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 0, which means health
  84  * checking is turned off by default.
  85  */
  86 unsigned int lnet_health_sensitivity = 0;
  87 static int sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp);
  88 static struct kernel_param_ops param_ops_health_sensitivity = {
  89         .set = sensitivity_set,
  90         .get = param_get_int,
  91 };
  92 #define param_check_health_sensitivity(name, p) \
  93                 __param_check(name, p, int)
  94 #ifdef HAVE_KERNEL_PARAM_OPS
  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 static int lnet_interfaces_max = LNET_INTERFACES_MAX_DEFAULT;
 104 static int intf_max_set(const char *val, cfs_kernel_param_arg_t *kp);
 105
 106 static struct kernel_param_ops param_ops_interfaces_max = {
 107         .set = intf_max_set,
 108         .get = param_get_int,
 109 };
 110
 111 #define param_check_interfaces_max(name, p) \
 112                 __param_check(name, p, int)
 113
 114 #ifdef HAVE_KERNEL_PARAM_OPS
 115 module_param(lnet_interfaces_max, interfaces_max, 0644);
 116 #else
 117 module_param_call(lnet_interfaces_max, intf_max_set, param_get_int,
 118                   &param_ops_interfaces_max, 0644);
 119 #endif
 120 MODULE_PARM_DESC(lnet_interfaces_max,
 121                 "Maximum number of interfaces in a node.");
 122
 123 unsigned lnet_peer_discovery_disabled = 0;
 124 static int discovery_set(const char *val, cfs_kernel_param_arg_t *kp);
 125
 126 static struct kernel_param_ops param_ops_discovery_disabled = {
 127         .set = discovery_set,
 128         .get = param_get_int,
 129 };
 130
 131 #define param_check_discovery_disabled(name, p) \
 132                 __param_check(name, p, int)
 133 #ifdef HAVE_KERNEL_PARAM_OPS
 134 module_param(lnet_peer_discovery_disabled, discovery_disabled, 0644);
 135 #else
 136 module_param_call(lnet_peer_discovery_disabled, discovery_set, param_get_int,
 137                   &param_ops_discovery_disabled, 0644);
 138 #endif
 139 MODULE_PARM_DESC(lnet_peer_discovery_disabled,
 140                 "Set to 1 to disable peer discovery on this node.");
 141
 142 unsigned lnet_transaction_timeout = 5;
 143 static int transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp);
 144 static struct kernel_param_ops param_ops_transaction_timeout = {
 145         .set = transaction_to_set,
 146         .get = param_get_int,
 147 };
 148
 149 #define param_check_transaction_timeout(name, p) \
 150                 __param_check(name, p, int)
 151 #ifdef HAVE_KERNEL_PARAM_OPS
 152 module_param(lnet_transaction_timeout, transaction_timeout, S_IRUGO|S_IWUSR);
 153 #else
 154 module_param_call(lnet_transaction_timeout, transaction_to_set, param_get_int,
 155                   &lnet_transaction_timeout, S_IRUGO|S_IWUSR);
 156 #endif
 157 MODULE_PARM_DESC(lnet_peer_discovery_disabled,
 158                 "Set to 1 to disable peer discovery on this node.");
 159
 160 unsigned lnet_retry_count = 0;
 161 static int retry_count_set(const char *val, cfs_kernel_param_arg_t *kp);
 162 static struct kernel_param_ops param_ops_retry_count = {
 163         .set = retry_count_set,
 164         .get = param_get_int,
 165 };
 166
 167 #define param_check_retry_count(name, p) \
 168                 __param_check(name, p, int)
 169 #ifdef HAVE_KERNEL_PARAM_OPS
 170 module_param(lnet_retry_count, retry_count, S_IRUGO|S_IWUSR);
 171 #else
 172 module_param_call(lnet_retry_count, retry_count_set, param_get_int,
 173                   &lnet_retry_count, S_IRUGO|S_IWUSR);
 174 #endif
 175 MODULE_PARM_DESC(lnet_retry_count,
 176                  "Maximum number of times to retry transmitting a message");
 177
 178
 179 unsigned lnet_lnd_timeout = LNET_LND_DEFAULT_TIMEOUT;
 180
 181 /*
 182  * This sequence number keeps track of how many times DLC was used to
 183  * update the local NIs. It is incremented when a NI is added or
 184  * removed and checked when sending a message to determine if there is
 185  * a need to re-run the selection algorithm. See lnet_select_pathway()
 186  * for more details on its usage.
 187  */
 188 static atomic_t lnet_dlc_seq_no = ATOMIC_INIT(0);
 189
 190 static int lnet_ping(struct lnet_process_id id, signed long timeout,
 191                      struct lnet_process_id __user *ids, int n_ids);
 192
 193 static int lnet_discover(struct lnet_process_id id, __u32 force,
 194                          struct lnet_process_id __user *ids, int n_ids);
 195
 196 static int
 197 sensitivity_set(const char *val, cfs_kernel_param_arg_t *kp)
 198 {
 199         int rc;
 200         unsigned *sensitivity = (unsigned *)kp->arg;
 201         unsigned long value;
 202
 203         rc = kstrtoul(val, 0, &value);
 204         if (rc) {
 205                 CERROR("Invalid module parameter value for 'lnet_health_sensitivity'\n");
 206                 return rc;
 207         }
 208
 209         /*
 210          * The purpose of locking the api_mutex here is to ensure that
 211          * the correct value ends up stored properly.
 212          */
 213         mutex_lock(&the_lnet.ln_api_mutex);
 214
 215         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
 216                 mutex_unlock(&the_lnet.ln_api_mutex);
 217                 return 0;
 218         }
 219
 220         if (value == *sensitivity) {
 221                 mutex_unlock(&the_lnet.ln_api_mutex);
 222                 return 0;
 223         }
 224
 225         *sensitivity = value;
 226
 227         mutex_unlock(&the_lnet.ln_api_mutex);
 228
 229         return 0;
 230 }
 231
 232 static int
 233 discovery_set(const char *val, cfs_kernel_param_arg_t *kp)
 234 {
 235         int rc;
 236         unsigned *discovery = (unsigned *)kp->arg;
 237         unsigned long value;
 238         struct lnet_ping_buffer *pbuf;
 239
 240         rc = kstrtoul(val, 0, &value);
 241         if (rc) {
 242                 CERROR("Invalid module parameter value for 'lnet_peer_discovery_disabled'\n");
 243                 return rc;
 244         }
 245
 246         value = (value) ? 1 : 0;
 247
 248         /*
 249          * The purpose of locking the api_mutex here is to ensure that
 250          * the correct value ends up stored properly.
 251          */
 252         mutex_lock(&the_lnet.ln_api_mutex);
 253
 254         if (value == *discovery) {
 255                 mutex_unlock(&the_lnet.ln_api_mutex);
 256                 return 0;
 257         }
 258
 259         *discovery = value;
 260
 261         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
 262                 mutex_unlock(&the_lnet.ln_api_mutex);
 263                 return 0;
 264         }
 265
 266         /* tell peers that discovery setting has changed */
 267         lnet_net_lock(LNET_LOCK_EX);
 268         pbuf = the_lnet.ln_ping_target;
 269         if (value)
 270                 pbuf->pb_info.pi_features &= ~LNET_PING_FEAT_DISCOVERY;
 271         else
 272                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
 273         lnet_net_unlock(LNET_LOCK_EX);
 274
 275         lnet_push_update_to_peers(1);
 276
 277         mutex_unlock(&the_lnet.ln_api_mutex);
 278
 279         return 0;
 280 }
 281
 282 static int
 283 transaction_to_set(const char *val, cfs_kernel_param_arg_t *kp)
 284 {
 285         int rc;
 286         unsigned *transaction_to = (unsigned *)kp->arg;
 287         unsigned long value;
 288
 289         rc = kstrtoul(val, 0, &value);
 290         if (rc) {
 291                 CERROR("Invalid module parameter value for 'lnet_transaction_timeout'\n");
 292                 return rc;
 293         }
 294
 295         /*
 296          * The purpose of locking the api_mutex here is to ensure that
 297          * the correct value ends up stored properly.
 298          */
 299         mutex_lock(&the_lnet.ln_api_mutex);
 300
 301         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
 302                 mutex_unlock(&the_lnet.ln_api_mutex);
 303                 return 0;
 304         }
 305
 306         if (value < lnet_retry_count || value == 0) {
 307                 mutex_unlock(&the_lnet.ln_api_mutex);
 308                 CERROR("Invalid value for lnet_transaction_timeout (%lu). "
 309                        "Has to be greater than lnet_retry_count (%u)\n",
 310                        value, lnet_retry_count);
 311                 return -EINVAL;
 312         }
 313
 314         if (value == *transaction_to) {
 315                 mutex_unlock(&the_lnet.ln_api_mutex);
 316                 return 0;
 317         }
 318
 319         *transaction_to = value;
 320         if (lnet_retry_count == 0)
 321                 lnet_lnd_timeout = value;
 322         else
 323                 lnet_lnd_timeout = value / lnet_retry_count;
 324
 325         mutex_unlock(&the_lnet.ln_api_mutex);
 326
 327         return 0;
 328 }
 329
 330 static int
 331 retry_count_set(const char *val, cfs_kernel_param_arg_t *kp)
 332 {
 333         int rc;
 334         unsigned *retry_count = (unsigned *)kp->arg;
 335         unsigned long value;
 336
 337         rc = kstrtoul(val, 0, &value);
 338         if (rc) {
 339                 CERROR("Invalid module parameter value for 'lnet_retry_count'\n");
 340                 return rc;
 341         }
 342
 343         /*
 344          * The purpose of locking the api_mutex here is to ensure that
 345          * the correct value ends up stored properly.
 346          */
 347         mutex_lock(&the_lnet.ln_api_mutex);
 348
 349         if (the_lnet.ln_state != LNET_STATE_RUNNING) {
 350                 mutex_unlock(&the_lnet.ln_api_mutex);
 351                 return 0;
 352         }
 353
 354         if (value > lnet_transaction_timeout) {
 355                 mutex_unlock(&the_lnet.ln_api_mutex);
 356                 CERROR("Invalid value for lnet_retry_count (%lu). "
 357                        "Has to be smaller than lnet_transaction_timeout (%u)\n",
 358                        value, lnet_transaction_timeout);
 359                 return -EINVAL;
 360         }
 361
 362         if (value == *retry_count) {
 363                 mutex_unlock(&the_lnet.ln_api_mutex);
 364                 return 0;
 365         }
 366
 367         *retry_count = value;
 368
 369         if (value == 0)
 370                 lnet_lnd_timeout = lnet_transaction_timeout;
 371         else
 372                 lnet_lnd_timeout = lnet_transaction_timeout / value;
 373
 374         mutex_unlock(&the_lnet.ln_api_mutex);
 375
 376         return 0;
 377 }
 378
 379 static int
 380 intf_max_set(const char *val, cfs_kernel_param_arg_t *kp)
 381 {
 382         int value, rc;
 383
 384         rc = kstrtoint(val, 0, &value);
 385         if (rc) {
 386                 CERROR("Invalid module parameter value for 'lnet_interfaces_max'\n");
 387                 return rc;
 388         }
 389
 390         if (value < LNET_INTERFACES_MIN) {
 391                 CWARN("max interfaces provided are too small, setting to %d\n",
 392                       LNET_INTERFACES_MAX_DEFAULT);
 393                 value = LNET_INTERFACES_MAX_DEFAULT;
 394         }
 395
 396         *(int *)kp->arg = value;
 397
 398         return 0;
 399 }
 400
 401 static char *
 402 lnet_get_routes(void)
 403 {
 404         return routes;
 405 }
 406
 407 static char *
 408 lnet_get_networks(void)
 409 {
 410         char   *nets;
 411         int     rc;
 412
 413         if (*networks != 0 && *ip2nets != 0) {
 414                 LCONSOLE_ERROR_MSG(0x101, "Please specify EITHER 'networks' or "
 415                                    "'ip2nets' but not both at once\n");
 416                 return NULL;
 417         }
 418
 419         if (*ip2nets != 0) {
 420                 rc = lnet_parse_ip2nets(&nets, ip2nets);
 421                 return (rc == 0) ? nets : NULL;
 422         }
 423
 424         if (*networks != 0)
 425                 return networks;
 426
 427         return "tcp";
 428 }
 429
 430 static void
 431 lnet_init_locks(void)
 432 {
 433         spin_lock_init(&the_lnet.ln_eq_wait_lock);
 434         spin_lock_init(&the_lnet.ln_msg_resend_lock);
 435         init_waitqueue_head(&the_lnet.ln_eq_waitq);
 436         init_waitqueue_head(&the_lnet.ln_mt_waitq);
 437         mutex_init(&the_lnet.ln_lnd_mutex);
 438 }
 439
 440 static void
 441 lnet_fini_locks(void)
 442 {
 443 }
 444
 445 struct kmem_cache *lnet_mes_cachep;        /* MEs kmem_cache */
 446 struct kmem_cache *lnet_small_mds_cachep;  /* <= LNET_SMALL_MD_SIZE bytes
 447                                             *  MDs kmem_cache */
 448
 449 static int
 450 lnet_descriptor_setup(void)
 451 {
 452         /* create specific kmem_cache for MEs and small MDs (i.e., originally
 453          * allocated in <size-xxx> kmem_cache).
 454          */
 455         lnet_mes_cachep = kmem_cache_create("lnet_MEs", sizeof(struct lnet_me),
 456                                             0, 0, NULL);
 457         if (!lnet_mes_cachep)
 458                 return -ENOMEM;
 459
 460         lnet_small_mds_cachep = kmem_cache_create("lnet_small_MDs",
 461                                                   LNET_SMALL_MD_SIZE, 0, 0,
 462                                                   NULL);
 463         if (!lnet_small_mds_cachep)
 464                 return -ENOMEM;
 465
 466         return 0;
 467 }
 468
 469 static void
 470 lnet_descriptor_cleanup(void)
 471 {
 472
 473         if (lnet_small_mds_cachep) {
 474                 kmem_cache_destroy(lnet_small_mds_cachep);
 475                 lnet_small_mds_cachep = NULL;
 476         }
 477
 478         if (lnet_mes_cachep) {
 479                 kmem_cache_destroy(lnet_mes_cachep);
 480                 lnet_mes_cachep = NULL;
 481         }
 482 }
 483
 484 static int
 485 lnet_create_remote_nets_table(void)
 486 {
 487         int               i;
 488         struct list_head *hash;
 489
 490         LASSERT(the_lnet.ln_remote_nets_hash == NULL);
 491         LASSERT(the_lnet.ln_remote_nets_hbits > 0);
 492         LIBCFS_ALLOC(hash, LNET_REMOTE_NETS_HASH_SIZE * sizeof(*hash));
 493         if (hash == NULL) {
 494                 CERROR("Failed to create remote nets hash table\n");
 495                 return -ENOMEM;
 496         }
 497
 498         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 499                 INIT_LIST_HEAD(&hash[i]);
 500         the_lnet.ln_remote_nets_hash = hash;
 501         return 0;
 502 }
 503
 504 static void
 505 lnet_destroy_remote_nets_table(void)
 506 {
 507         int i;
 508
 509         if (the_lnet.ln_remote_nets_hash == NULL)
 510                 return;
 511
 512         for (i = 0; i < LNET_REMOTE_NETS_HASH_SIZE; i++)
 513                 LASSERT(list_empty(&the_lnet.ln_remote_nets_hash[i]));
 514
 515         LIBCFS_FREE(the_lnet.ln_remote_nets_hash,
 516                     LNET_REMOTE_NETS_HASH_SIZE *
 517                     sizeof(the_lnet.ln_remote_nets_hash[0]));
 518         the_lnet.ln_remote_nets_hash = NULL;
 519 }
 520
 521 static void
 522 lnet_destroy_locks(void)
 523 {
 524         if (the_lnet.ln_res_lock != NULL) {
 525                 cfs_percpt_lock_free(the_lnet.ln_res_lock);
 526                 the_lnet.ln_res_lock = NULL;
 527         }
 528
 529         if (the_lnet.ln_net_lock != NULL) {
 530                 cfs_percpt_lock_free(the_lnet.ln_net_lock);
 531                 the_lnet.ln_net_lock = NULL;
 532         }
 533
 534         lnet_fini_locks();
 535 }
 536
 537 static int
 538 lnet_create_locks(void)
 539 {
 540         lnet_init_locks();
 541
 542         the_lnet.ln_res_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 543         if (the_lnet.ln_res_lock == NULL)
 544                 goto failed;
 545
 546         the_lnet.ln_net_lock = cfs_percpt_lock_alloc(lnet_cpt_table());
 547         if (the_lnet.ln_net_lock == NULL)
 548                 goto failed;
 549
 550         return 0;
 551
 552  failed:
 553         lnet_destroy_locks();
 554         return -ENOMEM;
 555 }
 556
 557 static void lnet_assert_wire_constants(void)
 558 {
 559         /* Wire protocol assertions generated by 'wirecheck'
 560          * running on Linux robert.bartonsoftware.com 2.6.8-1.521
 561          * #1 Mon Aug 16 09:01:18 EDT 2004 i686 athlon i386 GNU/Linux
 562          * with gcc version 3.3.3 20040412 (Red Hat Linux 3.3.3-7) */
 563
 564         /* Constants... */
 565         CLASSERT(LNET_PROTO_TCP_MAGIC == 0xeebc0ded);
 566         CLASSERT(LNET_PROTO_TCP_VERSION_MAJOR == 1);
 567         CLASSERT(LNET_PROTO_TCP_VERSION_MINOR == 0);
 568         CLASSERT(LNET_MSG_ACK == 0);
 569         CLASSERT(LNET_MSG_PUT == 1);
 570         CLASSERT(LNET_MSG_GET == 2);
 571         CLASSERT(LNET_MSG_REPLY == 3);
 572         CLASSERT(LNET_MSG_HELLO == 4);
 573
 574         /* Checks for struct lnet_handle_wire */
 575         CLASSERT((int)sizeof(struct lnet_handle_wire) == 16);
 576         CLASSERT((int)offsetof(struct lnet_handle_wire, wh_interface_cookie) == 0);
 577         CLASSERT((int)sizeof(((struct lnet_handle_wire *)0)->wh_interface_cookie) == 8);
 578         CLASSERT((int)offsetof(struct lnet_handle_wire, wh_object_cookie) == 8);
 579         CLASSERT((int)sizeof(((struct lnet_handle_wire *)0)->wh_object_cookie) == 8);
 580
 581         /* Checks for struct struct lnet_magicversion */
 582         CLASSERT((int)sizeof(struct lnet_magicversion) == 8);
 583         CLASSERT((int)offsetof(struct lnet_magicversion, magic) == 0);
 584         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->magic) == 4);
 585         CLASSERT((int)offsetof(struct lnet_magicversion, version_major) == 4);
 586         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->version_major) == 2);
 587         CLASSERT((int)offsetof(struct lnet_magicversion, version_minor) == 6);
 588         CLASSERT((int)sizeof(((struct lnet_magicversion *)0)->version_minor) == 2);
 589
 590         /* Checks for struct struct lnet_hdr */
 591         CLASSERT((int)sizeof(struct lnet_hdr) == 72);
 592         CLASSERT((int)offsetof(struct lnet_hdr, dest_nid) == 0);
 593         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->dest_nid) == 8);
 594         CLASSERT((int)offsetof(struct lnet_hdr, src_nid) == 8);
 595         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->src_nid) == 8);
 596         CLASSERT((int)offsetof(struct lnet_hdr, dest_pid) == 16);
 597         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->dest_pid) == 4);
 598         CLASSERT((int)offsetof(struct lnet_hdr, src_pid) == 20);
