764         cpt = lnet_net_lock_current();
1765
1766         list_for_each(tmp, &the_lnet.ln_nis) {
1767                 ni = list_entry(tmp, lnet_ni_t, ni_list);
1768                 if (idx-- == 0) {
1769                         rc = 0;
1770                         lnet_ni_lock(ni);
1771                         lnet_fill_ni_info(ni, config);
1772                         lnet_ni_unlock(ni);
1773                         break;
1774                 }
1775         }
1776
1777         lnet_net_unlock(cpt);
1778         return rc;
1779 }
1780
1781 int
1782 lnet_dyn_add_ni(lnet_pid_t requested_pid, struct lnet_ioctl_config_data *conf)
1783 {
1784         char                    *nets = conf->cfg_config_u.cfg_net.net_intf;
1785         lnet_ping_info_t        *pinfo;
1786         lnet_handle_md_t        md_handle;
1787         struct lnet_ni          *ni;
1788         struct list_head        net_head;
1789         int                     rc;
1790         lnet_remotenet_t        *rnet;
1791
1792         INIT_LIST_HEAD(&net_head);
1793
1794         /* Create a ni structure for the network string */
1795         rc = lnet_parse_networks(&net_head, nets);
1796         if (rc <= 0)
1797                 return rc == 0 ? -EINVAL : rc;
1798
1799         mutex_lock(&the_lnet.ln_api_mutex);
1800
1801         if (rc > 1) {
1802                 rc = -EINVAL; /* only add one interface per call */
1803                 goto failed0;
1804         }
1805
1806         ni = list_entry(net_head.next, struct lnet_ni, ni_list);
1807
1808         lnet_net_lock(LNET_LOCK_EX);
1809         rnet = lnet_find_net_locked(LNET_NIDNET(ni->ni_nid));
1810         lnet_net_unlock(LNET_LOCK_EX);
1811         /* make sure that the net added doesn't invalidate the current
1812          * configuration LNet is keeping */
1813         if (rnet != NULL) {
1814                 CERROR("Adding net %s will invalidate routing configuration\n",
1815                        nets);
1816                 rc = -EUSERS;
1817                 goto failed0;
1818         }
1819
1820         rc = lnet_ping_info_setup(&pinfo, &md_handle, 1 + lnet_get_ni_count(),
1821                                   false);
1822         if (rc != 0)
1823                 goto failed0;
1824
1825         list_del_init(&ni->ni_list);
1826
1827         rc = lnet_startup_lndni(ni, conf);
1828         if (rc != 0)
1829                 goto failed1;
1830
1831         if (ni->ni_lnd->lnd_accept != NULL) {
1832                 rc = lnet_acceptor_start();
1833                 if (rc < 0) {
1834                         /* shutdown the ni that we just started */
1835                         CERROR("Failed to start up acceptor thread\n");
1836                         lnet_shutdown_lndni(ni);
1837                         goto failed1;
1838                 }
1839         }
1840
1841         lnet_ping_target_update(pinfo, md_handle);
1842         mutex_unlock(&the_lnet.ln_api_mutex);
1843
1844         return 0;
1845
1846 failed1:
1847         lnet_ping_md_unlink(pinfo, &md_handle);
1848         lnet_ping_info_free(pinfo);
1849 failed0:
1850         mutex_unlock(&the_lnet.ln_api_mutex);
1851         while (!list_empty(&net_head)) {
1852                 ni = list_entry(net_head.next, struct lnet_ni, ni_list);
1853                 list_del_init(&ni->ni_list);
1854                 lnet_ni_free(ni);
1855         }
1856         return rc;
1857 }
1858
1859 int
1860 lnet_dyn_del_ni(__u32 net)
1861 {
1862         lnet_ni_t        *ni;
1863         lnet_ping_info_t *pinfo;
1864         lnet_handle_md_t  md_handle;
1865         int               rc;
1866
1867         /* don't allow userspace to shutdown the LOLND */
1868         if (LNET_NETTYP(net) == LOLND)
1869                 return -EINVAL;
1870
1871         mutex_lock(&the_lnet.ln_api_mutex);
1872         /* create and link a new ping info, before removing the old one */
1873         rc = lnet_ping_info_setup(&pinfo, &md_handle,
1874                                   lnet_get_ni_count() - 1, false);
1875         if (rc != 0)
1876                 goto out;
1877
1878         ni = lnet_net2ni(net);
1879         if (ni == NULL) {
1880                 rc = -EINVAL;
1881                 goto failed;
1882         }
1883
1884         /* decrement the reference counter taken by lnet_net2ni() */
1885         lnet_ni_decref_locked(ni, 0);
1886
1887         lnet_shutdown_lndni(ni);
1888
1889         if (lnet_count_acceptor_nis() == 0)
1890                 lnet_acceptor_stop();
1891
1892         lnet_ping_target_update(pinfo, md_handle);
1893         goto out;
1894 failed:
1895         lnet_ping_md_unlink(pinfo, &md_handle);
1896         lnet_ping_info_free(pinfo);
1897 out:
1898         mutex_unlock(&the_lnet.ln_api_mutex);
1899
1900         return rc;
1901 }
1902
1903 /**
1904  * LNet ioctl handler.
