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, 2017, 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
  35 #include <linux/ctype.h>
  36 #include <linux/log2.h>
  37 #include <linux/ktime.h>
  38 #include <linux/moduleparam.h>
  39 #include <linux/uaccess.h>
  40
  41 #include <lnet/lib-lnet.h>
  42
  43 #define D_LNI D_CONSOLE
  44
  45 /*
  46  * initialize ln_api_mutex statically, since it needs to be used in
  47  * discovery_set callback. That module parameter callback can be called
  48  * before module init completes. The mutex needs to be ready for use then.
  49  */
  50 struct lnet the_lnet = {
  51         .ln_api_mutex = __MUTEX_INITIALIZER(the_lnet.ln_api_mutex),
  52 };              /* THE state of the network */
  53 EXPORT_SYMBOL(the_lnet);
  54
  55 static char *ip2nets = "";
  56 module_param(ip2nets, charp, 0444);
  57 MODULE_PARM_DESC(ip2nets, "LNET network <- IP table");
  58
  59 static char *networks = "";
  60 module_param(networks, charp, 0444);
  61 MODULE_PARM_DESC(networks, "local networks");
  62
  63 static char *routes = "";
  64 module_param(routes, charp, 0444);
  65 MODULE_PARM_DESC(routes, "routes to non-local networks");
  66
  67 static int rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
  68 module_param(rnet_htable_size, int, 0444);
  69 MODULE_PARM_DESC(rnet_htable_size, "size of remote network hash table");
  70
  71 static int use_tcp_bonding = false;
  72 module_param(use_tcp_bonding, int, 0444);
  73 MODULE_PARM_DESC(use_tcp_bonding,
  74                  "Set to 1 to use socklnd bonding. 0 to use Multi-Rail");
  75
  76 unsigned int lnet_numa_range = 0;
  77 module_param(lnet_numa_range, uint, 0444);
  78 MODULE_PARM_DESC(lnet_numa_range,
  79                 "NUMA range to consider during Multi-Rail selection");
  80
  81 /*
  82  * lnet_health_sensitivity determines by how much we decrement the health
  83  * value on sending error. The value defaults to 100, which means health
  84  * interface health is decremented by 100 points every failure.
  85  */
  86 unsigned int lnet_health_sensitivity = 100;
  87 static int sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
  88 #ifdef HAVE_KERNEL_PARAM_OPS
  89 static struct kernel_param_ops param_ops_health_sensitivity = {
  90         .set = sensitivity_set,
  91         .get = param_get_int,
  92 };
  93 #define param_check_health_sensitivity(name, p) \
  94                 __param_check(name, p, int)
  95 module_param(lnet_health_sensitivity, health_sensitivity, S_IRUGO|S_IWUSR);
  96 #else
  97 module_param_call(lnet_health_sensitivity, sensitivity_set, param_get_int,
  98                   &lnet_health_sensitivity, S_IRUGO|S_IWUSR);
  99 #endif
 100 MODULE_PARM_DESC(lnet_health_sensitivity,
 101                 "Value to decrement the health value by on error");
 102
 103 /*
 104  * lnet_recovery_interval determines how often we should perform recovery
 105  * on unhealthy interfaces.
 106  */
 107 unsigned int lnet_recovery_interval = 1;
 108 static int recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp);
 109 #ifdef HAVE_KERNEL_PARAM_OPS
 110 static struct kernel_param_ops param_ops_recovery_interval = {
 111         .set = recovery_interval_set,
 112         .get = param_get_int,
 113 };
 114 #define param_check_recovery_interval(name, p) \
 115                 __param_check(name, p, int)
 116 module_param(lnet_recovery_interval, recovery_interval, S_IRUGO|S_IWUSR);
 117 #else
 118 module_param_call(lnet_recovery_interval, recovery_interval_set, param_get_int,
 119                   &lnet_recovery_interval, S_IRUGO|S_IWUSR);
 120 #endif
 121 MODULE_PARM_DESC(lnet_recovery_interval,
 122                 "Interval to recover unhealthy interfaces in seconds");
 123
 124 static int lnet_interfaces_max = LNET_INTERFACES_MAX_DEFAULT;
 125 static int intf_max_set(const char *val, cfs_kernel_param_arg_t *kp);
 126
 127 static struct kernel_param_ops param_ops_interfaces_max = {
 128         .set = intf_max_set,
 129         .get = param_get_int,
 130 };
 131
 132 #define param_check_interfaces_max(name, p) \
 133                 __param_check(name, p, int)
 134
 135 #ifdef HAVE_KERNEL_PARAM_OPS
 136 module_param(lnet_interfaces_max, interfaces_max, 0644);
 137 #else
 138 module_param_call(lnet_interfaces_max, intf_max_set, param_get_int,
 139                   &param_ops_interfaces_max, 0644);
 140 #endif
 141 MODULE_PARM_DESC(lnet_interfaces_max,
 142                 "Maximum number of interfaces in a node.");
 143
 144 unsigned lnet_peer_discovery_disabled = 0;
 145 static int discovery_set(const char *val, cfs_kernel_param_arg_t *kp);
 146
 147 static struct kernel_param_ops param_ops_discovery_disabled = {
 148         .set = discovery_set,
 149         .get = param_get_int,
 150 };
 151
 152 #define param_check_discovery_disabled(name, p) \
 153                 __param_check(name, p, int)
 154 #ifdef HAVE_KERNEL_PARAM_OPS
 155 module_param(lnet_peer_discovery_disabled, discovery_disabled, 0644);
 156 #else
 157 module_param_call(lnet_peer_discovery_disabled, discovery_set, param_get_int,
 158                   &param_ops_discovery_disabled, 0644);
 159 #endif
 160 MODULE_PARM_DESC(lnet_peer_discovery_disabled,
 161                 "Set to 1 to disable peer discovery on this node.");
 162
 163 unsigned int lnet_drop_asym_route;
 164 static int drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp);
 165
 166 static struct kernel_param_ops param_ops_drop_asym_route = {
 167         .set = drop_asym_route_set,
 168         .get = param_get_int,
 169 };
 170
 171 #define param_check_drop_asym_route(name, p)    \
 172         __param_check(name, p, int)
 173 #ifdef HAVE_KERNEL_PARAM_OPS
 174 module_param(lnet_drop_asym_route, drop_asym_route, 0644);
 175 #else
 176 module_param_call(lnet_drop_asym_route, drop_asym_route_set, param_get_int,
 177                   &param_ops_drop_asym_route, 0644);
 178 #endif
 179 MODULE_PARM_DESC(lnet_drop_asym_route,
 180                  "Set to 1 to drop asymmetrical route messages.");
 181
 182 #define LNET_TRANSACTION_TIMEOUT_NO_HEALTH_DEFAULT 50
 183 #define LNET_TRANSACTION_TIMEOUT_HEALTH_DEFAULT 10
 184
 185 unsigned lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_HEALTH_DEFAULT;
 186 static int transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp);
 187 #ifdef HAVE_KERNEL_PARAM_OPS
 188 static struct kernel_param_ops param_ops_transaction_timeout = {
 189         .set = transaction_to_set,
 190         .get = param_get_int,
 191 };
 192
 193 #define param_check_transaction_timeout(name, p) \
 194                 __param_check(name, p, int)
 195 module_param(lnet_transaction_timeout, transaction_timeout, S_IRUGO|S_IWUSR);
 196 #else
 197 module_param_call(lnet_transaction_timeout, transaction_to_set, param_get_int,
 198                   &lnet_transaction_timeout, S_IRUGO|S_IWUSR);
 199 #endif
 200 MODULE_PARM_DESC(lnet_transaction_timeout,
 201                 "Maximum number of seconds to wait for a peer response.");
 202
 203 #define LNET_RETRY_COUNT_HEALTH_DEFAULT 3
 204 unsigned lnet_retry_count = LNET_RETRY_COUNT_HEALTH_DEFAULT;
 205 static int retry_count_set(const char *val, cfs_kernel_param_arg_t *kp);
 206 #ifdef HAVE_KERNEL_PARAM_OPS
 207 static struct kernel_param_ops param_ops_retry_count = {
 208         .set = retry_count_set,
 209         .get = param_get_int,
 210 };
 211
 212 #define param_check_retry_count(name, p) \
 213                 __param_check(name, p, int)
 214 module_param(lnet_retry_count, retry_count, S_IRUGO|S_IWUSR);
 215 #else
 216 module_param_call(lnet_retry_count, retry_count_set, param_get_int,
 217                   &lnet_retry_count, S_IRUGO|S_IWUSR);
 218 #endif
 219 MODULE_PARM_DESC(lnet_retry_count,
 220                  "Maximum number of times to retry transmitting a message");
 221
 222
 223 unsigned lnet_lnd_timeout = LNET_LND_DEFAULT_TIMEOUT;
 224
 225 /*
 226  * This sequence number keeps track of how many times DLC was used to
 227  * update the local NIs. It is incremented when a NI is added or
 228  * removed and checked when sending a message to determine if there is
 229  * a need to re-run the selection algorithm. See lnet_select_pathway()
 230  * for more details on its usage.
 231  */
 232 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
 233
 234 static int lnet_ping(struct lnet_process_id id, signed long timeout,
 235                      struct lnet_process_id __user *ids, int n_ids);
 236
 237 static int lnet_discover(struct lnet_process_id id, __u32 force,
 238                          struct lnet_process_id __user *ids, int n_ids);
 239
 240 static int
 241 sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 242 {
 243         int rc;
 244         unsigned *sensitivity = (unsigned *)kp->arg;
 245         unsigned long value;
 246
 247         rc = kstrtoul(val, 0, &value);
 248         if (rc) {
 249                 CERROR("Invalid module parameter value for 'lnet_health_sensitivity'\n");
 250                 return rc;
 251         }
 252
 253         /*
 254          * The purpose of locking the api_mutex here is to ensure that
 255          * the correct value ends up stored properly.
 256          */
 257         mutex_lock(&the_lnet.ln_api_mutex);
 258
 259         if (value > LNET_MAX_HEALTH_VALUE) {
 260                 mutex_unlock(&the_lnet.ln_api_mutex);
 261                 CERROR("Invalid health value. Maximum: %d value = %lu\n",
 262                        LNET_MAX_HEALTH_VALUE, value);
 263                 return -EINVAL;
 264         }
 265
 266         /*
 267          * if we're turning on health then use the health timeout
 268          * defaults.
 269          */
 270         if (*sensitivity == 0 && value != 0) {
 271                 lnet_transaction_timeout = LNET_TRANSACTION_TIMEOUT_HEALTH_DEFAULT;
 272                 lnet_retry_count = LNET_RETRY_COUNT_HEALTH_DEFAULT;
 273         /*
 274          * if we're turning off health then use the no health timeout
 275          * default.
 276          */
 277         } else if (*sensitivity != 0 && value == 0) {
 278                 lnet_transaction_timeout =
 279                         LNET_TRANSACTION_TIMEOUT_NO_HEALTH_DEFAULT;
 280                 lnet_retry_count = 0;
 281         }
 282
 283         *sensitivity = value;
 284
 285         mutex_unlock(&the_lnet.ln_api_mutex);
 286
 287         return 0;
 288 }
 289
 290 static int
 291 recovery_interval_set(const char *val, cfs_kernel_param_arg_t *kp)
 292 {
 293         int rc;
 294         unsigned *interval = (unsigned *)kp->arg;
 295         unsigned long value;
 296
 297         rc = kstrtoul(val, 0, &value);
 298         if (rc) {
 299                 CERROR("Invalid module parameter value for 'lnet_recovery_interval'\n");
 300                 return rc;
 301         }
 302
 303         if (value < 1) {
 304                 CERROR("lnet_recovery_interval must be at least 1 second\n");
 305                 return -EINVAL;
 306         }
 307
 308         /*
 309          * The purpose of locking the api_mutex here is to ensure that
 310          * the correct value ends up stored properly.
 311          */
 312         mutex_lock(&the_lnet.ln_api_mutex);
 313
 314         *interval = value;
 315
 316         mutex_unlock(&the_lnet.ln_api_mutex);
 317
 318         return 0;
 319 }
 320
 321 static int
 322 discovery_set(const char *val, cfs_kernel_param_arg_t *kp)
 323 {
 324         int rc;
 325         unsigned *discovery = (unsigned *)kp->arg;
 326         unsigned long value;
 327         struct lnet_ping_buffer *pbuf;
 328
 329         rc = kstrtoul(val, 0, &value);
 330         if (rc) {
 331                 CERROR("Invalid module parameter value for 'lnet_peer_discovery_disabled'\n");
 332                 return rc;
 333         }
 334
 335         value = (value) ? 1 : 0;
 336
 337         /*
 338          * The purpose of locking the api_mutex here is to ensure that
 339          * the correct value ends up stored properly.
 340          */
 341         mutex_lock(&the_lnet.ln_api_mutex);
 342
 343         if (value == *discovery) {
 344                 mutex_unlock(&the_lnet.ln_api_mutex);
 345                 return 0;
 346         }
 347
 348         *discovery = value;
 349
 350         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
 351                 mutex_unlock(&the_lnet.ln_api_mutex);
 352                 return 0;
 353         }
 354
 355         /* tell peers that discovery setting has changed */
 356         lnet_net_lock(LNET_LOCK_EX);
 357         pbuf = the_lnet.ln_ping_target;
 358         if (value)
 359                 pbuf->pb_info.pi_features &= ~LNET_PING_FEAT_DISCOVERY;
 360         else
 361                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
 362         lnet_net_unlock(LNET_LOCK_EX);
 363
 364         lnet_push_update_to_peers(1);
 365
 366         mutex_unlock(&the_lnet.ln_api_mutex);
 367
 368         return 0;
 369 }
 370
 371 static int
 372 drop_asym_route_set(const char *val, cfs_kernel_param_arg_t *kp)
 373 {
 374         int rc;
 375         unsigned int *drop_asym_route = (unsigned int *)kp->arg;
 376         unsigned long value;
 377
 378         rc = kstrtoul(val, 0, &value);
 379         if (rc) {
 380                 CERROR("Invalid module parameter value for "
 381                        "'lnet_drop_asym_route'\n");
 382                 return rc;
 383         }
 384
 385         /*
 386          * The purpose of locking the api_mutex here is to ensure that
 387          * the correct value ends up stored properly.
 388          */
 389         mutex_lock(&the_lnet.ln_api_mutex);
 390
 391         if (value == *drop_asym_route) {
 392                 mutex_unlock(&the_lnet.ln_api_mutex);
 393                 return 0;
 394         }
 395
 396         *drop_asym_route = value;
 397
 398         mutex_unlock(&the_lnet.ln_api_mutex);
 399
 400         return 0;
 401 }
 402
 403 static int
 404 transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp)
 405 {
 406         int rc;
 407         unsigned *transaction_to = (unsigned *)kp->arg;
 408         unsigned long value;
 409
 410         rc = kstrtoul(val, 0, &value);
 411         if (rc) {
 412                 CERROR("Invalid module parameter value for 'lnet_transaction_timeout'\n");
 413                 return rc;
 414         }
 415
 416         /*
 417          * The purpose of locking the api_mutex here is to ensure that
 418          * the correct value ends up stored properly.
 419          */
 420         mutex_lock(&the_lnet.ln_api_mutex);
 421
 422         if (value < lnet_retry_count || value == 0) {
 423                 mutex_unlock(&the_lnet.ln_api_mutex);
 424                 CERROR("Invalid value for lnet_transaction_timeout (%lu). "
 425                        "Has to be greater than lnet_retry_count (%u)\n",
 426                        value, lnet_retry_count);
 427                 return -EINVAL;
 428         }
 429
 430         if (value == *transaction_to) {
 431                 mutex_unlock(&the_lnet.ln_api_mutex);
 432                 return 0;
 433         }
 434
 435         *transaction_to = value;
 436         if (lnet_retry_count == 0)
 437                 lnet_lnd_timeout = value;
 438         else
 439                 lnet_lnd_timeout = value / lnet_retry_count;
 440
 441         mutex_unlock(&the_lnet.ln_api_mutex);
 442
 443         return 0;
 444 }
 445
 446 static int
 447 retry_count_set(const char *val, cfs_kernel_param_arg_t *kp)
 448 {
 449         int rc;
 450         unsigned *retry_count = (unsigned *)kp->arg;
 451         unsigned long value;
 452
 453         rc = kstrtoul(val, 0, &value);
 454         if (rc) {
 455                 CERROR("Invalid module parameter value for 'lnet_retry_count'\n");
 456                 return rc;
 457         }
 458
 459         /*
 460          * The purpose of locking the api_mutex here is to ensure that
 461          * the correct value ends up stored properly.