 599         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->src_pid) == 4);
 600         CLASSERT((int)offsetof(struct lnet_hdr, type) == 24);
 601         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->type) == 4);
 602         CLASSERT((int)offsetof(struct lnet_hdr, payload_length) == 28);
 603         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->payload_length) == 4);
 604         CLASSERT((int)offsetof(struct lnet_hdr, msg) == 32);
 605         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg) == 40);
 606
 607         /* Ack */
 608         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.dst_wmd) == 32);
 609         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.dst_wmd) == 16);
 610         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.match_bits) == 48);
 611         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.match_bits) == 8);
 612         CLASSERT((int)offsetof(struct lnet_hdr, msg.ack.mlength) == 56);
 613         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.ack.mlength) == 4);
 614
 615         /* Put */
 616         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.ack_wmd) == 32);
 617         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.ack_wmd) == 16);
 618         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.match_bits) == 48);
 619         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.match_bits) == 8);
 620         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.hdr_data) == 56);
 621         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.hdr_data) == 8);
 622         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.ptl_index) == 64);
 623         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.ptl_index) == 4);
 624         CLASSERT((int)offsetof(struct lnet_hdr, msg.put.offset) == 68);
 625         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.put.offset) == 4);
 626
 627         /* Get */
 628         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.return_wmd) == 32);
 629         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.return_wmd) == 16);
 630         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.match_bits) == 48);
 631         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.match_bits) == 8);
 632         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.ptl_index) == 56);
 633         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.ptl_index) == 4);
 634         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.src_offset) == 60);
 635         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.src_offset) == 4);
 636         CLASSERT((int)offsetof(struct lnet_hdr, msg.get.sink_length) == 64);
 637         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.get.sink_length) == 4);
 638
 639         /* Reply */
 640         CLASSERT((int)offsetof(struct lnet_hdr, msg.reply.dst_wmd) == 32);
 641         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.reply.dst_wmd) == 16);
 642
 643         /* Hello */
 644         CLASSERT((int)offsetof(struct lnet_hdr, msg.hello.incarnation) == 32);
 645         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.hello.incarnation) == 8);
 646         CLASSERT((int)offsetof(struct lnet_hdr, msg.hello.type) == 40);
 647         CLASSERT((int)sizeof(((struct lnet_hdr *)0)->msg.hello.type) == 4);
 648
 649         /* Checks for struct lnet_ni_status and related constants */
 650         CLASSERT(LNET_NI_STATUS_INVALID == 0x00000000);
 651         CLASSERT(LNET_NI_STATUS_UP == 0x15aac0de);
 652         CLASSERT(LNET_NI_STATUS_DOWN == 0xdeadface);
 653
 654         /* Checks for struct lnet_ni_status */
 655         CLASSERT((int)sizeof(struct lnet_ni_status) == 16);
 656         CLASSERT((int)offsetof(struct lnet_ni_status, ns_nid) == 0);
 657         CLASSERT((int)sizeof(((struct lnet_ni_status *)0)->ns_nid) == 8);
 658         CLASSERT((int)offsetof(struct lnet_ni_status, ns_status) == 8);
 659         CLASSERT((int)sizeof(((struct lnet_ni_status *)0)->ns_status) == 4);
 660         CLASSERT((int)offsetof(struct lnet_ni_status, ns_unused) == 12);
 661         CLASSERT((int)sizeof(((struct lnet_ni_status *)0)->ns_unused) == 4);
 662
 663         /* Checks for struct lnet_ping_info and related constants */
 664         CLASSERT(LNET_PROTO_PING_MAGIC == 0x70696E67);
 665         CLASSERT(LNET_PING_FEAT_INVAL == 0);
 666         CLASSERT(LNET_PING_FEAT_BASE == 1);
 667         CLASSERT(LNET_PING_FEAT_NI_STATUS == 2);
 668         CLASSERT(LNET_PING_FEAT_RTE_DISABLED == 4);
 669         CLASSERT(LNET_PING_FEAT_MULTI_RAIL == 8);
 670         CLASSERT(LNET_PING_FEAT_DISCOVERY == 16);
 671         CLASSERT(LNET_PING_FEAT_BITS == 31);
 672
 673         /* Checks for struct lnet_ping_info */
 674         CLASSERT((int)sizeof(struct lnet_ping_info) == 16);
 675         CLASSERT((int)offsetof(struct lnet_ping_info, pi_magic) == 0);
 676         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_magic) == 4);
 677         CLASSERT((int)offsetof(struct lnet_ping_info, pi_features) == 4);
 678         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_features) == 4);
 679         CLASSERT((int)offsetof(struct lnet_ping_info, pi_pid) == 8);
 680         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_pid) == 4);
 681         CLASSERT((int)offsetof(struct lnet_ping_info, pi_nnis) == 12);
 682         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_nnis) == 4);
 683         CLASSERT((int)offsetof(struct lnet_ping_info, pi_ni) == 16);
 684         CLASSERT((int)sizeof(((struct lnet_ping_info *)0)->pi_ni) == 0);
 685 }
 686
 687 static struct lnet_lnd *lnet_find_lnd_by_type(__u32 type)
 688 {
 689         struct lnet_lnd *lnd;
 690         struct list_head *tmp;
 691
 692         /* holding lnd mutex */
 693         list_for_each(tmp, &the_lnet.ln_lnds) {
 694                 lnd = list_entry(tmp, struct lnet_lnd, lnd_list);
 695
 696                 if (lnd->lnd_type == type)
 697                         return lnd;
 698         }
 699         return NULL;
 700 }
 701
 702 unsigned int
 703 lnet_get_lnd_timeout(void)
 704 {
 705         return lnet_lnd_timeout;
 706 }
 707 EXPORT_SYMBOL(lnet_get_lnd_timeout);
 708
 709 void
 710 lnet_register_lnd(struct lnet_lnd *lnd)
 711 {
 712         mutex_lock(&the_lnet.ln_lnd_mutex);
 713
 714         LASSERT(libcfs_isknown_lnd(lnd->lnd_type));
 715         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == NULL);
 716
 717         list_add_tail(&lnd->lnd_list, &the_lnet.ln_lnds);
 718         lnd->lnd_refcount = 0;
 719
 720         CDEBUG(D_NET, "%s LND registered\n", libcfs_lnd2str(lnd->lnd_type));
 721
 722         mutex_unlock(&the_lnet.ln_lnd_mutex);
 723 }
 724 EXPORT_SYMBOL(lnet_register_lnd);
 725
 726 void
 727 lnet_unregister_lnd(struct lnet_lnd *lnd)
 728 {
 729         mutex_lock(&the_lnet.ln_lnd_mutex);
 730
 731         LASSERT(lnet_find_lnd_by_type(lnd->lnd_type) == lnd);
 732         LASSERT(lnd->lnd_refcount == 0);
 733
 734         list_del(&lnd->lnd_list);
 735         CDEBUG(D_NET, "%s LND unregistered\n", libcfs_lnd2str(lnd->lnd_type));
 736
 737         mutex_unlock(&the_lnet.ln_lnd_mutex);
 738 }
 739 EXPORT_SYMBOL(lnet_unregister_lnd);
 740
 741 void
 742 lnet_counters_get_common(struct lnet_counters_common *common)
 743 {
 744         struct lnet_counters *ctr;
 745         int i;
 746
 747         memset(common, 0, sizeof(*common));
 748
 749         lnet_net_lock(LNET_LOCK_EX);
 750
 751         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 752                 common->lcc_msgs_max     += ctr->lct_common.lcc_msgs_max;
 753                 common->lcc_msgs_alloc   += ctr->lct_common.lcc_msgs_alloc;
 754                 common->lcc_errors       += ctr->lct_common.lcc_errors;
 755                 common->lcc_send_count   += ctr->lct_common.lcc_send_count;
 756                 common->lcc_recv_count   += ctr->lct_common.lcc_recv_count;
 757                 common->lcc_route_count  += ctr->lct_common.lcc_route_count;
 758                 common->lcc_drop_count   += ctr->lct_common.lcc_drop_count;
 759                 common->lcc_send_length  += ctr->lct_common.lcc_send_length;
 760                 common->lcc_recv_length  += ctr->lct_common.lcc_recv_length;
 761                 common->lcc_route_length += ctr->lct_common.lcc_route_length;
 762                 common->lcc_drop_length  += ctr->lct_common.lcc_drop_length;
 763         }
 764         lnet_net_unlock(LNET_LOCK_EX);
 765 }
 766 EXPORT_SYMBOL(lnet_counters_get_common);
 767
 768 void
 769 lnet_counters_get(struct lnet_counters *counters)
 770 {
 771         struct lnet_counters *ctr;
 772         struct lnet_counters_health *health = &counters->lct_health;
 773         int             i;
 774
 775         memset(counters, 0, sizeof(*counters));
 776
 777         lnet_counters_get_common(&counters->lct_common);
 778
 779         lnet_net_lock(LNET_LOCK_EX);
 780
 781         cfs_percpt_for_each(ctr, i, the_lnet.ln_counters) {
 782                 health->lch_rst_alloc    += ctr->lct_health.lch_rst_alloc;
 783                 health->lch_resend_count += ctr->lct_health.lch_resend_count;
 784                 health->lch_response_timeout_count +=
 785                                 ctr->lct_health.lch_response_timeout_count;
 786                 health->lch_local_interrupt_count +=
 787                                 ctr->lct_health.lch_local_interrupt_count;
 788                 health->lch_local_dropped_count +=
 789                                 ctr->lct_health.lch_local_dropped_count;
 790                 health->lch_local_aborted_count +=
 791                                 ctr->lct_health.lch_local_aborted_count;
 792                 health->lch_local_no_route_count +=
 793                                 ctr->lct_health.lch_local_no_route_count;
 794                 health->lch_local_timeout_count +=
 795                                 ctr->lct_health.lch_local_timeout_count;
 796                 health->lch_local_error_count +=
 797                                 ctr->lct_health.lch_local_error_count;
 798                 health->lch_remote_dropped_count +=
 799                                 ctr->lct_health.lch_remote_dropped_count;
 800                 health->lch_remote_error_count +=
 801                                 ctr->lct_health.lch_remote_error_count;
 802                 health->lch_remote_timeout_count +=
 803                                 ctr->lct_health.lch_remote_timeout_count;
 804                 health->lch_network_timeout_count +=
 805                                 ctr->lct_health.lch_network_timeout_count;
 806         }
 807         lnet_net_unlock(LNET_LOCK_EX);
 808 }
 809 EXPORT_SYMBOL(lnet_counters_get);
 810
 811 void
 812 lnet_counters_reset(void)
 813 {
 814         struct lnet_counters *counters;
 815         int             i;
 816
 817         lnet_net_lock(LNET_LOCK_EX);
 818
 819         cfs_percpt_for_each(counters, i, the_lnet.ln_counters)
 820                 memset(counters, 0, sizeof(struct lnet_counters));
 821
 822         lnet_net_unlock(LNET_LOCK_EX);
 823 }
 824
 825 static char *
 826 lnet_res_type2str(int type)
 827 {
 828         switch (type) {
 829         default:
 830                 LBUG();
 831         case LNET_COOKIE_TYPE_MD:
 832                 return "MD";
 833         case LNET_COOKIE_TYPE_ME:
 834                 return "ME";
 835         case LNET_COOKIE_TYPE_EQ:
 836                 return "EQ";
 837         }
 838 }
 839
 840 static void
 841 lnet_res_container_cleanup(struct lnet_res_container *rec)
 842 {
 843         int     count = 0;
 844
 845         if (rec->rec_type == 0) /* not set yet, it's uninitialized */
 846                 return;
 847
 848         while (!list_empty(&rec->rec_active)) {
 849                 struct list_head *e = rec->rec_active.next;
 850
 851                 list_del_init(e);
 852                 if (rec->rec_type == LNET_COOKIE_TYPE_EQ) {
 853                         lnet_eq_free(list_entry(e, struct lnet_eq, eq_list));
 854
 855                 } else if (rec->rec_type == LNET_COOKIE_TYPE_MD) {
 856                         lnet_md_free(list_entry(e, struct lnet_libmd, md_list));
 857
 858                 } else { /* NB: Active MEs should be attached on portals */
 859                         LBUG();
 860                 }
 861                 count++;
 862         }
 863
 864         if (count > 0) {
 865                 /* Found alive MD/ME/EQ, user really should unlink/free
 866                  * all of them before finalize LNet, but if someone didn't,
 867                  * we have to recycle garbage for him */
 868                 CERROR("%d active elements on exit of %s container\n",
 869                        count, lnet_res_type2str(rec->rec_type));
 870         }
 871
 872         if (rec->rec_lh_hash != NULL) {
 873                 LIBCFS_FREE(rec->rec_lh_hash,
 874                             LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 875                 rec->rec_lh_hash = NULL;
 876         }
 877
 878         rec->rec_type = 0; /* mark it as finalized */
 879 }
 880
 881 static int
 882 lnet_res_container_setup(struct lnet_res_container *rec, int cpt, int type)
 883 {
 884         int     rc = 0;
 885         int     i;
 886
 887         LASSERT(rec->rec_type == 0);
 888
 889         rec->rec_type = type;
 890         INIT_LIST_HEAD(&rec->rec_active);
 891
 892         rec->rec_lh_cookie = (cpt << LNET_COOKIE_TYPE_BITS) | type;
 893
 894         /* Arbitrary choice of hash table size */
 895         LIBCFS_CPT_ALLOC(rec->rec_lh_hash, lnet_cpt_table(), cpt,
 896                          LNET_LH_HASH_SIZE * sizeof(rec->rec_lh_hash[0]));
 897         if (rec->rec_lh_hash == NULL) {
 898                 rc = -ENOMEM;
 899                 goto out;
 900         }
 901
 902         for (i = 0; i < LNET_LH_HASH_SIZE; i++)
 903                 INIT_LIST_HEAD(&rec->rec_lh_hash[i]);
 904
 905         return 0;
 906
 907 out:
 908         CERROR("Failed to setup %s resource container\n",
 909                lnet_res_type2str(type));
 910         lnet_res_container_cleanup(rec);
 911         return rc;
 912 }
 913
 914 static void
 915 lnet_res_containers_destroy(struct lnet_res_container **recs)
 916 {
 917         struct lnet_res_container       *rec;
 918         int                             i;
 919
 920         cfs_percpt_for_each(rec, i, recs)
 921                 lnet_res_container_cleanup(rec);
 922
 923         cfs_percpt_free(recs);
 924 }
 925
 926 static struct lnet_res_container **
 927 lnet_res_containers_create(int type)
 928 {
 929         struct lnet_res_container       **recs;
 930         struct lnet_res_container       *rec;
 931         int                             rc;
 932         int                             i;
 933
 934         recs = cfs_percpt_alloc(lnet_cpt_table(), sizeof(*rec));
 935         if (recs == NULL) {
 936                 CERROR("Failed to allocate %s resource containers\n",
 937                        lnet_res_type2str(type));
 938                 return NULL;
 939         }
 940
 941         cfs_percpt_for_each(rec, i, recs) {
 942                 rc = lnet_res_container_setup(rec, i, type);
 943                 if (rc != 0) {
 944                         lnet_res_containers_destroy(recs);
 945                         return NULL;
 946                 }
 947         }
 948
 949         return recs;
 950 }
 951
 952 struct lnet_libhandle *
 953 lnet_res_lh_lookup(struct lnet_res_container *rec, __u64 cookie)
 954 {
 955         /* ALWAYS called with lnet_res_lock held */
 956         struct list_head        *head;
 957         struct lnet_libhandle   *lh;
 958         unsigned int            hash;
 959
 960         if ((cookie & LNET_COOKIE_MASK) != rec->rec_type)
 961                 return NULL;
 962
 963         hash = cookie >> (LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS);
 964         head = &rec->rec_lh_hash[hash & LNET_LH_HASH_MASK];
 965
 966         list_for_each_entry(lh, head, lh_hash_chain) {
 967                 if (lh->lh_cookie == cookie)
 968                         return lh;
 969         }
 970
 971         return NULL;
 972 }
 973
 974 void
 975 lnet_res_lh_initialize(struct lnet_res_container *rec,
 976                        struct lnet_libhandle *lh)
 977 {
 978         /* ALWAYS called with lnet_res_lock held */
 979         unsigned int    ibits = LNET_COOKIE_TYPE_BITS + LNET_CPT_BITS;
 980         unsigned int    hash;
 981
 982         lh->lh_cookie = rec->rec_lh_cookie;
 983         rec->rec_lh_cookie += 1 << ibits;
 984
 985         hash = (lh->lh_cookie >> ibits) & LNET_LH_HASH_MASK;
 986
 987         list_add(&lh->lh_hash_chain, &rec->rec_lh_hash[hash]);
 988 }
 989
 990 static int lnet_unprepare(void);
 991
 992 static int
 993 lnet_prepare(lnet_pid_t requested_pid)
 994 {
 995         /* Prepare to bring up the network */
 996         struct lnet_res_container **recs;
 997         int                       rc = 0;
 998
 999         if (requested_pid == LNET_PID_ANY) {
1000                 /* Don't instantiate LNET just for me */
1001                 return -ENETDOWN;
1002         }
1003
1004         LASSERT(the_lnet.ln_refcount == 0);
1005
1006         the_lnet.ln_routing = 0;
1007
1008         LASSERT((requested_pid & LNET_PID_USERFLAG) == 0);
1009         the_lnet.ln_pid = requested_pid;
1010
1011         INIT_LIST_HEAD(&the_lnet.ln_test_peers);
1012         INIT_LIST_HEAD(&the_lnet.ln_remote_peer_ni_list);
1013         INIT_LIST_HEAD(&the_lnet.ln_nets);
1014         INIT_LIST_HEAD(&the_lnet.ln_routers);
1015         INIT_LIST_HEAD(&the_lnet.ln_drop_rules);
1016         INIT_LIST_HEAD(&the_lnet.ln_delay_rules);
1017         INIT_LIST_HEAD(&the_lnet.ln_dc_request);
1018         INIT_LIST_HEAD(&the_lnet.ln_dc_working);
1019         INIT_LIST_HEAD(&the_lnet.ln_dc_expired);
1020         INIT_LIST_HEAD(&the_lnet.ln_mt_localNIRecovq);
1021         INIT_LIST_HEAD(&the_lnet.ln_mt_peerNIRecovq);
1022         init_waitqueue_head(&the_lnet.ln_dc_waitq);
1023
1024         rc = lnet_descriptor_setup();
1025         if (rc != 0)
1026                 goto failed;
1027
1028         rc = lnet_create_remote_nets_table();
1029         if (rc != 0)
1030                 goto failed;
1031
1032         /*
1033          * NB the interface cookie in wire handles guards against delayed
1034          * replies and ACKs appearing valid after reboot.