1905  *
1906  */
1907 int
1908 LNetCtl(unsigned int cmd, void *arg)
1909 {
1910         struct libcfs_ioctl_data *data = arg;
1911         struct lnet_ioctl_config_data *config;
1912         lnet_process_id_t         id = {0};
1913         lnet_ni_t                *ni;
1914         int                       rc;
1915
1916         CLASSERT(LIBCFS_IOC_DATA_MAX >= sizeof(struct lnet_ioctl_net_config) +
1917                                         sizeof(struct lnet_ioctl_config_data));
1918
1919         switch (cmd) {
1920         case IOC_LIBCFS_GET_NI:
1921                 rc = LNetGetId(data->ioc_count, &id);
1922                 data->ioc_nid = id.nid;
1923                 return rc;
1924
1925         case IOC_LIBCFS_FAIL_NID:
1926                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
1927
1928         case IOC_LIBCFS_ADD_ROUTE:
1929                 config = arg;
1930
1931                 if (config->cfg_hdr.ioc_len < sizeof(*config))
1932                         return -EINVAL;
1933
1934                 mutex_lock(&the_lnet.ln_api_mutex);
1935                 rc = lnet_add_route(config->cfg_net,
1936                                     config->cfg_config_u.cfg_route.rtr_hop,
1937                                     config->cfg_nid,
1938                                     config->cfg_config_u.cfg_route.
1939                                         rtr_priority);
1940                 if (rc == 0) {
1941                         rc = lnet_check_routes();
1942                         if (rc != 0)
1943                                 lnet_del_route(config->cfg_net,
1944                                                config->cfg_nid);
1945                 }
1946                 mutex_unlock(&the_lnet.ln_api_mutex);
1947                 return rc;
1948
1949         case IOC_LIBCFS_DEL_ROUTE:
1950                 config = arg;
1951
1952                 if (config->cfg_hdr.ioc_len < sizeof(*config))
1953                         return -EINVAL;
1954
1955                 mutex_lock(&the_lnet.ln_api_mutex);
1956                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
1957                 mutex_unlock(&the_lnet.ln_api_mutex);
1958                 return rc;
1959
1960         case IOC_LIBCFS_GET_ROUTE:
1961                 config = arg;
1962
1963                 if (config->cfg_hdr.ioc_len < sizeof(*config))
1964                         return -EINVAL;
1965
1966                 return lnet_get_route(config->cfg_count,
1967                                       &config->cfg_net,
1968                                       &config->cfg_config_u.cfg_route.rtr_hop,
1969                                       &config->cfg_nid,
1970                                       &config->cfg_config_u.cfg_route.rtr_flags,
1971                                       &config->cfg_config_u.cfg_route.
