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 int
1540 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1541 {
1542         if (!pinfo)
1543                 return -EINVAL;
1544         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1545                 return -EPROTO;
1546         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1547                 return -EPROTO;
1548         /* Loopback is guaranteed to be present */
1549         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1550                 return -ERANGE;
1551         if (LNET_NETTYP(LNET_NIDNET(LNET_PING_INFO_LONI(pinfo))) != LOLND)
1552                 return -EPROTO;
1553         return 0;
1554 }
1555
1556 static void
1557 lnet_ping_target_destroy(void)
1558 {
1559         struct lnet_net *net;
1560         struct lnet_ni  *ni;
1561
1562         lnet_net_lock(LNET_LOCK_EX);
1563
1564         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1565                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1566                         lnet_ni_lock(ni);
1567                         ni->ni_status = NULL;
1568                         lnet_ni_unlock(ni);
1569                 }
1570         }
1571
1572         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1573         the_lnet.ln_ping_target = NULL;
1574
1575         lnet_net_unlock(LNET_LOCK_EX);
1576 }
1577
1578 static void
1579 lnet_ping_target_event_handler(struct lnet_event *event)
1580 {
1581         struct lnet_ping_buffer *pbuf = event->md.user_ptr;
1582
1583         if (event->unlinked)
1584                 lnet_ping_buffer_decref(pbuf);
1585 }
1586
1587 static int
1588 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1589                        struct lnet_handle_md *ping_mdh,
1590                        int ni_count, bool set_eq)
1591 {
1592         struct lnet_process_id id = {
1593                 .nid = LNET_NID_ANY,
1594                 .pid = LNET_PID_ANY
1595         };
1596         struct lnet_handle_me me_handle;
1597         struct lnet_md md = { NULL };
1598         int rc, rc2;
1599
1600         if (set_eq) {
1601                 rc = LNetEQAlloc(0, lnet_ping_target_event_handler,
1602                                  &the_lnet.ln_ping_target_eq);
1603                 if (rc != 0) {
1604                         CERROR("Can't allocate ping buffer EQ: %d\n", rc);
1605                         return rc;
1606                 }
1607         }
1608
1609         *ppbuf = lnet_ping_target_create(ni_count);
1610         if (*ppbuf == NULL) {
1611                 rc = -ENOMEM;
1612                 goto fail_free_eq;
1613         }
1614
1615         /* Ping target ME/MD */
1616         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1617                           LNET_PROTO_PING_MATCHBITS, 0,
1618                           LNET_UNLINK, LNET_INS_AFTER,
1619                           &me_handle);
1620         if (rc != 0) {
1621                 CERROR("Can't create ping target ME: %d\n", rc);
1622                 goto fail_decref_ping_buffer;
1623         }
1624
1625         /* initialize md content */
1626         md.start     = &(*ppbuf)->pb_info;
1627         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1628         md.threshold = LNET_MD_THRESH_INF;
1629         md.max_size  = 0;
1630         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1631                        LNET_MD_MANAGE_REMOTE;
1632         md.eq_handle = the_lnet.ln_ping_target_eq;
1633         md.user_ptr  = *ppbuf;
1634
1635         rc = LNetMDAttach(me_handle, md, LNET_RETAIN, ping_mdh);
1636         if (rc != 0) {
1637                 CERROR("Can't attach ping target MD: %d\n", rc);
1638                 goto fail_unlink_ping_me;
1639         }
1640         lnet_ping_buffer_addref(*ppbuf);
1641
1642         return 0;
1643
1644 fail_unlink_ping_me:
1645         rc2 = LNetMEUnlink(me_handle);
1646         LASSERT(rc2 == 0);
1647 fail_decref_ping_buffer:
1648         LASSERT(lnet_ping_buffer_numref(*ppbuf) == 1);
1649         lnet_ping_buffer_decref(*ppbuf);
1650         *ppbuf = NULL;
1651 fail_free_eq:
1652         if (set_eq) {
1653                 rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1654                 LASSERT(rc2 == 0);
1655         }
1656         return rc;
1657 }
1658
1659 static void
1660 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1661                     struct lnet_handle_md *ping_mdh)
1662 {
1663         sigset_t        blocked = cfs_block_allsigs();
1664
1665         LNetMDUnlink(*ping_mdh);
1666         LNetInvalidateMDHandle(ping_mdh);
1667
1668         /* NB the MD could be busy; this just starts the unlink */
1669         while (lnet_ping_buffer_numref(pbuf) > 1) {
1670                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
1671                 set_current_state(TASK_UNINTERRUPTIBLE);
1672                 schedule_timeout(cfs_time_seconds(1));
1673         }
1674
1675         cfs_restore_sigs(blocked);
1676 }
1677
1678 static void
1679 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
1680 {
1681         struct lnet_ni          *ni;
1682         struct lnet_net         *net;
1683         struct lnet_ni_status *ns;
1684         int                     i;
1685         int                     rc;
1686
1687         i = 0;
1688         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1689                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1690                         LASSERT(i < pbuf->pb_nnis);
1691
1692                         ns = &pbuf->pb_info.pi_ni[i];
1693
1694                         ns->ns_nid = ni->ni_nid;
1695
1696                         lnet_ni_lock(ni);
1697                         ns->ns_status = (ni->ni_status != NULL) ?
1698                                          ni->ni_status->ns_status :
1699                                                 LNET_NI_STATUS_UP;
1700                         ni->ni_status = ns;
1701                         lnet_ni_unlock(ni);
1702
1703                         i++;
1704                 }
1705         }
1706         /*
1707          * We (ab)use the ns_status of the loopback interface to
1708          * transmit the sequence number. The first interface listed
1709          * must be the loopback interface.
1710          */
1711         rc = lnet_ping_info_validate(&pbuf->pb_info);
1712         if (rc) {
1713                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
1714                 LBUG();
1715         }
1716         LNET_PING_BUFFER_SEQNO(pbuf) =
1717                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
1718 }
1719
1720 static void
1721 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
1722                         struct lnet_handle_md ping_mdh)
1723 {
1724         struct lnet_ping_buffer *old_pbuf = NULL;
1725         struct lnet_handle_md old_ping_md;
1726
1727         /* switch the NIs to point to the new ping info created */
1728         lnet_net_lock(LNET_LOCK_EX);
1729
1730         if (!the_lnet.ln_routing)
1731                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1732         if (!lnet_peer_discovery_disabled)
1733                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
1734
1735         /* Ensure only known feature bits have been set. */
1736         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
1737         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
1738
1739         lnet_ping_target_install_locked(pbuf);
1740
1741         if (the_lnet.ln_ping_target) {
1742                 old_pbuf = the_lnet.ln_ping_target;
1743                 old_ping_md = the_lnet.ln_ping_target_md;
1744         }
1745         the_lnet.ln_ping_target_md = ping_mdh;
1746         the_lnet.ln_ping_target = pbuf;
1747
1748         lnet_net_unlock(LNET_LOCK_EX);
1749
1750         if (old_pbuf) {
1751                 /* unlink and free the old ping info */
1752                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
1753                 lnet_ping_buffer_decref(old_pbuf);
1754         }
1755
1756         lnet_push_update_to_peers(0);
1757 }
1758
1759 static void
1760 lnet_ping_target_fini(void)
1761 {
1762         int             rc;
1763
1764         lnet_ping_md_unlink(the_lnet.ln_ping_target,
1765                             &the_lnet.ln_ping_target_md);
1766
1767         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1768         LASSERT(rc == 0);
1769
1770         lnet_ping_target_destroy();
1771 }
1772
1773 /* Resize the push target. */
1774 int lnet_push_target_resize(void)
1775 {
1776         struct lnet_process_id id = { LNET_NID_ANY, LNET_PID_ANY };
1777         struct lnet_md md = { NULL };
1778         struct lnet_handle_me meh;
1779         struct lnet_handle_md mdh;
1780         struct lnet_handle_md old_mdh;
1781         struct lnet_ping_buffer *pbuf;
1782         struct lnet_ping_buffer *old_pbuf;
1783         int nnis = the_lnet.ln_push_target_nnis;
1784         int rc;
1785
1786         if (nnis <= 0) {
1787                 rc = -EINVAL;
1788                 goto fail_return;
1789         }
1790 again:
1791         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1792         if (!pbuf) {
1793                 rc = -ENOMEM;
1794                 goto fail_return;
1795         }
1796
1797         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1798                           LNET_PROTO_PING_MATCHBITS, 0,
1799                           LNET_UNLINK, LNET_INS_AFTER,
1800                           &meh);
1801         if (rc) {
1802                 CERROR("Can't create push target ME: %d\n", rc);
1803                 goto fail_decref_pbuf;
1804         }
1805
1806         /* initialize md content */
1807         md.start     = &pbuf->pb_info;
1808         md.length    = LNET_PING_INFO_SIZE(nnis);
1809         md.threshold = LNET_MD_THRESH_INF;
1810         md.max_size  = 0;
1811         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE |
1812                        LNET_MD_MANAGE_REMOTE;
1813         md.user_ptr  = pbuf;
1814         md.eq_handle = the_lnet.ln_push_target_eq;
1815
1816         rc = LNetMDAttach(meh, md, LNET_RETAIN, &mdh);
1817         if (rc) {
1818                 CERROR("Can't attach push MD: %d\n", rc);
1819                 goto fail_unlink_meh;
1820         }
1821         lnet_ping_buffer_addref(pbuf);
1822
1823         lnet_net_lock(LNET_LOCK_EX);
1824         old_pbuf = the_lnet.ln_push_target;
1825         old_mdh = the_lnet.ln_push_target_md;
1826         the_lnet.ln_push_target = pbuf;
1827         the_lnet.ln_push_target_md = mdh;
1828         lnet_net_unlock(LNET_LOCK_EX);
1829
1830         if (old_pbuf) {
1831                 LNetMDUnlink(old_mdh);
1832                 lnet_ping_buffer_decref(old_pbuf);
1833         }
1834
1835         if (nnis < the_lnet.ln_push_target_nnis)
1836                 goto again;
1837
1838         CDEBUG(D_NET, "nnis %d success\n", nnis);
1839
1840         return 0;
1841
1842 fail_unlink_meh:
1843         LNetMEUnlink(meh);
1844 fail_decref_pbuf:
1845         lnet_ping_buffer_decref(pbuf);
1846 fail_return:
1847         CDEBUG(D_NET, "nnis %d error %d\n", nnis, rc);
1848         return rc;
1849 }
1850
1851 static void lnet_push_target_event_handler(struct lnet_event *ev)
1852 {
1853         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1854
1855         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
1856                 lnet_swap_pinginfo(pbuf);
1857
1858         lnet_peer_push_event(ev);
1859         if (ev->unlinked)
1860                 lnet_ping_buffer_decref(pbuf);
1861 }
1862
1863 /* Initialize the push target. */
1864 static int lnet_push_target_init(void)
1865 {
1866         int rc;
1867
1868         if (the_lnet.ln_push_target)
1869                 return -EALREADY;
1870
1871         rc = LNetEQAlloc(0, lnet_push_target_event_handler,
1872                          &the_lnet.ln_push_target_eq);
1873         if (rc) {
1874                 CERROR("Can't allocated push target EQ: %d\n", rc);
1875                 return rc;
1876         }
1877
1878         /* Start at the required minimum, we'll enlarge if required. */
1879         the_lnet.ln_push_target_nnis = LNET_INTERFACES_MIN;
1880
1881         rc = lnet_push_target_resize();
1882
1883         if (rc) {
1884                 LNetEQFree(the_lnet.ln_push_target_eq);
1885                 LNetInvalidateEQHandle(&the_lnet.ln_push_target_eq);
1886         }
1887
1888         return rc;
1889 }
1890
1891 /* Clean up the push target. */
1892 static void lnet_push_target_fini(void)
1893 {
1894         if (!the_lnet.ln_push_target)
1895                 return;
1896
1897         /* Unlink and invalidate to prevent new references. */