1035          */
1036         the_lnet.ln_interface_cookie = ktime_get_real_ns();
1037
1038         the_lnet.ln_counters = cfs_percpt_alloc(lnet_cpt_table(),
1039                                                 sizeof(struct lnet_counters));
1040         if (the_lnet.ln_counters == NULL) {
1041                 CERROR("Failed to allocate counters for LNet\n");
1042                 rc = -ENOMEM;
1043                 goto failed;
1044         }
1045
1046         rc = lnet_peer_tables_create();
1047         if (rc != 0)
1048                 goto failed;
1049
1050         rc = lnet_msg_containers_create();
1051         if (rc != 0)
1052                 goto failed;
1053
1054         rc = lnet_res_container_setup(&the_lnet.ln_eq_container, 0,
1055                                       LNET_COOKIE_TYPE_EQ);
1056         if (rc != 0)
1057                 goto failed;
1058
1059         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_ME);
1060         if (recs == NULL) {
1061                 rc = -ENOMEM;
1062                 goto failed;
1063         }
1064
1065         the_lnet.ln_me_containers = recs;
1066
1067         recs = lnet_res_containers_create(LNET_COOKIE_TYPE_MD);
1068         if (recs == NULL) {
1069                 rc = -ENOMEM;
1070                 goto failed;
1071         }
1072
1073         the_lnet.ln_md_containers = recs;
1074
1075         rc = lnet_portals_create();
1076         if (rc != 0) {
1077                 CERROR("Failed to create portals for LNet: %d\n", rc);
1078                 goto failed;
1079         }
1080
1081         return 0;
1082
1083  failed:
1084         lnet_unprepare();
1085         return rc;
1086 }
1087
1088 static int
1089 lnet_unprepare (void)
1090 {
1091         /* NB no LNET_LOCK since this is the last reference.  All LND instances
1092          * have shut down already, so it is safe to unlink and free all
1093          * descriptors, even those that appear committed to a network op (eg MD
1094          * with non-zero pending count) */
1095
1096         lnet_fail_nid(LNET_NID_ANY, 0);
1097
1098         LASSERT(the_lnet.ln_refcount == 0);
1099         LASSERT(list_empty(&the_lnet.ln_test_peers));
1100         LASSERT(list_empty(&the_lnet.ln_nets));
1101
1102         lnet_portals_destroy();
1103
1104         if (the_lnet.ln_md_containers != NULL) {
1105                 lnet_res_containers_destroy(the_lnet.ln_md_containers);
1106                 the_lnet.ln_md_containers = NULL;
1107         }
1108
1109         if (the_lnet.ln_me_containers != NULL) {
1110                 lnet_res_containers_destroy(the_lnet.ln_me_containers);
1111                 the_lnet.ln_me_containers = NULL;
1112         }
1113
1114         lnet_res_container_cleanup(&the_lnet.ln_eq_container);
1115
1116         lnet_msg_containers_destroy();
1117         lnet_peer_uninit();
1118         lnet_rtrpools_free(0);
1119
1120         if (the_lnet.ln_counters != NULL) {
1121                 cfs_percpt_free(the_lnet.ln_counters);
1122                 the_lnet.ln_counters = NULL;
1123         }
1124         lnet_destroy_remote_nets_table();
1125         lnet_descriptor_cleanup();
1126
1127         return 0;
1128 }
1129
1130 struct lnet_ni  *
1131 lnet_net2ni_locked(__u32 net_id, int cpt)
1132 {
1133         struct lnet_ni   *ni;
1134         struct lnet_net  *net;
1135
1136         LASSERT(cpt != LNET_LOCK_EX);
1137
1138         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1139                 if (net->net_id == net_id) {
1140                         ni = list_entry(net->net_ni_list.next, struct lnet_ni,
1141                                         ni_netlist);
1142                         return ni;
1143                 }
1144         }
1145
1146         return NULL;
1147 }
1148
1149 struct lnet_ni *
1150 lnet_net2ni_addref(__u32 net)
1151 {
1152         struct lnet_ni *ni;
1153
1154         lnet_net_lock(0);
1155         ni = lnet_net2ni_locked(net, 0);
1156         if (ni)
1157                 lnet_ni_addref_locked(ni, 0);
1158         lnet_net_unlock(0);
1159
1160         return ni;
1161 }
1162 EXPORT_SYMBOL(lnet_net2ni_addref);
1163
1164 struct lnet_net *
1165 lnet_get_net_locked(__u32 net_id)
1166 {
1167         struct lnet_net  *net;
1168
1169         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1170                 if (net->net_id == net_id)
1171                         return net;
1172         }
1173
1174         return NULL;
1175 }
1176
1177 unsigned int
1178 lnet_nid_cpt_hash(lnet_nid_t nid, unsigned int number)
1179 {
1180         __u64           key = nid;
1181         unsigned int    val;
1182
1183         LASSERT(number >= 1 && number <= LNET_CPT_NUMBER);
1184
1185         if (number == 1)
1186                 return 0;
1187
1188         val = hash_long(key, LNET_CPT_BITS);
1189         /* NB: LNET_CP_NUMBER doesn't have to be PO2 */
1190         if (val < number)
1191                 return val;
1192
1193         return (unsigned int)(key + val + (val >> 1)) % number;
1194 }
1195
1196 int
1197 lnet_cpt_of_nid_locked(lnet_nid_t nid, struct lnet_ni *ni)
1198 {
1199         struct lnet_net *net;
1200
1201         /* must called with hold of lnet_net_lock */
1202         if (LNET_CPT_NUMBER == 1)
1203                 return 0; /* the only one */
1204
1205         /*
1206          * If NI is provided then use the CPT identified in the NI cpt
1207          * list if one exists. If one doesn't exist, then that NI is
1208          * associated with all CPTs and it follows that the net it belongs
1209          * to is implicitly associated with all CPTs, so just hash the nid
1210          * and return that.
1211          */
1212         if (ni != NULL) {
1213                 if (ni->ni_cpts != NULL)
1214                         return ni->ni_cpts[lnet_nid_cpt_hash(nid,
1215                                                              ni->ni_ncpts)];
1216                 else
1217                         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1218         }
1219
1220         /* no NI provided so look at the net */
1221         net = lnet_get_net_locked(LNET_NIDNET(nid));
1222
1223         if (net != NULL && net->net_cpts != NULL) {
1224                 return net->net_cpts[lnet_nid_cpt_hash(nid, net->net_ncpts)];
1225         }
1226
1227         return lnet_nid_cpt_hash(nid, LNET_CPT_NUMBER);
1228 }
1229
1230 int
1231 lnet_cpt_of_nid(lnet_nid_t nid, struct lnet_ni *ni)
1232 {
1233         int     cpt;
1234         int     cpt2;
1235
1236         if (LNET_CPT_NUMBER == 1)
1237                 return 0; /* the only one */
1238
1239         cpt = lnet_net_lock_current();
1240
1241         cpt2 = lnet_cpt_of_nid_locked(nid, ni);
1242
1243         lnet_net_unlock(cpt);
1244
1245         return cpt2;
1246 }
1247 EXPORT_SYMBOL(lnet_cpt_of_nid);
1248
1249 int
1250 lnet_islocalnet(__u32 net_id)
1251 {
1252         struct lnet_net *net;
1253         int             cpt;
1254         bool            local;
1255
1256         cpt = lnet_net_lock_current();
1257
1258         net = lnet_get_net_locked(net_id);
1259
1260         local = net != NULL;
1261
1262         lnet_net_unlock(cpt);
1263
1264         return local;
1265 }
1266
1267 struct lnet_ni  *
1268 lnet_nid2ni_locked(lnet_nid_t nid, int cpt)
1269 {
1270         struct lnet_net  *net;
1271         struct lnet_ni   *ni;
1272
1273         LASSERT(cpt != LNET_LOCK_EX);
1274
1275         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1276                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1277                         if (ni->ni_nid == nid)
1278                                 return ni;
1279                 }
1280         }
1281
1282         return NULL;
1283 }
1284
1285 struct lnet_ni *
1286 lnet_nid2ni_addref(lnet_nid_t nid)
1287 {
1288         struct lnet_ni *ni;
1289
1290         lnet_net_lock(0);
1291         ni = lnet_nid2ni_locked(nid, 0);
1292         if (ni)
1293                 lnet_ni_addref_locked(ni, 0);
1294         lnet_net_unlock(0);
1295
1296         return ni;
1297 }
1298 EXPORT_SYMBOL(lnet_nid2ni_addref);
1299
1300 int
1301 lnet_islocalnid(lnet_nid_t nid)
1302 {
1303         struct lnet_ni  *ni;
1304         int             cpt;
1305
1306         cpt = lnet_net_lock_current();
1307         ni = lnet_nid2ni_locked(nid, cpt);
1308         lnet_net_unlock(cpt);
1309
1310         return ni != NULL;
1311 }
1312
1313 int
1314 lnet_count_acceptor_nets(void)
1315 {
1316         /* Return the # of NIs that need the acceptor. */
1317         int              count = 0;
1318         struct lnet_net  *net;
1319         int              cpt;
1320
1321         cpt = lnet_net_lock_current();
1322         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1323                 /* all socklnd type networks should have the acceptor
1324                  * thread started */
1325                 if (net->net_lnd->lnd_accept != NULL)
1326                         count++;
1327         }
1328
1329         lnet_net_unlock(cpt);
1330
1331         return count;
1332 }
1333
1334 struct lnet_ping_buffer *
1335 lnet_ping_buffer_alloc(int nnis, gfp_t gfp)
1336 {
1337         struct lnet_ping_buffer *pbuf;
1338
1339         LIBCFS_ALLOC_GFP(pbuf, LNET_PING_BUFFER_SIZE(nnis), gfp);
1340         if (pbuf) {
1341                 pbuf->pb_nnis = nnis;
1342                 atomic_set(&pbuf->pb_refcnt, 1);
1343         }
1344
1345         return pbuf;
1346 }
1347
1348 void
1349 lnet_ping_buffer_free(struct lnet_ping_buffer *pbuf)
1350 {
1351         LASSERT(lnet_ping_buffer_numref(pbuf) == 0);
1352         LIBCFS_FREE(pbuf, LNET_PING_BUFFER_SIZE(pbuf->pb_nnis));
1353 }
1354
1355 static struct lnet_ping_buffer *
1356 lnet_ping_target_create(int nnis)
1357 {
1358         struct lnet_ping_buffer *pbuf;
1359
1360         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1361         if (pbuf == NULL) {
1362                 CERROR("Can't allocate ping source [%d]\n", nnis);
1363                 return NULL;
1364         }
1365
1366         pbuf->pb_info.pi_nnis = nnis;
1367         pbuf->pb_info.pi_pid = the_lnet.ln_pid;
1368         pbuf->pb_info.pi_magic = LNET_PROTO_PING_MAGIC;
1369         pbuf->pb_info.pi_features =
1370                 LNET_PING_FEAT_NI_STATUS | LNET_PING_FEAT_MULTI_RAIL;
1371
1372         return pbuf;
1373 }
1374
1375 static inline int
1376 lnet_get_net_ni_count_locked(struct lnet_net *net)
1377 {
1378         struct lnet_ni  *ni;
1379         int             count = 0;
1380
1381         list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1382                 count++;
1383
1384         return count;
1385 }
1386
1387 static inline int
1388 lnet_get_net_ni_count_pre(struct lnet_net *net)
1389 {
1390         struct lnet_ni  *ni;
1391         int             count = 0;
1392
1393         list_for_each_entry(ni, &net->net_ni_added, ni_netlist)
1394                 count++;
1395
1396         return count;
1397 }
1398
1399 static inline int
1400 lnet_get_ni_count(void)
1401 {
1402         struct lnet_ni  *ni;
1403         struct lnet_net *net;
1404         int             count = 0;
1405
1406         lnet_net_lock(0);
1407
1408         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1409                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist)
1410                         count++;
1411         }
1412
1413         lnet_net_unlock(0);
1414
1415         return count;
1416 }
1417
1418 int
1419 lnet_ping_info_validate(struct lnet_ping_info *pinfo)
1420 {
1421         if (!pinfo)
1422                 return -EINVAL;
1423         if (pinfo->pi_magic != LNET_PROTO_PING_MAGIC)
1424                 return -EPROTO;
1425         if (!(pinfo->pi_features & LNET_PING_FEAT_NI_STATUS))
1426                 return -EPROTO;
1427         /* Loopback is guaranteed to be present */
1428         if (pinfo->pi_nnis < 1 || pinfo->pi_nnis > lnet_interfaces_max)
1429                 return -ERANGE;
1430         if (LNET_NETTYP(LNET_NIDNET(LNET_PING_INFO_LONI(pinfo))) != LOLND)
1431                 return -EPROTO;
1432         return 0;
1433 }
1434
1435 static void
1436 lnet_ping_target_destroy(void)
1437 {
1438         struct lnet_net *net;
1439         struct lnet_ni  *ni;
1440
1441         lnet_net_lock(LNET_LOCK_EX);
1442
1443         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1444                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1445                         lnet_ni_lock(ni);
1446                         ni->ni_status = NULL;
1447                         lnet_ni_unlock(ni);
1448                 }
1449         }
1450
1451         lnet_ping_buffer_decref(the_lnet.ln_ping_target);
1452         the_lnet.ln_ping_target = NULL;
1453
1454         lnet_net_unlock(LNET_LOCK_EX);
1455 }
1456
1457 static void
1458 lnet_ping_target_event_handler(struct lnet_event *event)
1459 {
1460         struct lnet_ping_buffer *pbuf = event->md.user_ptr;
1461
1462         if (event->unlinked)
1463                 lnet_ping_buffer_decref(pbuf);
1464 }
1465
1466 static int
1467 lnet_ping_target_setup(struct lnet_ping_buffer **ppbuf,
1468                        struct lnet_handle_md *ping_mdh,
1469                        int ni_count, bool set_eq)
1470 {
1471         struct lnet_process_id id = {
1472                 .nid = LNET_NID_ANY,
1473                 .pid = LNET_PID_ANY
1474         };
1475         struct lnet_handle_me me_handle;
1476         struct lnet_md md = { NULL };
1477         int rc, rc2;
1478
1479         if (set_eq) {
1480                 rc = LNetEQAlloc(0, lnet_ping_target_event_handler,
1481                                  &the_lnet.ln_ping_target_eq);
1482                 if (rc != 0) {
1483                         CERROR("Can't allocate ping buffer EQ: %d\n", rc);
1484                         return rc;
1485                 }
1486         }
1487
1488         *ppbuf = lnet_ping_target_create(ni_count);
1489         if (*ppbuf == NULL) {
1490                 rc = -ENOMEM;
1491                 goto fail_free_eq;
1492         }
1493
1494         /* Ping target ME/MD */
1495         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1496                           LNET_PROTO_PING_MATCHBITS, 0,
1497                           LNET_UNLINK, LNET_INS_AFTER,
1498                           &me_handle);
1499         if (rc != 0) {
1500                 CERROR("Can't create ping target ME: %d\n", rc);
1501                 goto fail_decref_ping_buffer;
1502         }
1503
1504         /* initialize md content */
1505         md.start     = &(*ppbuf)->pb_info;
1506         md.length    = LNET_PING_INFO_SIZE((*ppbuf)->pb_nnis);
1507         md.threshold = LNET_MD_THRESH_INF;
1508         md.max_size  = 0;
1509         md.options   = LNET_MD_OP_GET | LNET_MD_TRUNCATE |
1510                        LNET_MD_MANAGE_REMOTE;
1511         md.eq_handle = the_lnet.ln_ping_target_eq;
1512         md.user_ptr  = *ppbuf;
1513
1514         rc = LNetMDAttach(me_handle, md, LNET_RETAIN, ping_mdh);
1515         if (rc != 0) {
1516                 CERROR("Can't attach ping target MD: %d\n", rc);
1517                 goto fail_unlink_ping_me;
1518         }
1519         lnet_ping_buffer_addref(*ppbuf);
1520
1521         return 0;
1522
1523 fail_unlink_ping_me:
1524         rc2 = LNetMEUnlink(me_handle);
1525         LASSERT(rc2 == 0);
1526 fail_decref_ping_buffer:
1527         LASSERT(lnet_ping_buffer_numref(*ppbuf) == 1);
1528         lnet_ping_buffer_decref(*ppbuf);
1529         *ppbuf = NULL;
1530 fail_free_eq:
1531         if (set_eq) {
1532                 rc2 = LNetEQFree(the_lnet.ln_ping_target_eq);
1533                 LASSERT(rc2 == 0);
1534         }
1535         return rc;
1536 }
1537
1538 static void
1539 lnet_ping_md_unlink(struct lnet_ping_buffer *pbuf,
1540                     struct lnet_handle_md *ping_mdh)
1541 {
1542         sigset_t        blocked = cfs_block_allsigs();
1543
1544         LNetMDUnlink(*ping_mdh);
1545         LNetInvalidateMDHandle(ping_mdh);
1546
1547         /* NB the MD could be busy; this just starts the unlink */
1548         while (lnet_ping_buffer_numref(pbuf) > 1) {
1549                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
1550                 set_current_state(TASK_UNINTERRUPTIBLE);
1551                 schedule_timeout(cfs_time_seconds(1));
1552         }
1553
1554         cfs_restore_sigs(blocked);
1555 }
1556
1557 static void
1558 lnet_ping_target_install_locked(struct lnet_ping_buffer *pbuf)
1559 {
1560         struct lnet_ni          *ni;
1561         struct lnet_net         *net;
1562         struct lnet_ni_status *ns;
1563         int                     i;
1564         int                     rc;
1565
1566         i = 0;
1567         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
1568                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
1569                         LASSERT(i < pbuf->pb_nnis);
1570
1571                         ns = &pbuf->pb_info.pi_ni[i];
1572
1573                         ns->ns_nid = ni->ni_nid;
1574
1575                         lnet_ni_lock(ni);
1576                         ns->ns_status = (ni->ni_status != NULL) ?
1577                                          ni->ni_status->ns_status :
1578                                                 LNET_NI_STATUS_UP;
1579                         ni->ni_status = ns;
1580                         lnet_ni_unlock(ni);
1581
1582                         i++;
1583                 }
1584         }
1585         /*
1586          * We (ab)use the ns_status of the loopback interface to
1587          * transmit the sequence number. The first interface listed
1588          * must be the loopback interface.