1972                                         rtr_priority);
1973
1974         case IOC_LIBCFS_GET_NET: {
1975                 size_t total = sizeof(*config) +
1976                                sizeof(struct lnet_ioctl_net_config);
1977                 config = arg;
1978
1979                 if (config->cfg_hdr.ioc_len < total)
1980                         return -EINVAL;
1981
1982                 return lnet_get_net_config(config);
1983         }
1984
1985         case IOC_LIBCFS_GET_LNET_STATS:
1986         {
1987                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
1988
1989                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
1990                         return -EINVAL;
1991
1992                 lnet_counters_get(&lnet_stats->st_cntrs);
1993                 return 0;
1994         }
1995
1996         case IOC_LIBCFS_CONFIG_RTR:
1997                 config = arg;
1998
1999                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2000                         return -EINVAL;
2001
2002                 mutex_lock(&the_lnet.ln_api_mutex);
2003                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
2004                         rc = lnet_rtrpools_enable();
2005                         mutex_unlock(&the_lnet.ln_api_mutex);
2006                         return rc;
2007                 }
2008                 lnet_rtrpools_disable();
2009                 mutex_unlock(&the_lnet.ln_api_mutex);
2010                 return 0;
2011
2012         case IOC_LIBCFS_ADD_BUF:
2013                 config = arg;
2014
2015                 if (config->cfg_hdr.ioc_len < sizeof(*config))
2016                         return -EINVAL;
2017
2018                 mutex_lock(&the_lnet.ln_api_mutex);
2019                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
2020                                                 buf_tiny,
2021                                           config->cfg_config_u.cfg_buffers.
2022                                                 buf_small,
2023                                           config->cfg_config_u.cfg_buffers.
2024                                                 buf_large);
2025                 mutex_unlock(&the_lnet.ln_api_mutex);
2026                 return rc;
2027
2028         case IOC_LIBCFS_GET_BUF: {
2029                 struct lnet_ioctl_pool_cfg *pool_cfg;
2030                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
2031
2032                 config = arg;
2033
2034                 if (config->cfg_hdr.ioc_len < total)
2035                         return -EINVAL;
2036
2037                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
2038                 return lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
2039         }
2040
2041         case IOC_LIBCFS_GET_PEER_INFO: {
2042                 struct lnet_ioctl_peer *peer_info = arg;
2043
2044                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
2045                         return -EINVAL;
2046
2047                 return lnet_get_peer_info(
2048                    peer_info->pr_count,
2049                    &peer_info->pr_nid,
2050                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
2051                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
2052                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
2053                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
2054                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
2055                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
2056                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_rtr_credits,
2057                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
2058         }
2059
2060         case IOC_LIBCFS_NOTIFY_ROUTER:
2061                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
2062                                    cfs_time_current() -
2063                                    cfs_time_seconds(cfs_time_current_sec() -
2064                                                     (time_t)data->ioc_u64[0]));
2065
2066         case IOC_LIBCFS_LNET_DIST:
2067                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
2068                 if (rc < 0 && rc != -EHOSTUNREACH)
2069                         return rc;
2070
2071                 data->ioc_u32[0] = rc;
2072                 return 0;
2073
2074         case IOC_LIBCFS_TESTPROTOCOMPAT:
2075                 lnet_net_lock(LNET_LOCK_EX);
2076                 the_lnet.ln_testprotocompat = data->ioc_flags;
2077                 lnet_net_unlock(LNET_LOCK_EX);
2078                 return 0;
2079
2080         case IOC_LIBCFS_LNET_FAULT:
2081                 return lnet_fault_ctl(data->ioc_flags, data);
2082
2083         case IOC_LIBCFS_PING:
2084                 id.nid = data->ioc_nid;
2085                 id.pid = data->ioc_u32[0];
2086                 rc = lnet_ping(id, data->ioc_u32[1], /* timeout */
2087                                (lnet_process_id_t __user *)data->ioc_pbuf1,
2088                                data->ioc_plen1/sizeof(lnet_process_id_t));
2089                 if (rc < 0)
2090                         return rc;
2091                 data->ioc_count = rc;
2092                 return 0;
2093
2094         default:
2095                 ni = lnet_net2ni(data->ioc_net);
2096                 if (ni == NULL)
2097                         return -EINVAL;
2098
2099                 if (ni->ni_lnd->lnd_ctl == NULL)
2100                         rc = -EINVAL;
2101                 else
2102                         rc = ni->ni_lnd->lnd_ctl(ni, cmd, arg);
2103
2104                 lnet_ni_decref(ni);
2105                 return rc;
2106         }
2107         /* not reached */
2108 }
2109 EXPORT_SYMBOL(LNetCtl);
2110
2111 void LNetDebugPeer(lnet_process_id_t id)
2112 {
2113         lnet_debug_peer(id.nid);
2114 }
2115 EXPORT_SYMBOL(LNetDebugPeer);
2116
2117 /**
2118  * Retrieve the lnet_process_id_t ID of LNet interface at \a index. Note that
2119  * all interfaces share a same PID, as requested by LNetNIInit().