1898         LNetMDUnlink(the_lnet.ln_push_target_md);
1899         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
1900
1901         /* Wait for the unlink to complete. */
1902         while (lnet_ping_buffer_numref(the_lnet.ln_push_target) > 1) {
1903                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
1904                 set_current_state(TASK_UNINTERRUPTIBLE);
1905                 schedule_timeout(cfs_time_seconds(1));
1906         }
1907
1908         lnet_ping_buffer_decref(the_lnet.ln_push_target);
1909         the_lnet.ln_push_target = NULL;
1910         the_lnet.ln_push_target_nnis = 0;
1911
1912         LNetEQFree(the_lnet.ln_push_target_eq);
1913         LNetInvalidateEQHandle(&the_lnet.ln_push_target_eq);
1914 }
1915
1916 static int
1917 lnet_ni_tq_credits(struct lnet_ni *ni)
1918 {
1919         int     credits;
1920
1921         LASSERT(ni->ni_ncpts >= 1);
1922
1923         if (ni->ni_ncpts == 1)
1924                 return ni->ni_net->net_tunables.lct_max_tx_credits;
1925
1926         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
1927         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
1928         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
1929
1930         return credits;
1931 }
1932
1933 static void
1934 lnet_ni_unlink_locked(struct lnet_ni *ni)
1935 {
1936         /* move it to zombie list and nobody can find it anymore */
1937         LASSERT(!list_empty(&ni->ni_netlist));
1938         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
1939         lnet_ni_decref_locked(ni, 0);
1940 }
1941
1942 static void
1943 lnet_clear_zombies_nis_locked(struct lnet_net *net)
1944 {
1945         int             i;
1946         int             islo;
1947         struct lnet_ni  *ni;
1948         struct list_head *zombie_list = &net->net_ni_zombie;
1949
1950         /*
1951          * Now wait for the NIs I just nuked to show up on the zombie
1952          * list and shut them down in guaranteed thread context
1953          */
1954         i = 2;
1955         while (!list_empty(zombie_list)) {
1956                 int     *ref;
1957                 int     j;
1958
1959                 ni = list_entry(zombie_list->next,
1960                                 struct lnet_ni, ni_netlist);
1961                 list_del_init(&ni->ni_netlist);
1962                 /* the ni should be in deleting state. If it's not it's
1963                  * a bug */
1964                 LASSERT(ni->ni_state == LNET_NI_STATE_DELETING);
1965                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1966                         if (*ref == 0)
1967                                 continue;
1968                         /* still busy, add it back to zombie list */
1969                         list_add(&ni->ni_netlist, zombie_list);
1970                         break;
1971                 }
1972
1973                 if (!list_empty(&ni->ni_netlist)) {
1974                         lnet_net_unlock(LNET_LOCK_EX);
1975                         ++i;
1976                         if ((i & (-i)) == i) {
1977                                 CDEBUG(D_WARNING,
1978                                        "Waiting for zombie LNI %s\n",
1979                                        libcfs_nid2str(ni->ni_nid));
1980                         }
1981                         set_current_state(TASK_UNINTERRUPTIBLE);
1982                         schedule_timeout(cfs_time_seconds(1));
1983                         lnet_net_lock(LNET_LOCK_EX);
1984                         continue;
1985                 }
1986
1987                 lnet_net_unlock(LNET_LOCK_EX);
1988
1989                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
1990
1991                 LASSERT(!in_interrupt());
1992                 (net->net_lnd->lnd_shutdown)(ni);
1993
1994                 if (!islo)
1995                         CDEBUG(D_LNI, "Removed LNI %s\n",
1996                               libcfs_nid2str(ni->ni_nid));
1997
1998                 lnet_ni_free(ni);
1999                 i = 2;
2000                 lnet_net_lock(LNET_LOCK_EX);
2001         }
2002 }
2003
2004 /* shutdown down the NI and release refcount */
2005 static void
2006 lnet_shutdown_lndni(struct lnet_ni *ni)
2007 {
2008         int i;
2009         struct lnet_net *net = ni->ni_net;
2010
2011         lnet_net_lock(LNET_LOCK_EX);
2012         lnet_ni_lock(ni);
2013         ni->ni_state = LNET_NI_STATE_DELETING;
2014         lnet_ni_unlock(ni);
2015         lnet_ni_unlink_locked(ni);
2016         lnet_incr_dlc_seq();
2017         lnet_net_unlock(LNET_LOCK_EX);
2018
2019         /* clear messages for this NI on the lazy portal */
2020         for (i = 0; i < the_lnet.ln_nportals; i++)
2021                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
2022
2023         lnet_net_lock(LNET_LOCK_EX);
2024         lnet_clear_zombies_nis_locked(net);
2025         lnet_net_unlock(LNET_LOCK_EX);
2026 }
2027
2028 static void
2029 lnet_shutdown_lndnet(struct lnet_net *net)
2030 {
2031         struct lnet_ni *ni;
2032
2033         lnet_net_lock(LNET_LOCK_EX);
2034
2035         net->net_state = LNET_NET_STATE_DELETING;
2036
2037         list_del_init(&net->net_list);
2038
2039         while (!list_empty(&net->net_ni_list)) {
2040                 ni = list_entry(net->net_ni_list.next,
2041                                 struct lnet_ni, ni_netlist);
2042                 lnet_net_unlock(LNET_LOCK_EX);
2043                 lnet_shutdown_lndni(ni);
2044                 lnet_net_lock(LNET_LOCK_EX);
2045         }
2046
2047         lnet_net_unlock(LNET_LOCK_EX);
2048
2049         /* Do peer table cleanup for this net */
2050         lnet_peer_tables_cleanup(net);
2051
2052         lnet_net_lock(LNET_LOCK_EX);
2053         /*
2054          * decrement ref count on lnd only when the entire network goes
2055          * away
2056          */
2057         net->net_lnd->lnd_refcount--;
2058
2059         lnet_net_unlock(LNET_LOCK_EX);
2060
2061         lnet_net_free(net);
2062 }
2063
2064 static void
2065 lnet_shutdown_lndnets(void)
2066 {
2067         struct lnet_net *net;
2068         struct list_head resend;
2069         struct lnet_msg *msg, *tmp;
2070
2071         INIT_LIST_HEAD(&resend);
2072
2073         /* NB called holding the global mutex */
2074
2075         /* All quiet on the API front */
2076         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
2077         LASSERT(the_lnet.ln_refcount == 0);
2078
2079         lnet_net_lock(LNET_LOCK_EX);
2080         the_lnet.ln_state = LNET_STATE_STOPPING;
2081
2082         while (!list_empty(&the_lnet.ln_nets)) {
2083                 /*
2084                  * move the nets to the zombie list to avoid them being
2085                  * picked up for new work. LONET is also included in the
2086                  * Nets that will be moved to the zombie list
2087                  */
2088                 net = list_entry(the_lnet.ln_nets.next,
2089                                  struct lnet_net, net_list);
2090                 list_move(&net->net_list, &the_lnet.ln_net_zombie);
2091         }
2092
2093         /* Drop the cached loopback Net. */
2094         if (the_lnet.ln_loni != NULL) {
2095                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
2096                 the_lnet.ln_loni = NULL;
2097         }
2098         lnet_net_unlock(LNET_LOCK_EX);
2099
2100         /* iterate through the net zombie list and delete each net */
2101         while (!list_empty(&the_lnet.ln_net_zombie)) {
2102                 net = list_entry(the_lnet.ln_net_zombie.next,
2103                                  struct lnet_net, net_list);
2104                 lnet_shutdown_lndnet(net);
2105         }
2106
2107         spin_lock(&the_lnet.ln_msg_resend_lock);
2108         list_splice(&the_lnet.ln_msg_resend, &resend);
2109         spin_unlock(&the_lnet.ln_msg_resend_lock);
2110
2111         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
2112                 list_del_init(&msg->msg_list);
2113                 msg->msg_no_resend = true;
2114                 lnet_finalize(msg, -ECANCELED);
2115         }
2116
2117         lnet_net_lock(LNET_LOCK_EX);
2118         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
2119         lnet_net_unlock(LNET_LOCK_EX);
2120 }
2121
2122 static int
2123 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2124 {
2125         int                     rc = -EINVAL;
2126         struct lnet_tx_queue    *tq;
2127         int                     i;
2128         struct lnet_net         *net = ni->ni_net;
2129
2130         mutex_lock(&the_lnet.ln_lnd_mutex);
2131
2132         if (tun) {
2133                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2134                 ni->ni_lnd_tunables_set = true;
2135         }
2136
2137         rc = (net->net_lnd->lnd_startup)(ni);
2138
2139         mutex_unlock(&the_lnet.ln_lnd_mutex);
2140
2141         if (rc != 0) {
2142                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2143                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2144                 lnet_net_lock(LNET_LOCK_EX);
2145                 net->net_lnd->lnd_refcount--;
2146                 lnet_net_unlock(LNET_LOCK_EX);
2147                 goto failed0;
2148         }
2149
2150         lnet_ni_lock(ni);
2151         ni->ni_state = LNET_NI_STATE_ACTIVE;
2152         lnet_ni_unlock(ni);
2153
2154         /* We keep a reference on the loopback net through the loopback NI */
2155         if (net->net_lnd->lnd_type == LOLND) {
2156                 lnet_ni_addref(ni);
2157                 LASSERT(the_lnet.ln_loni == NULL);
2158                 the_lnet.ln_loni = ni;
2159                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2160                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2161                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2162                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2163                 return 0;
2164         }
2165
2166         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2167             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2168                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2169                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2170                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2171                                         "" : "per-peer ");
2172                 /* shutdown the NI since if we get here then it must've already
2173                  * been started
2174                  */
2175                 lnet_shutdown_lndni(ni);
2176                 return -EINVAL;
2177         }
2178
2179         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2180                 tq->tq_credits_min =
2181                 tq->tq_credits_max =
2182                 tq->tq_credits = lnet_ni_tq_credits(ni);
2183         }
2184
2185         atomic_set(&ni->ni_tx_credits,
2186                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2187         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2188
2189         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2190                 libcfs_nid2str(ni->ni_nid),
2191                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2192                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2193                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2194                 ni->ni_net->net_tunables.lct_peer_timeout);
2195
2196         return 0;
2197 failed0:
2198         lnet_ni_free(ni);
2199         return rc;
2200 }
2201
2202 static int
2203 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2204 {
2205         struct lnet_ni *ni;
2206         struct lnet_net *net_l = NULL;
2207         struct list_head        local_ni_list;
2208         int                     rc;
2209         int                     ni_count = 0;
2210         __u32                   lnd_type;
2211         struct lnet_lnd *lnd;
2212         int                     peer_timeout =
2213                 net->net_tunables.lct_peer_timeout;
2214         int                     maxtxcredits =
2215                 net->net_tunables.lct_max_tx_credits;
2216         int                     peerrtrcredits =
2217                 net->net_tunables.lct_peer_rtr_credits;
2218
2219         INIT_LIST_HEAD(&local_ni_list);
2220
2221         /*
2222          * make sure that this net is unique. If it isn't then
2223          * we are adding interfaces to an already existing network, and
2224          * 'net' is just a convenient way to pass in the list.
2225          * if it is unique we need to find the LND and load it if
2226          * necessary.