1589          */
1590         rc = lnet_ping_info_validate(&pbuf->pb_info);
1591         if (rc) {
1592                 LCONSOLE_EMERG("Invalid ping target: %d\n", rc);
1593                 LBUG();
1594         }
1595         LNET_PING_BUFFER_SEQNO(pbuf) =
1596                 atomic_inc_return(&the_lnet.ln_ping_target_seqno);
1597 }
1598
1599 static void
1600 lnet_ping_target_update(struct lnet_ping_buffer *pbuf,
1601                         struct lnet_handle_md ping_mdh)
1602 {
1603         struct lnet_ping_buffer *old_pbuf = NULL;
1604         struct lnet_handle_md old_ping_md;
1605
1606         /* switch the NIs to point to the new ping info created */
1607         lnet_net_lock(LNET_LOCK_EX);
1608
1609         if (!the_lnet.ln_routing)
1610                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_RTE_DISABLED;
1611         if (!lnet_peer_discovery_disabled)
1612                 pbuf->pb_info.pi_features |= LNET_PING_FEAT_DISCOVERY;
1613
1614         /* Ensure only known feature bits have been set. */
1615         LASSERT(pbuf->pb_info.pi_features & LNET_PING_FEAT_BITS);
1616         LASSERT(!(pbuf->pb_info.pi_features & ~LNET_PING_FEAT_BITS));
1617
1618         lnet_ping_target_install_locked(pbuf);
1619
1620         if (the_lnet.ln_ping_target) {
1621                 old_pbuf = the_lnet.ln_ping_target;
1622                 old_ping_md = the_lnet.ln_ping_target_md;
1623         }
1624         the_lnet.ln_ping_target_md = ping_mdh;
1625         the_lnet.ln_ping_target = pbuf;
1626
1627         lnet_net_unlock(LNET_LOCK_EX);
1628
1629         if (old_pbuf) {
1630                 /* unlink and free the old ping info */
1631                 lnet_ping_md_unlink(old_pbuf, &old_ping_md);
1632                 lnet_ping_buffer_decref(old_pbuf);
1633         }
1634
1635         lnet_push_update_to_peers(0);
1636 }
1637
1638 static void
1639 lnet_ping_target_fini(void)
1640 {
1641         int             rc;
1642
1643         lnet_ping_md_unlink(the_lnet.ln_ping_target,
1644                             &the_lnet.ln_ping_target_md);
1645
1646         rc = LNetEQFree(the_lnet.ln_ping_target_eq);
1647         LASSERT(rc == 0);
1648
1649         lnet_ping_target_destroy();
1650 }
1651
1652 /* Resize the push target. */
1653 int lnet_push_target_resize(void)
1654 {
1655         struct lnet_process_id id = { LNET_NID_ANY, LNET_PID_ANY };
1656         struct lnet_md md = { NULL };
1657         struct lnet_handle_me meh;
1658         struct lnet_handle_md mdh;
1659         struct lnet_handle_md old_mdh;
1660         struct lnet_ping_buffer *pbuf;
1661         struct lnet_ping_buffer *old_pbuf;
1662         int nnis = the_lnet.ln_push_target_nnis;
1663         int rc;
1664
1665         if (nnis <= 0) {
1666                 rc = -EINVAL;
1667                 goto fail_return;
1668         }
1669 again:
1670         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
1671         if (!pbuf) {
1672                 rc = -ENOMEM;
1673                 goto fail_return;
1674         }
1675
1676         rc = LNetMEAttach(LNET_RESERVED_PORTAL, id,
1677                           LNET_PROTO_PING_MATCHBITS, 0,
1678                           LNET_UNLINK, LNET_INS_AFTER,
1679                           &meh);
1680         if (rc) {
1681                 CERROR("Can't create push target ME: %d\n", rc);
1682                 goto fail_decref_pbuf;
1683         }
1684
1685         /* initialize md content */
1686         md.start     = &pbuf->pb_info;
1687         md.length    = LNET_PING_INFO_SIZE(nnis);
1688         md.threshold = LNET_MD_THRESH_INF;
1689         md.max_size  = 0;
1690         md.options   = LNET_MD_OP_PUT | LNET_MD_TRUNCATE |
1691                        LNET_MD_MANAGE_REMOTE;
1692         md.user_ptr  = pbuf;
1693         md.eq_handle = the_lnet.ln_push_target_eq;
1694
1695         rc = LNetMDAttach(meh, md, LNET_RETAIN, &mdh);
1696         if (rc) {
1697                 CERROR("Can't attach push MD: %d\n", rc);
1698                 goto fail_unlink_meh;
1699         }
1700         lnet_ping_buffer_addref(pbuf);
1701
1702         lnet_net_lock(LNET_LOCK_EX);
1703         old_pbuf = the_lnet.ln_push_target;
1704         old_mdh = the_lnet.ln_push_target_md;
1705         the_lnet.ln_push_target = pbuf;
1706         the_lnet.ln_push_target_md = mdh;
1707         lnet_net_unlock(LNET_LOCK_EX);
1708
1709         if (old_pbuf) {
1710                 LNetMDUnlink(old_mdh);
1711                 lnet_ping_buffer_decref(old_pbuf);
1712         }
1713
1714         if (nnis < the_lnet.ln_push_target_nnis)
1715                 goto again;
1716
1717         CDEBUG(D_NET, "nnis %d success\n", nnis);
1718
1719         return 0;
1720
1721 fail_unlink_meh:
1722         LNetMEUnlink(meh);
1723 fail_decref_pbuf:
1724         lnet_ping_buffer_decref(pbuf);
1725 fail_return:
1726         CDEBUG(D_NET, "nnis %d error %d\n", nnis, rc);
1727         return rc;
1728 }
1729
1730 static void lnet_push_target_event_handler(struct lnet_event *ev)
1731 {
1732         struct lnet_ping_buffer *pbuf = ev->md.user_ptr;
1733
1734         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC))
1735                 lnet_swap_pinginfo(pbuf);
1736
1737         lnet_peer_push_event(ev);
1738         if (ev->unlinked)
1739                 lnet_ping_buffer_decref(pbuf);
1740 }
1741
1742 /* Initialize the push target. */
1743 static int lnet_push_target_init(void)
1744 {
1745         int rc;
1746
1747         if (the_lnet.ln_push_target)
1748                 return -EALREADY;
1749
1750         rc = LNetEQAlloc(0, lnet_push_target_event_handler,
1751                          &the_lnet.ln_push_target_eq);
1752         if (rc) {
1753                 CERROR("Can't allocated push target EQ: %d\n", rc);
1754                 return rc;
1755         }
1756
1757         /* Start at the required minimum, we'll enlarge if required. */
1758         the_lnet.ln_push_target_nnis = LNET_INTERFACES_MIN;
1759
1760         rc = lnet_push_target_resize();
1761
1762         if (rc) {
1763                 LNetEQFree(the_lnet.ln_push_target_eq);
1764                 LNetInvalidateEQHandle(&the_lnet.ln_push_target_eq);
1765         }
1766
1767         return rc;
1768 }
1769
1770 /* Clean up the push target. */
1771 static void lnet_push_target_fini(void)
1772 {
1773         if (!the_lnet.ln_push_target)
1774                 return;
1775
1776         /* Unlink and invalidate to prevent new references. */
1777         LNetMDUnlink(the_lnet.ln_push_target_md);
1778         LNetInvalidateMDHandle(&the_lnet.ln_push_target_md);
1779
1780         /* Wait for the unlink to complete. */
1781         while (lnet_ping_buffer_numref(the_lnet.ln_push_target) > 1) {
1782                 CDEBUG(D_NET, "Still waiting for ping data MD to unlink\n");
1783                 set_current_state(TASK_UNINTERRUPTIBLE);
1784                 schedule_timeout(cfs_time_seconds(1));
1785         }
1786
1787         lnet_ping_buffer_decref(the_lnet.ln_push_target);
1788         the_lnet.ln_push_target = NULL;
1789         the_lnet.ln_push_target_nnis = 0;
1790
1791         LNetEQFree(the_lnet.ln_push_target_eq);
1792         LNetInvalidateEQHandle(&the_lnet.ln_push_target_eq);
1793 }
1794
1795 static int
1796 lnet_ni_tq_credits(struct lnet_ni *ni)
1797 {
1798         int     credits;
1799
1800         LASSERT(ni->ni_ncpts >= 1);
1801
1802         if (ni->ni_ncpts == 1)
1803                 return ni->ni_net->net_tunables.lct_max_tx_credits;
1804
1805         credits = ni->ni_net->net_tunables.lct_max_tx_credits / ni->ni_ncpts;
1806         credits = max(credits, 8 * ni->ni_net->net_tunables.lct_peer_tx_credits);
1807         credits = min(credits, ni->ni_net->net_tunables.lct_max_tx_credits);
1808
1809         return credits;
1810 }
1811
1812 static void
1813 lnet_ni_unlink_locked(struct lnet_ni *ni)
1814 {
1815         /* move it to zombie list and nobody can find it anymore */
1816         LASSERT(!list_empty(&ni->ni_netlist));
1817         list_move(&ni->ni_netlist, &ni->ni_net->net_ni_zombie);
1818         lnet_ni_decref_locked(ni, 0);
1819 }
1820
1821 static void
1822 lnet_clear_zombies_nis_locked(struct lnet_net *net)
1823 {
1824         int             i;
1825         int             islo;
1826         struct lnet_ni  *ni;
1827         struct list_head *zombie_list = &net->net_ni_zombie;
1828
1829         /*
1830          * Now wait for the NIs I just nuked to show up on the zombie
1831          * list and shut them down in guaranteed thread context
1832          */
1833         i = 2;
1834         while (!list_empty(zombie_list)) {
1835                 int     *ref;
1836                 int     j;
1837
1838                 ni = list_entry(zombie_list->next,
1839                                 struct lnet_ni, ni_netlist);
1840                 list_del_init(&ni->ni_netlist);
1841                 /* the ni should be in deleting state. If it's not it's
1842                  * a bug */
1843                 LASSERT(ni->ni_state & LNET_NI_STATE_DELETING);
1844                 cfs_percpt_for_each(ref, j, ni->ni_refs) {
1845                         if (*ref == 0)
1846                                 continue;
1847                         /* still busy, add it back to zombie list */
1848                         list_add(&ni->ni_netlist, zombie_list);
1849                         break;
1850                 }
1851
1852                 if (!list_empty(&ni->ni_netlist)) {
1853                         lnet_net_unlock(LNET_LOCK_EX);
1854                         ++i;
1855                         if ((i & (-i)) == i) {
1856                                 CDEBUG(D_WARNING,
1857                                        "Waiting for zombie LNI %s\n",
1858                                        libcfs_nid2str(ni->ni_nid));
1859                         }
1860                         set_current_state(TASK_UNINTERRUPTIBLE);
1861                         schedule_timeout(cfs_time_seconds(1));
1862                         lnet_net_lock(LNET_LOCK_EX);
1863                         continue;
1864                 }
1865
1866                 lnet_net_unlock(LNET_LOCK_EX);
1867
1868                 islo = ni->ni_net->net_lnd->lnd_type == LOLND;
1869
1870                 LASSERT(!in_interrupt());
1871                 (net->net_lnd->lnd_shutdown)(ni);
1872
1873                 if (!islo)
1874                         CDEBUG(D_LNI, "Removed LNI %s\n",
1875                               libcfs_nid2str(ni->ni_nid));
1876
1877                 lnet_ni_free(ni);
1878                 i = 2;
1879                 lnet_net_lock(LNET_LOCK_EX);
1880         }
1881 }
1882
1883 /* shutdown down the NI and release refcount */
1884 static void
1885 lnet_shutdown_lndni(struct lnet_ni *ni)
1886 {
1887         int i;
1888         struct lnet_net *net = ni->ni_net;
1889
1890         lnet_net_lock(LNET_LOCK_EX);
1891         lnet_ni_lock(ni);
1892         ni->ni_state |= LNET_NI_STATE_DELETING;
1893         ni->ni_state &= ~LNET_NI_STATE_ACTIVE;
1894         lnet_ni_unlock(ni);
1895         lnet_ni_unlink_locked(ni);
1896         lnet_incr_dlc_seq();
1897         lnet_net_unlock(LNET_LOCK_EX);
1898
1899         /* clear messages for this NI on the lazy portal */
1900         for (i = 0; i < the_lnet.ln_nportals; i++)
1901                 lnet_clear_lazy_portal(ni, i, "Shutting down NI");
1902
1903         lnet_net_lock(LNET_LOCK_EX);
1904         lnet_clear_zombies_nis_locked(net);
1905         lnet_net_unlock(LNET_LOCK_EX);
1906 }
1907
1908 static void
1909 lnet_shutdown_lndnet(struct lnet_net *net)
1910 {
1911         struct lnet_ni *ni;
1912
1913         lnet_net_lock(LNET_LOCK_EX);
1914
1915         net->net_state = LNET_NET_STATE_DELETING;
1916
1917         list_del_init(&net->net_list);
1918
1919         while (!list_empty(&net->net_ni_list)) {
1920                 ni = list_entry(net->net_ni_list.next,
1921                                 struct lnet_ni, ni_netlist);
1922                 lnet_net_unlock(LNET_LOCK_EX);
1923                 lnet_shutdown_lndni(ni);
1924                 lnet_net_lock(LNET_LOCK_EX);
1925         }
1926
1927         lnet_net_unlock(LNET_LOCK_EX);
1928
1929         /* Do peer table cleanup for this net */
1930         lnet_peer_tables_cleanup(net);
1931
1932         lnet_net_lock(LNET_LOCK_EX);
1933         /*
1934          * decrement ref count on lnd only when the entire network goes
1935          * away
1936          */
1937         net->net_lnd->lnd_refcount--;
1938
1939         lnet_net_unlock(LNET_LOCK_EX);
1940
1941         lnet_net_free(net);
1942 }
1943
1944 static void
1945 lnet_shutdown_lndnets(void)
1946 {
1947         struct lnet_net *net;
1948         struct list_head resend;
1949         struct lnet_msg *msg, *tmp;
1950
1951         INIT_LIST_HEAD(&resend);
1952
1953         /* NB called holding the global mutex */
1954
1955         /* All quiet on the API front */
1956         LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
1957         LASSERT(the_lnet.ln_refcount == 0);
1958
1959         lnet_net_lock(LNET_LOCK_EX);
1960         the_lnet.ln_state = LNET_STATE_STOPPING;
1961
1962         while (!list_empty(&the_lnet.ln_nets)) {
1963                 /*
1964                  * move the nets to the zombie list to avoid them being
1965                  * picked up for new work. LONET is also included in the
1966                  * Nets that will be moved to the zombie list
1967                  */
1968                 net = list_entry(the_lnet.ln_nets.next,
1969                                  struct lnet_net, net_list);
1970                 list_move(&net->net_list, &the_lnet.ln_net_zombie);
1971         }
1972
1973         /* Drop the cached loopback Net. */
1974         if (the_lnet.ln_loni != NULL) {
1975                 lnet_ni_decref_locked(the_lnet.ln_loni, 0);
1976                 the_lnet.ln_loni = NULL;
1977         }
1978         lnet_net_unlock(LNET_LOCK_EX);
1979
1980         /* iterate through the net zombie list and delete each net */
1981         while (!list_empty(&the_lnet.ln_net_zombie)) {
1982                 net = list_entry(the_lnet.ln_net_zombie.next,
1983                                  struct lnet_net, net_list);
1984                 lnet_shutdown_lndnet(net);
1985         }
1986
1987         spin_lock(&the_lnet.ln_msg_resend_lock);
1988         list_splice(&the_lnet.ln_msg_resend, &resend);
1989         spin_unlock(&the_lnet.ln_msg_resend_lock);
1990
1991         list_for_each_entry_safe(msg, tmp, &resend, msg_list) {
1992                 list_del_init(&msg->msg_list);
1993                 msg->msg_no_resend = true;
1994                 lnet_finalize(msg, -ECANCELED);
1995         }
1996
1997         lnet_net_lock(LNET_LOCK_EX);
1998         the_lnet.ln_state = LNET_STATE_SHUTDOWN;
1999         lnet_net_unlock(LNET_LOCK_EX);
2000 }
2001
2002 static int
2003 lnet_startup_lndni(struct lnet_ni *ni, struct lnet_lnd_tunables *tun)
2004 {
2005         int                     rc = -EINVAL;
2006         struct lnet_tx_queue    *tq;
2007         int                     i;
2008         struct lnet_net         *net = ni->ni_net;
2009
2010         mutex_lock(&the_lnet.ln_lnd_mutex);
2011
2012         if (tun) {
2013                 memcpy(&ni->ni_lnd_tunables, tun, sizeof(*tun));
2014                 ni->ni_lnd_tunables_set = true;
2015         }
2016
2017         rc = (net->net_lnd->lnd_startup)(ni);
2018
2019         mutex_unlock(&the_lnet.ln_lnd_mutex);
2020
2021         if (rc != 0) {
2022                 LCONSOLE_ERROR_MSG(0x105, "Error %d starting up LNI %s\n",
2023                                    rc, libcfs_lnd2str(net->net_lnd->lnd_type));
2024                 lnet_net_lock(LNET_LOCK_EX);
2025                 net->net_lnd->lnd_refcount--;
2026                 lnet_net_unlock(LNET_LOCK_EX);
2027                 goto failed0;
2028         }
2029
2030         lnet_ni_lock(ni);
2031         ni->ni_state |= LNET_NI_STATE_ACTIVE;
2032         ni->ni_state &= ~LNET_NI_STATE_INIT;
2033         lnet_ni_unlock(ni);
2034
2035         /* We keep a reference on the loopback net through the loopback NI */
2036         if (net->net_lnd->lnd_type == LOLND) {
2037                 lnet_ni_addref(ni);
2038                 LASSERT(the_lnet.ln_loni == NULL);
2039                 the_lnet.ln_loni = ni;
2040                 ni->ni_net->net_tunables.lct_peer_tx_credits = 0;
2041                 ni->ni_net->net_tunables.lct_peer_rtr_credits = 0;
2042                 ni->ni_net->net_tunables.lct_max_tx_credits = 0;
2043                 ni->ni_net->net_tunables.lct_peer_timeout = 0;
2044                 return 0;
2045         }
2046
2047         if (ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ||
2048             ni->ni_net->net_tunables.lct_max_tx_credits == 0) {
2049                 LCONSOLE_ERROR_MSG(0x107, "LNI %s has no %scredits\n",
2050                                    libcfs_lnd2str(net->net_lnd->lnd_type),
2051                                    ni->ni_net->net_tunables.lct_peer_tx_credits == 0 ?
2052                                         "" : "per-peer ");
2053                 /* shutdown the NI since if we get here then it must've already
2054                  * been started
2055                  */
2056                 lnet_shutdown_lndni(ni);
2057                 return -EINVAL;
2058         }
2059
2060         cfs_percpt_for_each(tq, i, ni->ni_tx_queues) {
2061                 tq->tq_credits_min =
2062                 tq->tq_credits_max =
2063                 tq->tq_credits = lnet_ni_tq_credits(ni);
2064         }
2065
2066         atomic_set(&ni->ni_tx_credits,
2067                    lnet_ni_tq_credits(ni) * ni->ni_ncpts);
2068         atomic_set(&ni->ni_healthv, LNET_MAX_HEALTH_VALUE);
2069
2070         CDEBUG(D_LNI, "Added LNI %s [%d/%d/%d/%d]\n",
2071                 libcfs_nid2str(ni->ni_nid),
2072                 ni->ni_net->net_tunables.lct_peer_tx_credits,
2073                 lnet_ni_tq_credits(ni) * LNET_CPT_NUMBER,
2074                 ni->ni_net->net_tunables.lct_peer_rtr_credits,
2075                 ni->ni_net->net_tunables.lct_peer_timeout);
2076
2077         return 0;
2078 failed0:
2079         lnet_ni_free(ni);
2080         return rc;
2081 }
2082
2083 static int
2084 lnet_startup_lndnet(struct lnet_net *net, struct lnet_lnd_tunables *tun)
2085 {
2086         struct lnet_ni *ni;
2087         struct lnet_net *net_l = NULL;
2088         struct list_head        local_ni_list;
2089         int                     rc;
2090         int                     ni_count = 0;
2091         __u32                   lnd_type;
2092         struct lnet_lnd *lnd;
2093         int                     peer_timeout =
2094                 net->net_tunables.lct_peer_timeout;
2095         int                     maxtxcredits =
2096                 net->net_tunables.lct_max_tx_credits;
2097         int                     peerrtrcredits =
2098                 net->net_tunables.lct_peer_rtr_credits;
2099
2100         INIT_LIST_HEAD(&local_ni_list);
2101
2102         /*
2103          * make sure that this net is unique. If it isn't then
2104          * we are adding interfaces to an already existing network, and
2105          * 'net' is just a convenient way to pass in the list.
2106          * if it is unique we need to find the LND and load it if
2107          * necessary.
2108          */
2109         if (lnet_net_unique(net->net_id, &the_lnet.ln_nets, &net_l)) {
2110                 lnd_type = LNET_NETTYP(net->net_id);
2111
2112                 mutex_lock(&the_lnet.ln_lnd_mutex);
2113                 lnd = lnet_find_lnd_by_type(lnd_type);
2114
2115                 if (lnd == NULL) {
2116                         mutex_unlock(&the_lnet.ln_lnd_mutex);
2117                         rc = request_module("%s", libcfs_lnd2modname(lnd_type));
2118                         mutex_lock(&the_lnet.ln_lnd_mutex);
2119
2120                         lnd = lnet_find_lnd_by_type(lnd_type);
2121                         if (lnd == NULL) {
2122                                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2123                                 CERROR("Can't load LND %s, module %s, rc=%d\n",
2124                                 libcfs_lnd2str(lnd_type),
2125                                 libcfs_lnd2modname(lnd_type), rc);
2126 #ifndef HAVE_MODULE_LOADING_SUPPORT
2127                                 LCONSOLE_ERROR_MSG(0x104, "Your kernel must be "
2128                                                 "compiled with kernel module "
2129                                                 "loading support.");
2130 #endif
2131                                 rc = -EINVAL;
2132                                 goto failed0;
2133                         }
2134                 }
2135
2136                 lnet_net_lock(LNET_LOCK_EX);
2137                 lnd->lnd_refcount++;
2138                 lnet_net_unlock(LNET_LOCK_EX);
2139
2140                 net->net_lnd = lnd;
2141
2142                 mutex_unlock(&the_lnet.ln_lnd_mutex);
2143
2144                 net_l = net;
2145         }
2146
2147         /*
2148          * net_l: if the network being added is unique then net_l
2149          *        will point to that network
2150          *        if the network being added is not unique then
2151          *        net_l points to the existing network.