 462          */
 463         mutex_lock(&the_lnet.ln_api_mutex);
 464
 465         if (lnet_health_sensitivity == 0) {
 466                 mutex_unlock(&the_lnet.ln_api_mutex);
 467                 CERROR("Can not set retry_count when health feature is turned off\n");
 468                 return -EINVAL;
 469         }
 470
 471         if (value > lnet_transaction_timeout) {
 472                 mutex_unlock(&the_lnet.ln_api_mutex);
 473                 CERROR("Invalid value for lnet_retry_count (%lu). "
 474                        "Has to be smaller than lnet_transaction_timeout (%u)\n",
 475                        value, lnet_transaction_timeout);
 476                 return -EINVAL;
 477         }
 478
 479         *retry_count = value;
 480
 481         if (value == 0)
 482                 lnet_lnd_timeout = lnet_transaction_timeout;
 483         else
 484                 lnet_lnd_timeout = lnet_transaction_timeout / value;
 485
 486         mutex_unlock(&the_lnet.ln_api_mutex);
 487
 488         return 0;
 489 }
 490
 491 static int
 492 intf_max_set(const char *val, cfs_kernel_param_arg_t *kp)
 493 {
 494         int value, rc;
 495
 496         rc = kstrtoint(val, 0, &value);
 497         if (rc) {
 498                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
 499                 return rc;
 500         }
 501
 502         if (value < LNET_INTERFACES_MIN) {
 503                 CWARN("max interfaces provided are too small, setting to %d\n",
 504                       LNET_INTERFACES_MAX_DEFAULT);
 505                 value = LNET_INTERFACES_MAX_DEFAULT;
 506         }
 507
 508         *(int *)kp->arg = value;
 509
 510         return 0;
 511 }
 512
 513 static char *
 514 lnet_get_routes(void)
 515 {
 516         return routes;
 517 }
 518
 519 static char *
 520 lnet_get_networks(void)
 521 {
 522         char   *nets;
 523         int     rc;
 524
 525         if (*networks != 0 && *ip2nets != 0) {
 526                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
 527                                    "'ip2nets' but not both at once\n");
 528                 return NULL;
 529         }
 530
 531         if (*ip2nets != 0) {
 532                 rc = lnet_parse_ip2nets(&nets, ip2nets);
 533                 return (rc == 0) ? nets : NULL;
 534         }
 535
 536         if (*networks != 0)
 537                 return networks;
 538
 539         return "tcp";
 540 }
 541
 542 static void
 543 lnet_init_locks(void)
 544 {
 545         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 546         spin_lock_init(&the_lnet.ln_msg_resend_lock);
 547         init_waitqueue_head(&the_lnet.ln_eq_waitq);
 548         init_waitqueue_head(&the_lnet.ln_mt_waitq);
 549         mutex_init(&the_lnet.ln_lnd_mutex);
 550 }
 551
 552 static void
 553 lnet_fini_locks(void)
 554 {
 555 }
 556
 557 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
 558 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
 559                                             *  MDs kmem_cache */
 560
 561 static int
 562 lnet_descriptor_setup(void)
 563 {
 564         /* create specific kmem_cache for MEs and small MDs (i.e., originally
 565          * allocated in <size-xxx> kmem_cache).
 566          */
 567         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
 568                                             0, 0, NULL);
 569         if (!lnet_mes_cachep)
 570                 return -ENOMEM;
 571
 572         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
 573                                                   LNET_SMALL_MD_SIZE, 0, 0,
 574                                                   NULL);
 575         if (!lnet_small_mds_cachep)
 576                 return -ENOMEM;
 577
 578         return 0;
 579 }
 580
 581 static void
 582 lnet_descriptor_cleanup(void)
 583 {
 584
 585         if (lnet_small_mds_cachep) {
 586                 kmem_cache_destroy(lnet_small_mds_cachep);
 587                 lnet_small_mds_cachep = NULL;
 588         }
 589
 590         if (lnet_mes_cachep) {
 591                 kmem_cache_destroy(lnet_mes_cachep);
 592                 lnet_mes_cachep = NULL;
 593         }
 594 }
 595
 596 static int
 597 lnet_create_remote_nets_table(void)
 598 {
 599         int               i;
 600         struct list_head *hash;
 601
 602         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 603         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 604         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 605         if (hash == NULL) {
 606                 CERROR("Failed to create remote nets hash table\n");
 607                 return -ENOMEM;
 608         }
 609
 610         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 611                 INIT_LIST_HEAD(&hash[i]);
 612         the_lnet.ln_remote_nets_hash = hash;
 613         return 0;
 614 }
 615
 616 static void
 617 lnet_destroy_remote_nets_table(void)
 618 {
 619         int i;
 620
 621         if (the_lnet.ln_remote_nets_hash == NULL)
 622                 return;
 623
 624         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 625                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 626
 627         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 628                     LNET_REMOTE_NETS_HASH_SIZE *
 629                     sizeof(the_lnet.ln_remote_nets_hash[0]));
 630         the_lnet.ln_remote_nets_hash = NULL;
 631 }
 632
 633 static void
 634 lnet_destroy_locks(void)
 635 {
 636         if (the_lnet.ln_res_lock != NULL) {
 637                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 638                 the_lnet.ln_res_lock = NULL;
 639         }
 640
 641         if (the_lnet.ln_net_lock != NULL) {
 642                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 643                 the_lnet.ln_net_lock = NULL;
 644         }
 645
 646         lnet_fini_locks();
 647 }
 648
 649 static int
 650 lnet_create_locks(void)
 651 {
 652         lnet_init_locks();
 653
 654         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 655         if (the_lnet.ln_res_lock == NULL)
 656                 goto failed;
 657
 658         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 659         if (the_lnet.ln_net_lock == NULL)
 660                 goto failed;
 661
 662         return 0;
 663
 664  failed:
 665         lnet_destroy_locks();
 666         return -ENOMEM;
 667 }
 668
 669 static void lnet_assert_wire_constants(void)
 670 {
 671         /* Wire protocol assertions generated by 'wirecheck'
 672          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 673          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 674          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 675
 676         /* Constants... */
 677         CLASSERT(LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 678         CLASSERT(LNET_PROTO_TCP_VERSION_MAJOR == 1);
 679         CLASSERT(LNET_PROTO_TCP_VERSION_MINOR == 0);
 680         CLASSERT(LNET_MSG_ACK == 0);
 681         CLASSERT(LNET_MSG_PUT == 1);
 682         CLASSERT(LNET_MSG_GET == 2);
 683         CLASSERT(LNET_MSG_REPLY == 3);
 684         CLASSERT(LNET_MSG_HELLO == 4);
 685
 686         /* Checks for struct lnet_handle_wire */
 687         CLASSERT((int)sizeof(struct lnet_handle_wire) == 16);
 688         CLASSERT((int)offsetof(struct lnet_handle_wire, wh_interface_cookie) == 0);
 689         CLASSERT((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) == 8);
 690         CLASSERT((int)offsetof(struct lnet_handle_wire, wh_object_cookie) == 8);
 691         CLASSERT((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) == 8);
 692
 693         /* Checks for struct struct lnet_magicversion */
 694         CLASSERT((int)sizeof(struct lnet_magicversion) == 8);
 695         CLASSERT((int)offsetof(struct lnet_magicversion, magic) == 0);
 696         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->magic) == 4);
 697         CLASSERT((int)offsetof(struct lnet_magicversion, version_major) == 4);
 698         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->version_major) == 2);
 699         CLASSERT((int)offsetof(struct lnet_magicversion, version_minor) == 6);
 700         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->version_minor) == 2);
 701
 702         /* Checks for struct struct lnet_hdr */
 703         CLASSERT((int)sizeof(struct lnet_hdr) == 72);
 704         CLASSERT((int)offsetof(struct lnet_hdr, dest_nid) == 0);
 705         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->dest_nid) == 8);
 706         CLASSERT((int)offsetof(struct lnet_hdr, src_nid) == 8);
 707         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->src_nid) == 8);
 708         CLASSERT((int)offsetof(struct lnet_hdr, dest_pid) == 16);
 709         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->dest_pid) == 4);
 710         CLASSERT((int)offsetof(struct lnet_hdr, src_pid) == 20);
 711         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->src_pid) == 4);
 712         CLASSERT((int)offsetof(struct lnet_hdr, type) == 24);
 713         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->type) == 4);
 714         CLASSERT((int)offsetof(struct lnet_hdr, payload_length) == 28);
 715         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->payload_length) == 4);
 716         CLASSERT((int)offsetof(struct lnet_hdr, msg) == 32);
 717         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg) == 40);
 718
 719         /* Ack */
 720         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.dst_wmd) == 32);
 721         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.dst_wmd) == 16);
 722         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.match_bits) == 48);
 723         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.match_bits) == 8);
 724         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.mlength) == 56);
 725         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.mlength) == 4);
 726
 727         /* Put */
 728         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.ack_wmd) == 32);
 729         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.ack_wmd) == 16);
 730         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.match_bits) == 48);
 731         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.match_bits) == 8);
 732         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.hdr_data) == 56);
 733         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.hdr_data) == 8);
 734         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.ptl_index) == 64);
 735         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.ptl_index) == 4);
 736         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.offset) == 68);
 737         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.offset) == 4);
 738
 739         /* Get */
 740         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.return_wmd) == 32);
 741         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.return_wmd) == 16);
 742         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.match_bits) == 48);
 743         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.match_bits) == 8);
 744         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.ptl_index) == 56);
 745         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.ptl_index) == 4);
 746         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.src_offset) == 60);
 747         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.src_offset) == 4);
 748         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.sink_length) == 64);
 749         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.sink_length) == 4);
 750
 751         /* Reply */
 752         CLASSERT((int)offsetof(struct lnet_hdr, msg.reply.dst_wmd) == 32);
 753         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.reply.dst_wmd) == 16);
 754
 755         /* Hello */
 756         CLASSERT((int)offsetof(struct lnet_hdr, msg.hello.incarnation) == 32);
 757         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.hello.incarnation) == 8);
 758         CLASSERT((int)offsetof(struct lnet_hdr, msg.hello.type) == 40);
 759         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.hello.type) == 4);
 760
 761         /* Checks for struct lnet_ni_status and related constants */
 762         CLASSERT(LNET_NI_STATUS_INVALID == 0x00000000);
 763         CLASSERT(LNET_NI_STATUS_UP == 0x15aac0de);
 764         CLASSERT(LNET_NI_STATUS_DOWN == 0xdeadface);
 765
 766         /* Checks for struct lnet_ni_status */
 767         CLASSERT((int)sizeof(struct lnet_ni_status) == 16);
 768         CLASSERT((int)offsetof(struct lnet_ni_status, ns_nid) == 0);
 769         CLASSERT((int)sizeof(((struct lnet_ni_status *)0)->ns_nid) == 8);
 770         CLASSERT((int)offsetof(struct lnet_ni_status, ns_status) == 8);
 771         CLASSERT((int)sizeof(((struct lnet_ni_status *)0)->ns_status) == 4);
 772         CLASSERT((int)offsetof(struct lnet_ni_status, ns_unused) == 12);
 773         CLASSERT((int)sizeof(((struct lnet_ni_status *)0)->ns_unused) == 4);
 774
 775         /* Checks for struct lnet_ping_info and related constants */
 776         CLASSERT(LNET_PROTO_PING_MAGIC == 0x70696E67);
 777         CLASSERT(LNET_PING_FEAT_INVAL == 0);
 778         CLASSERT(LNET_PING_FEAT_BASE == 1);
 779         CLASSERT(LNET_PING_FEAT_NI_STATUS == 2);
 780         CLASSERT(LNET_PING_FEAT_RTE_DISABLED == 4);
 781         CLASSERT(LNET_PING_FEAT_MULTI_RAIL == 8);
 782         CLASSERT(LNET_PING_FEAT_DISCOVERY == 16);
 783         CLASSERT(LNET_PING_FEAT_BITS == 31);
 784
 785         /* Checks for struct lnet_ping_info */
 786         CLASSERT((int)sizeof(struct lnet_ping_info) == 16);
 787         CLASSERT((int)offsetof(struct lnet_ping_info, pi_magic) == 0);
 788         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_magic) == 4);
 789         CLASSERT((int)offsetof(struct lnet_ping_info, pi_features) == 4);
 790         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_features) == 4);
 791         CLASSERT((int)offsetof(struct lnet_ping_info, pi_pid) == 8);
 792         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_pid) == 4);
 793         CLASSERT((int)offsetof(struct lnet_ping_info, pi_nnis) == 12);
 794         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_nnis) == 4);
 795         CLASSERT((int)offsetof(struct lnet_ping_info, pi_ni) == 16);
 796         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_ni) == 0);
 797 }
 798
 799 static struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
 800 {
 801         struct lnet_lnd *lnd;
 802         struct list_head *tmp;
 803
 804         /* holding lnd mutex */
 805         list_for_each(tmp, &the_lnet.ln_lnds) {
 806                 lnd = list_entry(tmp, struct lnet_lnd, lnd_list);
 807
 808                 if (lnd->lnd_type == type)
 809                         return lnd;
 810         }
 811         return NULL;
 812 }
 813
 814 unsigned int
 815 lnet_get_lnd_timeout(void)
 816 {
 817         return lnet_lnd_timeout;
 818 }
 819 EXPORT_SYMBOL(lnet_get_lnd_timeout);
 820
 821 void
 822 lnet_register_lnd(struct lnet_lnd *lnd)
 823 {
 824         mutex_lock(&the_lnet.ln_lnd_mutex);
 825
 826         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 827         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 828
 829         list_add_tail(&lnd->lnd_list, &the_lnet.ln_lnds);
 830         lnd->lnd_refcount = 0;
 831
 832         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 833
 834         mutex_unlock(&the_lnet.ln_lnd_mutex);
 835 }
 836 EXPORT_SYMBOL(lnet_register_lnd);
 837
 838 void
 839 lnet_unregister_lnd(struct lnet_lnd *lnd)
 840 {
 841         mutex_lock(&the_lnet.ln_lnd_mutex);
 842
 843         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 844         LASSERT(lnd->lnd_refcount == 0);
 845
 846         list_del(&lnd->lnd_list);
 847         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 848
 849         mutex_unlock(&the_lnet.ln_lnd_mutex);
 850 }
 851 EXPORT_SYMBOL(lnet_unregister_lnd);
 852
 853 void
 854 lnet_counters_get_common(struct lnet_counters_common *common)
 855 {
 856         struct lnet_counters *ctr;
 857         int i;
 858
 859         memset(common, 0, sizeof(*common));
 860
 861         lnet_net_lock(LNET_LOCK_EX);
 862
 863         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 864                 common->lcc_msgs_max     += ctr->lct_common.lcc_msgs_max;
 865                 common->lcc_msgs_alloc   += ctr->lct_common.lcc_msgs_alloc;
 866                 common->lcc_errors       += ctr->lct_common.lcc_errors;
 867                 common->lcc_send_count   += ctr->lct_common.lcc_send_count;
 868                 common->lcc_recv_count   += ctr->lct_common.lcc_recv_count;
 869                 common->lcc_route_count  += ctr->lct_common.lcc_route_count;
 870                 common->lcc_drop_count   += ctr->lct_common.lcc_drop_count;
 871                 common->lcc_send_length  += ctr->lct_common.lcc_send_length;
 872                 common->lcc_recv_length  += ctr->lct_common.lcc_recv_length;
 873                 common->lcc_route_length += ctr->lct_common.lcc_route_length;
 874                 common->lcc_drop_length  += ctr->lct_common.lcc_drop_length;
 875         }
 876         lnet_net_unlock(LNET_LOCK_EX);
 877 }
 878 EXPORT_SYMBOL(lnet_counters_get_common);
 879
 880 void
 881 lnet_counters_get(struct lnet_counters *counters)
 882 {
 883         struct lnet_counters *ctr;
 884         struct lnet_counters_health *health = &counters->lct_health;
 885         int             i;
 886
 887         memset(counters, 0, sizeof(*counters));
 888
 889         lnet_counters_get_common(&counters->lct_common);
 890
 891         lnet_net_lock(LNET_LOCK_EX);
 892
 893         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 894                 health->lch_rst_alloc    += ctr->lct_health.lch_rst_alloc;
 895                 health->lch_resend_count += ctr->lct_health.lch_resend_count;
 896                 health->lch_response_timeout_count +=
 897                                 ctr->lct_health.lch_response_timeout_count;
 898                 health->lch_local_interrupt_count +=
 899                                 ctr->lct_health.lch_local_interrupt_count;
 900                 health->lch_local_dropped_count +=
 901                                 ctr->lct_health.lch_local_dropped_count;
 902                 health->lch_local_aborted_count +=
 903                                 ctr->lct_health.lch_local_aborted_count;
 904                 health->lch_local_no_route_count +=
 905                                 ctr->lct_health.lch_local_no_route_count;
 906                 health->lch_local_timeout_count +=
 907                                 ctr->lct_health.lch_local_timeout_count;
 908                 health->lch_local_error_count +=
 909                                 ctr->lct_health.lch_local_error_count;
 910                 health->lch_remote_dropped_count +=
 911                                 ctr->lct_health.lch_remote_dropped_count;
 912                 health->lch_remote_error_count +=
 913                                 ctr->lct_health.lch_remote_error_count;
 914                 health->lch_remote_timeout_count +=
 915                                 ctr->lct_health.lch_remote_timeout_count;
 916                 health->lch_network_timeout_count +=
 917                                 ctr->lct_health.lch_network_timeout_count;
 918         }
 919         lnet_net_unlock(LNET_LOCK_EX);
 920 }
 921 EXPORT_SYMBOL(lnet_counters_get);
 922
 923 void
 924 lnet_counters_reset(void)
 925 {
 926         struct lnet_counters *counters;
 927         int             i;
 928
 929         lnet_net_lock(LNET_LOCK_EX);
 930
 931         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 932                 memset(counters, 0, sizeof(struct lnet_counters));
 933
 934         lnet_net_unlock(LNET_LOCK_EX);
 935 }
 936
 937 static char *
 938 lnet_res_type2str(int type)
 939 {
 940         switch (type) {
 941         default:
 942                 LBUG();
 943         case LNET_COOKIE_TYPE_MD:
 944                 return "MD";
 945         case LNET_COOKIE_TYPE_ME:
 946                 return "ME";
 947         case LNET_COOKIE_TYPE_EQ:
 948                 return "EQ";
 949         }
 950 }
 951
 952 static void
 953 lnet_res_container_cleanup(struct lnet_res_container *rec)
 954 {
 955         int     count = 0;
 956
 957         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 958                 return;
 959
 960         while (!list_empty(&rec->rec_active)) {
 961                 struct list_head *e = rec->rec_active.next;
 962
 963                 list_del_init(e);
 964                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 965                         lnet_eq_free(list_entry(e, struct lnet_eq, eq_list));
 966
 967                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 968                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
 969
 970                 } else { /* NB: Active MEs should be attached on portals */
 971                         LBUG();
 972                 }
 973                 count++;
 974         }
 975
 976         if (count > 0) {
 977                 /* Found alive MD/ME/EQ, user really should unlink/free
 978                  * all of them before finalize LNet, but if someone didn't,
 979                  * we have to recycle garbage for him */
 980                 CERROR("%d active elements on exit of %s container\n",
 981                        count, lnet_res_type2str(rec->rec_type));
 982         }
 983
 984         if (rec->rec_lh_hash != NULL) {
 985                 LIBCFS_FREE(rec->rec_lh_hash,
 986                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 987                 rec->rec_lh_hash = NULL;
 988         }
 989
 990         rec->rec_type = 0; /* mark it as finalized */
 991 }
 992
 993 static int
 994 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
 995 {
 996         int     rc = 0;
 997         int     i;
 998
 999         LASSERT(rec->rec_type == 0);
1000
1001         rec->rec_type = type;
1002         INIT_LIST_HEAD(&rec->rec_active);
1003
1004         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
1005
1006         /* Arbitrary choice of hash table size */
1007         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
1008                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
1009         if (rec->rec_lh_hash == NULL) {
1010                 rc = -ENOMEM;
1011                 goto out;
1012         }
1013
1014         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
1015                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
1016
1017         return 0;
1018
1019 out:
1020         CERROR("Failed to setup %s resource container\n",
1021                lnet_res_type2str(type));
1022         lnet_res_container_cleanup(rec);
1023         return rc;
1024 }
1025
1026 static void
1027 lnet_res_containers_destroy(struct lnet_res_container **recs)
1028 {
1029         struct lnet_res_container       *rec;
1030         int                             i;
1031
1032         cfs_percpt_for_each(rec, i, recs)
1033                 lnet_res_container_cleanup(rec);
1034
1035         cfs_percpt_free(recs);
1036 }
1037
1038 static struct lnet_res_container **
1039 lnet_res_containers_create(int type)
1040 {
1041         struct lnet_res_container       **recs;
1042         struct lnet_res_container       *rec;
1043         int                             rc;
1044         int                             i;
1045
1046         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
1047         if (recs == NULL) {
1048                 CERROR("Failed to allocate %s resource containers\n",
1049                        lnet_res_type2str(type));
1050                 return NULL;
1051         }
1052
1053         cfs_percpt_for_each(rec, i, recs) {
1054                 rc = lnet_res_container_setup(rec, i, type);
1055                 if (rc != 0) {
1056                         lnet_res_containers_destroy(recs);
1057                         return NULL;
1058                 }
1059         }
1060
1061         return recs;
1062 }
1063
1064 struct lnet_libhandle *
1065 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
1066 {
1067         /* ALWAYS called with lnet_res_lock held */
1068         struct list_head        *head;
1069         struct lnet_libhandle   *lh;
1070         unsigned int            hash;
1071
1072         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
1073                 return NULL;
1074
1075         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
1076         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
1077
1078         list_for_each_entry(lh, head, lh_hash_chain) {
1079                 if (lh->lh_cookie == cookie)
1080                         return lh;
1081         }
1082
1083         return NULL;
1084 }
1085
1086 void
1087 lnet_res_lh_initialize(struct lnet_res_container *rec,
1088                        struct lnet_libhandle *lh)
1089 {
1090         /* ALWAYS called with lnet_res_lock held */
1091         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
1092         unsigned int    hash;
1093
1094         lh->lh_cookie = rec->rec_lh_cookie;
1095         rec->rec_lh_cookie += 1 << ibits;
1096
1097         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
1098
1099         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
1100 }
1101
1102 static int lnet_unprepare(void);
1103
1104 static int
1105 lnet_prepare(lnet_pid_t requested_pid)
1106 {
1107         /* Prepare to bring up the network */
1108         struct lnet_res_container **recs;
1109         int                       rc = 0;
1110
1111         if (requested_pid == LNET_PID_ANY) {
1112                 /* Don't instantiate LNET just for me */
1113                 return -ENETDOWN;
1114         }
1115
1116         LASSERT(the_lnet.ln_refcount == 0);
1117
1118         the_lnet.ln_routing = 0;
1119
1120         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
1121         the_lnet.ln_pid = requested_pid;
1122
1123         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
1124         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
1125         INIT_LIST_HEAD(&the_lnet.ln_nets);
1126         INIT_LIST_HEAD(&the_lnet.ln_routers);
1127         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
1128         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
1129         INIT_LIST_HEAD(&the_lnet.ln_dc_request);
1130         INIT_LIST_HEAD(&the_lnet.ln_dc_working);
1131         INIT_LIST_HEAD(&the_lnet.ln_dc_expired);
1132         INIT_LIST_HEAD(&the_lnet.ln_mt_localNIRecovq);
1133         INIT_LIST_HEAD(&the_lnet.ln_mt_peerNIRecovq);
1134         init_waitqueue_head(&the_lnet.ln_dc_waitq);
1135         LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
1136
1137         rc = lnet_descriptor_setup();
1138         if (rc != 0)
1139                 goto failed;
1140
1141         rc = lnet_create_remote_nets_table();
1142         if (rc != 0)
1143                 goto failed;
1144
1145         /*
1146          * NB the interface cookie in wire handles guards against delayed
1147          * replies and ACKs appearing valid after reboot.