2120  *
2121  * \param index Index of the interface to look up.
2122  * \param id On successful return, this location will hold the
2123  * lnet_process_id_t ID of the interface.
2124  *
2125  * \retval 0 If an interface exists at \a index.
2126  * \retval -ENOENT If no interface has been found.
2127  */
2128 int
2129 LNetGetId(unsigned int index, lnet_process_id_t *id)
2130 {
2131         struct lnet_ni   *ni;
2132         struct list_head *tmp;
2133         int               cpt;
2134         int               rc = -ENOENT;
2135
2136         LASSERT(the_lnet.ln_refcount > 0);
2137
2138         cpt = lnet_net_lock_current();
2139
2140         list_for_each(tmp, &the_lnet.ln_nis) {
2141                 if (index-- != 0)
2142                         continue;
2143
2144                 ni = list_entry(tmp, lnet_ni_t, ni_list);
2145
2146                 id->nid = ni->ni_nid;
2147                 id->pid = the_lnet.ln_pid;
2148                 rc = 0;
2149                 break;
2150         }
2151
2152         lnet_net_unlock(cpt);
2153         return rc;
2154 }
2155 EXPORT_SYMBOL(LNetGetId);
2156
2157 /**
2158  * Print a string representation of handle \a h into buffer \a str of
2159  * \a len bytes.
2160  */
2161 void
2162 LNetSnprintHandle(char *str, int len, lnet_handle_any_t h)
2163 {
2164         snprintf(str, len, LPX64, h.cookie);
2165 }
2166 EXPORT_SYMBOL(LNetSnprintHandle);
2167
2168 static int
2169 lnet_ping(lnet_process_id_t id, int timeout_ms, lnet_process_id_t __user *ids,
2170           int n_ids)
2171 {
2172         lnet_handle_eq_t     eqh;
2173         lnet_handle_md_t     mdh;
2174         lnet_event_t         event;
2175         lnet_md_t            md = { NULL };
2176         int                  which;
2177         int                  unlinked = 0;
2178         int                  replied = 0;
2179         const int            a_long_time = 60000; /* mS */
2180         int                  infosz;
2181         lnet_ping_info_t    *info;
2182         lnet_process_id_t    tmpid;
2183         int                  i;
2184         int                  nob;
2185         int                  rc;
2186         int                  rc2;
2187         sigset_t         blocked;
2188
2189         infosz = offsetof(lnet_ping_info_t, pi_ni[n_ids]);
2190
2191         if (n_ids <= 0 ||
2192             id.nid == LNET_NID_ANY ||
2193             timeout_ms > 500000 ||              /* arbitrary limit! */
2194             n_ids > 20)                         /* arbitrary limit! */
2195                 return -EINVAL;
2196
2197         if (id.pid == LNET_PID_ANY)
2198                 id.pid = LNET_PID_LUSTRE;
2199
2200         LIBCFS_ALLOC(info, infosz);
2201         if (info == NULL)
2202                 return -ENOMEM;
2203
2204         /* NB 2 events max (including any unlink event) */
2205         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
2206         if (rc != 0) {
2207                 CERROR("Can't allocate EQ: %d\n", rc);
2208                 goto out_0;
2209         }
2210
2211         /* initialize md content */
2212         md.start     = info;
2213         md.length    = infosz;
2214         md.threshold = 2; /*GET/REPLY*/
2215         md.max_size  = 0;
2216         md.options   = LNET_MD_TRUNCATE;
2217         md.user_ptr  = NULL;
2218         md.eq_handle = eqh;
2219
2220         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
2221         if (rc != 0) {
2222                 CERROR("Can't bind MD: %d\n", rc);
2223                 goto out_1;
2224         }
2225
2226         rc = LNetGet(LNET_NID_ANY, mdh, id,
2227                      LNET_RESERVED_PORTAL,
2228                      LNET_PROTO_PING_MATCHBITS, 0);
2229
2230         if (rc != 0) {
2231                 /* Don't CERROR; this could be deliberate! */
2232
2233                 rc2 = LNetMDUnlink(mdh);
2234                 LASSERT(rc2 == 0);
2235
2236                 /* NB must wait for the UNLINK event below... */
2237                 unlinked = 1;
2238                 timeout_ms = a_long_time;
2239         }
2240
2241         do {
2242                 /* MUST block for unlink to complete */
2243                 if (unlinked)
2244                         blocked = cfs_block_allsigs();
2245
2246                 rc2 = LNetEQPoll(&eqh, 1, timeout_ms, &event, &which);
2247