2152          *
2153          * When we enter the loop below, we'll pick NIs off he
2154          * network beign added and start them up, then add them to
2155          * a local ni list. Once we've successfully started all
2156          * the NIs then we join the local NI list (of started up
2157          * networks) with the net_l->net_ni_list, which should
2158          * point to the correct network to add the new ni list to
2159          *
2160          * If any of the new NIs fail to start up, then we want to
2161          * iterate through the local ni list, which should include
2162          * any NIs which were successfully started up, and shut
2163          * them down.
2164          *
2165          * After than we want to delete the network being added,
2166          * to avoid a memory leak.
2167          */
2168
2169         /*
2170          * When a network uses TCP bonding then all its interfaces
2171          * must be specified when the network is first defined: the
2172          * TCP bonding code doesn't allow for interfaces to be added
2173          * or removed.
2174          */
2175         if (net_l != net && net_l != NULL && use_tcp_bonding &&
2176             LNET_NETTYP(net_l->net_id) == SOCKLND) {
2177                 rc = -EINVAL;
2178                 goto failed0;
2179         }
2180
2181         while (!list_empty(&net->net_ni_added)) {
2182                 ni = list_entry(net->net_ni_added.next, struct lnet_ni,
2183                                 ni_netlist);
2184                 list_del_init(&ni->ni_netlist);
2185
2186                 /* make sure that the the NI we're about to start
2187                  * up is actually unique. if it's not fail. */
2188                 if (!lnet_ni_unique_net(&net_l->net_ni_list,
2189                                         ni->ni_interfaces[0])) {
2190                         rc = -EINVAL;
2191                         goto failed1;
2192                 }
2193
2194                 /* adjust the pointer the parent network, just in case it
2195                  * the net is a duplicate */
2196                 ni->ni_net = net_l;
2197
2198                 rc = lnet_startup_lndni(ni, tun);
2199
2200                 LASSERT(ni->ni_net->net_tunables.lct_peer_timeout <= 0 ||
2201                         ni->ni_net->net_lnd->lnd_query != NULL);
2202
2203                 if (rc < 0)
2204                         goto failed1;
2205
2206                 lnet_ni_addref(ni);
2207                 list_add_tail(&ni->ni_netlist, &local_ni_list);
2208
2209                 ni_count++;
2210         }
2211
2212         lnet_net_lock(LNET_LOCK_EX);
2213         list_splice_tail(&local_ni_list, &net_l->net_ni_list);
2214         lnet_incr_dlc_seq();
2215         lnet_net_unlock(LNET_LOCK_EX);
2216
2217         /* if the network is not unique then we don't want to keep
2218          * it around after we're done. Free it. Otherwise add that
2219          * net to the global the_lnet.ln_nets */
2220         if (net_l != net && net_l != NULL) {
2221                 /*
2222                  * TODO - note. currently the tunables can not be updated
2223                  * once added
2224                  */
2225                 lnet_net_free(net);
2226         } else {
2227                 net->net_state = LNET_NET_STATE_ACTIVE;
2228                 /*
2229                  * restore tunables after it has been overwitten by the
2230                  * lnd
2231                  */
2232                 if (peer_timeout != -1)
2233                         net->net_tunables.lct_peer_timeout = peer_timeout;
2234                 if (maxtxcredits != -1)
2235                         net->net_tunables.lct_max_tx_credits = maxtxcredits;
2236                 if (peerrtrcredits != -1)
2237                         net->net_tunables.lct_peer_rtr_credits = peerrtrcredits;
2238
2239                 lnet_net_lock(LNET_LOCK_EX);
2240                 list_add_tail(&net->net_list, &the_lnet.ln_nets);
2241                 lnet_net_unlock(LNET_LOCK_EX);
2242         }
2243
2244         return ni_count;
2245
2246 failed1:
2247         /*
2248          * shutdown the new NIs that are being started up
2249          * free the NET being started
2250          */
2251         while (!list_empty(&local_ni_list)) {
2252                 ni = list_entry(local_ni_list.next, struct lnet_ni,
2253                                 ni_netlist);
2254
2255                 lnet_shutdown_lndni(ni);
2256         }
2257
2258 failed0:
2259         lnet_net_free(net);
2260
2261         return rc;
2262 }
2263
2264 static int
2265 lnet_startup_lndnets(struct list_head *netlist)
2266 {
2267         struct lnet_net         *net;
2268         int                     rc;
2269         int                     ni_count = 0;
2270
2271         /*
2272          * Change to running state before bringing up the LNDs. This
2273          * allows lnet_shutdown_lndnets() to assert that we've passed
2274          * through here.
2275          */
2276         lnet_net_lock(LNET_LOCK_EX);
2277         the_lnet.ln_state = LNET_STATE_RUNNING;
2278         lnet_net_unlock(LNET_LOCK_EX);
2279
2280         while (!list_empty(netlist)) {
2281                 net = list_entry(netlist->next, struct lnet_net, net_list);
2282                 list_del_init(&net->net_list);
2283
2284                 rc = lnet_startup_lndnet(net, NULL);
2285
2286                 if (rc < 0)
2287                         goto failed;
2288
2289                 ni_count += rc;
2290         }
2291
2292         return ni_count;
2293 failed:
2294         lnet_shutdown_lndnets();
2295
2296         return rc;
2297 }
2298
2299 /**
2300  * Initialize LNet library.
2301  *
2302  * Automatically called at module loading time. Caller has to call
2303  * lnet_lib_exit() after a call to lnet_lib_init(), if and only if the
2304  * latter returned 0. It must be called exactly once.
2305  *
2306  * \retval 0 on success
2307  * \retval -ve on failures.
2308  */
2309 int lnet_lib_init(void)
2310 {
2311         int rc;
2312
2313         lnet_assert_wire_constants();
2314
2315         /* refer to global cfs_cpt_table for now */
2316         the_lnet.ln_cpt_table   = cfs_cpt_table;
2317         the_lnet.ln_cpt_number  = cfs_cpt_number(cfs_cpt_table);
2318
2319         LASSERT(the_lnet.ln_cpt_number > 0);
2320         if (the_lnet.ln_cpt_number > LNET_CPT_MAX) {
2321                 /* we are under risk of consuming all lh_cookie */
2322                 CERROR("Can't have %d CPTs for LNet (max allowed is %d), "
2323                        "please change setting of CPT-table and retry\n",
2324                        the_lnet.ln_cpt_number, LNET_CPT_MAX);
2325                 return -E2BIG;
2326         }
2327
2328         while ((1 << the_lnet.ln_cpt_bits) < the_lnet.ln_cpt_number)
2329                 the_lnet.ln_cpt_bits++;
2330
2331         rc = lnet_create_locks();
2332         if (rc != 0) {
2333                 CERROR("Can't create LNet global locks: %d\n", rc);
2334                 return rc;
2335         }
2336
2337         the_lnet.ln_refcount = 0;
2338         LNetInvalidateEQHandle(&the_lnet.ln_rc_eqh);
2339         INIT_LIST_HEAD(&the_lnet.ln_lnds);
2340         INIT_LIST_HEAD(&the_lnet.ln_net_zombie);
2341         INIT_LIST_HEAD(&the_lnet.ln_rcd_zombie);
2342         INIT_LIST_HEAD(&the_lnet.ln_msg_resend);
2343         INIT_LIST_HEAD(&the_lnet.ln_rcd_deathrow);
2344
2345         /* The hash table size is the number of bits it takes to express the set
2346          * ln_num_routes, minus 1 (better to under estimate than over so we
2347          * don't waste memory). */
2348         if (rnet_htable_size <= 0)
2349                 rnet_htable_size = LNET_REMOTE_NETS_HASH_DEFAULT;
2350         else if (rnet_htable_size > LNET_REMOTE_NETS_HASH_MAX)
2351                 rnet_htable_size = LNET_REMOTE_NETS_HASH_MAX;
2352         the_lnet.ln_remote_nets_hbits = max_t(int, 1,
2353                                            order_base_2(rnet_htable_size) - 1);
2354
2355         /* All LNDs apart from the LOLND are in separate modules.  They
2356          * register themselves when their module loads, and unregister
2357          * themselves when their module is unloaded. */
2358         lnet_register_lnd(&the_lolnd);
2359         return 0;
2360 }
2361
2362 /**
2363  * Finalize LNet library.
2364  *
2365  * \pre lnet_lib_init() called with success.
2366  * \pre All LNet users called LNetNIFini() for matching LNetNIInit() calls.
2367  */
2368 void lnet_lib_exit(void)
2369 {
2370         LASSERT(the_lnet.ln_refcount == 0);
2371
2372         while (!list_empty(&the_lnet.ln_lnds))
2373                 lnet_unregister_lnd(list_entry(the_lnet.ln_lnds.next,
2374                                                struct lnet_lnd, lnd_list));
2375         lnet_destroy_locks();
2376 }
2377
2378 /**
2379  * Set LNet PID and start LNet interfaces, routing, and forwarding.
2380  *
2381  * Users must call this function at least once before any other functions.
2382  * For each successful call there must be a corresponding call to
2383  * LNetNIFini(). For subsequent calls to LNetNIInit(), \a requested_pid is
2384  * ignored.
2385  *
2386  * The PID used by LNet may be different from the one requested.
2387  * See LNetGetId().
2388  *
2389  * \param requested_pid PID requested by the caller.
2390  *
2391  * \return >= 0 on success, and < 0 error code on failures.
2392  */
2393 int
2394 LNetNIInit(lnet_pid_t requested_pid)
2395 {
2396         int                     im_a_router = 0;
2397         int                     rc;
2398         int                     ni_count;
2399         struct lnet_ping_buffer *pbuf;
2400         struct lnet_handle_md   ping_mdh;
2401         struct list_head        net_head;
2402         struct lnet_net         *net;
2403
2404         INIT_LIST_HEAD(&net_head);
2405
2406         mutex_lock(&the_lnet.ln_api_mutex);
2407
2408         CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
2409
2410         if (the_lnet.ln_refcount > 0) {
2411                 rc = the_lnet.ln_refcount++;
2412                 mutex_unlock(&the_lnet.ln_api_mutex);
2413                 return rc;
2414         }
2415
2416         rc = lnet_prepare(requested_pid);
2417         if (rc != 0) {
2418                 mutex_unlock(&the_lnet.ln_api_mutex);
2419                 return rc;
2420         }
2421
2422         /* create a network for Loopback network */
2423         net = lnet_net_alloc(LNET_MKNET(LOLND, 0), &net_head);
2424         if (net == NULL) {
2425                 rc = -ENOMEM;
2426                 goto err_empty_list;
2427         }
2428
2429         /* Add in the loopback NI */
2430         if (lnet_ni_alloc(net, NULL, NULL) == NULL) {
2431                 rc = -ENOMEM;
2432                 goto err_empty_list;
2433         }
2434
2435         /* If LNet is being initialized via DLC it is possible
2436          * that the user requests not to load module parameters (ones which
2437          * are supported by DLC) on initialization.  Therefore, make sure not
2438          * to load networks, routes and forwarding from module parameters
2439          * in this case.  On cleanup in case of failure only clean up
2440          * routes if it has been loaded */
2441         if (!the_lnet.ln_nis_from_mod_params) {
2442                 rc = lnet_parse_networks(&net_head, lnet_get_networks(),
2443                                          use_tcp_bonding);
2444                 if (rc < 0)
2445                         goto err_empty_list;
2446         }
2447
2448         ni_count = lnet_startup_lndnets(&net_head);
2449         if (ni_count < 0) {
2450                 rc = ni_count;
2451                 goto err_empty_list;
2452         }
2453
2454         if (!the_lnet.ln_nis_from_mod_params) {
2455                 rc = lnet_parse_routes(lnet_get_routes(), &im_a_router);
2456                 if (rc != 0)
2457                         goto err_shutdown_lndnis;
2458
2459                 rc = lnet_check_routes();
2460                 if (rc != 0)
2461                         goto err_destroy_routes;
2462
2463                 rc = lnet_rtrpools_alloc(im_a_router);
2464                 if (rc != 0)
2465                         goto err_destroy_routes;
2466         }
2467
2468         rc = lnet_acceptor_start();
2469         if (rc != 0)
2470                 goto err_destroy_routes;
2471
2472         the_lnet.ln_refcount = 1;
2473         /* Now I may use my own API functions... */
2474
2475         rc = lnet_ping_target_setup(&pbuf, &ping_mdh, ni_count, true);
2476         if (rc != 0)
2477                 goto err_acceptor_stop;
2478
2479         lnet_ping_target_update(pbuf, ping_mdh);
2480
2481         rc = lnet_monitor_thr_start();
2482         if (rc != 0)
2483                 goto err_stop_ping;
2484
2485         rc = lnet_push_target_init();
2486         if (rc != 0)
2487                 goto err_stop_monitor_thr;
2488
2489         rc = lnet_peer_discovery_start();
2490         if (rc != 0)
2491                 goto err_destroy_push_target;
2492
2493         lnet_fault_init();
2494         lnet_router_debugfs_init();
2495
2496         mutex_unlock(&the_lnet.ln_api_mutex);
2497
2498         return 0;
2499
2500 err_destroy_push_target:
2501         lnet_push_target_fini();
2502 err_stop_monitor_thr:
2503         lnet_monitor_thr_stop();
2504 err_stop_ping:
2505         lnet_ping_target_fini();
2506 err_acceptor_stop:
2507         the_lnet.ln_refcount = 0;
2508         lnet_acceptor_stop();
2509 err_destroy_routes:
2510         if (!the_lnet.ln_nis_from_mod_params)
2511                 lnet_destroy_routes();
2512 err_shutdown_lndnis:
2513         lnet_shutdown_lndnets();
2514 err_empty_list:
2515         lnet_unprepare();
2516         LASSERT(rc < 0);
2517         mutex_unlock(&the_lnet.ln_api_mutex);
2518         while (!list_empty(&net_head)) {
2519                 struct lnet_net *net;
2520
2521                 net = list_entry(net_head.next, struct lnet_net, net_list);
2522                 list_del_init(&net->net_list);
2523                 lnet_net_free(net);
2524         }
2525         return rc;
2526 }
2527 EXPORT_SYMBOL(LNetNIInit);
2528
2529 /**
2530  * Stop LNet interfaces, routing, and forwarding.
2531  *
2532  * Users must call this function once for each successful call to LNetNIInit().
2533  * Once the LNetNIFini() operation has been started, the results of pending
2534  * API operations are undefined.
2535  *
2536  * \return always 0 for current implementation.
2537  */
2538 int
2539 LNetNIFini()
2540 {
2541         mutex_lock(&the_lnet.ln_api_mutex);
2542
2543         LASSERT(the_lnet.ln_refcount > 0);
2544
2545         if (the_lnet.ln_refcount != 1) {
2546                 the_lnet.ln_refcount--;
2547         } else {
2548                 LASSERT(!the_lnet.ln_niinit_self);
2549
2550                 lnet_fault_fini();
2551
2552                 lnet_router_debugfs_init();
2553                 lnet_peer_discovery_stop();
2554                 lnet_push_target_fini();
2555                 lnet_monitor_thr_stop();
2556                 lnet_ping_target_fini();
2557
2558                 /* Teardown fns that use my own API functions BEFORE here */
2559                 the_lnet.ln_refcount = 0;
2560
2561                 lnet_acceptor_stop();
2562                 lnet_destroy_routes();
2563                 lnet_shutdown_lndnets();
2564                 lnet_unprepare();
2565         }
2566
2567         mutex_unlock(&the_lnet.ln_api_mutex);
2568         return 0;
2569 }
2570 EXPORT_SYMBOL(LNetNIFini);
2571
2572 /**
2573  * Grabs the ni data from the ni structure and fills the out
2574  * parameters
2575  *
2576  * \param[in] ni network        interface structure
2577  * \param[out] cfg_ni           NI config information
2578  * \param[out] tun              network and LND tunables
2579  */
2580 static void
2581 lnet_fill_ni_info(struct lnet_ni *ni, struct lnet_ioctl_config_ni *cfg_ni,
2582                    struct lnet_ioctl_config_lnd_tunables *tun,
2583                    struct lnet_ioctl_element_stats *stats,
2584                    __u32 tun_size)
2585 {
2586         size_t min_size = 0;
2587         int i;
2588
2589         if (!ni || !cfg_ni || !tun)
2590                 return;
2591
2592         if (ni->ni_interfaces[0] != NULL) {
2593                 for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2594                         if (ni->ni_interfaces[i] != NULL) {
2595                                 strncpy(cfg_ni->lic_ni_intf[i],
2596                                         ni->ni_interfaces[i],
2597                                         sizeof(cfg_ni->lic_ni_intf[i]));
2598                         }
2599                 }
2600         }
2601
2602         cfg_ni->lic_nid = ni->ni_nid;
2603         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2604                 cfg_ni->lic_status = LNET_NI_STATUS_UP;
2605         else
2606                 cfg_ni->lic_status = ni->ni_status->ns_status;
2607         cfg_ni->lic_tcp_bonding = use_tcp_bonding;
2608         cfg_ni->lic_dev_cpt = ni->ni_dev_cpt;
2609
2610         memcpy(&tun->lt_cmn, &ni->ni_net->net_tunables, sizeof(tun->lt_cmn));
2611
2612         if (stats) {
2613                 stats->iel_send_count = lnet_sum_stats(&ni->ni_stats,
2614                                                        LNET_STATS_TYPE_SEND);
2615                 stats->iel_recv_count = lnet_sum_stats(&ni->ni_stats,
2616                                                        LNET_STATS_TYPE_RECV);
2617                 stats->iel_drop_count = lnet_sum_stats(&ni->ni_stats,
2618                                                        LNET_STATS_TYPE_DROP);
2619         }
2620
2621         /*
2622          * tun->lt_tun will always be present, but in order to be
2623          * backwards compatible, we need to deal with the cases when
2624          * tun->lt_tun is smaller than what the kernel has, because it
2625          * comes from an older version of a userspace program, then we'll
2626          * need to copy as much information as we have available space.
2627          */
2628         min_size = tun_size - sizeof(tun->lt_cmn);
2629         memcpy(&tun->lt_tun, &ni->ni_lnd_tunables, min_size);
2630
2631         /* copy over the cpts */
2632         if (ni->ni_ncpts == LNET_CPT_NUMBER &&
2633             ni->ni_cpts == NULL)  {
2634                 for (i = 0; i < ni->ni_ncpts; i++)
2635                         cfg_ni->lic_cpts[i] = i;
2636         } else {
2637                 for (i = 0;
2638                      ni->ni_cpts != NULL && i < ni->ni_ncpts &&
2639                      i < LNET_MAX_SHOW_NUM_CPT;
2640                      i++)
2641                         cfg_ni->lic_cpts[i] = ni->ni_cpts[i];
2642         }
2643         cfg_ni->lic_ncpts = ni->ni_ncpts;
2644 }
2645
2646 /**
2647  * NOTE: This is a legacy function left in the code to be backwards
2648  * compatible with older userspace programs. It should eventually be
2649  * removed.