1148          */
1149         the_lnet.ln_interface_cookie = ktime_get_real_ns();
1150
1151         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
1152                                                 sizeof(struct lnet_counters));
1153         if (the_lnet.ln_counters == NULL) {
1154                 CERROR("Failed to allocate counters for LNet\n");
1155                 rc = -ENOMEM;
1156                 goto failed;
1157         }
1158
1159         rc = lnet_peer_tables_create();
1160         if (rc != 0)
1161                 goto failed;
1162
1163         rc = lnet_msg_containers_create();
1164         if (rc != 0)
1165                 goto failed;
1166
1167         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
1168                                       LNET_COOKIE_TYPE_EQ);
1169         if (rc != 0)
1170                 goto failed;
1171
1172         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME);
1173         if (recs == NULL) {
1174                 rc = -ENOMEM;
1175                 goto failed;
1176         }
1177
1178         the_lnet.ln_me_containers = recs;
1179
1180         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
1181         if (recs == NULL) {
1182                 rc = -ENOMEM;
1183                 goto failed;
1184         }
1185
1186         the_lnet.ln_md_containers = recs;
1187
1188         rc = lnet_portals_create();
1189         if (rc != 0) {
1190                 CERROR("Failed to create portals for LNet: %d\n", rc);
1191                 goto failed;
1192         }
1193
1194         return 0;
1195
1196  failed:
1197         lnet_unprepare();
1198         return rc;
1199 }
1200
1201 static int
1202 lnet_unprepare (void)
1203 {
1204         int rc;
1205
1206         /* NB no LNET_LOCK since this is the last reference.  All LND instances
1207          * have shut down already, so it is safe to unlink and free all
1208          * descriptors, even those that appear committed to a network op (eg MD
1209          * with non-zero pending count) */
1210
1211         lnet_fail_nid(LNET_NID_ANY, 0);
1212
1213         LASSERT(the_lnet.ln_refcount == 0);
1214         LASSERT(list_empty(&the_lnet.ln_test_peers));
1215         LASSERT(list_empty(&the_lnet.ln_nets));
1216
1217         if (!LNetEQHandleIsInvalid(the_lnet.ln_mt_eqh)) {
1218                 rc = LNetEQFree(the_lnet.ln_mt_eqh);
1219                 LNetInvalidateEQHandle(&the_lnet.ln_mt_eqh);
1220                 LASSERT(rc == 0);
1221         }
1222
1223         lnet_portals_destroy();
1224
1225         if (the_lnet.ln_md_containers != NULL) {
1226                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
1227                 the_lnet.ln_md_containers = NULL;
1228         }
1229
1230         if (the_lnet.ln_me_containers != NULL) {
1231                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
1232                 the_lnet.ln_me_containers = NULL;
1233         }
1234
1235         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
1236
1237         lnet_msg_containers_destroy();
1238         lnet_peer_uninit();
1239         lnet_rtrpools_free(0);
1240
1241         if (the_lnet.ln_counters != NULL) {
1242                 cfs_percpt_free(the_lnet.ln_counters);
1243                 the_lnet.ln_counters = NULL;
1244         }
1245         lnet_destroy_remote_nets_table();
1246         lnet_descriptor_cleanup();
1247
1248         return 0;
1249 }
1250
1251 struct lnet_ni  *
1252 lnet_net2ni_locked(__u32 net_id, int cpt)
1253 {
1254         struct lnet_ni   *ni;
1255         struct lnet_net  *net;
1256
1257         LASSERT(cpt != LNET_LOCK_EX);
1258
1259         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1260                 if (net->net_id == net_id) {
1261                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
1262                                         ni_netlist);
1263                         return ni;
1264                 }
1265         }
1266
1267         return NULL;
1268 }
1269
1270 struct lnet_ni *
1271 lnet_net2ni_addref(__u32 net)
1272 {
1273         struct lnet_ni *ni;
1274
1275         lnet_net_lock(0);
1276         ni = lnet_net2ni_locked(net, 0);
1277         if (ni)
1278                 lnet_ni_addref_locked(ni, 0);
1279         lnet_net_unlock(0);
1280
1281         return ni;
1282 }
1283 EXPORT_SYMBOL(lnet_net2ni_addref);
1284
1285 struct lnet_net *
1286 lnet_get_net_locked(__u32 net_id)
1287 {
1288         struct lnet_net  *net;
1289
1290         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1291                 if (net->net_id == net_id)
1292                         return net;
1293         }
1294
1295         return NULL;
1296 }
1297
1298 unsigned int
1299 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
1300 {
1301         __u64           key = nid;
1302         unsigned int    val;
1303
1304         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1305
1306         if (number == 1)
1307                 return 0;
1308
1309         val = hash_long(key, LNET_CPT_BITS);
1310         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
1311         if (val < number)
1312                 return val;
1313
1314         return (unsigned int)(key + val + (val >> 1)) % number;
1315 }
1316
1317 int
1318 lnet_cpt_of_nid_locked(lnet_nid_t nid, struct lnet_ni *ni)
1319 {
1320         struct lnet_net *net;
1321
1322         /* must called with hold of lnet_net_lock */
1323         if (LNET_CPT_NUMBER == 1)
1324                 return 0; /* the only one */
1325
1326         /*
1327          * If NI is provided then use the CPT identified in the NI cpt
1328          * list if one exists. If one doesn't exist, then that NI is
1329          * associated with all CPTs and it follows that the net it belongs
1330          * to is implicitly associated with all CPTs, so just hash the nid
1331          * and return that.
1332          */
1333         if (ni != NULL) {
1334                 if (ni->ni_cpts != NULL)
1335                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
1336                                                              ni->ni_ncpts)];
1337                 else
1338                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1339         }
1340
1341         /* no NI provided so look at the net */
1342         net = lnet_get_net_locked(LNET_NIDNET(nid));
1343
1344         if (net != NULL && net->net_cpts != NULL) {
1345                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
1346         }
1347
1348         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1349 }
1350
1351 int
1352 lnet_cpt_of_nid(lnet_nid_t nid, struct lnet_ni *ni)
1353 {
1354         int     cpt;
1355         int     cpt2;
1356
1357         if (LNET_CPT_NUMBER == 1)
1358                 return 0; /* the only one */
1359
1360         cpt = lnet_net_lock_current();
1361
1362         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
1363
1364         lnet_net_unlock(cpt);
1365
1366         return cpt2;
1367 }
1368 EXPORT_SYMBOL(lnet_cpt_of_nid);
1369
1370 int
1371 lnet_islocalnet(__u32 net_id)
1372 {
1373         struct lnet_net *net;
1374         int             cpt;
1375         bool            local;
1376
1377         cpt = lnet_net_lock_current();
1378
1379         net = lnet_get_net_locked(net_id);
1380
1381         local = net != NULL;
1382
1383         lnet_net_unlock(cpt);
1384
1385         return local;
1386 }
1387
1388 struct lnet_ni  *
1389 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
1390 {
1391         struct lnet_net  *net;
1392         struct lnet_ni   *ni;
1393
1394         LASSERT(cpt != LNET_LOCK_EX);
1395
1396         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1397                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1398                         if (ni->ni_nid == nid)
1399                                 return ni;
1400                 }
1401         }
1402
1403         return NULL;
1404 }
1405
1406 struct lnet_ni *
1407 lnet_nid2ni_addref(lnet_nid_t nid)
1408 {
1409         struct lnet_ni *ni;
1410
1411         lnet_net_lock(0);
1412         ni = lnet_nid2ni_locked(nid, 0);
1413         if (ni)
1414                 lnet_ni_addref_locked(ni, 0);
1415         lnet_net_unlock(0);
1416
1417         return ni;
1418 }
1419 EXPORT_SYMBOL(lnet_nid2ni_addref);
1420
1421 int
1422 lnet_islocalnid(lnet_nid_t nid)
1423 {
1424         struct lnet_ni  *ni;
1425         int             cpt;
1426
1427         cpt = lnet_net_lock_current();
1428         ni = lnet_nid2ni_locked(nid, cpt);
1429         lnet_net_unlock(cpt);
1430
1431         return ni != NULL;
1432 }
1433
1434 int
1435 lnet_count_acceptor_nets(void)
1436 {
1437         /* Return the # of NIs that need the acceptor. */
1438         int              count = 0;
1439         struct lnet_net  *net;
1440         int              cpt;
1441
1442         cpt = lnet_net_lock_current();
1443         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1444                 /* all socklnd type networks should have the acceptor
1445                  * thread started */
1446                 if (net->net_lnd->lnd_accept != NULL)
1447                         count++;
1448         }
1449
1450         lnet_net_unlock(cpt);
1451
1452         return count;
1453 }
1454
1455 struct lnet_ping_buffer *
1456 lnet_ping_buffer_alloc(int nnis, gfp_t gfp)
1457 {
1458         struct lnet_ping_buffer *pbuf;
1459
1460         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nnis), gfp);
1461         if (pbuf) {
1462                 pbuf->pb_nnis = nnis;
1463                 atomic_set(&pbuf->pb_refcnt, 1);
1464         }
1465
1466         return pbuf;
1467 }
1468
1469 void
1470 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
1471 {
1472         LASSERT(lnet_ping_buffer_numref(pbuf) == 0);
1473         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nnis));
1474 }
1475
1476 static struct lnet_ping_buffer *
1477 lnet_ping_target_create(int nnis)
1478 {
1479         struct lnet_ping_buffer *pbuf;
1480
1481         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1482         if (pbuf == NULL) {
1483                 CERROR("Can't allocate ping source [%d]\n", nnis);
1484                 return NULL;
1485         }
1486
1487         pbuf->pb_info.pi_nnis = nnis;
1488         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
1489         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
1490         pbuf->pb_info.pi_features =
1491                 LNET_PING_FEAT_NI_STATUS | LNET_PING_FEAT_MULTI_RAIL;
1492
1493         return pbuf;
1494 }
1495
1496 static inline int
1497 lnet_get_net_ni_count_locked(struct lnet_net *net)
1498 {
1499         struct lnet_ni  *ni;
1500         int             count = 0;
1501
1502         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1503                 count++;
1504
1505         return count;
1506 }
1507
1508 static inline int
1509 lnet_get_net_ni_count_pre(struct lnet_net *net)
1510 {
1511         struct lnet_ni  *ni;
1512         int             count = 0;
1513
1514         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
1515                 count++;
1516
1517         return count;
1518 }
1519
1520 static inline int
1521 lnet_get_ni_count(void)
1522 {
1523         struct lnet_ni  *ni;
1524         struct lnet_net *net;
1525         int             count = 0;
1526
1527         lnet_net_lock(0);
1528
1529         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1530                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1531                         count++;
1532         }
1533
1534         lnet_net_unlock(0);
1535
1536         return count;
1537 }
1538
1539 void
1540 lnet_swap_pinginfo(struct lnet_ping_buffer *pbuf)
1541 {
1542         struct lnet_ni_status *stat;
1543         int nnis;
1544         int i;
1545
1546         __swab32s(&pbuf->pb_info.pi_magic);
1547         __swab32s(&pbuf->pb_info.pi_features);
1548         __swab32s(&pbuf->pb_info.pi_pid);
1549         __swab32s(&pbuf->pb_info.pi_nnis);
1550         nnis = pbuf->pb_info.pi_nnis;
1551         if (nnis > pbuf->pb_nnis)
1552                 nnis = pbuf->pb_nnis;
1553         for (i = 0; i < nnis; i++) {
1554                 stat = &pbuf->pb_info.pi_ni[i];
1555                 __swab64s(&stat->ns_nid);
1556                 __swab32s(&stat->ns_status);
1557         }
1558         return;
1559 }
1560
1561 int
1562 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1563 {
1564         if (!pinfo)
1565                 return -EINVAL;
1566         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1567                 return -EPROTO;
1568         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1569                 return -EPROTO;
1570         /* Loopback is guaranteed to be present */
1571         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1572                 return -ERANGE;
1573         if (LNET_NETTYP(LNET_NIDNET(LNET_PING_INFO_LONI(pinfo))) != LOLND)
1574                 return -EPROTO;
1575         return 0;
1576 }
1577
1578 static void
1579 lnet_ping_target_destroy(void)
1580 {
1581         struct lnet_net *net;
1582         struct lnet_ni  *ni;
1583
1584         lnet_net_lock(LNET_LOCK_EX);
1585
1586         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1587                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1588                         lnet_ni_lock(ni);
1589                         ni->ni_status = NULL;
1590                         lnet_ni_unlock(ni);
1591                 }
1592         }
1593
1594         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1595         the_lnet.ln_ping_target = NULL;
1596
1597         lnet_net_unlock(LNET_LOCK_EX);
1598 }
1599
1600 static void
1601 lnet_ping_target_event_handler(struct lnet_event *event)
1602 {
1603         struct lnet_ping_buffer *pbuf = event->md.user_ptr;
1604
1605         if (event->unlinked)
1606                 lnet_ping_buffer_decref(pbuf);
1607 }
1608
1609 static int
1610 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1611                        struct lnet_handle_md *ping_mdh,
1612                        int ni_count, bool set_eq)
1613 {
1614         struct lnet_process_id id = {
1615                 .nid = LNET_NID_ANY,
1616                 .pid = LNET_PID_ANY
1617         };
1618         struct lnet_handle_me me_handle;
1619         struct lnet_md md = { NULL };
1620         int rc, rc2;
1621
1622         if (set_eq) {
1623                 rc = LNetEQAlloc(0, lnet_ping_target_event_handler,
1624                                  &the_lnet.ln_ping_target_eq);
1625                 if (rc != 0) {
1626                         CERROR("Can't allocate ping buffer EQ: %d\n", rc);
1627                         return rc;
1628                 }
1629         }
1630
1631         *ppbuf = lnet_ping_target_create(ni_count);
1632         if (*ppbuf == NULL) {
1633                 rc = -ENOMEM;
1634                 goto fail_free_eq;
1635         }
1636
1637         /* Ping target ME/MD */
1638         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1639                           LNET_PROTO_PING_MATCHBITS, 0,
1640                           LNET_UNLINK, LNET_INS_AFTER,
1641                           &me_handle);
1642         if (rc != 0) {
1643                 CERROR("Can't create ping target ME: %d\n", rc);
1644                 goto fail_decref_ping_buffer;
1645         }
1646
1647         /* initialize md content */
1648         md.start     = &(*ppbuf)->pb_info;
1649         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1650         md.threshold = LNET_MD_THRESH_INF;
1651         md.max_size  = 0;
1652         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1653                        LNET_MD_MANAGE_REMOTE;
1654         md.eq_handle = the_lnet.ln_ping_target_eq;
1655         md.user_ptr  = *ppbuf;
1656
1657         rc = LNetMDAttach(me_handle, md, LNET_RETAIN, ping_mdh);
1658         if (rc != 0) {
1659                 CERROR("Can't attach ping target MD: %d\n", rc);
1660                 goto fail_unlink_ping_me;
1661         }
1662         lnet_ping_buffer_addref(*ppbuf);
1663
1664         return 0;
1665
1666 fail_unlink_ping_me:
1667         rc2 = LNetMEUnlink(me_handle);
1668         LASSERT(rc2 == 0);
1669 fail_decref_ping_buffer:
1670         LASSERT(lnet_ping_buffer_numref(*ppbuf) == 1);
1671         lnet_ping_buffer_decref(*ppbuf);
1672         *ppbuf = NULL;
1673 fail_free_eq:
1674         if (set_eq) {
1675                 rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1676                 LASSERT(rc2 == 0);
1677         }
1678         return rc;
1679 }
1680
1681 static void
1682 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1683                     struct lnet_handle_md *ping_mdh)
1684 {
1685         sigset_t        blocked = cfs_block_allsigs();
1686
1687         LNetMDUnlink(*ping_mdh);
1688         LNetInvalidateMDHandle(ping_mdh);
1689
1690         /* NB the MD could be busy; this just starts the unlink */
1691         while (lnet_ping_buffer_numref(pbuf) > 1) {
1692                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
1693                 set_current_state(TASK_UNINTERRUPTIBLE);
1694                 schedule_timeout(cfs_time_seconds(1));
1695         }
1696
1697         cfs_restore_sigs(blocked);
1698 }
1699
1700 static void
1701 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
1702 {
1703         struct lnet_ni          *ni;
1704         struct lnet_net         *net;
1705         struct lnet_ni_status *ns;
1706         int                     i;
1707         int                     rc;
1708
1709         i = 0;
1710         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1711                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1712                         LASSERT(i < pbuf->pb_nnis);
1713
1714                         ns = &pbuf->pb_info.pi_ni[i];
1715
1716                         ns->ns_nid = ni->ni_nid;
1717
1718                         lnet_ni_lock(ni);
1719                         ns->ns_status = (ni->ni_status != NULL) ?