2227          */
2228         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2229                 lnd_type = LNET_NETTYP(net->net_id);
2230
2231                 mutex_lock(&the_lnet.ln_lnd_mutex);
2232                 lnd = lnet_find_lnd_by_type(lnd_type);
2233
2234                 if (lnd == NULL) {
2235                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2236                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2237                         mutex_lock(&the_lnet.ln_lnd_mutex);
2238
2239                         lnd = lnet_find_lnd_by_type(lnd_type);
2240                         if (lnd == NULL) {
2241                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2242                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
2243                                 libcfs_lnd2str(lnd_type),
2244                                 libcfs_lnd2modname(lnd_type), rc);
2245 #ifndef HAVE_MODULE_LOADING_SUPPORT
2246                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
2247                                                 "compiled with kernel module "
2248                                                 "loading support.");
2249 #endif
2250                                 rc = -EINVAL;
2251                                 goto failed0;
2252                         }
2253                 }
2254
2255                 lnet_net_lock(LNET_LOCK_EX);
2256                 lnd->lnd_refcount++;
2257                 lnet_net_unlock(LNET_LOCK_EX);
2258
2259                 net->net_lnd = lnd;
2260
2261                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2262
2263                 net_l = net;
2264         }
2265
2266         /*
2267          * net_l: if the network being added is unique then net_l
2268          *        will point to that network
2269          *        if the network being added is not unique then
2270          *        net_l points to the existing network.
2271          *
2272          * When we enter the loop below, we'll pick NIs off he
2273          * network beign added and start them up, then add them to
2274          * a local ni list. Once we've successfully started all
2275          * the NIs then we join the local NI list (of started up
2276          * networks) with the net_l->net_ni_list, which should
2277          * point to the correct network to add the new ni list to
2278          *
2279          * If any of the new NIs fail to start up, then we want to
2280          * iterate through the local ni list, which should include
2281          * any NIs which were successfully started up, and shut
2282          * them down.
2283          *
2284          * After than we want to delete the network being added,
2285          * to avoid a memory leak.
2286          */
2287
2288         /*
2289          * When a network uses TCP bonding then all its interfaces
2290          * must be specified when the network is first defined: the
2291          * TCP bonding code doesn't allow for interfaces to be added
2292          * or removed.
2293          */
2294         if (net_l != net && net_l != NULL && use_tcp_bonding &&
2295             LNET_NETTYP(net_l->net_id) == SOCKLND) {
2296                 rc = -EINVAL;
2297                 goto failed0;
2298         }
2299
2300         while (!list_empty(&net->net_ni_added)) {
2301                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
2302                                 ni_netlist);
2303                 list_del_init(&ni->ni_netlist);
2304
2305                 /* make sure that the the NI we're about to start
2306                  * up is actually unique. if it's not fail. */
2307                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2308                                         ni->ni_interfaces[0])) {
2309                         rc = -EINVAL;
2310                         goto failed1;
2311                 }
2312
2313                 /* adjust the pointer the parent network, just in case it
2314                  * the net is a duplicate */
2315                 ni->ni_net = net_l;
2316
2317                 rc = lnet_startup_lndni(ni, tun);
2318
2319                 LASSERT(ni->ni_net->net_tunables.lct_peer_timeout <= 0 ||
2320                         ni->ni_net->net_lnd->lnd_query != NULL);
2321
2322                 if (rc < 0)
2323                         goto failed1;
2324
2325                 lnet_ni_addref(ni);
2326                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2327
2328                 ni_count++;
2329         }
2330
2331         lnet_net_lock(LNET_LOCK_EX);
2332         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2333         lnet_incr_dlc_seq();
2334         lnet_net_unlock(LNET_LOCK_EX);
2335
2336         /* if the network is not unique then we don't want to keep
2337          * it around after we're done. Free it. Otherwise add that
2338          * net to the global the_lnet.ln_nets */
2339         if (net_l != net && net_l != NULL) {
2340                 /*
2341                  * TODO - note. currently the tunables can not be updated
2342                  * once added
2343                  */
2344                 lnet_net_free(net);
2345         } else {
2346                 net->net_state = LNET_NET_STATE_ACTIVE;
2347                 /*
2348                  * restore tunables after it has been overwitten by the
2349                  * lnd
2350                  */
2351                 if (peer_timeout != -1)
2352                         net->net_tunables.lct_peer_timeout = peer_timeout;
2353                 if (maxtxcredits != -1)
2354                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2355                 if (peerrtrcredits != -1)
2356                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2357
2358                 lnet_net_lock(LNET_LOCK_EX);
2359                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2360                 lnet_net_unlock(LNET_LOCK_EX);
2361         }
2362
2363         return ni_count;
2364
2365 failed1:
2366         /*
2367          * shutdown the new NIs that are being started up
2368          * free the NET being started
2369          */
2370         while (!list_empty(&local_ni_list)) {
2371                 ni = list_entry(local_ni_list.next, struct lnet_ni,
2372                                 ni_netlist);
2373
2374                 lnet_shutdown_lndni(ni);
2375         }
2376
2377 failed0:
2378         lnet_net_free(net);
2379
2380         return rc;
2381 }
2382
2383 static int
2384 lnet_startup_lndnets(struct list_head *netlist)
2385 {
2386         struct lnet_net         *net;
2387         int                     rc;
2388         int                     ni_count = 0;
2389
2390         /*
2391          * Change to running state before bringing up the LNDs. This
2392          * allows lnet_shutdown_lndnets() to assert that we've passed
2393          * through here.
2394          */
2395         lnet_net_lock(LNET_LOCK_EX);
2396         the_lnet.ln_state = LNET_STATE_RUNNING;
2397         lnet_net_unlock(LNET_LOCK_EX);
2398
2399         while (!list_empty(netlist)) {
2400                 net = list_entry(netlist->next, struct lnet_net, net_list);
2401                 list_del_init(&net->net_list);
2402
2403                 rc = lnet_startup_lndnet(net, NULL);
2404
2405                 if (rc < 0)
2406                         goto failed;
2407
2408                 ni_count += rc;
2409         }
2410
2411         return ni_count;
2412 failed:
2413         lnet_shutdown_lndnets();
2414
2415         return rc;
2416 }
2417
2418 /**
2419  * Initialize LNet library.
2420  *
2421  * Automatically called at module loading time. Caller has to call
2422  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2423  * latter returned 0. It must be called exactly once.
2424  *
2425  * \retval 0 on success
2426  * \retval -ve on failures.
2427  */
2428 int lnet_lib_init(void)
2429 {
2430         int rc;
2431
2432         lnet_assert_wire_constants();
2433
2434         /* refer to global cfs_cpt_table for now */
2435         the_lnet.ln_cpt_table   = cfs_cpt_table;
2436         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
2437
2438         LASSERT(the_lnet.ln_cpt_number > 0);
2439         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2440                 /* we are under risk of consuming all lh_cookie */
2441                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2442                        "please change setting of CPT-table and retry\n",
2443                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2444                 return -E2BIG;
2445         }
2446
2447         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2448                 the_lnet.ln_cpt_bits++;
2449
2450         rc = lnet_create_locks();
2451         if (rc != 0) {
2452                 CERROR("Can't create LNet global locks: %d\n", rc);
2453                 return rc;
2454         }
2455
2456         the_lnet.ln_refcount = 0;
2457         LNetInvalidateEQHandle(&the_lnet.ln_rc_eqh);
2458         INIT_LIST_HEAD(&the_lnet.ln_lnds);
2459         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2460         INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
2461         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2462         INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
2463
2464         /* The hash table size is the number of bits it takes to express the set
2465          * ln_num_routes, minus 1 (better to under estimate than over so we
2466          * don't waste memory). */
2467         if (rnet_htable_size <= 0)
2468                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2469         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2470                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2471         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2472                                            order_base_2(rnet_htable_size) - 1);
2473
2474         /* All LNDs apart from the LOLND are in separate modules.  They
2475          * register themselves when their module loads, and unregister
2476          * themselves when their module is unloaded. */
2477         lnet_register_lnd(&the_lolnd);
2478         return 0;
2479 }
2480
2481 /**
2482  * Finalize LNet library.
2483  *
2484  * \pre lnet_lib_init() called with success.
2485  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2486  */
2487 void lnet_lib_exit(void)
2488 {
2489         LASSERT(the_lnet.ln_refcount == 0);
2490
2491         while (!list_empty(&the_lnet.ln_lnds))
2492                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
2493                                                struct lnet_lnd, lnd_list));
2494         lnet_destroy_locks();
2495 }
2496
2497 /**
2498  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2499  *
2500  * Users must call this function at least once before any other functions.
2501  * For each successful call there must be a corresponding call to
2502  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2503  * ignored.
2504  *
2505  * The PID used by LNet may be different from the one requested.
2506  * See LNetGetId().
2507  *
2508  * \param requested_pid PID requested by the caller.
2509  *
2510  * \return >= 0 on success, and < 0 error code on failures.
2511  */
2512 int
2513 LNetNIInit(lnet_pid_t requested_pid)
2514 {
2515         int                     im_a_router = 0;
2516         int                     rc;
2517         int                     ni_count;
2518         struct lnet_ping_buffer *pbuf;
2519         struct lnet_handle_md   ping_mdh;
2520         struct list_head        net_head;
2521         struct lnet_net         *net;
2522
2523         INIT_LIST_HEAD(&net_head);
2524
2525         mutex_lock(&the_lnet.ln_api_mutex);
2526
2527         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
2528
2529         if (the_lnet.ln_refcount > 0) {
2530                 rc = the_lnet.ln_refcount++;
2531                 mutex_unlock(&the_lnet.ln_api_mutex);
2532                 return rc;
2533         }
2534
2535         rc = lnet_prepare(requested_pid);
2536         if (rc != 0) {
2537                 mutex_unlock(&the_lnet.ln_api_mutex);
2538                 return rc;
2539         }
2540
2541         /* create a network for Loopback network */
2542         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
2543         if (net == NULL) {
2544                 rc = -ENOMEM;
2545                 goto err_empty_list;
2546         }
2547
2548         /* Add in the loopback NI */
2549         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
2550                 rc = -ENOMEM;
2551                 goto err_empty_list;
2552         }
2553
2554         /* If LNet is being initialized via DLC it is possible
2555          * that the user requests not to load module parameters (ones which
2556          * are supported by DLC) on initialization.  Therefore, make sure not
2557          * to load networks, routes and forwarding from module parameters
2558          * in this case.  On cleanup in case of failure only clean up
2559          * routes if it has been loaded */
2560         if (!the_lnet.ln_nis_from_mod_params) {
2561                 rc = lnet_parse_networks(&net_head, lnet_get_networks(),
2562                                          use_tcp_bonding);
2563                 if (rc < 0)
2564                         goto err_empty_list;
2565         }
2566
2567         ni_count = lnet_startup_lndnets(&net_head);
2568         if (ni_count < 0) {
2569                 rc = ni_count;
2570                 goto err_empty_list;
2571         }
2572
2573         if (!the_lnet.ln_nis_from_mod_params) {
2574                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
2575                 if (rc != 0)
2576                         goto err_shutdown_lndnis;
2577
2578                 rc = lnet_check_routes();
2579                 if (rc != 0)
2580                         goto err_destroy_routes;
2581
2582                 rc = lnet_rtrpools_alloc(im_a_router);
2583                 if (rc != 0)
2584                         goto err_destroy_routes;
2585         }
2586
2587         rc = lnet_acceptor_start();
2588         if (rc != 0)
2589                 goto err_destroy_routes;
2590
2591         the_lnet.ln_refcount = 1;
2592         /* Now I may use my own API functions... */
2593
2594         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
2595         if (rc != 0)
2596                 goto err_acceptor_stop;
2597
2598         lnet_ping_target_update(pbuf, ping_mdh);
2599
2600         rc = LNetEQAlloc(0, lnet_mt_event_handler, &the_lnet.ln_mt_eqh);
2601         if (rc != 0) {
2602                 CERROR("Can't allocate monitor thread EQ: %d\n", rc);
2603                 goto err_stop_ping;
2604         }
2605
2606         rc = lnet_monitor_thr_start();
2607         if (rc != 0)
2608                 goto err_stop_ping;
2609
2610         rc = lnet_push_target_init();
2611         if (rc != 0)
2612                 goto err_stop_monitor_thr;
2613
2614         rc = lnet_peer_discovery_start();
2615         if (rc != 0)
2616                 goto err_destroy_push_target;
2617
2618         lnet_fault_init();
2619         lnet_router_debugfs_init();
2620
2621         mutex_unlock(&the_lnet.ln_api_mutex);
2622
2623         return 0;
2624
2625 err_destroy_push_target:
2626         lnet_push_target_fini();
2627 err_stop_monitor_thr:
2628         lnet_monitor_thr_stop();
2629 err_stop_ping:
2630         lnet_ping_target_fini();
2631 err_acceptor_stop:
2632         the_lnet.ln_refcount = 0;
2633         lnet_acceptor_stop();
2634 err_destroy_routes:
2635         if (!the_lnet.ln_nis_from_mod_params)
2636                 lnet_destroy_routes();
2637 err_shutdown_lndnis:
2638         lnet_shutdown_lndnets();
2639 err_empty_list:
2640         lnet_unprepare();
2641         LASSERT(rc < 0);
2642         mutex_unlock(&the_lnet.ln_api_mutex);
2643         while (!list_empty(&net_head)) {
2644                 struct lnet_net *net;
2645
2646                 net = list_entry(net_head.next, struct lnet_net, net_list);
2647                 list_del_init(&net->net_list);
2648                 lnet_net_free(net);
2649         }
2650         return rc;
2651 }
2652 EXPORT_SYMBOL(LNetNIInit);
2653
2654 /**
2655  * Stop LNet interfaces, routing, and forwarding.