2650  *
2651  * Grabs the ni data from the ni structure and fills the out
2652  * parameters
2653  *
2654  * \param[in] ni network        interface structure
2655  * \param[out] config           config information
2656  */
2657 static void
2658 lnet_fill_ni_info_legacy(struct lnet_ni *ni,
2659                          struct lnet_ioctl_config_data *config)
2660 {
2661         struct lnet_ioctl_net_config *net_config;
2662         struct lnet_ioctl_config_lnd_tunables *lnd_cfg = NULL;
2663         size_t min_size, tunable_size = 0;
2664         int i;
2665
2666         if (!ni || !config)
2667                 return;
2668
2669         net_config = (struct lnet_ioctl_net_config *) config->cfg_bulk;
2670         if (!net_config)
2671                 return;
2672
2673         BUILD_BUG_ON(ARRAY_SIZE(ni->ni_interfaces) !=
2674                      ARRAY_SIZE(net_config->ni_interfaces));
2675
2676         for (i = 0; i < ARRAY_SIZE(ni->ni_interfaces); i++) {
2677                 if (!ni->ni_interfaces[i])
2678                         break;
2679
2680                 strncpy(net_config->ni_interfaces[i],
2681                         ni->ni_interfaces[i],
2682                         sizeof(net_config->ni_interfaces[i]));
2683         }
2684
2685         config->cfg_nid = ni->ni_nid;
2686         config->cfg_config_u.cfg_net.net_peer_timeout =
2687                 ni->ni_net->net_tunables.lct_peer_timeout;
2688         config->cfg_config_u.cfg_net.net_max_tx_credits =
2689                 ni->ni_net->net_tunables.lct_max_tx_credits;
2690         config->cfg_config_u.cfg_net.net_peer_tx_credits =
2691                 ni->ni_net->net_tunables.lct_peer_tx_credits;
2692         config->cfg_config_u.cfg_net.net_peer_rtr_credits =
2693                 ni->ni_net->net_tunables.lct_peer_rtr_credits;
2694
2695         if (LNET_NETTYP(LNET_NIDNET(ni->ni_nid)) == LOLND)
2696                 net_config->ni_status = LNET_NI_STATUS_UP;
2697         else
2698                 net_config->ni_status = ni->ni_status->ns_status;
2699
2700         if (ni->ni_cpts) {
2701                 int num_cpts = min(ni->ni_ncpts, LNET_MAX_SHOW_NUM_CPT);
2702
2703                 for (i = 0; i < num_cpts; i++)
2704                         net_config->ni_cpts[i] = ni->ni_cpts[i];
2705
2706                 config->cfg_ncpts = num_cpts;
2707         }
2708
2709         /*
2710          * See if user land tools sent in a newer and larger version
2711          * of struct lnet_tunables than what the kernel uses.
2712          */
2713         min_size = sizeof(*config) + sizeof(*net_config);
2714
2715         if (config->cfg_hdr.ioc_len > min_size)
2716                 tunable_size = config->cfg_hdr.ioc_len - min_size;
2717
2718         /* Don't copy too much data to user space */
2719         min_size = min(tunable_size, sizeof(ni->ni_lnd_tunables));
2720         lnd_cfg = (struct lnet_ioctl_config_lnd_tunables *)net_config->cfg_bulk;
2721
2722         if (lnd_cfg && min_size) {
2723                 memcpy(&lnd_cfg->lt_tun, &ni->ni_lnd_tunables, min_size);
2724                 config->cfg_config_u.cfg_net.net_interface_count = 1;
2725
2726                 /* Tell user land that kernel side has less data */
2727                 if (tunable_size > sizeof(ni->ni_lnd_tunables)) {
2728                         min_size = tunable_size - sizeof(ni->ni_lnd_tunables);
2729                         config->cfg_hdr.ioc_len -= min_size;
2730                 }
2731         }
2732 }
2733
2734 struct lnet_ni *
2735 lnet_get_ni_idx_locked(int idx)
2736 {
2737         struct lnet_ni          *ni;
2738         struct lnet_net         *net;
2739
2740         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
2741                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
2742                         if (idx-- == 0)
2743                                 return ni;
2744                 }
2745         }
2746
2747         return NULL;
2748 }
2749
2750 struct lnet_ni *
2751 lnet_get_next_ni_locked(struct lnet_net *mynet, struct lnet_ni *prev)
2752 {
2753         struct lnet_ni          *ni;
2754         struct lnet_net         *net = mynet;
2755
2756         /*
2757          * It is possible that the net has been cleaned out while there is
2758          * a message being sent. This function accessed the net without
2759          * checking if the list is empty
2760          */
2761         if (prev == NULL) {
2762                 if (net == NULL)
2763                         net = list_entry(the_lnet.ln_nets.next, struct lnet_net,
2764                                         net_list);
2765                 if (list_empty(&net->net_ni_list))
2766                         return NULL;
2767                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2768                                 ni_netlist);
2769
2770                 return ni;
2771         }
2772
2773         if (prev->ni_netlist.next == &prev->ni_net->net_ni_list) {
2774                 /* if you reached the end of the ni list and the net is
2775                  * specified, then there are no more nis in that net */
2776                 if (net != NULL)
2777                         return NULL;
2778
2779                 /* we reached the end of this net ni list. move to the
2780                  * next net */
2781                 if (prev->ni_net->net_list.next == &the_lnet.ln_nets)
2782                         /* no more nets and no more NIs. */
2783                         return NULL;
2784
2785                 /* get the next net */
2786                 net = list_entry(prev->ni_net->net_list.next, struct lnet_net,
2787                                  net_list);
2788                 if (list_empty(&net->net_ni_list))
2789                         return NULL;
2790                 /* get the ni on it */
2791                 ni = list_entry(net->net_ni_list.next, struct lnet_ni,
2792                                 ni_netlist);
2793
2794                 return ni;
2795         }
2796
2797         if (list_empty(&prev->ni_netlist))
2798                 return NULL;
2799
2800         /* there are more nis left */
2801         ni = list_entry(prev->ni_netlist.next, struct lnet_ni, ni_netlist);
2802
2803         return ni;
2804 }
2805
2806 int
2807 lnet_get_net_config(struct lnet_ioctl_config_data *config)
2808 {
2809         struct lnet_ni *ni;
2810         int cpt;
2811         int rc = -ENOENT;
2812         int idx = config->cfg_count;
2813
2814         cpt = lnet_net_lock_current();
2815
2816         ni = lnet_get_ni_idx_locked(idx);
2817
2818         if (ni != NULL) {
2819                 rc = 0;
2820                 lnet_ni_lock(ni);
2821                 lnet_fill_ni_info_legacy(ni, config);
2822                 lnet_ni_unlock(ni);
2823         }
2824
2825         lnet_net_unlock(cpt);
2826         return rc;
2827 }
2828
2829 int
2830 lnet_get_ni_config(struct lnet_ioctl_config_ni *cfg_ni,
2831                    struct lnet_ioctl_config_lnd_tunables *tun,
2832                    struct lnet_ioctl_element_stats *stats,
2833                    __u32 tun_size)
2834 {
2835         struct lnet_ni          *ni;
2836         int                     cpt;
2837         int                     rc = -ENOENT;
2838
2839         if (!cfg_ni || !tun || !stats)
2840                 return -EINVAL;
2841
2842         cpt = lnet_net_lock_current();
2843
2844         ni = lnet_get_ni_idx_locked(cfg_ni->lic_idx);
2845
2846         if (ni) {
2847                 rc = 0;
2848                 lnet_ni_lock(ni);
2849                 lnet_fill_ni_info(ni, cfg_ni, tun, stats, tun_size);
2850                 lnet_ni_unlock(ni);
2851         }
2852
2853         lnet_net_unlock(cpt);
2854         return rc;
2855 }
2856
2857 int lnet_get_ni_stats(struct lnet_ioctl_element_msg_stats *msg_stats)
2858 {
2859         struct lnet_ni *ni;
2860         int cpt;
2861         int rc = -ENOENT;
2862
2863         if (!msg_stats)
2864                 return -EINVAL;
2865
2866         cpt = lnet_net_lock_current();
2867
2868         ni = lnet_get_ni_idx_locked(msg_stats->im_idx);
2869
2870         if (ni) {
2871                 lnet_usr_translate_stats(msg_stats, &ni->ni_stats);
2872                 rc = 0;
2873         }
2874
2875         lnet_net_unlock(cpt);
2876
2877         return rc;
2878 }
2879
2880 static int lnet_add_net_common(struct lnet_net *net,
2881                                struct lnet_ioctl_config_lnd_tunables *tun)
2882 {
2883         __u32                   net_id;
2884         struct lnet_ping_buffer *pbuf;
2885         struct lnet_handle_md   ping_mdh;
2886         int                     rc;
2887         struct lnet_remotenet *rnet;
2888         int                     net_ni_count;
2889         int                     num_acceptor_nets;
2890
2891         lnet_net_lock(LNET_LOCK_EX);
2892         rnet = lnet_find_rnet_locked(net->net_id);
2893         lnet_net_unlock(LNET_LOCK_EX);
2894         /*
2895          * make sure that the net added doesn't invalidate the current
2896          * configuration LNet is keeping
2897          */
2898         if (rnet) {
2899                 CERROR("Adding net %s will invalidate routing configuration\n",
2900                        libcfs_net2str(net->net_id));
2901                 lnet_net_free(net);
2902                 return -EUSERS;
2903         }
2904
2905         /*
2906          * make sure you calculate the correct number of slots in the ping
2907          * buffer. Since the ping info is a flattened list of all the NIs,
2908          * we should allocate enough slots to accomodate the number of NIs
2909          * which will be added.
2910          *
2911          * since ni hasn't been configured yet, use
2912          * lnet_get_net_ni_count_pre() which checks the net_ni_added list
2913          */
2914         net_ni_count = lnet_get_net_ni_count_pre(net);
2915
2916         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
2917                                     net_ni_count + lnet_get_ni_count(),
2918                                     false);
2919         if (rc < 0) {
2920                 lnet_net_free(net);
2921                 return rc;
2922         }
2923
2924         if (tun)
2925                 memcpy(&net->net_tunables,
2926                        &tun->lt_cmn, sizeof(net->net_tunables));
2927         else
2928                 memset(&net->net_tunables, -1, sizeof(net->net_tunables));
2929
2930         /*
2931          * before starting this network get a count of the current TCP
2932          * networks which require the acceptor thread running. If that
2933          * count is == 0 before we start up this network, then we'd want to
2934          * start up the acceptor thread after starting up this network
2935          */
2936         num_acceptor_nets = lnet_count_acceptor_nets();
2937
2938         net_id = net->net_id;
2939
2940         rc = lnet_startup_lndnet(net,
2941                                  (tun) ? &tun->lt_tun : NULL);
2942         if (rc < 0)
2943                 goto failed;
2944
2945         lnet_net_lock(LNET_LOCK_EX);
2946         net = lnet_get_net_locked(net_id);
2947         lnet_net_unlock(LNET_LOCK_EX);
2948
2949         LASSERT(net);
2950
2951         /*
2952          * Start the acceptor thread if this is the first network
2953          * being added that requires the thread.
2954          */
2955         if (net->net_lnd->lnd_accept && num_acceptor_nets == 0) {
2956                 rc = lnet_acceptor_start();
2957                 if (rc < 0) {
2958                         /* shutdown the net that we just started */
2959                         CERROR("Failed to start up acceptor thread\n");
2960                         lnet_shutdown_lndnet(net);
2961                         goto failed;
2962                 }
2963         }
2964
2965         lnet_net_lock(LNET_LOCK_EX);
2966         lnet_peer_net_added(net);
2967         lnet_net_unlock(LNET_LOCK_EX);
2968
2969         lnet_ping_target_update(pbuf, ping_mdh);
2970
2971         return 0;
2972
2973 failed:
2974         lnet_ping_md_unlink(pbuf, &ping_mdh);
2975         lnet_ping_buffer_decref(pbuf);
2976         return rc;
2977 }
2978
2979 static int lnet_handle_legacy_ip2nets(char *ip2nets,
2980                                       struct lnet_ioctl_config_lnd_tunables *tun)
2981 {
2982         struct lnet_net *net;
2983         char *nets;
2984         int rc;
2985         struct list_head net_head;
2986
2987         INIT_LIST_HEAD(&net_head);
2988
2989         rc = lnet_parse_ip2nets(&nets, ip2nets);
2990         if (rc < 0)
2991                 return rc;
2992
2993         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
2994         if (rc < 0)
2995                 return rc;
2996
2997         mutex_lock(&the_lnet.ln_api_mutex);
2998         while (!list_empty(&net_head)) {
2999                 net = list_entry(net_head.next, struct lnet_net, net_list);
3000                 list_del_init(&net->net_list);
3001                 rc = lnet_add_net_common(net, tun);
3002                 if (rc < 0)
3003                         goto out;
3004         }
3005
3006 out:
3007         mutex_unlock(&the_lnet.ln_api_mutex);
3008
3009         while (!list_empty(&net_head)) {
3010                 net = list_entry(net_head.next, struct lnet_net, net_list);
3011                 list_del_init(&net->net_list);
3012                 lnet_net_free(net);
3013         }
3014         return rc;
3015 }
3016
3017 int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
3018 {
3019         struct lnet_net *net;
3020         struct lnet_ni *ni;
3021         struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3022         int rc, i;
3023         __u32 net_id, lnd_type;
3024
3025         /* get the tunables if they are available */
3026         if (conf->lic_cfg_hdr.ioc_len >=
3027             sizeof(*conf) + sizeof(*tun))
3028                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3029                         conf->lic_bulk;
3030
3031         /* handle legacy ip2nets from DLC */
3032         if (conf->lic_legacy_ip2nets[0] != '\0')
3033                 return lnet_handle_legacy_ip2nets(conf->lic_legacy_ip2nets,
3034                                                   tun);
3035
3036         net_id = LNET_NIDNET(conf->lic_nid);
3037         lnd_type = LNET_NETTYP(net_id);
3038
3039         if (!libcfs_isknown_lnd(lnd_type)) {
3040                 CERROR("No valid net and lnd information provided\n");
3041                 return -EINVAL;
3042         }
3043
3044         net = lnet_net_alloc(net_id, NULL);
3045         if (!net)
3046                 return -ENOMEM;
3047
3048         for (i = 0; i < conf->lic_ncpts; i++) {
3049                 if (conf->lic_cpts[i] >= LNET_CPT_NUMBER)
3050                         return -EINVAL;
3051         }
3052
3053         ni = lnet_ni_alloc_w_cpt_array(net, conf->lic_cpts, conf->lic_ncpts,
3054                                        conf->lic_ni_intf[0]);
3055         if (!ni)
3056                 return -ENOMEM;
3057
3058         mutex_lock(&the_lnet.ln_api_mutex);
3059
3060         rc = lnet_add_net_common(net, tun);
3061
3062         mutex_unlock(&the_lnet.ln_api_mutex);
3063
3064         return rc;
3065 }
3066
3067 int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
3068 {
3069         struct lnet_net  *net;
3070         struct lnet_ni *ni;
3071         __u32 net_id = LNET_NIDNET(conf->lic_nid);
3072         struct lnet_ping_buffer *pbuf;
3073         struct lnet_handle_md  ping_mdh;
3074         int               rc;
3075         int               net_count;
3076         __u32             addr;
3077
3078         /* don't allow userspace to shutdown the LOLND */
3079         if (LNET_NETTYP(net_id) == LOLND)
3080                 return -EINVAL;
3081
3082         mutex_lock(&the_lnet.ln_api_mutex);
3083
3084         lnet_net_lock(0);
3085
3086         net = lnet_get_net_locked(net_id);
3087         if (!net) {
3088                 CERROR("net %s not found\n",
3089                        libcfs_net2str(net_id));
3090                 rc = -ENOENT;
3091                 goto unlock_net;
3092         }
3093
3094         addr = LNET_NIDADDR(conf->lic_nid);
3095         if (addr == 0) {
3096                 /* remove the entire net */
3097                 net_count = lnet_get_net_ni_count_locked(net);
3098
3099                 lnet_net_unlock(0);
3100
3101                 /* create and link a new ping info, before removing the old one */
3102                 rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3103                                         lnet_get_ni_count() - net_count,
3104                                         false);
3105                 if (rc != 0)
3106                         goto unlock_api_mutex;
3107
3108                 lnet_shutdown_lndnet(net);
3109
3110                 if (lnet_count_acceptor_nets() == 0)
3111                         lnet_acceptor_stop();
3112
3113                 lnet_ping_target_update(pbuf, ping_mdh);
3114
3115                 goto unlock_api_mutex;
3116         }
3117
3118         ni = lnet_nid2ni_locked(conf->lic_nid, 0);
3119         if (!ni) {
3120                 CERROR("nid %s not found\n",
3121                        libcfs_nid2str(conf->lic_nid));
3122                 rc = -ENOENT;
3123                 goto unlock_net;
3124         }
3125
3126         net_count = lnet_get_net_ni_count_locked(net);
3127
3128         lnet_net_unlock(0);
3129
3130         /* create and link a new ping info, before removing the old one */
3131         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3132                                   lnet_get_ni_count() - 1, false);
3133         if (rc != 0)
3134                 goto unlock_api_mutex;
3135
3136         lnet_shutdown_lndni(ni);
3137
3138         if (lnet_count_acceptor_nets() == 0)
3139                 lnet_acceptor_stop();
3140
3141         lnet_ping_target_update(pbuf, ping_mdh);
3142
3143         /* check if the net is empty and remove it if it is */
3144         if (net_count == 1)
3145                 lnet_shutdown_lndnet(net);
3146
3147         goto unlock_api_mutex;
3148
3149 unlock_net:
3150         lnet_net_unlock(0);
3151 unlock_api_mutex:
3152         mutex_unlock(&the_lnet.ln_api_mutex);
3153
3154         return rc;
3155 }
3156
3157 /*
3158  * lnet_dyn_add_net and lnet_dyn_del_net are now deprecated.
3159  * They are only expected to be called for unique networks.
3160  * That can be as a result of older DLC library
3161  * calls. Multi-Rail DLC and beyond no longer uses these APIs.