1720                                          ni->ni_status->ns_status :
1721                                                 LNET_NI_STATUS_UP;
1722                         ni->ni_status = ns;
1723                         lnet_ni_unlock(ni);
1724
1725                         i++;
1726                 }
1727         }
1728         /*
1729          * We (ab)use the ns_status of the loopback interface to
1730          * transmit the sequence number. The first interface listed
1731          * must be the loopback interface.
1732          */
1733         rc = lnet_ping_info_validate(&pbuf->pb_info);
1734         if (rc) {
1735                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
1736                 LBUG();
1737         }
1738         LNET_PING_BUFFER_SEQNO(pbuf) =
1739                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
1740 }
1741
1742 static void
1743 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
1744                         struct lnet_handle_md ping_mdh)
1745 {
1746         struct lnet_ping_buffer *old_pbuf = NULL;
1747         struct lnet_handle_md old_ping_md;
1748
1749         /* switch the NIs to point to the new ping info created */
1750         lnet_net_lock(LNET_LOCK_EX);
1751
1752         if (!the_lnet.ln_routing)
1753                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1754         if (!lnet_peer_discovery_disabled)
1755                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
1756
1757         /* Ensure only known feature bits have been set. */
1758         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
1759         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
1760
1761         lnet_ping_target_install_locked(pbuf);
1762
1763         if (the_lnet.ln_ping_target) {
1764                 old_pbuf = the_lnet.ln_ping_target;
1765                 old_ping_md = the_lnet.ln_ping_target_md;
1766         }
1767         the_lnet.ln_ping_target_md = ping_mdh;
1768         the_lnet.ln_ping_target = pbuf;
1769
1770         lnet_net_unlock(LNET_LOCK_EX);
1771
1772         if (old_pbuf) {
1773                 /* unlink and free the old ping info */
1774                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
1775                 lnet_ping_buffer_decref(old_pbuf);
1776         }
1777
1778         lnet_push_update_to_peers(0);
1779 }
1780
1781 static void
1782 lnet_ping_target_fini(void)
1783 {
1784         int             rc;
1785
1786         lnet_ping_md_unlink(the_lnet.ln_ping_target,
1787                             &the_lnet.ln_ping_target_md);
1788
1789         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1790         LASSERT(rc == 0);
1791
1792         lnet_ping_target_destroy();
1793 }
1794
1795 /* Resize the push target. */
1796 int lnet_push_target_resize(void)
1797 {
1798         struct lnet_process_id id = { LNET_NID_ANY, LNET_PID_ANY };
1799         struct lnet_md md = { NULL };
1800         struct lnet_handle_me meh;
1801         struct lnet_handle_md mdh;
1802         struct lnet_handle_md old_mdh;
1803         struct lnet_ping_buffer *pbuf;
1804         struct lnet_ping_buffer *old_pbuf;
1805         int nnis = the_lnet.ln_push_target_nnis;
1806         int rc;
1807
1808         if (nnis <= 0) {
1809                 rc = -EINVAL;
1810                 goto fail_return;
1811         }
1812 again:
1813         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1814         if (!pbuf) {
1815                 rc = -ENOMEM;
1816                 goto fail_return;
1817         }
1818
1819         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1820                           LNET_PROTO_PING_MATCHBITS, 0,
1821                           LNET_UNLINK, LNET_INS_AFTER,
1822                           &meh);
1823         if (rc) {
1824                 CERROR("Can't create push target ME: %d\n", rc);
1825                 goto fail_decref_pbuf;
1826         }
1827
1828         /* initialize md content */
1829         md.start     = &pbuf->pb_info;
1830         md.length    = LNET_PING_INFO_SIZE(nnis);
1831         md.threshold = LNET_MD_THRESH_INF;
1832         md.max_size  = 0;
1833         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE |
1834                        LNET_MD_MANAGE_REMOTE;
1835         md.user_ptr  = pbuf;
1836         md.eq_handle = the_lnet.ln_push_target_eq;
1837
1838         rc = LNetMDAttach(meh, md, LNET_RETAIN, &mdh);
1839         if (rc) {
1840                 CERROR("Can't attach push MD: %d\n", rc);
1841                 goto fail_unlink_meh;
1842         }
1843         lnet_ping_buffer_addref(pbuf);
1844
1845         lnet_net_lock(LNET_LOCK_EX);
1846         old_pbuf = the_lnet.ln_push_target;
1847         old_mdh = the_lnet.ln_push_target_md;
1848         the_lnet.ln_push_target = pbuf;
1849         the_lnet.ln_push_target_md = mdh;
1850         lnet_net_unlock(LNET_LOCK_EX);
1851
1852         if (old_pbuf) {
1853                 LNetMDUnlink(old_mdh);
1854                 lnet_ping_buffer_decref(old_pbuf);
1855         }
1856
1857         if (nnis < the_lnet.ln_push_target_nnis)
1858                 goto again;
1859
1860         CDEBUG(D_NET, "nnis %d success\n", nnis);
1861
1862         return 0;
1863
1864 fail_unlink_meh:
1865         LNetMEUnlink(meh);
1866 fail_decref_pbuf:
1867         lnet_ping_buffer_decref(pbuf);
1868 fail_return:
1869         CDEBUG(D_NET, "nnis %d error %d\n", nnis, rc);
1870         return rc;
1871 }
1872
1873 static void lnet_push_target_event_handler(struct lnet_event *ev)
1874 {
1875         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1876
1877         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
1878                 lnet_swap_pinginfo(pbuf);
1879
1880         lnet_peer_push_event(ev);
1881         if (ev->unlinked)
1882                 lnet_ping_buffer_decref(pbuf);
1883 }
1884
1885 /* Initialize the push target. */
1886 static int lnet_push_target_init(void)
1887 {
1888         int rc;
1889
1890         if (the_lnet.ln_push_target)
1891                 return -EALREADY;
1892
1893         rc = LNetEQAlloc(0, lnet_push_target_event_handler,
1894                          &the_lnet.ln_push_target_eq);
1895         if (rc) {
1896                 CERROR("Can't allocated push target EQ: %d\n", rc);
1897                 return rc;
1898         }
1899
1900         /* Start at the required minimum, we'll enlarge if required. */
1901         the_lnet.ln_push_target_nnis = LNET_INTERFACES_MIN;
1902
1903         rc = lnet_push_target_resize();
1904
1905         if (rc) {
1906                 LNetEQFree(the_lnet.ln_push_target_eq);
1907                 LNetInvalidateEQHandle(&the_lnet.ln_push_target_eq);
1908         }
1909
1910         return rc;
1911 }
1912
1913 /* Clean up the push target. */
1914 static void lnet_push_target_fini(void)
1915 {
1916         if (!the_lnet.ln_push_target)
1917                 return;
1918
1919         /* Unlink and invalidate to prevent new references. */
1920         LNetMDUnlink(the_lnet.ln_push_target_md);
1921         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
1922
1923         /* Wait for the unlink to complete. */
1924         while (lnet_ping_buffer_numref(the_lnet.ln_push_target) > 1) {
1925                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
1926                 set_current_state(TASK_UNINTERRUPTIBLE);
1927                 schedule_timeout(cfs_time_seconds(1));
1928         }
1929
1930         lnet_ping_buffer_decref(the_lnet.ln_push_target);
1931         the_lnet.ln_push_target = NULL;
1932         the_lnet.ln_push_target_nnis = 0;
1933
1934         LNetEQFree(the_lnet.ln_push_target_eq);
1935         LNetInvalidateEQHandle(&the_lnet.ln_push_target_eq);
1936 }
1937
1938 static int
1939 lnet_ni_tq_credits(struct lnet_ni *ni)
1940 {
1941         int     credits;
1942
1943         LASSERT(ni->ni_ncpts >= 1);
1944
1945         if (ni->ni_ncpts == 1)
1946                 return ni->ni_net->net_tunables.lct_max_tx_credits;
1947
1948         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
1949         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
1950         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
1951
1952         return credits;
1953 }
1954
1955 static void
1956 lnet_ni_unlink_locked(struct lnet_ni *ni)
1957 {
1958         /* move it to zombie list and nobody can find it anymore */
1959         LASSERT(!list_empty(&ni->ni_netlist));
1960         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
1961         lnet_ni_decref_locked(ni, 0);
1962 }
1963
1964 static void
1965 lnet_clear_zombies_nis_locked(struct lnet_net *net)
1966 {
1967         int             i;
1968         int             islo;
1969         struct lnet_ni  *ni;
1970         struct list_head *zombie_list = &net->net_ni_zombie;
1971
1972         /*
1973          * Now wait for the NIs I just nuked to show up on the zombie
1974          * list and shut them down in guaranteed thread context
1975          */
1976         i = 2;
1977         while (!list_empty(zombie_list)) {
1978                 int     *ref;
1979                 int     j;
1980
1981                 ni = list_entry(zombie_list->next,
1982                                 struct lnet_ni, ni_netlist);
1983                 list_del_init(&ni->ni_netlist);
1984                 /* the ni should be in deleting state. If it's not it's
1985                  * a bug */
1986                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
1987                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1988                         if (*ref == 0)
1989                                 continue;
1990                         /* still busy, add it back to zombie list */
1991                         list_add(&ni->ni_netlist, zombie_list);
1992                         break;
1993                 }
1994
1995                 if (!list_empty(&ni->ni_netlist)) {
1996                         lnet_net_unlock(LNET_LOCK_EX);
1997                         ++i;
1998                         if ((i & (-i)) == i) {
1999                                 CDEBUG(D_WARNING,
2000                                        "Waiting for zombie LNI %s\n",
2001                                        libcfs_nid2str(ni->ni_nid));
2002                         }
2003                         set_current_state(TASK_UNINTERRUPTIBLE);
2004                         schedule_timeout(cfs_time_seconds(1));
2005                         lnet_net_lock(LNET_LOCK_EX);
2006                         continue;
2007                 }
2008
2009                 lnet_net_unlock(LNET_LOCK_EX);
2010
2011                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
2012
2013                 LASSERT(!in_interrupt());
2014                 (net->net_lnd->lnd_shutdown)(ni);
2015
2016                 if (!islo)
2017                         CDEBUG(D_LNI, "Removed LNI %s\n",
2018                               libcfs_nid2str(ni->ni_nid));
2019
2020                 lnet_ni_free(ni);
2021                 i = 2;
2022                 lnet_net_lock(LNET_LOCK_EX);
2023         }
2024 }
2025
2026 /* shutdown down the NI and release refcount */
2027 static void
2028 lnet_shutdown_lndni(struct lnet_ni *ni)
2029 {
2030         int i;
2031         struct lnet_net *net = ni->ni_net;
2032
2033         lnet_net_lock(LNET_LOCK_EX);
2034         lnet_ni_lock(ni);
2035         ni->ni_state = LNET_NI_STATE_DELETING;
2036         lnet_ni_unlock(ni);
2037         lnet_ni_unlink_locked(ni);
2038         lnet_incr_dlc_seq();
2039         lnet_net_unlock(LNET_LOCK_EX);
2040
2041         /* clear messages for this NI on the lazy portal */
2042         for (i = 0; i < the_lnet.ln_nportals; i++)
2043                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
2044
2045         lnet_net_lock(LNET_LOCK_EX);
2046         lnet_clear_zombies_nis_locked(net);
2047         lnet_net_unlock(LNET_LOCK_EX);
2048 }
2049
2050 static void
2051 lnet_shutdown_lndnet(struct lnet_net *net)
2052 {
2053         struct lnet_ni *ni;
2054
2055         lnet_net_lock(LNET_LOCK_EX);
2056
2057         net->net_state = LNET_NET_STATE_DELETING;
2058
2059         list_del_init(&net->net_list);
2060
2061         while (!list_empty(&net->net_ni_list)) {
2062                 ni = list_entry(net->net_ni_list.next,
2063                                 struct lnet_ni, ni_netlist);
2064                 lnet_net_unlock(LNET_LOCK_EX);
2065                 lnet_shutdown_lndni(ni);
2066                 lnet_net_lock(LNET_LOCK_EX);
2067         }
2068
2069         lnet_net_unlock(LNET_LOCK_EX);
2070
2071         /* Do peer table cleanup for this net */
2072         lnet_peer_tables_cleanup(net);
2073
2074         lnet_net_lock(LNET_LOCK_EX);
2075         /*
2076          * decrement ref count on lnd only when the entire network goes
2077          * away
2078          */
2079         net->net_lnd->lnd_refcount--;
2080
2081         lnet_net_unlock(LNET_LOCK_EX);
2082
2083         lnet_net_free(net);
2084 }
2085
2086 static void
2087 lnet_shutdown_lndnets(void)
2088 {
2089         struct lnet_net *net;
2090         struct list_head resend;
2091         struct lnet_msg *msg, *tmp;
2092
2093         INIT_LIST_HEAD(&resend);
2094
2095         /* NB called holding the global mutex */
2096
2097         /* All quiet on the API front */
2098         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
2099         LASSERT(the_lnet.ln_refcount == 0);
2100
2101         lnet_net_lock(LNET_LOCK_EX);
2102         the_lnet.ln_state = LNET_STATE_STOPPING;
2103
2104         while (!list_empty(&the_lnet.ln_nets)) {
2105                 /*
2106                  * move the nets to the zombie list to avoid them being
2107                  * picked up for new work. LONET is also included in the
2108                  * Nets that will be moved to the zombie list
2109                  */
2110                 net = list_entry(the_lnet.ln_nets.next,
2111                                  struct lnet_net, net_list);
2112                 list_move(&net->net_list, &the_lnet.ln_net_zombie);
2113         }
2114
2115         /* Drop the cached loopback Net. */
2116         if (the_lnet.ln_loni != NULL) {
2117                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
2118                 the_lnet.ln_loni = NULL;
2119         }
2120         lnet_net_unlock(LNET_LOCK_EX);
2121
2122         /* iterate through the net zombie list and delete each net */
2123         while (!list_empty(&the_lnet.ln_net_zombie)) {
2124                 net = list_entry(the_lnet.ln_net_zombie.next,
2125                                  struct lnet_net, net_list);
2126                 lnet_shutdown_lndnet(net);
2127         }
2128
2129         spin_lock(&the_lnet.ln_msg_resend_lock);
2130         list_splice(&the_lnet.ln_msg_resend, &resend);
2131         spin_unlock(&the_lnet.ln_msg_resend_lock);
2132
2133         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
2134                 list_del_init(&msg->msg_list);
2135                 msg->msg_no_resend = true;
2136                 lnet_finalize(msg, -ECANCELED);
2137         }
2138
2139         lnet_net_lock(LNET_LOCK_EX);
2140         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
2141         lnet_net_unlock(LNET_LOCK_EX);
2142 }
2143
2144 static int
2145 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2146 {
2147         int                     rc = -EINVAL;
2148         struct lnet_tx_queue    *tq;
2149         int                     i;
2150         struct lnet_net         *net = ni->ni_net;
2151
2152         mutex_lock(&the_lnet.ln_lnd_mutex);
2153
2154         if (tun) {
2155                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2156                 ni->ni_lnd_tunables_set = true;
2157         }
2158
2159         rc = (net->net_lnd->lnd_startup)(ni);
2160
2161         mutex_unlock(&the_lnet.ln_lnd_mutex);
2162
2163         if (rc != 0) {
2164                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2165                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2166                 lnet_net_lock(LNET_LOCK_EX);
2167                 net->net_lnd->lnd_refcount--;
2168                 lnet_net_unlock(LNET_LOCK_EX);
2169                 goto failed0;
2170         }
2171
2172         lnet_ni_lock(ni);
2173         ni->ni_state = LNET_NI_STATE_ACTIVE;
2174         lnet_ni_unlock(ni);
2175
2176         /* We keep a reference on the loopback net through the loopback NI */
2177         if (net->net_lnd->lnd_type == LOLND) {
2178                 lnet_ni_addref(ni);
2179                 LASSERT(the_lnet.ln_loni == NULL);
2180                 the_lnet.ln_loni = ni;
2181                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2182                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2183                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2184                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2185                 return 0;
2186         }
2187
2188         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2189             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2190                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2191                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2192                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2193                                         "" : "per-peer ");
2194                 /* shutdown the NI since if we get here then it must've already
2195                  * been started
2196                  */
2197                 lnet_shutdown_lndni(ni);
2198                 return -EINVAL;
2199         }
2200
2201         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2202                 tq->tq_credits_min =
2203                 tq->tq_credits_max =
2204                 tq->tq_credits = lnet_ni_tq_credits(ni);
2205         }
2206
2207         atomic_set(&ni->ni_tx_credits,
2208                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2209         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2210
2211         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2212                 libcfs_nid2str(ni->ni_nid),
2213                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2214                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2215                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2216                 ni->ni_net->net_tunables.lct_peer_timeout);
2217
2218         return 0;
2219 failed0:
2220         lnet_ni_free(ni);
2221         return rc;
2222 }
2223
2224 static int
2225 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2226 {
2227         struct lnet_ni *ni;
2228         struct lnet_net *net_l = NULL;
2229         struct list_head        local_ni_list;
2230         int                     rc;
2231         int                     ni_count = 0;
2232         __u32                   lnd_type;
2233         struct lnet_lnd *lnd;
2234         int                     peer_timeout =
2235                 net->net_tunables.lct_peer_timeout;
2236         int                     maxtxcredits =
2237                 net->net_tunables.lct_max_tx_credits;
2238         int                     peerrtrcredits =
2239                 net->net_tunables.lct_peer_rtr_credits;
2240
2241         INIT_LIST_HEAD(&local_ni_list);
2242
2243         /*
2244          * make sure that this net is unique. If it isn't then
2245          * we are adding interfaces to an already existing network, and
2246          * 'net' is just a convenient way to pass in the list.