2656  *
2657  * Users must call this function once for each successful call to LNetNIInit().
2658  * Once the LNetNIFini() operation has been started, the results of pending
2659  * API operations are undefined.
2660  *
2661  * \return always 0 for current implementation.
2662  */
2663 int
2664 LNetNIFini()
2665 {
2666         mutex_lock(&the_lnet.ln_api_mutex);
2667
2668         LASSERT(the_lnet.ln_refcount > 0);
2669
2670         if (the_lnet.ln_refcount != 1) {
2671                 the_lnet.ln_refcount--;
2672         } else {
2673                 LASSERT(!the_lnet.ln_niinit_self);
2674
2675                 lnet_fault_fini();
2676
2677                 lnet_router_debugfs_fini();
2678                 lnet_peer_discovery_stop();
2679                 lnet_push_target_fini();
2680                 lnet_monitor_thr_stop();
2681                 lnet_ping_target_fini();
2682
2683                 /* Teardown fns that use my own API functions BEFORE here */
2684                 the_lnet.ln_refcount = 0;
2685
2686                 lnet_acceptor_stop();
2687                 lnet_destroy_routes();
2688                 lnet_shutdown_lndnets();
2689                 lnet_unprepare();
2690         }
2691
2692         mutex_unlock(&the_lnet.ln_api_mutex);
2693         return 0;
2694 }
2695 EXPORT_SYMBOL(LNetNIFini);
2696
2697 /**
2698  * Grabs the ni data from the ni structure and fills the out
2699  * parameters
2700  *
2701  * \param[in] ni network        interface structure
2702  * \param[out] cfg_ni           NI config information
2703  * \param[out] tun              network and LND tunables
2704  */
2705 static void
2706 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
2707                    struct lnet_ioctl_config_lnd_tunables *tun,
2708                    struct lnet_ioctl_element_stats *stats,
2709                    __u32 tun_size)
2710 {
2711         size_t min_size = 0;
2712         int i;
2713
2714         if (!ni || !cfg_ni || !tun)
2715                 return;
2716
2717         if (ni->ni_interfaces[0] != NULL) {
2718                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2719                         if (ni->ni_interfaces[i] != NULL) {
2720                                 strncpy(cfg_ni->lic_ni_intf[i],
2721                                         ni->ni_interfaces[i],
2722                                         sizeof(cfg_ni->lic_ni_intf[i]));
2723                         }
2724                 }
2725         }
2726
2727         cfg_ni->lic_nid = ni->ni_nid;
2728         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2729                 cfg_ni->lic_status = LNET_NI_STATUS_UP;
2730         else
2731                 cfg_ni->lic_status = ni->ni_status->ns_status;
2732         cfg_ni->lic_tcp_bonding = use_tcp_bonding;
2733         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
2734
2735         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
2736
2737         if (stats) {
2738                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
2739                                                        LNET_STATS_TYPE_SEND);
2740                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
2741                                                        LNET_STATS_TYPE_RECV);
2742                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
2743                                                        LNET_STATS_TYPE_DROP);
2744         }
2745
2746         /*
2747          * tun->lt_tun will always be present, but in order to be
2748          * backwards compatible, we need to deal with the cases when
2749          * tun->lt_tun is smaller than what the kernel has, because it
2750          * comes from an older version of a userspace program, then we'll
2751          * need to copy as much information as we have available space.
2752          */
2753         min_size = tun_size - sizeof(tun->lt_cmn);
2754         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
2755
2756         /* copy over the cpts */
2757         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
2758             ni->ni_cpts == NULL)  {
2759                 for (i = 0; i < ni->ni_ncpts; i++)
2760                         cfg_ni->lic_cpts[i] = i;
2761         } else {
2762                 for (i = 0;
2763                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
2764                      i < LNET_MAX_SHOW_NUM_CPT;
2765                      i++)
2766                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
2767         }
2768         cfg_ni->lic_ncpts = ni->ni_ncpts;
2769 }
2770
2771 /**
2772  * NOTE: This is a legacy function left in the code to be backwards
2773  * compatible with older userspace programs. It should eventually be
2774  * removed.
2775  *
2776  * Grabs the ni data from the ni structure and fills the out
2777  * parameters
2778  *
2779  * \param[in] ni network        interface structure
2780  * \param[out] config           config information
2781  */
2782 static void
2783 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
2784                          struct lnet_ioctl_config_data *config)
2785 {
2786         struct lnet_ioctl_net_config *net_config;
2787         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
2788         size_t min_size, tunable_size = 0;
2789         int i;
2790
2791         if (!ni || !config)
2792                 return;
2793
2794         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
2795         if (!net_config)
2796                 return;
2797
2798         BUILD_BUG_ON(ARRAY_SIZE(ni->ni_interfaces) !=
2799                      ARRAY_SIZE(net_config->ni_interfaces));
2800
2801         for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2802                 if (!ni->ni_interfaces[i])
2803                         break;
2804
2805                 strncpy(net_config->ni_interfaces[i],
2806                         ni->ni_interfaces[i],
2807                         sizeof(net_config->ni_interfaces[i]));
2808         }
2809
2810         config->cfg_nid = ni->ni_nid;
2811         config->cfg_config_u.cfg_net.net_peer_timeout =
2812                 ni->ni_net->net_tunables.lct_peer_timeout;
2813         config->cfg_config_u.cfg_net.net_max_tx_credits =
2814                 ni->ni_net->net_tunables.lct_max_tx_credits;
2815         config->cfg_config_u.cfg_net.net_peer_tx_credits =
2816                 ni->ni_net->net_tunables.lct_peer_tx_credits;
2817         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
2818                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
2819
2820         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2821                 net_config->ni_status = LNET_NI_STATUS_UP;
2822         else
2823                 net_config->ni_status = ni->ni_status->ns_status;
2824
2825         if (ni->ni_cpts) {
2826                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
2827
2828                 for (i = 0; i < num_cpts; i++)
2829                         net_config->ni_cpts[i] = ni->ni_cpts[i];
2830
2831                 config->cfg_ncpts = num_cpts;
2832         }
2833
2834         /*
2835          * See if user land tools sent in a newer and larger version
2836          * of struct lnet_tunables than what the kernel uses.
2837          */
2838         min_size = sizeof(*config) + sizeof(*net_config);
2839
2840         if (config->cfg_hdr.ioc_len > min_size)
2841                 tunable_size = config->cfg_hdr.ioc_len - min_size;
2842
2843         /* Don't copy too much data to user space */
2844         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
2845         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
2846
2847         if (lnd_cfg && min_size) {
2848                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
2849                 config->cfg_config_u.cfg_net.net_interface_count = 1;
2850
2851                 /* Tell user land that kernel side has less data */
2852                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
2853                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
2854                         config->cfg_hdr.ioc_len -= min_size;
2855                 }
2856         }
2857 }
2858
2859 struct lnet_ni *
2860 lnet_get_ni_idx_locked(int idx)
2861 {
2862         struct lnet_ni          *ni;
2863         struct lnet_net         *net;
2864
2865         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2866                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2867                         if (idx-- == 0)
2868                                 return ni;
2869                 }
2870         }
2871
2872         return NULL;
2873 }
2874
2875 struct lnet_ni *
2876 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
2877 {
2878         struct lnet_ni          *ni;
2879         struct lnet_net         *net = mynet;
2880
2881         /*
2882          * It is possible that the net has been cleaned out while there is
2883          * a message being sent. This function accessed the net without
2884          * checking if the list is empty
2885          */
2886         if (prev == NULL) {
2887                 if (net == NULL)
2888                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
2889                                         net_list);
2890                 if (list_empty(&net->net_ni_list))
2891                         return NULL;
2892                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2893                                 ni_netlist);
2894
2895                 return ni;
2896         }
2897
2898         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
2899                 /* if you reached the end of the ni list and the net is
2900                  * specified, then there are no more nis in that net */
2901                 if (net != NULL)
2902                         return NULL;
2903
2904                 /* we reached the end of this net ni list. move to the
2905                  * next net */
2906                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
2907                         /* no more nets and no more NIs. */
2908                         return NULL;
2909
2910                 /* get the next net */
2911                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
2912                                  net_list);
2913                 if (list_empty(&net->net_ni_list))
2914                         return NULL;
2915                 /* get the ni on it */
2916                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2917                                 ni_netlist);
2918
2919                 return ni;
2920         }
2921
2922         if (list_empty(&prev->ni_netlist))
2923                 return NULL;
2924
2925         /* there are more nis left */
2926         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
2927
2928         return ni;
2929 }
2930
2931 int
2932 lnet_get_net_config(struct lnet_ioctl_config_data *config)
2933 {
2934         struct lnet_ni *ni;
2935         int cpt;
2936         int rc = -ENOENT;
2937         int idx = config->cfg_count;
2938
2939         cpt = lnet_net_lock_current();
2940
2941         ni = lnet_get_ni_idx_locked(idx);
2942
2943         if (ni != NULL) {
2944                 rc = 0;
2945                 lnet_ni_lock(ni);
2946                 lnet_fill_ni_info_legacy(ni, config);
2947                 lnet_ni_unlock(ni);
2948         }
2949
2950         lnet_net_unlock(cpt);
2951         return rc;
2952 }
2953
2954 int
2955 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
2956                    struct lnet_ioctl_config_lnd_tunables *tun,
2957                    struct lnet_ioctl_element_stats *stats,
2958                    __u32 tun_size)
2959 {
2960         struct lnet_ni          *ni;
2961         int                     cpt;
2962         int                     rc = -ENOENT;
2963
2964         if (!cfg_ni || !tun || !stats)
2965                 return -EINVAL;
2966
2967         cpt = lnet_net_lock_current();
2968
2969         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
2970
2971         if (ni) {
2972                 rc = 0;
2973                 lnet_ni_lock(ni);
2974                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
2975                 lnet_ni_unlock(ni);
2976         }
2977
2978         lnet_net_unlock(cpt);
2979         return rc;
2980 }
2981
2982 int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
2983 {
2984         struct lnet_ni *ni;
2985         int cpt;
2986         int rc = -ENOENT;
2987
2988         if (!msg_stats)
2989                 return -EINVAL;
2990
2991         cpt = lnet_net_lock_current();
2992
2993         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
2994
2995         if (ni) {
2996                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
2997                 rc = 0;
2998         }
2999
3000         lnet_net_unlock(cpt);
3001
3002         return rc;
3003 }
3004
3005 static int lnet_add_net_common(struct lnet_net *net,
3006                                struct lnet_ioctl_config_lnd_tunables *tun)
3007 {
3008         __u32                   net_id;
3009         struct lnet_ping_buffer *pbuf;
3010         struct lnet_handle_md   ping_mdh;
3011         int                     rc;
3012         struct lnet_remotenet *rnet;
3013         int                     net_ni_count;
3014         int                     num_acceptor_nets;
3015
3016         lnet_net_lock(LNET_LOCK_EX);
3017         rnet = lnet_find_rnet_locked(net->net_id);
3018         lnet_net_unlock(LNET_LOCK_EX);
3019         /*
3020          * make sure that the net added doesn't invalidate the current
3021          * configuration LNet is keeping
3022          */
3023         if (rnet) {
3024                 CERROR("Adding net %s will invalidate routing configuration\n",
3025                        libcfs_net2str(net->net_id));
3026                 lnet_net_free(net);
3027                 return -EUSERS;
3028         }
3029
3030         /*
3031          * make sure you calculate the correct number of slots in the ping
3032          * buffer. Since the ping info is a flattened list of all the NIs,
3033          * we should allocate enough slots to accomodate the number of NIs
3034          * which will be added.