3162  */
3163 int
3164 lnet_dyn_add_net(struct lnet_ioctl_config_data *conf)
3165 {
3166         struct lnet_net         *net;
3167         struct list_head        net_head;
3168         int                     rc;
3169         struct lnet_ioctl_config_lnd_tunables tun;
3170         char *nets = conf->cfg_config_u.cfg_net.net_intf;
3171
3172         INIT_LIST_HEAD(&net_head);
3173
3174         /* Create a net/ni structures for the network string */
3175         rc = lnet_parse_networks(&net_head, nets, use_tcp_bonding);
3176         if (rc <= 0)
3177                 return rc == 0 ? -EINVAL : rc;
3178
3179         mutex_lock(&the_lnet.ln_api_mutex);
3180
3181         if (rc > 1) {
3182                 rc = -EINVAL; /* only add one network per call */
3183                 goto out_unlock_clean;
3184         }
3185
3186         net = list_entry(net_head.next, struct lnet_net, net_list);
3187         list_del_init(&net->net_list);
3188
3189         LASSERT(lnet_net_unique(net->net_id, &the_lnet.ln_nets, NULL));
3190
3191         memset(&tun, 0, sizeof(tun));
3192
3193         tun.lt_cmn.lct_peer_timeout =
3194           conf->cfg_config_u.cfg_net.net_peer_timeout;
3195         tun.lt_cmn.lct_peer_tx_credits =
3196           conf->cfg_config_u.cfg_net.net_peer_tx_credits;
3197         tun.lt_cmn.lct_peer_rtr_credits =
3198           conf->cfg_config_u.cfg_net.net_peer_rtr_credits;
3199         tun.lt_cmn.lct_max_tx_credits =
3200           conf->cfg_config_u.cfg_net.net_max_tx_credits;
3201
3202         rc = lnet_add_net_common(net, &tun);
3203
3204 out_unlock_clean:
3205         mutex_unlock(&the_lnet.ln_api_mutex);
3206         while (!list_empty(&net_head)) {
3207                 /* net_head list is empty in success case */
3208                 net = list_entry(net_head.next, struct lnet_net, net_list);
3209                 list_del_init(&net->net_list);
3210                 lnet_net_free(net);
3211         }
3212         return rc;
3213 }
3214
3215 int
3216 lnet_dyn_del_net(__u32 net_id)
3217 {
3218         struct lnet_net  *net;
3219         struct lnet_ping_buffer *pbuf;
3220         struct lnet_handle_md ping_mdh;
3221         int               rc;
3222         int               net_ni_count;
3223
3224         /* don't allow userspace to shutdown the LOLND */
3225         if (LNET_NETTYP(net_id) == LOLND)
3226                 return -EINVAL;
3227
3228         mutex_lock(&the_lnet.ln_api_mutex);
3229
3230         lnet_net_lock(0);
3231
3232         net = lnet_get_net_locked(net_id);
3233         if (net == NULL) {
3234                 lnet_net_unlock(0);
3235                 rc = -EINVAL;
3236                 goto out;
3237         }
3238
3239         net_ni_count = lnet_get_net_ni_count_locked(net);
3240
3241         lnet_net_unlock(0);
3242
3243         /* create and link a new ping info, before removing the old one */
3244         rc = lnet_ping_target_setup(&pbuf, &ping_mdh,
3245                                     lnet_get_ni_count() - net_ni_count, false);
3246         if (rc != 0)
3247                 goto out;
3248
3249         lnet_shutdown_lndnet(net);
3250
3251         if (lnet_count_acceptor_nets() == 0)
3252                 lnet_acceptor_stop();
3253
3254         lnet_ping_target_update(pbuf, ping_mdh);
3255
3256 out:
3257         mutex_unlock(&the_lnet.ln_api_mutex);
3258
3259         return rc;
3260 }
3261
3262 void lnet_incr_dlc_seq(void)
3263 {
3264         atomic_inc(&lnet_dlc_seq_no);
3265 }
3266
3267 __u32 lnet_get_dlc_seq_locked(void)
3268 {
3269         return atomic_read(&lnet_dlc_seq_no);
3270 }
3271
3272 static void
3273 lnet_ni_set_healthv(lnet_nid_t nid, int value, bool all)
3274 {
3275         struct lnet_net *net;
3276         struct lnet_ni *ni;
3277
3278         lnet_net_lock(LNET_LOCK_EX);
3279         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3280                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3281                         if (ni->ni_nid == nid || all) {
3282                                 atomic_set(&ni->ni_healthv, value);
3283                                 if (list_empty(&ni->ni_recovery) &&
3284                                     value < LNET_MAX_HEALTH_VALUE) {
3285                                         CERROR("manually adding local NI %s to recovery\n",
3286                                                libcfs_nid2str(ni->ni_nid));
3287                                         list_add_tail(&ni->ni_recovery,
3288                                                       &the_lnet.ln_mt_localNIRecovq);
3289                                         lnet_ni_addref_locked(ni, 0);
3290                                 }
3291                                 if (!all) {
3292                                         lnet_net_unlock(LNET_LOCK_EX);
3293                                         return;
3294                                 }
3295                         }
3296                 }
3297         }
3298         lnet_net_unlock(LNET_LOCK_EX);
3299 }
3300
3301 static int
3302 lnet_get_local_ni_hstats(struct lnet_ioctl_local_ni_hstats *stats)
3303 {
3304         int cpt, rc = 0;
3305         struct lnet_ni *ni;
3306         lnet_nid_t nid = stats->hlni_nid;
3307
3308         cpt = lnet_net_lock_current();
3309         ni = lnet_nid2ni_locked(nid, cpt);
3310
3311         if (!ni) {
3312                 rc = -ENOENT;
3313                 goto unlock;
3314         }
3315
3316         stats->hlni_local_interrupt = atomic_read(&ni->ni_hstats.hlt_local_interrupt);
3317         stats->hlni_local_dropped = atomic_read(&ni->ni_hstats.hlt_local_dropped);
3318         stats->hlni_local_aborted = atomic_read(&ni->ni_hstats.hlt_local_aborted);
3319         stats->hlni_local_no_route = atomic_read(&ni->ni_hstats.hlt_local_no_route);
3320         stats->hlni_local_timeout = atomic_read(&ni->ni_hstats.hlt_local_timeout);
3321         stats->hlni_local_error = atomic_read(&ni->ni_hstats.hlt_local_error);
3322         stats->hlni_health_value = atomic_read(&ni->ni_healthv);
3323
3324 unlock:
3325         lnet_net_unlock(cpt);
3326
3327         return rc;
3328 }
3329
3330 static int
3331 lnet_get_local_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3332 {
3333         struct lnet_ni *ni;
3334         int i = 0;
3335
3336         lnet_net_lock(LNET_LOCK_EX);
3337         list_for_each_entry(ni, &the_lnet.ln_mt_localNIRecovq, ni_recovery) {
3338                 list->rlst_nid_array[i] = ni->ni_nid;
3339                 i++;
3340                 if (i >= LNET_MAX_SHOW_NUM_NID)
3341                         break;
3342         }
3343         lnet_net_unlock(LNET_LOCK_EX);
3344         list->rlst_num_nids = i;
3345
3346         return 0;
3347 }
3348
3349 static int
3350 lnet_get_peer_ni_recovery_list(struct lnet_ioctl_recovery_list *list)
3351 {
3352         struct lnet_peer_ni *lpni;
3353         int i = 0;
3354
3355         lnet_net_lock(LNET_LOCK_EX);
3356         list_for_each_entry(lpni, &the_lnet.ln_mt_peerNIRecovq, lpni_recovery) {
3357                 list->rlst_nid_array[i] = lpni->lpni_nid;
3358                 i++;
3359                 if (i >= LNET_MAX_SHOW_NUM_NID)
3360                         break;
3361         }
3362         lnet_net_unlock(LNET_LOCK_EX);
3363         list->rlst_num_nids = i;
3364
3365         return 0;
3366 }
3367
3368 /**
3369  * LNet ioctl handler.
3370  *
3371  */
3372 int
3373 LNetCtl(unsigned int cmd, void *arg)
3374 {
3375         struct libcfs_ioctl_data *data = arg;
3376         struct lnet_ioctl_config_data *config;
3377         struct lnet_process_id    id = {0};
3378         struct lnet_ni           *ni;
3379         int                       rc;
3380
3381         BUILD_BUG_ON(sizeof(struct lnet_ioctl_net_config) +
3382                      sizeof(struct lnet_ioctl_config_data) > LIBCFS_IOC_DATA_MAX);
3383
3384         switch (cmd) {
3385         case IOC_LIBCFS_GET_NI:
3386                 rc = LNetGetId(data->ioc_count, &id);
3387                 data->ioc_nid = id.nid;
3388                 return rc;
3389
3390         case IOC_LIBCFS_FAIL_NID:
3391                 return lnet_fail_nid(data->ioc_nid, data->ioc_count);
3392
3393         case IOC_LIBCFS_ADD_ROUTE:
3394                 config = arg;
3395
3396                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3397                         return -EINVAL;
3398
3399                 mutex_lock(&the_lnet.ln_api_mutex);
3400                 rc = lnet_add_route(config->cfg_net,
3401                                     config->cfg_config_u.cfg_route.rtr_hop,
3402                                     config->cfg_nid,
3403                                     config->cfg_config_u.cfg_route.
3404                                         rtr_priority);
3405                 if (rc == 0) {
3406                         rc = lnet_check_routes();
3407                         if (rc != 0)
3408                                 lnet_del_route(config->cfg_net,
3409                                                config->cfg_nid);
3410                 }
3411                 mutex_unlock(&the_lnet.ln_api_mutex);
3412                 return rc;
3413
3414         case IOC_LIBCFS_DEL_ROUTE:
3415                 config = arg;
3416
3417                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3418                         return -EINVAL;
3419
3420                 mutex_lock(&the_lnet.ln_api_mutex);
3421                 rc = lnet_del_route(config->cfg_net, config->cfg_nid);
3422                 mutex_unlock(&the_lnet.ln_api_mutex);
3423                 return rc;
3424
3425         case IOC_LIBCFS_GET_ROUTE:
3426                 config = arg;
3427
3428                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3429                         return -EINVAL;
3430
3431                 mutex_lock(&the_lnet.ln_api_mutex);
3432                 rc = lnet_get_route(config->cfg_count,
3433                                     &config->cfg_net,
3434                                     &config->cfg_config_u.cfg_route.rtr_hop,
3435                                     &config->cfg_nid,
3436                                     &config->cfg_config_u.cfg_route.rtr_flags,
3437                                     &config->cfg_config_u.cfg_route.
3438                                         rtr_priority);
3439                 mutex_unlock(&the_lnet.ln_api_mutex);
3440                 return rc;
3441
3442         case IOC_LIBCFS_GET_LOCAL_NI: {
3443                 struct lnet_ioctl_config_ni *cfg_ni;
3444                 struct lnet_ioctl_config_lnd_tunables *tun = NULL;
3445                 struct lnet_ioctl_element_stats *stats;
3446                 __u32 tun_size;
3447
3448                 cfg_ni = arg;
3449
3450                 /* get the tunables if they are available */
3451                 if (cfg_ni->lic_cfg_hdr.ioc_len <
3452                     sizeof(*cfg_ni) + sizeof(*stats) + sizeof(*tun))
3453                         return -EINVAL;
3454
3455                 stats = (struct lnet_ioctl_element_stats *)
3456                         cfg_ni->lic_bulk;
3457                 tun = (struct lnet_ioctl_config_lnd_tunables *)
3458                                 (cfg_ni->lic_bulk + sizeof(*stats));
3459
3460                 tun_size = cfg_ni->lic_cfg_hdr.ioc_len - sizeof(*cfg_ni) -
3461                         sizeof(*stats);
3462
3463                 mutex_lock(&the_lnet.ln_api_mutex);
3464                 rc = lnet_get_ni_config(cfg_ni, tun, stats, tun_size);
3465                 mutex_unlock(&the_lnet.ln_api_mutex);
3466                 return rc;
3467         }
3468
3469         case IOC_LIBCFS_GET_LOCAL_NI_MSG_STATS: {
3470                 struct lnet_ioctl_element_msg_stats *msg_stats = arg;
3471
3472                 if (msg_stats->im_hdr.ioc_len != sizeof(*msg_stats))
3473                         return -EINVAL;
3474
3475                 mutex_lock(&the_lnet.ln_api_mutex);
3476                 rc = lnet_get_ni_stats(msg_stats);
3477                 mutex_unlock(&the_lnet.ln_api_mutex);
3478
3479                 return rc;
3480         }
3481
3482         case IOC_LIBCFS_GET_NET: {
3483                 size_t total = sizeof(*config) +
3484                                sizeof(struct lnet_ioctl_net_config);
3485                 config = arg;
3486
3487                 if (config->cfg_hdr.ioc_len < total)
3488                         return -EINVAL;
3489
3490                 mutex_lock(&the_lnet.ln_api_mutex);
3491                 rc = lnet_get_net_config(config);
3492                 mutex_unlock(&the_lnet.ln_api_mutex);
3493                 return rc;
3494         }
3495
3496         case IOC_LIBCFS_GET_LNET_STATS:
3497         {
3498                 struct lnet_ioctl_lnet_stats *lnet_stats = arg;
3499
3500                 if (lnet_stats->st_hdr.ioc_len < sizeof(*lnet_stats))
3501                         return -EINVAL;
3502
3503                 mutex_lock(&the_lnet.ln_api_mutex);
3504                 lnet_counters_get(&lnet_stats->st_cntrs);
3505                 mutex_unlock(&the_lnet.ln_api_mutex);
3506                 return 0;
3507         }
3508
3509         case IOC_LIBCFS_CONFIG_RTR:
3510                 config = arg;
3511
3512                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3513                         return -EINVAL;
3514
3515                 mutex_lock(&the_lnet.ln_api_mutex);
3516                 if (config->cfg_config_u.cfg_buffers.buf_enable) {
3517                         rc = lnet_rtrpools_enable();
3518                         mutex_unlock(&the_lnet.ln_api_mutex);
3519                         return rc;
3520                 }
3521                 lnet_rtrpools_disable();
3522                 mutex_unlock(&the_lnet.ln_api_mutex);
3523                 return 0;
3524
3525         case IOC_LIBCFS_ADD_BUF:
3526                 config = arg;
3527
3528                 if (config->cfg_hdr.ioc_len < sizeof(*config))
3529                         return -EINVAL;
3530
3531                 mutex_lock(&the_lnet.ln_api_mutex);
3532                 rc = lnet_rtrpools_adjust(config->cfg_config_u.cfg_buffers.
3533                                                 buf_tiny,
3534                                           config->cfg_config_u.cfg_buffers.
3535                                                 buf_small,
3536                                           config->cfg_config_u.cfg_buffers.
3537                                                 buf_large);
3538                 mutex_unlock(&the_lnet.ln_api_mutex);
3539                 return rc;
3540
3541         case IOC_LIBCFS_SET_NUMA_RANGE: {
3542                 struct lnet_ioctl_set_value *numa;
3543                 numa = arg;
3544                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
3545                         return -EINVAL;
3546                 lnet_net_lock(LNET_LOCK_EX);
3547                 lnet_numa_range = numa->sv_value;
3548                 lnet_net_unlock(LNET_LOCK_EX);
3549                 return 0;
3550         }
3551
3552         case IOC_LIBCFS_GET_NUMA_RANGE: {
3553                 struct lnet_ioctl_set_value *numa;
3554                 numa = arg;
3555                 if (numa->sv_hdr.ioc_len != sizeof(*numa))
3556                         return -EINVAL;
3557                 numa->sv_value = lnet_numa_range;
3558                 return 0;
3559         }
3560
3561         case IOC_LIBCFS_GET_BUF: {
3562                 struct lnet_ioctl_pool_cfg *pool_cfg;
3563                 size_t total = sizeof(*config) + sizeof(*pool_cfg);
3564
3565                 config = arg;
3566
3567                 if (config->cfg_hdr.ioc_len < total)
3568                         return -EINVAL;
3569
3570                 pool_cfg = (struct lnet_ioctl_pool_cfg *)config->cfg_bulk;
3571
3572                 mutex_lock(&the_lnet.ln_api_mutex);
3573                 rc = lnet_get_rtr_pool_cfg(config->cfg_count, pool_cfg);
3574                 mutex_unlock(&the_lnet.ln_api_mutex);
3575                 return rc;
3576         }
3577
3578         case IOC_LIBCFS_GET_LOCAL_HSTATS: {
3579                 struct lnet_ioctl_local_ni_hstats *stats = arg;
3580
3581                 if (stats->hlni_hdr.ioc_len < sizeof(*stats))
3582                         return -EINVAL;
3583
3584                 mutex_lock(&the_lnet.ln_api_mutex);
3585                 rc = lnet_get_local_ni_hstats(stats);
3586                 mutex_unlock(&the_lnet.ln_api_mutex);
3587
3588                 return rc;
3589         }
3590
3591         case IOC_LIBCFS_GET_RECOVERY_QUEUE: {
3592                 struct lnet_ioctl_recovery_list *list = arg;
3593                 if (list->rlst_hdr.ioc_len < sizeof(*list))
3594                         return -EINVAL;
3595
3596                 mutex_lock(&the_lnet.ln_api_mutex);
3597                 if (list->rlst_type == LNET_HEALTH_TYPE_LOCAL_NI)
3598                         rc = lnet_get_local_ni_recovery_list(list);
3599                 else
3600                         rc = lnet_get_peer_ni_recovery_list(list);
3601                 mutex_unlock(&the_lnet.ln_api_mutex);
3602                 return rc;
3603         }
3604
3605         case IOC_LIBCFS_ADD_PEER_NI: {
3606                 struct lnet_ioctl_peer_cfg *cfg = arg;
3607
3608                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3609                         return -EINVAL;
3610
3611                 mutex_lock(&the_lnet.ln_api_mutex);
3612                 rc = lnet_add_peer_ni(cfg->prcfg_prim_nid,
3613                                       cfg->prcfg_cfg_nid,
3614                                       cfg->prcfg_mr);
3615                 mutex_unlock(&the_lnet.ln_api_mutex);
3616                 return rc;
3617         }
3618
3619         case IOC_LIBCFS_DEL_PEER_NI: {
3620                 struct lnet_ioctl_peer_cfg *cfg = arg;
3621
3622                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3623                         return -EINVAL;
3624
3625                 mutex_lock(&the_lnet.ln_api_mutex);
3626                 rc = lnet_del_peer_ni(cfg->prcfg_prim_nid,
3627                                       cfg->prcfg_cfg_nid);
3628                 mutex_unlock(&the_lnet.ln_api_mutex);
3629                 return rc;
3630         }
3631
3632         case IOC_LIBCFS_GET_PEER_INFO: {
3633                 struct lnet_ioctl_peer *peer_info = arg;
3634
3635                 if (peer_info->pr_hdr.ioc_len < sizeof(*peer_info))
3636                         return -EINVAL;
3637
3638                 mutex_lock(&the_lnet.ln_api_mutex);
3639                 rc = lnet_get_peer_ni_info(
3640                    peer_info->pr_count,
3641                    &peer_info->pr_nid,
3642                    peer_info->pr_lnd_u.pr_peer_credits.cr_aliveness,
3643                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ncpt,
3644                    &peer_info->pr_lnd_u.pr_peer_credits.cr_refcount,
3645                    &peer_info->pr_lnd_u.pr_peer_credits.cr_ni_peer_tx_credits,
3646                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_credits,
3647                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_rtr_credits,
3648                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_min_tx_credits,
3649                    &peer_info->pr_lnd_u.pr_peer_credits.cr_peer_tx_qnob);
3650                 mutex_unlock(&the_lnet.ln_api_mutex);
3651                 return rc;
3652         }
3653
3654         case IOC_LIBCFS_GET_PEER_NI: {
3655                 struct lnet_ioctl_peer_cfg *cfg = arg;
3656
3657                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3658                         return -EINVAL;
3659
3660                 mutex_lock(&the_lnet.ln_api_mutex);
3661                 rc = lnet_get_peer_info(cfg,
3662                                         (void __user *)cfg->prcfg_bulk);
3663                 mutex_unlock(&the_lnet.ln_api_mutex);
3664                 return rc;
3665         }
3666
3667         case IOC_LIBCFS_GET_PEER_LIST: {
3668                 struct lnet_ioctl_peer_cfg *cfg = arg;
3669
3670                 if (cfg->prcfg_hdr.ioc_len < sizeof(*cfg))
3671                         return -EINVAL;
3672
3673                 mutex_lock(&the_lnet.ln_api_mutex);
3674                 rc = lnet_get_peer_list(&cfg->prcfg_count, &cfg->prcfg_size,
3675                                 (struct lnet_process_id __user *)cfg->prcfg_bulk);
3676                 mutex_unlock(&the_lnet.ln_api_mutex);
3677                 return rc;
3678         }
3679
3680         case IOC_LIBCFS_SET_HEALHV: {
3681                 struct lnet_ioctl_reset_health_cfg *cfg = arg;
3682                 int value;
3683                 if (cfg->rh_hdr.ioc_len < sizeof(*cfg))
3684                         return -EINVAL;
3685                 if (cfg->rh_value < 0 ||
3686                     cfg->rh_value > LNET_MAX_HEALTH_VALUE)
3687                         value = LNET_MAX_HEALTH_VALUE;
3688                 else
3689                         value = cfg->rh_value;
3690                 CDEBUG(D_NET, "Manually setting healthv to %d for %s:%s. all = %d\n",
3691                        value, (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI) ?