2247          * if it is unique we need to find the LND and load it if
2248          * necessary.
2249          */
2250         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2251                 lnd_type = LNET_NETTYP(net->net_id);
2252
2253                 mutex_lock(&the_lnet.ln_lnd_mutex);
2254                 lnd = lnet_find_lnd_by_type(lnd_type);
2255
2256                 if (lnd == NULL) {
2257                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2258                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2259                         mutex_lock(&the_lnet.ln_lnd_mutex);
2260
2261                         lnd = lnet_find_lnd_by_type(lnd_type);
2262                         if (lnd == NULL) {
2263                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2264                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
2265                                 libcfs_lnd2str(lnd_type),
2266                                 libcfs_lnd2modname(lnd_type), rc);
2267 #ifndef HAVE_MODULE_LOADING_SUPPORT
2268                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
2269                                                 "compiled with kernel module "
2270                                                 "loading support.");
2271 #endif
2272                                 rc = -EINVAL;
2273                                 goto failed0;
2274                         }
2275                 }
2276
2277                 lnet_net_lock(LNET_LOCK_EX);
2278                 lnd->lnd_refcount++;
2279                 lnet_net_unlock(LNET_LOCK_EX);
2280
2281                 net->net_lnd = lnd;
2282
2283                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2284
2285                 net_l = net;
2286         }
2287
2288         /*
2289          * net_l: if the network being added is unique then net_l
2290          *        will point to that network
2291          *        if the network being added is not unique then
2292          *        net_l points to the existing network.
2293          *
2294          * When we enter the loop below, we'll pick NIs off he
2295          * network beign added and start them up, then add them to
2296          * a local ni list. Once we've successfully started all
2297          * the NIs then we join the local NI list (of started up
2298          * networks) with the net_l->net_ni_list, which should
2299          * point to the correct network to add the new ni list to
2300          *
2301          * If any of the new NIs fail to start up, then we want to
2302          * iterate through the local ni list, which should include
2303          * any NIs which were successfully started up, and shut
2304          * them down.
2305          *
2306          * After than we want to delete the network being added,
2307          * to avoid a memory leak.
2308          */
2309
2310         /*
2311          * When a network uses TCP bonding then all its interfaces
2312          * must be specified when the network is first defined: the
2313          * TCP bonding code doesn't allow for interfaces to be added
2314          * or removed.
2315          */
2316         if (net_l != net && net_l != NULL && use_tcp_bonding &&
2317             LNET_NETTYP(net_l->net_id) == SOCKLND) {
2318                 rc = -EINVAL;
2319                 goto failed0;
2320         }
2321
2322         while (!list_empty(&net->net_ni_added)) {
2323                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
2324                                 ni_netlist);
2325                 list_del_init(&ni->ni_netlist);
2326
2327                 /* make sure that the the NI we're about to start
2328                  * up is actually unique. if it's not fail. */
2329                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2330                                         ni->ni_interfaces[0])) {
2331                         rc = -EINVAL;
2332                         goto failed1;
2333                 }
2334
2335                 /* adjust the pointer the parent network, just in case it
2336                  * the net is a duplicate */
2337                 ni->ni_net = net_l;
2338
2339                 rc = lnet_startup_lndni(ni, tun);
2340
2341                 LASSERT(ni->ni_net->net_tunables.lct_peer_timeout <= 0 ||
2342                         ni->ni_net->net_lnd->lnd_query != NULL);
2343
2344                 if (rc < 0)
2345                         goto failed1;
2346
2347                 lnet_ni_addref(ni);
2348                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2349
2350                 ni_count++;
2351         }
2352
2353         lnet_net_lock(LNET_LOCK_EX);
2354         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2355         lnet_incr_dlc_seq();
2356         lnet_net_unlock(LNET_LOCK_EX);
2357
2358         /* if the network is not unique then we don't want to keep
2359          * it around after we're done. Free it. Otherwise add that
2360          * net to the global the_lnet.ln_nets */
2361         if (net_l != net && net_l != NULL) {
2362                 /*
2363                  * TODO - note. currently the tunables can not be updated
2364                  * once added
2365                  */
2366                 lnet_net_free(net);
2367         } else {
2368                 net->net_state = LNET_NET_STATE_ACTIVE;
2369                 /*
2370                  * restore tunables after it has been overwitten by the
2371                  * lnd
2372                  */
2373                 if (peer_timeout != -1)
2374                         net->net_tunables.lct_peer_timeout = peer_timeout;
2375                 if (maxtxcredits != -1)
2376                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2377                 if (peerrtrcredits != -1)
2378                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2379
2380                 lnet_net_lock(LNET_LOCK_EX);
2381                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2382                 lnet_net_unlock(LNET_LOCK_EX);
2383         }
2384
2385         return ni_count;
2386
2387 failed1:
2388         /*
2389          * shutdown the new NIs that are being started up
2390          * free the NET being started
2391          */
2392         while (!list_empty(&local_ni_list)) {
2393                 ni = list_entry(local_ni_list.next, struct lnet_ni,
2394                                 ni_netlist);
2395
2396                 lnet_shutdown_lndni(ni);
2397         }
2398
2399 failed0:
2400         lnet_net_free(net);
2401
2402         return rc;
2403 }
2404
2405 static int
2406 lnet_startup_lndnets(struct list_head *netlist)
2407 {
2408         struct lnet_net         *net;
2409         int                     rc;
2410         int                     ni_count = 0;
2411
2412         /*
2413          * Change to running state before bringing up the LNDs. This
2414          * allows lnet_shutdown_lndnets() to assert that we've passed
2415          * through here.
2416          */
2417         lnet_net_lock(LNET_LOCK_EX);
2418         the_lnet.ln_state = LNET_STATE_RUNNING;
2419         lnet_net_unlock(LNET_LOCK_EX);
2420
2421         while (!list_empty(netlist)) {
2422                 net = list_entry(netlist->next, struct lnet_net, net_list);
2423                 list_del_init(&net->net_list);
2424
2425                 rc = lnet_startup_lndnet(net, NULL);
2426
2427                 if (rc < 0)
2428                         goto failed;
2429
2430                 ni_count += rc;
2431         }
2432
2433         return ni_count;
2434 failed:
2435         lnet_shutdown_lndnets();
2436
2437         return rc;
2438 }
2439
2440 /**
2441  * Initialize LNet library.
2442  *
2443  * Automatically called at module loading time. Caller has to call
2444  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2445  * latter returned 0. It must be called exactly once.
2446  *
2447  * \retval 0 on success
2448  * \retval -ve on failures.
2449  */
2450 int lnet_lib_init(void)
2451 {
2452         int rc;
2453
2454         lnet_assert_wire_constants();
2455
2456         /* refer to global cfs_cpt_table for now */
2457         the_lnet.ln_cpt_table   = cfs_cpt_table;
2458         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
2459
2460         LASSERT(the_lnet.ln_cpt_number > 0);
2461         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2462                 /* we are under risk of consuming all lh_cookie */
2463                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2464                        "please change setting of CPT-table and retry\n",
2465                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2466                 return -E2BIG;
2467         }
2468
2469         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2470                 the_lnet.ln_cpt_bits++;
2471
2472         rc = lnet_create_locks();
2473         if (rc != 0) {
2474                 CERROR("Can't create LNet global locks: %d\n", rc);
2475                 return rc;
2476         }
2477
2478         the_lnet.ln_refcount = 0;
2479         INIT_LIST_HEAD(&the_lnet.ln_lnds);
2480         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2481         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2482
2483         /* The hash table size is the number of bits it takes to express the set
2484          * ln_num_routes, minus 1 (better to under estimate than over so we
2485          * don't waste memory). */
2486         if (rnet_htable_size <= 0)
2487                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2488         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2489                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2490         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2491                                            order_base_2(rnet_htable_size) - 1);
2492
2493         /* All LNDs apart from the LOLND are in separate modules.  They
2494          * register themselves when their module loads, and unregister
2495          * themselves when their module is unloaded. */
2496         lnet_register_lnd(&the_lolnd);
2497         return 0;
2498 }
2499
2500 /**
2501  * Finalize LNet library.
2502  *
2503  * \pre lnet_lib_init() called with success.
2504  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2505  */
2506 void lnet_lib_exit(void)
2507 {
2508         LASSERT(the_lnet.ln_refcount == 0);
2509
2510         while (!list_empty(&the_lnet.ln_lnds))
2511                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
2512                                                struct lnet_lnd, lnd_list));
2513         lnet_destroy_locks();
2514 }
2515
2516 /**
2517  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2518  *
2519  * Users must call this function at least once before any other functions.
2520  * For each successful call there must be a corresponding call to
2521  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2522  * ignored.
2523  *
2524  * The PID used by LNet may be different from the one requested.
2525  * See LNetGetId().
2526  *
2527  * \param requested_pid PID requested by the caller.
2528  *
2529  * \return >= 0 on success, and < 0 error code on failures.
2530  */
2531 int
2532 LNetNIInit(lnet_pid_t requested_pid)
2533 {
2534         int                     im_a_router = 0;
2535         int                     rc;
2536         int                     ni_count;
2537         struct lnet_ping_buffer *pbuf;
2538         struct lnet_handle_md   ping_mdh;
2539         struct list_head        net_head;
2540         struct lnet_net         *net;
2541
2542         INIT_LIST_HEAD(&net_head);
2543
2544         mutex_lock(&the_lnet.ln_api_mutex);
2545
2546         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
2547
2548         if (the_lnet.ln_refcount > 0) {
2549                 rc = the_lnet.ln_refcount++;
2550                 mutex_unlock(&the_lnet.ln_api_mutex);
2551                 return rc;
2552         }
2553
2554         rc = lnet_prepare(requested_pid);
2555         if (rc != 0) {
2556                 mutex_unlock(&the_lnet.ln_api_mutex);
2557                 return rc;
2558         }
2559
2560         /* create a network for Loopback network */
2561         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
2562         if (net == NULL) {
2563                 rc = -ENOMEM;
2564                 goto err_empty_list;
2565         }
2566
2567         /* Add in the loopback NI */
2568         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
2569                 rc = -ENOMEM;
2570                 goto err_empty_list;
2571         }
2572
2573         /* If LNet is being initialized via DLC it is possible
2574          * that the user requests not to load module parameters (ones which
2575          * are supported by DLC) on initialization.  Therefore, make sure not
2576          * to load networks, routes and forwarding from module parameters
2577          * in this case.  On cleanup in case of failure only clean up
2578          * routes if it has been loaded */
2579         if (!the_lnet.ln_nis_from_mod_params) {
2580                 rc = lnet_parse_networks(&net_head, lnet_get_networks(),
2581                                          use_tcp_bonding);
2582                 if (rc < 0)
2583                         goto err_empty_list;
2584         }
2585
2586         ni_count = lnet_startup_lndnets(&net_head);
2587         if (ni_count < 0) {
2588                 rc = ni_count;
2589                 goto err_empty_list;
2590         }
2591
2592         if (!the_lnet.ln_nis_from_mod_params) {
2593                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
2594                 if (rc != 0)
2595                         goto err_shutdown_lndnis;
2596
2597                 rc = lnet_rtrpools_alloc(im_a_router);
2598                 if (rc != 0)
2599                         goto err_destroy_routes;
2600         }
2601
2602         rc = lnet_acceptor_start();
2603         if (rc != 0)
2604                 goto err_destroy_routes;
2605
2606         the_lnet.ln_refcount = 1;
2607         /* Now I may use my own API functions... */
2608
2609         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
2610         if (rc != 0)
2611                 goto err_acceptor_stop;
2612
2613         lnet_ping_target_update(pbuf, ping_mdh);
2614
2615         rc = LNetEQAlloc(0, lnet_mt_event_handler, &the_lnet.ln_mt_eqh);
2616         if (rc != 0) {
2617                 CERROR("Can't allocate monitor thread EQ: %d\n", rc);
2618                 goto err_stop_ping;
2619         }
2620
2621         rc = lnet_monitor_thr_start();
2622         if (rc != 0)
2623                 goto err_stop_ping;
2624
2625         rc = lnet_push_target_init();
2626         if (rc != 0)
2627                 goto err_stop_monitor_thr;
2628
2629         rc = lnet_peer_discovery_start();
2630         if (rc != 0)
2631                 goto err_destroy_push_target;
2632
2633         lnet_fault_init();
2634         lnet_router_debugfs_init();
2635
2636         mutex_unlock(&the_lnet.ln_api_mutex);
2637
2638         return 0;
2639
2640 err_destroy_push_target:
2641         lnet_push_target_fini();
2642 err_stop_monitor_thr:
2643         lnet_monitor_thr_stop();
2644 err_stop_ping:
2645         lnet_ping_target_fini();
2646 err_acceptor_stop:
2647         the_lnet.ln_refcount = 0;
2648         lnet_acceptor_stop();
2649 err_destroy_routes:
2650         if (!the_lnet.ln_nis_from_mod_params)
2651                 lnet_destroy_routes();
2652 err_shutdown_lndnis:
2653         lnet_shutdown_lndnets();
2654 err_empty_list:
2655         lnet_unprepare();
2656         LASSERT(rc < 0);
2657         mutex_unlock(&the_lnet.ln_api_mutex);
2658         while (!list_empty(&net_head)) {
2659                 struct lnet_net *net;
2660
2661                 net = list_entry(net_head.next, struct lnet_net, net_list);
2662                 list_del_init(&net->net_list);
2663                 lnet_net_free(net);
2664         }
2665         return rc;
2666 }
2667 EXPORT_SYMBOL(LNetNIInit);
2668
2669 /**
2670  * Stop LNet interfaces, routing, and forwarding.
2671  *
2672  * Users must call this function once for each successful call to LNetNIInit().
2673  * Once the LNetNIFini() operation has been started, the results of pending
2674  * API operations are undefined.
2675  *
2676  * \return always 0 for current implementation.
2677  */
2678 int
2679 LNetNIFini()
2680 {
2681         mutex_lock(&the_lnet.ln_api_mutex);
2682
2683         LASSERT(the_lnet.ln_refcount > 0);
2684
2685         if (the_lnet.ln_refcount != 1) {
2686                 the_lnet.ln_refcount--;
2687         } else {
2688                 LASSERT(!the_lnet.ln_niinit_self);
2689
2690                 lnet_fault_fini();
2691
2692                 lnet_router_debugfs_fini();
2693                 lnet_peer_discovery_stop();
2694                 lnet_push_target_fini();
2695                 lnet_monitor_thr_stop();
2696                 lnet_ping_target_fini();
2697
2698                 /* Teardown fns that use my own API functions BEFORE here */
2699                 the_lnet.ln_refcount = 0;
2700
2701                 lnet_acceptor_stop();
2702                 lnet_destroy_routes();
2703                 lnet_shutdown_lndnets();
2704                 lnet_unprepare();
2705         }
2706
2707         mutex_unlock(&the_lnet.ln_api_mutex);
2708         return 0;
2709 }
2710 EXPORT_SYMBOL(LNetNIFini);
2711
2712 /**
2713  * Grabs the ni data from the ni structure and fills the out
2714  * parameters
2715  *
2716  * \param[in] ni network        interface structure
2717  * \param[out] cfg_ni           NI config information
2718  * \param[out] tun              network and LND tunables
2719  */
2720 static void
2721 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
2722                    struct lnet_ioctl_config_lnd_tunables *tun,
2723                    struct lnet_ioctl_element_stats *stats,
2724                    __u32 tun_size)
2725 {
2726         size_t min_size = 0;
2727         int i;
2728
2729         if (!ni || !cfg_ni || !tun)
2730                 return;
2731
2732         if (ni->ni_interfaces[0] != NULL) {
2733                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2734                         if (ni->ni_interfaces[i] != NULL) {
2735                                 strncpy(cfg_ni->lic_ni_intf[i],
2736                                         ni->ni_interfaces[i],
2737                                         sizeof(cfg_ni->lic_ni_intf[i]));
2738                         }
2739                 }
2740         }
2741
2742         cfg_ni->lic_nid = ni->ni_nid;
2743         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2744                 cfg_ni->lic_status = LNET_NI_STATUS_UP;
2745         else
2746                 cfg_ni->lic_status = ni->ni_status->ns_status;
2747         cfg_ni->lic_tcp_bonding = use_tcp_bonding;
2748         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
2749
2750         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
2751
2752         if (stats) {
2753                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
2754                                                        LNET_STATS_TYPE_SEND);
2755                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
2756                                                        LNET_STATS_TYPE_RECV);
2757                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
2758                                                        LNET_STATS_TYPE_DROP);
2759         }
2760
2761         /*
2762          * tun->lt_tun will always be present, but in order to be
2763          * backwards compatible, we need to deal with the cases when
2764          * tun->lt_tun is smaller than what the kernel has, because it
2765          * comes from an older version of a userspace program, then we'll
2766          * need to copy as much information as we have available space.