3035          *
3036          * since ni hasn't been configured yet, use
3037          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
3038          */
3039         net_ni_count = lnet_get_net_ni_count_pre(net);
3040
3041         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3042                                     net_ni_count + lnet_get_ni_count(),
3043                                     false);
3044         if (rc < 0) {
3045                 lnet_net_free(net);
3046                 return rc;
3047         }
3048
3049         if (tun)
3050                 memcpy(&net->net_tunables,
3051                        &tun->lt_cmn, sizeof(net->net_tunables));
3052         else
3053                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
3054
3055         /*
3056          * before starting this network get a count of the current TCP
3057          * networks which require the acceptor thread running. If that
3058          * count is == 0 before we start up this network, then we'd want to
3059          * start up the acceptor thread after starting up this network
3060          */
3061         num_acceptor_nets = lnet_count_acceptor_nets();
3062
3063         net_id = net->net_id;
3064
3065         rc = lnet_startup_lndnet(net,
3066                                  (tun) ? &tun->lt_tun : NULL);
3067         if (rc < 0)
3068                 goto failed;
3069
3070         lnet_net_lock(LNET_LOCK_EX);
3071         net = lnet_get_net_locked(net_id);
3072         lnet_net_unlock(LNET_LOCK_EX);
3073
3074         LASSERT(net);
3075
3076         /*
3077          * Start the acceptor thread if this is the first network
3078          * being added that requires the thread.
3079          */
3080         if (net->net_lnd->lnd_accept && num_acceptor_nets == 0) {
3081                 rc = lnet_acceptor_start();
3082                 if (rc < 0) {
3083                         /* shutdown the net that we just started */
3084                         CERROR("Failed to start up acceptor thread\n");
3085                         lnet_shutdown_lndnet(net);
3086                         goto failed;
3087                 }
3088         }
3089
3090         lnet_net_lock(LNET_LOCK_EX);
3091         lnet_peer_net_added(net);
3092         lnet_net_unlock(LNET_LOCK_EX);
3093
3094         lnet_ping_target_update(pbuf, ping_mdh);
3095
3096         return 0;
3097
3098 failed:
3099         lnet_ping_md_unlink(pbuf, &ping_mdh);
3100         lnet_ping_buffer_decref(pbuf);
3101         return rc;
3102 }
3103
3104 static int lnet_handle_legacy_ip2nets(char *ip2nets,
3105                                       struct lnet_ioctl_config_lnd_tunables *tun)
3106 {
3107         struct lnet_net *net;
3108         char *nets;
3109         int rc;
3110         struct list_head net_head;
3111
3112         INIT_LIST_HEAD(&net_head);
3113
3114         rc = lnet_parse_ip2nets(&nets, ip2nets);
3115         if (rc < 0)
3116                 return rc;
3117
3118         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3119         if (rc < 0)
3120                 return rc;
3121
3122         mutex_lock(&the_lnet.ln_api_mutex);
3123         while (!list_empty(&net_head)) {
3124                 net = list_entry(net_head.next, struct lnet_net, net_list);
3125                 list_del_init(&net->net_list);
3126                 rc = lnet_add_net_common(net, tun);
3127                 if (rc < 0)
3128                         goto out;
3129         }
3130
3131 out:
3132         mutex_unlock(&the_lnet.ln_api_mutex);
3133
3134         while (!list_empty(&net_head)) {
3135                 net = list_entry(net_head.next, struct lnet_net, net_list);
3136                 list_del_init(&net->net_list);
3137                 lnet_net_free(net);
3138         }
3139         return rc;
3140 }
3141
3142 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
3143 {
3144         struct lnet_net *net;
3145         struct lnet_ni *ni;
3146         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3147         int rc, i;
3148         __u32 net_id, lnd_type;
3149
3150         /* get the tunables if they are available */
3151         if (conf->lic_cfg_hdr.ioc_len >=
3152             sizeof(*conf) + sizeof(*tun))
3153                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3154                         conf->lic_bulk;
3155
3156         /* handle legacy ip2nets from DLC */
3157         if (conf->lic_legacy_ip2nets[0] != '\0')
3158                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
3159                                                   tun);
3160
3161         net_id = LNET_NIDNET(conf->lic_nid);
3162         lnd_type = LNET_NETTYP(net_id);
3163
3164         if (!libcfs_isknown_lnd(lnd_type)) {
3165                 CERROR("No valid net and lnd information provided\n");
3166                 return -EINVAL;
3167         }
3168
3169         net = lnet_net_alloc(net_id, NULL);
3170         if (!net)
3171                 return -ENOMEM;
3172
3173         for (i = 0; i < conf->lic_ncpts; i++) {
3174                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
3175                         return -EINVAL;
3176         }
3177
3178         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
3179                                        conf->lic_ni_intf[0]);
3180         if (!ni)
3181                 return -ENOMEM;
3182
3183         mutex_lock(&the_lnet.ln_api_mutex);
3184
3185         rc = lnet_add_net_common(net, tun);
3186
3187         mutex_unlock(&the_lnet.ln_api_mutex);
3188
3189         return rc;
3190 }
3191
3192 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
3193 {
3194         struct lnet_net  *net;
3195         struct lnet_ni *ni;
3196         __u32 net_id = LNET_NIDNET(conf->lic_nid);
3197         struct lnet_ping_buffer *pbuf;
3198         struct lnet_handle_md  ping_mdh;
3199         int               rc;
3200         int               net_count;
3201         __u32             addr;
3202
3203         /* don't allow userspace to shutdown the LOLND */
3204         if (LNET_NETTYP(net_id) == LOLND)
3205                 return -EINVAL;
3206
3207         mutex_lock(&the_lnet.ln_api_mutex);
3208
3209         lnet_net_lock(0);
3210
3211         net = lnet_get_net_locked(net_id);
3212         if (!net) {
3213                 CERROR("net %s not found\n",
3214                        libcfs_net2str(net_id));
3215                 rc = -ENOENT;
3216                 goto unlock_net;
3217         }
3218
3219         addr = LNET_NIDADDR(conf->lic_nid);
3220         if (addr == 0) {
3221                 /* remove the entire net */
3222                 net_count = lnet_get_net_ni_count_locked(net);
3223
3224                 lnet_net_unlock(0);
3225
3226                 /* create and link a new ping info, before removing the old one */
3227                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3228                                         lnet_get_ni_count() - net_count,
3229                                         false);
3230                 if (rc != 0)
3231                         goto unlock_api_mutex;
3232
3233                 lnet_shutdown_lndnet(net);
3234
3235                 if (lnet_count_acceptor_nets() == 0)
3236                         lnet_acceptor_stop();
3237
3238                 lnet_ping_target_update(pbuf, ping_mdh);
3239
3240                 goto unlock_api_mutex;
3241         }
3242
3243         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
3244         if (!ni) {
3245                 CERROR("nid %s not found\n",
3246                        libcfs_nid2str(conf->lic_nid));
3247                 rc = -ENOENT;
3248                 goto unlock_net;
3249         }
3250
3251         net_count = lnet_get_net_ni_count_locked(net);
3252
3253         lnet_net_unlock(0);
3254
3255         /* create and link a new ping info, before removing the old one */
3256         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3257                                   lnet_get_ni_count() - 1, false);
3258         if (rc != 0)
3259                 goto unlock_api_mutex;
3260
3261         lnet_shutdown_lndni(ni);
3262
3263         if (lnet_count_acceptor_nets() == 0)
3264                 lnet_acceptor_stop();
3265
3266         lnet_ping_target_update(pbuf, ping_mdh);
3267
3268         /* check if the net is empty and remove it if it is */
3269         if (net_count == 1)
3270                 lnet_shutdown_lndnet(net);
3271
3272         goto unlock_api_mutex;
3273
3274 unlock_net:
3275         lnet_net_unlock(0);
3276 unlock_api_mutex:
3277         mutex_unlock(&the_lnet.ln_api_mutex);
3278
3279         return rc;
3280 }
3281
3282 /*
3283  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
3284  * They are only expected to be called for unique networks.