2248                 if (unlinked)
2249                         cfs_restore_sigs(blocked);
2250
2251                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
2252                        (rc2 <= 0) ? -1 : event.type,
2253                        (rc2 <= 0) ? -1 : event.status,
2254                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
2255
2256                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
2257
2258                 if (rc2 <= 0 || event.status != 0) {
2259                         /* timeout or error */
2260                         if (!replied && rc == 0)
2261                                 rc = (rc2 < 0) ? rc2 :
2262                                      (rc2 == 0) ? -ETIMEDOUT :
2263                                      event.status;
2264
2265                         if (!unlinked) {
2266                                 /* Ensure completion in finite time... */
2267                                 LNetMDUnlink(mdh);
2268                                 /* No assertion (racing with network) */
2269                                 unlinked = 1;
2270                                 timeout_ms = a_long_time;
2271                         } else if (rc2 == 0) {
2272                                 /* timed out waiting for unlink */
2273                                 CWARN("ping %s: late network completion\n",
2274                                       libcfs_id2str(id));
2275                         }
2276                 } else if (event.type == LNET_EVENT_REPLY) {
2277                         replied = 1;
2278                         rc = event.mlength;
2279                 }
2280
2281         } while (rc2 <= 0 || !event.unlinked);
2282
2283         if (!replied) {
2284                 if (rc >= 0)
2285                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
2286                               libcfs_id2str(id));
2287                 rc = -EIO;
2288                 goto out_1;
2289         }
2290
2291         nob = rc;
2292         LASSERT(nob >= 0 && nob <= infosz);
2293
2294         rc = -EPROTO;                           /* if I can't parse... */
2295
2296         if (nob < 8) {
2297                 /* can't check magic/version */
2298                 CERROR("%s: ping info too short %d\n",
2299                        libcfs_id2str(id), nob);
2300                 goto out_1;
2301         }
2302
2303         if (info->pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
2304                 lnet_swap_pinginfo(info);
2305         } else if (info->pi_magic != LNET_PROTO_PING_MAGIC) {
2306                 CERROR("%s: Unexpected magic %08x\n",
2307                        libcfs_id2str(id), info->pi_magic);
2308                 goto out_1;
2309         }
2310
2311         if ((info->pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
2312                 CERROR("%s: ping w/o NI status: 0x%x\n",
2313                        libcfs_id2str(id), info->pi_features);
2314                 goto out_1;
2315         }
2316
2317         if (nob < offsetof(lnet_ping_info_t, pi_ni[0])) {
2318                 CERROR("%s: Short reply %d(%d min)\n", libcfs_id2str(id),
2319                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[0]));
2320                 goto out_1;
2321         }
2322
2323         if (info->pi_nnis < n_ids)
2324                 n_ids = info->pi_nnis;
2325
2326         if (nob < offsetof(lnet_ping_info_t, pi_ni[n_ids])) {
2327                 CERROR("%s: Short reply %d(%d expected)\n", libcfs_id2str(id),
2328                        nob, (int)offsetof(lnet_ping_info_t, pi_ni[n_ids]));
2329                 goto out_1;
2330         }
2331
2332         rc = -EFAULT;                           /* If I SEGV... */
2333
2334         memset(&tmpid, 0, sizeof(tmpid));
2335         for (i = 0; i < n_ids; i++) {
2336                 tmpid.pid = info->pi_pid;
2337                 tmpid.nid = info->pi_ni[i].ns_nid;
2338                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
2339                         goto out_1;
2340         }
2341         rc = info->pi_nnis;
2342
2343  out_1:
2344         rc2 = LNetEQFree(eqh);
2345         if (rc2 != 0)
2346                 CERROR("rc2 %d\n", rc2);
2347         LASSERT(rc2 == 0);
2348
2349  out_0:
2350         LIBCFS_FREE(info, infosz);
2351         return rc;
2352 }