2767          */
2768         min_size = tun_size - sizeof(tun->lt_cmn);
2769         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
2770
2771         /* copy over the cpts */
2772         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
2773             ni->ni_cpts == NULL)  {
2774                 for (i = 0; i < ni->ni_ncpts; i++)
2775                         cfg_ni->lic_cpts[i] = i;
2776         } else {
2777                 for (i = 0;
2778                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
2779                      i < LNET_MAX_SHOW_NUM_CPT;
2780                      i++)
2781                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
2782         }
2783         cfg_ni->lic_ncpts = ni->ni_ncpts;
2784 }
2785
2786 /**
2787  * NOTE: This is a legacy function left in the code to be backwards
2788  * compatible with older userspace programs. It should eventually be
2789  * removed.
2790  *
2791  * Grabs the ni data from the ni structure and fills the out
2792  * parameters
2793  *
2794  * \param[in] ni network        interface structure
2795  * \param[out] config           config information
2796  */
2797 static void
2798 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
2799                          struct lnet_ioctl_config_data *config)
2800 {
2801         struct lnet_ioctl_net_config *net_config;
2802         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
2803         size_t min_size, tunable_size = 0;
2804         int i;
2805
2806         if (!ni || !config)
2807                 return;
2808
2809         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
2810         if (!net_config)
2811                 return;
2812
2813         BUILD_BUG_ON(ARRAY_SIZE(ni->ni_interfaces) !=
2814                      ARRAY_SIZE(net_config->ni_interfaces));
2815
2816         for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2817                 if (!ni->ni_interfaces[i])
2818                         break;
2819
2820                 strncpy(net_config->ni_interfaces[i],
2821                         ni->ni_interfaces[i],
2822                         sizeof(net_config->ni_interfaces[i]));
2823         }
2824
2825         config->cfg_nid = ni->ni_nid;
2826         config->cfg_config_u.cfg_net.net_peer_timeout =
2827                 ni->ni_net->net_tunables.lct_peer_timeout;
2828         config->cfg_config_u.cfg_net.net_max_tx_credits =
2829                 ni->ni_net->net_tunables.lct_max_tx_credits;
2830         config->cfg_config_u.cfg_net.net_peer_tx_credits =
2831                 ni->ni_net->net_tunables.lct_peer_tx_credits;
2832         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
2833                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
2834
2835         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2836                 net_config->ni_status = LNET_NI_STATUS_UP;
2837         else
2838                 net_config->ni_status = ni->ni_status->ns_status;
2839
2840         if (ni->ni_cpts) {
2841                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
2842
2843                 for (i = 0; i < num_cpts; i++)
2844                         net_config->ni_cpts[i] = ni->ni_cpts[i];
2845
2846                 config->cfg_ncpts = num_cpts;
2847         }
2848
2849         /*
2850          * See if user land tools sent in a newer and larger version
2851          * of struct lnet_tunables than what the kernel uses.
2852          */
2853         min_size = sizeof(*config) + sizeof(*net_config);
2854
2855         if (config->cfg_hdr.ioc_len > min_size)
2856                 tunable_size = config->cfg_hdr.ioc_len - min_size;
2857
2858         /* Don't copy too much data to user space */
2859         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
2860         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
2861
2862         if (lnd_cfg && min_size) {
2863                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
2864                 config->cfg_config_u.cfg_net.net_interface_count = 1;
2865
2866                 /* Tell user land that kernel side has less data */
2867                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
2868                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
2869                         config->cfg_hdr.ioc_len -= min_size;
2870                 }
2871         }
2872 }
2873
2874 struct lnet_ni *
2875 lnet_get_ni_idx_locked(int idx)
2876 {
2877         struct lnet_ni          *ni;
2878         struct lnet_net         *net;
2879
2880         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2881                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2882                         if (idx-- == 0)
2883                                 return ni;
2884                 }
2885         }
2886
2887         return NULL;
2888 }
2889
2890 struct lnet_ni *
2891 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
2892 {
2893         struct lnet_ni          *ni;
2894         struct lnet_net         *net = mynet;
2895
2896         /*
2897          * It is possible that the net has been cleaned out while there is
2898          * a message being sent. This function accessed the net without
2899          * checking if the list is empty
2900          */
2901         if (prev == NULL) {
2902                 if (net == NULL)
2903                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
2904                                         net_list);
2905                 if (list_empty(&net->net_ni_list))
2906                         return NULL;
2907                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2908                                 ni_netlist);
2909
2910                 return ni;
2911         }
2912
2913         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
2914                 /* if you reached the end of the ni list and the net is
2915                  * specified, then there are no more nis in that net */
2916                 if (net != NULL)
2917                         return NULL;
2918
2919                 /* we reached the end of this net ni list. move to the
2920                  * next net */
2921                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
2922                         /* no more nets and no more NIs. */
2923                         return NULL;
2924
2925                 /* get the next net */
2926                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
2927                                  net_list);
2928                 if (list_empty(&net->net_ni_list))
2929                         return NULL;
2930                 /* get the ni on it */
2931                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2932                                 ni_netlist);
2933
2934                 return ni;
2935         }
2936
2937         if (list_empty(&prev->ni_netlist))
2938                 return NULL;
2939
2940         /* there are more nis left */
2941         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
2942
2943         return ni;
2944 }
2945
2946 int
2947 lnet_get_net_config(struct lnet_ioctl_config_data *config)
2948 {
2949         struct lnet_ni *ni;
2950         int cpt;
2951         int rc = -ENOENT;
2952         int idx = config->cfg_count;
2953
2954         cpt = lnet_net_lock_current();
2955
2956         ni = lnet_get_ni_idx_locked(idx);
2957
2958         if (ni != NULL) {
2959                 rc = 0;
2960                 lnet_ni_lock(ni);
2961                 lnet_fill_ni_info_legacy(ni, config);
2962                 lnet_ni_unlock(ni);
2963         }
2964
2965         lnet_net_unlock(cpt);
2966         return rc;
2967 }
2968
2969 int
2970 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
2971                    struct lnet_ioctl_config_lnd_tunables *tun,
2972                    struct lnet_ioctl_element_stats *stats,
2973                    __u32 tun_size)
2974 {
2975         struct lnet_ni          *ni;
2976         int                     cpt;
2977         int                     rc = -ENOENT;
2978
2979         if (!cfg_ni || !tun || !stats)
2980                 return -EINVAL;
2981
2982         cpt = lnet_net_lock_current();
2983
2984         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
2985
2986         if (ni) {
2987                 rc = 0;
2988                 lnet_ni_lock(ni);
2989                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
2990                 lnet_ni_unlock(ni);
2991         }
2992
2993         lnet_net_unlock(cpt);
2994         return rc;
2995 }
2996
2997 int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
2998 {
2999         struct lnet_ni *ni;
3000         int cpt;
3001         int rc = -ENOENT;
3002
3003         if (!msg_stats)
3004                 return -EINVAL;
3005
3006         cpt = lnet_net_lock_current();
3007
3008         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
3009
3010         if (ni) {
3011                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
3012                 rc = 0;
3013         }
3014
3015         lnet_net_unlock(cpt);
3016
3017         return rc;
3018 }
3019
3020 static int lnet_add_net_common(struct lnet_net *net,
3021                                struct lnet_ioctl_config_lnd_tunables *tun)
3022 {
3023         __u32                   net_id;
3024         struct lnet_ping_buffer *pbuf;
3025         struct lnet_handle_md   ping_mdh;
3026         int                     rc;
3027         struct lnet_remotenet *rnet;
3028         int                     net_ni_count;
3029         int                     num_acceptor_nets;
3030
3031         lnet_net_lock(LNET_LOCK_EX);
3032         rnet = lnet_find_rnet_locked(net->net_id);
3033         lnet_net_unlock(LNET_LOCK_EX);
3034         /*
3035          * make sure that the net added doesn't invalidate the current
3036          * configuration LNet is keeping
3037          */
3038         if (rnet) {
3039                 CERROR("Adding net %s will invalidate routing configuration\n",
3040                        libcfs_net2str(net->net_id));
3041                 lnet_net_free(net);
3042                 return -EUSERS;
3043         }
3044
3045         /*
3046          * make sure you calculate the correct number of slots in the ping
3047          * buffer. Since the ping info is a flattened list of all the NIs,
3048          * we should allocate enough slots to accomodate the number of NIs
3049          * which will be added.
3050          *
3051          * since ni hasn't been configured yet, use
3052          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
3053          */
3054         net_ni_count = lnet_get_net_ni_count_pre(net);
3055
3056         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3057                                     net_ni_count + lnet_get_ni_count(),
3058                                     false);
3059         if (rc < 0) {
3060                 lnet_net_free(net);
3061                 return rc;
3062         }
3063
3064         if (tun)
3065                 memcpy(&net->net_tunables,
3066                        &tun->lt_cmn, sizeof(net->net_tunables));
3067         else
3068                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
3069
3070         /*
3071          * before starting this network get a count of the current TCP
3072          * networks which require the acceptor thread running. If that
3073          * count is == 0 before we start up this network, then we'd want to
3074          * start up the acceptor thread after starting up this network
3075          */
3076         num_acceptor_nets = lnet_count_acceptor_nets();
3077
3078         net_id = net->net_id;
3079
3080         rc = lnet_startup_lndnet(net,
3081                                  (tun) ? &tun->lt_tun : NULL);
3082         if (rc < 0)
3083                 goto failed;
3084
3085         lnet_net_lock(LNET_LOCK_EX);
3086         net = lnet_get_net_locked(net_id);
3087         lnet_net_unlock(LNET_LOCK_EX);
3088
3089         LASSERT(net);
3090
3091         /*
3092          * Start the acceptor thread if this is the first network
3093          * being added that requires the thread.
3094          */
3095         if (net->net_lnd->lnd_accept && num_acceptor_nets == 0) {
3096                 rc = lnet_acceptor_start();
3097                 if (rc < 0) {
3098                         /* shutdown the net that we just started */
3099                         CERROR("Failed to start up acceptor thread\n");
3100                         lnet_shutdown_lndnet(net);
3101                         goto failed;
3102                 }
3103         }
3104
3105         lnet_net_lock(LNET_LOCK_EX);
3106         lnet_peer_net_added(net);
3107         lnet_net_unlock(LNET_LOCK_EX);
3108
3109         lnet_ping_target_update(pbuf, ping_mdh);
3110
3111         return 0;
3112
3113 failed:
3114         lnet_ping_md_unlink(pbuf, &ping_mdh);
3115         lnet_ping_buffer_decref(pbuf);
3116         return rc;
3117 }
3118
3119 static int lnet_handle_legacy_ip2nets(char *ip2nets,
3120                                       struct lnet_ioctl_config_lnd_tunables *tun)
3121 {
3122         struct lnet_net *net;
3123         char *nets;
3124         int rc;
3125         struct list_head net_head;
3126
3127         INIT_LIST_HEAD(&net_head);
3128
3129         rc = lnet_parse_ip2nets(&nets, ip2nets);
3130         if (rc < 0)
3131                 return rc;
3132
3133         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3134         if (rc < 0)
3135                 return rc;
3136
3137         mutex_lock(&the_lnet.ln_api_mutex);
3138         while (!list_empty(&net_head)) {
3139                 net = list_entry(net_head.next, struct lnet_net, net_list);
3140                 list_del_init(&net->net_list);
3141                 rc = lnet_add_net_common(net, tun);
3142                 if (rc < 0)
3143                         goto out;
3144         }
3145
3146 out:
3147         mutex_unlock(&the_lnet.ln_api_mutex);
3148
3149         while (!list_empty(&net_head)) {
3150                 net = list_entry(net_head.next, struct lnet_net, net_list);
3151                 list_del_init(&net->net_list);
3152                 lnet_net_free(net);
3153         }
3154         return rc;
3155 }
3156
3157 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
3158 {
3159         struct lnet_net *net;
3160         struct lnet_ni *ni;
3161         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3162         int rc, i;
3163         __u32 net_id, lnd_type;
3164
3165         /* get the tunables if they are available */
3166         if (conf->lic_cfg_hdr.ioc_len >=
3167             sizeof(*conf) + sizeof(*tun))
3168                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3169                         conf->lic_bulk;
3170
3171         /* handle legacy ip2nets from DLC */
3172         if (conf->lic_legacy_ip2nets[0] != '\0')
3173                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
3174                                                   tun);
3175
3176         net_id = LNET_NIDNET(conf->lic_nid);
3177         lnd_type = LNET_NETTYP(net_id);
3178
3179         if (!libcfs_isknown_lnd(lnd_type)) {
3180                 CERROR("No valid net and lnd information provided\n");
3181                 return -EINVAL;
3182         }
3183
3184         net = lnet_net_alloc(net_id, NULL);
3185         if (!net)
3186                 return -ENOMEM;
3187
3188         for (i = 0; i < conf->lic_ncpts; i++) {
3189                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
3190                         return -EINVAL;
3191         }
3192
3193         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
3194                                        conf->lic_ni_intf[0]);
3195         if (!ni)
3196                 return -ENOMEM;
3197
3198         mutex_lock(&the_lnet.ln_api_mutex);
3199
3200         rc = lnet_add_net_common(net, tun);
3201
3202         mutex_unlock(&the_lnet.ln_api_mutex);
3203
3204         return rc;
3205 }
3206
3207 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
3208 {
3209         struct lnet_net  *net;
3210         struct lnet_ni *ni;
3211         __u32 net_id = LNET_NIDNET(conf->lic_nid);
3212         struct lnet_ping_buffer *pbuf;
3213         struct lnet_handle_md  ping_mdh;
3214         int               rc;
3215         int               net_count;
3216         __u32             addr;
3217
3218         /* don't allow userspace to shutdown the LOLND */
3219         if (LNET_NETTYP(net_id) == LOLND)
3220                 return -EINVAL;
3221
3222         mutex_lock(&the_lnet.ln_api_mutex);
3223
3224         lnet_net_lock(0);
3225
3226         net = lnet_get_net_locked(net_id);
3227         if (!net) {
3228                 CERROR("net %s not found\n",
3229                        libcfs_net2str(net_id));
3230                 rc = -ENOENT;
3231                 goto unlock_net;
3232         }
3233
3234         addr = LNET_NIDADDR(conf->lic_nid);
3235         if (addr == 0) {
3236                 /* remove the entire net */
3237                 net_count = lnet_get_net_ni_count_locked(net);
3238
3239                 lnet_net_unlock(0);
3240
3241                 /* create and link a new ping info, before removing the old one */
3242                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3243                                         lnet_get_ni_count() - net_count,
3244                                         false);
3245                 if (rc != 0)
3246                         goto unlock_api_mutex;
3247
3248                 lnet_shutdown_lndnet(net);
3249
3250                 if (lnet_count_acceptor_nets() == 0)
3251                         lnet_acceptor_stop();
3252
3253                 lnet_ping_target_update(pbuf, ping_mdh);
3254
3255                 goto unlock_api_mutex;
3256         }
3257
3258         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
3259         if (!ni) {
3260                 CERROR("nid %s not found\n",
3261                        libcfs_nid2str(conf->lic_nid));
3262                 rc = -ENOENT;
3263                 goto unlock_net;
3264         }
3265
3266         net_count = lnet_get_net_ni_count_locked(net);
3267
3268         lnet_net_unlock(0);
3269
3270         /* create and link a new ping info, before removing the old one */
3271         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3272                                   lnet_get_ni_count() - 1, false);
3273         if (rc != 0)
3274                 goto unlock_api_mutex;
3275
3276         lnet_shutdown_lndni(ni);
3277
3278         if (lnet_count_acceptor_nets() == 0)
3279                 lnet_acceptor_stop();
3280
3281         lnet_ping_target_update(pbuf, ping_mdh);
3282
3283         /* check if the net is empty and remove it if it is */
3284         if (net_count == 1)
3285                 lnet_shutdown_lndnet(net);
3286
3287         goto unlock_api_mutex;
3288
3289 unlock_net:
3290         lnet_net_unlock(0);
3291 unlock_api_mutex:
3292         mutex_unlock(&the_lnet.ln_api_mutex);
3293
3294         return rc;
3295 }
3296
3297 /*
3298  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
3299  * They are only expected to be called for unique networks.