3285  * That can be as a result of older DLC library
3286  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
3287  */
3288 int
3289 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
3290 {
3291         struct lnet_net         *net;
3292         struct list_head        net_head;
3293         int                     rc;
3294         struct lnet_ioctl_config_lnd_tunables tun;
3295         char *nets = conf->cfg_config_u.cfg_net.net_intf;
3296
3297         INIT_LIST_HEAD(&net_head);
3298
3299         /* Create a net/ni structures for the network string */
3300         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3301         if (rc <= 0)
3302                 return rc == 0 ? -EINVAL : rc;
3303
3304         mutex_lock(&the_lnet.ln_api_mutex);
3305
3306         if (rc > 1) {
3307                 rc = -EINVAL; /* only add one network per call */
3308                 goto out_unlock_clean;
3309         }
3310
3311         net = list_entry(net_head.next, struct lnet_net, net_list);
3312         list_del_init(&net->net_list);
3313
3314         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
3315
3316         memset(&tun, 0, sizeof(tun));
3317
3318         tun.lt_cmn.lct_peer_timeout =
3319           conf->cfg_config_u.cfg_net.net_peer_timeout;
3320         tun.lt_cmn.lct_peer_tx_credits =
3321           conf->cfg_config_u.cfg_net.net_peer_tx_credits;
3322         tun.lt_cmn.lct_peer_rtr_credits =
3323           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
3324         tun.lt_cmn.lct_max_tx_credits =
3325           conf->cfg_config_u.cfg_net.net_max_tx_credits;
3326
3327         rc = lnet_add_net_common(net, &tun);
3328
3329 out_unlock_clean:
3330         mutex_unlock(&the_lnet.ln_api_mutex);
3331         while (!list_empty(&net_head)) {
3332                 /* net_head list is empty in success case */
3333                 net = list_entry(net_head.next, struct lnet_net, net_list);
3334                 list_del_init(&net->net_list);
3335                 lnet_net_free(net);
3336         }
3337         return rc;
3338 }
3339
3340 int
3341 lnet_dyn_del_net(__u32 net_id)
3342 {
3343         struct lnet_net  *net;
3344         struct lnet_ping_buffer *pbuf;
3345         struct lnet_handle_md ping_mdh;
3346         int               rc;
3347         int               net_ni_count;
3348
3349         /* don't allow userspace to shutdown the LOLND */
3350         if (LNET_NETTYP(net_id) == LOLND)
3351                 return -EINVAL;
3352
3353         mutex_lock(&the_lnet.ln_api_mutex);
3354
3355         lnet_net_lock(0);
3356
3357         net = lnet_get_net_locked(net_id);
3358         if (net == NULL) {
3359                 lnet_net_unlock(0);
3360                 rc = -EINVAL;
3361                 goto out;
3362         }
3363
3364         net_ni_count = lnet_get_net_ni_count_locked(net);
3365
3366         lnet_net_unlock(0);
3367
3368         /* create and link a new ping info, before removing the old one */
3369         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3370                                     lnet_get_ni_count() - net_ni_count, false);
3371         if (rc != 0)
3372                 goto out;
3373
3374         lnet_shutdown_lndnet(net);
3375
3376         if (lnet_count_acceptor_nets() == 0)
3377                 lnet_acceptor_stop();
3378
3379         lnet_ping_target_update(pbuf, ping_mdh);
3380
3381 out:
3382         mutex_unlock(&the_lnet.ln_api_mutex);
3383
3384         return rc;
3385 }
3386
3387 void lnet_incr_dlc_seq(void)
3388 {
3389         atomic_inc(&lnet_dlc_seq_no);
3390 }
3391
3392 __u32 lnet_get_dlc_seq_locked(void)
3393 {
3394         return atomic_read(&lnet_dlc_seq_no);
3395 }
3396
3397 static void
3398 lnet_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3399 {
3400         struct lnet_net *net;
3401         struct lnet_ni *ni;
3402
3403         lnet_net_lock(LNET_LOCK_EX);
3404         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3405                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3406                         if (ni->ni_nid == nid || all) {
3407                                 atomic_set(&ni->ni_healthv, value);
3408                                 if (list_empty(&ni->ni_recovery) &&
3409                                     value < LNET_MAX_HEALTH_VALUE) {
3410                                         CERROR("manually adding local NI %s to recovery\n",
3411                                                libcfs_nid2str(ni->ni_nid));
3412                                         list_add_tail(&ni->ni_recovery,
3413                                                       &the_lnet.ln_mt_localNIRecovq);
3414                                         lnet_ni_addref_locked(ni, 0);
3415                                 }
3416                                 if (!all) {
3417                                         lnet_net_unlock(LNET_LOCK_EX);
3418                                         return;
3419                                 }
3420                         }
3421                 }
3422         }
3423         lnet_net_unlock(LNET_LOCK_EX);
3424 }
3425
3426 static int
3427 lnet_get_local_ni_hstats(struct lnet_ioctl_local_ni_hstats *stats)
3428 {
3429         int cpt, rc = 0;
3430         struct lnet_ni *ni;
3431         lnet_nid_t nid = stats->hlni_nid;
3432
3433         cpt = lnet_net_lock_current();
3434         ni = lnet_nid2ni_locked(nid, cpt);
3435
3436         if (!ni) {
3437                 rc = -ENOENT;
3438                 goto unlock;
3439         }
3440
3441         stats->hlni_local_interrupt = atomic_read(&ni->ni_hstats.hlt_local_interrupt);
3442         stats->hlni_local_dropped = atomic_read(&ni->ni_hstats.hlt_local_dropped);
3443         stats->hlni_local_aborted = atomic_read(&ni->ni_hstats.hlt_local_aborted);
3444         stats->hlni_local_no_route = atomic_read(&ni->ni_hstats.hlt_local_no_route);
3445         stats->hlni_local_timeout = atomic_read(&ni->ni_hstats.hlt_local_timeout);
3446         stats->hlni_local_error = atomic_read(&ni->ni_hstats.hlt_local_error);
3447         stats->hlni_health_value = atomic_read(&ni->ni_healthv);
3448
3449 unlock:
3450         lnet_net_unlock(cpt);
3451
3452         return rc;
3453 }
3454
3455 static int
3456 lnet_get_local_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3457 {
3458         struct lnet_ni *ni;
3459         int i = 0;
3460
3461         lnet_net_lock(LNET_LOCK_EX);
3462         list_for_each_entry(ni, &the_lnet.ln_mt_localNIRecovq, ni_recovery) {
3463                 list->rlst_nid_array[i] = ni->ni_nid;
3464                 i++;
3465                 if (i >= LNET_MAX_SHOW_NUM_NID)
3466                         break;
3467         }
3468         lnet_net_unlock(LNET_LOCK_EX);
3469         list->rlst_num_nids = i;
3470
3471         return 0;
3472 }
3473
3474 static int
3475 lnet_get_peer_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3476 {
3477         struct lnet_peer_ni *lpni;
3478         int i = 0;
3479
3480         lnet_net_lock(LNET_LOCK_EX);
3481         list_for_each_entry(lpni, &the_lnet.ln_mt_peerNIRecovq, lpni_recovery) {
3482                 list->rlst_nid_array[i] = lpni->lpni_nid;
3483                 i++;
3484                 if (i >= LNET_MAX_SHOW_NUM_NID)
3485                         break;
3486         }
3487         lnet_net_unlock(LNET_LOCK_EX);
3488         list->rlst_num_nids = i;
3489
3490         return 0;
3491 }
3492
3493 /**
3494  * LNet ioctl handler.
3495  *
3496  */
3497 int
3498 LNetCtl(unsigned int cmd, void *arg)
3499 {
3500         struct libcfs_ioctl_data *data = arg;
3501         struct lnet_ioctl_config_data *config;
3502         struct lnet_process_id    id = {0};
3503         struct lnet_ni           *ni;
3504         int                       rc;
3505
3506         BUILD_BUG_ON(sizeof(struct lnet_ioctl_net_config) +
3507                      sizeof(struct lnet_ioctl_config_data) > LIBCFS_IOC_DATA_MAX);
3508
3509         switch (cmd) {
3510         case IOC_LIBCFS_GET_NI:
3511                 rc = LNetGetId(data->ioc_count, &id);
3512                 data->ioc_nid = id.nid;
3513                 return rc;
3514
3515         case IOC_LIBCFS_FAIL_NID:
3516                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
3517
3518         case IOC_LIBCFS_ADD_ROUTE:
3519                 config = arg;
3520
3521                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3522                         return -EINVAL;
3523
3524                 mutex_lock(&the_lnet.ln_api_mutex);
3525                 rc = lnet_add_route(config->cfg_net,
3526                                     config->cfg_config_u.cfg_route.rtr_hop,
3527                                     config->cfg_nid,
3528                                     config->cfg_config_u.cfg_route.
3529                                         rtr_priority);
3530                 if (rc == 0) {
3531                         rc = lnet_check_routes();
3532                         if (rc != 0)
3533                                 lnet_del_route(config->cfg_net,
3534                                                config->cfg_nid);
3535                 }
3536                 mutex_unlock(&the_lnet.ln_api_mutex);
3537                 return rc;
3538
3539         case IOC_LIBCFS_DEL_ROUTE:
3540                 config = arg;
3541
3542                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3543                         return -EINVAL;
3544
3545                 mutex_lock(&the_lnet.ln_api_mutex);
3546                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
3547                 mutex_unlock(&the_lnet.ln_api_mutex);
3548                 return rc;
3549
3550         case IOC_LIBCFS_GET_ROUTE:
3551                 config = arg;
3552
3553                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3554                         return -EINVAL;
3555
3556                 mutex_lock(&the_lnet.ln_api_mutex);
3557                 rc = lnet_get_route(config->cfg_count,
3558                                     &config->cfg_net,
3559                                     &config->cfg_config_u.cfg_route.rtr_hop,
3560                                     &config->cfg_nid,
3561                                     &config->cfg_config_u.cfg_route.rtr_flags,
3562                                     &config->cfg_config_u.cfg_route.
3563                                         rtr_priority);
3564                 mutex_unlock(&the_lnet.ln_api_mutex);
3565                 return rc;
3566
3567         case IOC_LIBCFS_GET_LOCAL_NI: {
3568                 struct lnet_ioctl_config_ni *cfg_ni;
3569                 struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3570                 struct lnet_ioctl_element_stats *stats;
3571                 __u32 tun_size;
3572
3573                 cfg_ni = arg;
3574
3575                 /* get the tunables if they are available */
3576                 if (cfg_ni->lic_cfg_hdr.ioc_len <
3577                     sizeof(*cfg_ni) + sizeof(*stats) + sizeof(*tun))
3578                         return -EINVAL;
3579
3580                 stats = (struct lnet_ioctl_element_stats *)
3581                         cfg_ni->lic_bulk;
3582                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3583                                 (cfg_ni->lic_bulk + sizeof(*stats));
3584
3585                 tun_size = cfg_ni->lic_cfg_hdr.ioc_len - sizeof(*cfg_ni) -
3586                         sizeof(*stats);
3587
3588                 mutex_lock(&the_lnet.ln_api_mutex);
3589                 rc = lnet_get_ni_config(cfg_ni, tun, stats, tun_size);
3590                 mutex_unlock(&the_lnet.ln_api_mutex);
3591                 return rc;
3592         }
3593
3594         case IOC_LIBCFS_GET_LOCAL_NI_MSG_STATS: {
3595                 struct lnet_ioctl_element_msg_stats *msg_stats = arg;
3596
3597                 if (msg_stats->im_hdr.ioc_len != sizeof(*msg_stats))
3598                         return -EINVAL;
3599
3600                 mutex_lock(&the_lnet.ln_api_mutex);
3601                 rc = lnet_get_ni_stats(msg_stats);
3602                 mutex_unlock(&the_lnet.ln_api_mutex);
3603
3604                 return rc;
3605         }
3606
3607         case IOC_LIBCFS_GET_NET: {
3608                 size_t total = sizeof(*config) +
3609                                sizeof(struct lnet_ioctl_net_config);
3610                 config = arg;
3611
3612                 if (config->cfg_hdr.ioc_len < total)
3613                         return -EINVAL;
3614
3615                 mutex_lock(&the_lnet.ln_api_mutex);
3616                 rc = lnet_get_net_config(config);
3617                 mutex_unlock(&the_lnet.ln_api_mutex);
3618                 return rc;
3619         }
3620
3621         case IOC_LIBCFS_GET_LNET_STATS:
3622         {
3623                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
3624
3625                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
3626                         return -EINVAL;
3627
3628                 mutex_lock(&the_lnet.ln_api_mutex);
3629                 lnet_counters_get(&lnet_stats->st_cntrs);
3630                 mutex_unlock(&the_lnet.ln_api_mutex);
3631                 return 0;
3632         }
3633
3634         case IOC_LIBCFS_CONFIG_RTR:
3635                 config = arg;
3636
3637                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3638                         return -EINVAL;
3639
3640                 mutex_lock(&the_lnet.ln_api_mutex);
3641                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
3642                         rc = lnet_rtrpools_enable();
3643                         mutex_unlock(&the_lnet.ln_api_mutex);
3644                         return rc;
3645                 }
3646                 lnet_rtrpools_disable();
3647                 mutex_unlock(&the_lnet.ln_api_mutex);
3648                 return 0;
3649
3650         case IOC_LIBCFS_ADD_BUF:
3651                 config = arg;
3652
3653                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3654                         return -EINVAL;
3655
3656                 mutex_lock(&the_lnet.ln_api_mutex);
3657                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
3658                                                 buf_tiny,
3659                                           config->cfg_config_u.cfg_buffers.
3660                                                 buf_small,
3661                                           config->cfg_config_u.cfg_buffers.
