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