3300  * That can be as a result of older DLC library
3301  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
3302  */
3303 int
3304 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
3305 {
3306         struct lnet_net         *net;
3307         struct list_head        net_head;
3308         int                     rc;
3309         struct lnet_ioctl_config_lnd_tunables tun;
3310         char *nets = conf->cfg_config_u.cfg_net.net_intf;
3311
3312         INIT_LIST_HEAD(&net_head);
3313
3314         /* Create a net/ni structures for the network string */
3315         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3316         if (rc <= 0)
3317                 return rc == 0 ? -EINVAL : rc;
3318
3319         mutex_lock(&the_lnet.ln_api_mutex);
3320
3321         if (rc > 1) {
3322                 rc = -EINVAL; /* only add one network per call */
3323                 goto out_unlock_clean;
3324         }
3325
3326         net = list_entry(net_head.next, struct lnet_net, net_list);
3327         list_del_init(&net->net_list);
3328
3329         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
3330
3331         memset(&tun, 0, sizeof(tun));
3332
3333         tun.lt_cmn.lct_peer_timeout =
3334           conf->cfg_config_u.cfg_net.net_peer_timeout;
3335         tun.lt_cmn.lct_peer_tx_credits =
3336           conf->cfg_config_u.cfg_net.net_peer_tx_credits;
3337         tun.lt_cmn.lct_peer_rtr_credits =
3338           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
3339         tun.lt_cmn.lct_max_tx_credits =
3340           conf->cfg_config_u.cfg_net.net_max_tx_credits;
3341
3342         rc = lnet_add_net_common(net, &tun);
3343
3344 out_unlock_clean:
3345         mutex_unlock(&the_lnet.ln_api_mutex);
3346         while (!list_empty(&net_head)) {
3347                 /* net_head list is empty in success case */
3348                 net = list_entry(net_head.next, struct lnet_net, net_list);
3349                 list_del_init(&net->net_list);
3350                 lnet_net_free(net);
3351         }
3352         return rc;
3353 }
3354
3355 int
3356 lnet_dyn_del_net(__u32 net_id)
3357 {
3358         struct lnet_net  *net;
3359         struct lnet_ping_buffer *pbuf;
3360         struct lnet_handle_md ping_mdh;
3361         int               rc;
3362         int               net_ni_count;
3363
3364         /* don't allow userspace to shutdown the LOLND */
3365         if (LNET_NETTYP(net_id) == LOLND)
3366                 return -EINVAL;
3367
3368         mutex_lock(&the_lnet.ln_api_mutex);
3369
3370         lnet_net_lock(0);
3371
3372         net = lnet_get_net_locked(net_id);
3373         if (net == NULL) {
3374                 lnet_net_unlock(0);
3375                 rc = -EINVAL;
3376                 goto out;
3377         }
3378
3379         net_ni_count = lnet_get_net_ni_count_locked(net);
3380
3381         lnet_net_unlock(0);
3382
3383         /* create and link a new ping info, before removing the old one */
3384         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3385                                     lnet_get_ni_count() - net_ni_count, false);
3386         if (rc != 0)
3387                 goto out;
3388
3389         lnet_shutdown_lndnet(net);
3390
3391         if (lnet_count_acceptor_nets() == 0)
3392                 lnet_acceptor_stop();
3393
3394         lnet_ping_target_update(pbuf, ping_mdh);
3395
3396 out:
3397         mutex_unlock(&the_lnet.ln_api_mutex);
3398
3399         return rc;
3400 }
3401
3402 void lnet_incr_dlc_seq(void)
3403 {
3404         atomic_inc(&lnet_dlc_seq_no);
3405 }
3406
3407 __u32 lnet_get_dlc_seq_locked(void)
3408 {
3409         return atomic_read(&lnet_dlc_seq_no);
3410 }
3411
3412 static void
3413 lnet_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3414 {
3415         struct lnet_net *net;
3416         struct lnet_ni *ni;
3417
3418         lnet_net_lock(LNET_LOCK_EX);
3419         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3420                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3421                         if (ni->ni_nid == nid || all) {
3422                                 atomic_set(&ni->ni_healthv, value);
3423                                 if (list_empty(&ni->ni_recovery) &&
3424                                     value < LNET_MAX_HEALTH_VALUE) {
3425                                         CERROR("manually adding local NI %s to recovery\n",
3426                                                libcfs_nid2str(ni->ni_nid));
3427                                         list_add_tail(&ni->ni_recovery,
3428                                                       &the_lnet.ln_mt_localNIRecovq);
3429                                         lnet_ni_addref_locked(ni, 0);
3430                                 }
3431                                 if (!all) {
3432                                         lnet_net_unlock(LNET_LOCK_EX);
3433                                         return;
3434                                 }
3435                         }
3436                 }
3437         }
3438         lnet_net_unlock(LNET_LOCK_EX);
3439 }
3440
3441 static int
3442 lnet_get_local_ni_hstats(struct lnet_ioctl_local_ni_hstats *stats)
3443 {
3444         int cpt, rc = 0;
3445         struct lnet_ni *ni;
3446         lnet_nid_t nid = stats->hlni_nid;
3447
3448         cpt = lnet_net_lock_current();
3449         ni = lnet_nid2ni_locked(nid, cpt);
3450
3451         if (!ni) {
3452                 rc = -ENOENT;
3453                 goto unlock;
3454         }
3455
3456         stats->hlni_local_interrupt = atomic_read(&ni->ni_hstats.hlt_local_interrupt);
3457         stats->hlni_local_dropped = atomic_read(&ni->ni_hstats.hlt_local_dropped);
3458         stats->hlni_local_aborted = atomic_read(&ni->ni_hstats.hlt_local_aborted);
3459         stats->hlni_local_no_route = atomic_read(&ni->ni_hstats.hlt_local_no_route);
3460         stats->hlni_local_timeout = atomic_read(&ni->ni_hstats.hlt_local_timeout);
3461         stats->hlni_local_error = atomic_read(&ni->ni_hstats.hlt_local_error);
3462         stats->hlni_health_value = atomic_read(&ni->ni_healthv);
3463
3464 unlock:
3465         lnet_net_unlock(cpt);
3466
3467         return rc;
3468 }
3469
3470 static int
3471 lnet_get_local_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3472 {
3473         struct lnet_ni *ni;
3474         int i = 0;
3475
3476         lnet_net_lock(LNET_LOCK_EX);
3477         list_for_each_entry(ni, &the_lnet.ln_mt_localNIRecovq, ni_recovery) {
3478                 list->rlst_nid_array[i] = ni->ni_nid;
3479                 i++;
3480                 if (i >= LNET_MAX_SHOW_NUM_NID)
3481                         break;
3482         }
3483         lnet_net_unlock(LNET_LOCK_EX);
3484         list->rlst_num_nids = i;
3485
3486         return 0;
3487 }
3488
3489 static int
3490 lnet_get_peer_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3491 {
3492         struct lnet_peer_ni *lpni;
3493         int i = 0;
3494
3495         lnet_net_lock(LNET_LOCK_EX);
3496         list_for_each_entry(lpni, &the_lnet.ln_mt_peerNIRecovq, lpni_recovery) {
3497                 list->rlst_nid_array[i] = lpni->lpni_nid;
3498                 i++;
3499                 if (i >= LNET_MAX_SHOW_NUM_NID)
3500                         break;
3501         }
3502         lnet_net_unlock(LNET_LOCK_EX);
3503         list->rlst_num_nids = i;
3504
3505         return 0;
3506 }
3507
3508 /**
3509  * LNet ioctl handler.
3510  *
3511  */
3512 int
3513 LNetCtl(unsigned int cmd, void *arg)
3514 {
3515         struct libcfs_ioctl_data *data = arg;
3516         struct lnet_ioctl_config_data *config;
3517         struct lnet_process_id    id = {0};
3518         struct lnet_ni           *ni;
3519         int                       rc;
3520
3521         BUILD_BUG_ON(sizeof(struct lnet_ioctl_net_config) +
3522                      sizeof(struct lnet_ioctl_config_data) > LIBCFS_IOC_DATA_MAX);
3523
3524         switch (cmd) {
3525         case IOC_LIBCFS_GET_NI:
3526                 rc = LNetGetId(data->ioc_count, &id);
3527                 data->ioc_nid = id.nid;
3528                 return rc;
3529
3530         case IOC_LIBCFS_FAIL_NID:
3531                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
3532
3533         case IOC_LIBCFS_ADD_ROUTE:
3534                 config = arg;
3535
3536                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3537                         return -EINVAL;
3538
3539                 mutex_lock(&the_lnet.ln_api_mutex);
3540                 rc = lnet_add_route(config->cfg_net,
3541                                     config->cfg_config_u.cfg_route.rtr_hop,
3542                                     config->cfg_nid,
3543                                     config->cfg_config_u.cfg_route.
3544                                         rtr_priority);
3545                 mutex_unlock(&the_lnet.ln_api_mutex);
3546                 return rc;
3547
3548         case IOC_LIBCFS_DEL_ROUTE:
3549                 config = arg;
3550
3551                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3552                         return -EINVAL;
3553
3554                 mutex_lock(&the_lnet.ln_api_mutex);
3555                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
3556                 mutex_unlock(&the_lnet.ln_api_mutex);
3557                 return rc;
3558
3559         case IOC_LIBCFS_GET_ROUTE:
3560                 config = arg;
3561
3562                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3563                         return -EINVAL;
3564
3565                 mutex_lock(&the_lnet.ln_api_mutex);
3566                 rc = lnet_get_route(config->cfg_count,
3567                                     &config->cfg_net,
3568                                     &config->cfg_config_u.cfg_route.rtr_hop,
3569                                     &config->cfg_nid,
3570                                     &config->cfg_config_u.cfg_route.rtr_flags,
3571                                     &config->cfg_config_u.cfg_route.
3572                                         rtr_priority);
3573                 mutex_unlock(&the_lnet.ln_api_mutex);
3574                 return rc;
3575
3576         case IOC_LIBCFS_GET_LOCAL_NI: {
3577                 struct lnet_ioctl_config_ni *cfg_ni;
3578                 struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3579                 struct lnet_ioctl_element_stats *stats;
3580                 __u32 tun_size;
3581
3582                 cfg_ni = arg;
3583
3584                 /* get the tunables if they are available */
3585                 if (cfg_ni->lic_cfg_hdr.ioc_len <
3586                     sizeof(*cfg_ni) + sizeof(*stats) + sizeof(*tun))
3587                         return -EINVAL;
3588
3589                 stats = (struct lnet_ioctl_element_stats *)
3590                         cfg_ni->lic_bulk;
3591                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3592                                 (cfg_ni->lic_bulk + sizeof(*stats));
3593
3594                 tun_size = cfg_ni->lic_cfg_hdr.ioc_len - sizeof(*cfg_ni) -
3595                         sizeof(*stats);
3596
3597                 mutex_lock(&the_lnet.ln_api_mutex);
3598                 rc = lnet_get_ni_config(cfg_ni, tun, stats, tun_size);
3599                 mutex_unlock(&the_lnet.ln_api_mutex);
3600                 return rc;
3601         }
3602
3603         case IOC_LIBCFS_GET_LOCAL_NI_MSG_STATS: {
3604                 struct lnet_ioctl_element_msg_stats *msg_stats = arg;
3605
3606                 if (msg_stats->im_hdr.ioc_len != sizeof(*msg_stats))
3607                         return -EINVAL;
3608
3609                 mutex_lock(&the_lnet.ln_api_mutex);
3610                 rc = lnet_get_ni_stats(msg_stats);
3611                 mutex_unlock(&the_lnet.ln_api_mutex);
3612
3613                 return rc;
3614         }
3615
3616         case IOC_LIBCFS_GET_NET: {
3617                 size_t total = sizeof(*config) +
3618                                sizeof(struct lnet_ioctl_net_config);
3619                 config = arg;
3620
3621                 if (config->cfg_hdr.ioc_len < total)
3622                         return -EINVAL;
3623
3624                 mutex_lock(&the_lnet.ln_api_mutex);
3625                 rc = lnet_get_net_config(config);
3626                 mutex_unlock(&the_lnet.ln_api_mutex);
3627                 return rc;
3628         }
3629
3630         case IOC_LIBCFS_GET_LNET_STATS:
3631         {
3632                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
3633
3634                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
3635                         return -EINVAL;
3636
3637                 mutex_lock(&the_lnet.ln_api_mutex);
3638                 lnet_counters_get(&lnet_stats->st_cntrs);
3639                 mutex_unlock(&the_lnet.ln_api_mutex);
3640                 return 0;
3641         }
3642
3643         case IOC_LIBCFS_CONFIG_RTR:
3644                 config = arg;
3645
3646                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3647                         return -EINVAL;
3648
3649                 mutex_lock(&the_lnet.ln_api_mutex);
3650                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
3651                         rc = lnet_rtrpools_enable();
3652                         mutex_unlock(&the_lnet.ln_api_mutex);
3653                         return rc;
3654                 }
3655                 lnet_rtrpools_disable();
3656                 mutex_unlock(&the_lnet.ln_api_mutex);
3657                 return 0;
3658
3659         case IOC_LIBCFS_ADD_BUF:
3660                 config = arg;
3661
3662                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3663                         return -EINVAL;
3664
3665                 mutex_lock(&the_lnet.ln_api_mutex);
3666                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
3667                                                 buf_tiny,
3668                                           config->cfg_config_u.cfg_buffers.
3669                                                 buf_small,
3670                                           config->cfg_config_u.cfg_buffers.
3671                                                 buf_large);
3672                 mutex_unlock(&the_lnet.ln_api_mutex);
3673                 return rc;
3674
3675         case IOC_LIBCFS_SET_NUMA_RANGE: {
3676                 struct lnet_ioctl_set_value *numa;
3677                 numa = arg;
3678                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
3679                         return -EINVAL;
3680                 lnet_net_lock(LNET_LOCK_EX);
3681                 lnet_numa_range = numa->sv_value;
3682                 lnet_net_unlock(LNET_LOCK_EX);
3683                 return 0;
3684         }
3685
3686         case IOC_LIBCFS_GET_NUMA_RANGE: {
3687                 struct lnet_ioctl_set_value *numa;
3688                 numa = arg;
3689                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
3690                         return -EINVAL;
3691                 numa->sv_value = lnet_numa_range;
3692                 return 0;
3693         }
3694
3695         case IOC_LIBCFS_GET_BUF: {
3696                 struct lnet_ioctl_pool_cfg *pool_cfg;
3697                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
3698
3699                 config = arg;
3700
3701                 if (config->cfg_hdr.ioc_len < total)
3702                         return -EINVAL;
3703
3704                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
3705
3706                 mutex_lock(&the_lnet.ln_api_mutex);
3707                 rc = lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
3708                 mutex_unlock(&the_lnet.ln_api_mutex);
3709                 return rc;
3710         }
3711
3712         case IOC_LIBCFS_GET_LOCAL_HSTATS: {
3713                 struct lnet_ioctl_local_ni_hstats *stats = arg;
3714
3715                 if (stats->hlni_hdr.ioc_len < sizeof(*stats))
3716                         return -EINVAL;
3717
3718                 mutex_lock(&the_lnet.ln_api_mutex);
3719                 rc = lnet_get_local_ni_hstats(stats);
3720                 mutex_unlock(&the_lnet.ln_api_mutex);
3721
3722                 return rc;
3723         }
3724
3725         case IOC_LIBCFS_GET_RECOVERY_QUEUE: {
3726                 struct lnet_ioctl_recovery_list *list = arg;
3727                 if (list->rlst_hdr.ioc_len < sizeof(*list))
3728                         return -EINVAL;
3729
3730                 mutex_lock(&the_lnet.ln_api_mutex);
3731                 if (list->rlst_type == LNET_HEALTH_TYPE_LOCAL_NI)
3732                         rc = lnet_get_local_ni_recovery_list(list);
3733                 else
3734                         rc = lnet_get_peer_ni_recovery_list(list);
3735                 mutex_unlock(&the_lnet.ln_api_mutex);
3736                 return rc;
3737         }
3738
3739         case IOC_LIBCFS_ADD_PEER_NI: {
3740                 struct lnet_ioctl_peer_cfg *cfg = arg;
3741
3742                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3743                         return -EINVAL;
3744
3745                 mutex_lock(&the_lnet.ln_api_mutex);
3746                 rc = lnet_add_peer_ni(cfg->prcfg_prim_nid,
3747                                       cfg->prcfg_cfg_nid,
3748                                       cfg->prcfg_mr);
3749                 mutex_unlock(&the_lnet.ln_api_mutex);
3750                 return rc;
3751         }
3752
3753         case IOC_LIBCFS_DEL_PEER_NI: {
3754                 struct lnet_ioctl_peer_cfg *cfg = arg;
3755
3756                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3757                         return -EINVAL;
3758
3759                 mutex_lock(&the_lnet.ln_api_mutex);
3760                 rc = lnet_del_peer_ni(cfg->prcfg_prim_nid,
3761                                       cfg->prcfg_cfg_nid);
3762                 mutex_unlock(&the_lnet.ln_api_mutex);
3763                 return rc;
3764         }
3765
3766         case IOC_LIBCFS_GET_PEER_INFO: {
3767                 struct lnet_ioctl_peer *peer_info = arg;
3768
3769                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
3770                         return -EINVAL;
3771
3772                 mutex_lock(&the_lnet.ln_api_mutex);
3773                 rc = lnet_get_peer_ni_info(
3774                    peer_info->pr_count,
3775                    &peer_info->pr_nid,
3776                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
3777                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
3778                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
3779                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
3780                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
3781                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
3782                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_tx_credits,
3783                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
3784                 mutex_unlock(&the_lnet.ln_api_mutex);
3785                 return rc;
3786         }
3787
3788         case IOC_LIBCFS_GET_PEER_NI: {
3789                 struct lnet_ioctl_peer_cfg *cfg = arg;
3790
3791                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3792                         return -EINVAL;
3793
3794                 mutex_lock(&the_lnet.ln_api_mutex);
3795                 rc = lnet_get_peer_info(cfg,
3796                                         (void __user *)cfg->prcfg_bulk);
3797                 mutex_unlock(&the_lnet.ln_api_mutex);
3798                 return rc;
3799         }
3800
3801         case IOC_LIBCFS_GET_PEER_LIST: {
3802                 struct lnet_ioctl_peer_cfg *cfg = arg;
3803
3804                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3805                         return -EINVAL;
3806
3807                 mutex_lock(&the_lnet.ln_api_mutex);
3808                 rc = lnet_get_peer_list(&cfg->prcfg_count, &cfg->prcfg_size,
3809                                 (struct lnet_process_id __user *)cfg->prcfg_bulk);
3810                 mutex_unlock(&the_lnet.ln_api_mutex);
3811                 return rc;
3812         }
3813
3814         case IOC_LIBCFS_SET_HEALHV: {
3815                 struct lnet_ioctl_reset_health_cfg *cfg = arg;
3816                 int value;
3817                 if (cfg->rh_hdr.ioc_len < sizeof(*cfg))
3818                         return -EINVAL;
3819                 if (cfg->rh_value < 0 ||
3820                     cfg->rh_value > LNET_MAX_HEALTH_VALUE)
3821                         value = LNET_MAX_HEALTH_VALUE;
3822                 else
3823                         value = cfg->rh_value;
3824                 CDEBUG(D_NET, "Manually setting healthv to %d for %s:%s. all = %d\n",
3825                        value, (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI) ?