3662                                                 buf_large);
3663                 mutex_unlock(&the_lnet.ln_api_mutex);
3664                 return rc;
3665
3666         case IOC_LIBCFS_SET_NUMA_RANGE: {
3667                 struct lnet_ioctl_set_value *numa;
3668                 numa = arg;
3669                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
3670                         return -EINVAL;
3671                 lnet_net_lock(LNET_LOCK_EX);
3672                 lnet_numa_range = numa->sv_value;
3673                 lnet_net_unlock(LNET_LOCK_EX);
3674                 return 0;
3675         }
3676
3677         case IOC_LIBCFS_GET_NUMA_RANGE: {
3678                 struct lnet_ioctl_set_value *numa;
3679                 numa = arg;
3680                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
3681                         return -EINVAL;
3682                 numa->sv_value = lnet_numa_range;
3683                 return 0;
3684         }
3685
3686         case IOC_LIBCFS_GET_BUF: {
3687                 struct lnet_ioctl_pool_cfg *pool_cfg;
3688                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
3689
3690                 config = arg;
3691
3692                 if (config->cfg_hdr.ioc_len < total)
3693                         return -EINVAL;
3694
3695                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
3696
3697                 mutex_lock(&the_lnet.ln_api_mutex);
3698                 rc = lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
3699                 mutex_unlock(&the_lnet.ln_api_mutex);
3700                 return rc;
3701         }
3702
3703         case IOC_LIBCFS_GET_LOCAL_HSTATS: {
3704                 struct lnet_ioctl_local_ni_hstats *stats = arg;
3705
3706                 if (stats->hlni_hdr.ioc_len < sizeof(*stats))
3707                         return -EINVAL;
3708
3709                 mutex_lock(&the_lnet.ln_api_mutex);
3710                 rc = lnet_get_local_ni_hstats(stats);
3711                 mutex_unlock(&the_lnet.ln_api_mutex);
3712
3713                 return rc;
3714         }
3715
3716         case IOC_LIBCFS_GET_RECOVERY_QUEUE: {
3717                 struct lnet_ioctl_recovery_list *list = arg;
3718                 if (list->rlst_hdr.ioc_len < sizeof(*list))
3719                         return -EINVAL;
3720
3721                 mutex_lock(&the_lnet.ln_api_mutex);
3722                 if (list->rlst_type == LNET_HEALTH_TYPE_LOCAL_NI)
3723                         rc = lnet_get_local_ni_recovery_list(list);
3724                 else
3725                         rc = lnet_get_peer_ni_recovery_list(list);
3726                 mutex_unlock(&the_lnet.ln_api_mutex);
3727                 return rc;
3728         }
3729
3730         case IOC_LIBCFS_ADD_PEER_NI: {
3731                 struct lnet_ioctl_peer_cfg *cfg = arg;
3732
3733                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3734                         return -EINVAL;
3735
3736                 mutex_lock(&the_lnet.ln_api_mutex);
3737                 rc = lnet_add_peer_ni(cfg->prcfg_prim_nid,
3738                                       cfg->prcfg_cfg_nid,
3739                                       cfg->prcfg_mr);
3740                 mutex_unlock(&the_lnet.ln_api_mutex);
3741                 return rc;
3742         }
3743
3744         case IOC_LIBCFS_DEL_PEER_NI: {
3745                 struct lnet_ioctl_peer_cfg *cfg = arg;
3746
3747                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3748                         return -EINVAL;
3749
3750                 mutex_lock(&the_lnet.ln_api_mutex);
3751                 rc = lnet_del_peer_ni(cfg->prcfg_prim_nid,
3752                                       cfg->prcfg_cfg_nid);
3753                 mutex_unlock(&the_lnet.ln_api_mutex);
3754                 return rc;
3755         }
3756
3757         case IOC_LIBCFS_GET_PEER_INFO: {
3758                 struct lnet_ioctl_peer *peer_info = arg;
3759
3760                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
3761                         return -EINVAL;
3762
3763                 mutex_lock(&the_lnet.ln_api_mutex);
3764                 rc = lnet_get_peer_ni_info(
3765                    peer_info->pr_count,
3766                    &peer_info->pr_nid,
3767                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
3768                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
3769                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
3770                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
3771                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
3772                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
3773                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_tx_credits,
3774                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
3775                 mutex_unlock(&the_lnet.ln_api_mutex);
3776                 return rc;
3777         }
3778
3779         case IOC_LIBCFS_GET_PEER_NI: {
3780                 struct lnet_ioctl_peer_cfg *cfg = arg;
3781
3782                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3783                         return -EINVAL;
3784
3785                 mutex_lock(&the_lnet.ln_api_mutex);
3786                 rc = lnet_get_peer_info(cfg,
3787                                         (void __user *)cfg->prcfg_bulk);
3788                 mutex_unlock(&the_lnet.ln_api_mutex);
3789                 return rc;
3790         }
3791
3792         case IOC_LIBCFS_GET_PEER_LIST: {
3793                 struct lnet_ioctl_peer_cfg *cfg = arg;
3794
3795                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3796                         return -EINVAL;
3797
3798                 mutex_lock(&the_lnet.ln_api_mutex);
3799                 rc = lnet_get_peer_list(&cfg->prcfg_count, &cfg->prcfg_size,
3800                                 (struct lnet_process_id __user *)cfg->prcfg_bulk);
3801                 mutex_unlock(&the_lnet.ln_api_mutex);
3802                 return rc;
3803         }
3804
3805         case IOC_LIBCFS_SET_HEALHV: {
3806                 struct lnet_ioctl_reset_health_cfg *cfg = arg;
3807                 int value;
3808                 if (cfg->rh_hdr.ioc_len < sizeof(*cfg))
3809                         return -EINVAL;
3810                 if (cfg->rh_value < 0 ||
3811                     cfg->rh_value > LNET_MAX_HEALTH_VALUE)
3812                         value = LNET_MAX_HEALTH_VALUE;
3813                 else
3814                         value = cfg->rh_value;
3815                 CDEBUG(D_NET, "Manually setting healthv to %d for %s:%s. all = %d\n",
3816                        value, (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI) ?
3817                        "local" : "peer", libcfs_nid2str(cfg->rh_nid), cfg->rh_all);
3818                 mutex_lock(&the_lnet.ln_api_mutex);
3819                 if (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI)
3820                         lnet_ni_set_healthv(cfg->rh_nid, value,
3821                                              cfg->rh_all);
3822                 else
3823                         lnet_peer_ni_set_healthv(cfg->rh_nid, value,
3824                                                   cfg->rh_all);
3825                 mutex_unlock(&the_lnet.ln_api_mutex);
3826                 return 0;
3827         }
3828
3829         case IOC_LIBCFS_NOTIFY_ROUTER: {
3830                 time64_t deadline = ktime_get_real_seconds() - data->ioc_u64[0];
3831
3832                 /* The deadline passed in by the user should be some time in
3833                  * seconds in the future since the UNIX epoch. We have to map
3834                  * that deadline to the wall clock.
3835                  */
3836                 deadline += ktime_get_seconds();
3837                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
3838                                    deadline);
3839         }
3840
3841         case IOC_LIBCFS_LNET_DIST:
3842                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
3843                 if (rc < 0 && rc != -EHOSTUNREACH)
3844                         return rc;
3845
3846                 data->ioc_u32[0] = rc;
3847                 return 0;
3848
3849         case IOC_LIBCFS_TESTPROTOCOMPAT:
3850                 lnet_net_lock(LNET_LOCK_EX);
3851                 the_lnet.ln_testprotocompat = data->ioc_flags;
3852                 lnet_net_unlock(LNET_LOCK_EX);
3853                 return 0;
3854
3855         case IOC_LIBCFS_LNET_FAULT:
3856                 return lnet_fault_ctl(data->ioc_flags, data);
3857
3858         case IOC_LIBCFS_PING: {
3859                 signed long timeout;
3860
3861                 id.nid = data->ioc_nid;
3862                 id.pid = data->ioc_u32[0];
3863
3864                 /* If timeout is negative then set default of 3 minutes */
3865                 if (((s32)data->ioc_u32[1] <= 0) ||
3866                     data->ioc_u32[1] > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
3867                         timeout = msecs_to_jiffies(DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC);
3868                 else
3869                         timeout = msecs_to_jiffies(data->ioc_u32[1]);
3870
3871                 rc = lnet_ping(id, timeout, data->ioc_pbuf1,
3872                                data->ioc_plen1 / sizeof(struct lnet_process_id));
3873
3874                 if (rc < 0)
3875                         return rc;
3876
3877                 data->ioc_count = rc;
3878                 return 0;
3879         }
3880
3881         case IOC_LIBCFS_PING_PEER: {
3882                 struct lnet_ioctl_ping_data *ping = arg;
3883                 struct lnet_peer *lp;
3884                 signed long timeout;
3885
3886                 /* If timeout is negative then set default of 3 minutes */
3887                 if (((s32)ping->op_param) <= 0 ||
3888                     ping->op_param > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
3889                         timeout = msecs_to_jiffies(DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC);
3890                 else
3891                         timeout = msecs_to_jiffies(ping->op_param);
3892
3893                 rc = lnet_ping(ping->ping_id, timeout,
3894                                ping->ping_buf,
3895                                ping->ping_count);
3896                 if (rc < 0)
3897                         return rc;
3898
3899                 mutex_lock(&the_lnet.ln_api_mutex);
3900                 lp = lnet_find_peer(ping->ping_id.nid);
3901                 if (lp) {
3902                         ping->ping_id.nid = lp->lp_primary_nid;
3903                         ping->mr_info = lnet_peer_is_multi_rail(lp);
3904                         lnet_peer_decref_locked(lp);
3905                 }
3906                 mutex_unlock(&the_lnet.ln_api_mutex);
3907
3908                 ping->ping_count = rc;
3909                 return 0;
3910         }
3911
3912         case IOC_LIBCFS_DISCOVER: {
3913                 struct lnet_ioctl_ping_data *discover = arg;
3914                 struct lnet_peer *lp;
3915
3916                 rc = lnet_discover(discover->ping_id, discover->op_param,
3917                                    discover->ping_buf,
3918                                    discover->ping_count);
3919                 if (rc < 0)
3920                         return rc;
3921
3922                 mutex_lock(&the_lnet.ln_api_mutex);
3923                 lp = lnet_find_peer(discover->ping_id.nid);
3924                 if (lp) {
3925                         discover->ping_id.nid = lp->lp_primary_nid;
3926                         discover->mr_info = lnet_peer_is_multi_rail(lp);
3927                         lnet_peer_decref_locked(lp);
3928                 }
3929                 mutex_unlock(&the_lnet.ln_api_mutex);
3930
3931                 discover->ping_count = rc;
3932                 return 0;
3933         }
3934
3935         default:
3936                 ni = lnet_net2ni_addref(data->ioc_net);
3937                 if (ni == NULL)
3938                         return -EINVAL;
3939
3940                 if (ni->ni_net->net_lnd->lnd_ctl == NULL)
3941                         rc = -EINVAL;
3942                 else
3943                         rc = ni->ni_net->net_lnd->lnd_ctl(ni, cmd, arg);
3944
3945                 lnet_ni_decref(ni);
3946                 return rc;
3947         }
3948         /* not reached */
3949 }
3950 EXPORT_SYMBOL(LNetCtl);
3951
3952 void LNetDebugPeer(struct lnet_process_id id)
3953 {
3954         lnet_debug_peer(id.nid);
3955 }
3956 EXPORT_SYMBOL(LNetDebugPeer);
3957
3958 /**
3959  * Determine if the specified peer \a nid is on the local node.
3960  *
3961  * \param nid   peer nid to check
3962  *
3963  * \retval true         If peer NID is on the local node.
3964  * \retval false        If peer NID is not on the local node.
3965  */
3966 bool LNetIsPeerLocal(lnet_nid_t nid)
3967 {
3968         struct lnet_net *net;
3969         struct lnet_ni *ni;
3970         int cpt;
3971
3972         cpt = lnet_net_lock_current();
3973         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3974                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3975                         if (ni->ni_nid == nid) {
3976                                 lnet_net_unlock(cpt);
3977                                 return true;
3978                         }
3979                 }
3980         }
3981         lnet_net_unlock(cpt);
3982
3983         return false;
3984 }
3985 EXPORT_SYMBOL(LNetIsPeerLocal);
3986
3987 /**
3988  * Retrieve the struct lnet_process_id ID of LNet interface at \a index.