3692                        "local" : "peer", libcfs_nid2str(cfg->rh_nid), cfg->rh_all);
3693                 mutex_lock(&the_lnet.ln_api_mutex);
3694                 if (cfg->rh_type == LNET_HEALTH_TYPE_LOCAL_NI)
3695                         lnet_ni_set_healthv(cfg->rh_nid, value,
3696                                              cfg->rh_all);
3697                 else
3698                         lnet_peer_ni_set_healthv(cfg->rh_nid, value,
3699                                                   cfg->rh_all);
3700                 mutex_unlock(&the_lnet.ln_api_mutex);
3701                 return 0;
3702         }
3703
3704         case IOC_LIBCFS_NOTIFY_ROUTER: {
3705                 time64_t deadline = ktime_get_real_seconds() - data->ioc_u64[0];
3706
3707                 /* The deadline passed in by the user should be some time in
3708                  * seconds in the future since the UNIX epoch. We have to map
3709                  * that deadline to the wall clock.
3710                  */
3711                 deadline += ktime_get_seconds();
3712                 return lnet_notify(NULL, data->ioc_nid, data->ioc_flags,
3713                                    deadline);
3714         }
3715
3716         case IOC_LIBCFS_LNET_DIST:
3717                 rc = LNetDist(data->ioc_nid, &data->ioc_nid, &data->ioc_u32[1]);
3718                 if (rc < 0 && rc != -EHOSTUNREACH)
3719                         return rc;
3720
3721                 data->ioc_u32[0] = rc;
3722                 return 0;
3723
3724         case IOC_LIBCFS_TESTPROTOCOMPAT:
3725                 lnet_net_lock(LNET_LOCK_EX);
3726                 the_lnet.ln_testprotocompat = data->ioc_flags;
3727                 lnet_net_unlock(LNET_LOCK_EX);
3728                 return 0;
3729
3730         case IOC_LIBCFS_LNET_FAULT:
3731                 return lnet_fault_ctl(data->ioc_flags, data);
3732
3733         case IOC_LIBCFS_PING: {
3734                 signed long timeout;
3735
3736                 id.nid = data->ioc_nid;
3737                 id.pid = data->ioc_u32[0];
3738
3739                 /* If timeout is negative then set default of 3 minutes */
3740                 if (((s32)data->ioc_u32[1] <= 0) ||
3741                     data->ioc_u32[1] > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
3742                         timeout = msecs_to_jiffies(DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC);
3743                 else
3744                         timeout = msecs_to_jiffies(data->ioc_u32[1]);
3745
3746                 rc = lnet_ping(id, timeout, data->ioc_pbuf1,
3747                                data->ioc_plen1 / sizeof(struct lnet_process_id));
3748
3749                 if (rc < 0)
3750                         return rc;
3751
3752                 data->ioc_count = rc;
3753                 return 0;
3754         }
3755
3756         case IOC_LIBCFS_PING_PEER: {
3757                 struct lnet_ioctl_ping_data *ping = arg;
3758                 struct lnet_peer *lp;
3759                 signed long timeout;
3760
3761                 /* If timeout is negative then set default of 3 minutes */
3762                 if (((s32)ping->op_param) <= 0 ||
3763                     ping->op_param > (DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC))
3764                         timeout = msecs_to_jiffies(DEFAULT_PEER_TIMEOUT * MSEC_PER_SEC);
3765                 else
3766                         timeout = msecs_to_jiffies(ping->op_param);
3767
3768                 rc = lnet_ping(ping->ping_id, timeout,
3769                                ping->ping_buf,
3770                                ping->ping_count);
3771                 if (rc < 0)
3772                         return rc;
3773
3774                 mutex_lock(&the_lnet.ln_api_mutex);
3775                 lp = lnet_find_peer(ping->ping_id.nid);
3776                 if (lp) {
3777                         ping->ping_id.nid = lp->lp_primary_nid;
3778                         ping->mr_info = lnet_peer_is_multi_rail(lp);
3779                         lnet_peer_decref_locked(lp);
3780                 }
3781                 mutex_unlock(&the_lnet.ln_api_mutex);
3782
3783                 ping->ping_count = rc;
3784                 return 0;
3785         }
3786
3787         case IOC_LIBCFS_DISCOVER: {
3788                 struct lnet_ioctl_ping_data *discover = arg;
3789                 struct lnet_peer *lp;
3790
3791                 rc = lnet_discover(discover->ping_id, discover->op_param,
3792                                    discover->ping_buf,
3793                                    discover->ping_count);
3794                 if (rc < 0)
3795                         return rc;
3796
3797                 mutex_lock(&the_lnet.ln_api_mutex);
3798                 lp = lnet_find_peer(discover->ping_id.nid);
3799                 if (lp) {
3800                         discover->ping_id.nid = lp->lp_primary_nid;
3801                         discover->mr_info = lnet_peer_is_multi_rail(lp);
3802                         lnet_peer_decref_locked(lp);
3803                 }
3804                 mutex_unlock(&the_lnet.ln_api_mutex);
3805
3806                 discover->ping_count = rc;
3807                 return 0;
3808         }
3809
3810         default:
3811                 ni = lnet_net2ni_addref(data->ioc_net);
3812                 if (ni == NULL)
3813                         return -EINVAL;
3814
3815                 if (ni->ni_net->net_lnd->lnd_ctl == NULL)
3816                         rc = -EINVAL;
3817                 else
3818                         rc = ni->ni_net->net_lnd->lnd_ctl(ni, cmd, arg);
3819
3820                 lnet_ni_decref(ni);
3821                 return rc;
3822         }
3823         /* not reached */
3824 }
3825 EXPORT_SYMBOL(LNetCtl);
3826
3827 void LNetDebugPeer(struct lnet_process_id id)
3828 {
3829         lnet_debug_peer(id.nid);
3830 }
3831 EXPORT_SYMBOL(LNetDebugPeer);
3832
3833 /**
3834  * Determine if the specified peer \a nid is on the local node.
3835  *
3836  * \param nid   peer nid to check
3837  *
3838  * \retval true         If peer NID is on the local node.
3839  * \retval false        If peer NID is not on the local node.
3840  */
3841 bool LNetIsPeerLocal(lnet_nid_t nid)
3842 {
3843         struct lnet_net *net;
3844         struct lnet_ni *ni;
3845         int cpt;
3846
3847         cpt = lnet_net_lock_current();
3848         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3849                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3850                         if (ni->ni_nid == nid) {
3851                                 lnet_net_unlock(cpt);
3852                                 return true;
3853                         }
3854                 }
3855         }
3856         lnet_net_unlock(cpt);
3857
3858         return false;
3859 }
3860 EXPORT_SYMBOL(LNetIsPeerLocal);
3861
3862 /**
3863  * Retrieve the struct lnet_process_id ID of LNet interface at \a index.
3864  * Note that all interfaces share a same PID, as requested by LNetNIInit().
3865  *
3866  * \param index Index of the interface to look up.
3867  * \param id On successful return, this location will hold the
3868  * struct lnet_process_id ID of the interface.
3869  *
3870  * \retval 0 If an interface exists at \a index.
3871  * \retval -ENOENT If no interface has been found.
3872  */
3873 int
3874 LNetGetId(unsigned int index, struct lnet_process_id *id)
3875 {
3876         struct lnet_ni   *ni;
3877         struct lnet_net  *net;
3878         int               cpt;
3879         int               rc = -ENOENT;
3880
3881         LASSERT(the_lnet.ln_refcount > 0);
3882
3883         cpt = lnet_net_lock_current();
3884
3885         list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
3886                 list_for_each_entry(ni, &net->net_ni_list, ni_netlist) {
3887                         if (index-- != 0)
3888                                 continue;
3889
3890                         id->nid = ni->ni_nid;
3891                         id->pid = the_lnet.ln_pid;
3892                         rc = 0;
3893                         break;
3894                 }
3895         }
3896
3897         lnet_net_unlock(cpt);
3898         return rc;
3899 }
3900 EXPORT_SYMBOL(LNetGetId);
3901
3902 static int lnet_ping(struct lnet_process_id id, signed long timeout,
3903                      struct lnet_process_id __user *ids, int n_ids)
3904 {
3905         struct lnet_handle_eq eqh;
3906         struct lnet_handle_md mdh;
3907         struct lnet_event event;
3908         struct lnet_md md = { NULL };
3909         int which;
3910         int unlinked = 0;
3911         int replied = 0;
3912         const signed long a_long_time = msecs_to_jiffies(60 * MSEC_PER_SEC);
3913         struct lnet_ping_buffer *pbuf;
3914         struct lnet_process_id tmpid;
3915         int i;
3916         int nob;
3917         int rc;
3918         int rc2;
3919         sigset_t blocked;
3920
3921         /* n_ids limit is arbitrary */
3922         if (n_ids <= 0 || id.nid == LNET_NID_ANY)
3923                 return -EINVAL;
3924
3925         /*
3926          * if the user buffer has more space than the lnet_interfaces_max
3927          * then only fill it up to lnet_interfaces_max
3928          */
3929         if (n_ids > lnet_interfaces_max)
3930                 n_ids = lnet_interfaces_max;
3931
3932         if (id.pid == LNET_PID_ANY)
3933                 id.pid = LNET_PID_LUSTRE;
3934
3935         pbuf = lnet_ping_buffer_alloc(n_ids, GFP_NOFS);
3936         if (!pbuf)
3937                 return -ENOMEM;
3938
3939         /* NB 2 events max (including any unlink event) */
3940         rc = LNetEQAlloc(2, LNET_EQ_HANDLER_NONE, &eqh);
3941         if (rc != 0) {
3942                 CERROR("Can't allocate EQ: %d\n", rc);
3943                 goto fail_ping_buffer_decref;
3944         }
3945
3946         /* initialize md content */
3947         md.start     = &pbuf->pb_info;
3948         md.length    = LNET_PING_INFO_SIZE(n_ids);
3949         md.threshold = 2; /* GET/REPLY */
3950         md.max_size  = 0;
3951         md.options   = LNET_MD_TRUNCATE;
3952         md.user_ptr  = NULL;
3953         md.eq_handle = eqh;
3954
3955         rc = LNetMDBind(md, LNET_UNLINK, &mdh);
3956         if (rc != 0) {
3957                 CERROR("Can't bind MD: %d\n", rc);
3958                 goto fail_free_eq;
3959         }
3960
3961         rc = LNetGet(LNET_NID_ANY, mdh, id,
3962                      LNET_RESERVED_PORTAL,
3963                      LNET_PROTO_PING_MATCHBITS, 0, false);
3964
3965         if (rc != 0) {
3966                 /* Don't CERROR; this could be deliberate! */
3967                 rc2 = LNetMDUnlink(mdh);
3968                 LASSERT(rc2 == 0);
3969
3970                 /* NB must wait for the UNLINK event below... */
3971                 unlinked = 1;
3972                 timeout = a_long_time;
3973         }
3974
3975         do {
3976                 /* MUST block for unlink to complete */
3977                 if (unlinked)
3978                         blocked = cfs_block_allsigs();
3979
3980                 rc2 = LNetEQPoll(&eqh, 1, timeout, &event, &which);
3981
3982                 if (unlinked)
3983                         cfs_restore_sigs(blocked);
3984
3985                 CDEBUG(D_NET, "poll %d(%d %d)%s\n", rc2,
3986                        (rc2 <= 0) ? -1 : event.type,
3987                        (rc2 <= 0) ? -1 : event.status,
3988                        (rc2 > 0 && event.unlinked) ? " unlinked" : "");
3989
3990                 LASSERT(rc2 != -EOVERFLOW);     /* can't miss anything */
3991
3992                 if (rc2 <= 0 || event.status != 0) {
3993                         /* timeout or error */
3994                         if (!replied && rc == 0)
3995                                 rc = (rc2 < 0) ? rc2 :
3996                                      (rc2 == 0) ? -ETIMEDOUT :
3997                                      event.status;
3998
3999                         if (!unlinked) {
4000                                 /* Ensure completion in finite time... */
4001                                 LNetMDUnlink(mdh);
4002                                 /* No assertion (racing with network) */
4003                                 unlinked = 1;
4004                                 timeout = a_long_time;
4005                         } else if (rc2 == 0) {
4006                                 /* timed out waiting for unlink */
4007                                 CWARN("ping %s: late network completion\n",
4008                                       libcfs_id2str(id));
4009                         }
4010                 } else if (event.type == LNET_EVENT_REPLY) {
4011                         replied = 1;
4012                         rc = event.mlength;
4013                 }
4014         } while (rc2 <= 0 || !event.unlinked);
4015
4016         if (!replied) {
4017                 if (rc >= 0)
4018                         CWARN("%s: Unexpected rc >= 0 but no reply!\n",
4019                               libcfs_id2str(id));
4020                 rc = -EIO;
4021                 goto fail_free_eq;
4022         }
4023
4024         nob = rc;
4025         LASSERT(nob >= 0 && nob <= LNET_PING_INFO_SIZE(n_ids));
4026
4027         rc = -EPROTO;           /* if I can't parse... */
4028
4029         if (nob < 8) {
4030                 CERROR("%s: ping info too short %d\n",
4031                        libcfs_id2str(id), nob);
4032                 goto fail_free_eq;
4033         }
4034
4035         if (pbuf->pb_info.pi_magic == __swab32(LNET_PROTO_PING_MAGIC)) {
4036                 lnet_swap_pinginfo(pbuf);
4037         } else if (pbuf->pb_info.pi_magic != LNET_PROTO_PING_MAGIC) {
4038                 CERROR("%s: Unexpected magic %08x\n",
4039                        libcfs_id2str(id), pbuf->pb_info.pi_magic);
4040                 goto fail_free_eq;
4041         }
4042
4043         if ((pbuf->pb_info.pi_features & LNET_PING_FEAT_NI_STATUS) == 0) {
4044                 CERROR("%s: ping w/o NI status: 0x%x\n",
4045                        libcfs_id2str(id), pbuf->pb_info.pi_features);
4046                 goto fail_free_eq;
4047         }
4048
4049         if (nob < LNET_PING_INFO_SIZE(0)) {
4050                 CERROR("%s: Short reply %d(%d min)\n",
4051                        libcfs_id2str(id),
4052                        nob, (int)LNET_PING_INFO_SIZE(0));
4053                 goto fail_free_eq;
4054         }
4055
4056         if (pbuf->pb_info.pi_nnis < n_ids)
4057                 n_ids = pbuf->pb_info.pi_nnis;
4058
4059         if (nob < LNET_PING_INFO_SIZE(n_ids)) {
4060                 CERROR("%s: Short reply %d(%d expected)\n",
4061                        libcfs_id2str(id),
4062                        nob, (int)LNET_PING_INFO_SIZE(n_ids));
4063                 goto fail_free_eq;
4064         }
4065
4066         rc = -EFAULT;           /* if I segv in copy_to_user()... */
4067
4068         memset(&tmpid, 0, sizeof(tmpid));
4069         for (i = 0; i < n_ids; i++) {
4070                 tmpid.pid = pbuf->pb_info.pi_pid;
4071                 tmpid.nid = pbuf->pb_info.pi_ni[i].ns_nid;
4072                 if (copy_to_user(&ids[i], &tmpid, sizeof(tmpid)))
4073                         goto fail_free_eq;
4074         }
4075         rc = pbuf->pb_info.pi_nnis;
4076
4077  fail_free_eq:
4078         rc2 = LNetEQFree(eqh);
4079         if (rc2 != 0)
4080                 CERROR("rc2 %d\n", rc2);
4081         LASSERT(rc2 == 0);
4082
4083  fail_ping_buffer_decref:
4084         lnet_ping_buffer_decref(pbuf);
4085         return rc;
4086 }
4087
4088 static int
4089 lnet_discover(struct lnet_process_id id, __u32 force,
4090               struct lnet_process_id __user *ids, int n_ids)
4091 {
4092         struct lnet_peer_ni *lpni;
4093         struct lnet_peer_ni *p;
4094         struct lnet_peer *lp;
4095         struct lnet_process_id *buf;
4096         int cpt;
4097         int i;
4098         int rc;
4099         int max_intf = lnet_interfaces_max;
4100         size_t buf_size;
4101
4102         if (n_ids <= 0 ||
4103             id.nid == LNET_NID_ANY)
4104                 return -EINVAL;
4105
4106         if (id.pid == LNET_PID_ANY)
4107                 id.pid = LNET_PID_LUSTRE;
4108
4109         /*
4110          * if the user buffer has more space than the max_intf
4111          * then only fill it up to max_intf
4112          */
4113         if (n_ids > max_intf)
4114                 n_ids = max_intf;
4115
4116         buf_size = n_ids * sizeof(*buf);
4117
4118         LIBCFS_ALLOC(buf, buf_size);
4119         if (!buf)
4120                 return -ENOMEM;
4121
4122         cpt = lnet_net_lock_current();
4123         lpni = lnet_nid2peerni_locked(id.nid, LNET_NID_ANY, cpt);
4124         if (IS_ERR(lpni)) {
4125                 rc = PTR_ERR(lpni);
4126                 goto out;
4127         }
4128
4129         /*
4130          * Clearing the NIDS_UPTODATE flag ensures the peer will
4131          * be discovered, provided discovery has not been disabled.
4132          */
4133         lp = lpni->lpni_peer_net->lpn_peer;
4134         spin_lock(&lp->lp_lock);
4135         lp->lp_state &= ~LNET_PEER_NIDS_UPTODATE;
4136         /* If the force flag is set, force a PING and PUSH as well. */
4137         if (force)
4138                 lp->lp_state |= LNET_PEER_FORCE_PING | LNET_PEER_FORCE_PUSH;
4139         spin_unlock(&lp->lp_lock);
4140         rc = lnet_discover_peer_locked(lpni, cpt, true);
4141         if (rc)
4142                 goto out_decref;
4143
4144         /* Peer may have changed. */
4145         lp = lpni->lpni_peer_net->lpn_peer;
4146         if (lp->lp_nnis < n_ids)
4147                 n_ids = lp->lp_nnis;
4148
4149         i = 0;
4150         p = NULL;
4151         while ((p = lnet_get_next_peer_ni_locked(lp, NULL, p)) != NULL) {
4152                 buf[i].pid = id.pid;
4153                 buf[i].nid = p->lpni_nid;
4154                 if (++i >= n_ids)
4155                         break;
4156         }
4157
4158         lnet_net_unlock(cpt);
4159
4160         rc = -EFAULT;
4161         if (copy_to_user(ids, buf, n_ids * sizeof(*buf)))
4162                 goto out_relock;
4163         rc = n_ids;
4164 out_relock:
4165         lnet_net_lock(cpt);
4166 out_decref:
4167         lnet_peer_ni_decref_locked(lpni);
4168 out:
4169         lnet_net_unlock(cpt);
4170
4171         LIBCFS_FREE(buf, buf_size);
4172
4173         return rc;
4174 }
4175
4176 /**
4177  * Retrieve peer discovery status.
4178  *
4179  * \retval 1 if lnet_peer_discovery_disabled is 0
4180  * \retval 0 if lnet_peer_discovery_disabled is 1
4181  */
4182 int
4183 LNetGetPeerDiscoveryStatus(void)
4184 {
4185         return !lnet_peer_discovery_disabled;
4186 }
4187 EXPORT_SYMBOL(LNetGetPeerDiscoveryStatus);