3826                        "local" : "peer", libcfs_nid2str(cfg->rh_nid), cfg->rh_all);
3827                 mutex_lock(&the_lnet.ln_api_mutex);
3828                 if (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI)
3829                         lnet_ni_set_healthv(cfg->rh_nid, value,
3830                                              cfg->rh_all);
3831                 else
3832                         lnet_peer_ni_set_healthv(cfg->rh_nid, value,
3833                                                   cfg->rh_all);
3834                 mutex_unlock(&the_lnet.ln_api_mutex);
3835                 return 0;
3836         }
3837
3838         case IOC_LIBCFS_NOTIFY_ROUTER: {
3839                 time64_t deadline = ktime_get_real_seconds() - data->ioc_u64[0];
3840
3841                 /* The deadline passed in by the user should be some time in
3842                  * seconds in the future since the UNIX epoch. We have to map
3843                  * that deadline to the wall clock.
3844                  */
3845                 deadline += ktime_get_seconds();
3846                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
3847                                    deadline);
3848         }
3849
3850         case IOC_LIBCFS_LNET_DIST:
3851                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
3852                 if (rc < 0 && rc != -EHOSTUNREACH)
3853                         return rc;
3854
3855                 data->ioc_u32[0] = rc;
3856                 return 0;
3857
3858         case IOC_LIBCFS_TESTPROTOCOMPAT:
3859                 lnet_net_lock(LNET_LOCK_EX);
3860                 the_lnet.ln_testprotocompat = data->ioc_flags;
3861                 lnet_net_unlock(LNET_LOCK_EX);
3862                 return 0;
3863
3864         case IOC_LIBCFS_LNET_FAULT:
3865                 return lnet_fault_ctl(data->ioc_flags, data);
3866
3867         case IOC_LIBCFS_PING: {
3868                 signed long timeout;
3869
3870                 id.nid = data->ioc_nid;
3871                 id.pid = data->ioc_u32[0];
3872
3873                 /* If timeout is negative then set default of 3 minutes */
3874                 if (((s32)data->ioc_u32[1] <= 0) ||
3875                     data->ioc_u32[1] > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
3876                         timeout = msecs_to_jiffies(DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC);
3877                 else
3878                         timeout = msecs_to_jiffies(data->ioc_u32[1]);
3879
3880                 rc = lnet_ping(id, timeout, data->ioc_pbuf1,
3881                                data->ioc_plen1 / sizeof(struct lnet_process_id));
3882
3883                 if (rc < 0)
3884                         return rc;
3885
3886                 data->ioc_count = rc;
3887                 return 0;
3888         }
3889
3890         case IOC_LIBCFS_PING_PEER: {
3891                 struct lnet_ioctl_ping_data *ping = arg;
3892                 struct lnet_peer *lp;
3893                 signed long timeout;
3894
3895                 /* If timeout is negative then set default of 3 minutes */
3896                 if (((s32)ping->op_param) <= 0 ||
3897                     ping->op_param > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
3898                         timeout = msecs_to_jiffies(DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC);
3899                 else
3900                         timeout = msecs_to_jiffies(ping->op_param);
3901
3902                 rc = lnet_ping(ping->ping_id, timeout,
3903                                ping->ping_buf,
3904                                ping->ping_count);
3905                 if (rc < 0)
3906                         return rc;
3907
3908                 mutex_lock(&the_lnet.ln_api_mutex);
3909                 lp = lnet_find_peer(ping->ping_id.nid);
3910                 if (lp) {
3911                         ping->ping_id.nid = lp->lp_primary_nid;
3912                         ping->mr_info = lnet_peer_is_multi_rail(lp);
3913                         lnet_peer_decref_locked(lp);
3914                 }
3915                 mutex_unlock(&the_lnet.ln_api_mutex);
3916
3917                 ping->ping_count = rc;
3918                 return 0;
3919         }
3920
3921         case IOC_LIBCFS_DISCOVER: {
3922                 struct lnet_ioctl_ping_data *discover = arg;
3923                 struct lnet_peer *lp;
3924
3925                 rc = lnet_discover(discover->ping_id, discover->op_param,
3926                                    discover->ping_buf,
3927                                    discover->ping_count);
3928                 if (rc < 0)
3929                         return rc;
3930
3931                 mutex_lock(&the_lnet.ln_api_mutex);
3932                 lp = lnet_find_peer(discover->ping_id.nid);
3933                 if (lp) {
3934                         discover->ping_id.nid = lp->lp_primary_nid;
3935                         discover->mr_info = lnet_peer_is_multi_rail(lp);
3936                         lnet_peer_decref_locked(lp);
3937                 }
3938                 mutex_unlock(&the_lnet.ln_api_mutex);
3939
3940                 discover->ping_count = rc;
3941                 return 0;
3942         }
3943
3944         default:
3945                 ni = lnet_net2ni_addref(data->ioc_net);
3946                 if (ni == NULL)
3947                         return -EINVAL;
3948
3949                 if (ni->ni_net->net_lnd->lnd_ctl == NULL)
3950                         rc = -EINVAL;
3951                 else
3952                         rc = ni->ni_net->net_lnd->lnd_ctl(ni, cmd, arg);
3953
3954                 lnet_ni_decref(ni);
3955                 return rc;
3956         }
3957         /* not reached */
3958 }
3959 EXPORT_SYMBOL(LNetCtl);
3960
3961 void LNetDebugPeer(struct lnet_process_id id)
3962 {
3963         lnet_debug_peer(id.nid);
3964 }
3965 EXPORT_SYMBOL(LNetDebugPeer);
3966
3967 /**
3968  * Determine if the specified peer \a nid is on the local node.
3969  *
3970  * \param nid   peer nid to check
3971  *
3972  * \retval true         If peer NID is on the local node.
3973  * \retval false        If peer NID is not on the local node.
3974  */
3975 bool LNetIsPeerLocal(lnet_nid_t nid)
3976 {
3977         struct lnet_net *net;
3978         struct lnet_ni *ni;
3979         int cpt;
3980
3981         cpt = lnet_net_lock_current();
3982         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3983                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3984                         if (ni->ni_nid == nid) {
3985                                 lnet_net_unlock(cpt);
3986                                 return true;
3987                         }
3988                 }
3989         }
3990         lnet_net_unlock(cpt);
3991
3992         return false;
3993 }
3994 EXPORT_SYMBOL(LNetIsPeerLocal);
3995
3996 /**
3997  * Retrieve the struct lnet_process_id ID of LNet interface at \a index.
3998  * Note that all interfaces share a same PID, as requested by LNetNIInit().
3999  *
4000  * \param index Index of the interface to look up.
4001  * \param id On successful return, this location will hold the
4002  * struct lnet_process_id ID of the interface.
4003  *
4004  * \retval 0 If an interface exists at \a index.
4005  * \retval -ENOENT If no interface has been found.
4006  */
4007 int
4008 LNetGetId(unsigned int index, struct lnet_process_id *id)
4009 {
4010         struct lnet_ni   *ni;
4011         struct lnet_net  *net;
4012         int               cpt;
4013         int               rc = -ENOENT;
4014
4015         LASSERT(the_lnet.ln_refcount > 0);
4016
4017         cpt = lnet_net_lock_current();
4018
4019         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
4020                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
4021                         if (index-- != 0)
4022                                 continue;
4023
4024                         id->nid = ni->ni_nid;
4025                         id->pid = the_lnet.ln_pid;
4026                         rc = 0;
4027                         break;
4028                 }
4029         }
4030
4031         lnet_net_unlock(cpt);
4032         return rc;
4033 }
4034 EXPORT_SYMBOL(LNetGetId);
4035
4036 static int lnet_ping(struct lnet_process_id id, signed long timeout,
4037                      struct lnet_process_id __user *ids, int n_ids)
4038 {
4039         struct lnet_handle_eq eqh;
4040         struct lnet_handle_md mdh;
4041         struct lnet_event event;
4042         struct lnet_md md = { NULL };
4043         int which;
4044         int unlinked = 0;
4045         int replied = 0;
4046         const signed long a_long_time = msecs_to_jiffies(60 * MSEC_PER_SEC);
4047         struct lnet_ping_buffer *pbuf;
4048         struct lnet_process_id tmpid;
4049         int i;
4050         int nob;
4051         int rc;
4052         int rc2;
4053         sigset_t blocked;
4054
4055         /* n_ids limit is arbitrary */
4056         if (n_ids <= 0 || id.nid == LNET_NID_ANY)
4057                 return -EINVAL;
4058
4059         /*
4060          * if the user buffer has more space than the lnet_interfaces_max
4061          * then only fill it up to lnet_interfaces_max
4062          */
4063         if (n_ids > lnet_interfaces_max)
4064                 n_ids = lnet_interfaces_max;
4065
4066         if (id.pid == LNET_PID_ANY)
4067                 id.pid = LNET_PID_LUSTRE;
4068
4069         pbuf = lnet_ping_buffer_alloc(n_ids, GFP_NOFS);
4070         if (!pbuf)
4071                 return -ENOMEM;
4072
4073         /* NB 2 events max (including any unlink event) */
4074         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
4075         if (rc != 0) {
4076                 CERROR("Can't allocate EQ: %d\n", rc);
4077                 goto fail_ping_buffer_decref;
4078         }
4079
4080         /* initialize md content */
4081         md.start     = &pbuf->pb_info;
4082         md.length    = LNET_PING_INFO_SIZE(n_ids);
4083         md.threshold = 2; /* GET/REPLY */
4084         md.max_size  = 0;
4085         md.options   = LNET_MD_TRUNCATE;
4086         md.user_ptr  = NULL;
4087         md.eq_handle = eqh;
4088
4089         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
4090         if (rc != 0) {
4091                 CERROR("Can't bind MD: %d\n", rc);
4092                 goto fail_free_eq;
4093         }
4094
4095         rc = LNetGet(LNET_NID_ANY, mdh, id,
4096                      LNET_RESERVED_PORTAL,
4097                      LNET_PROTO_PING_MATCHBITS, 0, false);
4098
4099         if (rc != 0) {
4100                 /* Don't CERROR; this could be deliberate! */
4101                 rc2 = LNetMDUnlink(mdh);
4102                 LASSERT(rc2 == 0);
4103
4104                 /* NB must wait for the UNLINK event below... */
4105                 unlinked = 1;
4106                 timeout = a_long_time;
4107         }
4108
4109         do {
4110                 /* MUST block for unlink to complete */
4111                 if (unlinked)
4112                         blocked = cfs_block_allsigs();
4113
4114                 rc2 = LNetEQPoll(&eqh, 1, timeout, &event, &which);
4115
4116                 if (unlinked)
4117                         cfs_restore_sigs(blocked);
4118
4119                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
4120                        (rc2 <= 0) ? -1 : event.type,
4121                        (rc2 <= 0) ? -1 : event.status,
4122                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
4123
4124                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
4125
4126                 if (rc2 <= 0 || event.status != 0) {
4127                         /* timeout or error */
4128                         if (!replied && rc == 0)
4129                                 rc = (rc2 < 0) ? rc2 :
4130                                      (rc2 == 0) ? -ETIMEDOUT :
4131                                      event.status;
4132
4133                         if (!unlinked) {
4134                                 /* Ensure completion in finite time... */
4135                                 LNetMDUnlink(mdh);
4136                                 /* No assertion (racing with network) */
4137                                 unlinked = 1;
4138                                 timeout = a_long_time;
4139                         } else if (rc2 == 0) {
4140                                 /* timed out waiting for unlink */
4141                                 CWARN("ping %s: late network completion\n",
4142                                       libcfs_id2str(id));
4143                         }
4144                 } else if (event.type == LNET_EVENT_REPLY) {
4145                         replied = 1;
4146                         rc = event.mlength;
4147                 }
4148         } while (rc2 <= 0 || !event.unlinked);
4149
4150         if (!replied) {
4151                 if (rc >= 0)
4152                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
4153                               libcfs_id2str(id));
4154                 rc = -EIO;
4155                 goto fail_free_eq;
4156         }
4157
4158         nob = rc;
4159         LASSERT(nob >= 0 && nob <= LNET_PING_INFO_SIZE(n_ids));
4160
4161         rc = -EPROTO;           /* if I can't parse... */
4162
4163         if (nob < 8) {
4164                 CERROR("%s: ping info too short %d\n",
4165                        libcfs_id2str(id), nob);
4166                 goto fail_free_eq;
4167         }
4168
4169         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
4170                 lnet_swap_pinginfo(pbuf);
4171         } else if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
4172                 CERROR("%s: Unexpected magic %08x\n",
4173                        libcfs_id2str(id), pbuf->pb_info.pi_magic);
4174                 goto fail_free_eq;
4175         }
4176
4177         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
4178                 CERROR("%s: ping w/o NI status: 0x%x\n",
4179                        libcfs_id2str(id), pbuf->pb_info.pi_features);
4180                 goto fail_free_eq;
4181         }
4182
4183         if (nob < LNET_PING_INFO_SIZE(0)) {
4184                 CERROR("%s: Short reply %d(%d min)\n",
4185                        libcfs_id2str(id),
4186                        nob, (int)LNET_PING_INFO_SIZE(0));
4187                 goto fail_free_eq;
4188         }
4189
4190         if (pbuf->pb_info.pi_nnis < n_ids)
4191                 n_ids = pbuf->pb_info.pi_nnis;
4192
4193         if (nob < LNET_PING_INFO_SIZE(n_ids)) {
4194                 CERROR("%s: Short reply %d(%d expected)\n",
4195                        libcfs_id2str(id),
4196                        nob, (int)LNET_PING_INFO_SIZE(n_ids));
4197                 goto fail_free_eq;
4198         }
4199
4200         rc = -EFAULT;           /* if I segv in copy_to_user()... */
4201
4202         memset(&tmpid, 0, sizeof(tmpid));
4203         for (i = 0; i < n_ids; i++) {
4204                 tmpid.pid = pbuf->pb_info.pi_pid;
4205                 tmpid.nid = pbuf->pb_info.pi_ni[i].ns_nid;
4206                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
4207                         goto fail_free_eq;
4208         }
4209         rc = pbuf->pb_info.pi_nnis;
4210
4211  fail_free_eq:
4212         rc2 = LNetEQFree(eqh);
4213         if (rc2 != 0)
4214                 CERROR("rc2 %d\n", rc2);
4215         LASSERT(rc2 == 0);
4216
4217  fail_ping_buffer_decref:
4218         lnet_ping_buffer_decref(pbuf);
4219         return rc;
4220 }
4221
4222 static int
4223 lnet_discover(struct lnet_process_id id, __u32 force,
4224               struct lnet_process_id __user *ids, int n_ids)
4225 {
4226         struct lnet_peer_ni *lpni;
4227         struct lnet_peer_ni *p;
4228         struct lnet_peer *lp;
4229         struct lnet_process_id *buf;
4230         int cpt;
4231         int i;
4232         int rc;
4233         int max_intf = lnet_interfaces_max;
4234         size_t buf_size;
4235
4236         if (n_ids <= 0 ||
4237             id.nid == LNET_NID_ANY)
4238                 return -EINVAL;
4239
4240         if (id.pid == LNET_PID_ANY)
4241                 id.pid = LNET_PID_LUSTRE;
4242
4243         /*
4244          * if the user buffer has more space than the max_intf
4245          * then only fill it up to max_intf
4246          */
4247         if (n_ids > max_intf)
4248                 n_ids = max_intf;
4249
4250         buf_size = n_ids * sizeof(*buf);
4251
4252         LIBCFS_ALLOC(buf, buf_size);
4253         if (!buf)
4254                 return -ENOMEM;
4255
4256         cpt = lnet_net_lock_current();
4257         lpni = lnet_nid2peerni_locked(id.nid, LNET_NID_ANY, cpt);
4258         if (IS_ERR(lpni)) {
4259                 rc = PTR_ERR(lpni);
4260                 goto out;
4261         }
4262
4263         /*
4264          * Clearing the NIDS_UPTODATE flag ensures the peer will
4265          * be discovered, provided discovery has not been disabled.
4266          */
4267         lp = lpni->lpni_peer_net->lpn_peer;
4268         spin_lock(&lp->lp_lock);
4269         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
4270         /* If the force flag is set, force a PING and PUSH as well. */
4271         if (force)
4272                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
4273         spin_unlock(&lp->lp_lock);
4274         rc = lnet_discover_peer_locked(lpni, cpt, true);
4275         if (rc)
4276                 goto out_decref;
4277
4278         /* Peer may have changed. */
4279         lp = lpni->lpni_peer_net->lpn_peer;
4280         if (lp->lp_nnis < n_ids)
4281                 n_ids = lp->lp_nnis;
4282
4283         i = 0;
4284         p = NULL;
4285         while ((p = lnet_get_next_peer_ni_locked(lp, NULL, p)) != NULL) {
4286                 buf[i].pid = id.pid;
4287                 buf[i].nid = p->lpni_nid;
4288                 if (++i >= n_ids)
4289                         break;
4290         }
4291
4292         lnet_net_unlock(cpt);
4293
4294         rc = -EFAULT;
4295         if (copy_to_user(ids, buf, n_ids * sizeof(*buf)))
4296                 goto out_relock;
4297         rc = n_ids;
4298 out_relock:
4299         lnet_net_lock(cpt);
4300 out_decref:
4301         lnet_peer_ni_decref_locked(lpni);
4302 out:
4303         lnet_net_unlock(cpt);
4304
4305         LIBCFS_FREE(buf, buf_size);
4306
4307         return rc;
4308 }
4309
4310 /**
4311  * Retrieve peer discovery status.
4312  *
4313  * \retval 1 if lnet_peer_discovery_disabled is 0
4314  * \retval 0 if lnet_peer_discovery_disabled is 1
4315  */
4316 int
4317 LNetGetPeerDiscoveryStatus(void)
4318 {
4319         return !lnet_peer_discovery_disabled;
4320 }
4321 EXPORT_SYMBOL(LNetGetPeerDiscoveryStatus);