3989  * Note that all interfaces share a same PID, as requested by LNetNIInit().
3990  *
3991  * \param index Index of the interface to look up.
3992  * \param id On successful return, this location will hold the
3993  * struct lnet_process_id ID of the interface.
3994  *
3995  * \retval 0 If an interface exists at \a index.
3996  * \retval -ENOENT If no interface has been found.
3997  */
3998 int
3999 LNetGetId(unsigned int index, struct lnet_process_id *id)
4000 {
4001         struct lnet_ni   *ni;
4002         struct lnet_net  *net;
4003         int               cpt;
4004         int               rc = -ENOENT;
4005
4006         LASSERT(the_lnet.ln_refcount > 0);
4007
4008         cpt = lnet_net_lock_current();
4009
4010         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
4011                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
4012                         if (index-- != 0)
4013                                 continue;
4014
4015                         id->nid = ni->ni_nid;
4016                         id->pid = the_lnet.ln_pid;
4017                         rc = 0;
4018                         break;
4019                 }
4020         }
4021
4022         lnet_net_unlock(cpt);
4023         return rc;
4024 }
4025 EXPORT_SYMBOL(LNetGetId);
4026
4027 static int lnet_ping(struct lnet_process_id id, signed long timeout,
4028                      struct lnet_process_id __user *ids, int n_ids)
4029 {
4030         struct lnet_handle_eq eqh;
4031         struct lnet_handle_md mdh;
4032         struct lnet_event event;
4033         struct lnet_md md = { NULL };
4034         int which;
4035         int unlinked = 0;
4036         int replied = 0;
4037         const signed long a_long_time = msecs_to_jiffies(60 * MSEC_PER_SEC);
4038         struct lnet_ping_buffer *pbuf;
4039         struct lnet_process_id tmpid;
4040         int i;
4041         int nob;
4042         int rc;
4043         int rc2;
4044         sigset_t blocked;
4045
4046         /* n_ids limit is arbitrary */
4047         if (n_ids <= 0 || id.nid == LNET_NID_ANY)
4048                 return -EINVAL;
4049
4050         /*
4051          * if the user buffer has more space than the lnet_interfaces_max
4052          * then only fill it up to lnet_interfaces_max
4053          */
4054         if (n_ids > lnet_interfaces_max)
4055                 n_ids = lnet_interfaces_max;
4056
4057         if (id.pid == LNET_PID_ANY)
4058                 id.pid = LNET_PID_LUSTRE;
4059
4060         pbuf = lnet_ping_buffer_alloc(n_ids, GFP_NOFS);
4061         if (!pbuf)
4062                 return -ENOMEM;
4063
4064         /* NB 2 events max (including any unlink event) */
4065         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
4066         if (rc != 0) {
4067                 CERROR("Can't allocate EQ: %d\n", rc);
4068                 goto fail_ping_buffer_decref;
4069         }
4070
4071         /* initialize md content */
4072         md.start     = &pbuf->pb_info;
4073         md.length    = LNET_PING_INFO_SIZE(n_ids);
4074         md.threshold = 2; /* GET/REPLY */
4075         md.max_size  = 0;
4076         md.options   = LNET_MD_TRUNCATE;
4077         md.user_ptr  = NULL;
4078         md.eq_handle = eqh;
4079
4080         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
4081         if (rc != 0) {
4082                 CERROR("Can't bind MD: %d\n", rc);
4083                 goto fail_free_eq;
4084         }
4085
4086         rc = LNetGet(LNET_NID_ANY, mdh, id,
4087                      LNET_RESERVED_PORTAL,
4088                      LNET_PROTO_PING_MATCHBITS, 0, false);
4089
4090         if (rc != 0) {
4091                 /* Don't CERROR; this could be deliberate! */
4092                 rc2 = LNetMDUnlink(mdh);
4093                 LASSERT(rc2 == 0);
4094
4095                 /* NB must wait for the UNLINK event below... */
4096                 unlinked = 1;
4097                 timeout = a_long_time;
4098         }
4099
4100         do {
4101                 /* MUST block for unlink to complete */
4102                 if (unlinked)
4103                         blocked = cfs_block_allsigs();
4104
4105                 rc2 = LNetEQPoll(&eqh, 1, timeout, &event, &which);
4106
4107                 if (unlinked)
4108                         cfs_restore_sigs(blocked);
4109
4110                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
4111                        (rc2 <= 0) ? -1 : event.type,
4112                        (rc2 <= 0) ? -1 : event.status,
4113                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
4114
4115                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
4116
4117                 if (rc2 <= 0 || event.status != 0) {
4118                         /* timeout or error */
4119                         if (!replied && rc == 0)
4120                                 rc = (rc2 < 0) ? rc2 :
4121                                      (rc2 == 0) ? -ETIMEDOUT :
4122                                      event.status;
4123
4124                         if (!unlinked) {
4125                                 /* Ensure completion in finite time... */
4126                                 LNetMDUnlink(mdh);
4127                                 /* No assertion (racing with network) */
4128                                 unlinked = 1;
4129                                 timeout = a_long_time;
4130                         } else if (rc2 == 0) {
4131                                 /* timed out waiting for unlink */
4132                                 CWARN("ping %s: late network completion\n",
4133                                       libcfs_id2str(id));
4134                         }
4135                 } else if (event.type == LNET_EVENT_REPLY) {
4136                         replied = 1;
4137                         rc = event.mlength;
4138                 }
4139         } while (rc2 <= 0 || !event.unlinked);
4140
4141         if (!replied) {
4142                 if (rc >= 0)
4143                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
4144                               libcfs_id2str(id));
4145                 rc = -EIO;
4146                 goto fail_free_eq;
4147         }
4148
4149         nob = rc;
4150         LASSERT(nob >= 0 && nob <= LNET_PING_INFO_SIZE(n_ids));
4151
4152         rc = -EPROTO;           /* if I can't parse... */
4153
4154         if (nob < 8) {
4155                 CERROR("%s: ping info too short %d\n",
4156                        libcfs_id2str(id), nob);
4157                 goto fail_free_eq;
4158         }
4159
4160         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
4161                 lnet_swap_pinginfo(pbuf);
4162         } else if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
4163                 CERROR("%s: Unexpected magic %08x\n",
4164                        libcfs_id2str(id), pbuf->pb_info.pi_magic);
4165                 goto fail_free_eq;
4166         }
4167
4168         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
4169                 CERROR("%s: ping w/o NI status: 0x%x\n",
4170                        libcfs_id2str(id), pbuf->pb_info.pi_features);
4171                 goto fail_free_eq;
4172         }
4173
4174         if (nob < LNET_PING_INFO_SIZE(0)) {
4175                 CERROR("%s: Short reply %d(%d min)\n",
4176                        libcfs_id2str(id),
4177                        nob, (int)LNET_PING_INFO_SIZE(0));
4178                 goto fail_free_eq;
4179         }
4180
4181         if (pbuf->pb_info.pi_nnis < n_ids)
4182                 n_ids = pbuf->pb_info.pi_nnis;
4183
4184         if (nob < LNET_PING_INFO_SIZE(n_ids)) {
4185                 CERROR("%s: Short reply %d(%d expected)\n",
4186                        libcfs_id2str(id),
4187                        nob, (int)LNET_PING_INFO_SIZE(n_ids));
4188                 goto fail_free_eq;
4189         }
4190
4191         rc = -EFAULT;           /* if I segv in copy_to_user()... */
4192
4193         memset(&tmpid, 0, sizeof(tmpid));
4194         for (i = 0; i < n_ids; i++) {
4195                 tmpid.pid = pbuf->pb_info.pi_pid;
4196                 tmpid.nid = pbuf->pb_info.pi_ni[i].ns_nid;
4197                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
4198                         goto fail_free_eq;
4199         }
4200         rc = pbuf->pb_info.pi_nnis;
4201
4202  fail_free_eq:
4203         rc2 = LNetEQFree(eqh);
4204         if (rc2 != 0)
4205                 CERROR("rc2 %d\n", rc2);
4206         LASSERT(rc2 == 0);
4207
4208  fail_ping_buffer_decref:
4209         lnet_ping_buffer_decref(pbuf);
4210         return rc;
4211 }
4212
4213 static int
4214 lnet_discover(struct lnet_process_id id, __u32 force,
4215               struct lnet_process_id __user *ids, int n_ids)
4216 {
4217         struct lnet_peer_ni *lpni;
4218         struct lnet_peer_ni *p;
4219         struct lnet_peer *lp;
4220         struct lnet_process_id *buf;
4221         int cpt;
4222         int i;
4223         int rc;
4224         int max_intf = lnet_interfaces_max;
4225         size_t buf_size;
4226
4227         if (n_ids <= 0 ||
4228             id.nid == LNET_NID_ANY)
4229                 return -EINVAL;
4230
4231         if (id.pid == LNET_PID_ANY)
4232                 id.pid = LNET_PID_LUSTRE;
4233
4234         /*
4235          * if the user buffer has more space than the max_intf
4236          * then only fill it up to max_intf
4237          */
4238         if (n_ids > max_intf)
4239                 n_ids = max_intf;
4240
4241         buf_size = n_ids * sizeof(*buf);
4242
4243         LIBCFS_ALLOC(buf, buf_size);
4244         if (!buf)
4245                 return -ENOMEM;
4246
4247         cpt = lnet_net_lock_current();
4248         lpni = lnet_nid2peerni_locked(id.nid, LNET_NID_ANY, cpt);
4249         if (IS_ERR(lpni)) {
4250                 rc = PTR_ERR(lpni);
4251                 goto out;
4252         }
4253
4254         /*
4255          * Clearing the NIDS_UPTODATE flag ensures the peer will
4256          * be discovered, provided discovery has not been disabled.
4257          */
4258         lp = lpni->lpni_peer_net->lpn_peer;
4259         spin_lock(&lp->lp_lock);
4260         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
4261         /* If the force flag is set, force a PING and PUSH as well. */
4262         if (force)
4263                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
4264         spin_unlock(&lp->lp_lock);
4265         rc = lnet_discover_peer_locked(lpni, cpt, true);
4266         if (rc)
4267                 goto out_decref;
4268
4269         /* Peer may have changed. */
4270         lp = lpni->lpni_peer_net->lpn_peer;
4271         if (lp->lp_nnis < n_ids)
4272                 n_ids = lp->lp_nnis;
4273
4274         i = 0;
4275         p = NULL;
4276         while ((p = lnet_get_next_peer_ni_locked(lp, NULL, p)) != NULL) {
4277                 buf[i].pid = id.pid;
4278                 buf[i].nid = p->lpni_nid;
4279                 if (++i >= n_ids)
4280                         break;
4281         }
4282
4283         lnet_net_unlock(cpt);
4284
4285         rc = -EFAULT;
4286         if (copy_to_user(ids, buf, n_ids * sizeof(*buf)))
4287                 goto out_relock;
4288         rc = n_ids;
4289 out_relock:
4290         lnet_net_lock(cpt);
4291 out_decref:
4292         lnet_peer_ni_decref_locked(lpni);
4293 out:
4294         lnet_net_unlock(cpt);
4295
4296         LIBCFS_FREE(buf, buf_size);
4297
4298         return rc;
4299 }
4300
4301 /**
4302  * Retrieve peer discovery status.
4303  *
4304  * \retval 1 if lnet_peer_discovery_disabled is 0
4305  * \retval 0 if lnet_peer_discovery_disabled is 1
4306  */
4307 int
4308 LNetGetPeerDiscoveryStatus(void)
4309 {
4310         return !lnet_peer_discovery_disabled;
4311 }
4312 EXPORT_SYMBOL(LNetGetPeerDiscoveryStatus);